JP7074709B2 - Power consumption prediction system, method and program - Google Patents

Description

The present invention relates to a system, method and program for predicting power consumption on a predicted date for each predetermined area, and in particular, considers power consumption fluctuation due to a change in the number of people in each area caused by the movement of people between areas. Concerning power consumption prediction systems, methods and programs.

The number of electrical appliances and appliances is steadily increasing, and as a result, overall power consumption is on the rise. Power companies must deal with such situations. On the other hand, the technology for storing the generated power is not cost-effective. Therefore, it is important to generate electricity in a just-in-time manner and balance the amount of electricity generated with the demand for electricity. In addition, utilities need to plan consumption profiles in a short period of time in order to properly manage the distribution network and reduce operating costs.

The ratio of power generation to power demand can be optimized if the power consumption in the near future (for example, one day ahead) can be accurately predicted. Conventionally, Patent Document 1-4 discloses a technique focusing on the power consumption of a device such as a computer or a server, or an endpoint (household or factory) in order to execute a short-term power consumption prediction. Further, Non-Patent Documents 1-3 disclose a technique for determining an endpoint behavior pattern in order to predict power consumption one day in advance. For example, each household will be informed about the normal behavior of the resident (daytime presence on weekdays or weekends, nighttime habits, etc.), and power consumption will be estimated based on this.

Further, Non-Patent Document 4 discloses a method of constructing a prediction model by machine learning including parameters related to various factors related to power consumption and predicting future power consumption using this prediction model.

US Pat. No. 6577962 US Published Patent 20170371308 US Published Patent No. 20140229026 US Published Patent No. 20100179704

J. Campillo, F. Wallin, D. Torstensson, and I. Vassileva. Energy demand model design for forecasting electricity consumption and simulating demand response scenarios in Sweden. International Conference on Applied Energy, Jul 2012. P. Day, M. Fabian, D. Noble, G. Ruwisch, R. Spencer, J. Stevenson and R. Thoppay. Residential Power Load Forecasting. Procedia Computer Science, Volume 28, 2014, Pages 457-464. A. Ozawa and Y. Yoshida. Residential Energy Demand Modeling based on Questionnaire Surveys. Journal of Japan Society of Energy and Resources, Vol.36, No.5, Sep 2015. A. D. Papalexopoulos and T. C. Hesterberg, "A regression-based approach to short-term system load forecasting," in IEEE Transactions on Power Systems, vol. 5, no. 4, pp. 1535-1547, Nov. 1990.

In the conventional method of predicting power consumption based on notification of action schedule and presence information from endpoints related to power consumption of homes and factories, for example, if information related to home schedule is leaked to the outside, There was a problem that the possibility of being exposed to threats such as burglary increased.

In addition, if you try to build a prediction model by machine learning including parameters related to various external factors that affect power consumption, the calculation cost will increase as the parameters increase, and the processing load will increase, making predictions in a short time. There was a problem that it became difficult.

Furthermore, according to the analysis of the inventors, it was confirmed that the movement of people occupies a large weight as an extrinsic factor that affects the fluctuation of power consumption, and even if a prediction model is constructed by machine learning considering various extrinsic factors, humans There was a problem that the power consumption fluctuated greatly and the prediction accuracy decreased in the situation where there was a lot of movement.

An object of the present invention is to solve the above technical problems and to provide a power consumption prediction system, a method and a program which are safe, have high prediction accuracy, and can make predictions in a short time with a low processing load.

In order to achieve the above object, the present invention is characterized in that it has the following configuration in a power consumption prediction system that predicts power consumption for each area.

(1) For history-based prediction means that calculates history-based forecast values by applying the forecast date calendar information to the history-based forecast model constructed by learning the power consumption history associated with the calendar information in each area, and for calendar information. The movement prediction means that predicts the movement of people by applying the calendar information of the prediction date to the movement prediction model constructed by learning the movement history of people between the linked areas, and the power consumption associated with the number of people in each area. A movement-based prediction means that calculates a movement-based prediction value by applying the predicted number of locations after movement to a movement-based prediction model constructed by learning historical information, and the history-based prediction value and the movement-based prediction value. It is equipped with a prediction accuracy improving means for predicting the power consumption on the predicted date based on the prediction difference.

(2) The history-based prediction means calculates the history-based prediction value by applying the calendar information and extrinsic factors of the prediction date to the history-based prediction model constructed by learning the history information of power consumption and extrinsic factors associated with the calendar information. I did it.

(3) As the extrinsic factors, at least one of weather information such as weather, temperature, humidity, precipitation, solar radiation, and wind speed, or event information such as fireworks display and sports is used.

(4) The history-based prediction means calculates the history-based prediction value in the first cycle, and the movement prediction means predicts the movement of a person in the second cycle longer than the first cycle.

(5) The prediction date is divided into multiple time zones, and the movement prediction means predicts the movement of people in each time zone.

(6) The movement-based prediction means constructed a movement-based prediction model for each time zone, and applied the movement of the person predicted by the movement prediction means for each time zone to the corresponding movement-based prediction model.

(7) As a means for improving prediction accuracy, the prediction difference between the time zone average of the history-based predicted value obtained for each time zone and the movement-based predicted value predicted for each time zone is applied to each history-based predicted value for each time zone. I tried to do it.

According to the present invention, the following effects are achieved.

(1) Correct the history-based predicted value that predicted the power consumption of the predicted date based on the power consumption history, based on the difference from the movement-based predicted value that predicted the power consumption of the predicted day based on the movement history of people. By predicting the power consumption on the predicted date, it is possible to prevent the prediction accuracy from deteriorating due to the movement of people, and the power consumption is safe, has high prediction accuracy, and can be predicted in a short time with a low processing load. Be able to provide prediction systems, methods and programs.

(2) The history-based prediction means learns the history information of the power consumption and extrinsic factors associated with the calendar information to build a history-based prediction model, so that the extrinsic factors can be reflected in the power consumption prediction and the consumption is highly accurate. Power prediction becomes possible.

(3) By considering external factors such as weather information and event information, which are major factors of power consumption fluctuations, it becomes possible to mitigate the influence of these external factors on the power consumption forecast.

(4) By calculating the history-based prediction value in the first cycle and predicting the movement of a person in the second cycle, which is longer than the first cycle, the burden of prediction calculation can be reduced. ..

(5) By dividing the predicted date into multiple time zones and predicting the movement of people in each time zone, even if the movement tendency of people is unique according to the time zone, the movement of people is predicted with high accuracy. become able to.

(6) The movement-based prediction means builds a movement-based prediction model for each time zone, and applies the movement of the person predicted by the movement prediction means for each time zone to the corresponding movement-based prediction model. Even if the trend is unique depending on the time zone, the movement-based prediction value can be predicted with high accuracy.

(7) By applying the prediction difference between the time zone average of the history-based predicted value and the movement-based predicted value predicted for each time zone to each history-based predicted value for each time zone, the power consumption tendency depends on the time zone. Even in this case, highly accurate power consumption prediction becomes possible.

It is a block diagram which showed the structure of the main part of the power consumption prediction system which concerns on one Embodiment of this invention. It is a figure which showed the prediction method of the power consumption by the history-based prediction model. It is a figure which showed the method of predicting the flow of people between areas. It is a figure which showed the example of the time zone set in the predicted day. It is a figure which showed the method of the movement prediction by the movement-based prediction model. It is a figure which showed the calculation method of the prediction difference. It is a figure which showed the method of improving the accuracy of the history-based predicted value based on the prediction difference. It is a functional block diagram of one Embodiment of this invention. It is a flowchart which showed the operation of the functional block diagram. It is a figure which showed the example which improved the prediction accuracy based on the prediction difference.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the main part of the power consumption prediction system according to the embodiment of the present invention, based on the history information of power consumption for a relatively large area such as a municipality or a prefecture. The power prediction unit 1 that predicts future power consumption (history-based predicted value) and the power consumption (movement-based predicted value) are predicted based on the number of people in each area, and the history-based according to the prediction difference of each predicted value. It is composed of a prediction accuracy improving unit 2 that adjusts a predicted value and improves the prediction accuracy.

In the power prediction unit 1, the past power consumption associated with the calendar information for each area is registered as history information ci-1 in the power consumption history database iDB-h1. Such power consumption history ci-1 can be collected using the Advanced Metering Infrastructure (AMI).

External factors databases eDB-i ₁ to eDB-i _N include past weather information such as weather, temperature, humidity, precipitation, solar radiation, and wind speed, as well as future weather forecasts, as external factors that affect power consumption in each area. Alternatively, event information such as fireworks festivals and sports is registered as history information ei-i ₁ to ei-i _N linked to calendar information. Such extrinsic history ei-i ₁ to ei-i _N can be collected via API (Application Programming Interface).

The first ML algorithm a1 is registered in the power consumption estimation server SV1. The first ML algorithm a1 has a power consumption history ci-1 associated with the calendar information in the power consumption history database iDB-h1 and an extrinsic factor history associated with the calendar information in the extrinsic factors databases eDB-i ₁ to eDB-i _N. Multiple regression analysis is performed by machine learning (ML) on the relationship between ei-i ₁ and ei-i _N , and the power consumption of the forecast area on the forecast date is based on external factors such as weather forecasts and event information on the forecast date. Build a history-based forecasting model M ₁ for forecasting.

As shown in FIG. 2, the history-based prediction model M ₁ sets the power consumption Yh of the prediction area in the prediction area based on the following equation (1) based on external factors such as calendar information of the prediction date and weather forecast. Predict every 30 minutes. In this embodiment, the power consumption of the next day is predicted at 10 am every day, and 48 history-based predicted values Yhi (i ∈ {0, 1…, 47}) are accumulated in the history-based predicted value database iDB-f1. To.

Where Yhi is the history-based prediction value in the i-th cycle, β _{0, i} is the section of the regression line applied to the i-th cycle, N is the number of machine learning parameters, and β _{k, i} is the i-th cycle. The regression coefficients of the k-th machine learning parameter, p _{k, i} , are the values of the k-th machine learning parameter applied to the i-th period.

In this way, in this embodiment, the power consumption on the predicted day is predicted every 30 minutes, and the history-based predicted value Yhi of 48 pieces / day can be obtained. Can be predicted.

In the prediction accuracy improvement unit 2, the movement history database iDB-h2 has the history information mi-1 in which the flow of people from other areas and the flow of people to other areas are linked to the calendar information for each area. It is registered. Such movement history mi-1 can be collected via the location identification function of the smartphone of the user who has agreed to share his / her location.

The third ML algorithm a3 is registered in the movement prediction server SV2. The third ML algorithm a3 uses the movement history mi-1 of people between each area registered in the movement history database iDB-h2 as an element for the flow of people between areas on the predicted date, as shown in FIG. We construct a movement prediction model M ₃ that makes predictions based on the transition matrix to be performed and then calculates the number of people in each area after movement.

FIG. 3 shows the flow of people moving between the four areas A, B, C, and D. In the transition matrix, the element m _AD represents the flow of people from area A to area D, and the elements. m _DA represents the flow of people from Area D to Area A, and element m _A A represents the flow of people from Area A to Area A. The prediction result mf of the flow of people is registered in the movement prediction database iDB-f2.

Here, for example, considering the general movement pattern of people, after moving to work by commuting in the morning, working at work until evening, and moving to home at night after work, for a certain period of time between movements. A movement pattern that stays in the same place is assumed. Further, if the prediction cycle of the flow of people is shortened, a slight flow of people near the boundary of each area also affects the prediction, so that the load of the prediction calculation increases.

Therefore, in the present embodiment, the movement prediction based on the movement history mi-1 is made independent without being incorporated as a parameter of the multiple regression analysis used for the history-based prediction, and the prediction cycle is set to the history-based prediction cycle (30 minutes). The processing load is reduced by making it longer than.

In this embodiment, as will be described in detail later, the predicted date is divided into _four time zones T ₁ , T ₂ , T ₃ , and T ₄ , and consumed in each time zone T ₁ , T ₂ , T ₃ , and T 4. Where the power prediction Yh is calculated, the movement prediction model M ₃ also uses one prediction model to determine the movement of people between areas and the number of people in each area during the time zones T ₁ , T ₂ , T ₃ , and T ₄ . Predict each. The time zones T ₁ , T ₂ , T ₃ , and T ₄ do not have to be evenly spaced, and may be divided at time intervals that have similar movement patterns of people.

The second ML algorithm a2 is further registered in the power consumption estimation server SV1. The second ML algorithm a2 is registered in the movement history database iDB-h2, which is the number of people in the area estimated based on the predicted value mf of the flow of people in each area registered in the movement prediction database iDB-f2. By machine learning (ML) the relationship between the movement history mi-1 and the power consumption history ci-1 registered in the power consumption history database iDB-h1, each time zone T ₁ , T ₂ , T ₃ , T ₄ We construct four movement-based prediction models M _2j (M ₂₁ , M ₂₂ , M ₂₃ , M ₂₄ ) corresponding to.

As shown in Figures 4 and 5, each movement-based prediction model M ₂₁ , M ₂₂ , M ₂₃ , and M ₂₄ predicts the prediction dates for each time zone T ₁ , T ₂ , T ₃ , and T ₄ for each area. Calculate the mobile-based predicted value Ymj (j ∈ {1, 2, 3, 4}) of power consumption for each time zone T ₁ , T ₂ , T ₃ , and T ₄ based on the number of people in each location.

As shown in FIG. 6, the power consumption estimation server SV1 further has a history-based prediction value _Yhi predicted by the history _- based prediction model M ₁ in a _{30 -} minute cycle. Average value for each (time zone average) Yh_aj (Yh_a1, Yh2_a2, Yh3_a3, Yh4_a4) is calculated, and each time zone average Yh_aj and each movement-based prediction model M _2j are time zones T ₁ , T ₂ , T ₃ , Applying the movement-based predicted values Ymj (Ym1, Ym2, Ym3, Ym4) predicted for each T ₄ to the following equation (2), the prediction difference γj for each time zone T ₁ , T ₂ , T ₃ , T ₄ Calculate (γ1, γ2, γ3, γ4).

γj = Ymj --Yh_aj… (2)

As shown in FIG. 7, the power consumption estimation server SV1 further records the predicted differences γ1, γ2, γ3, and _γ4 calculated for _each time zone T1, T2 _, T3, and T4 in _each time zone. From the base predicted value Yhi, the power consumption predicted value Yi with improved prediction accuracy is calculated by adjusting based on the following equation (3). The predicted power consumption value Yi with improved accuracy is stored in the database iDB-ef.

Yi = Yhi + γj… (3)

In FIG. 8, the power consumption estimation server SV1 and the mobile estimation server SV2 cooperate to predict the power consumption on the predicted date while constructing each prediction model M using the history information stored in each database DB, and further predict the power consumption. It is a functional block diagram which showed the structure which improves the prediction accuracy, and FIG. 9 is a flowchart which showed the procedure.

The history-based prediction unit 101 uses the history-based prediction model M ₁ constructed by machine learning the power consumption history ci-1 and extrinsic history ei-i ₁ to ei-i _N linked to the calendar information of each area, and the prediction date. Calculate the history-based predicted value Yhi of power consumption by applying the calendar information and extrinsic factors of. As the extrinsic factor, meteorological information such as weather, temperature, humidity, precipitation, solar radiation, and wind speed, or event information such as fireworks display and sports can be used. The history-based prediction unit 101 can apply the calendar information and extrinsic factors of the prediction date to the history-based prediction model M ₁ and calculate the history-based prediction value Yhi in the first cycle (for example, 30 minutes).

The movement prediction unit 102 applies the calendar information of the prediction date to the movement prediction model M ₃ constructed by learning the movement history mi-1 of the person associated with the calendar information between areas, and the flow of people between each area. Predict mf. The movement prediction model M ₃ can predict the movement mf of a person in a second cycle longer than the first cycle.

The movement-based prediction unit 103 divides the prediction date into a plurality of time zones (in this embodiment, four time zones T ₁ , T ₂ , T ₃ , and T ₄ ), and power consumption associated with the number of people in each area. The history ci-1 and the extrinsic factors ei-i ₁ to ei-i _N are learned, and the movement-based prediction models M ₂₁ , M ₂₂ , M ₂₃ , and M ₂₄ are constructed for each time zone. Then, the movement-based prediction value Ymj of power consumption is calculated by applying the predicted number of people mf in each area after the flow of people to each movement-based prediction model M ₂₁ , M ₂₂ , M ₂₃ , M ₂₄ . ..

The prediction difference calculation unit 104 calculates the prediction difference γj between the time zone average Yhi_aj of the history-based prediction value Yhi and the movement-based prediction value Ymj. The prediction difference calculation unit 104 has a time zone average Yh_aj (Yh_a1, Yh_a2, Yh_a3, Yh_a4) of the history-based predicted value Yhi obtained for each time zone T ₁ , T ₂ , T ₃ , and T ₄ , and a time zone T ₁ , The prediction difference γj (γ1, γ2, γ3, γ4) from the movement-based predicted value Ymj (Ym1, Ym2, Ym3, Ym4) calculated for each of T ₂ , T ₃ , and T ₄ can be calculated.

The prediction accuracy improvement unit 105 corrects the history-based prediction value Yhi with the prediction difference γmj, and outputs the power consumption prediction value Yi with improved prediction accuracy. The prediction accuracy improvement unit 105 applies the prediction difference γ1 to the history _- based prediction value Yhi in the time zone T1. Similarly, the prediction difference γ2 is applied to the history-based predicted value Yhi of the time zone T ₂ , the prediction difference γ3 is applied to the history-based predicted value Yhi of the time zone T ₃ , and the history-based predicted value of the time zone T ₄ is applied. The predicted difference γ4 can be applied to Yhi. As a result, it is possible to output the power consumption predicted value Yi with improved prediction accuracy.

Next, it will be described with reference to the flowchart of FIG. In step S1, the power consumption prediction area and the prediction date are designated. In step S2, extrinsic factors such as weather forecasts and event information related to the forecast date of the forecast area are acquired. In step S3, it is determined whether or not the prediction model for predicting the power consumption on the prediction date of the prediction area has been registered, and if not registered, the process proceeds to step S4.

In step S4, the history-based prediction model M ₁ is constructed in the history-based prediction unit 101. The history-based prediction unit 101 further applies external factors such as the forecast area, the calendar information of the forecast date, and the weather forecast and the event information regarding the forecast date of the forecast area to the history-based forecast model M ₁ , so that the history-based forecast value is obtained. Yhi is calculated every 30 minutes.

In parallel with this, in step S5, the movement prediction unit 102 constructs a movement prediction model M ₃ that predicts the number of inflows and the number of outflows on the prediction date of the prediction area. The movement prediction unit 102 further applies the calendar information of the prediction area and the prediction date to the movement prediction model M ₃ to calculate the movement prediction value mf of the person on the prediction date of the prediction area.

In step S6, the movement-based prediction unit 103 predicts the power consumption of the predicted area on the predicted day based on the number of people in each area estimated based on the number of inflows / outflows on the predicted day. Models M ₂₁ , M ₂₂ , M ₂₃ , and M ₂₄ are constructed for each of the time zones T ₁ , T ₂ , T ₃ , and T ₄ . The history-based prediction unit 103 further applies the movement prediction predicted for each time zone T ₁ , T ₂ , T ₃ , and T ₄ to the corresponding movement-based prediction models M ₂₁ , M ₂₂ , M ₂₃ , and M ₂₄ . Therefore, the movement-based predicted value Ymj on the predicted date of the predicted area is calculated for each time zone (Ym1, Ym2, Ym3, Ym4).

As described above, in the present embodiment, the construction and prediction of the history-based prediction model M ₁ (step S4), the construction and prediction of the movement prediction model M ₃ (step S5), and the movement-based prediction models M ₂₁ , M ₂₂ , and M ₂₃ . , M ₂₄ is constructed and predicted (step S6) in parallel, so that the overall processing cost can be reduced.

In step S7, the time zone average Yh_aj (Yh_a1, Yh_a2, Yh_a3, Yh_a4) of the history-based predicted value Yhi is calculated. In step S8, as shown in FIG. 7, for each time zone T ₁ , T ₂ , T ₃ , and T ₄ , the time zone average Yh_aj (Yh_a1, Yh_a2, Yh_a3, Yh_a4) of the history-based predicted value and the time zone are described. The prediction difference γ from the movement-based prediction value Ymj (Ym1, Ym2, Ym3, Ym4) predicted for each of T ₁ , T ₂ , T ₃ , and T ₄ is calculated by the prediction difference calculation unit 104.

In step S9, the prediction accuracy improvement unit 105 reflects the prediction differences γ1, γ2, γ3, and _γ4 according to the _time zones T1, T2 _, T3, and T4 in _each history-based prediction value Yhi for prediction. Improve accuracy. If it is determined in step S4 that the prediction model has been registered, the process proceeds to step S10, and each registered prediction model is updated based on the history information acquired thereafter.

FIG. 10 is a diagram showing an example of comparing the relationship between the predicted value of power consumption and the measured value for each time zone T ₁ , T ₂ , T ₃ , T ₄ (column V), and is a diagram showing a history-based predicted value (history-based predicted value). The predicted values (Y column) after correcting the X column) based on the predicted difference (γ1, γ2, γ3, γ4) from the movement-based predicted values are all close to the measured values (Z column) and have a history. It can be seen that the accuracy of the base prediction value Yhi is improved by the movement base prediction value Ymj.

1 ... Power prediction unit, 2 ... Accuracy improvement unit, 101 ... History-based prediction unit, 102 ... Movement prediction unit, 103 ... Movement-based prediction unit, 104 ... Prediction difference calculation unit, 105 ... Prediction accuracy improvement unit, SV1 ... Consumption Power estimation server, SV2 ... Mobile estimation server

Claims (12)

In a power consumption prediction system that predicts power consumption for each area
A history-based prediction means that calculates the history-based prediction value by applying the calendar information of the prediction date to the history-based prediction model constructed by learning the power consumption history associated with the calendar information of each area.
A movement prediction means that predicts the movement of people by applying the calendar information of the prediction date to the movement prediction model constructed by learning the movement history of people between areas linked to the calendar information.
A movement-based prediction means that calculates a movement-based prediction value by applying the predicted number of locations after movement to a movement-based prediction model constructed by learning historical information on power consumption associated with the number of people in each area.
The power consumption prediction system is provided with a prediction accuracy improving means for predicting the power consumption on the prediction date based on the prediction difference between the history-based prediction value and the movement-based prediction value. The history-based prediction means applies the calendar information of the prediction date and the extrinsic factors to the history-based prediction model constructed by learning the history information of the power consumption and the extrinsic factors associated with the calendar information to calculate the history-based prediction value. The power consumption prediction system according to claim 1, which is characterized. The power consumption prediction system according to claim 2, wherein the extrinsic factor is meteorological information including at least one of weather, temperature, humidity, precipitation, solar radiation, and wind speed. The power consumption prediction system according to claim 2 or 3, wherein the extrinsic factor is information on an event that affects the power consumption prediction. The history-based prediction means calculates the history-based prediction value in the first cycle, and the movement prediction means predicts the movement of a person in a second cycle longer than the first cycle. The power consumption prediction system according to any one of claims 1 to 4. The predicted date is divided into multiple time zones,
The power consumption prediction system according to any one of claims 1 to 5, wherein the movement prediction means predicts the movement of a person for each time zone. The movement-based prediction means builds a movement-based prediction model for each time zone.
The power consumption prediction system according to claim 6, wherein the movement predicting means applies the movement of a person predicted for each time zone to a corresponding movement-based prediction model. The prediction accuracy improving means applies the prediction difference between the time zone average of the history-based predicted values obtained for each time zone and the movement-based predicted value predicted for each time zone to each history-based predicted value for each time zone. The power consumption prediction system according to claim 7. In the power consumption prediction method in which the computer predicts the power consumption for each area
The procedure to calculate the history-based prediction value by applying the calendar information of the prediction date to the history-based prediction model constructed by learning the power consumption history associated with the calendar information of each area.
The procedure for predicting the movement of people by applying the calendar information of the prediction date to the movement prediction model constructed by learning the movement history of people between areas linked to the calendar information.
A procedure for calculating a movement-based predicted value by applying the predicted number of locations after movement to a movement-based prediction model constructed by learning historical information on power consumption associated with the number of people in each area.
A power consumption prediction method comprising a procedure for predicting power consumption on a prediction date based on a prediction difference between a history-based prediction value and a movement-based prediction value. In the procedure for calculating the history-based predicted value, the history-based predicted value is calculated by applying the calendar information and the extrinsic factor of the predicted date to the prediction model constructed by learning the history information of the power consumption and the extrinsic factor associated with the calendar information. The power consumption prediction method according to claim 9, wherein the power consumption is predicted. In the power consumption prediction program that the power consumption predicts for each area
The procedure to calculate the history-based forecast value by applying the calendar information of the forecast date to the forecast model constructed by learning the power consumption history associated with the calendar information of each area.
The procedure for predicting the movement of people by applying the calendar information of the prediction date to the prediction model constructed by learning the movement history of people between areas linked to the calendar information.
A procedure for calculating a movement-based predicted value by applying the predicted number of locations after movement to a prediction model constructed by learning historical information on power consumption associated with the number of people in each area.
A power consumption prediction program characterized by causing a computer to execute a procedure for predicting power consumption on a predicted date based on a prediction difference between a history-based predicted value and a movement-based predicted value. In the procedure for calculating the history-based predicted value, the history-based predicted value is calculated by applying the calendar information and the extrinsic factor of the predicted date to the prediction model constructed by learning the history information of the power consumption and the extrinsic factor associated with the calendar information. The power consumption prediction program according to claim 11, wherein the power consumption is predicted.

Images