JP4735973B2 - Power price zone prediction method and power price zone prediction program - Google Patents

Description

  The present invention relates to a power price zone prediction method using a neural network to which a clipping method is applied, and a power price zone prediction program.

  With the recent expansion of electricity liberalization in Japan, competition in the electric power business has intensified. In the future liberalized electricity market, electricity price forecasts will be very important for each company to make more profits. Conventionally, as a power transaction prediction system, power supply side and demand side bid information, generator data, power system data and demand data are input, and a set of these input data is registered as scenario data. A system is known in which a power price and a power transaction amount, which are simulated market transaction results, are calculated, and a simulation is performed in consideration of differences in system constraints and bidding conditions (Patent Document 1).

Japanese Patent Laying-Open No. 2005-25377 (FIG. 1)

  However, since electricity market prices are determined from various factors such as weather and fuel costs, the nonlinearity is very strong, and as shown in FIG. Unlike stock prices, it is a complex problem that is very difficult to predict. In the above-described conventional power price prediction system, since there is a lack of information correlated with the power price, there is a risk that a large price error of several tens to 200% may occur, and only a very inaccurate prediction may be made. could not.

  The present invention has been made in view of these points, and an object thereof is to provide a power price zone prediction method and a power price zone prediction program capable of predicting a power price with high accuracy.

  A power price zone prediction method according to the present invention is a power price zone prediction method for predicting a range of power prices using a computer including an input device, a storage device, a calculation device, and an output device, and is input by the input device. A step of determining that the computing device classifies the price range of the power price into a plurality of zones based on the information of the zone for classifying the price range of the power price, and sets a boundary between the plurality of zones as a threshold value. A step of learning a neural network constructed on the storage device by obtaining output data obtained by clipping an output value for learning data including a predetermined input variable based on the determined threshold, and the learned neural network Is used to determine the output value for the input data including the predetermined input variable based on the determined threshold value. A step of obtaining a tapped output value, and a step of using the output value obtained in this step to predict whether the price ranges of the plurality of zones classified by a binary tree are to be output via the output device It is characterized by comprising.

  A power price zone prediction program according to the present invention is a power price zone prediction program for causing a computer including an input device, a storage device, a calculation device, and an output device to predict a range of power prices, The computing device classifies the price range of the power price into a plurality of zones based on the information of the zone that classifies the price range of the power price input by the input device, and sets a boundary between the zones of the plurality of zones as a threshold value. And a step of learning a neural network constructed on the storage device by obtaining output data obtained by clipping an output value for learning data including a predetermined input variable based on the determined threshold value. And input data including the predetermined input variable using the learned neural network A step of obtaining an output value obtained by clipping the output value with respect to the determined threshold value, and using the output value obtained in this step to predict whether the price ranges of the plurality of zones classified by a binary tree And the step of outputting via the output device.

  According to the present invention, it is possible to provide a power price zone prediction method capable of predicting a power price range by a neural network to which a clipping method is applied and predicting a possible range of the power price with high accuracy.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a diagram showing a computer system for realizing the power price zone prediction method according to the present embodiment. This system includes an input device (or receiving device) 1, an arithmetic device 2, a storage device 3, and an output device (or transmitting device) 4. This system includes a control program for causing the input device 1, the arithmetic device 2, the storage device 3, and the output device 4 to execute various processes, and executes various processes described later according to the control program. The input device 1 is used when inputting information on a zone for classifying a price range of electric power prices. The arithmetic device 2 classifies the price range of the electric power price into a plurality of zones and determines to use the boundaries of the plurality of zones as a threshold value. Further, the arithmetic device 2 obtains output data obtained by clipping the output value for the learning data including the predetermined input variable based on the determined threshold value, so that the neural network constructed on the storage device 3 learns. To do. Further, the arithmetic unit 2 uses the learned neural network to select one of the plurality of classified zones using the output value clipped based on the threshold value that determines the output value for the input data including the predetermined input variable. Determine by binary tree, and determine the price range forecast result. The output device 4 is used when displaying or transmitting a price range prediction result.

  In this embodiment, the price range of the power price is predicted by a neural network to which a clipping method is applied. Then, a simulation is performed using each neural network of MLP (multilayer perse proton), RBFN (radial basis function network), and NRBFN (normalized radial basis function network), and comparison verification of recognition rates is performed.

  Here, each neural network of MLP, RBFN, and NRBFN will be briefly described.

  FIG. 2 is an explanatory diagram showing a general configuration of the MLP. As shown in FIG. 2, an MLP is generally composed of three layers: an input layer, an intermediate layer, and an output layer. The output of the intermediate layer and the output layer is generally expressed by a sigmoid function. The MLP is updated by iteratively learning the connection weights between the layers, thereby constructing an optimum network.

  FIG. 3 is an explanatory diagram showing a general configuration of RBFN. As shown in FIG. 3, the RBFN is composed of a three-layer network having a radial basis function as an intermediate layer. As the radial basis function is closer to the center, a stronger output is obtained, and it is generally expressed by Gaussian. The combination of the input layer and the intermediate layer has a function of simply transmitting the input, and the weight is arranged only in the combination of the intermediate layer and the output layer. As described above, since the weight is only a combination of the intermediate layer and the output layer, learning can be performed at a speed higher than that of the MLP, and an arbitrary nonlinear function can be approximated with high accuracy if a sufficient basis function is given.

  FIG. 4 is an explanatory diagram showing a general configuration of the NRBFN. NRBFN is an evolution of radial basis function networks. The basic structure of the NRBFN is the same as that of the RBFN, but the NRBFN normalizes the output of the RBFN with the sum of the outputs of the respective basis functions. By performing output normalization in this way, generalization ability can be improved.

  The output of the RBFN is local as shown by the dotted line in FIG. 5, whereas the output of the NRBFN is obtained by complementing the centers of the basis functions in a sigmoid manner as shown by the solid line in FIG. Thus, the NRBFN that can obtain a global output is excellent in generalization ability and behaves like a fuzzy inference system.

  Next, clipping will be described. Clipping means conversion from a real value series to a binary series, as shown in FIG. In the present embodiment, by clipping, the value of the time-series data at time t is expressed as 1 when it is larger than the threshold μ, and is expressed as 0 when it is smaller than the threshold μ. In the present embodiment, clipping is applied to the output value of the prediction model to predict whether it is above or below the threshold value. Prediction is performed for each set threshold value, and the prediction result is used to determine the price range in the next step.

  In the present embodiment, as shown in FIG. 7, power price zone prediction using a clipping method and a binary tree is performed. In the present embodiment, the price range itself is predicted, not the price value itself. In this embodiment, it is determined whether or not a threshold value set in each prediction model is exceeded, and a price range is determined by a binary tree using the determination result. By predicting the possible range of the price in this way, a highly accurate prediction result can be obtained.

  Note that the prediction model is constructed by an artificial neural network (ANN). In this embodiment, each neural network of MLP, RBFN, and NRBFN is used as a prediction model.

  FIG. 8 shows a flowchart in the case where the clipping method and the binary tree are applied to power price zone prediction. First, initial setting of each prediction model 1-i is performed (step S1). Specifically, as initialization, each variable is initialized, and the parameters of each prediction model 1 to i are determined and set in the prediction model. Since the output of the sigmoid function or the radial basis function is from 0 to 1, the data input to each prediction model 1 to i is normalized so as to be within the range of 0 to 1. In addition, the parameter set to each prediction model 1-i is determined by trial and error, for example.

  Next, the power price range is divided into n zones, and n-1 threshold values serving as boundaries between the zones are determined (step S2). Specifically, for example, as shown in FIG. 9B, the power price range is classified into four categories of zone A, zone B, zone C, and zone D, and μ1, μ2, and μ3 that are boundaries of each zone. Is determined as a threshold value. Note that the number of classifications (number of zones) and the classification width (range of each zone) may be determined according to the purpose, for example.

  Next, in this embodiment, the prediction models 1 to i are learned using the steepest descent method (step S3). Here, input variables at the time of actual power prices included in each zone, such as power demand, temperature, maximum daily load, wet / dry temperature at peak load, maximum temperature in region A, average temperature, cooling A temperature deviation, the maximum temperature in the regions B and C, etc. are given as learning data to each prediction model 1 to i. Then, clipping is performed on the output value of the learning data with reference to the threshold set in step S1. In the present embodiment, in order to predict the price range, the relationship between the given learning data and the output data obtained by clipping the output value of the learning data with the respective threshold values μi determined in step S2. Each prediction model 1 to i learns.

  Then, the input variables described above are given as input data to each prediction model 1 to i learned in step S3, and binary prediction above or below the threshold set in step S1 is performed (step S4). Here, the value obtained by each prediction model 1 to i is the final result of clipping (set of 0 or 1) for each threshold value set in step S1 with respect to real values from 0 to 1. Estimated value. That is, the final predicted value is 1 for the threshold μ1, 1 for the threshold μ2, 0 for the threshold μ3, and so on.

  Finally, a price range is determined by a binary tree using the prediction result in step S4 (step S5). Here, the method for determining the expected price range using the binary tree in step S5 will be described. FIG. 9A is an explanatory diagram showing a price range determination method when the price range is set as shown in FIG. 9B.

  As shown in FIG. 9A, in each branch, it is determined whether it is above or below the threshold value using the result predicted in the previous step. In this way, it is possible to determine the price range by determining whether each branch is above or below the threshold and limiting the range that the price can take. For example, when the final predicted value in step S4 is 1 for the threshold μ1, 1 for the threshold μ2, and 0 for the threshold μ3, the price range prediction result is between μ2 and μ3, that is, the zone C. It becomes.

  A simulation result of power price zone prediction using the clipping method and binary tree according to this embodiment will be described below.

  In this simulation, the hourly price, demand, and temperature data in the New England power pool published on the ISO New England web were used. Specifically, 774 pieces of learning data from July 1, 2004 to the same month 31st of the same year, and 168 test data from August 1, 2004 to the same day 7th of the same year as shown in FIG. It was used. This simulation is performed using three prediction models of MLP, RBFN, and NRBFN, and the recognition rate for each prediction model is compared and verified.

  In this simulation, it is assumed that the parameters of each prediction model are determined by trial and error to the values shown in FIG. 11 and the determined parameters are initially set.

  Further, as shown in FIG. 12A, in this simulation, the input data is price, demand, and temperature data for the past two hours, and the output is clipped price data. These numbers of input data are arbitrary, and may be other numbers such as data for 3 hours. In addition, as shown in FIG. 12B, a price range is set for every 50 dollars in case 1, every 10 dollars in case 2, and every 5 dollars in case 3, and a simulation was performed in each case. .

  For calculating the recognition rate, two equations shown in FIG. 13 were used. Formula 1 is a formula for calculating the recognition rate (r1) when the recognition is successful when the actual price and the predicted price range completely match. Formula 2 is a formula for calculating the recognition rate (r2) when the allowable error between the actual price and the predicted price range is 1.

  The simulation results under the above simulation conditions are shown in FIGS. FIG. 14 is an explanatory diagram showing a recognition rate when Expression 1 is used. In Case 1, the recognition rate is not so different between the models, but in Case 2, the recognition rate was 2.98% higher for RBFN and 4.17% higher for NRBFN than MLP. In Case 3, the recognition rate of RBFN was 3.57% and NRBFN was 8.93% higher than MLP. From these results, it can be seen that NRBFN is the most effective model.

  FIG. 15 is an explanatory diagram showing the recognition rate when Expression 2 is used. In Case 2, MLP was 99.40%, and RBFN and NRBFN were 100%, which is a very high recognition rate. The recognition rate in case 3 also showed a high recognition rate of 90% or more. From these results, it was found that the predicted price range is less likely to deviate by more than two from the actual price range.

  As described above, in this embodiment, since the power price range is predicted by the neural network to which the clipping method is applied, the range that the power price can take is predicted using the neural network having a high nonlinear approximation capability. The power price can be predicted with high accuracy. In particular, as shown in the above-described simulation results, it is possible to predict the range of the power price with the highest accuracy by using NRBFN to predict the range that the power price can take.

1 is a block diagram of a power price zone prediction system according to an embodiment of the present invention. It is explanatory drawing which shows the structural example of a multilayer perse proton (MLP). It is explanatory drawing which shows the structural example of a radial basis function network (RBFN). It is explanatory drawing which shows the structural example of a normalization radial basis function network (NRBFN). It is explanatory drawing which shows the example of an output of a normalization radial basis function network. It is explanatory drawing which shows the example of the data before and behind clipping. It is a conceptual diagram of the prediction process of the zone of the electric power price in this embodiment. It is a flowchart at the time of applying a clipping method and a binary tree to zone prediction of an electric power price. FIG. 9A is an explanatory diagram showing a method for determining a price range. FIG. 9B is an explanatory diagram illustrating an example of setting a price range. It is explanatory drawing which shows test data. It is explanatory drawing which shows the parameter initially set to each prediction model. FIG. 12A is an explanatory diagram showing simulation conditions for a prediction model. FIG. 12B is an explanatory diagram showing the setting contents of the price range in the simulation. It is explanatory drawing which shows the example of the calculation formula of a recognition rate. It is explanatory drawing which shows the example of a simulation result. It is explanatory drawing which shows the other example of a simulation result. It is a graph which shows the time-sequential fluctuation | variation of an electric power price.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Input device 2 ... Arithmetic device 3 ... Memory | storage device 4 ... Output device

Claims (2)

A power price zone prediction method for predicting a range of power prices using a computer including an input device, a storage device, a computing device, and an output device,
Based on the information of the zone that classifies the price range of the power price input by the input device, the computing device classifies the price range of the power price into a plurality of zones, and sets the boundaries of the plurality of zones as threshold values. Steps to decide to do,
Learning a neural network built on the storage device by obtaining output data obtained by clipping an output value for learning data including a predetermined input variable based on the determined threshold;
Using the learned neural network to obtain an output value obtained by clipping an output value for input data including the predetermined input variable based on the determined threshold;
Using the output value obtained in this step, predicting whether it is the price range of the plurality of zones classified by a binary tree, and outputting via the output device Electricity price zone prediction method. A power price zone prediction program for causing a computer including an input device, a storage device, a computing device, and an output device to predict a range of power prices,
In the computer,
Based on the information of the zone that classifies the price range of the power price input by the input device, the computing device classifies the price range of the power price into a plurality of zones, and sets the boundaries of the plurality of zones as threshold values. Steps to decide to do,
Learning a neural network built on the storage device by obtaining output data obtained by clipping an output value for learning data including a predetermined input variable based on the determined threshold;
Using the learned neural network to obtain an output value obtained by clipping an output value for input data including the predetermined input variable based on the determined threshold;
Using the output value obtained in this step, predicting whether it is the price range of the plurality of zones classified by a binary tree, and outputting the output via the output device Prediction program.