Keywords

1 Introduction

A cyber-physical system (CPS) analyzes and uses data obtained from the real world. The use of various data and functions allows improvements to the performance of the social system and efficient operation management. In the CPS, a huge amount of real world data is collected and analyzed to construct a virtual world on a computer. Here, it is important to blend information of the physical world with information of the cyber world. The CPS finds the optimal solution by simulating the future world in the constructed virtual world and feeds it back to society. This makes it possible to efficiently solve various urban problems, such as those pertaining to traffic and energy.

CPSs are expected to become a common part of social infrastructure in the energy sector [1]. Because of recent developments in this sector, there is a need to provide a stable energy supply to realize a low-carbon society. A smart grid is a power grid with a communication function and control function; e.g., the grid may use smart meters. The purpose of the smart grid is to optimize the supply-demand balance within a small community. In Japan, a large amount of renewable energy is scheduled for installation by 2020 [2]. Against such a background, analyses and optimization of the energy supply and demand balance are required. Because the supply side has a limited ability to control the power demand, attempts to optimize the energy problem have been made on the demand side. For instance, energy management systems optimize a wide variety of energy sources to carry out local production for the local consumption of energy [3].

2 Related Work

In the conventional energy management system, power peaks are cut and shifted mainly using solar power generation and batteries. A reduction in the emission of carbon dioxide can be expected from the optimum operation of batteries according to the prediction technologies of the power demand and photovoltaic power generation using weather information. However, the effect of the reduction of carbon dioxide largely depends on the photovoltaic power generation and battery capacity. It is a problem that the costs of solar panels and battery installation are high. Additionally, it is a problem that power consumption is strongly affected by the behavior of people even if the generation of photovoltaic power is accurately forecast.

Meanwhile, a demand response [4, 5] can be used to manage energy by controlling the electric power demand itself. This method saves electricity by changing the electricity price dynamically. Energy is conserved by setting the electric unit price high when the electric power demand is large. Although a large power saving can be expected depending on the setting of the unit price of electricity, it is not an ideal method because there is a possibility that the living comfort of the user will be sacrificed. Additionally, at institutions where there are many people who are not conscious of electricity charges, such as universities and complex commercial facilities, the energy saving will be small.

The present research therefore attempts to cut and shift the power peak of an entire target area by controlling the congestion of each building without a compulsion to demonstrate a strong energy management effect even for universities and complex commercial facilities. Targets such as universities and compound commercial facilities comprise multiple facilities. Some areas of these targets are sometimes locally crowded and uncomfortable. Meanwhile, there are wide spaces containing only a small number of users, where energy is used inefficiently. We attempt to change crowd behavior to solve these problems. A conventional study [6] showed that the power consumption of each facility depends on the congestion of the facility. It is assumed here that it is possible to control the congestion of each facility any time by presenting information that prompts behavioral change via social networking services or e-mail. In fact, we control the congestion of each facility to level the electric power use. This makes it possible to reduce the emission of carbon dioxide. It is thus conceivable that we can manage energy without the user being consciously aware of the power savings.

3 Proposed Method

The purpose of our research is to realize energy management aimed at reducing the emission of carbon dioxide by changing people’s behavior in small and medium-sized electricity communities. We assume that small and medium-sized electricity communities comprise facilities with multiple power demand characteristics. We attempt to cut and shift the power peak of the entire target area by controlling the crowd congestion of each building and thus demonstrate a strong energy management effect even for small and medium-sized electricity communities. Under the above assumption, the power demand is leveled in the following procedure.

  1. 1.

    Crowd congestion in each building is changed by the presentation of information prompting a behavioral change. Some areas of targets are sometimes locally crowded and uncomfortable. Meanwhile, there are wide spaces containing only a small number of users, where energy is used inefficiently. We attempt to change crowd behavior to solve these problems. It is assumed here that it is possible to control the congestion of each facility any time by presenting information that prompts a behavior change via social networking services or e-mail. Here, even if congestion changes, it is important that the total number of people does not change.

  2. 2.

    The power consumption of each building changes with the crowd congestion. We here assume that the power demand will change with the change in crowd congestion. A conventional study [6] showed that the power consumption of each facility depends on the congestion of the facility. We control the crowd congestion of each facility to change the power consumption of each building by using this knowledge.

  3. 3.

    The total power demand is leveled by aggregating the power consumption of multiple buildings. The goal of the system is to level the total power demand in the community. The changed power consumption of multiple buildings is aggregated. As a result, it is conceivable that we can manage energy without the user being consciously aware of power savings.

Functions necessary for a CPS to level the power demand as mentioned above are listed as follows.

  • Acquisition of heterogeneous data and construction of a database

  • Construction of a power demand prediction model for each facility

  • Creation of presentation information that promotes a behavioral change and realizes the optimum crowd congestion for each facility for energy management

The following sections present the details of each process.

3.1 Acquisition of Heterogeneous Data and Construction of the Database

Various sensors were installed on the campus of a university to acquire heterogeneous data. In the construction of the database, each datum was synchronized with respect to time. In this database, data were normalized and missing data filled in. The experimental environment selected for this study encompasses multiple buildings on the Ito Campus of Kyushu University in Fukuoka, Japan. Students are free to access and stay in these buildings during the day.

Acquisition of Power Consumption Data for Each Facility. There are six targets (buildings A–F) in the central zone of the university campus, as shown in Fig. 1. Power consumption within each building is measured hourly.

Fig. 1.
figure 1

Target buildings A–F at Ito Campus, Kyushu University. Blue circles indicate the positions of 14 P-sen units. (Color figure online)

Full size image

Acquisition of Meteorological Data. Each sensor is installed on the roof of a building on the campus and measures temperature, humidity, solar irradiance, wind speed, and wind direction every minute. In addition, the Japan Meteorological Agency publishes various local data, such as temperature, humidity, and solar irradiance data, every hour [7], and we can thus alternatively use data recorded at the position closest to the university. This approach has data redundancy.

Acquisition of Crowd Congestion. Crowd flow data can be measured using a pole-type small sensor node (P-sen) installed at several locations on the campus. Figure 2 shows the appearance of a P-sen, which has a network camera, wireless LAN access point, and range finder. Data for analyzing human behavior can be redundantly acquired using multiple sensors. In this study, crowd flow was measured using only the range finder. This measurement method is not accurate when there is occlusion in the crowd. However, this approach preserves privacy and can measure congestion under various illumination conditions. Figure 1 shows the positions of the 14 P-sens (Nos. 1–14). It is possible to measure the positions of moving objects in the area in front of each P-sen at about 10 Hz. The 14 P-sen units cover the entire area over which people can move in the zone.

Fig. 2.
figure 2

P-sen unit for crowd flow measurement. In this study, only the range finder was used to measure crowd flow.

Full size image

Crowd flow data are generated from the acquired moving object information using a Kalman filter [8]. Furthermore, the number of unique users observed per minute is calculated from the crowd flow data for each P-sen. In this study, the mean number of unique users per hour is defined as the degree of congestion for each zone. It is possible to estimate the degree of congestion robustly taking this approach, even if there is some short-term data loss.

3.2 Construction of a Power Demand Prediction Model for Each Facility

It is necessary to construct a model to estimate the power consumption when the crowd congestion changes. Because it is known that crowd congestion affects the power consumption of buildings [6], we build a power demand prediction model using this knowledge. We use a vector autoregression (VAR) model as a power demand prediction model. We here explain VAR, which is a technique that can be used to forecast power demand from heterogeneous data. VAR is represented by the model

$$\begin{aligned} \mathbf{Y}_\mathbf{t}=\sum _{\mathbf{m}={\mathbf {1}}}^\mathbf{{M}} \mathbf {A_mY_{t-m}}+\mathbf {\epsilon _t}, \end{aligned}$$
(1)

where the inputs are n kinds of time series data, \(\mathbf {Y_t}=(\mathbf{{y}}_{\mathbf {1},\mathbf{{t}}},\mathbf{{y}}_{\mathbf {2},\mathbf{{t}}},\cdots ,\mathbf{{y}}_\mathbf{{n,t}})\) with M lags. We use Akaike’s information criterion (AIC) [9] to determine parameter M, calculated as

$$\begin{aligned} AIC=-2\ln L + 2k, \end{aligned}$$
(2)

where L is the maximum likelihood and k is the number of free parameters. The AIC is used to evaluate the goodness of the statistical model maintaining a balance between the fitting of the data and the complexity of the model. The number of lags M is determined so as to minimize the AIC. The constructed VAR model is used in forecasting by inputting n kinds of time series data, \(\mathbf {Y_t}=(\mathbf{{y}}_{\mathbf {1},\mathbf{{t}}},\mathbf{{y}}_{\mathbf {2},\mathbf{{t}}},\cdots ,\mathbf{{y}}_\mathbf{{n,t}})\) with M lags. It is possible to predict the power demand for each building when changing the crowd congestion using these models.

3.3 Creation of Presentation Information that Promotes a Behavioral Change and Realizes Optimum Crowd Congestion for Each Facility in Energy Management

In the case of providing information that promotes behavioral changes, it is ideal to provide the most effective information for leveling energy at the most effective time according to the congestion situation of the facility. However, it is difficult to judge whether the timing and content of the information to be provided are appropriate for leveling energy by simply observing crowd congestion after the presentation of information. For example, even if congestion is temporarily eliminated immediately after information is provided, if the energy is not leveled as a result, the information provided is not appropriate. Conversely, even if congestion increases temporarily after information is provided, if the energy is leveled as a result, the information to be provided is appropriate. In the case of solving the congestion problem of commercial facilities, if the total number of users decreases as a result of congestion mitigation, the presented information is inappropriate on the facility side. In this way, it cannot be decided whether the provided information is appropriate unless the result of how congestion changes in the real world. We therefore determine the content and timing of information to be provided in a reinforcement learning framework for this problem.

Reinforcement learning is a method of learning appropriate behavior in unknown circumstances. The learner in reinforcement learning is called an agent and learns appropriate behavior rules through interaction with the environment. The agent observes the state of the environment and executes an action according to that state. Selecting an action for the state of the environment is called a policy. The action affects the state of the environment, and the environment rewards the agent as a form of behavior evaluation. The objective of reinforcement learning is to seek the optimal policy that finally obtains maximum rewards through the repetition of this process. Reinforcement learning has been particularly successful for game tasks [10, 11]. However, in the learning process of reinforcement learning, because agents perform actions through trial and error according to environmental conditions and rewards, they take wrong actions in the early stage of learning in many cases. Although this is not a problem in game tasks, in the case of learning by interaction with the real world as in this research, taking a wrong action may cause confusion in the real world. To solve this problem, we construct a model that is close to the real-world environment as a preliminary stage of the experiment in the real world, and conduct a simulation using the model.

4 Experiments

This paper focuses on the construction of a power demand prediction model for each facility. In this experiment conducted to validate the proposed method, power demand models are constructed by analyzing the acquired heterogeneous data for Ito Campus, Kyushu University. Data were acquired hourly over a period of 6 months from September 8, 2015 to March 8, 2016. Appropriate locations and periods were selected for the experiment so that there was no long-term data loss that would affect the accuracy of data interpolation. The experimental targets were buildings A, B, and C shown in Fig. 1. Building A has lecture rooms, many of which are used for daytime lectures. Building B is not used for lectures but is made available to students at all times. Building C contains university administration offices and rooms used for lectures. The mean and standard deviation of the power consumption for each building are given in Table 1. The mean power consumption of building A is larger than the mean consumptions of buildings B and C. The standard deviation of the power consumption increases from building C to building B to building A.

Table 1. Mean and standard deviation of the power consumption of each building.
Full size table

To analyze heterogeneous data, congestion information from the P-sens closest to the entrances of the target buildings was used. Specifically, as illustrated in Fig. 3, P-sen 4 was used to forecast the power demand of building A, P-sen 1 was used for building B, and P-sen 6 was used for building C under the assumption that the power demand is affected by the congestion information obtained from these P-sens. It was found in preliminary experiments that this assumption provides good results. Table 2 gives the mean and standard deviation of congestion data acquired by the P-sens. The mean degree of congestion is highest for P-sen 1, the standard deviation of the degree of congestion is highest for P-sen 4, and both the mean and standard deviation are lowest for P-sen 6. We used temperature, humidity, and solar irradiance data for Fukuoka, made available by the Japan Meteorological Agency.

Fig. 3.
figure 3

Congestion information used for constructing power demand models.

Full size image
Table 2. Means and standard deviations of the congestion data acquired by the P-sens.
Full size table

In the analysis, we verified whether a statistically suitable model can be constructed by combining the congestion and meteorological data. We forecast the power demand for each building in a one-sample future. In the analysis, the degree of congestion, temperature, humidity, and solar irradiance were selected to construct the statistical model. The results of the power demand forecasts of buildings A, B, and C are respectively given in Tables 3, 4, and 5. The results obtained without using the congestion data are also shown for comparison.

Table 3. Results of the power demand forecast for building A. The mean error is smallest when temperature and congestion information are used. There is a significant difference with respect to all other results.
Full size table
Table 4. Results of the power demand forecast for building B. The mean error is smallest when temperature and congestion information are used. There is a significant difference with respect to all other results.
Full size table
Table 5. Results of the power demand forecast for building C. The mean error is smallest when temperature and congestion information are used. There is a significant difference with respect to all other results.
Full size table

It was found that for all buildings, the mean difference between the estimation results and actual measurement values was smallest when temperature and congestion data were used. It is important to note here that the parameters constituting the model vary greatly for each building.

5 Discussion

The experimental results showed that the power consumption characteristics differ for each building. In particular, the effect of the change in crowd congestion varies from building to building. This means that there is a possibility that the total power consumption will change with a change in crowd congestion. It is thought that this relationship can be used for energy management by changing crowd congestion appropriately. We will determine the content and timing of information to be provided in a reinforcement learning framework.

6 Conclusion

We proposed a method of realizing energy management aimed at reducing the emission of carbon dioxide by changing people’s behavior in small and medium-sized electricity communities. We constructed a model with which to estimate the power consumption of each building when the crowd congestion changes. In experiments, we found that the effect of the change in crowd congestion varies from building to building. Future works are to estimate the optimum crowd congestion for energy management and to create presentation information that promotes behavioral changes that realize the estimated optimum crowd congestion.