Smart-building forecasting energy using spatio-temporal CNN-LSTM approach: Case of smart hospital

Abstract

Capturing complex spatio-temporal dependencies in energy data of connected infrastructures remains a significant challenge. Indeed, most of the proposed approaches focus on either spatial or temporal dependencies, without effectively modeling their joint influence. This paper presents a comprehensive spatio-temporal framework for energy forecasting in smart buildings, with structured representation of entities, events and relationships in the IoT ecosystem. For energy consumption prediction, we propose a hybrid CNN2D-LSTM model that combines convolutional layers to extract spatial features from multi-variable data with LSTM layers to capture temporal dependencies. The proposed model is instantiated on a smart hospital building which is characterized by continuous operation, high energy consumption and heterogeneous medical equipment. Under realistic constraints, a simulation-based dataset was generated and used for evaluation, reflecting dynamic behaviors in the event-driven system and realistic scenarios of energy consumption. The experimental evaluation demonstrates that the proposed model outperforms conventional approaches such as ANN, RNN, CNN, and LSTM and also hybrid models from the literature in terms of prediction accuracy. Comparisons were also conducted under different operating scenarios, including normal consumption, overconsumption, and under-consumption conditions, while maintaining a reasonable processing time. Beyond prediction, the proposed framework supports Green IoT principles by enabling energy management systems to monitor consumption, detect anomalies, and support proactive decision-making. It provides a scalable foundation for energy management in complex smart buildings.

Keywords

Smart building energy consumption machine-learning spatio-temporal forecasting smart Hospital Green IoT CNN2D-LSTM multi-layer architecture

1. Introduction

The increasing integration of digital technologies into buildings has given rise to smart buildings, where IoT-based sensing, communication, and control are leveraged to optimize resource consumption and enhance user comfort. In this context, the Green Internet of Things (Green IoT) has emerged as a central paradigm, emphasizing energy-efficient data acquisition and intelligent resource management to minimize the environmental footprint of connected infrastructures.¹ Smart buildings represent an essential convergence between technological innovation and sustainable development, offering effective solutions for resource management in increasingly complex environments.²

Energy management in smart buildings is a complex task influenced by technological, behavioral, and contextual factors.³ One of the main difficulties lies in the high heterogeneity of energy sources and loads, as smart buildings often integrate grid electricity, solar panels, batteries or backup generators alongside diverse energy-consuming systems such as HVAC, lighting, elevators, and IoT devices.⁴ Coordinating these heterogeneous components in real time to ensure both efficiency and reliability is a significant technical challenge. Additionally, the dynamic and unpredictable nature of energy demand, influenced by occupancy patterns, weather conditions, and operational schedules, can lead to sudden peaks that increase operational costs and affect system stability.⁵ Another major issue concerns the management of critical and non-critical equipment, especially in environments where life-support systems or refrigeration units must remain operational at all times. Balancing energy allocation without compromising reliability is therefore crucial. Smart buildings must also contend with event-driven and contextual conditions such as heatwaves or equipment failures that can drastically alter energy consumption profiles, demanding adaptability and resilience from building systems. Finally, the integration of sustainable and green IoT objectives intensifies these challenges, as buildings are now expected to minimize waste, optimize energy use, and contribute to carbon neutrality goals, all while maintaining comfort, resilience, and operational performance. These challenges are even more pronounced in critical infrastructures such as hospitals, where ensuring uninterrupted operation of life-support systems and medical equipment is vital. Smart hospitals thus constitute an ideal case study to evaluate intelligent energy management strategies. Nevertheless, while the hospital scenario is particularly relevant, the proposed approach remains general and can be extended to any smart building context with strong spatial and temporal interdependencies.^2,6

Predicting energy consumption is a crucial step towards proactive energy management. According to the IEA Tracking Report Buildings, the building sector accounted for approximately 30% of global final energy consumption and around 27% of total energy-related CO $_{2}$ emissions in recent years.⁷ A review of existing literature on smart building energy forecasting as Mariano-Hernández et al.,⁸ Divina et al.,⁹ Zini and Carcasci¹⁰ reveals several limitations. Most approaches focus on either spatial or temporal dependencies, without effectively modeling their joint influence, which limits their ability to capture complex interactions within buildings. In addition, many studies rely on simplified datasets that do not reflect dynamic and event-driven operating conditions, leading to limited robustness in real-world scenarios. Furthermore, although hybrid deep learning models such as CNN–LSTM have been explored,¹¹ they are often not specifically designed to capture the structural complexity of smart buildings, including multi-zone configurations and heterogeneous equipment interactions.

To overcome these limitations, we propose a conceptual and spatio-temporal framework for energy consumption prediction in smart buildings. The conceptual framework provides a structured representation of entities, such as sensors, actuators, equipment, and events, and their relationships across the different layers of an IoT ecosystem. It bridges the gap between physical infrastructure and intelligent analytics, ensuring consistency between data collection, prediction, and control. The contributions of this work are as follows:

–
A conceptual framework that models interactions between IoT components, building entities, and energy management processes, ensuring consistency between data acquisition and predictive analysis.
–
A hybrid CNN2D-LSTM model designed to jointly capture spatial correlations across building zones and temporal dependencies in energy consumption data.
–
The use of a multi-variable, event-driven dataset reflecting realistic smart building conditions to evaluate the robustness of the proposed approach.

The novelty of this work lies in their integration within a structured smart building framework and their adaptation to multi-zone, event-driven environments, enabling more realistic and robust energy consumption modeling.

The rest of the paper is organized as follows: Section 2 introduces basic concepts related to smart buildings and energy management challenges. Section 3 reviews existing works in the domain and provides a comparison and a discussion about existing approaches, while clearly identifying the main gaps in the domain. Section 4 presents conceptual elements of the framework and implementation details, mainly the detailed architecture and functioning of the hybrid predictive model. Section 5 presents conceptual and technical elements of our case study relative to a smart hospital, and describes data generation rules and constraints, aiming to produce realistic datasets for training and evaluating the proposed prediction model. Section 6 focuses on the evaluation of our solution through experimental results. Finally, Section 7 concludes the paper and evokes future works.
2. Basic concepts

In this section, we introduce some notions relative to smart buildings and energy consumption management.

2.1. Smart building

Smart buildings integrate advanced technologies such as the Internet of Things (IoT), smart sensors and automation systems to optimize resources management and improve the overall performance of the infrastructure.^3,12 By connecting various devices and centralizing data collection, smart buildings enable optimal energy allocation and real-time response to variations in demand; this enhances energy efficiency and reduces the carbon footprint. Several categories of smart buildings can be found, among them:

Smart Public Infrastructure: includes government buildings, transportation hubs, libraries, and other public service facilities that integrate IoT-based energy management, security systems, and sustainability measures.¹³

Smart Residential Buildings: refer to homes integrating smart equipment such as thermostats, energy-efficient appliances, automated lighting, and home security systems, in order to enhance comfort, security, and energy efficiency.¹⁴

Smart Commercial Buildings: offices, shopping malls, and hotels employ smart energy management systems and predictive maintenance to optimize operational costs, enabling real-time monitoring and data-driven decision-making to enhance energy efficiency and sustainability.¹⁵

Smart Industrial Buildings: factories leverage industrial IoT (IIoT) solutions for process automation, predictive maintenance, and intelligent resource allocation. Energy management in these buildings focuses on optimizing machinery operations and minimizing energy waste through real-time monitoring and smart grid interaction.¹⁶

Smart Educational Buildings: schools, universities, and research institutions implement smart solutions for energy efficiency, environmental monitoring, and adaptive learning spaces. IoT-driven automation regulates lighting, ventilation, and classroom conditions based on environmental conditions.¹⁷

Smart Hospitals: incorporate IoT-based monitoring and control systems to ensure the efficient operation of medical equipment, climate control, and emergency power supply. Due to their high and continuous energy demands, smart hospitals require advanced energy optimization strategies to maintain operational reliability, minimize costs, and enhance patient care.¹⁰

2.2. Energy consumption challenges in smart buildings

Managing energy consumption in smart buildings presents several challenges due to the complexity and dynamic nature of these environments. The main challenges are summarized in the following points:

–
Smart buildings use a wide range of IoT devices (sensors, actuators, smart meters, …), which generate large volumes of heterogeneous data. Also, the coexistence of multiple energy sources (electricity, solar panels, batteries, etc.) complicates energy distribution and management.
–
Energy demand varies based on occupancy, weather conditions, and operational requirements. Sudden consumption peaks, especially during extreme weather or high occupancy, require dynamic load balancing to avoid wasted energy and network overload.
–
The coexistence of critical systems (life-support machines, emergency power supply) and non-critical loads (lighting, HVAC) requires uninterrupted energy supply to critical equipment while optimizing overall consumption.
–
Adaptive energy management systems must continuously learn and adjust to user preferences, occupancy patterns, and external conditions to enhance comfort while minimizing waste.

2.3. Green IoT and smart buildings

Green IoT refers to the integration of energy-efficient principles and sustainability objectives into IoT systems, aiming to reduce the overall energy footprint of IoT devices, communication processes, and data management infrastructures, while maintaining high levels of connectivity, intelligence, and responsiveness.¹⁸ In the context of smart buildings, Green IoT leverages real-time data collection and advanced analytics to optimize energy consumption, adapt operations to occupants’ needs, and support the integration of diverse energy sources. This synergy promotes efficient resource allocation, reduces operational costs and preserves the environment.²

Smart hospitals are a specific type of smart buildings, where advanced IoT and automation technologies are mobilized to meet the demands of the healthcare sector. By integrating intelligent sensors and proactive energy management systems, hospitals optimize energy consumption to ensure the uninterrupted operation of medical equipment and guarantee patient comfort and safety.

3. Literature review on energy consumption forecasting in smart buildings

After exploration and analysis of the literature, we identified two main categories of approaches for energy management: mathematical approaches and artificial intelligence (AI)-based approaches.

3.1. Mathematical-based approaches

The main objective of mathematical-based approaches is to model energy consumption prediction and capture linear relationships in time series, using mathematical equations and statistical models (e.g., regressions, ARIMA, TBATS).

In Goudarzi et al.,¹⁹ the authors proposed an optimized hybrid model, named AIK-EWMA, combining ARIMA (Auto-Regressive Integrated Moving Average) and ICA (Imperialist Competitive Algorithm), to accurately predict energy consumption in smart buildings and identify potential failures. The authors also used the EWMA (Exponentially Weighted Moving Average) model to monitor and detect minor variations in the predictions. The work of Alduailij et al.²⁰ proposed a forecasting model based on different statistical and machine learning approaches (ARIMA, TBATS, ANN, and LSTM) to predict the short-term (24 h) electricity consumption of smart government buildings, such as libraries, schools, and halls, in order to reduce energy costs and carbon emissions. In Fumo and Biswas,²¹ the authors explored simple, multiple and quadratic regression models, using meteorological data and energy consumption to predict energy consumption in residential buildings, specifically a research house (TxAIRE Research House). The authors of Korolija et al.²² developed simplified regression models to predict the annual energy requirements (heating, cooling, and auxiliaries) of HVAC systems in office buildings.

3.2. Artificial Intelligence-based approaches

AI-based approaches rely on artificial intelligence techniques capable of automatically identifying complex patterns and correlations within energy consumption data; this category include machine learning (ML) methods and deep learning (DL) methods.

3.2.1. Machine-learning-based approaches

The primary goal of this category of approaches is to predict energy consumption patterns in smart building systems, using machine learning methods that seek to automatically learncomplex relationships from consumption data and explanatory variables (climate, occupancy, etc.), in order to improve the accuracy of prediction compared to purely mathematical models, while remaining relatively efficient.

A predictive energy consumption model for smart commercial buildings is proposed in Keytingan et al.,¹⁵ using machine learning algorithms. The authors evaluate three methodologies: k-Nearest Neighbor (k-NN), Support Vector Machine (SVM), and Artificial Neural Network (ANN). The SVM performed best for two of the buildings considered, while k-NN was superior in the other cases. The authors of Nie et al.²³ proposed an innovative model to predict the energy consumption of smart homes, with simulations based on common home appliances such as HVAC, water heaters, washing machines, etc., using gradient boosting regression trees (GBRT), and a hybrid ARIMA-GBRT version. The proposed models reduce the prediction error (RMSE) compared to existing models (SVM, ARIMA-RNN, etc.) and allow better estimation of electricity consumption. The authors of González-Vidal et al.²⁴ compared energy consumption prediction models in a smart university building, using two approaches: models based solely on data (black-box models) and hybrid models using both data and physical knowledge (grey-box models). Their objective is to evaluate the performance of these approaches to predict energy consumption. In Cao et al.,²⁵ the authors compared and evaluated several machine learning models, simple (linear regression, SVR) and ensemble models (Random Forest, XGBoost), to predict electricity consumption in smart hospitals, more precisely in Shanghai Tenth People’s Hospital. They focus on the accuracy of predictions based on climate, occupancy and operation variables to improve energy management. In Guo et al.,²⁶the authors proposed an ANN prediction model that predicts the amount of energy needed to heat and cool a smart residential building, in order to enable early energy analysis, optimize building design and improve HVAC system performance.

3.2.2. Deep-learning-based approaches

This category includes works that use deep-learning methods leveraging advanced neural architectures (ANN, CNN, LSTM, hybrid) to simultaneously capture nonlinear, temporal and sometimes spatial dependencies in energy data. This is to offer superior predictive capacity, particularly for complex and high-dimensional series.

A deep-learning framework named kCNN-LSTM was proposed in Somu et al.²⁷ for energy consumption prediction of an academic building, including smart classrooms, offices, research labs, and servers. This model combines k-means clustering, convolutional neural networks (CNNs), and LSTM networks to analyze trends, identify energy-related features, and model temporal dependencies in energy consumption data. In Zini and Carcasci,¹⁰ the authors presented an energy monitoring method for HVAC systems of a hospital using predictive models based on multilayer perceptron and artificial neural networks (ANNs), to monitor electricity consumption and detect anomalies in energy behaviors. The models achieve high performances, with an R $^{2}$ value exceeding 0.92. Furthermore, An ANN-based methodology was proposed in Attoue et al.¹⁷ to predict the indoor temperature in a smart university building, improve occupant comfort and reduce energy consumption. The ANN model used provides an accurate prediction of indoor temperature with a high correlation coefficient ( $R > 0.95$ ). The authors of Durand et al.²⁸ developed a framework based on LSTM networks, to analyze and predict energy consumption in a smart institutional building. The work includes extensive time series analysis, feature engineering, and the creation of forecasting models based on these series. Also, a framework for predicting electricity consumption in smart residential buildings, using LSTM and Sequence to Sequence Learning method (Seq2Seq) was described in Rehman and Iqbal,²⁹ in order to optimize energy management, enable more efficient use of resources, and anticipate energy demand. In the same vein, Anan et al.³⁰ used an LSTM prediction model to improve the accuracy of energy consumption prediction in office buildings by integrating occupancy into prediction models, in order to enable better energy management, efficiency, planning and demand response. Other works focused on the use of hybrid models as Zhiqian et al.¹¹ who proposed a hybrid deep learning model (CNN–LSTM) to predict HVAC energy consumption in public buildings; the goal was to handle the nonlinear and multi-scale nature of energy data and improve prediction accuracy. The study demonstrates the effectiveness of the model for intelligent energy management and optimization. A learning-based framework using CNN and Bi-LSTM models with Attention mechanism was presented in Binbusayyis and Sha³¹ for efficient energy management and decision-making in smart building systems; the results showed good accuracy and robustness of the models.

Table 1 provides a comparison of the different prediction approaches mentioned above, according to a set of criteria we defined: type of Building, IoT devices, category of approach, hybrid model, methods used, data source, performance metrics and results obtained.

Table 1.
Comparison of approaches - prediction for energy consumption.

Type of IoT Hybrid Used Data Performance Results

Reference Building devices Category model methods Source metrics obtained

Somu et al.²⁷ Academic building (smart rooms, offices, laboratories) Raspberry Pi, sensors Deep-learning Yes K-means, CNN, LSTM Real historical data MSE, RMSE, MAE, MAPE Increased accuracy of predictions

Nie et al.²³ Smart home Temperature, humidity sensors and household appliances Machine-learning Yes Gradient Boosting Regression Tree (GBRT), ARIMA-GBRT Real historical data RMSE, MAE, MAPE Reducing prediction errors

Keytingan et al.¹⁵ Commercial building Smart meters Machine-learning Yes K-NN, SVM, ANN Real-time data RMSE, NRMSE, MAPE Increased accuracy

Zini and Carcasci¹⁰ Smart hospital Smart meters, wireless sensors Deep-learning Yes ANN, MLP Real-time data RMSE, R $^{2}$ Effective monitoring and anomaly detection

Attoue et al.¹⁷ University building Temperature, humidity and solar radiation sensors Deep-learning No ANN Real historical data MSE, R $^{2}$ Accurate temperature prediction, improved comfort

Goudarzi et al.¹⁹ Commercial building Raspberry Pi, sensors Mathematical Yes ARMI, ICA, EWMA Real-time data RMSE, MAE, ARE Reduction of energy loss

Alduailij et al.²⁰ Government building Smart meters Mathematical Yes ARIMA, TBATS, ANN, LSTM Real-time data MAPE, Accuracy Efficient management of energy peaks

Fumo and Biswas²¹ Smart home / Mathematical Yes Simple, multiple and quadratic regression Real-time data RMSD, R $^{2}$ More accurate estimation of energy consumption

Korolija et al.²² Office building / Mathematical No Regression model (quadratic and power) Simulation (Enegry Plus) R $^{2}$ , RMSD, Residual errors Fast and accurate prediction of energy needs

González-Vidal et al.²⁴ Smart building / Machine-learning Yes SVM, RF, XGBoost Real historical data / Improved prediction accuracy

Cao et al.²⁵ Smart hospital Smart meters Machine-learning Yes ANN, RF Real-time data MAPE, R $^{2}$ Increased accuracy

Durand et al.²⁸ Institutional building / Deep-learning No LSTM Real historical data RMSE, MAPE, R $^{2}$ Increased accuracy of predictions

Rehman and Iqbal²⁹ Residential Building Smart meters Deep-learning Yes LSTM, Seq2Seq Historical data MAE, RMSE Efficient energy management

Anan et al.³⁰ Commercial building Sensors Deep-learning No LSTM Real-time data MAPE, MAE, RMSE Improved prediction accuracy

Guo et al.²⁶ Residential Building / Machine-learning No ANN Simulation (Tsanas & Xifara) RMSE, MAE, MAPE Improved prediction accuracy

Binbusayyis and Sha³¹ Residential Building Sensors, household appliances Deep-learning Yes CNN, Bi LSTM, Attention Historical data MSE, MAE, MAPE, RMSE Improved prediction accuracy and energy management optimization

Zhiqian et al.¹¹ Smart Building HVAC sensors, environmental sensors Deep-learning Yes CNN, LSTM Historical data MAE, MAPE, R $^{2}$ Improved prediction accuracy

	Type of	IoT		Hybrid	Used	Data	Performance	Results
Somu et al.²⁷	Academic building (smart rooms, offices, laboratories)	Raspberry Pi, sensors	Deep-learning	Yes	K-means, CNN, LSTM	Real historical data	MSE, RMSE, MAE, MAPE	Increased accuracy of predictions
Nie et al.²³	Smart home	Temperature, humidity sensors and household appliances	Machine-learning	Yes	Gradient Boosting Regression Tree (GBRT), ARIMA-GBRT	Real historical data	RMSE, MAE, MAPE	Reducing prediction errors
Keytingan et al.¹⁵	Commercial building	Smart meters	Machine-learning	Yes	K-NN, SVM, ANN	Real-time data	RMSE, NRMSE, MAPE	Increased accuracy
Zini and Carcasci¹⁰	Smart hospital	Smart meters, wireless sensors	Deep-learning	Yes	ANN, MLP	Real-time data	RMSE, R $^{2}$	Effective monitoring and anomaly detection
Attoue et al.¹⁷	University building	Temperature, humidity and solar radiation sensors	Deep-learning	No	ANN	Real historical data	MSE, R $^{2}$	Accurate temperature prediction, improved comfort
Goudarzi et al.¹⁹	Commercial building	Raspberry Pi, sensors	Mathematical	Yes	ARMI, ICA, EWMA	Real-time data	RMSE, MAE, ARE	Reduction of energy loss
Alduailij et al.²⁰	Government building	Smart meters	Mathematical	Yes	ARIMA, TBATS, ANN, LSTM	Real-time data	MAPE, Accuracy	Efficient management of energy peaks
Fumo and Biswas²¹	Smart home	/	Mathematical	Yes	Simple, multiple and quadratic regression	Real-time data	RMSD, R $^{2}$	More accurate estimation of energy consumption
Korolija et al.²²	Office building	/	Mathematical	No	Regression model (quadratic and power)	Simulation (Enegry Plus)	R $^{2}$ , RMSD, Residual errors	Fast and accurate prediction of energy needs
González-Vidal et al.²⁴	Smart building	/	Machine-learning	Yes	SVM, RF, XGBoost	Real historical data	/	Improved prediction accuracy
Cao et al.²⁵	Smart hospital	Smart meters	Machine-learning	Yes	ANN, RF	Real-time data	MAPE, R $^{2}$	Increased accuracy
Durand et al.²⁸	Institutional building	/	Deep-learning	No	LSTM	Real historical data	RMSE, MAPE, R $^{2}$	Increased accuracy of predictions
Rehman and Iqbal²⁹	Residential Building	Smart meters	Deep-learning	Yes	LSTM, Seq2Seq	Historical data	MAE, RMSE	Efficient energy management
Anan et al.³⁰	Commercial building	Sensors	Deep-learning	No	LSTM	Real-time data	MAPE, MAE, RMSE	Improved prediction accuracy
Guo et al.²⁶	Residential Building	/	Machine-learning	No	ANN	Simulation (Tsanas & Xifara)	RMSE, MAE, MAPE	Improved prediction accuracy
Binbusayyis and Sha³¹	Residential Building	Sensors, household appliances	Deep-learning	Yes	CNN, Bi LSTM, Attention	Historical data	MSE, MAE, MAPE, RMSE	Improved prediction accuracy and energy management optimization
Zhiqian et al.¹¹	Smart Building	HVAC sensors, environmental sensors	Deep-learning	Yes	CNN, LSTM	Historical data	MAE, MAPE, R $^{2}$	Improved prediction accuracy

Analysis and discussion

Regarding the categories of approaches, we can say that mathematical and statistical models such as ARIMA,²³ TBATS²⁰ or regression²² are suitable for capturing linear temporal dependencies, but they often fail to handle nonlinear and high-dimensional data typical of IoT environments; this limits their applicability in complex environments.

Machine learning methods like SVM, ensemble techniques (XGBoost, GBRT) or Random Forests (RF), improve predictive performance by learning from historical data. However, they remain sensitive to feature selection and do not exploit sequential dependencies, which are fundamental in energy consumption time series.

Deep learning approaches emerged as the most promising solutions due to their ability to model complex nonlinear and sequential patterns. Architectures such as CNN, ANN, and LSTM^11,17,27,29 demonstrate improved performance compared to traditional methods. However, each architecture individually remains limited: CNNs efficiently extract spatial features but ignore temporal evolution, while LSTM capture long-term temporal dependencies but neglect spatial interactions among building zones.³⁰ This structural limitation prevents these models from fully capturing the spatio-temporal nature of energy consumption.

Moreover, a critical gap in the literature lies in the application context; most studies focus on residential or office buildings, where energy dynamics are relatively regular. In contrast, most studies rarely address event-driven and complex heterogeneous environments such as hospitals,^10,25 where energy dynamics are influenced simultaneously by heterogeneous equipment (critical vs. non-critical), dynamic occupancy patterns, and event-driven conditions (e.g., heatwaves, pandemics, system failures), making prediction significantly more challenging.

Also, we noticed that few studies explore the use of critical equipment such as medical equipment. The management of critical equipment is often a priority because of its direct impact on the health, safety and comfort of the occupants.

From the table, we can see that most of the works use either real-time data^10,15 or historical data,^17,27 often complemented by simulations to validate the models. Moreover, to validate the effectiveness of the proposed approaches, the performance metrics like RMSE, MAE, and R $^{2}$ are the most commonly used.

Another observation is that hybrid models found in the literature (e.g., ANN–LSTM²⁰ or ANN–RF²⁵) do not explicitly encode the spatial structure of the building, and thus cannot fully represent inter-room dependencies, which are essential for accurate forecasting in multi-room infrastructures. In addition, few works Durand et al.²⁸ are supported by a conceptual or system-level framework, meaning that prediction models remain disconnected from the system-level operation, data flow, and IoT integration. These findings clearly reveal that existing models are fragmented, addressing specific aspects of energy forecasting problem without providing an integrated framework capable of adapting to dynamic system conditions.

This observation motivates our work to design a spatio-temporal deep learning framework that integrates both conceptual modeling and hybrid model-based prediction, capable of handling spatial heterogeneity, temporal dependencies, and contextual variability within smart buildings. Our proposed CNN2D–LSTM model directly responds to the identified gaps by:

–

Jointly modeling spatial and temporal dependencies, leveraging 2D convolution for room-to-room interactions and LSTM layers for sequential temporal learning.

–

Operating within a conceptual and system architecture, ensuring that predictions are consistent with the building’s IoT data flow and energy management processes.

–

Adapting to dynamic conditions, including event-driven scenarios (e.g., heatwaves, pandemics, outages).

4. Proposed solution

In this section, we present the conceptual and technical aspects of our framework. First, we describe a meta-model of concepts supporting a smart building EMS, and a software architecture of the system. Second, on the technical side, we propose a hybrid deep learning model (CNN2D+LSTM) for accurate energy consumption prediction aiming to support proactive energy management, improve operational efficiency, and ensure reliable energy supply for critical building functions.

4.1. Conceptual meta-model for smart building EMS

The meta-model described in Figure 1 captures both the spatial organization of the infrastructure and the temporal dynamics of its operation. On the spatial dimension, the smart building is represented as a hierarchical structure composed of ensemble of buildings, floors, and rooms. Each room is characterized by attributes such as orientation (North, South, East, West), geometric position (pos_x, pos_y), and area, which together define the spatial context influencing energy behavior, for example, solar exposure or proximity to HVAC zones. Regarding the temporal dimension, the model integrates the evolution of energy-related variables collected by IoT devices. Sensors continuously measure time-dependent data such as temperature, humidity, occupancy, and energy consumption, while actuators dynamically respond to these variations by adjusting lighting, HVAC, or other control systems. This interaction between spatial configuration and temporal variation allows the system to represent how energy consumption evolves over time and across physical zones of the building.

Figure 1.

Metamodel of concepts - Smart building EMS.

A smart building uses a variety of IoT devices including environmental sensors and energy monitoring sensors, actuators like smart switches and smart thermostats, to operate effectively and optimize its overall performance. It also contains a set of equipment, which can be categorized into two types: critical equipment, such as fire safety systems and security cameras, and non-critical equipment, such as lighting systems and HVAC units. Energy sources such as electric energy, solar panels, batteries, or emergency generators, supply power to the building’s equipment. These are managed through an Energy Management System (EMS) that processes spatio-temporal data, detects events (e.g., demand peaks, outages), generates alerts and triggers adaptive control or prediction mechanisms.

The metamodel thus provides a unified representation linking the “where” (spatial structure) and the “when” (temporal dynamics) aspects of energy management within a smart building.

4.2. System architecture

The proposed smart building EMS follows a four-layer architecture (see Figure 2) reflecting the fundamental components of an IoT environment and interactions between them. Each layer plays a distinct role, from data acquisition to intelligent decision-making and user interaction. The architecture includes a perception layer, a network layer, a data processing layer, and an application layer.

Figure 2.

Multi-layer EMS architecture.

The perception layer (also called the physical layer) captures real-time data from the building environment. It includes environmental sensors (humidity, temperature, and motion), HVAC and lighting actuators, critical and non-critical equipment, and the structural layout of smart buildings. This layer monitors and transmits spatio-temporal data such as temperature, humidity, and energy usage consumption parameters, to higher layers.

The network layer supports wireless communication protocols such as Wi-Fi and Bluetooth, and leverages application communication protocol such as HTTP and MQTT, to ensure lightweight, secure, and efficient message exchange. It represents the communication relationship between the IoT devices, equipment and the EMS.

The data processing layer includes storage and intelligent components of the EMS. The cloud data storage component provides scalable and persistent storage of historical and real-time data. This layer applies prediction models like CNN2D-LSTM model to predict future energy consumption, and extract temporal and spatial patterns. In this layer, the spatio-temporal reasoning, event detection, and alert generation, are effectively implemented.

The application layer represents the interface between the EMS and the user. It visualizes real-time and forecasted energy consumption by room, floor, or building, displays detected events (e.g., heatwave, outage), and issues alerts to facility managers.

4.3. Hybrid prediction model for energy consumption

As already said, we adopt a hybrid deep learning model that combines Convolutional Neural Networks (CNN2D) and Long Short-Term Memory (LSTM) networks for energy consumption prediction. While LSTM models are known for their capacity to model temporal dependencies in sequential data, the addition of CNN layers allows the model to also capture spatial correlations between rooms and floors, which is particularly relevant in multi-room facilities. This model enables the network to jointly learn spatial features (such as room location and energy usage pattern distribution) and temporal dynamics (e.g., daily load profiles, occupancy evolution, HVAC cycles) of energy consumption, leading to more accurate and context-aware predictions.

4.3.1. Problem formulation

Energy consumption in smart buildings evolves over time and exhibits temporal dependencies influenced by seasonal patterns, occupancy behavior, equipment cycles, and contextual conditions.

Let: $X_{t} \in R^{d}$ be a multivariate feature vector at time step t, describing the conditions in a given room:

–
Environmental features: indoor/outdoor temperature, humidity
–
Operational features: occupancy, HVAC and lighting usage, medical equipment consumption (critical and non-critical equipment)
–
Contextual features: hour of day, day of week, weekend, night, event indicators (e.g., heatwave, pandemic, breakdown)
–
Temporal dependencies: lagged values of the global energy consumption

To capture both temporal and spatial dependencies, we define a sliding window of length $L = 18$ ( $\approx$ 4.5 days at 6 h resolution). At each time step, the hospital rooms are projected onto a dynamic 2D grid of dimension $W \times H$ , forming a tensor: $X = {X_{t - L + 1}, \dots, X_{t}} \in R^{W \times H \times f}$ . Where each cell corresponds to a room and contains its feature vector. The target is the aggregated energy consumption the next time instant (t + 1), expressed in log-space as $y_{t + 1} = l o g (1 + {e n e r g y}_{t + 1})$ .
4.3.2. Model architecture: CNN2D - LSTM

The proposed architecture consists of the following components:

Input layer: A spatio-temporal tensor of shape $(L, W, H, f)$ , where each time instant is a 2D spatial grid of room features.

Spatial encoding (applied TimeDistributed across time steps):

–
Two CNN2D blocks where each block consists of $$
Conv2D with 16 filters: Applies 2D convolution to each spatial frame independently to extract spatial features, such as localized consumption patterns, or room location and orientation.
$$
BatchNormalization & MaxPooling2D: Used to stabilize training and reduce spatial dimensions.
$$
Conv2D with 24 filters
$$
BatchNormalization & MaxPooling2D

–
GlobalAveragePooling2D: applied per time step.

Temporal modeling: LSTM layer with 128 units, returning a sequence embedding.

Dense Layers: Fully connected layers of sizes $256 \to 128 \to 64 \to 32$ , each with ReLU activation and Dropout (rate $=$ 0.3), for high-level integration of learned features.

Output Layer: A single neuron predicting the next global energy consumption value at time t+1

A series of controlled experiments were conducted to systematically evaluate the impact of key hyper-parameters, including the number of convolutional filters, LSTM units, and dropout rate. Multiple configurations were tested, and their performance was assessed on a validation dataset. The final configuration was selected based on its ability to minimize prediction error while ensuring a good trade-off between model complexity and generalization.

Figure 3 illustrates the full architecture of the hybrid predictive model, showing the flow of information from the input spatial-temporal window to the prediction of the next building-wide energy value.

Figure 3.
CNN2D $+$ LSTM architecture.

The spatial feature extraction process begins with the TimeDistributed Conv2D layer (top block in Figure 3), which applies 2D convolutions to each time step in order to capture spatial correlations between different zones of the building. The extracted features are then passed through a BatchNormalization and MaxPooling2D block, which stabilizes the training process and reduces spatial dimensions. Subsequently, a TimeDistributed GlobalAveragePooling2D layer is applied to flatten the spatial representations at each time step into compact feature vectors. These features are then fed into the LSTM layers (128 units), which model temporal dependencies across time. Finally, the learned temporal representations are processed through Dropout and fully connected Dense layers to produce the final energy consumption prediction.
4.3.3. CNN2D - LSTM process

Figure 4 illustrates the overall data processing and prediction pipeline of our framework. Room data, including spatial descriptors (location, orientation, area) and environmental variables (temperature, humidity, occupancy, energy use), are first preprocessed and mapped onto a 2D spatial grid. Sliding temporal windows are then constructed and passed through a CNN block to extract spatial correlations among neighboring rooms, followed by an LSTM block to capture temporal dependencies, to finally obtain the final prediction of energy consumption. These forecasts provide actionable inputs to the EMS, supporting proactive decision-making and operational efficiency in smart buildings.

Figure 4.

Functional schema of the prediction process.

5. Case study: A smart hospital

Smart hospital represents one of the most complex and energy-intensive types of smart buildings. Studies reported that healthcare facilities consume on average 2.5 to 3 times more energy per square meter than office buildings.³² Unlike residential or commercial facilities, hospitals operate continuously, host a wide variety of critical and non-critical medical equipment, and must maintain strict indoor environmental conditions for patient safety. This combination of heterogeneous energy loads, continuous operation, and event-driven variability like emergencies, pandemics and equipment failures, makes energy management particularly challenging.³³ Thus, modeling and forecasting the energy consumption of smart hospitals is essential not only to reduce operational costs, but also to maintain a high quality of service and promote environmental sustainability in a field where reliability is essential.

5.1. Specialization of concepts to a smart-hospital

Based on the metamodel described in Section 4 (Figure 1), a smart hospital building preserves the same overall structure composed of ensemble of buildings, floors, rooms, equipment, and IoT devices, but specializes certain entities and relationships to reflect the specificities of the healthcare context, Table 2 summarizes these specifics.

Table 2.
Specialization of the metamodel to a smart-hospital.

Generic metamodel concept Specialization in smart hospital

Building Medical buildings (cardiology, general surgery)

Room Operating rooms, laboratories, intensive care units, patients’ rooms

Equipment Medical equipment (anesthesia machines, laboratory devices, infusion pumps, sterilization machines, patients’ monitoring equipment)

Sensors Biomedical equipment sensors, bed occupancy sensors

Actuators Cold chain actuators, hospital bed actuators

Event Pandemic, heatwaves, medical equipment failure

Generic metamodel concept	Specialization in smart hospital
Building	Medical buildings (cardiology, general surgery)
Room	Operating rooms, laboratories, intensive care units, patients’ rooms
Equipment	Medical equipment (anesthesia machines, laboratory devices, infusion pumps, sterilization machines, patients’ monitoring equipment)
Sensors	Biomedical equipment sensors, bed occupancy sensors
Actuators	Cold chain actuators, hospital bed actuators
Event	Pandemic, heatwaves, medical equipment failure

5.2. Data generation for smart hospital

To evaluate the proposed model, we generated a dataset using a custom-built simulation model designed to replicate the energy consumption behavior of a smart hospital composed of five interconnected buildings: Cardiology, Periodontology, Pediatrics, General Surgery, and Dermatology. These buildings exhibit structural and functional diversity, with varying numbers of floors, rooms, and medical equipment. Table 3 presents the smart hospital infrastructure data, we considered.

Table 3.
Smart hospital infrastructure data.

Building Floors Rooms

Cardiology 3 23

Periodontology 2 13

Pediatrics 3 22

General Surgery 3 21

Dermatology 2 13

Building	Floors	Rooms
Cardiology	3	23
Periodontology	2	13
Pediatrics	3	22
General Surgery	3	21
Dermatology	2	13

Each building is equipped with a variety of critical and non-critical medical devices, lighting systems, HVAC units, elevators, and ambient sensors. The simulation spans a period of 60 days, with data collected at 6-hour intervals, resulting in detailed temporal granularity. Table 4 presents the dataset features.

Table 4.

Dataset features.

Category	Features
Timestamp	Date and time of data entry
Spatial features	Building, Floor, Room: location (pos_x, pos_y), orientation (North, South, East, West), area (m $^{2}$ )
Environmental features	temp_outdoor_C, temp_indoor_C, humidity_pct
Operational features	occupancy_count, hvac_kwh, lighting_kwh, medical_critique_kwh, medical_non_critique_kwh, medical_kwh, energy_kwh
Contextual features	event_heatwave, event_pandemic, event_outage, hour, weekday, is_weekend, is_night, n_critique, n_non_critique

Note that medical_kwh is defined as the sum of critical and non-critical medical equipment energy consumption. And energy_kwh $=$ hvac_kwh $+$ lighting_kwh $+$ medical_kwh, with stochastic noise occasionally added to reflect measurement variability.

Due to the limited availability of publicly accessible smart hospital datasets, a simulation-based approach was adopted. The simulation parameters were defined according to energy consumption patterns reported in previous studies on smart buildings and healthcare facilities. The generated ranges and operational scenarios were selected to approximate realistic hospital energy profiles described in the literature. To ensure realism, the dataset was generated under a set of physical, behavioral, and operational constraints described below:

Temporal Variations Rules –

From 08:00 to 17:00, occupancy levels tend to be higher, resulting in lower lighting consumption due to daylight availability.

–

During nighttime (e.g., 00:00 to 06:00), occupancy decreases and lighting consumption drops overall, except in critical zones like emergency rooms that remain illuminated for continuous operation.

Room-Type Specific Energy Profiles Each room type in the hospital follows a distinct energy consumption profile reflecting its function and operational pattern. –

Patient rooms show stable consumption with occasional increases due to medical activity.

–

Laboratories and surgical operation rooms exhibit high daytime peaks due to intensive medical activities and HVAC usage for sterilization.

–

Waiting areas have high variability with strong dependence on daytime occupancy.

–

Rest areas have minimal energy usage except for lighting during off-hours.

–

Radiology rooms maintain a high and constant baseline consumption because of imaging equipment that must remain operational or in standby mode.

Cross-feature Correlations:

Increases in hvac_kwh are correlated with higher energy_kwh.

Heatwave events (event_canicule) result in elevated indoor temperatures and consequently increased HVAC usage, which can also affect lighting and occupancy behavior due to discomfort.

Environmental Value Ranges: –

temp_outdoor_C: 20–34 $^{\circ}$ C in normal conditions, 29–40 $^{\circ}$ C during heatwaves.

–

temp_indoor_C: 20–28 $^{\circ}$ C normally, 27–35 $^{\circ}$ C during heatwaves.

–

humidity_pct: 40%–60% in all cases.

Operational Value Ranges: –

occupancy_count: 0–20 depending on room type, capped at 3 in consultation or resting rooms, up to 20 in waiting areas.

–

hvac_kwh: 1.5–2.5 kWh in normal conditions, 2.5–4.0 kWh during heatwaves.

–

lighting_kwh: 0–3.0 kWh depending on time and occupancy.

–

medical_kwh: 1–20 kWh depending on equipment type and usage.

–

energy_kwh: Computed as hvac_kwh + lighting_kwh + medical_kwh).

Temporal and Event-driven Patterns: –

Realistic simulation of critical events such as heatwaves, pandemics, and power outages, which were introduced following pre-defined frequency patterns to simulate operational stress conditions.

–

During weekends, increased public and patient activity leads to higher energy consumption, particularly in outpatient and emergency services.

The simulated dataset was designed to reproduce realistic energy consumption behaviors commonly observed in smart hospitals. The generated data integrates variations related to occupancy levels, temperature conditions, lighting systems, HVAC operation, and the use of medical equipment. In addition, the simulation reproduces daily fluctuations in energy demand, including daytime peaks caused by intensive equipment usage and reduced nighttime consumption. Critical hospital devices were also considered as continuously operating loads to better reflect real hospital environments. This synthetic dataset reflects the spatio-temporal, behavioral, and contextual dynamics observed in smart hospitals, providing a robust basis for evaluating energy prediction algorithms under diverse operational scenarios.

5.3. Application overview of the smart hospital EMS

By integrating advanced machine learning algorithms into an intuitive platform, the system enables hospital managers, energy engineers, and facility operators to monitor and forecast energy consumption in real time. For framework development, we used Python to implement our predictive model, Firebase for data storage, and HTML, JavaScript, CSS3, and Express.js for the web application including the following key features:

–
Interactive dashboards to visualize energy consumption by room, floor, or building over time (see Figure 5).
–
Real-time alerts for unusual consumption, peak risks, or equipment anomalies.
–
Scenario simulation tools to test and compare different energy management strategies (e.g., crisis mode vs. normal operation).
–
Configuration interface to customize parameters such as thresholds of temperature, humidity, occupancy levels, etc.

Figure 5.
Energy consumption dashboard.

The application helps ensure the continuity of essential healthcare services by minimizing energy-related risks and inefficiencies for both critical and non-critical equipment. It also supports Green IoT principles by promoting sustainable energy management within the smart hospital, reducing unnecessary energy usage and preventing waste.
6. Experimental results of the predictive model

To evaluate the effectiveness of the proposed CNN2D–LSTM model, we defined several test scenarios based on the smart hospital dataset to capture different operational and environmental conditions. The experiments aim to assess four main aspects:

Baseline energy prediction accuracy: comparing the CNN2D–LSTM model with conventional approaches (LSTM, CNN, etc.), and against a reference model from the literature Somu et al.²⁷ and Zhiqian et al.¹¹

Normal Scenario: representing standard hospital operation during regular weekdays, with typical occupancy, temperature, and energy demand patterns.

Under-Consumption Scenario: corresponding to periods of reduced activity, such as nighttime and weekends, where occupancy and energy usage significantly decrease.

Heatwave Scenario: simulating a rise in outdoor temperature leading to increased cooling demand and HVAC system activity.

These scenarios collectively demonstrate how the CNN2D–LSTM framework captures both spatial and temporal dependencies, ensuring reliable energy prediction across varying hospital contexts.

The dataset was chronologically split into 70% training, 15% validation, and 15% test sets. Input features were normalized using Min-Max scaling fitted on the training set. Our model was trained with the following settings:

–
Optimizer: AdamW (fallback Adam), chosen for fast convergence and better regularization, reducing overfitting on time-dependent data.
–
Loss function: Huber loss, to improve robustness against outliers and sudden consumption peaks.
–
Batch size: 16, as a compromise between computational efficiency and model stability.
–
Maximum epochs: 240, with early stopping (patience $=$ 20) to ensure convergence while preventing overfitting once the validation loss plateaued.
–
Evaluation metrics: RMSE, MAE, MAPE, and R $^{2}$ . Evaluation is restricted to samples with energy consumption $\geq$ 2.0, to avoid high relative error during very low-consumption periods (e.g., night, rest areas).

6.1. Prediction performance metrics

To evaluate the accuracy and robustness of our proposed model, we rely on four widely used regression metrics: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared (R $^{2}$ ) as described in Table 5.

Table 5.
Prediction performance metrics.

Performance metrics Formula Use justification

RMSE $\sqrt{\frac{1}{n} \sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}$ It helps assess how much energy prediction deviates from actual consumption in absolute magnitude, especially in energy-intensive zones like operating rooms or treatment areas.

MAE $\frac{1}{n} \sum_{i = 1}^{n} | y_{i} - {\hat{y}}_{i} |$ It is particularly useful in hospital settings where consistent forecasting across rooms is critical.

MAPE $\frac{100}{n} \sum_{i = 1}^{n} | \frac{y_{i} - {\hat{y}}_{i}}{y_{i}} |$ It is especially relevant in smart hospitals where rooms vary greatly in energy use (e.g., consultation rooms vs intensive care).

R $^{2}$ $1 - \frac{\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}{\sum_{i = 1}^{n} (y_{i} - \bar{y})^{2}}$ It is useful for understanding how well the model captures the underlying dynamics of energy consumption across time and space.

Performance metrics	Formula	Use justification
RMSE	$\sqrt{\frac{1}{n} \sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}$	It helps assess how much energy prediction deviates from actual consumption in absolute magnitude, especially in energy-intensive zones like operating rooms or treatment areas.
MAE	$\frac{1}{n} \sum_{i = 1}^{n} \| y_{i} - {\hat{y}}_{i} \|$	It is particularly useful in hospital settings where consistent forecasting across rooms is critical.
MAPE	$\frac{100}{n} \sum_{i = 1}^{n} \| \frac{y_{i} - {\hat{y}}_{i}}{y_{i}} \|$	It is especially relevant in smart hospitals where rooms vary greatly in energy use (e.g., consultation rooms vs intensive care).
R $^{2}$	$1 - \frac{\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}{\sum_{i = 1}^{n} (y_{i} - \bar{y})^{2}}$	It is useful for understanding how well the model captures the underlying dynamics of energy consumption across time and space.

Where $y_{i}$ denotes the actual value, ${\hat{y}}_{i}$ the predicted value, $\bar{y}$ the mean of actual values, and $n$ the number of observations.

As already said, we compared the proposed CNN2D–LSTM model against six baseline prediction models widely used in energy forecasting and two works from literature using hybrid models Somu et al.²⁷ and Zhiqian et al.,¹¹ the models are:

Artificial neural Network (ANN): a fully connected architecture used as a baseline without explicit spatial or temporal modeling.

Recurrent Neural Network (RNN): designed to capture short-term temporal dependencies in sequential data.

Convolutional Neural Network (CNN): implemented as a 1D temporal convolutional model to extract local temporal patterns.

Long Short-Term Memory (LSTM): designed to model long-term temporal dependenciesin time-series data.

Hybrid ANN–LSTM and RNN–LSTM: combining dense or recurrent feature extraction with temporal memory modeling.

K-means, CNN, LSTM²⁷: a hybrid clustering-based approach that combines data partitioning with deep learning models.

CNN, LSTM¹¹: a hybrid architecture combining convolutional layers for feature extraction and LSTM layers for temporal dependency modeling.

To ensure a fair comparison, all models were trained using the same simulated smart hospital dataset, identical preprocessing steps, and the same experimental conditions. Specifically, the same input features, data preprocessing steps, and training–testing splits were used across all methods. In addition, the number of training epochs and batch size were kept identical for all models. Hyper-parameters for each model were selected using a similar empirical tuning strategy to ensure a balanced and unbiased evaluation.

The comparative evaluation can also be interpreted as an ablation study, where the contribution of each component is progressively analyzed. The CNN model includes only convolutional layers applied to the input data to capture local patterns, without temporal modeling. The LSTM model processes the same input features as sequential data to capture temporal dependencies without spatial encoding. The proposed CNN2D–LSTM model combines both components by applying spatial feature extraction at each time step followed by temporal modeling, ensuring that performance differences are solely due to architectural variations rather than data discrepancies.

Several observations can be drawn from these results reported on Table 6:

ANN achieves limited accuracy, as it ignores temporal dependencies inherent to energy consumption time series.

CNN performs poorly (negative $R^{2}$ ), confirming that temporal convolutions alone cannot capture long-range dependencies in hospital data.

LSTM improves over ANN and CNN by modeling sequential dynamics, but its performance remains moderate ( $R^{2} = 0.37$ ). This suggests that temporal modeling alone is insufficient in environments with strong spatial heterogeneity.

RNN attains slightly explanatory power ( $R^{2} = 0.09$ ), but its accuracy is hindered by vanishing gradients and its limited capacity to capture complex non-linearities.

Hybrid models (ANN-LSTM, RNN-LSTM) yield inconsistent results; although they integrate memory, they lack explicit spatial encoding, resulting in only marginal improvements.

The accuracy of Somu et al.²⁷ is lower than our CNN2D–LSTM accuracy, indicating that while clustering helps to reduce data variability, it does not fully address the intrinsic spatio-temporal complexity of hospital energy dynamics.

The CNN–LSTM model¹¹ shows moderate performance ( $R^{2} = 0.18$ ), indicating its ability to capture temporal patterns through sequential modeling. However, its relatively high RMSE (3.91) and MAE (3.58) suggest limited accuracy in predicting complex energy variations. This can be explained by the absence of explicit spatial modeling, as the 1D convolution focuses on feature extraction without preserving the spatial structure of building data. As a result, the model struggles to accurately represent the heterogeneous and spatially distributed energy consumption patterns characteristic of smart hospital environments.

Our proposed model CNN2D-LSTM clearly outperforms all baselines, with the lowest RMSE (1.61) and MAE (1.12), a competitive MAPE (9.31%), and the highest R $^{2}$ (0.76), so it gives an accuracy of 90.69 %. Its superior performance confirms that jointly learning spatial and temporal dependencies is essential for accurate forecasting in complex hospital environments.

Table 6.

Performance of prediction models.

Model	RMSE	MAE	MAPE(%)	R $^{2}$	Accuracy (%)
ANN	4.87	1.12	29.63	0.15	70.37
RNN	4,22	3.52	20.90	0.09	79.1
CNN	5.25	4.36	30.13	−1.13	69.87
LSTM	4.26	3.37	27.76	0.37	72.24
ANN-LSTM	5.30	4.40	30.09	0.11	69.91
RNN-LSTM	5.32	4.43	32.30	−0.63	67.7
CNN-LSTM-K-means²⁷	5.37	4.32	31.18	−1.33	68.82
CNN-LSTM¹¹	3.91	3.58	28.59	0.18	71.41
CNN2D-LSTM (our proposed model)	1.61	1.12	9.31	0.76	90.69

The results show that combining CNN2D and LSTM is particularly effective for hospital energy prediction; CNN layers extract localized spatial patterns, such as correlated HVAC zones or clusters of high-energy medical equipment, and LSTM layers capture long-term dependencies across time, including daily/weekly cycles and event-driven variations. So, their integration enables the model to generalize across heterogeneous rooms and equipment while preserving temporal coherence.

6.2. Comparison with hybrid models (predicted and realistic scenarios)

The objective of this comparison is to demonstrate the effectiveness of the proposed CNN2D–LSTM model in accurately predicting energy consumption, and highlight the ability of the proposed approach to better capture complex spatial–temporal dependencies, giving more realistic results then the two hybrid models used in Somu et al.²⁷ and Zhiqian et al.¹¹

Figure 6 presents a comparison between the real and predicted energy consumption values obtained using the proposed CNN2D + LSTM architecture. The x-axis represents the test samples over time, while the y-axis indicates energy consumption in kilowatt-hours (kWh). The blue curve corresponds to the real values, whereas the orange curve shows the model’s predictions.

Figure 6.

Real vs Predicted energy consumption using CNN2D $+$ LSTM architecture.

We can observe that the two curves lie in the same range and display peaks and troughs at the same periods, indicating good tracking of the cycles. Minor discrepancies are observed, particularly in peak and drop regions, where the model tends to slightly overestimate or underestimate consumption due to its smoothing effect. Nevertheless, the close proximity of the two curves confirms the model’s ability to jointly capture spatial and temporal dependencies in hospital energy dynamics. This demonstrates that the CNN2D–LSTM provides reliable forecasts of total energy demand, which is crucial for proactive energy management in smart hospitals. Figure 7 compares the real hospital energy consumption with the predictions obtained from our proposed CNN2D–LSTM model and from the model proposed in Somu et al.²⁷

Figure 7.

Real vs Predicted energy consumption - Comparison between CNN2D + LSTM and Somu et al.²⁷ work.

We observe that the predictions of Somu et al.²⁷ (green dotted curve) remain consistently below the real consumption, indicating a systematic underestimation of energy demand. In contrast, our model (orange curve) captures the overall dynamics more effectively, staying closer to the real curve throughout the test horizon and showing improved alignment in the final intervals. This demonstrates that while both models can follow the general trend, the spatial-temporal integration of our CNN2D-LSTM provides a more realistic approximation of hospital energy patterns. The gap between the curves highlights the limitations of clustering-based baselines when applied to complex environments with strong spatial heterogeneity, such as smart hospitals.

Toward the end of the observation period, the three curves (real, CNN2D–LSTM, and Somu et al.²⁷) appear almost superimposed. This convergence, however, mainly results from the higher predictive stability of our model, which consistently aligns with the real consumption values as the system reaches a steady state. In contrast, the model of Somu et al.²⁷ only approaches the real curve at this stage after having systematically underestimated energy demand throughout most of the horizon. This behavior confirms that our spatio-temporal modeling ensures both faster convergence and higher long-term accuracy, particularly when the energy dynamics stabilize.

In contrast, the method proposed by Somu et al.²⁷ tends to perform poorly during periods of high energy consumption, particularly in peak demand intervals. This behavior can be explained by its reliance on clustering-based preprocessing combined with deep learning models, which may smooth out extreme values and reduce sensitivity to sudden variations. As a result, the model struggles to accurately capture sharp increases in energy demand that are typically driven by complex and nonlinear interactions between occupancy, equipment usage, and environmental conditions. In high-consumption scenarios, such as heatwaves or intensive medical activity, energy patterns exhibit strong variability and nonlinearity. Models that do not explicitly integrate both spatial heterogeneity and temporal dependencies may fail to adapt to these rapid fluctuations. In contrast, the proposed CNN2D–LSTM model effectively captures these dynamics by combining spatial feature extraction with temporal sequence modeling, enabling more accurate prediction of peak consumption periods. Figure 8 compares the real hospital energy consumption with the predictions obtained from our proposed CNN2D-LSTM model and from the CNN–LSTM model inspired by Zhiqian et al.¹¹

Figure 8.

Real vs Predicted energy consumption - Comparison between CNN2D $+$ LSTM and Zhiqian et al.¹¹ work.

We observe that the predictions of the CNN–LSTM model (green dashed curve) remain consistently below the real consumption values, indicating a systematic underestimation of energy demand throughout the test period. In contrast, our CNN2D–LSTM model (orange curve) more accurately follows the real consumption (blue curve), capturing both the overall trend and the temporal variations with higher fidelity. This demonstrates that while both models are able to learn temporal patterns, the integration of spatial information in the CNN2D–LSTM significantly improves prediction accuracy in complex environments such as smart hospitals.

The gap between the CNN–LSTM predictions and the real curve is particularly noticeable during periods of moderate to high energy consumption, where the model fails to capture the amplitude of variations. This limitation can be attributed to the use of 1D convolution, which processes features sequentially without preserving the spatial relationships between rooms and equipment. As a result, the model is unable to fully represent the heterogeneous distribution of energy consumption across different hospital zones.

Toward the end of the observation period, although the CNN–LSTM curve shows slight alignment with the decreasing trend, it still underestimates the actual consumption levels. In contrast, the CNN2D–LSTM model maintains closer proximity to the real values, reflecting better generalization and stability. This behavior confirms that incorporating spatial–temporal dependencies enables the model to better adapt to both gradual and abrupt changes in energy demand.

Furthermore, during periods of significant variation, the CNN–LSTM model exhibits reduced sensitivity to fluctuations, leading to smoother but less accurate predictions. This smoothing effect limits its ability to capture sudden changes driven by contextual events or operational dynamics. In contrast, the proposed CNN2D–LSTM model effectively captures these variations by combining spatial feature extraction with temporal modeling, resulting in more reliable predictions across different operating conditions.

6.3. Experimentation of our CNN2D-LSTM model for three scenarios)

We also compared the model performance across three operational scenarios: heat wave, under-consumption (night/weekend) and normal daily activity. Figure 9 present the comparison between real energy consumption and the predictions obtained from our model, in the three scenarios.

Figure 9.

Real vs Predicted energy consumption - different scenarios. (a) Heat wave scenario, (b) under-consumption scenario, and (c) normal activity scenario.

Figure 9(a) shows the real vs. predicted time series during heat waves, we can see that the prediction clearly follows the general trend and consumption peaks, confirming that the model accurately captures the increase in demand during extreme conditions. Figure 9(b) shows under-consumption (night/weekend), we can see that the model maintains good stability at low activity levels, and the predicted values remain more or less close to the real values. Figure 9(c) presents the normal activity case (business day), it shows a good overall correspondence: the shape of the daily variations and the average levels are well reproduced, indicating that the model provides reliable predictions for daily operations.

Computational complexity. The proposed CNN2D–LSTM model introduces higher computational complexity compared to standalone models due to the integration of convolutional and recurrent layers. The training process was performed offline and required approximately 10–15 min on a standard machine (e.g., Intel i7 CPU with 8 GB RAM). However, once trained, the model achieves fast inference, with an average prediction time of approximately 10–20 ms per sample, making it suitable for real-time or near real-time deployment within an EMS. These results demonstrate that the proposed model remains computationally feasible and can be deployed in large-scale IoT-based smart building environments. For instance, the training time of Somu et al.²⁷ is approximately 21 min with a prediction time of 43 ms per sample. While Zhiqian et al.¹¹ achieved a much shorter training time of 32 s with a prediction time of 38 ms per sample, highlighting that the computational cost of deep learning models varies significantly depending on the architecture and implementation choices.

7. Conclusions and future works

The work presented in this paper deals with forecasting energy consumption in smart buildings, by addressing a key gap in the literature related to the modeling of spatio-temporal dependencies in complex IoT-enabled environments, particularly in large-scale and heterogeneous smart buildings. Indeed, existing approaches often fail to jointly capture spatial interactions between zones and temporal evolution of energy consumption. Consequently, we proposed a spatio-temporal deep learning framework for predicting energy consumption in IoT-based enabled smart buildings. The framework integrates a conceptual support capturing the spatial hierarchy (building–floor–room–equipment) and temporal dynamics of IoT entities, a multi-layer system architecture linking perception, network, data processing, and application layers through an Energy Management System (EMS) capable of alert generation, thresholds configuration, and visualization, and a hybrid CNN2D–LSTM prediction model, that learns spatial correlations between zones and temporal dependencies in energy use, enabling accurate and adaptive forecasting.

Although the framework is generic and applicable to any building type, it was instantiated and validated through a smart hospital case study, chosen for its energy-intensive and operationally critical nature. Hospitals represent one of the most demanding building typologies, where continuous operation, medical equipment heterogeneity, and strict comfort constraints make energy management a particularly challenging task. Using realistic simulated datasets covering five blocks of hospital building, during a period of 60-days, we demonstrated that our approach consistently outperforms conventional models such as ANN, CNN, RNN, LSTM, etc., as well as recent hybrid models from the literature; CNN-LSTM-K-means²⁷ and CNN-LSTM.¹¹ The CNN2D–LSTM achieved high predictive accuracy ( $RMSE = 1.61$ , $MAE = 1.12$ , $MAPE = 9.31 %$ , $R^{2} = 0.76$ ), proving its effectiveness across diverse scenarios including normal operations, under-consumption periods (night/weekend), and extreme events such as heatwaves and pandemics. The model not only followed the overall energy dynamics more closely but also converged faster toward real consumption values, showing high robustness across different operational contexts.

Beyond quantitative performance, this work contributes to the realization of Green IoT principles by supporting sustainable and adaptive energy management. The conceptual EMS integrates alert mechanisms, dynamic thresholds, and dashboard functionalities that help managers reduce unnecessary consumption, detect anomalies early, and maintain comfort while minimizing environmental impact. The predictive capabilities of the CNN2D-LSTM framework thus enable proactive control of energy, fostering both operational efficiency and sustainability within healthcare infrastructures. Despite the promising results obtained by the proposed CNN2D-LSTM model, this study presents certain limitations. The dataset used in this work was generated through simulation and may not fully represent the complexity and variability of real smart hospital environments. Factors such as unexpected equipment failures, emergency situations, human behavioral variability, and real-time operational constraints remain difficult to reproduce accurately in simulated scenarios.

For future work, several research directions are envisaged:

–

Test the CNN2D–LSTM framework on real hospital datasets, including long-term monitoring data, to validate scalability and generalization, under real operational conditions.

–

Extend the prediction framework to also incorporate other operational dimensions such as indoor air quality, comfort levels, and equipment reliability.

–

Extend the framework towards real-time optimization and control of energy flows using bio-inspired algorithms such as the Slime Mould Algorithm to dynamically adjust system parameters based on predictive insights.

–

Couple prediction with models of solar energy generation and battery storage to enable cleaner and more resilient energy ecosystems in hospitals.

–

Model the behavior of energy resources in smart building systems, because it allows to simulate scenarios involving system failures or high demand, and thus to design robust strategies. By taking a systemic approach to energy resource behavior, we can reduce both costs and the carbon footprint, thus contributing to sustainability, with is part of Green IoT.

Overall, this study provides a robust, interpretable, and sustainable foundation for intelligent energy prediction and management in IoT-enabled buildings. By demonstrating the practical implementation of a Green IoT approach through the smart hospital case, it bridges the gap between conceptual modeling, deep learning prediction, and environmentally conscious energy optimization, paving the way for the modern self-adaptive and sustainable smart infrastructures.

Footnotes

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

ORCID iDs

Benmessaoud Lylia

Boukhedouma Saida

References

Vadruccio

Siragusa

Tumino

. Increasing energy efficiency in smart buildings through Internet of Things retrofitting intervention. Procedia Comput Sci 2023; 219: 263–270.

Silva

BVF

Holm-Nielsen

Sadrizadeh

, et al. Sustainable, green, or smart? Pathways for energy-efficient healthcare buildings. Sustain Cities Soc 2024; 103: 105511.

Aliero

Asif

Ghani

, et al. Systematic review analysis on smart building: challenges and opportunities. Sustainability 2022; 14: 3009.

Ebeed

Hassan

Kamel

, et al. Smart building energy management with renewables and storage systems using a modified weighted mean of vectors algorithm. Sci Rep 2025; 15: 4733.

Agarwal

Balaji

Gupta

, et al. Occupancy-driven energy management for smart building automation. In: Proceedings of the 2nd ACM workshop on embedded sensing systems for energy-efficiency in buildings (BuildSys ’10), New York, USA, 2010, pp. 1–6. DOI: 10.1145/1878431.1878433.

Agency

Buildings

. Construction: tracking report-buildings, https://globalabc.org/resources/publications/iea-tracking-report-buildings (2022, accessed 20 November 2025).

Mariano-Hernández

Hernández-Callejo

García

, et al. A review of energy consumption forecasting in smart buildings: methods, input variables, forecasting horizon and metrics. Appl Sci 2020; 10: 8323.

BENMESSAOUD

BOUKHEDOUMA

. Real-time monitoring and optimization of energy consumption in smart buildings: case of hospital. In: Proceedings of the 5th International Conference on Emerging Smart Technologies and Applications, Ibb, Yemen 2025: 1–8.DOI: 10.1109/eSmarTA66764.2025.11132162

Divina

García Torres

Goméz Vela

, et al. A comparative study of time series forecasting methods for short-term electric energy consumption prediction in smart buildings. Energies 2019; 12: 1934.

10.

Zini

Carcasci

. Machine learning-based energy monitoring method applied to the HVAC systems electricity demand of an Italian healthcare facility. Smart Energy 2024; 14: 100294.

11.

Zhiqian

Yan

Chenglong

, et al. Energy consumption prediction for HVAC systems in public buildings based on SSA-CNN-LSTM neural network. J Phys: Conf Ser 2025; 3074: 012019.

12.

Zarco-Soto

Sharif Ali

, et al. Energy consumption in buildings: a compilation of current studies. Energy Rep 2025; 13: 1293–1307.

13.

Shah

Nasir

Fayaz

, et al. A review on energy consumption optimization techniques in IoT-based smart building environments. Information 2019; 10: 108.

14.

Hakawati

Mousa

Draidi

. Smart energy management in residential buildings: the impact of knowledge and behavior. Sci Rep 2024; 14: 1702.

15.

Keytingan

Ramli

Awalin

. Energy consumption prediction by using machine learning for smart building: case study in Malaysia. Dev Built Environ 2021; 5: 100042.

16.

Yang

, et al. Energy-efficiency-oriented scheduling in smart manufacturing. J Ambient Intell Humaniz Comput 2019; 10: 969–978.

17.

Attoue

Shahrour

Younes

. Smart building: Use of the artificial neural network approach for indoor temperature forecasting. Energies 2018; 11: 395.

18.

Liu

Ansari

. Toward green IoT: energy solutions and key challenges. IEEE Commun Mag 2019; 57: 104–110.

19.

Goudarzi

Anisi

Soleymani

, et al. An IoT-based prediction technique for efficient energy consumption in buildings. IEEE Trans Green Commun Netwo 2021; 5: 2076–2088.

20.

Alduailij

Petri

Rana

, et al. Forecasting peak energy demand for smart buildings. J Supercomput 2021; 77: 6356–6380.

21.

Fumo

Biswas

MAR

. Regression analysis for prediction of residential energy consumption. Renew Sustain Energy Rev 2015; 47: 332–343.

22.

Korolija

Zhang

Marjanovic-Halburd

, et al. Regression models for predicting uk office building energy consumption from heating and cooling demands. Energy Build 2013; 59: 214–227.

23.

Nie

Roccotelli

Fanti

, et al. Prediction of home energy consumption based on gradient boosting regression tree. Energy Rep 2021; 7: 205–213.

24.

González-Vidal

Ramallo-González

Terroso-Sáenz

, et al. Data driven modeling for energy consumption prediction in smart buildings. In: Proceedings of the IEEE international conference on big data (big data), Boston, USA, 2017, pp. 4562–4569. DOI: 10.1109/BigData.2017.8258499.

25.

Cao

Zhang

, et al. Electrical load prediction of healthcare buildings through single and ensemble learning. Energy Rep 2020; 6: 1681–1693.

26.

Guo

Liu

Zheng

. Early energy performance analysis of smart buildings by consolidated artificial neural network paradigms. Heliyon 2024; 10: 25848.

27.

Somu

Raman

MRG

Ramamritham

. A deep learning framework for building energy consumption forecast. Renew Sustain Energy Rev 2021; 137: 110618.

28.

Durand

Aguilar

R-Moreno

. An analysis of the energy consumption forecasting problem in smart buildings using LSTM. Sustainability 2022; 14: 13358.

29.

Rehman

Iqbal

. Electricity consumption prediction in smart buildings using deep learning approaches. Energy Inf 2025; 8: 131.

30.

Anan

Kanaan

Benhaddou

, et al. Occupant-aware energy consumption prediction in smart buildings using a LSTM model and time series data. Energies 2024; 17: 6451.

31.

Binbusayyis

Sha

. Energy consumption prediction using modified deep CNN-Bi LSTM with attention mechanism. Heliyon 2025; 11: e41507.

32.

Bawaneh

Ghazi Nezami

Rasheduzzaman

, et al. Energy consumption analysis and characterization of healthcare facilities in the united states. Energies 2019; 12: 3775.

33.

Nourdine

Saad

. Energy consumption in hospitals. In: Proceedings of the 2020 international conference on electrical and information technologies (ICEIT), Rabat, Morocco, 2020, pp. 1–6. DOI: 10.1109/ICEIT48248.2020.9113177.

	Type of	IoT		Hybrid	Used	Data	Performance	Results
Reference	Building	devices	Category	model	methods	Source	metrics	obtained
Somu et al.²⁷	Academic building (smart rooms, offices, laboratories)	Raspberry Pi, sensors	Deep-learning	Yes	K-means, CNN, LSTM	Real historical data	MSE, RMSE, MAE, MAPE	Increased accuracy of predictions
Nie et al.²³	Smart home	Temperature, humidity sensors and household appliances	Machine-learning	Yes	Gradient Boosting Regression Tree (GBRT), ARIMA-GBRT	Real historical data	RMSE, MAE, MAPE	Reducing prediction errors
Keytingan et al.¹⁵	Commercial building	Smart meters	Machine-learning	Yes	K-NN, SVM, ANN	Real-time data	RMSE, NRMSE, MAPE	Increased accuracy
Zini and Carcasci¹⁰	Smart hospital	Smart meters, wireless sensors	Deep-learning	Yes	ANN, MLP	Real-time data	RMSE, R $^{2}$	Effective monitoring and anomaly detection
Attoue et al.¹⁷	University building	Temperature, humidity and solar radiation sensors	Deep-learning	No	ANN	Real historical data	MSE, R $^{2}$	Accurate temperature prediction, improved comfort
Goudarzi et al.¹⁹	Commercial building	Raspberry Pi, sensors	Mathematical	Yes	ARMI, ICA, EWMA	Real-time data	RMSE, MAE, ARE	Reduction of energy loss
Alduailij et al.²⁰	Government building	Smart meters	Mathematical	Yes	ARIMA, TBATS, ANN, LSTM	Real-time data	MAPE, Accuracy	Efficient management of energy peaks
Fumo and Biswas²¹	Smart home	/	Mathematical	Yes	Simple, multiple and quadratic regression	Real-time data	RMSD, R $^{2}$	More accurate estimation of energy consumption
Korolija et al.²²	Office building	/	Mathematical	No	Regression model (quadratic and power)	Simulation (Enegry Plus)	R $^{2}$ , RMSD, Residual errors	Fast and accurate prediction of energy needs
González-Vidal et al.²⁴	Smart building	/	Machine-learning	Yes	SVM, RF, XGBoost	Real historical data	/	Improved prediction accuracy
Cao et al.²⁵	Smart hospital	Smart meters	Machine-learning	Yes	ANN, RF	Real-time data	MAPE, R $^{2}$	Increased accuracy
Durand et al.²⁸	Institutional building	/	Deep-learning	No	LSTM	Real historical data	RMSE, MAPE, R $^{2}$	Increased accuracy of predictions
Rehman and Iqbal²⁹	Residential Building	Smart meters	Deep-learning	Yes	LSTM, Seq2Seq	Historical data	MAE, RMSE	Efficient energy management
Anan et al.³⁰	Commercial building	Sensors	Deep-learning	No	LSTM	Real-time data	MAPE, MAE, RMSE	Improved prediction accuracy
Guo et al.²⁶	Residential Building	/	Machine-learning	No	ANN	Simulation (Tsanas & Xifara)	RMSE, MAE, MAPE	Improved prediction accuracy
Binbusayyis and Sha³¹	Residential Building	Sensors, household appliances	Deep-learning	Yes	CNN, Bi LSTM, Attention	Historical data	MSE, MAE, MAPE, RMSE	Improved prediction accuracy and energy management optimization
Zhiqian et al.¹¹	Smart Building	HVAC sensors, environmental sensors	Deep-learning	Yes	CNN, LSTM	Historical data	MAE, MAPE, R $^{2}$	Improved prediction accuracy