Analytical device, analytical method, and analytical program

Abstract

To provide an analysis apparatus for efficiently analyzing error patterns on output prediction, an analysis method, and an analysis program.SOLUTION: A patten extraction unit 13 calculates time-series first errors between actual values and prediction values for first output amounts output from an energy generation apparatus for every multiple days in the past, and extracts multiple error patterns which represent representative temporal change pattern of errors from the calculated first errors. A label management unit 14 calculates time-series second errors between actual values and prediction values for second output amount output from the energy generation apparatus for each day in an analysis target period, specifies error patterns corresponding to temporal change of the second errors for each day, and allocates the specified error patterns to the days in the analysis target period. An analysis unit 16 analyzes errors between the actual values and the prediction values for the second output amount in the analysis target period, using the error patterns allocated by the label management unit 14 to each day in the analysis target period.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an analysis device, an analysis method, and an analysis program. [Background technology]

[0002] As awareness of environmental issues grows worldwide, the amount of power generation methods derived from natural energy sources, such as photovoltaic (PV), has increased rapidly in recent years, and is expected to continue to expand in the future. However, because PV output is significantly affected by the season and weather, it is important to predict PV output with high accuracy in advance, secure adjustment reserve capacity, and maintain the supply and demand balance in preparation for fluctuations in PV output and prediction errors. Research on PV output prediction has been conducted worldwide to date.

[0003] Conventional PV output forecasts are mostly based on hourly forecasts, and the main forecasting methods are numerical weather forecasts that predict meteorological phenomena, including solar radiation, from current weather values, and machine learning based on the forecast results to predict solar radiation and PV output.

[0004] However, compared to conventional power demand forecasts for distribution systems, the error in PV output forecasts is large. This is because predicting the amount of solar radiation, which is the main factor that determines PV output, is more difficult than predicting other meteorological factors such as temperature. Therefore, in order to ensure an appropriate reserve capacity, it is important not only to improve the accuracy of the forecast, but also to analyze the PV output forecast error and understand its characteristics.

[0005] Research into PV output forecast errors includes technology that targets solar radiation forecasts using the Japan Meteorological Agency's numerical weather forecast model, and technology that spatially analyzes forecast errors in multiple wide-area areas. The former evaluates error trends on a daily basis, and it has been pointed out that, for example, the global model has a tendency to overpredict throughout the year. This suggests the importance of securing reserve capacity that is greater than the average error. The latter compares seasonal trends in over / underprediction in five areas, and shows the results of error analysis along the spatial axis, for example, that while the tendency for underprediction is stronger in the summer, the tendency for underprediction in July is not as strong in only some areas.

[0006] In addition, the following technologies have been proposed as solar radiation intensity prediction technologies. For example, there is a technology that uses machine learning to map solar radiation intensity at multiple points onto a two-dimensional plane, identifies nodes on the two-dimensional plane that correspond to each node to which time-series data is input, and predicts solar radiation intensity at a prediction target point based on the similarity between the identified nodes. Similarly, there is a technology that identifies nodes corresponding to each node to which time-series data is input from nodes on a two-dimensional plane onto which solar radiation intensity at multiple points is mapped, classifies change patterns into groups based on the similarity between the identified nodes, and extracts a representative of each group. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] JP 2017-146126 A [Patent Document 2] JP 2017-146674 A [Non-patent literature]

[0008] [Non-Patent Document 1] Hideaki Ohtake, Kenichi Shimose, Joao Gari da Silva Fonseca Junior, Takumi Takashima, Takashi Ozeki, Yoshinori Yamada, “Error evaluation of solar radiation prediction of the weekly forecast model of the Japan Meteorological Agency”, Journal of Electrical Engineering B, 2014, Vol.134, No.6, pp.501-509, [Non-Patent Document 2] Kenichi Shimose, Hideaki Ohtake, Joao Gari da Silva Fonseca Junior, Takumi Takashima, Takashi Ozeki, Yoshinori Yamada, "Analysis of the factors of solar radiation forecast error in the mesoscale model of the Japan Meteorological Agency", Journal of Electrical Engineering B, 2014, Vol. 134, No. 6, pp. 518-526 [Non-Patent Document 3] JGS Fonseca, Jr., T. Oozeki, H. Ohtake, K. Shimose, T. Takashima, K. Ogimoto, “Characterizing the Regional Photovoltaic Power Forecast Error in Japan:a Study of 5 Regions”, IEEJ Trans. PE,2014,Vol.134,No.6,pp.537-544 Summary of the Invention [Problem to be solved by the invention]

[0009] However, conventional research on PV output forecast errors is limited to a bird's-eye view of the PV output forecast error. For example, among days with a tendency toward over-prediction, there are days when the forecast is slightly over-predicted throughout the day, and days when the forecast is significantly over-estimated for only a few hours in the afternoon. Even if the total error for the entire day is about the same, the appropriate way to prepare for the error may be different for both of these days. Thus, in order to optimize the reserve capacity, a detailed error analysis is required that takes into account the specific time-varying pattern of the forecast error during the day, that is, the "error pattern." However, conventional analyses are limited to evaluating the error on a daily or monthly basis, and no analysis of the error pattern is performed, and it is also difficult to efficiently perform error pattern analysis using conventional methods.

[0010] In addition, a technology for predicting solar radiation intensity based on the similarity of nodes generated by mapping solar radiation intensity onto a two-dimensional plane, which correspond to each node to which time-series data is input, is a technology for improving the accuracy of prediction. In addition, a technology for classifying patterns of change in solar radiation intensity using similarities calculated in a similar manner is an information aggregation technology aimed at improving the calculation efficiency of simulations that use prediction results. Therefore, although both technologies focus on patterns, they cannot be used for the purpose of analyzing prediction errors in detail and clarifying their characteristics.

[0011] The disclosed technology has been made in consideration of the above, and aims to provide an analysis device, an analysis method, and an analysis program that efficiently perform error pattern analysis related to output prediction. [Means for solving the problem]

[0012] In one embodiment of the analysis device, analysis method, and analysis program disclosed in the present application, the pattern extraction unit calculates a first error in a time series between actual measured values ​​and predicted values ​​of a first output amount from an energy generating device for each of a plurality of days in the past, and extracts a plurality of error patterns representing a representative time change pattern of the error from the calculated first error. The label management unit obtains a second error in a time series between actual measured values ​​and predicted values ​​of a second output amount from the energy generating device for each day of an analysis target period, identifies the error pattern corresponding to the time change of the second error for each day, and assigns the identified error pattern to each day of the analysis target period. The analysis unit analyzes the error between the actual measured values ​​and predicted values ​​of the second output amount in the analysis target period using the error pattern assigned to each day of the analysis target period by the label management unit. Effect of the Invention

[0013] In one aspect, the present invention is capable of efficiently performing error analysis regarding output amount prediction. [Brief description of the drawings]

[0014] [Figure 1]FIG. 1 is a block diagram of a prediction error analysis device according to an embodiment. [Diagram 2] FIG. 2 is a diagram showing an overview of the prediction error analysis. [Diagram 3] FIG. 3 is a diagram showing an example of a clustering result into two clusters. [Figure 4] FIG. 4 is a diagram showing the relationship between the actual measured value and the predicted value of each PV output when classified into two clusters. [Diagram 5] Figure 5 shows the proportion of clusters for each month when classified into two clusters. [Figure 6] FIG. 6 is a diagram showing an example of the clustering result into five clusters. [Figure 7] FIG. 7 is a diagram showing the relationship between the actual measured value and the predicted value of each PV output when classified into five clusters. [Figure 8] FIG. 8 shows the proportion of clusters for each month when classified into five clusters. [Figure 9] FIG. 9 is a flowchart of a prediction error analysis process performed by the prediction error analysis device according to the embodiment. [Figure 10] FIG. 10 is a diagram showing the hardware configuration of the prediction error analysis device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Hereinafter, examples of the analytical device, analytical method, and analytical program disclosed in the present application will be described in detail with reference to the drawings. Note that the analytical device, analytical method, and analytical program disclosed in the present application are not limited to the following examples. EXAMPLES

[0016] 1 is a block diagram of a prediction error analysis device according to an embodiment of the present invention. The prediction error analysis device 1 according to the present embodiment is connected to an output amount database 2, a forecast value database 3, and a management terminal device 4.

[0017] The output amount database 2 holds past information on the actual measured value of the PV output output from the PV, which is the energy generating device to be predicted. The actual measured value of the PV output is time-series data. In other words, the actual measured value of the PV output is the actual measured value of the PV output for each hour of the day.

[0018] The predicted value database 3 holds the prediction results of a PV output prediction device (not shown) for the output amount of the PV to be predicted. The prediction results are time-series data corresponding to each PV output stored in the output amount database 2.

[0019] The management terminal device 4 is a terminal used by a manager of the prediction error analysis device 1. The manager may be a technician who analyzes the prediction errors.

[0020] Next, a detailed description will be given of the prediction error analysis device 1. As shown in Fig. 1, the prediction error analysis device 1 has an information acquisition unit 11, a vectorization unit 12, a pattern extraction unit 13, a label management unit 14, a data storage unit 15 and an analysis unit 16.

[0021] The information acquiring unit 11 acquires actual values ​​of past PV output from the output amount database 2. It is preferable that the actual values ​​of past PV output acquired by the information acquiring unit 11 include all patterns of prediction errors of PV output occurring throughout one year as prediction errors, which are errors from corresponding predicted values. For example, the information acquiring unit 11 may acquire actual values ​​of PV output for each day of the past year, or may acquire actual values ​​of PV output for each day over multiple years. Furthermore, if the patterns for each day can be estimated to some extent, the information acquiring unit 11 may extract several days that are estimated to include different patterns in one year, and acquire actual values ​​of PV output for each extracted day.

[0022] Furthermore, the information acquiring unit 11 acquires, from the predicted value database 3, predicted values ​​of the PV output, which are prediction results corresponding to the respective acquired actual measured values ​​of the PV output. Next, the information acquiring unit 11 associates each actual measured value of the PV output with each predicted value of the PV output. Then, the information acquiring unit 11 outputs information including the associated actual measured value and predicted value of the PV output for each day to the vectorizing unit 12. The actual measured value and predicted value of the PV output acquired by the information acquiring unit 11 are an example of "the actual measured value and predicted value for the first output amount".

[0023] The vectorization unit 12 receives an input of information including the associated actual measured value of PV output and the predicted value of PV output for each day from the information acquisition unit 11. Then, the vectorization unit 12 calculates the error between the actual measured value of PV output and the predicted value of PV output for each day. Here, the vectorization unit 12 calculates time series data representing the error for each hour of each day using the following formula (1) as the error for each day.

[0024]

number

[0025] Here, err on the left side of equation (1) h d represents the normalized error value for each time of each day. The first term in the numerator on the right side of formula (1) represents the predicted PV output at time h on the dth day when the days are arranged in chronological order. The second term is the actual PV output at time h on the same dth day. The Y in the denominator on the right side d h,max is the maximum value of the actual measured value of the PV output at time h during one week before and after the dth day, and is expressed by the following formula (2).

[0026]

number

[0027] That is, the vectorization unit 12 according to this embodiment normalizes the actual error between the actual measurement value and the predicted value by the maximum value of the actual measurement value of the PV output in the preceding and following one week, and sets the normalized value as the error between the actual measurement value and the predicted value. This time series error is an example of the "first error".

[0028] Fig. 2 is a diagram showing an overview of the prediction error analysis. The vectorization unit 12 calculates the prediction error 101 for each day including, for example, Day#1 to Day#3 in Fig. 2 using Equation (1). Fig. 2 shows an example in which the PV output was measured between 5:00 and 19:00 on each day.

[0029] Next, the vectorization unit 12 arranges the errors for each day, which are time-series data, by time and vectorizes them to generate a vector representing the prediction error for each hour of each day. Hereinafter, the vector representing the prediction error for each hour is referred to as the "error vector." For example, the vectorization unit 12 vectorizes the prediction error 101 for each day in FIG. 2 to generate a vector 102 of the errors for each day. Then, the vectorization unit 12 outputs the generated vector of the errors for each day to the pattern extraction unit 13.

[0030] The pattern extraction unit 13 receives an input of the vector of the error for each day from the vectorization unit 12. Next, from the acquired vector data, the pattern extraction unit 13 extracts data for the period from 8:00 to 15:00, which is an hour's margin both before and after the period from 7:00 to 16:00, when the error of formula (1) has a stable value on average throughout the year.

[0031] Moreover, the pattern extraction unit 13 receives an input of information specifying how many clusters to generate when clustering is performed from the management terminal device 4. Here, the number of clusters to be generated can be specified as long as it is two or more. The pattern extraction unit 13 performs clustering by classifying all vectors representing the prediction errors for each time between 8:00 and 15:00 on each day into the specified number of clusters. For example, the pattern extraction unit 13 uses k-means clustering as a clustering method. Furthermore, in k-means clustering, the pattern extraction unit 13 uses a metric based on cos similarity that can emphasize the tendency of fluctuation as a distance metric. Thereafter, the pattern extraction unit 13 outputs information on each cluster generated as information on an error pattern to the label management unit 14 and the data storage unit 15. For example, the average vector of the error vectors belonging to the cluster can be used as the information on the error pattern.

[0032] For example, Fig. 2 shows a case where the prediction is classified into two types of clusters. The pattern extraction unit 13 can classify the error vectors of each day into clusters C1 and C2 by clustering the vectors 102 in Fig. 2. In this case, cluster C1 is a pattern of "all-day excessive" in which the prediction results are excessive compared to the actual PV output throughout the day. Cluster C2 is a pattern of "morning under-small / afternoon over-large" in which the predicted value of the PV output is excessively small compared to the actual measured value of the PV output in the morning and the predicted value of the PV output is excessive compared to the actual measured value of the PV output in the afternoon.

[0033] The label management unit 14 receives an input of information on each generated cluster from the pattern extraction unit 13. Next, the label management unit 14 acquires the actual measured value of the PV output for each day of the one year period to be analyzed from the output amount database 2. The label management unit 14 also acquires the predicted value of the PV output for each day of the one year period to be analyzed from the predicted value database 3. The actual measured value and predicted value of the PV output for each day of the one year acquired by the label management unit 14 are an example of the "actual measured value and predicted value for the second output amount". Here, in this embodiment, the label management unit 14 acquires information from the output amount database 2 and the predicted value database 3, but if the information acquisition unit 11 acquires the actual measured value and predicted value of the PV output for the analysis period, for example, the label management unit 14 may acquire the information.

[0034] Next, the label management unit 14 obtains the time series error between the actual measured value of the PV output and the predicted value of the PV output for each hour of each day of the one year period to be analyzed. This error is an example of the "second error". It is preferable to calculate the second error in the same manner as the first error. Then, the label management unit 14 identifies the closest error pattern among the multiple error patterns extracted by the pattern extraction unit 13 to the pattern of the time change of the error between the actual measured value of the PV output and the predicted value of the PV output for each hour of each day. Then, the label management unit 14 assigns a label representing a cluster corresponding to the identified error pattern to each day of the one year period to be analyzed. Then, the label management unit 14 stores information in which a label is assigned to each day of the one year in the data storage unit 15. Here, in this embodiment, the label management unit 14 uses information of one year as the period to be analyzed, but is not limited to this, and may use information of multiple years.

[0035] For example, when classified into two clusters as shown in FIG. 2, the label management unit 14 assigns the label of cluster C1 to Day #1 and #3, and the label of cluster C2 to Day #2, as shown in table 103. The label management unit 14 also assigns the label of either cluster C1 or C2 to other days. Then, the label management unit 14 stores the generated table 103 in the data storage unit 15. Furthermore, the label management unit 14 stores information of each cluster, such as an error pattern, in the data storage unit 15.

[0036] The data storage unit 15 is a storage device. The data storage unit 15 receives and stores information on each generated cluster from the pattern extraction unit 13. The data storage unit 15 also stores information generated by the label management unit 14 in which a label is assigned to each day of the year to be analyzed.

[0037] The analysis unit 16 acquires information in which a label is assigned to each day of the one-year period to be analyzed from the data storage unit 15. Furthermore, the analysis unit 16 acquires information on the error pattern from the data storage unit 15. The analysis unit 16 then performs analysis using the information in which a label is assigned to each day of the one-year period to be analyzed and the information on each error pattern. For example, the analysis unit 16 determines the content rate of the error pattern in each month, etc. Then, the analysis unit 16 analyzes seasonal changes in the error pattern and weather conditions according to the error pattern, etc., using the content rate of the error pattern in each month, etc. After that, the analysis unit 16 notifies the administrator of the analysis result by, for example, transmitting the analysis result to the management terminal device 4.

[0038] Next, a specific example of the forecast error analysis by the forecast error analysis device 1 will be described. Here, the analysis period was set to one year. In addition, the k-means method using a distance metric based on cosine similarity was adopted as the clustering method. In addition, the number of clusters was changed to 2, 3, 4, and 5. In addition, the following conditions were used as the PV output forecast model. As the forecast method, a forecast model was created for each hour using Support Vector Regression (SVR). In addition, as the explanatory variables, the Meso Scale Model (MSM) of the Japan Meteorological Agency was used, and the initial time was set to 6:00, and the amount of solar radiation at a single point in Kumagaya City, Saitama Prefecture was used. The forecast target time was set to 5:00 to 19:00 on the following day. In addition, the learning period was set to one year from the previous year.

[0039] First, an example of classification into two clusters will be described. Fig. 3 is a diagram showing an example of the clustering result into two clusters. Fig. 4 is a diagram showing the relationship between the actual measured value and the predicted value of each PV output when classified into two clusters. Fig. 5 is a diagram showing the cluster ratio for each month when classified into two clusters.

[0040] Through clustering by the pattern extraction unit 13, the error vectors for each day in the one-year period under analysis are classified into clusters C1 and C2 as shown in Fig. 3. In each graph in Fig. 3, the horizontal axis represents time and the vertical axis represents error.

[0041] For each of these two clusters, the average measured PV output value and the average predicted value are calculated, and their relationship over time is plotted as shown in the two graphs in Figure 4. In the graph in Figure 4, the horizontal axis represents time and the vertical axis represents PV output. Additionally, the solid line represents the predicted value, and the dashed line represents the measured value.

[0042] As shown in the graph of cluster C1 in FIG. 4, in cluster C1, the predicted values ​​exceed the actual measured values ​​throughout the day. In other words, cluster C1 can be said to be a group of daily excessive error patterns. Also, as shown in the graph of cluster C2 in FIG. 4, in cluster C2, the predicted values ​​fall below the actual measured values ​​throughout the day. In other words, cluster C2 can be said to be a group of daily undersized error patterns. In this case, the pattern extraction unit 13 classifies the error vectors of each day into two clusters: a daily excessive error pattern and a daily undersized error pattern.

[0043] The label management unit 14 labels each day to be analyzed according to whether it is included in cluster C1 or cluster C2. The analysis unit 16 then calculates the inclusion rate of each label for each month. In this case, the inclusion rate of each label for each month is shown in FIG. 5.

[0044] For example, the analysis unit 16 can analyze that there is a tendency for under-prediction to occur throughout the year, and that there is a tendency for the rate of over-prediction to be high in June and October during the rainy season and typhoon season, based on Fig. 5. In this way, it can be confirmed that an analysis that is not inconsistent with already known knowledge is possible.

[0045] Next, a case where the number of clusters is changed from 3 to 5 in order will be described. FIG. 6 is a diagram showing an example of the clustering result into 5 clusters. FIG. 7 is a diagram showing the relationship between the actual measured value and the predicted value of each PV output when classified into 5 clusters. FIG. 8 is a diagram showing the proportion of clusters for each month when classified into 5 clusters.

[0046] When the number of clusters is 3, the vector of the error for each day in the one year period to be analyzed is classified into clusters C1, C3, and C5 in FIG. 6 by clustering by the pattern extraction unit 13. In the graph of FIG. 6, the horizontal axis represents time, and the vertical axis represents error. When the number of clusters is 3, the clusters to which the vector representing the error for each day in the one year period to be analyzed is classified are increased to cluster C2 in addition to clusters C1, C3, and C5 when the number of clusters is 3. When the number of clusters is 5, the vector representing the error for each day in the one year period to be analyzed is classified into clusters C1 to C5 by clustering by the pattern extraction unit 13.

[0047] For each of the five clusters C1 to C5, the average value of the actual measured PV output and the average value of the predicted PV output are calculated, and their relationships over time are plotted as shown in the five graphs in Figure 7. In the graphs in Figure 7, the horizontal axis represents time and the vertical axis represents PV output. The solid line represents the predicted value, and the dashed line represents the actual measured value.

[0048] Clusters C1 and C5 correspond to the two clusters, and are a daily over-large error pattern and a daily under-small error pattern. Furthermore, in cluster C3, which is increased when the number of clusters is 3, the predicted value is lower than the actual measurement value in the morning. That is, cluster C3 can be said to be a group of morning under-small error patterns. Furthermore, in cluster C2, which is increased when the number of clusters is 3, the predicted value is higher than the actual measurement value in the afternoon. That is, cluster C2 can be said to be a group of afternoon over-large error patterns. Furthermore, in cluster C4, which is increased when the number of clusters is 5, the predicted value is lower than the actual measurement value in the afternoon. That is, cluster C5 can be said to be a group of afternoon under-small error patterns. That is, in this case, the pattern extraction unit 13 classifies the vectors representing the errors into five clusters, a daily over-large error pattern, an afternoon over-large error pattern, a morning under-small error pattern, an afternoon under-small error pattern, and a daily under-small error pattern.

[0049] The label management unit 14 labels each day in the analysis period according to which of the clusters C1 to C5 the day is included in. The analysis unit 16 then calculates the inclusion rate of each label for each month. In this case, the inclusion rate of each label for each month is shown in FIG. 8.

[0050] For example, the analysis unit 16 can analyze based on Fig. 8 that error patterns of morning under-prediction and afternoon over-prediction occur frequently from October to January. In addition, the analysis unit 16 can analyze based on Fig. 8 that error patterns of afternoon over-prediction occur more frequently during the high temperature period from July to September, and that afternoon under-prediction increases during the short sunshine period from October to January. In this way, by increasing the number of clusters, the number of types of classified error patterns increases, knowledge about the error between the actual measured value and the predicted value becomes more detailed, and it can be confirmed that seasonality also appears prominently.

[0051] Here, in this embodiment, the analysis unit 16 performed the analysis based on the ratio of the error pattern for each day in a specified period, but this is not limiting. By using a similar method, it is possible to efficiently clarify the relationship between meteorological elements such as weather, temperature, and air pressure, and the error between the actual measured value and the predicted value.

[0052] In this case, for example, the analysis unit 16 may transmit a graph showing the content rate of the error pattern for each month shown in Fig. 5 or 8 as the analysis result to the management terminal device 4 and notify the administrator. Furthermore, the analysis unit 16 may further add the analysis result to the graph shown in Fig. 5 or 8 and transmit it to the management terminal device 4.

[0053] 9 is a flowchart of the prediction error analysis process by the prediction error analysis device according to the embodiment. Next, the flow of the prediction error analysis process by the prediction error analysis device 1 according to the embodiment will be described with reference to FIG.

[0054] The information acquiring unit 11 acquires past actual measured values ​​of PV output from the output amount database 2. The information acquiring unit 11 also acquires predicted values ​​corresponding to each of the acquired PV outputs from the predicted value database 3 (step S1). Next, the information acquiring unit 11 associates the actual measured values ​​of each PV output with the predicted values ​​of each PV output. Then, the information acquiring unit 11 outputs information including the associated actual measured values ​​and predicted values ​​of PV output for each day to the vectorizing unit 12.

[0055] The vectorization unit 12 receives information including the actual measured value and the predicted value of the PV output associated with each day from the information acquisition unit 11. Then, the vectorization unit 12 calculates the error between the actual measured value of the PV output and the predicted value of the PV output. Next, the vectorization unit 12 vectorizes the calculated error to generate a vector of the error for each day (step S2). Thereafter, the vectorization unit 12 outputs the generated vector of the error for each day to the pattern extraction unit 13.

[0056] The pattern extraction unit 13 receives an input of the error vector for each day from the vectorization unit 12. Next, the pattern extraction unit 13 performs clustering using all of the error vectors for each day, and classifies them into a specified number of clusters. Then, the pattern extraction unit 13 calculates the average of the error vectors for each day that belong to each cluster, and extracts information on the error pattern (step S3). Thereafter, the pattern extraction unit 13 outputs the information on the error pattern to the label management unit 14 and the data storage unit 15.

[0057] The label management unit 14 receives an input of information on the error pattern from the pattern extraction unit 13. Next, the label management unit 14 acquires information on the actual measured value of the PV output for the one year to be analyzed from the output amount database 2. Furthermore, the pattern extraction unit 13 acquires the prediction results corresponding to each of the acquired PV outputs from the predicted value database 3. Next, the label management unit 14 calculates the error between the actual measured value and the predicted value of the PV output for the one year to be analyzed for each day. Then, the label management unit 14 identifies an error pattern that is closest to the pattern of the change in the error between the actual measured value and the predicted value of the PV output for each day for the one year to be analyzed, and performs labeling to represent the identified error pattern for each day (step S4). After that, the label management unit 14 stores information in which a label is assigned to each day for the one year to be analyzed in the data storage unit 15.

[0058] The analysis unit 16 acquires information in which a label is assigned to each day of the year to be analyzed from the data storage unit 15. Then, the analysis unit 16 obtains the content rate of the error pattern in each month, etc. After that, the analysis unit 16 performs an analysis using the content rate of the error pattern in each month, etc. (Step S5).

[0059] Thereafter, the analysis unit 16 notifies the administrator of the analysis result by, for example, transmitting the analysis result to the management terminal device 4 (step S6).

[0060] As described above, the prediction error analysis device according to the present embodiment vectorizes the prediction errors, which are time-series data, and performs clustering to extract error patterns. The prediction error analysis device then determines which of the extracted error patterns corresponds to the pattern of change in error for each hour of the day being analyzed, and uses the determination result to analyze the prediction errors.

[0061] By increasing the number of clusters to be classified, it is possible to classify the forecast errors for each day of the year in more detail, making it easier to check seasonal fluctuations in forecast errors. In addition, by reducing the number of clusters to be classified, it becomes easier to grasp the overall picture of forecast errors in a relatively large framework. This makes it possible to perform detailed and efficient analysis of error patterns related to output forecasts.

[0062] In the above explanation, the prediction error of the PV output when a photovoltaic power generation system is used as the energy generating system is taken as an example, but other energy generating systems may be used as long as the output amount varies due to natural influences. For example, it is possible to perform prediction error analysis for the output amount of other natural energy sources such as wind power using a similar configuration.

[0063] (Hardware configuration) Fig. 10 is a hardware configuration diagram of a prediction error analysis device. As shown in Fig. 10, the prediction error analysis device 1 includes, for example, a CPU (Central Processing Unit) 91, a memory 92, a hard disk 93, and a network interface 94. The CPU 91 is connected to the memory 92, the hard disk 93, and the network interface 94 via a bus.

[0064] The network interface 94 is an interface for communication between the prediction error analysis device 1 and an external device. For example, the network interface 94 relays communication between the CPU 91 and the output amount database 2, the prediction value database 3, and the management terminal device 4.

[0065] The hard disk 93 is an auxiliary storage device. The hard disk 93 realizes the function of the data storage unit 15 illustrated in Fig. 1. The hard disk 93 also stores various programs including programs that realize the functions of the information acquisition unit 11, the vectorization unit 12, the pattern extraction unit 13, the label management unit 14, and the analysis unit 16 illustrated in Fig. 1.

[0066] The memory 92 is a main storage device, and is, for example, a direct random access memory (DRAM).

[0067] The CPU 91 reads out various programs from the hard disk 93, expands them in the memory 92, and executes them. In this way, the CPU 91 realizes the functions of the information acquisition unit 11, the vectorization unit 12, the pattern extraction unit 13, the label management unit 14, and the analysis unit 16 illustrated in FIG. [Explanation of symbols]

[0068] 1. Forecast error analyzer 2 Output volume database 3. Prediction database 4. Management terminal 11 Information acquisition department 12 Vectorization Department 13 Pattern Extraction Section 14 Label Management Department 15 Data storage section 16 Analysis Department

Claims

1. a pattern extraction unit that calculates a first error in a time series between an actual measurement value and a predicted value for a first output amount from the energy generating device for each of a plurality of days in the past, and extracts a plurality of error patterns that represent a representative time change pattern of the error from the calculated first error; a label management unit that calculates a second error in a time series between an actual value and a predicted value for a second output amount from the energy generating device for each day of an analysis period, identifies the error pattern corresponding to a time change of the second error for each day, and assigns the identified error pattern to each day of the analysis period; an analysis unit that analyzes an error between an actual measurement value and a predicted value of the second output amount during the analysis period by using the error pattern assigned to each day of the analysis period by the label management unit; An analytical device comprising:

2. The analysis device according to claim 1 , wherein the pattern extraction unit acquires the error patterns appearing as each cluster by applying a clustering technique to the first errors represented by the time series for each of the past multiple days.

3. The analysis apparatus according to claim 2 , wherein the pattern extraction unit classifies the first errors into a designated number of clusters.

4. 4. The analysis device according to claim 3, wherein the pattern extraction unit performs the classification by k-means clustering using a metric based on cosine similarity as a distance metric.

5. 5. The analysis device according to claim 1, wherein the analysis unit calculates a proportion of the error pattern assigned to each day for each predetermined period included in the analysis target period.

6. The analysis device according to any one of claims 1 to 4, characterized in that the analysis unit calculates the proportion of the error pattern assigned to each day during the analysis period for which a value of a specified meteorological element falls within a specified range.

7. 6. The analysis device according to claim 1, wherein the analysis unit analyzes a relationship between the error pattern and meteorological elements.

8. The analysis device according to claim 6 , further comprising a notification unit that notifies a graph showing a ratio of the error pattern assigned to each day for each predetermined period included in the analysis target period.

9. 8. The analyzer according to claim 1, wherein the energy generating device is a solar power generating device.

10. calculating a first error in a time series between an actual measured value and a predicted value for a first output amount from the energy generating device for each of a plurality of days in the past, and extracting a plurality of error patterns representing a representative time change pattern of the error from the calculated first error; determining a second error in a time series between an actual value and a predicted value for a second output amount from the energy generating device for each day of an analysis period, identifying the error pattern corresponding to a time change of the second error for each day, and allocating the identified error pattern to each day of the analysis period; The error pattern assigned to each day of the analysis period is used to analyze the error between the actual measurement value and the predicted value of the second output amount during the analysis period.

13. An analytical method comprising:

11. calculating a first error in a time series between an actual measured value and a predicted value for a first output amount from the energy generating device for each of a plurality of days in the past, and extracting a plurality of error patterns representing a representative time change pattern of the error from the calculated first error; determining a second error in a time series between an actual value and a predicted value for a second output amount from the energy generating device for each day of an analysis period, identifying the error pattern corresponding to a time change of the second error for each day, and allocating the identified error pattern to each day of the analysis period; An analysis of an error between an actual measurement value and a predicted value of the second output amount during the analysis period is performed using the error pattern assigned to each day of the analysis period. An analysis program that causes a computer to execute a process.

Images