JPS6053904B2 - Demand forecasting method - Google Patents

Description

[Detailed Description of the Invention] The present invention is applicable to water supply demand, electricity demand, city gas demand,
This relates to a prediction method for predicting temporal changes in the daily amount of wastewater, such as the amount of wastewater flowing into the sewer system.

It is known that the demand for water supply, electricity, and city gas, as well as the amount of sewage flowing into sewers, fluctuate in a close relationship with the level of activity of human life and socio-economic activities.

To give an example, the phenomenon of increased demand for hot rice fields, tap water and electricity in the summer is a reflection of human lifestyles.
The decline in demand for water and electricity in urban centers reflects the slowdown in socio-economic activity.

Hereinafter, for the sake of simplicity, the explanation will be given using water demand for water supply as an example.

It is generally known that demand fluctuations within a day have two peaks, as shown in the typical demand curve in FIG.

However, the time at which the peak occurs, the size of the peak, the relative size of the morning peak and the afternoon peak, etc., vary widely.

Figure 2 is a superimposition of the demand curves for Mondays arbitrarily selected from each month from January to 12J1, and it can be seen that the demand curves change very rapidly.

As mentioned above, these demand fluctuations are based on various changes in human life and social and economic activities due to factors such as season, day of the week, time of day, weather, and temperature.

Therefore, in order to operate equipment economically and deal with fluctuations in demand without causing problems, prediction technology that uses factors with clear measurement basis as indicators is required. Conventional forecasting methods have been based on capturing daily demand changes as patterns.

In other words, you can prepare many patterns in advance, such as patterns by day of the week, patterns by weather, and patterns by season, and combine them, or prepare many patterns that are combined from the beginning, and then select one of them. How to select one, etc.

Although this method is easy to understand at first glance,
The problem is that all activities between two bridges are evaluated based on the same factors. For example, the amount of demand in the morning is strongly influenced by the weather in the morning, and the amount of demand in the afternoon is affected by both the morning and afternoon weather, so weather changes should also be patterned, but the types of patterns There is a limit because the number of data increases and the amount of original data per pattern to create one pattern decreases, resulting in inaccuracies as statistical data.

There is also a method of correcting past patterns using exponential smoothing to create a predicted pattern, but this is not preferable because the mathematical basis is unclear and there is always a delay.

Therefore, it has been difficult to improve prediction accuracy with such conventional methods.

The present invention was developed in view of the above circumstances, and aims to improve prediction accuracy by dividing a day into several time periods, creating the most appropriate demand forecast model for each time period, and using this model. The purpose of this study is to provide a method for predicting demand.

An embodiment of the present invention will be described below with reference to FIGS. 3 and 4.

Since the contents of human life and social activities, which are the sources of demand, differ over time, it is natural that the demand structure differs depending on the time period, and the method used in the present invention is a very natural way of thinking.

Hereinafter, the procedure for identifying a demand forecast model used in the present invention will be explained in detail. It is assumed that demand data for the past N days is measured by time zone.

This time period will be defined as one hour, and hereinafter this will be referred to as the hourly demand amount. In order to extract the characteristics of time demand,
Find the average value of the integrated value of the time demand amount for one day, and divide the time demand amount by this. This value is called the standardized time demand and is expressed by the first equation.

For convenience of explanation, the divisor was taken as the average value, but of course,
Any value obtained by multiplying the integrated value for one day by a coefficient including 1.0 may be used. 1-1hF No. 1A
Aν υν ツ A1ノHere, ZiJ
is the JE3l time demand, ZIJ is the standardized JEli
is the time demand at . By doing the above operations,
A demand distribution in which the average value of the hourly demand amount is RlJ can be obtained by selecting any one field.

When this is applied to all the data for N days, the average value of the time demand for all days becomes Rl. A daily demand data group of J is obtained. Next, the time is fixed, and the standardized hourly demand data at that time is collected for N days to form a group of hourly demand data. FIG. 3 is a diagram showing the relationship between the daily demand data group and the timely demand data group.

The data on curves 1, 2, 3, and 4 in the figure are 7
This is a group of daily demand data for E1, 2nd day, and Nth day.

Further, points 11, 12, 13, and 14 in the figure are a group of time-specific demand data at time i.

In the past, we were so focused on the temporal changes in demand that we mainly considered the time series as daily demand data, making it difficult to predict complex and subtly changing demand. This is as stated above.

In the present invention, as shown in FIG. 3, we focus on a group of demand data for each fixed time and identify the optimal model for that time.

Model identification methods include autoregressive models and multiple regression models, but GMDH (GrOupMethOdOfDa
The taHandllng) method is superior. That is, this G
The MDH method is constructed by repeating this procedure with a procedure called one layer as one unit.

What is done in this unit is to use K types and N sets of input variables, and generate KC2 types and N sets of variables called intermediate variables.
Among these, K types and N sets of intermediate variables may be selected.
At this time, the intermediate variable is selected so that it is closest to the desired variable (objective variable) in a least squares sense, so if you set an appropriate stopping criterion and stop this repeated operation, the intermediate variable will become the objective variable. The identification value of However, in the first layer, the input variable is the original human power variable, and from the next layer on, the generated and selected intermediate variable is used as the input variable. Once the predictive model structure is determined in this way, predictions can be made using that model.

More specifically, the basic GMDH method begins by dividing the input variables into training data and checking data.

Also, now the input variables are x1(t), X2(t),...
The expression KOlnlOgOrOv-GabOr is assumed, and a combination of two variables is considered.

That is, based on the above preparation, apply the least squares method to the training data for a certain combination of (1, j), find the coefficient of Mo~tun, and use the coefficient and checking data to create the model. Examine the quality using the following criteria. That is, here, NlCi'5-Etsu♀g data number, Yl, tree (
t) is the checking data as variables xl(t)Xj(t)
I asked for it first.

The value obtained from equation (4) using the coefficient of ~A5, z(
t) is an actual measured value. These operations are performed using 2 variables from K types.
Combinations KC2 for selecting types of variables are executed for all the combinations, K types of combinations are selected from those with the smallest value of equation (B), and all intermediate variables of K types and N sets are generated. In this way, inputting K types and N sets of data and outputting K types and N sets of intermediate variables is repeated as one single position layer. ) does not improve, ie, reduce, that of the previous layer, this iteration is stopped.

The above is a method called basic GMDH, and it is possible to obtain statistical models of events with unknown correlations, such as social and economic phenomena, water distribution measurements, etc., using relatively simple algorithms. A model identification method that uses this and further improves the degree of identification is disclosed in, for example, Japanese Patent Application No. 54-32953 (Japanese Patent Application Laid-open No. 55-32953).
124855) Water demand forecasting method, Patent No. 1230
No. 161, No. 1, Water Demand Forecasting Method, is detailed, and the methods shown therein may be used when implementing the method.

In this way, the demand forecast model (F1(X),...,F
24(X)).

Input variable vector X (Xl
,...,Xe) are as shown in the table below. The weather is digitized as 3 if it is sunny, 2 if it is cloudy or snowy, and 1 if it is rainy.

The snowfall influence coefficient is defined by the following formula assuming the amount of remaining snow.
However, in the UK, the amount of snowfall i days before the forecast date, and T is the snow melting time constant, which varies depending on the region, but usually about 3 to 5 is suitable.

If the time period is divided into hourly units, the number of models is 2.
The number of models can be reduced to 4, but the number of models can be reduced to 2 to 3 hours during times when demand is relatively stable, such as at night.

Now, the model F7 (.) that measures the hourly demand from 6:00 to 7:00 is based on the obtained demand forecast model, for example.

The coefficient A5 is zero. Similarly, from 8 o'clock to 9 o'clock 1
The model F9 (.) that predicts the time demand is as follows.

The former prediction model is an example in which a one-layer configuration is optimal, and the latter prediction model is an example in which a two-layer configuration is optimal. Next, the procedure for predicting demand using the model identified in this way will be described.

(1) Assign values to the input variables Xl, . . . , X9 of the prediction model.

Xl, X2, X, , X7 are based on the weather forecast, XG,
6 gives each numerical value based on the observed value of the previous day. j is known because it usually occurs at dawn. X7 is obtained from the previous day's observed value and Z. X, is determined from the observed values for the past four days. (2) Identified predictive model (F1(X),
・, F24(X)) predicted value of time demand standardized? Find (1=1,...,24).

That is, F, (X), (1=1, . . . , 24) is a function of the demand forecasting model.

The above equation (3) is when i=7, and the equation (4) is also when i=7.
This is the case of 9. (3) From the predicted value of the standardized hourly demand amount and the daily total demand quantum value D, calculate the hourly demand quantum value estimate using the following equation.

1 day = total blindness i Jingtan: j Example: N, patent Il23Ol
It can be determined by the method detailed in No. 6l No. 1 Water Demand Forecasting Method.

FIG. 4 shows an embodiment of the demand forecasting device according to the present invention.
This aims to optimize the flow rate of water sent to the distribution reservoir.

In the figure, numeral 21 is a time period demand input device for inputting time period demand (time demand) data, and 22 is an output of the time period demand input device 21 which is subjected to the calculation of the first equation. There is a time demand standardization device that standardizes the time demand.

Reference numeral 23 denotes a prediction model identification device, which inputs the standardized hourly demand output from the hourly demand standardization device 22, converts it into a time-based demand data group, and converts it into a group of time-specific demand data. A prediction model is identified using a method such as GMDH.

24 is a time demand quantum measuring device, and this time demand quantum measuring device 24 inputs input information 25 for prediction (for example, weather, temperature, rainfall, etc.) to the model identified by the predictive model identification device 23. The predicted value of the input and standardized time demand is calculated.

Reference numeral 26 denotes a conversion device, and this conversion device 26 calculates the time demand quantum value from the predicted value and the output of the total demand quantum measurement device 21 which calculates the total demand quantum value using the second equation.

28 is a water transmission system optimal operation device, and this water transmission system optimal operation device 28 operates the water treatment plant 2 based on the above-mentioned hourly demand quantum measurement value.
Optimization of water flow rate for the purpose of smoothing the operation of water supply system 9, reducing the number of starts and stops of the water pump 31 that sends water from the water purification plant 29 to the water distribution reservoir 30, and maintaining the water level of the plurality of water distribution reservoirs 30 within the upper and lower limit ranges. It is a device that aims to

A concrete implementation method is, for example, based on predicted water demand and pond water level, as detailed in Japanese Patent Application No. 50-33593 (Japanese Unexamined Patent Publication No. 51-109145) r. The inflow rate is determined to keep the pond water level at a medium level, and the pump or electric valve is controlled to maintain the determined flow rate, or the predicted water demand fluctuation curve is determined and the adjacent two
A method such as setting a change time at approximately the midpoint of the two extreme values and giving a flow rate adjustment change command to the pump or motor-operated valve may be used. 32 is a water pipe connecting the water purification plant 29 and the water distribution reservoir 30; 33 is an electric valve for adjusting the amount of water flowing under gravity; controlled by

34 is a water distribution pipe for distributing water from the water distribution reservoir 30 to the consumers 35; 36 is a flow meter; this flow meter 36 is connected to the water treatment plant 2;
9 and the water distribution reservoir 30, which measures the outflow flow rate, that is, the demand amount.

Reference numeral 37 denotes a flow rate integrating device, which obtains the measured output of the flow meter 36 installed in the water distribution pipe 34, integrates it by time zone, and accumulates past cumulative data, and calculates the demand by time zone. Output to the amount input device 21.

This device with such a configuration measures the flow rate of water distributed from the water distribution reservoir 30 to the consumers 35 with the flow meter 36 provided in the water distribution pipe 34, and provides the measured value to the flow rate integration device 37.

As a result, the flow rate integration device 37 integrates the amount of water demanded for each time period and accumulates past integration data. These integrated data are given to the time period demand input device 21, and based on these inputs, the time period demand amount input device 21 generates the time period demand data corresponding to each time period and the past N Output daily demand data. This demand data by time zone and for the past N days is collected by the hourly demand standardization device 2.
2, whereby the time demand standardization device 22
Based on this input data, the above-mentioned first equation is calculated to calculate and output the standardized time demand amount for that time period. The output of this time demand standardization device 22 is given to a prediction model identification device 23, where a statistical prediction model for demand quantum measurement is identified. Using the input information 25 for prediction and the identified prediction model, the time demand quantum measuring device 24 calculates the predicted value of the standardized time demand amount.

From the output of the total demand forecasting device 27 for calculating the total demand quantum measurement value and the output of the time demand quantum measurement device 24, the conversion device 26 calculates the time demand quantum measurement value according to the second equation. The time demand quantum measurement value obtained in this way is given to the water transmission system optimal operation device 28, and based on this, the water transmission system optimal operation device 28 smooths the operation of the water purification plant 29 and starts and stops the water transmission pump 31. The optimal value of the water supply flow rate necessary to reduce the number of times and maintain the water level of the plurality of water distribution reservoirs 30 within the upper and lower limit ranges is determined, and a control output is generated to control the water supply pump 31 and the electric valve 33. This allows for optimal water distribution control that corresponds to fluctuations in demand depending on the time of day. In this way, we obtain the hourly demand amount, which is the demand amount for each time period, and calculate 1 of this hourly demand amount.
Calculate the average value of the daily integrated values, calculate the ratio of this to the hourly demand amount, obtain the standardized hourly demand amount for each time period, and further calculate the standardized hourly demand amount data for each time of day N Obtain a group of hourly demand data for each day, identify the optimal model for demand at each hour, and standardize the identified model by providing input information such as weather and temperature necessary for forecasting. The predicted value of hourly demand was obtained based on the predicted value of hourly demand, and the predicted value of hourly demand was obtained from this and the predicted value of total demand. We set up an appropriate demand prediction model for each time period that matches trends in human life and socio-economic activities, add information such as natural conditions to this model, and predict standardized hourly demand. Furthermore, since the hourly demand amount is obtained by correcting it according to the total demand amount, it is possible to perform highly accurate demand forecasts for each time period, thereby facilitating the smooth supply of electricity, water, city gas, etc. It is possible to provide a demand forecasting method that has excellent features such as being able to perform operations in a timely and appropriate manner.

It should be noted that the present invention is not limited to the embodiments described above and shown in the drawings, but can be implemented with appropriate modifications within the scope without changing the gist thereof.

[Brief explanation of drawings]

Figure 1 is a diagram showing fluctuations in demand in one day, Figure 2 is a diagram showing an example of fluctuations in demand in one day, and Figure 3 is the relationship between daily demand data group and timely demand data group. FIG. 4 is a block diagram showing the configuration when the present invention is applied to water supply distribution equipment. 21... Time-slot demand input device, 22... Hourly demand standardization device, 23... Prediction model identification device, 24
... Time demand quantum measurement device, 26 ... Conversion device, 27
... Total demand forecasting device, 28 ... Water transmission system optimum operation device, 30 ... Water distribution reservoir, 31 ... Water transmission pump, 33.
...Electric valve, 36...Flow meter, 37...Flow rate integrating device.

Claims (1)

[Claims] 1. In a method for predicting demand changes in events with unknown correlations such as demand for water supply, electricity, city gas, etc. or the amount of sewage flowing into sewers, etc., obtain the measured output of the target demand change and calculate it by time period. In addition, the past demand time series for each time period is obtained, and the past demand time series for each time period is converted into the total daily demand amount for the same day. Standardize by the value multiplied by a coefficient, obtain a standardized hourly demand prediction model using this standardized hourly demand time series and observed values of weather information, and input the weather information at each prediction timing. A demand forecasting method characterized in that a predicted value of hourly demand is determined from a standardized predicted value of hourly demand obtained by forecasting hourly demand and the predicted total demand.