JPH11126102A - Predictive method for demand of gas - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a predictive method for demand of gas, by which the construction of a predictive model and the maintenance of its parameter can be easily performed and the demand of gas can be highly accurately predicted even on a gas supply line of a large scale. SOLUTION: As the input of a hierarchical neural circuit, a data set for learning for a prescribed period in the past is prepared (process 4) by using actual demand pattern a prescribed time after a day (n) while taking the demand pattern result for a prescribed period before the day (n), the atmospheric temperature pattern result for a prescribed period before the day (n), the actual atmospheric temperature pattern a prescribed time after the day (n) and the day-of-the-week flag of the day (n) as teacher data. In this case, the day (n) expresses a day in the past. Next, the learning data set is inputted to a hierarchical neural circuit network 8 and learning is performed (process 5). Afterwards, the demand pattern result for the prescribed period before a predictive day, the atmospheric temperature pattern result for the prescribed period before the predictive day, the predictive atmospheric temperature pattern from the predictive day to the future for the prescribed time and the day-of-the-week flag of the predictive day are inputted to the learnt hierarchical neural circuit network 8 as input data and the demand from the predictive day to the future for the prescribed time is predicted (processes 6 and 7).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for estimating a gas demand of a gas supply line, which fluctuates depending on weather conditions, using a neural network.

[0002]

2. Description of the Related Art In the operation of a gas supply line, it is necessary to predict the daily gas demand in order to maintain the supply reliability and improve the operation efficiency. In particular, due to the recent increase in natural gas demand and the difficulty of constructing new gas supply lines, the demand for efficient operation of gas supply lines is increasing year by year, and more accurate forecasts of total gas demand are required. I have. Forecasting the current total gas demand is based on forecasts based on the experience and intuition of skilled operators, and on prediction models based on statistical analysis methods such as multiple regression analysis and time series analysis, which correlate past demand results with temperature, day of the week, etc. There is a method in which a computer is prepared, the predicted temperature of the prediction target day is input to the data, and a computer calculates the temperature. Further, in fields such as power peak demand and stock prices, forecasts using neural networks (neural networks) have been made.

[0003]

All of the conventional prediction methods described above predict the daily demand or the peak demand for each day to be predicted, and are applicable when the gas supply line is relatively small. It is an effective method. However, when applied to a long-distance or network-like gas supply line having a plurality of demand ends, there are the following problems.

[0004] Since gas is a compressible fluid, the effect of pressure fluctuations in the gas supply line due to a change in demand at the demand end of the gas supply line is transmitted to the supply end of the gas supply line with a time delay. The degree of the time delay in which the influence of the fluctuation in the demand amount at the demand end is transmitted to the supply end increases as the gas transport line becomes larger due to factors such as longer distance and networking.

In general, the demand end of a long-distance gas transport line is often a gas company, but if the demand end is a gas company, the gas company owns the gas company in order to secure a stable supply. Since the gas is temporarily stored in the gas holder and then supplied to the terminal consumer, there is a time delay between the consumption of the terminal consumer and the amount withdrawn from the gas transportation line. As described above, when the influence of the gas demand fluctuation is transmitted to the supply end with a large time delay, the supply amount at a certain time reflects the influence of the past demand fluctuation at the demand end. Therefore, in order to determine the supply amount at a certain time, it is necessary to take into account the past demand end fluctuations in demand and to take in past temperature information having a strong correlation with the past demand end fluctuations.
However, since the conventional forecast method does not incorporate past temperature information that has a strong correlation with the past demand-end demand fluctuation and the past demand-end demand fluctuation described above,
In a large-scale gas supply line in which the influence of the gas demand change at the demand end is transmitted to the supply end with a large time delay, there is a problem that it is not possible to appropriately predict the demand amount.

[0006] In addition, statistical analysis methods such as multiple regression analysis and time series analysis which have been conventionally used as a prediction method include:
In order to create a forecast model, it is necessary to be familiar with both the demand fluctuation characteristics and the statistical analysis method of the gas supply line to be forecasted. Therefore, it is necessary to maintain the parameters of the prediction model, and the operation is complicated.

In recent years, prediction using a neural network, which has begun to be used for peak demand forecasting of electric power, has been complicated in handling, such as preparing a prediction model for each season.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and does not require specialized knowledge such as a demand fluctuation characteristic of a gas supply line to be predicted and a statistical analysis method.
It is an object of the present invention to provide a gas demand forecasting method capable of easily constructing a forecast model and maintaining its parameters, and enabling highly accurate forecasting of a gas demand even in a large-scale gas supply line. Aim.

[0009]

According to the present invention, there is provided a method for estimating a gas demand according to the present invention, wherein a demand pattern result from n days to a predetermined time in the past, and a temperature pattern result from n days to a predetermined time in the past. , Using the actual temperature pattern from the nth day to the predetermined time ahead and the day of the week flag on the nth day, using the actual demand pattern from the nth day to the predetermined time ahead as teacher data,
A learning data set for a predetermined period in the past is created. Here, n days represent past days. Next, learning is performed by inputting the learning data set to the hierarchical neural network. For this learning, a back propagation learning algorithm is used. Thereafter, as input data to the learned hierarchical neural network, the demand pattern results from the predicted date to a predetermined time in the past, the temperature patterns from the predicted date to a predetermined time in the past, the predicted temperature patterns from the predicted date to a predetermined time in the future, and The day of the week flag is input for the forecast date, and the demand amount from the forecast date to a predetermined time ahead is predicted.

In another method for predicting gas demand according to the present invention (claim 2), the above-mentioned prediction method (claim 1)
In, the start time and the end time of the demand prediction pattern obtained as the output of the neural network are fixed, and a part of the expected temperature pattern input to the learned neural network is changed from the demand prediction pattern start time to the expected execution time. By replacing with the actual temperature pattern obtained during the period, the effect of the prediction error included in the predicted temperature is reduced, and the demand prediction accuracy is improved.

In another gas demand forecasting method according to the present invention (claim 3), a smoothed demand pattern and a smoothed temperature pattern are calculated in advance from the past demand and the actual temperature prior to the creation of the learning data set. create. Next, the difference between the demand pattern actual and the smoothed demand pattern from the nth day to a predetermined time in the past, the difference between the actual temperature pattern and the smoothed air temperature pattern from the nth to the predetermined time in the past, The difference between the actual temperature pattern and the smoothed air temperature pattern and the day of the week flag on the nth day are used as learning data. Create a dataset. Here, n days represent past days. Next, the learning data set is input to the hierarchical neural network to perform learning. For this learning, a back propagation learning algorithm is used. Then, as input data to the learned hierarchical neural network, the difference between the demand pattern actual and the smoothed demand pattern from the predicted date to a predetermined time in the past, the temperature pattern actual and the smoothed temperature pattern from the predicted date to the predetermined time in the past , The difference between the predicted temperature pattern and the smoothed temperature pattern from the predicted date to a predetermined time ahead and the day of the week flag of the predicted date are input, and the predicted demand amount and the smoothed demand pattern from the predicted date to a predetermined time ahead are To predict the difference between next,
The smoothed demand pattern is added to the difference between the forecast demand amount and the smoothed demand pattern from the forecast date to a predetermined time ahead to obtain the forecast demand amount from the forecast date to a predetermined time ahead.

In another method for predicting gas demand according to the present invention (claim 4), the above-mentioned prediction method (claim 3)
In, the start time and the end time of the difference between the demand forecast pattern obtained as the output of the neural network and the smoothed demand pattern are fixed, and the predicted temperature pattern and the smoothed temperature pattern to be input to the learned neural network are fixed. It is included in the expected temperature by replacing a part of the difference with the difference between the actual temperature pattern and the smoothed temperature pattern obtained from the start time of the difference between the demand forecast pattern and the smoothed demand pattern to the expected execution time. Reduce the effect of forecast errors and improve demand forecast accuracy.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION

(Embodiment 1) FIG. 1 is a block diagram showing a configuration of a system according to Embodiment 1 of the present invention. This system is
Actual demand database 1, Actual temperature database 2, Forecast temperature database 3, Learning data set creation process 4,
It comprises a learning step 5, a prediction data creation step 6, a prediction step 7, and a neural network 8.

First, an outline of the system shown in FIG. 1 will be described.
The actual demand database 1 is automatically loaded with hourly actual values for each customer along the gas supply line. The actual temperature database 2 automatically captures hourly actual temperature values at points having a strong correlation with customers along the gas supply line. The predicted temperature database 3 automatically captures the predicted temperature at the same point as the actual temperature database 2 and interpolates the hourly data if necessary. In the learning data set creation step 4, input data and teacher data necessary for learning of the neural network 8 are extracted from the actual demand database 1 and the actual temperature database 2 to create a learning data set for a predetermined period. In the learning step 5, the data set created in the learning data set creation step 4 is given to the neural network 8, and learning is performed so that the output of the neural network 8 and the teacher data match well. This learning uses a back propagation method.

In the prediction data creating step 6, input data necessary for the prediction of the neural network 8 is extracted from the actual demand database 1, the actual temperature database 2, and the expected temperature database 3 to create the prediction data. In the forecasting step 7, the forecasting data created in the forecasting data creating step 6 is given to the neural network 8, and a demand pattern from the day of the forecast to a predetermined time ahead is calculated.

FIG. 2 is a diagram showing an example of input / output data of the neural network 8 of FIG. This embodiment will be described more specifically with reference to FIG. The neural network 8 shown in FIG. 2 is a three-layer ramelhart-type neural network. The actual demand pattern for a predetermined period before the nth day is a 24 hour actual demand for one hour interval on the day before the nth day. n of points that have a strong correlation with demand as the actual temperature pattern for the given period before n days
The actual temperature for 24 hours at one hour intervals of the previous day, the expected temperature pattern for 48 hours at one hour intervals after n days, a day of the week flag indicating a holiday, etc., and the output layer has n Forecast demand patterns for 48 hours at one hour intervals after the day are output. Note that it is necessary to select an appropriate period for the period of the actual demand pattern and the actual temperature pattern before the nth day in consideration of the degree of a time delay in which the influence of the demand fluctuation at the demand end is transmitted to the supply end. . Further, the periods of the predicted temperature pattern and the predicted demand pattern are determined from the period in which the predicted value is required.

In the neural network 8 shown in FIG. 2, the values entered from the input layer neurons are weighted and passed to the intermediate layer neurons. Then, it is further weighted and passed to the neuron of the output layer to become an output value. Therefore,
By providing data to the input and output layers and adjusting the weights so that they match, the correlation between the data in the input and output layers can be captured in the weights.

Next, a specific prediction procedure will be described. (A) In the learning data set creation step 4, input layer data and output layer data for the past several tens of days that satisfy the time relationship between the input and output layers in FIG. A learning data set is created by extracting the input layer data and the output layer data as one set. (B) In the learning step 5, of the learning data sets for the past several tens of days created in the learning data set creation step 4, the input layer data is given to the input layer of the neural network 8, and the output layer data is sent to the teacher signal. To the output layer of the neural network 8 to determine the weight of each neuron. (C) In the prediction data creation step 6, data necessary for the input layer of the neural network 8 shown in FIG. 2 is extracted from the actual demand database 1, the actual temperature database 2, and the expected temperature database 3. (D) In the prediction step 7, the prediction data created in the prediction data creation step 6 is given to the input layer of the neural network 8 to obtain a forecast demand pattern.

(Embodiment 2) FIG. 3 is a diagram showing an example of input / output data of a neural network 8 according to Embodiment 2 of the present invention. The configuration of the neural network and up to the learning step 5 are the same as in the first embodiment. Of the input data of the neural network 8 shown in FIG. 3, the predicted temperature pattern is stored in the actual temperature database 2 with the lapse of time from 0:00 on the current day, which is the predicted start time. Temperature can be replaced with actual temperature. In the present embodiment, by replacing the predicted temperature that can be replaced with the actual temperature with the actual temperature, error factors included in the expected temperature can be reduced, and the prediction accuracy of the demand pattern can be improved.
Note that the specific prediction procedure is the same as in the first embodiment.

(Embodiment 3) FIG. 4 is a diagram showing an example of input / output data of a neural network 8 according to Embodiment 3 of the present invention. The neural network 8 in FIG. 4 is a three-layer ramelhart-type neural network. The input layer has a difference between the actual demand pattern of the predetermined period before the nth day and the smoothed demand pattern at one hour intervals of the day before the nth day. The difference between the actual demand for 24 hours and the smoothed demand at each time, the difference between the actual temperature pattern and the smoothed temperature pattern for the predetermined period before n days, and 1 day before the point n days before the point having a strong correlation with the demand. The difference between the actual temperature for 24 hours of the time interval and the smoothed air temperature at each time, and the difference between the expected air temperature pattern and the smoothed air temperature pattern, which is 4 for one hour interval after n days
It has a day of the week flag indicating the difference between the predicted temperature for 8 hours and the smoothed temperature at each time, a holiday, and the like. The difference from the smoothing demand is output. Thereafter, the smoothed demand at each time is added to the obtained difference between the forecast demand for 48 hours at the hour interval after the nth day and the smoothed demand at each time, and 48 hours at the hourly interval after the nth day. Get forecast demand.

In the period of the actual demand pattern and the actual temperature pattern before the nth day, it is necessary to select an appropriate period in consideration of the degree of time delay in which the influence of the demand fluctuation at the demand end is transmitted to the supply end. There is. Further, the periods of the predicted temperature pattern and the predicted demand pattern are determined from the period in which the predicted value is required. Simple average and exponential smoothing methods are used for smoothing temperature patterns to reduce the effects of temperature changes due to seasonal changes and for smoothing demand patterns to clarify the relationship between temperature fluctuations and demand fluctuations at each time. Used. In order to consider the influence of short-term temperature fluctuations, the smoothing degree is usually selected to be an average value of one week to one month. Note that, also in the present embodiment, the specific prediction procedure is the same as in the above-described first embodiment.

(Embodiment 4) FIG. 5 is a diagram showing an example of input / output data of a neural network 8 according to Embodiment 4 of the present invention. The configuration of the neural network and the learning step 5 are the same as those of the third embodiment. Among the input data of the neural network 8 shown in FIG. 5, the difference between the predicted temperature pattern and the smoothed temperature pattern indicates that the actual temperature is changed with the lapse of time from 0:00, which is the expected start time, on the actual temperature database 2. , The estimated temperature can be replaced with the actual temperature. In the present embodiment, by replacing the expected temperature that can be replaced with the actual temperature with the actual temperature, the error factors included in the expected temperature can be reduced, and the prediction accuracy of the demand pattern can be improved.

[0023]

EXAMPLES Table 1 is a table showing a comparison of prediction errors between the present invention and the conventional method. The prediction error used was an absolute average error (MAPE) represented by the following equation.

[0024]

[Table 1]

[0025]

(Equation 1)

Here, Qd ₀ is the sum of the demand values for 24 hours, that is, the actual value of the daily demand, Qd _p is the sum of the demand values for 24 hours, ie, the predicted value of the daily demand, n is the number of data.

As shown in Table 1, according to the present invention, it can be seen that the prediction error is reduced while a larger amount of information is predicted than in the prior art.

FIG. 6 is a diagram showing an example of a demand prediction result according to the present invention (Embodiment 4). In FIG.
An example of prediction results for 43 hours at one-hour intervals from the hour to 0:00 of the next day is shown.

FIG. 7 is a characteristic diagram showing the effect of taking in the actual temperature in the second and fourth embodiments. In the present embodiment, in the prediction example of 43 hours at an interval of 1 hour from 6:00 on the current day to 0:00 on the next day shown in FIG.
The absolute average error distribution (MAPE) is obtained for the prediction results for 24 hours up to the hour. In the figure, no actual temperature acquisition is a prediction error at 6:00 on the day (corresponding to the first and third embodiments), and actual temperature intake is from 0:00 on the next day, that is, from 6:00 on the day to 0:00 on the next day. This is the case where the actual temperature is taken in (corresponding to the second and fourth embodiments). As can be seen from the figure, according to the present invention, by taking in the actual temperature, the error factors included in the expected temperature can be reduced, and the prediction error can be greatly reduced.

[0030]

As described above, according to the present invention, in a demand forecasting method for predicting gas demand of a gas supply line fluctuating according to weather conditions using a neural network, an output of the neural network and a demand pattern actual value are provided. Learning is performed so as to match the demand pattern, and the demand pattern from the prediction date to a predetermined time ahead is predicted using the neural network after the learning is completed. It can be easily constructed without the need for specialized knowledge such as fluctuation characteristics and statistical analysis methods. In addition, the parameters of the neural network are automatically maintained by the above learning, so that maintenance can be easily performed. It can be carried out.
In addition, because it captures past temperature information that has a strong correlation with past demand end fluctuations and past demand end fluctuations, the effect of gas demand fluctuations at the demand end causes a large time delay. Accordingly, it is possible to perform appropriate gas demand prediction in a line for a large-scale gas supply transmitted to a supply end.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a system to which a gas demand prediction method according to a first embodiment of the present invention is applied.

FIG. 2 is a diagram showing an example of input / output data of the neural network of FIG. 1;

FIG. 3 is a diagram illustrating an example of input / output data of a neural network according to a second embodiment of the present invention.

FIG. 4 is a diagram illustrating an example of input / output data of a neural network according to a third embodiment of the present invention.

FIG. 5 is a diagram illustrating an example of input / output data of a neural network according to a fourth embodiment of the present invention.

FIG. 6 is a diagram showing a demand prediction result.

FIG. 7 is a diagram showing the effect of taking in actual temperature.

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁶ Identification symbol FI // F17D 3/18 G06F 15/20 F

Claims (4)

[Claims] 1. A demand forecasting method for forecasting gas demand of a gas supply line using a neural network, wherein input data of input / output data input / output to / from the neural network is a predetermined time from a forecast date to a predetermined time in the past. The demand pattern results, the temperature pattern results from the forecast day to a predetermined time in the past, the forecast temperature patterns from the forecast day to the predetermined time ahead, and the day of the week flag are used as output data, and the demand pattern from the day to the predetermined time ahead is used as output data. The learning value is obtained by using the actual value of the input / output data during the learning as learning data so that the output of the neural network and the actual value of the demand pattern from the current day to a predetermined time ahead match. Demand pattern results from the forecast day to the past for a predetermined time, temperature pattern results from the forecast day to a predetermined time in the past, and a predetermined time from the day A demand pattern from a predicted date to a predetermined time ahead by inputting a predicted temperature pattern up to and a day of the week flag. 2. A demand forecast pattern obtained as an output of the neural network is fixed with a start time and an end time, and a part of the forecast temperature pattern of the day, which is input data of the neural network, is replaced with the demand forecast pattern. 2. The method for predicting a gas demand according to claim 1, wherein the actual temperature pattern obtained from the start time to the predicted execution time is replaced. 3. A demand forecasting method for predicting gas demand of a gas supply line using a neural network, wherein a smoothed demand pattern and a smoothed temperature pattern within a predetermined period are created in advance, and the neural network is provided with the smoothed demand pattern. As input / output data to be input and output, the difference between the demand pattern actual from the forecast day to a predetermined time in the past and the smoothed demand pattern, the difference between the temperature pattern actual from the forecast day to a predetermined time in the past and the smoothed temperature pattern, The difference between the expected temperature pattern and the smoothed temperature pattern from the day to a predetermined time ahead and the day of the week flag,
Using the difference between the demand pattern from the current day to a predetermined time ahead and the smoothed value of the demand pattern as output data, using the actual value of the input / output data for the past several days as learning data as learning data, the output of the neural network and The learning is performed so that the difference between the demand pattern actual value from the day and a predetermined time ahead and the difference between the smoothed demand pattern coincides with each other. The difference between the pattern result and the smoothed demand pattern, the difference between the temperature pattern result and the smoothed temperature pattern from the predicted day to a predetermined time in the past, and the predicted temperature pattern and the smoothed temperature pattern from the day to the predetermined time ahead. After inputting the difference between the forecast date and the day of the week flag and predicting the difference between the demand amount from the forecast date and a predetermined time ahead and the smoothed demand pattern, the difference between Gas demand prediction method characterized by calculating the demand patterns up. 4. An expected temperature pattern of the day as input data of the neural network by fixing a start time and an end time of a difference between a demand forecast pattern obtained as an output of the neural network and the smoothed demand pattern. And a part of the difference between the smoothed temperature pattern and the difference between the actual temperature pattern obtained from the start time of the difference between the demand forecast pattern and the smoothed demand pattern and the predicted execution time and the smoothed temperature pattern. 4. The method for predicting a gas demand according to claim 3, wherein: