Multidimensional situational information fusion method for energy saving on campus

Abstract

In order to solve the problem of resource waste in colleges and universities, this paper proposes a multidimensional situational information fusion method, which can be used to normalize, analyze and predict the multi-source data such as natural, humanistic and spatio-temporal data on campus so as to meet the application requirements for high-level decision-making. With this method, firstly, the event object model is used to normalize multi-source data. Then, the multidimensional situational information fusion mechanism of twice reasoning is used to obtain the real-time situation and equipment control scheme of the campus so that real-time intelligent semantic understanding is realized. In the process of reasoning, the improved KNN prediction model is used to predict situational trends, and the prediction information is used to continue deep reasoning and mining. Finally, the real-time energy-saving regulation is carried out through control instructions. In addition, through simulation verification, experimental results show that this method can quickly identify, integrate, and predict the current real-time situation and generate reasoning results, and finally achieve the goal of intelligent control for energy saving on campus.

Keywords

Smart campus energy saving situational information fusion intelligent control

1 Introduction

With the rapid development of global economy, the problem with energy utilization has gradually aroused the attention of all countries in the world [1, 2]. As the campus has a large population and large building area, it has always been a major energy consumer [3]. In most schools, waste of resources is often seen. Researches show that more than two-thirds of campus energy consumption is used for room heating (cooling), water heating and lighting [4, 5]. The utilization efficiency of resources depends on whether a school manages energy consumption properly. For example, in respect of real-time control, if indoor temperature and brightness in classrooms, libraries, and offices can be adjusted intelligently, not only pleasant learning and work environment can be created, but also waste of resources can be avoided.

There have been a lot of researches on issues such as campus energy-saving and intelligent control. These researches focus on the analysis and application of six major factors, such as nature, humanity, energy-saving materials (or renewable energy), time, space and energy operation. Natural factors include weather [6, 7], solar irradiance [8] and indoor physical conditions [9 –11]. Humanity factors include behavioral habits [12 –15], views and values [16, 17] and policies and systems [18]. Factors of energy-saving materials (or renewable energy) include whether solar energy resources [19] or LED energy-saving bulbs [20] are used. Time factors include course arrangement [21], academic calendar [22] and time and steps [23]. Spatial factors include spatial location relationship between building and equipment, between buildings, and between equipment [24 –26] and space occupancy rate [23]. Factors of energy operation include whether energy operates exceptionally [27] and the mode and law of energy operation [28, 29].

Some scholars have made comprehensive analysis and research of two or three of the above-mentioned influence factors. For example, (1) Nature+Energy Operation: P. Shen et al. [7] proposed a fast multi-target optimization algorithm based on future weather conditions, used for the transformation and planning of the campus buildings of UPenn. W.-J. Shyr et al. [9] developed an energy management system (EMS) and used light sensors to sense campus brightness, which is used for lighting control in Taiwan University. L. Yung-Hsin et al. [11] used the analytic hierarchy process (AHP) to analyze physical condition factors, proposed an initial cloud energy management system, and used it to manage energy consumption in colleges and universities. (2) Humanity+Time+Space: J. Han et al developed a data driver to detect the behavior of teachers and students and the temporal and spatial factors involved in the behavior and found that the behavior changes regularly, which is of great significance for campus energy saving [30]. (3) Humanity+Energy operation: C. Sheng-Luen et al proposed a GAM requirement model based on human routine activities and trained the model by using the historical power consumption data of a university campus in 2014. The research can accurately predict the power consumption from 2015 to 2017 [22]. (4) Energy-saving materials (or renewable energy)+Space+Nature: H. Ahmed A. A. et al. [8] measured and analyzed the solar irradiance on the main campus of the university and found that the photovoltaic (PV) system installed on the roofs of the buildings on the main campus had good energy-saving effect. (5) Energy-saving materials (or renewable energy)+Space: J. Hoyo-Montaño transformed and upgraded the light equipment and air conditioning equipment in the university, changed the previous fluorescent T8 tubes into efficient LED T8 tubes, and rearranged the spatial locations of the equipment, thus saving 36.42% of energy, compared with the past [20]. (6) Humanity+Nature: B. Hamzah et al. [10] surveyed students’ comfort at different temperatures, collected 175 questionnaires, and computed energy consumption for cooling by using the EnergyPlus software. According to the research, if the air conditioning temperature is increased from the minimum 25.0°C to the minimum 26.0°C, it can not only make teachers and students feel comfortable, but also save energy. E. Mohd Ahnuar et al. [16] conducted a questionnaire survey on the problems that might arise in the construction of green campus (divided into five aspects including economy, organization, natural conditions, behavior and technology) and calculated that the main factors hindering the sustainable development of the university were the shortage of funds and the lack of management experience. S. Allen et al. [18] investigated the impact of students’ behavior, natural environment and policy support on energy conservation, built the Attitude-Behavior-Constraint (ABC) model, and applied the model to the researches of students’ behavior of saving energy. To sum up, the diversified campus environment are not only related to natural factors, but also to space, time and human factors. If the system does not take them into full consideration, it will lead to a significant decline in energy saving effect, and even waste. For example, if the system only considers that the classroom is not bright enough and turns the lights on, but it does not perceive that the classroom is not in class (i.e., the classroom is not occupied), it will cause a waste of electric energy. Therefore, the integration of multi-dimensional information will be conducive to the comprehensive analysis of various situations so as to save campus energy as much as possible. However, current researches focus on the comprehensive analysis of three or fewer kinds of influencing factors, but fail to fully integrate and utilize more relevant information. In addition, current researches fail to deal extensively and deep enough with the factors affecting energy conservation. Especially, there are few researches on the quantitative analysis of humanistic factors.

The rapid development of internet of things and contextual computing technology has provided good support for solving the problem of insufficient situational information fusion and made it possible to realize real-time resource control and energy-saving prediction [31 –33]. Therefore, an architecture based on contextual computing is constructed in this paper to comprehensively integrate data that may affect campus energy consumption. However, considering that (1) Upgrading facilities will cost a lot of money, which is only applicable to schools with more funds; (2) Energy operation factors are mainly used for detection of abnormal behavior of energy utilization and analysis of energy use characteristics in a wide range, which have little impact on researches of situational energy conservation in a small range, the factors of energy-saving materials (or renewable energy) and energy operation are not included in the discussion of this paper for the time being.

For the reasons above-mentioned, this paper will propose a multi-dimensional information fusion method based on the nature, humanity, time and space and build a set of system approaches ranging from underlying data acquisition and processing, multidimensional information analysis and mining, real-time situational fusion, situational trend prediction to real-time control of energy use equipment through condition factors such as physical laws of nature, laws of climate change, behavior and customs, geographical constraints and temporal and spatial transformation. The method proposed in this paper can not only realize the effective energy-saving management of intelligent campus and reduce the occurrence of waste, but also provide teachers and students with intelligent environmental services in an active and real-time manner and create a good learning atmosphere.

The main contributions of this study are as follows: (1) An efficient reasoning mechanism of multi-dimensional information fusion is proposed; (2) An improved machine learning algorithm is proposed and applied to prediction of situations; (3) The overall framework of campus energy saving control is given. The following content is organized as follows: Firstly, in the second part, the overall framework and operational mechanism of energy-saving control based on situation fusion reasoning will be introduced. Then, the three layers of the frame are described respectively in bottom-up order. That is, in the third part, the method of collecting and processing the underlying data are introduced. In the fourth part, the internal structure and operation mechanism of multi-dimensional situation information fusion reasoning are given, and at the same time specific examples of energy-saving control are given. Finally, the fifth part is about our detailed experimental verification, and the research work is summarized in the sixth part.

2 Overall architecture of intelligent campus energy conservation control

Figure 1 shows the overall architecture of intelligent campus energy conservation control. The architecture is divided into three layers: The data acquisition layer first collects the underlying original data through hardware and software interfaces [34, 35], and at the same time normalizes the underlying data, providing data in standard input format for the upper data analysis. The data analysis layer analyzes, deduces and forecasts multi-source data to obtain real-time situation and trend after fusion. The equipment control layer generates the final control scheme through the real-time situation and its trend state and the equipment control rules, and controls the equipment accurately and in real time through the hardware and software interfaces [36].

3 Collection and processing of underlying data

3.1 Collection of underlying data

The multidimensional data sources used in this paper are real-time sensor data, computer spatio-temporal data, internet environment data and human geographic data. Real-time sensor data include indoor temperature, indoor brightness, soil moisture, electricity consumption and water flow velocity, coming respectively from intelligent remote monitor (See Fig. 2, embedded temperature sensor, brightness sensor and soil moisture meter), electricity meter and flow velocity sensor; the computer spatio-temporal data include the information of current time, regional location, classroom usage and holidays, coming respectively from the computer time system, regional location labeling table, classroom usage schedule (e.g., class schedule and examination classroom arrangement) and national holidays table. Internet environmental data include outdoor temperature, outdoor humidity, weather conditions, atmospheric pressure, wind direction and wind speed, which come from the web weather forecast WEB service; human geographic data include customs, institutional culture, climate and geographical location, which are generated by semi-automatic rule modeling.

3.2 Object-based data stream model

Since the underlying data come from various sources and the structure is heterogeneous, it is necessary to normalize the data format before making data analysis [37, 38]. This paper provides that formatted data is stored in the form of event object. The basic format is: event={“measurement device number”: DEVICE_NUM, “measurement information type”: DEVICE_TYPE, “measurement device location”: DEVICE_LOCATION, “measurement time”: DEVICE_TIME, “measurement value”: DEVICE_VALUE}, where, curly braces ({}) represent an event; quotes (“ ”) indicate the event property/state; the colon (:) is followed by the attribute value/status value; multiple event properties, including their attribute values, are separated by commas (,). Multiple event objects of the same type in the same situation are encapsulated as a stream of events by brackets ([]), separated by commas (,). The value types of the event objects are qualified to character strings and numbers. The event stream structure is shown in Fig. 3.

For example, the data stream structure based on temperature and brightness is: [{“Device-ID”: 000100cG, “Type”: temperature, “Location”: Classroom001, “Time-Stamp”: 2019-10-22 08 : 38 : 39, “Value”: 15°C}, {“Device-ID”: 020707bN, “Type”: luminance, “Location”: Classroom001, “Time-Stamp”: 2019-10-22 08 : 38 : 02, “Value”: 448.62lux}].

Since the normalized data stream contains a variety of unordered data objects, there must be a problem of low object extraction efficiency. Therefore, the data acquisition layer must classify, sort and reorganize the data stream. First, the data acquisition layer classifies each object in the data stream by type attribute and stores the results in the corresponding parallel type queue. Then, the layer goes through each type queue one by one, takes out data objects in the fixed sliding window (e.g., 5 minutes for one window) from a single type queue for spatial classification, puts objects of the same area or floor into the same spatial queue and sort them by location number. Finally, the objects in each spatial queue, with one sliding window as a unit, are extracted as the information in a certain situation.

4 The multidimensional situational information fusion mechanism of twice reasoning

4.1 The classification and establishment of the situation of intelligent campus

Due to great difference in the utilization patterns of campus space resources, classification of situations according to the spatial nature will be conducive to the refined and personalized treatment of campus situations. Therefore, this paper sets intelligent campus as the situational subject, and divides it into eight sub-classes of second-level situational subject such as classroom, dormitory, school road, library, dining hall, bath hall, greenbelt and office area. These sub-classes are composed of five types of basic state information such as nature, humanity, time, space and equipment. The specific classification of situations is shown in Fig. 4.

Meanwhile, campus environment subjects in reality are all in their unique natural, humanistic and space-time background, and these factors can directly or indirectly determine people’s activities, thus affecting the use of resources [39]. For instance, in winter, because of the big difference in climate and living habits between the south and the north of China, people in the north use central heating while people in the south use air conditioning. So college libraries in the south need to pay attention to the energy-saving behavior of air conditioners in winter, while college libraries in the north need to pay attention to the energy-saving scheme of central heating [40, 41]. Therefore, in the process of multidimensional information fusion, it is necessary to consider its hierarchical structure and reason according to priority relationships. As shown in Fig. 5, the information in the outermost layer has the highest priority, which can be referred to as the background of inner layer information. Only when the inner layer information is fused with the result of the outer layer information fusion, can it continue to be processed.

4.2 Construction of intelligent campus situational ontology

As ontology can well express the relevant knowledge of the campus [42], this paper establishes the campus situational ontology as shown in Fig. 6, including five major parts: time, space, nature, humanity and equipment ontology. (1) The time ontology includes three aspects such as ‘unit’, ‘day and night’ and ‘season’, among which, the unit includes year, month, day, hour, minute and second; day and night include day (morning or afternoon) and night; and seasons include spring, summer, autumn and winter. (2) Space ontology includes four attributes such as space number, space type, space occupancy and space name, among which, space name is defined by space number and space type. (3) The natural ontology is composed of two second-level ontologies, physical measurement and geographical nature, and has two types of values, namely real time value and trend value, which coexist in a specific space-time. The physical measurement body consists of five measurement aspects: flow velocity, electric quantity, humidity, temperature and brightness, which all have three basic states (high, medium and low). Geographical nature ontology consists of geographical features and weather features. (4) Humanistic ontology consists of four second-level ontologies, namely, living habits, human perception, time arrangement and human behavior. There are two types of values, namely, real time value and trend value, which coexist in specific space-time and natural environment. Life habit ontology includes eating habit (including eating peak and eating low peak), bathing habit (including bathing peak and bathing low peak) and self-study habit (including self-study peak and self-study low peak). Body perception includes temperature perception (in three states: hot, comfortable and cold), brightness perception (in three states: bright, comfortable and dark), soil moisture perception (in three states: dry, comfortable and wet) and none (perception with nobody there). The ontology of time arrangement also includes three three-level ontologies such as course arrangement, holiday arrangement and office arrangement. Among them, the course arrangement ontology includes three attributes, namely course name, class location and class starting and ending time; the holiday arrangement ontology includes two attributes, namely holiday name and holiday starting and ending time; and the office arrangement includes two attributes, namely office location and starting and ending time of office. The ontology of students’ behavior consists of five behavioral states, namely, class, meal, bath, self-study and office, which all have two basic attributes (proceeding and stop). (5) The device ontology is composed of four parts such as number, type, state and use condition. The state includes on and off. The equipment can be turned up or down in the state of ‘on’, and the use condition includes normal and abnormal.

4.3 Generation method of semi-automatic inference rules

In this paper, ontological reasoning [42] is applied to intelligent campus multidimensional situation fusion, which can derive high-level situations from simple situations, thus obtaining rich and valuable information [43]. Therefore, based on the above situation subjects, this paper constructs the following general expression of rules for the construction of inference rules. This general expression means that under a certain usage situation (UsageSituation), if the known condition is CONDITION, then the inference result is RESULT.

IF { [?CONDITION] } THEN {[?RESULT]}.

As a great number of rules are required for campus situation reasoning, if the semi-automatic method is used to generate reasoning rules, it will be conducive to reducing the workload and improving the operability of the system. Three functional modules to add reasoning rules are designed and realized in this paper. The specific principles are as follows:

(1) Automatic parsing module: for text information with regular syntax, the parser is used to construct the syntax tree, and then CONDITION and RESULT information are extracted; (2) Information extraction module: for web page information (class table, holiday table, etc.), BeautifulSoup library [44] (i.e., webpage crawling library of python) and regular expressions are used to parse web page and extract key information (CONDITION and RESULT); (3) Manual input module: for the information that cannot be entered automatically, such as unstructured text information such as survey data, the system administrator can input it manually through the software interface. The following are two representative examples in the rule base respectively established automatically and manually. See Table 1.

Table 1
Some examples in the rule base

Serial number UsageSituation CONDITION RESULT Rule-building Approach

1 [@CLASSROOM] nature.temperature.value≥29 nature.temperature.state.value=‘high’ manually

2 [@CLASSROOM] time.hour.value BETWEEN [9, 10] AND location.regionName=‘classroom001’ humanity.classing.value=’classing’ automatically

Serial number	UsageSituation	CONDITION	RESULT	Rule-building Approach
1	[@CLASSROOM]	nature.temperature.value≥29	nature.temperature.state.value=‘high’	manually
2	[@CLASSROOM]	time.hour.value BETWEEN [9, 10] AND location.regionName=‘classroom001’	humanity.classing.value=’classing’	automatically

4.4 Multidimensional situational information fusion mechanism

In this paper, a multidimensional situational information fusion mechanism twice reasoning process (TRP-MSFM) is designed to achieve campus energy conservation control. Its operation process is shown in Fig. 7. Firstly, the normalized event stream is input into the Situational Reasoning Engine (SRE) for reasoning. If there is a situation development trend and the trend can affect the control of the equipment, the Situation Trend Prediction Machine (STPM) is called for computation, and the generated situation trend state is constructed as a new event stream, which is re-input to SRE to continue reasoning. The purpose of constructing this feedback process is to solve the conflict between current and future situation states. For example, if the library is now closed, it is recommended to set the HVAC to ‘off’ (the real-time situation makes the energy saving decision), but if the library will be open ten minutes later and the room temperature is low, it is recommended to set the HVAC to ‘on’ (the future trend situation makes the energy use decision). Obviously, two opposite conclusions will be generated in the system at this time. Therefore, this feedback process needs to be used to solve this problem: Suppose that SRE reasoning results suggest that device_x be set to a, and STPM reasoning results suggest that device_x be set to b (values of a and b are 0 and 1, 0 meaning off, 1 meaning on), then the final setting of device_x should be a∪b. That is, the final setting of the HVAV in the library is 0∪1 = 1 (on). The processing allows the library to adjust the temperature to the right range before it opens. Then, TRP-MSFM then compares the reasoning results with the cache information (save other situation information except the state of the device, used for judging the change of the situation): If equal, the situation and the execution device state remain unchanged and no operation is done on the device; if not equal, the reasoning result will be updated to the situation and the cache information will be updated accordingly. Finally, the situation information will be pushed to the Equipment Control Reasoning Engine (ECRE) for reasoning, and the equipment will be controlled according to the results. Meanwhile, the new state of the equipment will be updated to the situation.

In addition, SRE and ECRE are supported by different rule bases. The applicable rules of SRE are situational inference rules, the condition is known situational information, and the result is implied situational information. The applicable rules of ECRE are equipment control rules, the condition is situational information, and the result is equipment control scheme.

In order to simplify writing, many abbreviations are used in this paper. For the convenience of memory and search, the following summarizes the abbreviations used throughout the paper. See Table 2.

Table 2

Comparison table of full names and their abbreviations

Full name	Abbreviation
a Multidimensional Situational information Fusion Mechanism Twice Reasoning Process	TRP-MSFM
the Situational Reasoning Engine	SRE
the Situation Trend Prediction Machine	STPM
the Equipment Control Reasoning Engine	ECRE
an improved KNN prediction model combined with Kmeans	KM-KNN
Prediction Model Name	PMN
the Model Input Feature Preprocessor	MIFP

4.5 The internal structure of STPM

Machine learning (ML) can significantly reduce the time of prediction calculation, and it is increasingly applied to building situation analysis [45, 46]. Therefore, this paper uses the K-Nearest Neighbor prediction model (KNN) in ML library that supports regression prediction well to predict various possible situational trends.

As KNN is a lazy learning method, its prediction process takes a long time when the training set is large. Therefore, based on the hierarchical relationship among multidimensional situational information, this paper proposes an improved KNN prediction model combined with Kmeans KM-KNN to accelerate KNN’s prediction speed. The core idea is as follows: First, Kmeans algorithm is used to divide the training set D into n classes such as D₁, D₂, ... , D_n, and at the same time store their clustering centers (d₁, d₂, ... , d_n) and classification data set (D₁, D₂..., D_n) in the cache. The clustered information and the clustering centers do not contain output tags in order to classify and recognize geographical background features (such as geographical location and weather). Then read the cluster information in the cache and calculate the Euclidean distance dist₁, dist₂, ... , dist_n between the test sample and the clustering centers d₁, d₂, ... , d_n, and take D_i corresponding to the minimal dist_i, i∈[1,n]. Finally, put test samples into D_i for KNN prediction, namely, sort in ascending order the Euclidean distances calculated between test samples and all training samples in D_i, take the first k training samples (k value is the quadratic root of D_i), and take the average value of their real value output tags as the prediction result.

Take Kmeans’ clustering of weather information in different regions as an example. According to Fig. 8, the outdoor temperature and humidity of Xi ‘an and Guangzhou in January (standardized by z-score) can be effectively divided into two categories with different weather characteristics by Kmeans for KNN to use later. Therefore, the improved KNN algorithm is as follows:

The internal structure diagram of KM-KNN-based STPM is given below.

As shown in Fig. 9, the current situation information and Prediction Model Name (PMN) invoked by the rules will be input to STPM, and STPM will match the corresponding Model Input Feature Preprocessor (MIFP) according to PMN. MIFP sorts the current situation information into model input feature vector. STPM inputs data according to KM-KNN model matching and corresponding to PMN and returns the final prediction result to SRE.

Different types of KM-KNN predict different situational trend objects, which can include adjusting the air conditioning/heating equipment to the expected time required to reach the comfortable temperature in the target area and adjusting the lawn sprinkler to the expected time needed for proper humidity in the target area.

Analyze the example of TRP-MSFM with the campus library as the object, as shown in Fig. 10.

SRE first reasons the situational information and learns that the library is currently closed. Meanwhile, in the process of reasoning, it finds that the library is about to open at the next moment and perceives that the temperature is hot. Therefore, the library STPM1 needs to be called to predict the time required for air conditioning to adjust indoor temperature to the comfortable temperature. The training set is derived from the historical data of a certain period of time in a certain year (such as summer in Xi ‘an). Among them, the operational mode of the library STPM1 is as follows: First, use Kmeans to complete weather characteristics classification and recognition; then use KNN to predict ‘the time needed for the indoor temperature to be adjusted to the proper temperature’ (rounded up). Historical time is the time required by air conditioner to adjust indoor temperature to the target temperature, and the most recent matching value is taken. This feature can avoid the influence of external factors such as the aging of the air conditioner on the prediction results, and finally returns to the result that ‘temperature adjustment can be completed in 15 minutes’. For this purpose, ECRE suggests that the air-conditioner be turned on at 8 : 15.

5 Experiments

Python and Java are used as the main development tools in the experiment. Python is used for the construction of the ontology library, rule library, inference engine and neural network, and Java is used for the development of the system interfaces. The experimental testing period is four weeks, and the experimental location is the author’s school. The source and acquisition method of the original data are described in section 3.1.

Weather forecast, humanity, and situational data are stored using Mysql databases, while rules are stored using CSV files.

The hardware environment is the desktop computer with 4GB memory and 3.30 GHz CPU i5-4590. For the software environment, pycharm+anaconda under windows7 are used as the compiling and running environment, and the prediction model is built by using the sklearn library. In addition, in order to verify that the system software can process complex campus situation information in real time, so as to achieve the goal of timely response, we will focus this group of experiments on the operational efficiency of the fusion and prediction method proposed in this paper.

In terms of data set, data of four common situations on smart campus such as classroom, school road, library and lawn are extracted in this group of experiments, including (1) data of the tested classroom and library area: temperature (two temperature sensors in each area), brightness (three brightness sensors in each area), time, weather, curriculum schedule and whether there are people (two infrared sensors in each area); (2) data of the tested lawn area: soil moisture (four or more soil moisture sensors in each area), weather and time information. In addition to the above data, Experiment 3 also includes a historical database, which stores historical data over a period of four weeks in each of the four situations, as well as the time required for the refrigeration or heating equipment to reach the desired goal.

5.1 Experiment 1

Due to the concurrency of the underlying data in the real-time transmission process, the data acquisition layer processing described in Section 3.2 is simulated in this experiment to verify the timeliness of the module. Figure 11 shows two processing methods (i.e., the square represents method 1 and the dot represents method 2). Method 1 firstly classifies events by type, then sorts events by their occurrence location (number), and finally reorganizes events in the same situation with a sliding window. Method 2 directly extracts the events in the cache queue to reorganize. As the number of events increases, the processing time of the two methods also increases, but Method 1 takes less time, increases at a smaller rate, and improves the operation efficiency by about 46.82%. In addition, it can be seen from Fig. 11 that the running time of Method 1 is composed of the primary cache of system I/O, but it takes less time to compute for the processing of data stream.

5.2 Experiment 2

Figures 12 and 13 describe respectively the average reasoning time for SRE+ECRE to process streams of different numbers of events with caches done and with caches undone. By comparing the processing time curves with caches done and with caches undone, we can see that the reasoning speed with caches done is significantly faster than that with caches undone. In the case of 300 rules and 2,500 event streams to be processed, the reasoning speed with caches done is less than 10 seconds, which can achieve the goal of fast reasoning. As it takes some time to initialize the reasoning engine, the broken line slope is larger when the rule number is 0-50. This experiment shows that we can greatly reduce the number of situational reasoning by caching the situational information in the previous period, thus saving the total reasoning time of TRP-MSFM.

5.3 Experiment 3

In this experiment, the performance of KM-KNN is simulated and compared with traditional KNN, Bayesian Ridge and Decision Tree. Approximately 7,000 pieces of relevant situational data are used for training of each prediction model, and the prediction time consumption, memory occupation, prediction accuracy (R-squared), Mean Square Error (MSE) and Mean Absolute Error (MAE) are recorded.

Where, $R - square = 1 - \frac{\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}{\sum_{i = 1}^{n} (y_{i} - \bar{y})^{2}}$ , $MSE = \frac{\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}{n}$ , $MAE = \frac{\sum_{i = 1}^{n} | y_{i} - {\hat{y}}_{i} |}{n}$ , n represents the total number of the situations, y_i represents the real value of tag i (from the historical database, such as the time required for refrigeration or heating equipment to reach the expected goal), ${\hat{y}}_{i}$ represents the predicted value of tag i (from the predicted result of the machine learning algorithm), and $\bar{y}$ represents the mean value of the tag, $\bar{y} = \frac{\sum_{i = 1}^{n} y_{i}}{n}$ , 1 ≤ i ≤ n, i ∈ Z, n ≥ 1, n ∈ Z.

According to Figs. 14 and 15, in terms of speed of prediction, KM-KNN is obviously faster than traditional KNN. With the increase of the test set, its running time increases slower than traditional KNN. In terms of memory usage, KM-KNN uses significantly less memory than traditional KNN, Bayesian Ridge and Decision Tree. This is because the Kmeans algorithm has classified the training set. KNN only needs to calculate the training samples belonging to a certain class, which reduces KNN’s calculation amount and further reduces the overall running time and memory space. The running memory size is related to the classification number of the Kmeans algorithm, and the more classification number, the less running memory.

The rates of curves in Fig. 15 are all close to 0, which indicates that the memory usage of these four types of models in the prediction stage mainly focuses on the loading and calling of models. Fig. 16 describes the results of 100 prediction experiments. Among them, the average prediction accuracy, average MSE and average MAE of KM-KNN are close to those of traditional KNN, and the average prediction accuracy is higher than that of Bayesian Ridge and Decision Tree. Meanwhile, the average MSE of KM-KNN is lower than that of Bayesian Ridge and Decision Tree. The average MAE is lower than that of Bayesian Ridge, but close to that of Decision Tree.

The results of this group of experiments show that KM-KNN is more efficient than traditional KNN and saves more memory than the other three algorithms. Meanwhile, KM-KNN has higher prediction accuracy and better generalization ability. In conclusion, KM-KNN can accurately predict the required implicit situational information in a short time.

6 Conclusion

In this research, we have analyzed and integrated spatio-temporal, natural and human information through semantic web and situational fusion technologies, so as to integrate the campus with nature, and achieve campus energy conservation through high-level understanding of the natural environment. Firstly, this paper has normalized the underlying data to ensure the consistency of data formats. Then, this paper has proposed a multi-dimensional situational information fusion mechanism based on space-time, nature and humanity, which includes the process of twice reasoning (situational reasoning and device control reasoning) and can generate semantically rich situational information and intelligent device control scheme. In addition, we have constructed STPM based on KM-KNN prediction method to predict the situation trend, and the prediction results are used for subsequent reasoning and equipment control. Moreover, we have presented in this paper the overall architecture of intelligent campus energy conservation control integrating the above comprehensive schemes. We have also tested the key software modules, and analyzed the data processing performance of three modules, namely the normalized processing, integrated reasoning and training prediction modules. The results of the experiments show that the processing performance of each test module is good. They can control energy using equipment quickly and accurately according to different situations, improve the efficiency of water and electricity utilization on campus, so as to achieve the expected energy saving goal. However, this paper does not consider the influence of individual behavior and habits on energy conservation, and the next research will be carried out in this direction.

Footnotes

Acknowledgment

This work was supported by Scientific Research Plan Projects of Shaanxi Education Department (Grant No. 17JK0376), Natural Science Basic Research Plan in Shaanxi Province of China (Grant No. 2019JM-484), Key Research and Development Project in Shaanxi Province of China (Grant No.2018GY-023), and Scientific Research Program Funded by Shaanxi Provincial Education Department (Grant No.19JC021) respectively.

References

Albayyaa

, Hagare

and Saha

, Energy conservation in residential buildings by incorporating passive solar and energy efficiency design strategies and higher thermal mass, Energy and Buildings 182 (2019), 205–213.

Wanjing

, Dongxu

, Yajun

, Wenqiang

and Kechang

, Research on the composite index of the modern chinese energy system, Sustainability 11 (2019), 150–167.

Popoola

S.I.

, Atayero

A.A.

, Okanlawon

T.T.

, Omopariola

B.I.

and Takpor

O.A.

, Smart campus: Data on energy consumption in an ICT-driven university, Data in Brief 16 (2018), 780–793.

Emeakarohaa

, Siang Anga

, Yana

and Hopthrowb

, A persuasive feedback support system for energy conservation and carbon emission reduction in campus residential buildings, Energy and Buildings 82 (2014), 719–732.

Schmidt

and Åhlund

, Smart buildings as cyber-physical systems: Data-driven predictive control strategies for energy efficiency, Renewable and Sustainable Energy Reviews 90 (2018), 742–756.

Ghojogh Najad

, Ahmad

and Safitri Zen

, Approach to environmental sustainability and green campus at Universiti Teknologi Malaysia: A Review, Environment and Ecology Research 6 (2018), 203–209.

Shen

, Braham

, Yi

and Eaton

, Rapid multi-objective optimization with multi-year future weather condition and decision-making support for building retrofit, Energy 172 (2019), 892–912.

Ahmed

, A.A. and Adel

, A feasibility study of PV installation: Case study at Shaqra University, 2018 9th International Renewable Energy Congress (IREC) (2018), 1–5.

Shyr

W.-J.

, Zeng

L.-W.

, Lin

C.-K.

, Lin

C.-M.

and Hsieh

W.-Y.

, Application of an Energy Management System via the Internet of Things on a University Campus, 14 (2018), 1759–1766.

10.

Hamzah

, Kusno

and Mulyadi

, Design of energy efficient and thermally comfortable air-conditioned university classrooms in the tropics, International Journal of Sustainable Energy 38 (2018), 382–397.

11.

Yung-Hsin

and Hsi-Chi

, Establishment of cloud energy management evaluation framework of campus buildings, 2018 IEEE International Conference on Applied System Invention (ICASI) (2018), 78–81.

12.

Konstantakopoulos

I.C.

, Barkan

A.R.

, He

, Veeravalli

, Liu

and Spanos

, A deep learning and gamification approach to improving human-building interaction and energy efficiency in smart infrastructure, Applied Energy 237 (2019), 810–821.

13.

Yao-Chung

and Huan

, The design and implementation of a low-cost power monitoring system for campus power monitoring, 2018 IEEE International Conference on Applied System Invention (ICASI) (2018), 720–721.

14.

Lin

and Azar

, Mixing work and leisure? Energy conservation actions and spillovers between building occupants at work and at home in the UAE, Energy Research & Social Science 47 (2019), 215–223.

15.

Kar

, Shareef

, Kumar

, Harn

K.T.

, Kalluri

and Panda

S.K.

, ReViCEE: A recommendation based approach for personalized control, visual comfort & energy efficiency in buildings, Building and Environment 152 (2019), 135–144.

16.

Mohd Ahnuar

, Gomez

C.P.

, Yin Yin

, Mohd Nordin

, Yunus

and Abdul Rahman

N.A.

, Development of a progressive green university campus maturity assessment tool and framework for Malaysian universities, MATEC Web of Conferences 266 (2019), 1–6.

17.

Fitri Sari

, Omar

, Abdul Rahman

, Fadhil Md Din

, Mat Taib

, Krishnan

, Safitri Zen

, Hanafi

, Hadiyanto

and Suwartha

, Greening campus experience: moving towards living laboratory action plan, E3 S Web of Conferences 48 (2018), 1–5.

18.

Allen

and Marquart-Pyatt

S.T.

, Workplace energy conservation at Michigan State University, International Journal of Sustainability in Higher Education 19 (2018), 114–129.

19.

Chowdhury

, Hossain

, Longo

and Yaïci

, Optimization of solar energy system for the electric vehicle at University Campus in Dhaka, Bangladesh, Energies 11 (2018), 1–10.

20.

Hoyo-Montaño

, Valencia-Palomo

, Galaz-Bustamante

, García-Barrientos

and Espejel-Blanco

, Environmental impacts of energy saving actions in an academic building, Sustainability 11 (2019), 989–1009.

21.

Jindal

, Kumar

and Rodrigues

J.J.P.C.

, A heuristic-based smart HVAC energy management scheme for university buildings, IEEE Transactions on Industrial Informatics 14 (2018), 5074–5086.

22.

Sheng-Luen

, Yi-Xun

, Ming-Wei

, GAM modeling on power consumption for campus buildings, 2018 37th Chinese Control Conference (CCC) (2018), 8079–8084.

23.

Bourdeau

, Guo

and Nefzaoui

, Buildings energy consumption generation gap: A post-occupancy assessment in a case study of three higher education buildings, Energy and Buildings 159 (2018), 600–611.

24.

Zhang

and Gao

, Application research of campus operation and maintenance management based on BIM, IOP Conference Series: Materials Science and Engineering 439 (2018), 1–5.

25.

Feng

, Zhang

, Chen

, Wang

, Zhang

and Li

, Design of Intelligent bus positioning based on internet of things for smart campus, IEEE Access 6 (2018), 60005–60015.

26.

Matthew R

, Michael G.

, Dimitri N.

, Dynamic implementation of resource management strategies for a resilient advanced life support system, AIAA Scitech 2019 Forum (2019), 1–11.

27.

Yimin

, Zhile

, Yuanjun

and Qingtian

, The college compus energy monitoring platform for artificial intelligence application, 2018 IEEE Innovative Smart Grid Technologies - Asia (ISGT Asia) (2018), 699–704.

28.

Salvia

A.L.

, Reginatto

, Brandli

L.L.

, da Rocha

V.T.

, Daneli

R.C.

and Frandoloso

M.A.L.

, Analysis of energy consumption and efficiency at University of Passo Fundo—Brazil, Towards Green Campus Operations (2018), 519–533.

29.

Ling

, Xiangwei

, Jie

, Weixin

, Yan

and Junyi

, Research and design of energy efficiency monitoring system for resource-saving campus, Advances in Engineering Research 149 (2018), 765–769.

30.

Han

, Lee

, Cho

, Yoon

, Lee

and Rhee

, Improving the energy saving process with high-resolution data: A case study in a university building, Sensors (Basel) 18 (2018), 1606–1621.

31.

Perera

, Zaslavsky

, Christen

and Georgakopoulos

, Context aware computing for the internet of things: A survey, IEEE Communications Surveys & Tutorials 16 (2014), 414–454.

32.

Sezer

O.B.

, Dogdu

and Ozbayoglu

A.M.

, Context-aware computing, learning, and big data in internet of things: A survey, IEEE Internet of Things Journal 5 (2018), 1–27.

33.

Curry

, Derguech

, Hasan

, Kouroupetroglou

and ul Hassan

, A real-time linked dataspace for the internet of things: Enabling “Pay-As-You-Go” data management in smart environments, Future Generation Computer Systems 90 (2019), 405–422.

34.

Kabáč

, Consel

and Volanschi

, Designing parallel data processing for enabling large-scale sensor applications, Personal and Ubiquitous Computing 21 (2017), 457–473.

35.

de Alencar Bezerra

, Jackson dos Santos

, Rogerio Pinheiro

and Rocha Barbosa

, Dynamic evaluation of the energy efficiency of environments in Brazilian University Classrooms Using DEA, Sustainability 9 (2017), 2373–2387.

36.

Shahzad

, Yang

, Ahmad

A.W.

and Lee

, Energy-efficient intelligent street lighting system using traffic-adaptive control, IEEE Sensors Journal 16 (2016), 5397–5405.

37.

Pérez-Chacón

, Luna-Romera

, Troncoso

, Martínez-Álvarez

and Riquelme

, Big data analytics for discovering electricity consumption patterns in smart cities, Energies 11 (2018), 1–19.

38.

Alvarez-Campana

, Lopez

, Vazquez

, Villagra

V.A.

and Berrocal

, Smart CEI Moncloa: An IoT-based platform for people flow and environmental monitoring on a Smart University Campus, Sensors (Basel) 17 (2017), 2856–2880.

39.

Amasyali

and El-Gohary

N.M.

, A review of data-driven building energy consumption prediction studies, Renewable and Sustainable Energy Reviews 81 (2018), 1192–1205.

40.

Leung

and Rosenthal

, Explicating perceived sustainability-related climate: A situational motivator of pro-environmental behavior, Sustainability 11 (2019), 1–17.

41.

Dan

, Lei

and Zhenxia

, What affects individual energy conservation behavior: Personal habits, external conditions or values? An empirical study based on a survey of college students, Energy Policy 128 (2019), 150–161.

42.

Somsuphaprungyos

, Buranaruch

and Boonbrahm

, A framework of decision support system based on integrated data for electricity management in campus, Science and Technology RMUTT Journal 7 (2017), 183–193.

43.

Tamani

, Ahvar

, Santos

, Istasse

, Praca

, Brun

P.-E.

, Ghamri

, Crespi

and Becue

, Rule-based model for smart building supervision and management, 2018 IEEE International Conference on Services Computing (SCC) (2018), 9–16.

44.

https://www.crummy.com/software/BeautifulSoup/.

45.

Geyer

and Singaravel

, Component-based machine learning for performance prediction in building design, Applied Energy 228 (2018), 1439–1453.

46.

Singaravel

, Suykens

and Geyer

, Deep-learning neural-network architectures and methods: Using component-based models in building-design energy prediction, Advanced Engineering Informatics 38 (2018), 81–90.