JP2000163394A - Method for predicting water distribution demand amount - Google Patents

Abstract

PROBLEM TO BE SOLVED: To predict a water distribution demand amount with high precision by displaying a graph of the correspondence relation between the water distribution demand amount of learnt data and temperature with output unit values for respective values of an input unit of a neural network while the neural network is learning. SOLUTION: While the neural network used to predict the water distribution demand amount is learning, the learnt data 9 including items of the water distribution demand amount, temperature, dates, days of the week, weather, etc., are stored in a working memory 5 by a neural network learning control part 2. Here, the neural network learning control part 2 learns according to the learnt data 9 in the working memory 5 and various set values in a condition table and updates the combination coefficient between an input layer and an intermediate layer and the combination coefficient between the intermediate layer and an output layer. After the learning is finished when learning end conditions in the condition table are met, the graph of the correspondence relation between the learnt data 9 and water distribution demand amount is displayed on the device 12 through an output control part 7.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a water distribution amount prediction method for improving the accuracy of a water distribution amount prediction device which is a support device for planning an optimal operation plan for water supply and sewerage facilities.

[0002]

2. Description of the Related Art With the advancement and diversification of management in water and sewage systems, particularly water purification plants in water supply facilities, and the progress of automation to reduce the burden on operators, the monitoring equipment of water purification plants is also simply a state of equipment. The planning support function that provides information for planning the optimal operation plan of the water treatment plant, which has been developed from the function to monitor the water supply, has been occupying a large position.

In order to formulate an operation plan for a water purification plant, it is important to predict the amount of demanded water in a demand network (water distribution system). For example, there is a case where water is effectively distributed to a wide area demand network from a municipal large water distribution station, a prefectural large water distribution station, a booster water supply station, or the like.
The amount of demanded water is affected and fluctuated by weather and natural conditions such as weather and temperature, calendars such as days of the week, holidays, dates and seasons, and social conditions of festivals and events in the area. Therefore, many years of experience and know-how of skilled operators, numerical programming, mathematical statistical methods, and estimation and prediction methods using simple neural networks are used.

[0004]

The above-mentioned prior art has the following problems. (1) In many cases, forecasts of water demand are generally made by referring to similar past data. However, it is said that it is difficult to accurately predict the demanded water amount because of the influence of factors such as weather and temperature, natural phenomena and social life factors such as festivals and events. (2) When using a neural network for water demand forecasting and learning and using the neural network for forecasting, it is necessary to easily understand the structure and input / output relationship of the neural network from the coupling coefficient and error rate of the learning result. Can not be verified and the value cannot be verified in advance until the actual input value is entered and the predicted value is obtained. (3) When forecasting water demand using neural networks or mathematical statistical methods, past weather data does not include all combinations of weather, calendar (day of the week, etc.), temperature and actual demand water. When there is no input data that is the basis of the prediction, that is, when there is no combination of the weather, the calendar (day of the week, etc.) and the temperature close to the past actual data, the prediction error tends to increase.

The present invention has been made in consideration of the above-mentioned problems of the prior art, and an object of the present invention is to display a prediction line of a learning result and add missing data of each section. Since learning is performed, it is an object of the present invention to create a neural network having excellent generalization ability, and to provide a highly accurate water distribution demand forecasting method based on the neural network.

[0006]

Means for attaining the object are as follows: (1) In claim 1, during and / or after learning of the neural network used for the prediction of the demand for water distribution, the demand for water distribution is determined. Graphs showing the correspondence between distribution demand and temperature of learning data including items such as quantity, temperature, date, day of the week, weather, etc., and colors or other items in the learning data such as date, day of the week, weather, etc. This is a water distribution demand forecasting method for displaying an output unit value for each value of an input unit of a neural network expressed by a line type, that is, a prediction line indicating a relationship between a temperature and a predicted water demand.

(2) In the second aspect, the learning data of the neural network used for the prediction of the demand for water distribution is divided into a weather section and a month section from January to December for distinguishing between sunny, cloudy, rainy, and the like. , Has day of the week divisions from Sunday to Saturday, tallies and analyzes in each division, simulates missing data with no data in the division by multiple regression analysis, etc. according to the number of days table for each division and the division value table for each division This is a water distribution demand forecasting method in which a learning coefficient is added to learning data to learn a coupling coefficient of a neural network.

Next, the operation will be described. According to the first and second aspects of the present invention, as described in (1) and (2) above, the temperature and water distribution with respect to the year, month, day, day of the week, weather division, and the like of the learning result of the neural network constituting the water distribution amount prediction device are obtained. Since the prediction line of the demand amount is displayed and the missing data of each section is added and learned, a highly accurate water distribution amount demand prediction method can be obtained. In predicting the water distribution amount, each value of the daily results, that is, the water distribution demand amount, the temperature, the date, the day of the week, the weather, and the like are recorded and used for the prediction.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below in detail with reference to the drawings. (1) According to claim 1, during and / or after learning of a neural network used for water demand forecasting, water demand for learning data including items such as water demand, temperature, year, month, day, day of the week, and weather. Predict the output unit value, that is, temperature, for each value of the input unit of the neural network, which represents the graph of the correspondence relationship between the amount and the temperature by expressing the items in the learning data such as date, day of the week, weather, etc. in color or line type 1 to 8 show an embodiment of a water distribution demand prediction method for displaying a prediction line indicating a relationship between water distribution demands. The detailed procedure of FIGS. 1 to 8 will be described later.

(2) In claim 2, it is preferable that the training data of the neural network used for the prediction of the water distribution demand is past data for about one year or more, but the prediction is possible if it is for one month before the prediction date. is there. The training data has weather divisions to distinguish between sunny, cloudy, rainy, etc., month divisions from January to December, and day divisions from Sunday to Saturday. In accordance with the total number of days table for each section and the section value table for each section, missing data with no data in the section is pseudo-generated by multiple regression analysis etc. described below, added to the learning data, and learned the coupling coefficient of the neural network. 1 to 8 show an embodiment of the method for predicting the distribution demand. The detailed procedure of FIGS. 1 to 8 will be described later.

FIG. 1 is an overall configuration diagram of one embodiment of a water distribution amount prediction apparatus including a water distribution amount prediction method according to an embodiment of the present invention.
Is a water distribution amount prediction device, 1 is a learning data control unit, 2 is a neural network learning control unit, 3 is a neural network (coupling coefficient), 4 is a neural network prediction control unit, 5 is a working memory (learning data control unit 1, neural network). A place where temporarily input or calculated data used by the network learning control unit 2, the neural network prediction control unit 4, the input control unit 6, the output control unit 7, and the multiple regression analysis control unit 10 are stored, and various conditions. , Including a section value table having values for weather, month, and day of the week), 6 is an input control section, 7 is an output control section, 8 is actual data, 9 is learning data, and 10 is detail Is a multiple regression analysis control unit described later, 11 is an input device,
2 is an output / display device (or output to another device)
It is.

In the water distribution amount prediction device 13, when the data and various set values which are the basis of the performance data 8 are input to the input device 11, the input values are stored in the working memory 5 via the input control unit 6. The data is stored. Each of the input values described above is input to the learning data control unit 1. Here, based on the actual data in the working memory 5 and various setting values in the condition table, that is, the learning period, the range of the learning data is determined by the date. Is selected, and the corresponding performance data is assigned to the learning data 9 based on the learning data item selection conditions.
Output to The learning data creation result can be output and displayed on the output and display device 12 via the output control unit 7.

[0013] The learning data control unit 1 totalizes all data in the learning data for each section or according to the correspondence between the sections, and creates a day totaling table. The first method of adding missing data with no data in the section to the learning data is to focus on the day count table of the month section versus the weather section for each section, and for a section whose number of days is equal to or less than a specified value, for example, 10 or less. Then, seven data are generated from Sunday to Saturday. For the month and weather of these data, use the value of the corresponding section, select the day corresponding to the day of the week using random numbers, etc.The temperature is the average value of the month or the temperature between the maximum temperature and the minimum temperature of the month Use the appropriate temperature generated by random numbers, etc., and assume that the water distribution demand is the value obtained by multiplying the preset reference water distribution demand by the corresponding month value, weather value and day value in the section value table. Water distribution demand. However, since the value of each section value table indicates a ratio, it is used after dividing by 100. Months with no data in the month segment are not subject to this process.

As a second method of adding missing data having no data in a section to the learning data, the table of the number of days of the month section, the day section, and the weather section is focused on each section. To generate one data. For these data, the month, day of the week, and weather use the values of the corresponding section, and the date is selected by a random number, etc., for the day of the day.The temperature is the average of the month or the maximum and minimum temperature of the month. The appropriate water temperature is generated by random numbers and used, and the water distribution demand is a value obtained by multiplying the preset reference water distribution demand by the corresponding month value, weather value and day value in the section value table. Is assumed water demand. However, a month for which there is no data within the month segment is not subject to this processing.

As a third method for adding insufficient data having no data in the section to the learning data, the method other than the water demand is the same as the first or second method, and the water demand is obtained by multiple regression analysis. Is the way. The data of the multiple regression analysis is existing learning data, and the regression coefficient is obtained by using the values of the section value table shown in FIG. The water distribution demand is obtained from the regression coefficient, the weather, month, and day of the corresponding section converted from the section value table and the temperature. However, a month for which there is no data within the month segment is not subject to this processing. FIG. 8 shows a configuration example of the multiple regression analysis. The data obtained in this way is added to the learning data. The learning data creation result can be output and displayed on the output and display device 12 via the output control unit 7.

In the water distribution amount prediction device 13, when various set values are input to the input device 11, the input values are stored in the working memory 5 via the input control unit 6. The learning data 9 is stored in the working memory 5 by the neural network learning control unit 2. Here, learning is performed by the neural network learning control unit 2 based on the learning data on the working memory 5 and various setting values in the condition table, and the coupling coefficient between the input layer and the intermediate layer, the intermediate layer And the coupling coefficient between the output layers are updated. If the learning end condition in the condition table is met, the learning ends. The coupling coefficient after the learning is stored in the neural network (coupling coefficient) 3. The learning result is output to the output / display device 12 via the output control unit 7.
It can also be displayed.

In the water distribution amount predicting device 13, when the predicted value of the weather division and the predicted temperature value of the predicted date are inputted from the input device 11, the values inputted together with the calendar data of the preset date, month and day are input. Are stored in the working memory 5 via the input control unit 6. The neural network prediction control unit 4 is a neural network (coupling coefficient)
An output value is calculated based on the value obtained by normalizing the coupling coefficient and the input value of No. 3 by the partition value table, the value is inversely calculated, the value is used as an estimated value, and the result is output via the output control unit 7. And output / display to the device 12 for output / display.

FIG. 2 is a diagram showing an example of designating whether or not to use which of the performance data items in the learning data. In FIG. 2, among the performance data items,
It indicates that the month, day, weather, maximum temperature and actual water demand are used in the learning data.

FIG. 3 is a diagram showing an example of the condition table. As described in FIG. 2, the learning data item selection condition as a condition item indicates that the month, the day of the week, the weather, the maximum temperature, and the actual water demand are used in the learning data. The learning period indicates that data within one year from the learning date is used. The learning end condition is that the average value of the absolute value error% of the neural network is smaller than 1.5% or the number of times of learning reaches 10,000. The value of 1.5 and the value of 10000 can be changed by designation. The reference water supply demand indicates that the maximum water supply demand in the learning data is used.

FIG. 4 is an example of a section value table for designating numerical values for weather, month, and day of the week. In FIG. 4, in the case where the weather classification is sunny, cloudy, and rain (snow), three quantified values are entered for each classification. This value may be calculated from the actual data based on the ratio to the actual water demand, or may be set by an expert based on experience. Even when the weather classification is other than three, the same setting is made even if the classification method is changed. The value and method for each section for month and day are processed in the same way as for the weather section.

FIG. 5 is a view showing an example of a totaling table of the number of days in the month section versus the weather section. In this figure, the month section is 12 sections from January to December, and the weather section is three sections of sunny, cloudy, and rainy (snow), and indicates the number of days corresponding to the grid-like view. For example, a sunny day in January is 27 days, a cloudy day in January is 4 days, and a rainy (snow) day is 0 day.

FIG. 6 is a diagram showing an example of a correspondence table of symbols and colors for the sections. These symbols and colors are used when displaying and outputting, and either one of the symbols and colors may be used, or both may be used simultaneously. Further, for example, the weather division and the day division may be combined, and the weather division may be represented by a color and the day division may be represented by a symbol. FIG. 6A shows the symbols and colors for the weather sections. When the weather classification is sunny, the symbol is a circle and the color is red. When the weather classification is cloudy, the symbol is a triangle and the color is light blue. When the weather classification is rain (snow), the symbol is a diamond and the color is blue. It is. FIG. 6B shows symbols and colors for day of the week divisions. Although not shown in the figure, symbols and colors can be shown for month divisions.

FIG. 7 is a diagram showing an example of a prediction line and learning data for a weather section. In the figure, the learning data for September is displayed. Furthermore, based on the learning results of the neural network, enter September for the month section, Monday for the day section, and sunny for the weather section, and gradually increase the temperature value from the lowest value to the highest value in small increments. What is obtained by connecting the demand amount with a line is a sunny prediction line. Similarly, for the weather section, a cloudy prediction line is input when cloudy is input, and a rainy (snow) prediction line is input for rainy (snow) weather section. The symbols and colors used in FIG. 7 are those shown in FIG. In this figure, only Monday is shown, but other days can also be represented in this case. In this case, they may be displayed overlapping on the same place (coordinate system) or may be displayed at another place. Other months can be similarly represented. Furthermore, the whole learning data and figures for several months and one year can be represented. Furthermore, these charts can display and output only the designated chart.

FIG. 8 is a conceptual diagram showing an example of a multiple regression analysis. In the example of FIG. 8, the input variables (independent variables and explanatory variables) are the first power of the temperature, the second power of the temperature, the third power of the temperature, the weather, the month, and the day of the week, and the output variables (the dependent variables and the dependent variables) are This is the case of actual water demand. The relational expression between input variables and output variables is Y
i = ΣBjXij + ui, Yi is actual water demand,
Xij is an input variable, Bj is a regression coefficient, ui is an error term, i
Indicates the number of data, and j indicates the number of input variables (including the constant 1). The multiple regression analysis is a mathematical (statistical) method of calculating a regression coefficient Bj so as to minimize the sum of error terms for all data. Further, if Yi is calculated by inputting data other than the data used for the calculation, it can be used as the predicted water distribution demand. In FIG. 8, the values of the section value table of FIG. 4 corresponding to the weather, the month, and the day of the week are used.

FIG. 9 is a conceptual diagram showing an example of a neural network in the water distribution amount prediction device. As a structure and learning means of the neural network, there is a typical error back propagation method (back propagation method). (For example, "Neural Network Model and Connectionism", written by Shunichi Amari, published by the University of Tokyo Press.) The input / output relationship of the unit i other than the input layer, that is, the intermediate layer and the output layer, is as follows (1) , (2) and (3). Inputs for unit i are Oj (j = 1 to N),
The output is represented by Yi, and the coupling coefficient for each Oj is represented by Wij. Input product sum Xi = Σ WijOj (1) Formula (1) is applied to function f (Xi) for conversion. In general, a sigmoid function such as the following equation (2) that is differentiable is generally used as a function. f (Xi) = 1 / {1 + exp (−Xi)} (2) Output Yi = f (Xi) (3) Here, the value of Yi is a number between 0 and 1. On the other hand, the units in the input layer use the input value as it is as the output value.

In FIG. 9, the input layers are a total of 26 units of temperature 1st, temperature 2nd, temperature 3rd, weather division (3 units), month division (12 units), and day division (8 units). The intermediate layer has 9 units. The output layer is one unit of actual water demand. It should be noted that the input numerical value and the output numerical value are normalized by the set maximum value, the set minimum value, and the like so as to fall within a range between 0 and 1. In FIG. 9, there are three units in the case of weather, and each unit corresponds to the case of sunny, cloudy, and rainy (snow), and when the input value is sunny, 1, 0, 0, respectively.
In the case of cloudy, 0, 1, and 0 are input, and in the case of rain (snow), 0, 0, and 1 are input. In the case of the month and the day of the week, processing is performed in the same manner as the weather. When actually using the output numerical value, use the reverse calculation.

[0027]

As described above, according to the water distribution predicting apparatus of the present invention, the prediction line of the temperature and the water demand for the month, the day of the week, and the weather division of the learning result of the neural network constituting the water distribution predicting apparatus is obtained. Since the data is displayed and the missing data of each section is calculated and added by multiple regression analysis and learned, the water distribution prediction device is realized with extremely high accuracy, and is extremely effective in practical use. .

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a system according to an embodiment of the present invention.

FIG. 2 is a diagram showing whether or not a performance data item is used in learning data according to an embodiment of the present invention.

FIG. 3 is a diagram showing a condition table according to an embodiment of the present invention.

FIG. 4 is a diagram showing a section value table according to an embodiment of the present invention.

FIG. 5 is a diagram showing a table of the number of days of a month section versus a weather section according to an embodiment of the present invention.

FIG. 6 is a diagram showing a correspondence table of display symbols and colors for divisions according to an embodiment of the present invention.

FIG. 7 is a diagram showing prediction lines and learning data for weather sections according to one embodiment of the present invention.

FIG. 8 is a conceptual diagram showing a multiple regression analysis according to one embodiment of the present invention.

FIG. 9 is a conceptual diagram showing a neural network according to an embodiment of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 Learning data control part 2 Neural network learning control part 3 Neural network (coupling coefficient) 4 Neural network prediction control part 5 Working memory (condition table, section value table) 6 Input control part 7 Output control part 8 Result data 9 Learning data 10 Multiple regression analysis control unit 11 Input device 12 Output / display device 13 Water distribution prediction device

Claims (2)

[Claims] 1. During and / or after learning of a neural network used for water demand forecasting, water demand and temperature of learning data including items such as water demand, temperature, date, day of the week, weather, etc. The output unit values for each value of the input unit of the neural network, in which the items in the learning data such as date, day of the week, weather, etc. are expressed by color or line type, that is, the graph of the correspondence relation of A water demand forecasting method that displays a forecast line showing the relationship of 2. The learning data of the neural network used for the prediction of the water demand is divided into a weather section, a month section from January to December, a month section from January to December, and a sun section from Sunday to Saturday. It has day of the week divisions, tabulates and analyzes in each section, and generates missing data with no data in the section by multiple regression analysis etc. according to the number of days table for each section and the section value table for each section, and adds it to the learning data And a water distribution demand forecasting method for learning a coupling coefficient of a neural network.