Demand prediction device, demand prediction method, and program

Abstract

To provide a device, a method and a program for predicting a demand amount of water with a predefined level of accuracy or better even in the case where a demand amount increases or decreases due to a sudden factor and where at least a portion of a demand result value shows missing or abnormality.SOLUTION: A demand prediction device is provided with a demand amount prediction unit for predicting a demand amount of water of a prediction object period by using one prediction model selected from among a plurality of prediction models according to a situation. The prediction models include a first prediction model for accepting, as input, result demand amount data of a demand amount of water of a past fixed period, weather data showing weather of past and future fixed periods and calendar day data showing at least a day of the week, and outputting prediction demand amount data showing a prediction value of a prediction amount of water of a prediction object period, a second prediction model for accepting, as input, the result demand amount data without accepting the weather data and the calendar day data, and outputting the prediction demand amount data, and a third prediction model for accepting, as input, the weather data and the calendar day data without accepting the result demand amount data and outputting the prediction demand amount data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] An embodiment of the present invention relates to a demand forecasting device, a demand forecasting method, and a program. [Background technology]

[0002] At water treatment facilities such as water purification plants, water demand forecasts are periodically made to plan water operations over a certain period. For example, accurately forecasting water demand for a given day is important for efficiently operating waterworks facilities that supply purified water and for supplying the right amount of purified water to consumers, without excess or shortage.

[0003] The amount of purified water consumed by consumers is thought to be affected by the weather and day of the week on that day, and a method has been proposed for building a demand forecasting model that inputs information on weather and day of the week along with actual water demand. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2001-216287 [Patent Document 2] Japanese Patent Application Publication No. 2020-201609 Summary of the Invention [Problem to be solved by the invention]

[0005] For example, if water demand increases or decreases due to unexpected factors such as a water pipe leak or a temporary additional supply to another water distribution area, the accuracy of water demand forecasts (demand forecasts) may decrease. Furthermore, if the error in the demand forecast increases and the forecasts are repeatedly revised, this may hinder the efficient operation of waterworks facilities and the stable supply of purified water.

[0006] Furthermore, if at least some of the actual demand values used for the demand forecast are missing or abnormal due to a sensor abnormality or the like, the accuracy of the demand forecast may be significantly reduced.

[0007] The problem that the invention aims to solve is to provide a demand forecasting device, a demand forecasting method, and a program that can predict water demand with a certain level of accuracy or higher, even if demand increases or decreases due to unexpected factors, or if at least some of the actual demand values are missing or abnormal. [Means for solving the problem]

[0008] The demand forecasting device of the embodiment includes a demand forecasting unit that forecasts the demand for water for a forecast period using one forecasting model selected from a plurality of forecasting models depending on the situation, and the plurality of forecasting models include a first forecasting model that inputs actual demand data indicating actual values of water demand for a certain period in the past, weather data indicating the weather for a certain period in the past and future, and calendar data indicating at least the days of the week for each day, and outputs forecast demand data indicating predicted values of water demand for the forecast period; a second forecasting model that does not input the weather data and the calendar data, but inputs the actual demand data, and outputs the forecast demand data; and a third forecasting model that does not input the actual demand data, but inputs the weather data and the calendar data, and outputs the forecast demand data. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram showing an example of the overall configuration of a water management system according to an embodiment. [Figure 2] FIG. 2 is a diagram showing a modification of the system configuration shown in FIG. [Figure 3] FIG. 3 is a diagram illustrating an example of the configuration of the demand prediction device 105. [Figure 4] FIG. 4 is a diagram showing a schematic relationship between information on the demand prediction unit 207 side and information on the water management planning device 106 side. [Figure 5] FIG. 5 is a diagram schematically showing specific examples of input data and output data of the first prediction model M1, the second prediction model M2, and the third prediction model M3. [Figure 6] FIG. 6 is a diagram showing an example of the temporal arrangement relationship between input data used in predicting water demand and output data obtained as a prediction result. [Figure 7] FIG. 7 is a diagram showing an example of the form of the input data shown in FIG. 6 when the input data is input to the first prediction model M1. [Figure 8] FIG. 8 is a diagram showing an example of a case where there exists a result value of a demand amount that exceeds an upper limit value or a result value of a demand amount that falls below a lower limit value in the result demand amount data. [Figure 9] FIG. 9 shows examples of two methods for determining the upper and lower limit values of the water distribution volume upper and lower limiters based on the distribution of actual water distribution volume values at a specific time (time T) every day over a certain period in the past. [Figure 10] FIG. 10 is a diagram showing an example of the temporal arrangement relationship between the "actual water distribution amount data" which is one of the input data used for learning and the "correct water distribution amount data" which is the correct data. [Figure 11A] FIG. 11A is a diagram showing an example of an integrated value per unit time of the actual water distribution amount value to be added to the "correct water distribution amount data." [Figure 11B] FIG. 11B is a diagram showing an example of the range (accumulation range) of the accumulated value of the actual water distribution amount to be added to the "correct water distribution amount data." [Figure 12] FIG. 12 is a diagram showing an example of dividing learning data into training data and test data. [Figure 13] FIG. 13 is a diagram illustrating an example of the basic operation of the demand forecasting device 105. [Figure 14A] FIG. 14A is a flowchart (first half) showing an example of the operation of prediction model learning unit 203. [Figure 14B] FIG. 14B is a flowchart showing an example of the operation of prediction model learning unit 203 (second half). [Figure 15] FIG. 15 is a diagram showing an example of the operation of the demand amount prediction unit 207. [Figure 16A] FIG. 16A is a flowchart (first half) showing a modified example of the operation of prediction model learning unit 203. [Figure 16B] FIG. 16B is a flowchart (second half) showing a modified example of the operation of prediction model learning unit 203. [Figure 17] FIG. 17 is a diagram showing a modified example of the operation of the demand quantity prediction unit 207 corresponding to the modified example of the operation of the prediction model learning unit 203 shown in FIGS. 16A and 16B. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments will be described with reference to the drawings.

[0011] (Overall system configuration) FIG. 1 shows an example of the overall configuration of a water management system according to an embodiment.

[0012] The water management system shown in FIG. 1 includes a water purification plant 1, a water reservoir 2, a water distribution area 3, a central monitoring and control system 4, and a plant equipment control device 5.

[0013] The water purification plant 1 performs various treatments on raw water to produce purified water in a purified water reservoir 1a. The produced purified water is transported by a water supply pump 11 and sent to a distribution reservoir 2. The number of water supply pumps 11 is not limited to one, and may be multiple. The purified water stored in the distribution reservoir 2 is distributed by the water supply pump 11 to consumers (not shown) in a distribution area 3 by a distribution pump (not shown) or by gravity feeding.

[0014] The distributing reservoir 2 is provided with equipment 10 including sensors related to the water treatment process. The equipment 10 is provided with a water level meter 21 that measures the water level of the distributing reservoir 2, as well as a flow meter 22 that measures the flow rate of water sent from the distributing reservoir 2 to the distribution area 3.

[0015] The central monitoring and control system 4 includes a calendar information database (DB) 210, a water reservoir water level database (DB) 101, an actual demand database (DB) 102, a weather information database (DB) 103, a water operation parameter database (DB) 104, a demand forecasting device 105, a water operation planning device 106, an operation plan DB database (DB) 107, and the like.

[0016] Some or all of the functions of the demand forecasting device 105 and the water operation planning device 106 may be realized as a program executed by a processor such as a central processing unit provided in one or more computers.

[0017] The calendar day information DB 201 stores calendar day information data (calendar day data) that indicates the day type, day of the week, number of days since the beginning of the year, etc. Day types are types of weekdays, holidays, and special days. Special days indicate types of special days, such as the New Year holidays, consecutive holidays in May, Obon, and public holidays. The calendar day data can be used to determine, for example, whether a certain day corresponds to a specific day of the week or a specific public holiday.

[0018] The reservoir water level DB 101 accumulates data (water level information) indicating the actual water level value of the reservoir 2 measured by the water level gauge 21 at predetermined intervals.

[0019] The actual demand DB 102 accumulates data indicating actual water distribution amounts (also referred to as actual demand amounts) measured by the flowmeter 22 at predetermined intervals.

[0020] The weather information DB 103 stores weather information data (weather data) indicating past and future weather conditions that are input by an operator via a user interface (not shown) or automatically input by a weather information processing device (not shown). The past weather data indicates actual values of weather and temperature, while the future weather data indicates forecast values of weather and temperature. Weather may be classified, for example, as "sunny," "cloudy," or "rainy," and may be divided into morning and afternoon. The maximum and minimum temperatures are stored. Forecast values are stored, for example, for up to one week in advance. The actual values and forecast values are updated at least once a day.

[0021] The water operation parameter DB104 stores data (water operation parameter information) indicating various parameters such as the regular plan start time (e.g., "0:00"), regular plan length (e.g., "24 hours"), and plan unit time (e.g., "1 hour").

[0022] The demand prediction device 105 is a device that performs processing related to prediction of the water demand (hereinafter, sometimes abbreviated as "demand prediction") for the water distribution area 3. When performing demand prediction, the demand prediction device 105 uses the actual water distribution volume values stored in the actual demand DB 102, the weather data stored in the weather information DB 103, the calendar date data stored in the calendar date information DB 201, the learning parameters stored in the learning parameter DB 202, the regular plan length and the plan unit time stored in the water operation parameter DB 104, etc.

[0023] When the demand prediction device 105 receives information indicating a calculation trigger and a plan start time, which will be described later, transmitted from the water operation planning device 106, it calculates a predicted water distribution amount (for example, a predicted water distribution amount for 25 hours starting from one hour before the plan start time). The functional configuration of the demand prediction device 105 will be described later.

[0024] The water operation planning device 106 processes water operation plans, specifically, operation plans for the water pump 11 that delivers water. Water operation plans are divided into "daily fixed plans" and "re-plans." In "daily fixed plans," a pump operation plan is implemented for a certain period of time (e.g., 24 hours) from a predetermined time (e.g., midnight) to a predetermined time of the day (e.g., midnight) or a predetermined time of the next day (e.g., midnight). In "re-plans," the pump operation plan is redone if the water level of the reservoir 2 falls outside the dead zone set above and below the planned value. For example, if the water level of the reservoir 2 is within the "planned value ± dead zone," re-planning is not necessary. In such cases, water operation plans are generally made once per day. On the other hand, if the water level of the reservoir 2 deviates from the "planned value ± dead zone," the water operation plan is redone (re-planned). To redo the water operation plan, the demand forecasting device 105 must redo (reforecast) the demand forecast. In such cases, water management planning will occur multiple times per day.

[0025] Specifically, once a day, before the regular plan start time read from the water management parameter DB 104 (for example, one hour before the regular plan start time), the water management planning device 106 transmits a trigger (calculation trigger) to the demand forecasting device 105 requesting transmission of the calculation results of the water distribution volume forecast value (also called the demand volume forecast value), as well as information indicating the plan start date and time. The trigger transmitted at this time is called the regular calculation calculation trigger.

[0026] When the water operation planning device 106 receives information indicating the predicted water distribution amount from the demand forecasting device 105, it calculates and updates the operation plan (target flow rate) of the water supply pump 11 and the planned water level value of the distributing reservoir 2 based on the received information indicating the predicted water distribution amount. Specifically, it also reads the actual water level value of the distributing reservoir 2 from the distributing reservoir water level DB 101 and the regular planned length and planned unit time from the water operation parameter DB 104, and calculates the operation plan (target flow rate) of the water supply pump 11 and the planned water level value of the distributing reservoir 2 so that the water level of the distributing reservoir 2 falls within the upper and lower operational limits while checking the operating status of the water supply pump 11 and suppressing fluctuations in the treatment volume at the water purification plant 1.

[0027] Furthermore, the water operation planning device 106 monitors the deviation between the actual water level value read from the reservoir water level DB 101 and the calculated planned water level value, and if the deviation falls outside a predetermined range, it again transmits a calculation trigger and information indicating the plan start date and time to the demand forecasting device 105. The trigger transmitted at this time is called a calculation trigger for recalculation. The calculation trigger for recalculation may also be transmitted to the demand forecasting device 105 in response to an instruction from an operator.

[0028] The operation plan DB 107 stores the operation plans calculated by the water operation planning device 106.

[0029] The plant equipment control device 5 controls the water pump 11 based on the operation plan (target flow rate value) received from the operation plan DB 107.

[0030] (Modification of the configuration of FIG. 1) Figure 2 shows a modified example of the system configuration shown in Figure 1. The following will mainly explain the differences from Figure 1.

[0031] 2 includes the water purification plant 1, water reservoir 2, water distribution district 3, central monitoring and control system 4, and plant equipment control device 5, as well as a wide-area monitoring and control system 6 located on the cloud. In this case, the calendar information database (DB) 210, water reservoir water level DB 101, actual demand DB 102, weather information DB 103, water operation parameter DB 104, demand forecasting device 105, and water operation planning device 106 are not installed in the central monitoring and control system 4, but are installed in the wide-area monitoring and control system 6.

[0032] The wide-area monitoring and control system 6 includes a server 108 that transmits and receives information to and from each element within the wide-area monitoring and control system 6. The central monitoring and control system 4 also includes a server 109 that transmits and receives information to and from each unit. The servers 108 and 109 transmit and receive information via the network 7.

[0033] Information indicating the operating status of the water supply pump 11 is transmitted to the water operation planning device 106 via the server 108, the network 7, and the server 109. Data indicating the actual water level of the distributing reservoir 2, measured at a predetermined interval by the water level meter 21, is transmitted to the distributing reservoir water level DB 101 via the server 108, the network 7, and the server 109. Data indicating the actual distributed water amount, measured at a predetermined interval by the flow meter 22, is transmitted to the actual demand DB 102 via the server 108, the network 7, and the server 109. The operation plan calculated by the water operation planning device 106 is transmitted to the operation plan DB 107 via the server 109, the network 7, and the server 108.

[0034] In this configuration, the load of information processing and information storage on the central monitoring and control system 4 side can be reduced.

[0035] FIG. 3 shows an example of the configuration of the demand forecasting device 105.

[0036] The demand forecasting device 105 shown in Figure 3 includes a learning parameter database (DB) 202, a prediction model learning unit 203, a water distribution upper and lower limit limiter database (DB) 204, a prediction model database (DB) 205, a prediction parameter database (DB) 206, a demand prediction unit 207, and a water distribution prediction value database (DB) 208.

[0037] The learning parameter DB202 stores learning parameters such as the input data length (the number of hours indicating how many hours of actual water distribution volume values should be included in the input data, which will be described later), the number of days of learning data (the number of days indicating how many days of input data and correct answer data, which will be described later, should be included in the learning data), data correction parameters (parameters used to correct data), the number of days between re-learning (information indicating how many days between re-learning), the scheduled date for learning update, and the training / test data ratio (the ratio of learning data used for training to learning data used for testing (checking)).

[0038] The prediction model learning unit 203 reads the actual water distribution volume from the actual demand volume DB102, weather data from the weather information DB103, calendar date data from the calendar date information DB201, learning parameters from the learning parameter DB202, and the regular planning length and planning unit time from the water operation parameter DB104, and uses these data appropriately to learn and generate three types of prediction models described below for predicting the water demand (the amount of water to be distributed from the water distribution reservoir 2 to the water distribution area 3) for at least the prediction period (for example, the prediction target date).

[0039] The water distribution amount upper and lower limiter DB 204 stores upper and lower limit values (hereinafter, sometimes referred to as "upper and lower limit values of the water distribution amount upper and lower limiter") used to determine whether any of the actual water distribution amount values is missing or indicates an abnormality. The upper and lower limit values of the water distribution amount upper and lower limiter will be described in detail later.

[0040] The prediction model DB 205 stores three types of prediction models learned by the prediction model learning unit 203.

[0041] The prediction parameter DB 206 stores information indicating prediction parameters such as a threshold value for the number of predictions, a threshold value for the proportion of corrected data, and the like.

[0042] The demand prediction unit 207 predicts the water demand for the prediction period (for example, the prediction date) using one prediction model selected depending on the situation from the three types of learned prediction models.

[0043] When the demand prediction unit 207 receives the calculation trigger and information indicating the plan start date and time sent from the water operation planning device 106, it reads out the actual water distribution volume value from the actual demand volume DB 102, weather data from the weather information DB 103, calendar date data from the calendar date information DB 201, the regular plan length and planning unit time from the water operation parameter DB 104, the upper and lower limit values of the water distribution volume upper and lower limit limiter from the water distribution volume upper and lower limit limiter DB 204, the prediction model from the prediction model DB 205, and the prediction parameters from the prediction parameter DB 206, and calculates the water distribution volume prediction value.

[0044] The water distribution amount prediction value DB 208 stores the water distribution amount prediction value predicted by the demand amount prediction unit 207.

[0045] FIG. 4 shows a schematic diagram of the relationship between the information on the demand prediction unit 207 side and the information on the water operation planning device 106 side.

[0046] 4, once a day, for example, 30 minutes before the regular plan start time, the water management planning device 106 transmits a calculation trigger to the demand prediction device 105 along with information indicating the plan start date and time. When the demand prediction device 105 receives the calculation trigger and information indicating the plan start time from the water management planning device 106, it calculates the predicted water distribution amount (predicted demand amount) for each hour of the prediction period (for example, the 25 hours from 11:00 PM, which is one hour before the regular plan start time, to midnight the following day) using one of the three trained prediction models, and transmits the calculation results of the predicted water distribution amount (prediction results) to the water management planning device 106.

[0047] When the water management planning device 106 receives the calculation result of the predicted water distribution amount from the demand prediction device 105, it calculates and updates the operation plan (target flow rate) of the water supply pump 11 and the planned water level value of the distributing reservoir 2 based on the predicted water distribution amount, as described above. Furthermore, if the difference between the actual water level value and the planned water level value falls outside the specified range, the water management planning device 106 again transmits information indicating the calculation trigger and the plan start date and time to the demand prediction device 105.

[0048] In the example of Figure 4, the regular time is shown as midnight, but it is not limited to midnight and can be other times. Plans are basically made in one-hour increments, but can also be made in 30-minute or 15-minute increments. Plan lengths are basically made in 24-hour increments, but can also be 36-hour or 48-hour increments.

[0049] (3 types of prediction models) The three types of prediction models are referred to as a first prediction model M1, a second prediction model M2, and a third prediction model M3, respectively.

[0050] The first prediction model M1 inputs "actual demand data" (actual water distribution data) that indicates the actual water demand value (actual demand value) for a certain period in the past, "weather data" that indicates the weather for a certain period in the past and future, and "calendar data" that indicates at least the day of the week for each day, and outputs "predicted demand data" (predicted water distribution data) that indicates the predicted water demand value for the prediction period.

[0051] The second prediction model M2 does not receive input of "weather data" and "calendar date data", but receives input of "actual demand data" and outputs "predicted demand data".

[0052] The third forecasting model M3 does not input "actual demand data", but inputs "weather data" and "calendar date data", and outputs "forecast demand data".

[0053] FIG. 5 shows a schematic diagram of specific examples of input data and output data of the first prediction model M1, the second prediction model M2, and the third prediction model M3.

[0054] Specifically, the first prediction model M1 generates input data using "weather data" indicating the weather and temperature, "calendar data" indicating the type of day, day of the week, etc., "prediction start time (date and time)" indicating the time to start demand prediction, and "actual demand data" indicating the actual water distribution volume for the most recent 24 hours. The first prediction model M1 is obtained by learning the input data and correct answer data for a certain period of time in the past.

[0055] In the second prediction model M2, input data is formed using the "prediction start time (date and time)" indicating the time when demand prediction starts, and "actual demand data" indicating the actual values of water distribution volume for the most recent 24 hours, for example, from among the various data used for learning by the prediction model M1. The second prediction model M2 is obtained by learning the input data and correct answer data for a certain period of time in the past.

[0056] Of the various types of data used by the prediction model M1 for learning, the input data for the third prediction model M3 is formed using "weather data" indicating the weather and temperature, "calendar data" indicating the type of day, day of the week, etc., and the "prediction start time (date and time)" indicating the time when demand prediction should begin. The third prediction model M3 is obtained by learning the input data and correct answer data from a certain period in the past.

[0057] The output data includes information on the predicted water distribution volume (predicted demand volume) for 25 hours from the prediction start time.

[0058] A specific example of learning using input data and correct answer data will be described later.

[0059] (Demand forecast using each forecast model) When making a demand prediction, the demand prediction unit 207 uses three types of prediction models depending on the situation.

[0060] Prediction using the first prediction model M1 The demand prediction unit 207 predicts the water demand using the first prediction model M1 if the rate at which missing or abnormal values are included in the actual demand data for a certain period in the past does not exceed a predetermined threshold, and if the number of prediction redoes (number of predictions) for the prediction period does not exceed a predetermined threshold. As a result, if normal data is obtained for normal demand, it is possible to perform an accurate demand prediction by taking into account the most recent actual demand value, the season, weather, day of the week, etc.

[0061] - Forecast using the second forecasting model M2 The demand prediction unit 207 predicts the water demand using the second prediction model M2 if the proportion of missing or abnormal values included in the actual demand data for a certain period in the past does not exceed a predetermined threshold and if the number of prediction redoes (number of predictions) for the prediction period exceeds a predetermined threshold. As a result, even if normal data is obtained, if the demand is different from usual, for example, the actual demand data is provided to the second prediction model M2 and the demand prediction is performed without taking into account the season, weather, day of the week, etc., thereby preventing a decrease in prediction accuracy.

[0062] - Forecast using the third forecasting model, M3 The demand prediction unit 207 predicts the water demand using the third prediction model M3 when the proportion of missing or abnormal values included in the actual demand data for a certain period in the past exceeds a predetermined threshold. As a result, when normal data is not available, the third prediction model M3 is provided with weather data and calendar date data to predict the demand, thereby avoiding a situation where prediction cannot be made.

[0063] FIG. 6 shows a schematic diagram of an example of the temporal arrangement of input data used to predict water demand and output data obtained as a prediction result.

[0064] As described above, when the demand prediction unit 207 receives the calculation trigger and information indicating the plan start date and time sent from the water operation planning device 106, it reads out "actual water distribution data" indicating the actual water distribution volume values for a certain period from the actual demand DB 102, "weather data" for a certain period from the weather information DB 103, "calendar date data" for a certain period from the calendar date information DB 201, data indicating the regular plan length and planning unit time from the water operation parameter DB 104, data indicating the upper and lower limit values of the water distribution volume upper and lower limit limiters from the water distribution volume upper and lower limit limiter DB 204, the relevant prediction model from the prediction model DB 205, and prediction parameters from the prediction parameter DB 206, and selectively uses one of the three prediction models to generate "predicted water distribution volume data" indicating the predicted water distribution volume value, and transmits the generated predicted water distribution volume data to the water operation planning device 106 as the prediction result. FIG. 6 shows an example of the temporal arrangement of the "actual water distribution amount data," "predicted water distribution amount data," "weather data," and "calendar date data."

[0065] Here, a case will be considered in which a demand amount is predicted using the first prediction model M1 after learning. Details of the learning method will be described later.

[0066] As shown in Figure 6, the water operation plan period is "from midnight to midnight on October 26, 2021" (24 hours), and the regular plan start time notified in advance by the water operation planning device 106 to the demand prediction unit 207 is "midnight on October 26, 2021."

[0067] In this case, the demand prediction unit 207 sets the prediction start time (date and time) to 23:00 on October 25, 2021 to generate "predicted water distribution data" indicating the predicted water distribution value for the period from 23:00 on October 25, 2021 to 24:00 on October 26, 2021 (25 hours) as output data, and creates "actual water distribution data" for the period from 23:00 on October 24, 2021 to 23:00 on October 25, 2021 (24 hours) as input data, as well as "weather data" and "calendar data" for the period from 24th October 2021 to 26th October 2021 (3 days).

[0068] The periods for "weather data" and "calendar data" shall encompass (cover) the periods for "forecasted water distribution data" and "actual water distribution data." As mentioned above, the weather in "weather data" shall divide a day into morning and afternoon and categorize it as "sunny," "cloudy," "rainy," etc., and shall indicate the maximum and minimum temperatures. "Calendar data" shall be shown on a daily basis.

[0069] FIG. 7 shows an example of the form of the input data shown in FIG. 6 when the input data is input to the first prediction model M1.

[0070] The input data shown in FIG. 7 is arranged in the following order from the left: "prediction start date and time," "weather data," "calendar date data," and "actual water distribution amount data."

[0071] As mentioned above, the "Predicted start date and time" indicates "11:00 PM on October 25, 2021."

[0072] The "weather data" shows the "morning weather," "afternoon weather," "minimum temperature," and "maximum temperature" for each of the three days: the previous day, the current day, and the following day. In this case, the current day corresponds to the start date and time of the forecast, i.e., October 25, 2021.

[0073] The "Calendar Day Data" shows the "Day Type," "Day of the Week," and "Day of the Year" for each of the three days: the previous day, the current day, and the following day.

[0074] "Actual water distribution data" is the water distribution amount per hour for 24 hours [m 3 ] shows the actual value.

[0075] (Learning each prediction model) The prediction model learning unit 203 uses learning data to learn a prediction model using AI (Artificial Intelligence). For example, the prediction model learning unit 203 provides the AI with "a set of input data and correct answer data for the number of days of learning data" as learning data, and trains the prediction model so that it can output output data close to the correct answer data. As a result of learning, a new prediction model is obtained.

[0076] Below, we will explain the learning methods for each of the three types of predictive models.

[0077] The prediction model learning unit 203 has the function of learning using correct demand data (correct data) that indicates correct values for predicted values (demand prediction values) of water demand for at least the prediction period as correct data for a certain period in the past, and generating a first prediction model M1, a second prediction model M2, and a third prediction model M3 for predicting water demand for at least the prediction period.

[0078] Specifically, the prediction model learning unit 203 learns a first prediction model M1 using at least actual demand data, weather data, calendar date data, and correct demand data from a certain period of the past, learns a second prediction model M2 using at least actual demand data and correct demand data from a certain period of the past, and learns a third prediction model M3 using at least weather data, calendar date data, and correct demand data from a certain period of the past.

[0079] If the actual demand data used for learning contains missing or abnormal values, the accuracy of the demand prediction may be significantly reduced. Therefore, the prediction model learning unit 203 has a function to determine upper and lower limits for the water distribution volume upper and lower limiters based on the distribution of actual demand values over a certain period of time in the past, and to determine that actual demand values in the actual demand data that exceed the upper limit or fall below the lower limit are values that indicate an abnormality. This makes it possible to determine whether normal actual demand data is being obtained when learning is performed.

[0080] Figure 8 shows an example of a case where the actual demand data contains actual demand values that exceed the upper limit or fall below the lower limit. The graph in Figure 8 shows actual demand data for 365 days. In the graph in Figure 8, the horizontal axis represents time [hours], and the vertical axis represents the water distribution volume [m 3 In this graph, upper and lower limits are set based on the distribution of actual demand values over a certain period of time in the past, and there are actual demand values that exceed the upper limit or fall below the lower limit. Such actual demand values are determined to be abnormal.

[0081] FIG. 9 shows two examples of methods for determining the upper and lower limit values of the water distribution amount upper and lower limiters based on the distribution of actual values of water distribution amount at a predetermined time (time T) every day for a certain period in the past.

[0082] Figure 9(a) shows the concept of the first method. In the example of Figure 9(a), the mean and standard deviation (σ) of the distribution of actual water distribution volume values at a predetermined time (time T) are found, and the upper and lower limits are determined by calculating "the mean of the distribution ± the standard deviation (σ) × a predetermined multiple." In the example of Figure 9(a), the predetermined multiple is set to 3, but it may be a number other than 3. It also need not be an integer multiple.

[0083] Figure 9(b) shows the concept of the second method. In the example of Figure 9(b), the upper and lower limits of a data range (w) that excludes a few percent (for example, 3%) of the data at the top and bottom of the distribution of actual water distribution volume values at a specified time (T) are shifted up or down by a coefficient (k) times the data range (w).

[0084] In this way, the upper and lower limit values for determining obvious outliers are set. The parameters required for calculating the upper and lower limit values of the upper and lower limiters for the water distribution volume are stored in the learning parameter DB 202 as data correction parameters, and are read out and used.

[0085] The prediction model learning unit 203 also has a function of learning the first and second prediction models M1 and M2 after correcting actual demand values that exceed the upper limit of the upper and lower limit water distribution limiters, actual demand values that fall below the lower limit, and missing values of actual demand values. This prevents the first and second prediction models M1 and M2 from being affected by missing or abnormal values.

[0086] For example, the demand amount may be corrected by replacing actual demand amounts that exceed the upper limit of the upper and lower limit limiters with the same value as the upper limit, and replacing actual demand amounts that fall below the lower limit and missing values of the actual demand amount with the same value as the lower limit.

[0087] As another correction method, actual demand values that fall outside the range of the upper and lower limits of the upper and lower limiters for water distribution volume and missing values for actual water distribution volume may be linearly interpolated using valid demand volume data before and after.

[0088] Figure 8 shows an example in which linear interpolation is performed on individual actual demand values (values indicating an abnormality) that are outside the upper and lower limits of the upper and lower water distribution limiters, using the previous and following valid actual demand values (actual demand values that are not outside the upper and lower limits of the upper and lower water distribution limiters).

[0089] Furthermore, when the actual demand data contains actual demand values that have been assigned information indicating that they are irregular, the prediction model training unit 203 has a function of correcting the actual demand values and training the first prediction model M1, and training the second prediction model M2 without correcting the actual demand values. The information indicating "irregular" indicates that the demand has increased or decreased due to an unexpected factor and is assigned in advance by an operator via a user interface (not shown) or automatically by the demand prediction device 105. As a result, even when the demand has increased or decreased due to an unexpected factor, the second prediction model M2 can predict the demand similar to that on a day when the actual water distribution volume values showed a similar trend. Meanwhile, when the demand shows a normal trend, the first prediction model M1 can predict the demand similar to that on a day when the demand trend is similar to the weather and day of the week conditions.

[0090] The demand prediction unit 207 has a function of comparing each predicted value of demand contained in the predicted demand data output from the first or third prediction model M1, M3 with each actual value of demand for the corresponding period, and if the difference between the two exceeds a predetermined value, notifying information indicating that the demand is irregular. This allows the operator who receives the notification to know without delay that the demand has increased or decreased due to an unexpected factor, and can add information indicating that the actual value of the water distribution volume is "irregular" when the demand has increased or decreased due to an unexpected factor.

[0091] Specifically, when making the first prediction each day, the demand prediction unit 207 reads the first predicted water distribution amount value of the previous day from the predicted water distribution amount DB 208, compares it with the actual water distribution amount value for the same period read from the actual demand amount DB 102, and if the difference is equal to or greater than a predetermined threshold, notifies the operator that the actual water distribution amount value of the previous day is "irregular" via a GUI (Graphical User Interface) screen on a display device (not shown) or an audio device. The difference between the actual water distribution amount value and the predicted water distribution amount value can be evaluated by the error in the 24-hour integrated value or the mean square error per unit time.

[0092] (Predictive model learning) When carrying out learning, there are two methods for learning each predictive model, as shown below.

[0093] ·First learning method When the first learning method is applied, the prediction model learning unit 203 registers input data and correct answer data shifted by one hour per day in the learning data for each time, and learns the first, second, and third prediction models M1, M2, and M3 using the learning data for each time, with 24 models for each time. This method makes it possible to distribute the load by shifting the timing of re-learning for each target time.

[0094] Second learning method When the second learning method is applied, the prediction model learning unit 203 registers 24 sets of input data and correct answer data shifted by one hour per day in the learning data, and learns the first, second, and third prediction models M1, M2, and M3 using the learning data, one common set for each time period. This method reduces the management burden for each prediction model.

[0095] In the following explanation of learning, an example in which the first learning method is mainly applied will be shown.

[0096] FIG. 10 shows a schematic example of the temporal arrangement relationship between the "actual water distribution amount data," which is one of the input data used for learning, and the "correct water distribution amount data," which is the correct data.

[0097] The temporal positional relationship between the "actual water distribution amount data" and the "correct water distribution amount data" is the same as the temporal positional relationship between the "actual water distribution amount data" and the "predicted water distribution amount data" shown in FIG.

[0098] The input data used for learning includes "actual water distribution volume data," as well as "weather data" and "calendar data" as shown in Figure 6. However, to avoid complicating the illustration, the "weather data" and "calendar data" have been omitted here.

[0099] The temporal positional relationship between the "actual water distribution volume data" and "correct water distribution volume data" shown in Figure 10 and the "weather data" and "calendar date data" not shown is the same as the temporal positional relationship between the "actual water distribution volume data" and "predicted water distribution volume data" and the "weather data" and "calendar date data" shown in Figure 6.

[0100] As described above, the prediction model learning unit 203 learns a first prediction model M1 using at least actual demand data, weather data, calendar date data, and correct demand data, learns a second prediction model M2 using at least actual demand data and correct demand data, and learns a third prediction model M3 using at least weather data, calendar date data, and correct demand data.

[0101] Before learning begins, the upper and lower limits of the water distribution upper and lower limiter are calculated for each time period based on the distribution of actual water distribution values for a certain period of time, for example, the year up to midnight on October 24, 2021.

[0102] Learning is carried out using input data and correct answer data for a predetermined number of days of learning data.

[0103] For example, when the latest confirmed time of the actual water distribution volume is "24:00 on October 24, 2021," the prediction model learning unit 203 sets the most recent hour prior to that, "24:00 on October 24, 2021," as the reference date and time, and then performs the following processing.

[0104] The prediction model learning unit 203 sets time T as the date and time (corresponding to the prediction start date and time) that is located back from the reference date and time by the regular schedule length read from the water operation parameter DB 104 plus one hour, and reads the actual water distribution volume values from the reference date and time to time T from the actual demand DB 102. After correcting any values that exceed the upper and lower limit values of the upper and lower limit values of the water distribution volume upper and lower limit limiters and any missing values, the prediction model learning unit 203 registers the data for time T as "correct water distribution volume data." Furthermore, the prediction model learning unit 203 reads from the actual demand DB 102 the actual water distribution volume values from the start date and time of the registered "correct water distribution volume data" to a date and time that is located back by the input data length read from the learning parameter DB 202, and after correcting any values that exceed the upper and lower limit values of the water distribution volume upper and lower limit limiters and any missing values, the prediction model learning unit 203 registers the data for time T as "actual water distribution volume data." In addition, the weather and temperature in the range covering the date and time of the "actual water distribution volume data" and the "correct water distribution volume data" are read from the weather information DB 103, and the day type, day of the week, etc. in the same range are read from the calendar day information DB 201. For each day in the same range, the type of day of the week and whether it is a public holiday are determined based on the information read from the calendar day information DB 201, and the determination result, together with the read information, are registered as "condition data" in the learning data for time T.

[0105] The "correct water distribution volume data" may be added with the accumulated value of the actual water distribution volume read from the actual demand volume DB 102 as the accumulated water distribution volume. In this case, the accumulated value of the water distribution volume is added to the output data of the prediction model, and therefore the accumulated value of the water distribution volume is also added to the "predicted water distribution volume data." By explicitly learning the accumulated water distribution volume value, the prediction accuracy of the total daily water distribution volume (daily water distribution volume) obtained by accumulating the predicted water distribution volume values is improved.

[0106] FIG. 11A shows an example of the accumulated value per unit time of the actual water distribution amount to be added to the "correct water distribution amount data." FIG. 11B shows an example of the range (accumulation range) of the accumulated value of the actual water distribution amount to be added to the "correct water distribution amount data." As shown in the examples of FIGS. 11A and 11B, the accumulated range of the water distribution amount may be set to 24 hours from 1 hour after time T (24 hours from T+1:00, T+2:00, ..., T+25:00). In this way, when the regular plan length is set to 36 or 48 hours, it is possible to avoid a decrease in the prediction accuracy of the daily water distribution amount by taking into account the accumulated value of the water distribution amount for more than 24 hours in the future.

[0107] The same process as above is performed for other multiple reference dates and times, which are obtained by moving the reference date and time back by one hour at a time, for the number of days of learning data specified in advance.If any of the actual water distribution volume values in the "correct water distribution volume data" have been corrected, that "correct water distribution volume data" and the actual water distribution volume data and condition data corresponding to the same reference date and time are deleted from the learning data.

[0108] Based on the learning data for each time created in this way, the prediction model learning unit 203 learns the first, second, and third prediction models M1, M2, and M3, 24 of each for each time, as follows.

[0109] First, a prediction model M2 for time T is trained using input data including actual water distribution amount data registered in the training data for time T and correct water distribution amount data.

[0110] Thereafter, if any of the actual water distribution volume values in the correct water distribution volume data is given information indicating that it is "irregular," that correct water distribution volume data and the actual water distribution volume data and condition data corresponding to the same reference date and time are deleted from the learning data for time T. Also, if any of the actual water distribution volume data is given information indicating that it is "irregular," that actual water distribution volume value is considered to be missing, and that actual water distribution volume value is corrected. The correction may be to the lower limit value of the lower limit water distribution volume limiter, or linear interpolation using the data before and after.

[0111] Then, a prediction model M3 for time T is trained using input data including the condition data registered in the training data for time T and the correct water distribution amount data.

[0112] Thereafter, if the proportion of the actual water distribution volume data that includes corrected values exceeds the threshold proportion of corrected data read from the prediction parameter DB206, the actual water distribution volume data and the correct water distribution volume data and condition data that correspond to the reference date and time are deleted from the learning data for time T.

[0113] Then, a prediction model M1 for time T is trained using input data including the actual water distribution amount data and condition data registered in the training data for time T, and the correct water distribution amount data.

[0114] The above process is carried out 24 times, T=0 to 23.

[0115] When training a prediction model that is common to all time periods, the training data is also one that is common to all time periods, and in addition to the actual water distribution volume data, condition data, and correct water distribution volume data, the start time of the correct water distribution volume data (equivalent to the time of the prediction start date and time) is registered in the training data as time data. Prediction models M1 to M3 are trained in the same way as when training prediction models for each time period. However, when training any prediction model, time data is additionally used.

[0116] FIG. 11 shows an example of dividing learning data into training data and test data.

[0117] The learning data for each time (weather data, calendar date data, actual water distribution data, correct water distribution data) can be divided into training data and test data as shown in Figure 11, and if the prediction accuracy when the test data is given to the prediction model obtained by learning using the training data does not meet the standard, it can be not adopted.

[0118] The ratio of training data to test data is determined by the training / test data ratio stored in the learning parameter DB 202. For example, if the ratio of training data to test data is "3:1", then of the set of input data and correct answer data in the learning data for time T, 3 / 4 is allocated to training data and 1 / 4 is allocated to test data.

[0119] In addition, when the prediction model learning unit 203 or the demand prediction unit 207 reads out the actual water distribution volume value from the actual demand volume DB 102, if the recording period of the actual water distribution volume value is shorter than the planned unit time read out from the water operation parameter DB 104 and the value is converted into an integrated value for the planned unit time, the integrated value containing missing values may be treated as a missing value.

[0120] (An example of the basic operation of the demand forecasting device 105) FIG. 12 shows an example of the basic operation of the demand forecasting device 105.

[0121] An example of the basic operation of the demand forecasting device 105 will be described with reference to the flowchart of FIG.

[0122] First, by referring to the "scheduled learning update date" stored in the learning parameter DB 202, it is determined whether the current date and time has reached the scheduled learning update date (S1).

[0123] If the current date and time has reached the scheduled learning update date, learning of each prediction model is carried out (S2). A specific example of the learning process will be described later. When learning is complete, the scheduled learning update date is updated based on the "number of days between re-learning" stored in the learning parameter DB 202 (S3).

[0124] Thereafter, it is determined whether or not a calculation trigger has been sent from the water management planning device 106 (S4).

[0125] When a calculation trigger is sent from the water operation planning device 106, a demand prediction is carried out (S5). A specific example of the demand prediction process will be described later.

[0126] The demand forecasting device 105 repeats this process.

[0127] (An example of the operation of the prediction model learning unit 203) 13A and 13B show an example of the operation of the prediction model learning unit 203. In FIG.

[0128] An example of the operation of the prediction model learning unit 203 will be described with reference to the flowcharts of FIGS. 13A and 13B.

[0129] First, the actual water distribution volume values for the latest fixed period (for example, one year) are read from the actual demand volume DB 102. If the recording cycle of the actual water distribution volume values is shorter than the planned unit time read from the water operation parameter DB 104, they are converted into an integrated value for the planned unit time (for example, converting one minute to one hour), and then the upper and lower limit values of the upper and lower water distribution volume limiter are calculated for each hour based on the distribution of the actual water distribution volume values at each hour (S101).

[0130] Next, the latest hour before the latest data determination date and time is substituted for the reference date and time for generating learning data (S102).

[0131] Next, the following process is repeated for the number of days of learning data read from the learning parameter DB202 (S103), and further 24 times per day (every hour) (S104) to generate time data indicating the time, correct water distribution data (hereinafter abbreviated as "correct data"), actual water distribution data, and condition data for learning the prediction model at each time.

[0132] First, the time that is “regular schedule length + 1 hour” back from the reference date and time is set to T (a value between 0 and 23), and T is registered in the Nth row of the time data for time T (one of multiple rows prepared for the number of days of learning data) (S105).

[0133] Next, the actual water distribution volume values from the reference date and time up to the date and time preceding the regular plan length read from the water operation parameter DB 104 are read from the actual demand DB 102, and among the read actual water distribution volume values, those outside the range of the upper and lower limits of the upper and lower water distribution volume limiters and missing values are corrected, and then registered in the Nth row of the correct data for time T (S106). At this time, the integrated value of the actual water distribution volume values may be added to the correct data. If the recording period of the actual water distribution volume values is shorter than the planned unit time read from the water operation parameter DB 104, they are converted to an integrated value for the planned unit time and then registered.

[0134] Next, the actual water distribution volume values from the start date and time of the registered correct answer data to the date and time going back by the input data length read from the learning parameter DB202 are read from the actual demand volume DB102, and among the read actual water distribution volume values, those that fall outside the range of the upper and lower limits of the upper and lower water distribution volume limiters and missing values are corrected, and then registered in the Nth row of the actual water distribution volume data for time T (S107). If the recording period of the actual water distribution volume values is shorter than the planned unit time read from the water operation parameter DB104, they are converted into an integrated value for the planned unit time and then registered.

[0135] Next, the weather and temperature in the range covering the date and time of the registered actual water distribution amount data and correct data is read from the weather information DB 103, and for each day in the same range, it is determined whether it is a day of the week or a public holiday (hereinafter abbreviated as day of the week) based on the calendar day data read from the calendar day information DB 201, and the determination result is registered in the Nth row of the condition data for time T together with the read information (S108). At this time, it may also be determined which day of the year each day falls on, and the determination result may also be registered in the condition data.

[0136] Finally, the reference date and time is updated to a value that is one hour earlier (S109).

[0137] The above process is repeated 24 times per day (S110) for the number of days of learning data (S111), and 24 sets of time data, condition data, actual water distribution data, and correct answer data are obtained for each time from midnight to 11pm. If there is a row in the correct answer data for each time that contains a corrected actual water distribution value, the corresponding row is deleted from the time data, condition data, actual water distribution data, and correct answer data for the same time (S112).

[0138] Next, prediction models M1, M2, and M3 for each time are trained.

[0139] Here, the following process is repeated 24 times (S113) from T=0 o'clock to 23 o'clock.

[0140] First, input data in which time data for time T and actual water distribution amount data are linked horizontally, and correct answer data for time T are given to train a prediction model M2 for time T (S114). To train the prediction model, training of a regression model using a neural network or machine learning may be applied.

[0141] Next, if there is a row in the correct data for time T that contains information indicating that the actual water distribution volume value is "irregular," the corresponding row is deleted from the time data, condition data, actual water distribution volume data, and correct data for time T.

[0142] Furthermore, if the actual water distribution volume data for time T includes an actual water distribution volume value with information indicating that it is "irregular," that actual water distribution volume value is regarded as missing and corrected (S115). Here, the information indicating that it is "irregular" is linked to an actual water distribution volume value that the operator has determined to be a demand volume different from the demand volume on a day with similar weather and day of the week conditions, and is registered in advance via a GUI screen (not shown) or the like.

[0143] Then, input data in which the time data and condition data for time T are linked horizontally, and correct answer data for time T are given to train a prediction model M3 for time T (S116).

[0144] Furthermore, if there is a row in which the proportion of the actual water distribution volume data for time T that includes values corrected to the upper and lower limit values of the water distribution volume upper and lower limit limiters exceeds the threshold proportion of corrected data read from the prediction parameter DB206, the corresponding row is deleted from the time data, condition data, actual water distribution volume data, and correct answer data for time T (S117).

[0145] Then, input data in which time data, actual water distribution data, and condition data for time T are linked horizontally, and correct answer data for time T are provided to train a prediction model M1 for time T (S118).

[0146] After learning prediction models M1, M2, and M3 for 0:00 to 23:00 (S119), prediction model learning is terminated. As a result, 3 types x 24 = 72 trained prediction models are created.

[0147] (An example of the operation of the demand prediction unit 207) FIG. 14 shows an example of the operation of the demand prediction unit 207.

[0148] An example of the operation of the demand amount prediction unit 207 will be described with reference to the flowchart of FIG.

[0149] First, it is determined whether the calculation trigger received from the water operation planning device 106 is for a scheduled calculation (S201). If it is for a scheduled calculation, the number of predictions for the target prediction day is reset (S202). If it is for a recalculation, the number of predictions for the target prediction day is incremented (S203).

[0150] Next, the following process is performed to generate new time data, actual water distribution amount data, and condition data for demand prediction.

[0151] First, the time one hour before the plan start date and time received from the water operation planning device 106 is set to T, and T is registered as time data (S204).

[0152] Next, the actual water distribution volume values from the planning start date and time to the date and time going back by the input data length read from the learning parameter DB202 are read from the actual demand DB102, and if the recording period of the actual water distribution volume values is shorter than the planning unit time read from the water operation parameter DB104, they are converted to an integrated value for the planning unit time, and those that exceed the range of the upper and lower limit values of the water distribution volume upper and lower limit limiter calculated in S101 are linearly interpolated using the valid actual water distribution volume values before and after, and then registered in the actual water distribution volume data (S205).

[0153] Next, the weather and temperature data covering the range from the date and time of the registered actual water distribution data and the plan start date and time to the date and time that is the regular plan length are read from the weather information DB 103, and the day of the week for each day in the same range is determined based on the calendar date data read from the calendar date information DB 201, and the determination result is registered in the condition data together with the read information (S206). At this time, the number of days since the start of the year for each day may also be calculated and registered in the condition data.

[0154] Next, input data for demand forecasting is created by the process shown below, and the input data is given to each forecasting model to perform demand forecasting.

[0155] It is determined whether the proportion of the actual water distribution amount data that includes corrected values exceeds the threshold value of the proportion of corrected data read from the prediction parameter DB 206 (S207). If so, the input data, which is the horizontal concatenation of time data and condition data, is provided to the prediction model M3 for time T to predict the water distribution amount (demand) (S211).

[0156] If the proportion of the actual water distribution amount data that includes corrected values does not exceed the threshold for the corrected data proportion, it is determined whether the number of predictions for the prediction target day exceeds the threshold for the number of predictions read from the prediction parameter DB 206 (S208).If so, the input data, in which the time data and the actual water distribution amount data are linked horizontally, is provided to the prediction model M2 for time T to predict the water distribution amount (demand) (S210).

[0157] If the number of predictions for the target day does not exceed the threshold number of predictions, the input data, which is a horizontal link of time data, actual water distribution data, and condition data, is provided to the prediction model M1 for time T to predict the water distribution volume (demand) (S209).

[0158] Furthermore, when the water distribution volume is predicted using the prediction model M2 (S210), when the scheduled prediction is made for the next day, a calculation trigger for the scheduled calculation sent from the water operation planning device 106 is received and the number of predictions is reset (S202), so that the prediction will be made again using the prediction model M1.

[0159] (Modification of the operation of the prediction model learning unit 203) 13A and 13B are examples in which the "first learning method" is adopted, but the "second learning method" may be adopted instead. An example of the operation of the prediction model learning unit 203 in this case is shown below.

[0160] 15A and 15B show modified examples of the operation of the prediction model learning unit 203.

[0161] A modified example of the operation of the prediction model learning unit 203 will be described with reference to the flowcharts of FIGS. 15A and 15B.

[0162] First, the actual water distribution volume values for the latest fixed period (e.g., one year) are read from the actual demand volume DB 102. If the recording cycle of the actual water distribution volume values is shorter than the planned unit time read from the water operation parameter DB 104, they are converted into an integrated value for the planned unit time (e.g., one minute converted into one hour), and then the upper and lower limit values of the upper and lower water distribution volume limiter are calculated for each hour based on the distribution of the actual water distribution volume values at each hour (S301).

[0163] Next, the latest hour before the latest data determination date and time is substituted for the reference date and time for generating learning data (S302).

[0164] Next, the following process is repeated for the number of days of learning data read from the learning parameter DB202 (S303), and further 24 times per day (every hour) (S304) to generate time data indicating the time, condition data, actual water distribution data, and correct answer data for learning the prediction model for each time.

[0165] First, the time that is "regular schedule length + 1 hour" before the reference date and time is set to T (any value between 0 and 23), and T is registered in the {(N-1)×24+H}th row of the time data for T (S305).

[0166] Next, the actual water distribution volume values from the reference date and time up to the date and time preceding the regular plan length read from the water operation parameter DB 104 are read from the actual demand DB 102, and among the read actual water distribution volume values, those outside the range of the upper and lower limits of the upper and lower water distribution volume limiters and missing values are corrected, and then registered in the {(N-1) x 24 + H}th row of the correct data (S306). At this time, the integrated value of the actual water distribution volume values may be added to the correct data. If the recording period of the actual water distribution volume values is shorter than the planned unit time read from the water operation parameter DB 104, they are converted to an integrated value for the planned unit time and then registered.

[0167] Next, the actual water distribution volume values from the start date and time of the registered correct answer data to the date and time going back by the input data length read from the learning parameter DB202 are read from the actual demand volume DB102, and among the read actual water distribution volume values, those outside the range of the upper and lower limits of the upper and lower water distribution volume limiters and missing values are corrected, and then registered in the {(N-1)×24+H}th row of the actual water distribution volume data (S307). If the recording period of the actual water distribution volume values is shorter than the planned unit time read from the water operation parameter DB104, they are converted into an integrated value for the planned unit time and then registered.

[0168] Next, the weather and temperature data for the range covering the date and time of the registered actual water distribution data and correct data are read from the weather information DB 103, and for each day in the same range, it is determined whether it is a day of the week or a public holiday (hereinafter referred to as a day of the week) based on the calendar day data read from the calendar day information DB 201, and the determination result is registered in the {(N-1) x 24 + H}th row of the condition data together with the read information (S308). At this time, it may also be determined which day of the year each day falls on, and the determination result may also be registered in the condition data.

[0169] Finally, the reference date and time is updated to a value that is one hour earlier (S309).

[0170] The above process is repeated 24 times per day (S310) for the number of days of learning data (S311) to obtain 24 hours of time data, condition data, actual water distribution data, and correct answer data. If there is a row in the correct answer data for each time that contains a corrected actual water distribution value, the corresponding row is deleted from the time data, condition data, actual water distribution data, and correct answer data for the same time (S312).

[0171] In the following process, prediction models M1, M2, and M3 are each trained once.

[0172] First, the prediction model M2 is trained by providing input data in which 24-hour time data and actual water distribution amount data are linked horizontally, and correct answer data (S313).

[0173] Next, if there is a row in the 24-hour correct data where the actual water distribution volume value is marked with information indicating that it is "irregular," the corresponding row is deleted from the time data, condition data, actual water distribution volume data, and correct data.

[0174] Furthermore, if the 24-hour actual water distribution data includes an actual water distribution value that has been assigned information indicating that it is "irregular," that actual water distribution value is considered to be missing and is corrected (S314). Here, the information indicating that it is "irregular" is linked to an actual water distribution value that the operator has determined to be a demand amount that is different from the demand amount on a day with similar weather and day of the week conditions, and is registered in advance via a GUI screen (not shown) or the like.

[0175] Then, the prediction model M3 is trained by providing input data in which 24 hours of time data and condition data are linked horizontally, and 24 hours of correct answer data (S315).

[0176] Furthermore, if there is a row in which the proportion of the actual water distribution volume data for 24 hours that includes values corrected to the upper and lower limit values of the water distribution volume upper and lower limit limiters exceeds the threshold proportion of corrected data read from the prediction parameter DB206, the corresponding row is deleted from the 24-hour time data, condition data, actual water distribution volume data, and correct answer data (S316).

[0177] Then, the prediction model M1 is trained by providing input data in which 24 hours of time data, actual water distribution amount data, and condition data are linked horizontally, and 24 hours of correct answer data (S317).

[0178] This results in trained predictive models M1, M2, and M3.

[0179] (Modification of the operation of the demand prediction unit 207) FIG. 16 shows a modified example of the operation of the demand quantity prediction unit 207 corresponding to the modified example of the operation of the prediction model learning unit 203 shown in FIGS. 15A and 15B.

[0180] The processing of S401 to S411 shown in FIG. 16 is almost the same as the processing of S201 to S211 shown in FIG. 14 described above, and can be understood from the above description, so description thereof will be omitted here.

[0181] (Effects of the embodiment) As described above, according to the embodiment, it is possible to predict water demand with a certain level of accuracy or higher, even if demand increases or decreases due to unexpected factors, or even if at least some of the actual demand values are missing or abnormal. Specifically, the following effects can be obtained, for example.

[0182] By learning prediction models M1, M2, and M3 with different input data, it is possible to predict water distribution volume using a prediction model that is appropriate for the situation at the time of prediction, thereby preventing a decline in prediction accuracy.

[0183] If the proportion of missing or abnormal values in the actual demand data for a certain period in the past does not exceed a predetermined threshold, and if the number of re-predictions (number of predictions) for the prediction period exceeds a predetermined threshold, the second prediction model M2 can be used to predict water demand, thereby preventing a decline in prediction accuracy. For example, even if normal data is obtained, if the demand is different from usual, the actual demand data can be fed into the second prediction model M2 to predict demand without taking into account the season, weather, day of the week, etc., making it possible to make predictions similar to those for similar days regardless of weather or day of the week conditions, thereby preventing a decline in prediction accuracy.

[0184] If the percentage of missing or abnormal values in the actual demand data for a certain period in the past exceeds a predetermined threshold, the third prediction model M3 can be used to predict water demand, thereby avoiding situations where predictions cannot be made. For example, when normal data is not available, weather data and calendar day data can be provided to the third prediction model M3 to predict demand, making it possible to make a prediction close to that of a day with similar weather and day of the week conditions.

[0185] By calculating the upper and lower limit values of the water distribution volume upper and lower limiters from the distribution of actual water distribution volume values at each time over a certain period in the past, it is possible to determine whether the actual water distribution volume values are missing or indicate an abnormality.

[0186] By training the prediction models M1 and M2 using actual water distribution volume values in which missing or anomalous values have been replaced with the lower or upper limit values of the upper and lower water distribution volume limiters, if the actual water distribution volume value corrected to the upper or lower limit values of the upper and lower water distribution volume limiters is small (i.e., if the actual water distribution volume value that was missing or anomalous is small), demand prediction can be performed using the first prediction model M1 or the second M2. Also, when the demand prediction unit 207 repeatedly receives calculation triggers from the water operation planning device 106, even if the most recent actual water distribution volume value is missing or anomalous, prediction model M2 can be used to perform a prediction similar to that for a similar day, regardless of weather or day of the week conditions.

[0187] Among the actual water distribution volume data, those with information indicating that they are "irregular" are used as is to train prediction model M2, and those with information indicating that they are "irregular" are treated as missing values and corrected before training prediction models M1 and M3.As a result, prediction model M2 can make predictions that are close to when the actual water distribution volume values show similar movements, even when demand increases or decreases due to sudden factors, and prediction models M1 and M3 can make predictions that are close to when the weather and day of the week conditions are similar.

[0188] When the difference between the actual water distribution volume and the predicted water distribution volume from the previous day is greater than a threshold, the system determines that the demand volume is irregular and notifies the operator. This allows the operator to know without delay that the demand volume has increased or decreased due to a sudden factor, and can add information indicating that the actual water distribution volume is "irregular" when the demand volume has increased or decreased due to a sudden factor.

[0189] By registering input data and correct answer data shifted by one hour per day in the learning data for each time, and training the first, second, and third prediction models M1, M2, and M3 using 24 pieces of learning data for each time, the timing of re-training for each target time can be shifted, thereby distributing the load.On the other hand, by registering 24 pieces of input data and correct answer data shifted by one hour per day in the learning data, and training the first, second, and third prediction models M1, M2, and M3 using one piece of learning data that is common to all times, the management load of the prediction models can be reduced.

[0190] By providing correct water distribution data, including the accumulated water distribution value, and training the prediction models M1, M2, and M3, it is expected that the prediction accuracy of the daily water distribution volume (daily water distribution volume), which is the accumulated water distribution volume forecast value for 24 hours from the start time of the regular schedule, will improve.

[0191] By providing not only the day of the week but also the number of days since the beginning of the year as calendar date information and training the forecasting models M1 and M3, it is possible to make forecasts that take the season into account.

[0192] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0193] 1...water purification plant, 2...water reservoir, 3...water distribution area, 4...central monitoring and control system, 5...plant equipment control device, 6...wide-area monitoring and control system, 7...network, 10...equipment, 11...water transmission pump, 21...water level meter, 22...flow meter, 101...water reservoir water level database (DB), 102...actual demand database (DB), 103...weather information database (DB), 104...water operation parameter database (DB), 105...demand forecasting device, 106...water operation Planning device, 107...operation plan DB database (DB), 108...server, 109...server, 201...calendar information database (DB), 202...learning parameter database (DB), 203...prediction model learning unit, 204...water distribution upper and lower limit limiter database (DB), 205...prediction model database (DB), 206...prediction parameter database (DB), 207...demand prediction unit, 208...water distribution prediction value database (DB).

Claims

1. The system includes a demand prediction unit that predicts the water demand for a prediction period using one prediction model selected from a plurality of prediction models depending on the situation, The plurality of predictive models are a first prediction model that receives input of actual demand data indicating actual values of water demand for a certain period in the past, weather data indicating weather for certain periods in the past and future, and calendar data indicating at least the days of the week of each day, and outputs predicted demand data indicating predicted values of water demand for the prediction period; a second forecasting model that does not input the weather data and the calendar date data, but inputs the actual demand data, and outputs the forecast demand data; a third forecasting model that does not input the actual demand data but inputs the weather data and the calendar date data and outputs the forecast demand data; A demand forecasting device comprising:

2. The demand prediction unit If the rate at which the actual demand data contains missing or abnormal values does not exceed a predetermined threshold, and if the number of times that predictions are redone for the prediction period does not exceed a predetermined threshold, predict the water demand using the first prediction model. The demand forecasting device according to claim 1 .

3. The demand prediction unit If the rate at which the actual demand data contains missing or abnormal values does not exceed a predetermined threshold and if the number of times the prediction has been redone for the prediction period exceeds a predetermined threshold, the water demand is predicted using the second prediction model. The demand forecasting device according to claim 1 .

4. The demand prediction unit If the rate at which the actual demand data contains missing or abnormal values exceeds a predetermined threshold, the water demand is predicted using the third prediction model. The demand forecasting device according to claim 1 .

5. a prediction model learning unit that performs learning using correct answer demand data that indicates correct answer values for predicted values of water demand for at least the prediction period as correct answer data for a certain period in the past, and generates the first prediction model, the second prediction model, and the third prediction model for predicting water demand for at least the prediction period; The prediction model learning unit training the first forecasting model using at least the actual demand data, the weather data, the calendar date data, and the correct demand data for a certain period in the past; training the second prediction model using at least the actual demand quantity data and the correct demand quantity data for a certain period of time in the past; training the third forecasting model using at least the weather data, the calendar date data, and the correct demand data for a certain period of time in the past; The demand forecasting device according to claim 1 .

6. The prediction model learning unit determining an upper limit value and a lower limit value based on a distribution of actual values of the demand amount for a certain period in the past, and determining that an actual value of the demand amount in the actual demand amount data that exceeds the upper limit value or that is below the lower limit value corresponds to a value indicating an abnormality; The demand forecasting device according to claim 5 .

7. The prediction model learning unit correcting actual values of the demand amount that exceed the upper limit value, actual values of the demand amount that are below the lower limit value, and missing values of actual values of the demand amount, and then training the first and second prediction models; The demand forecasting device according to claim 6.

8. the correction includes performing linear interpolation on the actual demand value that exceeds the upper limit value, the actual demand value that is below the lower limit value, and the missing actual demand value using valid actual demand values before and after the actual demand value; The demand forecasting device according to claim 7.

9. The prediction model learning unit If the actual demand data includes an actual value of a demand to which information indicating that the demand is irregular is previously assigned, the actual value of the demand is corrected and then the first and third prediction models are trained, and the second prediction model is trained without correcting the actual value of the demand. The demand forecasting device according to claim 6.

10. The demand prediction unit comparing a predicted value of each demand included in the predicted demand data output from the first or third prediction model with an actual value of each demand for a corresponding period, and if the difference between the two exceeds a predetermined value, reporting information indicating an irregularity; The demand forecasting device according to claim 1 .

11. The integrated value of the actual value of the actual demand quantity data is added to the correct demand quantity data. The demand forecasting device according to claim 5 .

12. The integrated value is obtained by integrating the actual values of the actual demand data for 24 hours. The demand forecasting device according to claim 11.

13. The calendar date data further includes information indicating the number of days since the beginning of the year. The demand forecasting device according to claim 1 .

14. The prediction model learning unit For each day, input data and correct answer data shifted by one hour are registered in learning data for each time period, and 24 of the first, second, and third prediction models are trained for each time period using the learning data for each time period. The demand forecasting device according to any one of claims 5 to 9.

15. The prediction model learning unit For each day, 24 sets of input data and correct answer data shifted by one hour are registered in the learning data, and the first, second, and third prediction models are each trained using the learning data, one set common to all times. The demand forecasting device according to any one of claims 5 to 9.

16. a demand prediction unit predicting the water demand for a prediction period using one prediction model selected from a plurality of prediction models depending on the situation; The plurality of predictive models are a first prediction model that receives input of actual demand data indicating actual values of water demand for a certain period in the past, weather data indicating weather for certain periods in the past and future, and calendar data indicating at least the days of the week of each day, and outputs predicted demand data indicating predicted values of water demand for the prediction period; a second forecasting model that does not input the weather data and the calendar date data, but inputs the actual demand data, and outputs the forecast demand data; a third forecasting model that does not input the actual demand data but inputs the weather data and the calendar date data and outputs the forecast demand data; Demand forecasting methods, including:

17. On one or more computers, A program for realizing a function of predicting water demand for a forecast period using one prediction model selected from a plurality of prediction models depending on the situation, The plurality of predictive models are a first prediction model that receives input of actual demand data indicating actual values of water demand for a certain period in the past, weather data indicating weather for certain periods in the past and future, and calendar data indicating at least the days of the week of each day, and outputs predicted demand data indicating predicted values of water demand for the prediction period; a second forecasting model that does not input the weather data and the calendar date data, but inputs the actual demand data, and outputs the forecast demand data; a third forecasting model that does not input the actual demand data but inputs the weather data and the calendar date data and outputs the forecast demand data; Including, the program.

Images