JP2013074695A - Device, method and program for predicting photovoltaic generation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide device, method and program for predicting photovoltaic generation, capable of estimating a photovoltaic power generation amount with high accuracy.SOLUTION: The photovoltaic generation prediction device for predicting a photovoltaic power generation amount includes: data input means for inputting data necessary for predicting the power generation amount; clustering processing means for classifying the input data by clustering so as to obtain a neural network learning data; model construction means for constructing a neural network model on the basis of each clustered cluster, to classify into each cluster having a cluster representation nearest to a test data; and prediction means for predicting the photovoltaic power generation amount using the neural network in the cluster having the classified test data, to output the predicted power generation amount.

Description

  The present invention relates to a solar power generation prediction apparatus, a solar power generation prediction method, and a solar power generation prediction program that predict solar power generation with high accuracy.

  In recent years, due to the problem of global warming, the introduction and expansion of photovoltaic power generation is expected to realize a low-carbon society. When solar power generation that is effective in reducing carbon dioxide is introduced into the power grid, there is uncertainty that the generated power varies depending on the weather. In order to operate the power system stably, it is necessary to predict solar power generation with high accuracy. To date, research on the prediction of photovoltaic power generation includes methods such as time series models and neural networks (hereinafter referred to as ANN) for solar radiation amount prediction. In solar power generation prediction, methods using meteorological data, methods using predicted values of solar radiation, and the like have been proposed.

  It also estimates the illuminance received on the surface of the photovoltaic module, estimates the electricity production of the photovoltaic generator, and corrects the estimate based on whether the past relationship period is clear or cloudy A method for predicting the amount of electricity produced by a photovoltaic power generation apparatus is known (see, for example, Patent Document 1).

JP 2010-239856 A

  By the way, in a power plant that supplies power, a power generation plan is made one day in advance, and the output of the day is adjusted according to the power demand. However, solar power generation has input uncertainty due to weather changes. Therefore, in order to realize stable operation of the power system at the power plant, highly accurate photovoltaic power generation prediction is essential. However, the conventional prediction method has a problem that the power generation amount of solar power generation cannot be predicted with high accuracy.

  The present invention has been made in view of such circumstances, and provides a solar power generation prediction device, a solar power generation prediction method, and a solar power generation prediction program capable of predicting the power generation amount of solar power generation with high accuracy. The purpose is to do.

  The photovoltaic power generation prediction apparatus of the present invention is a photovoltaic power generation prediction apparatus that predicts the power generation amount of solar power generation, and uses data input means for inputting data necessary for power generation amount prediction, and learning data of a neural network. Therefore, a clustering processing means for classifying the input data by clustering, a model construction means for constructing a neural network model for each clustered cluster, and classifying the cluster having a cluster representative closest to the test data, and the test Predicting means for predicting the power generation amount of photovoltaic power generation using a neural network in the cluster into which the data is classified and outputting the predicted power generation amount is provided.

  The clustering processing means of the present invention is characterized by performing clustering using deterministic annealing.

  The neural network model of the present invention uses a radial basis function network.

  The solar power generation prediction method of the present invention is a solar power generation prediction method for predicting the power generation amount of solar power generation, and uses a data input step for inputting data necessary for power generation amount prediction and learning data of a neural network. Therefore, a clustering process for classifying the input data by clustering, a model construction step for constructing a neural network model for each clustered cluster and classifying the cluster having the cluster representative closest to the test data, and the test A prediction step of predicting a power generation amount of photovoltaic power generation using a neural network in a cluster into which data is classified and outputting a predicted power generation amount.

  The photovoltaic power generation prediction program of the present invention is a photovoltaic power generation prediction program for causing a computer on a photovoltaic power generation prediction apparatus that predicts the amount of photovoltaic power generation to predict the amount of photovoltaic power generation. A data input step for inputting necessary data, a clustering step for classifying the input data by clustering in order to obtain neural network learning data, a neural network model is constructed for each clustered cluster, and test data And a model building step for classifying the cluster having the nearest cluster representative, and a prediction step for predicting the power generation amount of photovoltaic power generation using a neural network in the cluster into which the test data is classified and outputting the predicted power generation amount It is characterized by making it perform.

  According to the present invention, input data necessary for power generation amount prediction is classified by clustering, a neural network model is constructed for each cluster, a cluster having the cluster representative closest to the test data is classified, and the test data is classified. Since the power generation amount of solar power generation is predicted using the neural network in the cluster, an effect that the power generation amount of solar power generation can be predicted with high accuracy is obtained.

It is a block diagram which shows the structure of one Embodiment of this invention. It is a flowchart which shows the processing operation of the apparatus shown in FIG. It is a flowchart which shows the processing operation of the apparatus shown in FIG. It is explanatory drawing which shows the structure of RBFN. It is a figure which shows the concept of the process which estimates the electric power generated of a solar power generation device using DA clustering as a method of pre-processing. It is a figure which shows the change of an evaluation function when the update width | variety (DELTA) (beta) of DA is changed. It is a figure which shows the change of the data allocated to each cluster when changing update width (DELTA) (beta) of DA. It is a figure which shows the relationship between (DELTA) (beta) in DA clustering required for RBFN construction | assembly in the case of the cluster 1, and an evaluation function. It is a figure which shows the relationship between (DELTA) (beta) in DA clustering in the case of the cluster 1, and the number of data which belonged to the cluster. It is a figure which shows the relationship between (DELTA) (beta) in DA clustering required for RBFN construction in the case of the cluster 2, and an evaluation function. It is a figure which shows the relationship between (DELTA) (beta) in DA clustering in the case of the cluster 2, and the number of data which belonged to the cluster. It is a figure which shows the relationship between (DELTA) (beta) in DA clustering required for RBFN construction in the case of the cluster 3, and an evaluation function. It is a figure which shows the relationship between (DELTA) (beta) in DA clustering in the case of the cluster 3, and the number of data which belonged to the cluster. It is the figure which compared the prediction result output by this embodiment with the number of clusters, and a measured value. It is a figure which shows transition of the error in each time.

  Hereinafter, a photovoltaic power generation prediction apparatus according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the embodiment. In this figure, the code | symbol 1 is a solar power generation prediction apparatus comprised with a computer apparatus. Reference numeral 2 denotes an input unit that is configured by an input device such as a keyboard and performs an information input operation. Reference numeral 3 denotes a display unit that is configured by a display device and displays power generation amount prediction result information and the like. Reference numeral 11 denotes a data input unit for inputting data necessary for power generation amount prediction. In addition to data input from the input unit 2, the data input unit 11 inputs data output from other devices and sensor output data using communication means. Reference numeral 12 denotes a clustering processing unit that performs a clustering process to classify input data input in the data input unit 11. Reference numeral 13 denotes a model construction unit that constructs a prediction model for each cluster classified in the clustering processing unit 12. Reference numeral 14 denotes a prediction processing unit that predicts the generated power of photovoltaic power generation based on the constructed prediction model and outputs a prediction result.

  Next, the operation principle of the photovoltaic power generation prediction apparatus according to this embodiment will be described. In this embodiment, in order to apply ANN in photovoltaic power generation prediction and obtain a more accurate output, a prediction method using a neural network using clustering is used as data preprocessing. In the preprocessing, the input data is classified into several clusters by clustering. Thereafter, similar data is extracted in each cluster, and a prediction model is constructed. The prediction model uses a highly accurate nonlinear approximate radial basis function network (hereinafter referred to as RBFN). ANN is characterized in that the smaller the variance of input data, the higher the output that can be obtained. That is, a highly accurate solution can be obtained by using learning data more similar to test data. By using similar input data in input / output data having a strong correlation, the output data is similar with high probability. Therefore, a preprocessed model by clustering is used to create a set of similar data. Since a prediction model is constructed for each cluster from which similar data is extracted, the dispersion of input data of each prediction model is suppressed, and a higher-accuracy solution than a single prediction model can be obtained.

  Methods used as preprocessing include clustering methods such as k-means. k-means calculates at high speed, but has a drawback that the solution depends on the initial value. Here, deterministic annealing (hereinafter referred to as DA) capable of high-accuracy and global clustering is used as a clustering method. DA uses free energy that depends on temperature in the concept of statistical mechanics, and converges to an optimal solution by gradually decreasing the temperature parameter from a high temperature state. Since the objective function changes at each temperature and optimization is performed deterministically, a global optimal solution can be obtained efficiently without depending on the initial value. Also, for determining the center vector of RBFN, by using DA clustering, it is possible to construct an RBFN having a global structure and perform highly accurate prediction.

  Here, RBFN will be described with reference to FIG. FIG. 4 is a diagram showing the structure of RBFN. The RBFN is a hierarchical neural network having a three-layer structure including an input layer, an intermediate layer, and an output layer. Input data is allocated to each unit of the input layer, and passed to each intermediate layer without weight. The unit of the intermediate layer has a radial basis function called a response function, and performs the same function as a sigmoid function in a multilayer perceptron (hereinafter referred to as MLP). The radial basis function is expressed by a Gaussian density function defined by the position and width of the center. The Gaussian density function gives the maximum output when the input data is closest to the center position, and decreases monotonically as it moves away from the center. The width of the radial basis function serves to control the rate of decrease. By using this radial basis function, RBFN has a higher nonlinear approximation capability than MLP. Further, since only the weight between the intermediate layer and the output layer is learned, there is an advantage that the calculation time can be greatly reduced as compared with the MLP.

ＲＢＦＮの出力値ｙは、中間層のユニットの重み付けされた値となり、（２）式で求められる。

Mathematically, the output of the RBFN intermediate layer unit is expressed as shown in equation (1).

The output value y of the RBFN is a weighted value of the unit of the intermediate layer, and is obtained by the equation (2).

  For the determination of the center vector of the radial basis function, the center can be determined by using a clustering method. In the prior art, k-means, which is a representative clustering method, is used in determining the center vector. This method applies each cluster representative obtained by k-means to the center of the basis function of RBFN. However, k-means is a method that depends on the initial value, and since optimal clustering cannot always be performed, the RBFN has a local structure. In the present embodiment, DA clustering is used for determining the center of the RBFN basis function. This method can perform optimal clustering without depending on the initial value. An RBFN having a global structure can be obtained by using the cluster representative obtained in this way.

  Next, clustering will be described. Clustering is a technique for finding a characteristic pattern of data from a data distribution in a set of data and dividing similar data into subsets called clusters. That is, the clustering is a process of classifying a set of vectors of m samples of given data into clusters having k representative vectors. The representative vector is the center vector of each classified cluster and represents the feature of the cluster.

K-means, which is one of clustering methods, searches for a division that minimizes an evaluation function representing a distance. The evaluation function is shown in equation (3).

  In k-means, first, the centroids of k initial clusters are arbitrarily set, the distance between each data and the centroid of each cluster is obtained, and assigned to the cluster having the centroid that minimizes the distance. If the allocation to the cluster does not change, the process ends. If not, the center of gravity of each cluster is obtained, and the cluster allocation process is repeated as a representative point of each cluster.

  On the other hand, the DA clustering method is an annealing method, and converges to an optimal solution by gradually decreasing the temperature parameter from a high temperature state. Since DA uses the free energy in statistical mechanics and changes the objective function at each temperature, optimization can be performed efficiently without depending on the initial value of the solution or falling into a local solution. DA gradually decreases the temperature, and the objective function approaches a given nonlinear function as the temperature decreases. This gives DA the original objective function at sufficiently low temperatures. In addition, the DA performs deterministic optimization at each temperature, and uses the optimal solution obtained in the current state as the initial value of the next state to perform iterative calculation. The smaller the temperature change, the smaller the change in the objective function, so deterministic optimization is performed at each temperature by slowly lowering the temperature from a high temperature state. By repeating this work, a global optimum solution is found at a sufficiently low temperature.

DA clustering is a clustering technique that applies the principle of DA. Since DA clustering uses probabilities in the affiliation state, certain data can belong to a plurality of clusters. Thereby, in the clustering process, since the cluster representative for each data is searched flexibly, it is possible to suppress dependence on the initial value and falling into a local solution. Therefore, in DA, the affiliation state is expressed by a probability to obtain equation (4). DA clustering uses the principle of maximum entropy for the affiliation probability in equation (4). As a result, each data can belong to all clusters. The affiliation probability of the equation (4) using the maximum entropy is a Boltzmann distribution as in the equation (5).

Β in equation (5) represents the reciprocal of the temperature parameter in the annealing algorithm, and optimization is performed at each temperature. When β = 0, each data belongs uniformly to all clusters. On the other hand, as β increases, that is, the temperature decreases, each data belongs to one cluster little by little. Since the probability of belonging to a certain cluster is 1 at β = ∞ and 0 to other clusters, all data belong to one cluster. In an equilibrium state in statistical mechanics, the free energy is minimized, and therefore the equation (6) is obtained for the representative of the cluster at each temperature. The cluster representative is updated by solving equation (6). In order to optimally optimize the clustering at each temperature, the equations (5) and (6) are repeatedly calculated. Equation (7) represents a new membership probability that can be obtained from the cluster representative at the number of iterations t + 1 obtained in Equation (6).

  Thus, in DA clustering, the objective function changes each time the temperature is lowered, and the cluster representative is deterministically optimized at each temperature. Here, with reference to FIG. 2, the DA clustering processing operation in the clustering processing unit 12 shown in FIG. 1 will be described. FIG. 2 is a flowchart showing the DA clustering processing operation in the clustering processing unit 12 shown in FIG.

First, the clustering processing unit 12 sets an initial temperature parameter β ₀ , an upper limit value β _max of the temperature parameter, and a convergence criterion ε (step S1). Subsequently, the clustering processing unit 12 selects an initial cluster representative from the data (step S2), and obtains an affiliation probability at t using equation (7) (step S3).

  Next, the clustering processing unit 12 obtains a new cluster representative from the obtained affiliation probability using the equation (6) (step S4), and obtains a new affiliation probability from the new cluster representative using the equation (7) (step S5).

Next, the clustering processing unit 12 evaluates d ^{(t + 1)} from the evaluation function of equation (4) (step S6). As a result of this evaluation, it is determined whether or not the condition | d ^{(t + 1)} −d ^(t) | / d ^(t) <ε is satisfied (step S7). If the condition is not satisfied as a result of this determination, the process returns to step S4 and the process is repeated.

On the other hand, if this condition is satisfied, the clustering processing unit 12 determines whether or not the end condition (β ≧ β _max ) is satisfied (step S8). If the end condition is not satisfied as a result of this determination, the clustering processing unit 12 increases β (step S9), returns to step S4, repeats the process until the end condition is satisfied, and when the end condition is satisfied End the process.

  Next, the processing operation for predicting the generated power of the photovoltaic power generation apparatus by the photovoltaic power generation prediction apparatus 1 shown in FIG. 1 will be described. Here, the input data preprocessing and the prediction method by RBFN are used for the prediction of the generated power of the photovoltaic power generation. The ANN has a feature that a highly accurate output can be obtained as the variance of input data is small, and a highly accurate solution can be obtained by using learning data more similar to test data. That is, by using similar input data in input / output data having a strong correlation, output data is similar with high probability. Therefore, if a clustering method capable of extracting similar data is performed as preprocessing, the nonlinear approximation capability of ANN can be utilized efficiently. Therefore, in this embodiment, DA clustering capable of highly accurate clustering is used as a preprocessing method.

With reference to FIG. 5, the process which estimates the electric power generated of a solar power generation device using DA clustering as a pre-processing method is demonstrated. FIG. 5 is a diagram illustrating a concept of processing for predicting the generated power of the photovoltaic power generation apparatus using DA clustering as a preprocessing technique. In FIG. 5, x represents input data, Cl _i (i = 1, 2,..., N) represents a cluster classified by DA clustering, and y _i (i = 1, 2,..., N) represents an output. When the test data is input, the cluster having the closest cluster representative is selected, and prediction is performed by the RBFN constructed in the cluster. Also, DA clustering is used to determine the center of the RBFN basis function. Since DA clustering can perform optimal clustering, an RBFN having a global structure is obtained by using the obtained cluster representative.

  Next, the processing operation of the photovoltaic power generation prediction apparatus 1 shown in FIG. 1 will be described with reference to FIG. FIG. 3 is a flowchart showing the processing operation of the photovoltaic power generation prediction apparatus 1 shown in FIG. First, the data input unit 11 inputs data necessary for power generation amount prediction, and delivers the input data to the clustering processing unit 12. The data entered here includes time, solar radiation, temperature, solar panel temperature, humidity, sunshine duration, total solar radiation, direct solar radiation, diffuse solar radiation, solar radiation index, cloud cover, ultraviolet light, illuminance, and rainfall. , Solar altitude, solar orientation, etc. In response to this, the clustering processing unit 12 classifies the input data by DA clustering by the processing operation described above (step S11), and sets the data included in each cluster as RBFN learning data (step S12).

  Next, the model construction unit 13 constructs an RBFN model for each cluster clustered by the clustering processing unit 12 (step S13), and classifies the cluster having the cluster representative closest to the test data (step S14). Then, the prediction processing unit 14 uses the RBFN in the cluster into which the test data is classified to predict the generated power of the photovoltaic power generation (Step S15) and displays the prediction result on the display unit 3. The prediction processing unit 14 may output the prediction result information to another device. By this processing operation, the predicted power generation amount for each predetermined elapsed time is related and output.

Next, an increase in the parameter β for DA clustering in the preprocessing will be described. Here, the increase amount of the parameter β in “increase the parameter” in step S9 shown in FIG. 2 is referred to as an update width Δβ. DA clustering is a technique for obtaining an optimal clustering result by gradually changing the parameter β. However, depending on how β is changed, there may be no solution. Therefore, in this embodiment, the parameter β is changed as shown in the equation (8).
β _{k + 1} = β _k + Δβ (8)
Where k is the number of iterations of the algorithm, β _{k is} the parameter β when the number of iterations is k, and Δβ is the update width.

In order to obtain a good solution, Δβ in equation (8) is important. The upper and lower limit values of Δβ for converging all data to the cluster in the preprocessing will be described. FIG. 6 shows changes in the evaluation function when the DA update width Δβ is changed. The smaller the evaluation function on the horizontal axis, the better the cluster result. It can be seen from the figure that a good evaluation function is obtained when the update width Δβ is 0.01 or less. Therefore, the upper and lower limits of Δβ are
0.03 ≦ Δβ ≦ 1.3 (9)
However, when Δβ ≦ 0.02 and 1.4 ≦ Δβ, the evaluation function seems to be good, but since all the data has not converged to any cluster, it is determined to be inappropriate.

  FIG. 7 shows changes in data assigned to each cluster when the DA update width Δβ is changed. The vertical axis indicates the number of data belonging to the cluster. If clustering is performed appropriately, the number of data belonging to the cluster is 2700. That is, when the expression (9) is satisfied, all data converged to the cluster.

Next, the update width Δβ of the data clustering DA necessary for constructing the RBFN in the cluster obtained after the preprocessing will be described. In the preprocessing, all data was classified into 3 clusters. The result of the number of clusters of 3 was the best than the number of clusters of 2 and 4. Results of building RBFN in cluster 1 Δβ ≦ 4 (10)
However, an appropriate result was obtained (see FIGS. 8 and 9). FIG. 8 shows the relationship between Δβ and the evaluation function in DA clustering necessary for RBFN construction in the case of cluster 1. FIG. 9 shows the relationship between Δβ in DA clustering in the case of cluster 1 and the number of data belonging to the cluster.
In the figure, when Δβ ≦ 4, all data converged to one of the clusters. On the other hand, when 4 <Δβ, some data did not converge to any cluster.

Similarly, the appropriate ranges of Δβ in the clusters 2 and 3 are respectively the expressions (11) and (12) (see FIGS. 10, 11, 12, and 13).
Δβ ≦ 4 (11)
Δβ ≦ 20 (12)

Therefore, from the formulas (10), (11), and (12), the update width Δβ of the data clustering DA necessary for constructing the RBFN in the cluster obtained after the preprocessing is
Δβ ≦ 4
Is desirable.

Next, in order to verify the effect of the photovoltaic power generation prediction apparatus according to the present embodiment, a result of simulation is described. For the determination of the center vector of RBFN, k-RBFN using k-means, which is a conventional clustering technique, and D-RBFN using DA clustering are compared. As a preprocessing method, k-means and DA clustering are compared, and the following method is defined here in order to examine the effectiveness of the method according to the present embodiment.
Method A: MLP
Method B: k-RBFN (determination of the center vector: k-means, no preprocessing)
Method C: k-means-k-RBFN (determination of the center vector: k-means, preprocessing: k-means)
Method D: k-means-D-RBFN (determination of center vector: DA clustering, preprocessing: k-means)
Method E: DA-D-RBFN (determination of center vector: DA clustering, preprocessing: DA clustering)

The simulation conditions are as follows.
(1) The generation output prediction after 30 minutes is performed using the time series data of the past photovoltaic power generation and the panel temperature every minute. At the same time, the solar power generation and the panel temperature are calculated for 5 minutes of dispersion, 10 minutes of dispersion, difference, and second-order difference, respectively, and these are used as ANN input data, so the number of input layers is 10 dimensions. . In addition, Table 1 shows the number of clusters MLP and RBFN, the learning coefficient α, the moment coefficient η, the moment increment value Δη, the number of intermediate layers, and the number of learnings. The values of these parameters were determined by prior simulation.

(2) As data used here, May 2nd to 4th of 2005 is used as learning data, and 5th is used as test data. During this period, since there was no power generation during the night from 20:00 to 5:00, power generation prediction from 5:00 to 20:00 is performed excluding that period. Therefore, there are 3700 learning patterns since there are 3 days of 15-hour data for every minute. The clustering data is 10 dimensions shown in (1) and 11 dimensions obtained by adding the power generation output after 30 minutes to the vector. Table 2 shows the DA clustering parameters determined by the preliminary simulation.

  (3) The center vector of clustering is applied to the determination of the center of the RBFN basis function, and the width is the distance from the center vector to the farthest point included in the cluster.

The results of simulation under the above simulation conditions are as follows. In order to show the effectiveness of k-means and DA clustering, Table 3 shows the best value, average value, worst value, and standard deviation of evaluation functions for 100 initial values.

  The evaluation functions for k-means and DA clustering are given by equations (3) and (4), respectively. Comparing the evaluation functions of k-means and DA clustering, the same value is obtained when the number of clusters is two. On the other hand, when the number of clusters is 3 or 4, DA clustering has obtained good results in the best value, average value, worst value, and standard deviation. Although there is only a slight difference with respect to the best value, the average value, the worst value, and the standard deviation are greatly improved. Regarding the standard deviation of DA clustering, the number of clusters was 0 in any case. This indicates that DA clustering performs global clustering. Therefore, it was confirmed that DA clustering is effective in the clustering method.

Table 4 shows the average value (Ave.Error), maximum value (MAX Error), standard deviation (Ab.Error) of the absolute value error between the predicted output and the teacher data of the power generation amount after 30 minutes obtained from Method A by Method E. Standard Deviation). In Table 4, [pu] means per unit. For example, Method A of Ave. Error of 6.3435 × 10 ⁻² indicates that the average error from the teacher data is 6.3435%. The values in parentheses in each column of the average value, the maximum value, and the standard deviation of the absolute value error are the Method A to the Method A when the average value, the maximum value, and the standard deviation of the method A are 1, respectively. The relative value of E is shown. According to Table 4, when Method B (k-RBFN) and Method C (k-means-k-RBFN) are compared, the average value, maximum value, and standard deviation of the absolute value error from Method B teacher data are 5. 2238 × 10 ⁻² , 2.9534 × 10 ⁻¹ , 5.6678 × 10 ⁻² , and when the number of clusters of Method C is 3, the average value, the maximum value, and the standard deviation of the absolute value error from the teacher data Are 4.8989 × 10 ⁻² , 2.7838 × 10 ⁻¹ , 4.7521 × 10 ⁻² , and Method D has an average value of absolute value error, a maximum value, and a standard deviation of 6 respectively. .2196%, 5.7425%, 16.156% Good results are obtained, and the prediction accuracy is better when the preprocessing by clustering is used for the input data. Hunt. This is because ANN has the feature that the more accurate the output is obtained as the variance of the input data is smaller, the ANN obtained the input / output data having a stronger correlation by extracting similar data by clustering. It is.

When the number of clusters of Method D (k-means-D-RBFN) and Method E (DA-D-RBFN) is 3 respectively, the average value and the maximum value of the absolute value error with Method D teacher data The standard deviations are 4.6636 × 10 ⁻² , 2.7687 × 10 ⁻¹ , 4.6388 × 10 ⁻² , respectively, and the average value, the maximum value, and the standard deviation of the absolute value error from the Method E teacher data Are 4.3892 × 10 ⁻² , 2.7547 × 10 ⁻¹ , and 4.5734 × 10 ⁻² , respectively, and Method E is 5 in terms of the average value, the maximum value, and the standard deviation of the absolute value error. Good results were obtained: 8839%, 0.5057%, 1.4098%. In addition, when Method D and Method E each have 4 clusters, the average value, maximum value, and standard deviation of absolute value errors from Method D teacher data are 4.8733 × 10 ⁻² and 2.877 × 10 respectively. ⁻¹ , 4.8562 × 10 ⁻² , and the average value, the maximum value, and the standard deviation of the absolute value error from Method E teacher data are 4.6500 × 10 ⁻² , 2.7576 × 10 ⁻¹ , respectively. Since it is 4.6866 × 10 ⁻² , Method E is 4.5821%, 3.5028%, 3.4924 in terms of average value, maximum value, and standard deviation of absolute value error even when the number of clusters is 4. % Good results. This indicates that it is more effective to adopt DA clustering than k-means as a method for preprocessing input data in the present invention. Moreover, the best results are obtained when the number of clusters is 3 among Methods C to Method E, which are pre-processed methods. The reason why the number of clusters is 3 is superior to the case where the number of clusters is 4, because the more similar data gathers when the number of clusters is increased, but learning becomes difficult due to the decrease of the learning data. is there. That is, when clustering is performed as preprocessing for input data, high-precision prediction can be performed by using an appropriate number of clusters.

Next, when the number of clusters of Method C (k-means-k-RBFN) and Method D (k-means-D-RBFN) is 3 respectively, the absolute value error of Method C teaching data is compared. The average value, the maximum value, and the standard deviation are 4.889 × 10 ⁻² , 2.7838 × 10 ⁻¹ , 4.7521 × 10 ⁻² , respectively, and the average value of the absolute value error with the Method D teacher data, Since the maximum value and the standard deviation are 4.6636 × 10 ⁻² , 2.7687 × 10 ⁻¹ , 4.6388 × 10 ⁻² respectively, Method D is the average value, the maximum value, and the absolute value error. Good results were obtained with standard deviations of 4.8031%, 0.5424%, and 2.3842%, respectively. That is, the prediction error is smaller when DA clustering is used to determine the center of the RBFN basis function. This is because D-RBFN has a global structure, whereas k-RBFN has a local structure.

  As described above, in the present invention, it is most effective to use DA clustering for the determination of the center vector of RBFN, and to use DA clustering as a preprocessing method, and Method E (DA-D-) with 3 clusters. In the case of RBFN), the average error is improved by 30.808%, the maximum error is improved by 16.3233%, and the standard deviation is improved by 29.323% compared to the conventional ANN method Method A (MLP). Compared to Method B (k-RBFN), which is an RBFN method, the average error was improved by 15.977%, the maximum error was 6.7270%, and the standard deviation was improved by 19.309%.

  Here, FIG. 14 shows the prediction result output and the actual measurement value when Method E with 3 clusters is used, and FIG. 15 shows the transition of the error at each time. 14 and 15, since the data used is data every minute, the horizontal axis represents from 5 o'clock to 20 o'clock as 0 minute to 900 minutes. 14 and 15, it is easy to approximate when the measured value is stably generated from 0 to 400 minutes, but the measured value fluctuates from 500 to 700 minutes, and the error at that time is large. You can see that Therefore, the maximum error has become large. However, even in the maximum error, the method according to the present embodiment has obtained better results than the comparative method, and it has been confirmed that the method is effective in the method of photovoltaic power generation prediction.

  As described above, in photovoltaic power generation prediction, high-precision clustering is performed on input data using DA-clustering, and D-RBFN using DA clustering is also used for determining the center of the basis function of RBFN. By using it, it is possible to predict the generated power of the solar power generation device with high accuracy.

  A program for realizing the function of the processing unit in FIG. 1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed, thereby executing solar power generation prediction. Processing may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  It can be applied to applications in which it is essential to use a neural network to predict the power generated by a solar power generation device.

  DESCRIPTION OF SYMBOLS 1 ... Solar power generation prediction apparatus, 11 ... Data input part, 12 ... Clustering process part, 13 ... Model construction part, 14 ... Prediction process part, 2 ... Input part, 3 ... Display section

Claims (5)

A solar power generation prediction device for predicting the amount of power generated by solar power generation,
Data input means for inputting data necessary for power generation amount prediction;
Clustering processing means for classifying the input data by clustering in order to obtain neural network learning data;
A model construction means for constructing a neural network model for each clustered cluster and classifying the cluster into a cluster having a cluster representative closest to the test data;
A photovoltaic power generation prediction apparatus comprising: a prediction unit that predicts a power generation amount of solar power generation using a neural network in a cluster into which the test data is classified, and outputs the predicted power generation amount.   The photovoltaic power generation prediction apparatus according to claim 1, wherein the clustering processing unit performs clustering using deterministic annealing.   The photovoltaic power generation prediction apparatus according to claim 1, wherein the neural network model uses a radial basis function network. A photovoltaic power generation prediction method for predicting the amount of solar power generation,
A data input step for inputting data necessary for power generation amount prediction;
A clustering step for classifying the input data by clustering to obtain neural network learning data;
A model construction step of constructing a neural network model for each clustered cluster and classifying the cluster into a cluster having a cluster representative closest to the test data;
A prediction step of predicting a power generation amount of solar power generation using a neural network in a cluster into which the test data is classified, and outputting the predicted power generation amount. A photovoltaic power generation prediction program for causing a computer on a photovoltaic power generation prediction device that predicts the amount of photovoltaic power generation to predict the amount of photovoltaic power generation,
A data input step for inputting data necessary for power generation amount prediction;
A clustering step for classifying the input data by clustering to obtain neural network learning data;
A model construction step of constructing a neural network model for each clustered cluster and classifying the cluster into a cluster having a cluster representative closest to the test data;
A photovoltaic power generation prediction program, comprising: predicting a power generation amount of solar power generation using a neural network in a cluster into which the test data is classified, and outputting a predicted power generation amount.

Images