JP2001216287A - Short term demand prediction device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a short term demand prediction device capable of constructing an appropriate predication model by a neural network and suppressing an overfitting of prediction model. SOLUTION: A neural network prediction model construction means 11 performs learning based on result values classified as learning data in a result value classification means 13 and constructs a prediction model. At the time, the learning is stopped by learning stoppage instruction from a learning stoppage judgment means 12 and the prediction model for which an overfitting is suppressed is constructed. A neural network demand prediction means 14 predicts the demand amount of the day by using the prediction model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to, for example, a short-term demand predicting apparatus for predicting short-term consumption of purified water in demand for water supply.

[0002]

2. Description of the Related Art For example, to accurately predict the amount of purified water consumed on the day of consumption by a customer of a water supply system, it is necessary to efficiently operate a water supply facility for supplying purified water or to provide an appropriate amount of purified water to the customer. It is important to supply without.

Since the amount of purified water consumed by consumers is considered to be affected by the weather of the day, in predicting the amount of purified water of the day, weather information of the day is input and the amount of purified water consumed is output. It is common to build a demand forecasting model.

Hitherto, in constructing such a demand prediction model, a statistical method or a method using a neural network has been proposed (JP-A-56-124855, JP-A-05-17976). In a statistical method, for example, an autoregressive model or a multiple regression model is constructed and used, and in a method using a neural network, a causal relationship between input and output is modeled and used. When building a demand forecasting model applying the latter neural network, a model is built using past demand actual data.

[0005]

However, at this time, if a good prediction model is constructed for past performance data, the prediction accuracy may be degraded at the time of prediction. This is overfitting of the prediction model due to over-learning (a phenomenon in which the predicted value outputs a value that is significantly different from the actual value), and poses a problem in actual application.

There is a demand to provide a more accurate prediction model by suppressing the overfitting and constructing a prediction model using an appropriate neural network.

[0007] It is an object of the present invention to provide a short-term demand prediction device capable of constructing an appropriate prediction model using a neural network and suppressing overfitting of the prediction model.

[0008]

According to a first aspect of the present invention, there is provided a short-term demand predicting apparatus for categorizing accumulated actual values into learning data and check data, and learning by the actual value classifying means. A neural network prediction model construction unit that learns and builds a prediction model based on the actual values classified as data, and the neural network prediction model construction unit based on the actual values classified as check data by the actual value classification unit. And a neural network demand prediction means for predicting the demand on the day using a prediction model constructed by the neural network construction means.

[0009] In the short-term demand forecasting apparatus according to the first aspect of the present invention, the neural network forecasting model constructing means includes:
Learning is performed based on the performance values classified as learning data by the performance value classifying means, and a prediction model is constructed. In this case, the learning is stopped by the learning stop from the learning stop determination means, and a prediction model in which overfitting is suppressed is constructed.
The neural network demand prediction means predicts the demand on the day using the prediction model.

According to a second aspect of the present invention, in the short-term demand forecasting apparatus according to the first aspect of the present invention, the actual value classifying means classifies the newly added actual value as check data, or learns the accumulated actual value. The method is characterized in that data and check data are alternately classified, or actual values of a certain period among accumulated actual values are classified as check data.

[0011] In the short-term demand prediction device according to the second aspect of the invention, in addition to the operation of the first aspect, the learning stop determining means includes a newly added actual value or an alternately selected actual value, or Using the actual value of a certain period as the check data, the learning stop is determined.

According to a third aspect of the present invention, in the short-term demand predicting apparatus according to the first or second aspect of the present invention, the learning stop judging means predicts a demand amount by using a prediction model constructed by the neural network prediction model constructing means. Comparing the predicted demand with the corresponding actual value,
When the error is minimized, the learning of the neural network prediction model constructing means is stopped.

[0013] In the short-term demand prediction device according to the third aspect of the present invention, in addition to the operation of the first or second aspect of the present invention, the learning stop judging means predicts by the prediction model constructed by the neural network prediction model constructing means. The learning is stopped when the error between the calculated demand amount and the corresponding actual value is minimized. Thereby, the neural network prediction model construction means suppresses overfitting and constructs a prediction model with high accuracy.

According to a fourth aspect of the present invention, there is provided the short-term demand forecasting apparatus according to any one of the first to third aspects, wherein the accumulated actual values are a weather information actual value and a demand actual value. The model is a demand forecast model for water supply.

According to a fourth aspect of the present invention, in addition to the operation of any one of the first to third aspects of the present invention, the neural network predictive model constructing means includes a learning value classifying means, A demand forecast model for water supply is constructed based on the weather information actual value and demand actual value classified as. Further, the learning stop determining means inputs the actual weather information value classified as check data by the actual value classifying means to the demand forecasting model, thereby minimizing an error between the predicted demand amount and the demand actual value of the check data. When it becomes, learning of the neural network prediction model construction means is stopped. Thereby, the neural network prediction model construction means suppresses overfitting and constructs a prediction model with high accuracy.

[0016]

Embodiments of the present invention will be described below. FIG. 1 is a configuration diagram of a short-term demand prediction device according to an embodiment of the present invention. As shown in FIG. 1, a weather forecast value a is input to a neural network demand prediction unit 14 of the short-term demand prediction device 1 by an operator. The neural network demand prediction means 14 predicts the demand on the day using the weather forecast value a set by the operator 31 using the prediction model stored as the prediction model in the learned weight coefficient storage means 23. Then, the predicted demand on the day is output to the operator as a demand predicted value b.

The demand actual value accumulating means 21 accumulates daily demand actual values (purified water consumption) c which are acquired,
In addition, the weather information actual value storage means 22 stores the weather information actual value d set by the operator. The actual values accumulated in the demand actual value accumulating means 21 and the weather information actual value accumulating means 22 are classified into learning data A and check data B by the actual value classifying means 13 as described later. The check data B is input to the network prediction model construction means 11, and the check data B is input to the learning stop determination means 12.

The neural network prediction model constructing means 11 learns by the neural network based on the weather information actual value d and the demand actual value c classified as the learning data A by the actual value classifying means 13 and constructs a prediction model. That is, the prediction model is constructed by updating the weight coefficient W of the neural network, and the updated weight coefficient W is stored in the learned weight coefficient storage unit 23 as the prediction model.

On the other hand, the learning stop judging means 12 uses the weather information actual value d and the actual demand value c classified as the check data B by the actual value classifying means 13 and the prediction model calculated by the neural network prediction model constructing means 11. Based on the demand forecast value e, when constructing a prediction model in the neural network prediction model construction means 11, it is determined to stop learning for suppressing overfitting of the prediction model. When it is determined that learning is to be stopped, a learning stop signal f is output to the neural network prediction model construction means 11.

The actual value classifying means 13 classifies the actual values stored in the demand actual value accumulating means 21 and the weather information actual value accumulating means 22 into learning data and check data as shown in FIG. FIG. 2A shows a classification in a case where newly added actual values are used as check data, and FIG. 2B shows the accumulated actual values used alternately for learning data and check data. The classification of the case is shown. FIG. 2C shows a case where, among the accumulated actual values, the actual values in a certain period are classified so as to be used as check data. The classification of the actual demand value and the actual weather information value into the learning data and the check data may be performed by other classification methods.

The demand actual value and the weather information actual value classified as the learning data are used when the neural network prediction model constructing means 11 constructs the prediction model by the neural network, and the demand actual value and the weather information classified as the check data are used. The information result value is used when learning stop is determined by the learning stop determination 12.

The learning stop determining means 12 is a neural network predictive model constructing means for suppressing overfitting of the predictive model at the time of constructing the predictive model based on the demand actual value and the weather information actual value classified as the check data. Eleven learning stops are determined. In this case, the demand amount is predicted by inputting the actual weather information value of the check data to the prediction model constructed for each learning. Then, the predicted demand amount is compared with the demand actual value of the check data, and learning is stopped at a point where the error between the predicted demand amount and the actual demand value is minimized.

As described above, the actual demand value and the actual weather information value are classified into learning data and check data, and the neural network prediction model construction 11 constructs a prediction model based on the learning data. The learning error in this case decreases as the number of times of learning increases, as shown in FIG.

On the other hand, as shown in FIG. 3B, the difference between the demand amount predicted by inputting the actual weather information value of the check data and the actual demand value of the check data is determined when the number of times of learning is a predetermined number n. It becomes the minimum and becomes larger after that. Therefore, the learning stop signal f is output at a point where the error is minimized. As a result, it is possible to prevent the construction of a prediction model that is applicable only to the learning data, and it is possible to construct a more accurate prediction model.

Next, the neural network prediction model construction means 11 will be described. The neural network used in the neural network prediction model construction means 11 according to the embodiment of the present invention has a three-layer structure including an input layer, an intermediate layer, and an output layer, as shown in FIG.
Now, it is assumed that the input layer has l neurons, the intermediate layer has m neurons, and the output layer has n neurons. The sigmoid function shown in the equation (1) is used for the intermediate layer.

[0026]

(Equation 1)

Here, sa, sb, sc, sd: parameters x: input values of the intermediate layer f (x): sigmoid function

A prediction model is constructed by updating the weight coefficient W between the connections of the respective layers shown in FIG. 4 using the back propagation method.

(1) Step 1 From the signal given to the input layer, the intermediate layer and the output layer are operated according to the following neuron model.

<Hidden Layer Calculation> The output Hj of the jth neuron in the hidden layer is obtained by equation (2).

[0031]

(Equation 2)

Here, Ii: the output of the i-th neuron in the input layer Hj: the output of the j-th neuron in the intermediate layer Wji: the weight coefficient l between the i-th neuron in the input layer and the j-th neuron in the intermediate layer: the number of neurons in the input layer f0: threshold function of the intermediate layer {Equation (1)}

<Calculation of Output Layer> The output Ok of the k-th neuron in the output layer is obtained by equation (3).

[0034]

(Equation 3)

Here, Hj: output of the j-th neuron in the intermediate layer Ok: output of the k-th neuron in the output layer Wkj: weight coefficient m between the j-th neuron in the intermediate layer and the k-th neuron in the output layer m: number of neurons in the intermediate layer

(2) Step 2 A prediction model is constructed (the weight coefficient W is corrected by learning). The output Ok of the k-th neuron in the output layer and the k-th output in the output layer
The learning is performed by modifying the weighting coefficient W of the network so as to minimize the sum of the square error with the teaching signal yk of the neuron.

The correction of the weight coefficient Wkj between the intermediate layer and the output layer is performed using equation (4).

[0038]

(Equation 4)

Here, t: number of learning times ε: parameter for determining the magnitude of one correction dk: error of output layer

The correction of the weight coefficient Wji between the input layer and the intermediate layer is performed using equation (5).

[0041]

(Equation 5)

Here, dj: back propagation error of the intermediate layer f'0: differentiation of f0

In order to further reduce the oscillation at the time of learning and accelerate the convergence, the equation (6) is used.

[0044]

(Equation 6)

Where α is a parameter for stability.

The weighting factor W obtained as a result of this learning is stored in the learned weighting factor storage means 23 as a prediction model.

The data used as input / output of the neural network may be input weather (morning and afternoon of the previous day, that day) temperature (minimum, maximum temperature) water distribution of the previous day output water distribution of the day on the previous day Is used as described below.

For example, the weather stored in the weather information actual value storage means 22 is fine: a, cloudy: b, rain: c, snow: d
(A, b, c, d are values of [−1 to 1]),
The temperature accumulated in the weather information actual value accumulating means 22 and the actual demand value accumulated in the actual demand accumulating means 21 are processed by equation (7).

[0049]

(Equation 7)

Here, θ (i): data on the day i θ: average value of data σθ: variance of data θ: deviation from the average value n: number of data

In this way, a prediction model is constructed by the neural network prediction model construction means 11 and stored in the learned weight coefficient accumulation means 23. Then, the neural network demand prediction unit 14 predicts the demand on the day using the weather forecast value set by the operator 31 using the prediction model stored as the prediction model in the learned weight coefficient storage unit 23. The demand prediction result is provided to the operator 31 and is used as information for determining the water supply facility operation method on the day.

According to this embodiment, it is possible to efficiently operate the water supply facility and accurately predict the purified water consumption on the day, which is important for supplying an appropriate amount of purified water to the customer without excess or shortage. Becomes

In the above description, the prediction regarding the purified water consumption has been described. However, the short-term demand prediction device according to the embodiment of the present invention is not limited to the prediction of the purified water consumption. Is also applicable.

[0054]

As described above, according to the present invention, it is possible to accurately predict the consumption on the day, so that the facility supplying the demand can be operated efficiently, and The quantity can be supplied to consumers without excess or shortage.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a short-term demand prediction device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a method of classifying learning data and check data by an actual value classifying unit according to the embodiment of the present invention.

FIG. 3 is an explanatory diagram of a method of determining learning stop by learning stop determining means according to the embodiment of the present invention.

FIG. 4 is an explanatory diagram of a neural network prediction model construction unit according to the embodiment of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Short-term demand prediction apparatus, 11 ... Neural network prediction model construction means, 12 ... Learning stop determination means, 13 ...
Actual value classifying means, 14: Neural network demand forecasting means, 21: Demand actual value accumulating means, 22: Weather information actual value accumulating means, 23: Learned weight coefficient accumulating means, 31
… Operator

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Yasuyuki Miyajima 1-1-1, Shibaura, Minato-ku, Tokyo Inside Toshiba Corporation Head Office (72) Inventor Yukihiko Tomizawa 1-1-1, Shibaura, Minato-ku, Tokyo Shares F-term in the Toshiba head office (reference) 5B049 AA00 BB00 CC11 EE14 EE31 GG07 5H004 GB08 HB02 HB20 KC23 KC25 KC28 KD42 9A001 HH03 HH06 JJ73 KK55

Claims (4)

[Claims] 1. A performance value classifying means for classifying accumulated performance values into learning data and check data, and learning based on performance values classified as learning data by the performance value classification means to construct a prediction model. A neural network prediction model constructing unit, a learning stop determining unit for stopping learning in the neural network prediction model constructing unit based on the actual value classified as check data by the actual value classifying unit, and a neural network constructing unit. A short-term demand forecasting device, comprising: a neural network demand forecasting unit for forecasting the demand on the day using the constructed forecast model. 2. The actual value classifying means classifies the newly added actual value as check data, alternately classifies the accumulated actual value into learning data and check data, or sorts the accumulated actual value. The short-term demand prediction device according to claim 1, wherein the actual value of a certain period is classified as check data. 3. The learning stop judging means predicts a demand amount by using a prediction model constructed by the neural network prediction model constructing means, compares the predicted demand amount with an actual value corresponding thereto, and calculates an error. The short-term demand prediction device according to claim 1 or 2, wherein the learning of the neural network prediction model construction means is stopped when the minimum is reached. 4. The accumulated actual value is a weather information actual value and a demand actual value, and the forecasting model is a demand forecasting model in a water supply system. Short-term demand forecasting device as described.