Model generation device and model generation method
The model generation device and method efficiently select an appropriate learning model for photovoltaic power generation by using a predetermined optimization method, reducing learning time and improving accuracy.
Patent Information
- Application Number
- JP2024020500
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-02-14
- Publication Date
- 2025-08-26
AI Technical Summary
Existing technologies require significant learning time to find an appropriate learning model for estimating the amount of power generated by photovoltaic power generation.
A model generation device and method that acquires features for estimating solar panel power generation, generates multiple learning models using machine learning algorithms, and selects an appropriate model based on index values derived through a predetermined optimization method, such as the Nelder-Mead method, to reduce learning time.
Reduces the time required to find an appropriate learning model for estimating photovoltaic power generation by efficiently searching through hyperparameter settings and machine learning algorithms.
Smart Images

Figure 2025124438000001_ABST
Abstract
Description
[Technical Field]
[0001] One aspect of the present invention relates to a technology for generating a learning model for estimating the amount of power generated by photovoltaic power generation. [Background technology]
[0002] One technology for predicting the amount of power generated by solar power generation is disclosed in Patent Document 1 below. Patent Document 1 calculates the degree of clearness corresponding to weather forecast data based on the relationship between the degree of clearness and weather stored in a storage means, and calculates the amount of solar radiation at each time by multiplying the degree of clearness obtained as a result of the calculation by the theoretical maximum amount of solar radiation at each time during the forecast time period. Then, the power generation output is calculated by multiplying the amount of solar radiation by information related to conversion efficiency, which indicates the relationship between the amount of solar radiation and power generation output. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2013-164286 Summary of the Invention [Problem to be solved by the invention]
[0004] An object of one aspect of the present invention is to reduce the learning time required to find an appropriate learning model for estimating the amount of power generated by photovoltaic power generation. [Means for solving the problem]
[0005] In order to solve the above problem, a model generation device according to one embodiment of the present invention comprises an acquisition unit that acquires features for estimating the power generation amount of a solar panel, and a learning unit that generates multiple learning models based on the features by executing a machine learning algorithm and derives multiple index values that indicate the quality of each of the multiple generated learning models, and the learning unit searches for a learning model having a better index value based on the multiple index values according to a predetermined optimization method, and selects the learning model having the better index value as an appropriate learning model.
[0006] In addition, in order to solve the above-mentioned problems, a model generation method according to one embodiment of the present invention includes an acquisition process for acquiring features for estimating the power generation amount of a solar panel, and a learning process for generating multiple learning models based on the features by executing a machine learning algorithm and deriving multiple index values indicating the quality of each of the multiple generated learning models, wherein in the learning process, a learning model having a better index value is searched for based on the multiple index values according to a predetermined optimization method, and the learning model having the better index value is selected as an appropriate learning model. [Effects of the Invention]
[0007] According to one aspect of the present invention, it is possible to reduce the learning time required to find an appropriate learning model for estimating the amount of power generated by photovoltaic power generation. [Brief explanation of the drawings]
[0008] [Figure 1] FIG. 1 is a functional block diagram showing a schematic configuration of a model generating device according to a reference embodiment. [Figure 2] 10 is a table showing evaluation results of a plurality of learning models generated by a model generation device according to a reference embodiment. [Figure 3] 10 is a flowchart illustrating a processing procedure of a model generating device according to a reference embodiment. [Figure 4] FIG. 1 is a functional block diagram showing a schematic configuration of a power generation amount estimation device according to a reference embodiment. [Figure 5] 1 is a functional block diagram showing a schematic configuration of a model generating device according to a first embodiment. [Figure 6] 4 is a flowchart illustrating a processing procedure of the model generating device according to the first embodiment. [Figure 7] 4 is a flowchart showing an outline of a main processing flow of a learning device of the model generation device according to the first embodiment. [Figure 8] This is a diagram for explaining the concept of the Nelder-Mead method. [Figure 9] 10 is a flowchart showing a flow of a process for executing optimization. DETAILED DESCRIPTION OF THE INVENTION
[0009] [Reference form] Before describing the learning device 10 of the first embodiment, a learning device 10s will be described as a reference embodiment. For ease of explanation, components having the same functions as those described in the reference embodiment will be denoted by the same reference numerals in the following embodiments, and their descriptions will not be repeated. Also, for the sake of brevity, descriptions of matters similar to those in known technologies will be omitted as appropriate.
[0010] <Configuration example of model generation device 1s> 1 is a functional block diagram showing a schematic configuration of a model generation device 1s according to a reference embodiment. The model generation device 1s includes a learning device 10s, a solar radiation measurement device 12 (acquisition unit), a temperature measurement device 13 (acquisition unit), an RTC (Real Time Clock) 14 (acquisition unit), and a power generation amount measurement device 15 (acquisition unit).
[0011] The solar radiation measuring device 12 is installed near the solar panels included in the PV (Photovoltaic Power Generation) system. It measures the amount of solar radiation around the solar panels (hereinafter referred to as solar radiation) and outputs the measurement result to the learning device 10s. The temperature measuring device 13 is installed near the solar panels. It measures the temperature around the solar panels (hereinafter referred to as temperature) and outputs the measurement result to the learning device 10s. The RTC 14 keeps time and outputs at least the current date and time as time information to the learning device 10s. The time information includes the month, day, hour, minute, and second, but any of these pieces of information may be used. The power generation amount measuring device 15 is installed near the solar panels. It measures the current amount of power generated by the solar panels (hereinafter referred to as solar power generation amount) and outputs the measurement result to the learning device 10s. The amount of solar radiation, temperature, time information, and solar power generation amount are stored as a set in a memory unit (not shown) of the model generation device 1s as performance data DZ. Hereinafter, the performance data DZ will be abbreviated as simply DZ.
[0012] The learning device 10s outputs a trained model 11s based on feature quantities for estimating the amount of solar power generation. The feature quantities include the amount of solar radiation, temperature, and time information. The learning model is constructed using the amount of solar radiation, temperature, and time information as explanatory variables and the amount of solar power generation as a target variable. In other words, the learning device 10s trains the learning model using the combination of the amount of solar radiation, temperature, time information, and amount of solar power generation as training data, and outputs the trained model 11s.
[0013] The processing in the learning device 10s can be roughly divided into a learning phase and a verification phase. In the learning phase, the learning device 10s acquires a DZ. Then, the learning device 10s divides the acquired DZ into training data and verification data.
[0014] In a reference embodiment, the learning device 10s generates multiple learning models based on training data by exhaustively (brute-force) executing each of a predetermined number of types of machine learning algorithms. The predetermined number of types of machine learning algorithms are not particularly limited as long as they can solve a regression task. Examples of machine learning algorithms in the reference embodiment include neural networks (NNs), support vector regressions (SVRs), lasso regressions, and ridge regressions.
[0015] Additionally, the learning device 10s may change each hyperparameter set (hyperparameter setting) of a certain machine learning algorithm. In this case, the learning device 10s executes the same machine learning algorithm using the changed hyperparameter set to generate a further learning model. In this way, the learning device 10s may generate a plurality of learning models by comprehensively applying the hyperparameter sets.
[0016] Next, the verification phase will be described. The learning device 10s derives multiple index values indicating the quality of each of the multiple learning models generated in the learning phase. Specifically, the learning device 10s inputs feature quantities in the verification data into each learning model, causing the learning model to output a predicted value of solar power generation. The learning device 10s then derives an index value for each learning model based on the predicted value and the true value of solar power generation indicated in the verification data.
[0017] The index value is not particularly limited as long as it is used in the field of machine learning to evaluate the prediction accuracy in a regression task. The index value may be any statistical quantity derived based on the mean absolute error (MAE) or the root mean square error (RMSE), etc. In the reference example, the higher the index value, the better the prediction accuracy of the prediction model. As an example, the index value may be derived as a quantity in percentage.
[0018] The learning device 10s selects a best learning model from a plurality of learning models based on the index values of each of the plurality of learning models. The learning device 10s outputs the best learning model as a trained model 11s. For example, the learning device 10s identifies the maximum value (maximum index value) among the plurality of index values. Then, the learning device 10s selects the learning model having the maximum index value as the best learning model.
[0019] The learning device 10s may generate a table (evaluation result table) showing the evaluation results in the verification phase. Table TB in FIG. 2 is an example of the evaluation result table. In TB, one index value is recorded in one-to-one correspondence with one hyperparameter set (e.g., Para1) in one machine learning algorithm.
[0020] In the example of Figure 2, Para1 and Para2 each represent a set of hyperparameters (a series of hyperparameters) applied to a certain machine learning algorithm (e.g., NN). NN Para1: Batch size = 256, Maximum number of training iterations = 1000, ... NN Para2: Batch size = 128, Maximum number of training iterations = 1000, ... LASSO Para1: Regularization coefficient = 1.0, Maximum number of training iterations = 1000, ... LASSO Para2: Regularization coefficient = 0.5, Maximum number of training iterations = 1000, ... RIDGE Para1: Regularization coefficient = 1.0, Maximum number of training iterations = 1000, ... RIDGE Para2: Regularization coefficient = 0.5, Maximum number of training iterations = 1000, ... ·SVR Para1: Kernel type="rbf", Regularization coefficient=1.0, … ·SVR Para2: Kernel type="rbf", Regularization coefficient=0.5, … As will be apparent to those skilled in the art, the number of hyperparameter sets is not limited to two. For example, five hyperparameter sets, Para1 to Para5, may be assigned to each machine learning algorithm.
[0021] In the example of FIG. 2, the index value (81.6%) corresponding to Para1 of NN is the maximum value among the index values in TB (see the hatched cells in TB).
[0022] As will be apparent to those skilled in the art, the method for selecting the best learning model is not limited to the above example. The learning device 10s may select the best learning model from among multiple learning models based on multiple index values. For example, the learning device 10s may derive a statistical value based on multiple index values and select the best learning model based on the statistical value. The maximum index value in the reference embodiment is an example of a statistical value.
[0023] The learning device 10s may be configured to train the learning model using as training data a set of the amount of solar radiation actually measured by the solar radiation measuring device 12, the temperature actually measured by the temperature measuring device 13, and the amount of solar power generation actually measured by the power generation measuring device 15. The learning device 10s may also be configured to train the learning model using as training data a set of the amount of solar radiation actually measured by the solar radiation measuring device 12, time information at the time of measurement output from the RTC 14, and the amount of solar power generation actually measured by the power generation measuring device 15.
[0024] The learning device 10s trains the learning model using, for example, one year's worth of data covering six hours per day (time periods when solar power generation is possible). The amount of solar power generation has the strongest correlation with the amount of solar radiation, but it also correlates with temperature, season, time of day, etc. Therefore, by training the learning model including these data, the prediction accuracy of the trained model can be improved.
[0025] <Processing procedure of the model generation device 1s> 3 is a flowchart for explaining the processing steps of the model generating device 1s. First, the solar radiation measuring device 12 measures the amount of solar radiation and outputs it to the learning device 10s (acquisition step S1). As described above, depending on the type of solar radiation measuring device 12, one of the following is output: global solar radiation on a horizontal surface, global solar radiation on an inclined surface, direct solar radiation on a horizontal surface, or direct solar radiation on an inclined surface.
[0026] Next, the temperature measuring device 13 measures the temperature on the back side of the solar panel and outputs it to the learning device 10s (acquisition step S2). The RTC 14 also acquires current time information and outputs it to the learning device 10s (S3). The power generation amount measuring device 15 also measures the amount of solar power generation and outputs it to the learning device 10s (acquisition step S4).
[0027] Next, the learning device 10s trains a learning model using as training data a set of the amount of solar radiation measured by the solar radiation measuring device 12, the temperature measured by the temperature measuring device 13, the time information at the time of measurement output from the RTC 14, and the amount of solar power generation measured by the power generation measuring device 15. In a reference embodiment, the learning device 10s generates multiple learning models by comprehensively executing each of a predetermined number of types of machine learning algorithms. The learning device 10s selects the best learning model by performing a grid search on the generated multiple learning models (S5). That is, the learning device 10s evaluates the quality of each of the generated multiple learning models and selects the best learning model from the multiple learning models.
[0028] Then, the learning device 10s determines whether or not to end the process (S16). If the process is not to be ended (S16, No), the learning device 10s returns to step S11 and repeats the subsequent processes. If the process is to be ended (S16, Yes), the learning device 10s outputs the trained model 11s and ends the process.
[0029] In addition, one year's worth of actual data (measured solar radiation, measured temperature, time information at the time of measurement, measured solar power generation amount) may be acquired and stored, and the learning device 10s may learn from that one year's worth of actual data to create a trained model.
[0030] <Configuration example of power generation amount estimation device 2> 4 is a functional block diagram showing a schematic configuration of a power generation amount estimation device 2 according to a reference embodiment. The power generation amount estimation device 2 estimates the amount of solar power generation corresponding to the actual measured values of the amount of solar radiation and the actual measured values of air temperature using a trained model 11s (in the example of the reference embodiment, the best trained model) generated by a learning device 10s. The power generation amount estimation device 2 includes an acquisition unit 21 and an estimation unit 22. The estimation unit 22 also includes a control unit 223 and the trained model 11s generated by the learning device 10s.
[0031] The acquisition unit 21 is a functional unit that acquires the measured values of the amount of solar radiation and the measured values of the temperature. The measured value of the amount of solar radiation is, for example, the amount of solar radiation around the solar panel measured by the solar radiation measurement device 12. The measured value of the temperature is, for example, the temperature measured by the temperature measurement device 13. The acquisition unit 21 may acquire the measured values of the amount of solar radiation and the measured values of the temperature from outside in response to receiving a predetermined user operation.
[0032] The acquisition unit 21 may externally acquire weather forecast data instead of the actual measured values of solar radiation and temperature. The weather forecast data includes predicted data of solar radiation, predicted data of temperature, and time information of the forecast target. For example, if the acquisition unit 21 is configured with a network adapter or the like, it receives weather forecast data from a server or the like owned by a weather forecast data provider, such as the Japan Meteorological Agency or a private weather forecasting company, via a wide area communication network such as the Internet. Furthermore, if the acquisition unit 21 is configured with a man-machine interface or the like, the user may input the weather forecast data.
[0033] When estimating the amount of solar power generation, the control unit 223 outputs the actual measured value of the amount of solar radiation, the actual measured value of the temperature, and the time information at the time of the actual measurement to the trained model 11s. The control unit 223 may output the actual data of the amount of solar power generation to the trained model 11s.
[0034] The trained model 11s receives as input the actual measured value of solar radiation, the actual measured value of temperature, and time information at the time of the actual measurement, and outputs an estimated value of the amount of solar power generation. The estimated value of the amount of solar power generation is displayed, for example, on a display device. The trained model 11s may further receive as input actual data on the amount of solar power generation, and output an estimated value of the amount of solar power generation.
[0035] [Embodiment 1] FIG. 5 is a functional block diagram showing a schematic configuration of a model generation device 1 according to the first embodiment. The model generation device 1 includes a learning device 10 (learning unit), a solar radiation measuring device 12, a temperature measuring device 13, an RTC 14, and a power generation amount measuring device 15. In other words, the model generation device 1 according to the first embodiment differs from the model generation device 1s according to the reference embodiment in that the model generation device 1 includes a learning device 10 instead of the learning device 10s. Prior to describing the operation of the learning device 10s, possible improvements in the reference embodiment will be described below. Note that the power generation amount estimation device 2 according to the first embodiment has the same functions as the power generation amount estimation device 2 according to the reference embodiment, and therefore its description will be omitted.
[0036] As described above, in the reference embodiment, multiple learning models are generated by exhaustively (brute-force) executing multiple types of machine learning algorithms. Then, the best learning model is selected by exhaustively evaluating each of the generated learning models. That is, the best learning model is selected by performing a grid search on the generated learning models.
[0037] However, in the reference embodiment, multiple learning models are generated by applying all patterns of hyperparameter sets for multiple types of machine learning algorithms, which may result in an extension of the calculation time (learning time) required for the learning phase. To prevent this extension of calculation time, for example, a method can be considered in which a search is performed by narrowing down the machine learning algorithm selection candidates and hyperparameter adjustment candidates. However, it is not certain whether the narrowed down candidates are appropriate, and the predictive accuracy of the learning model found from the narrowed down candidates may be significantly lower than the predictive accuracy of the best learning model found by the above-mentioned grid search.
[0038] In this embodiment, a learning device is provided that can find an appropriate learning model with good prediction accuracy (for example, a learning model with an index value greater than a predetermined threshold) while reducing calculation time.
[0039] <Example of processing by the learning device 10> 6 is a flowchart for explaining the processing procedure of the model generation device 1. In the processing of the model generation device 1, the processing (S1 to S4, S6) other than the processing of the learning device 10 (learning step S15) is the same as that of the model generation device 1s. Therefore, only the processing of the learning device 10 will be explained below.
[0040] Similar to the reference embodiment, the learning device 10 executes a machine learning algorithm to generate multiple learning models based on features for estimating solar power generation. The learning device 10 then derives multiple index values indicating the quality of each of the multiple learning models by verifying each of the multiple learning models.
[0041] Here, the learning device 10 searches for a learning model (appropriate learning model) having a better index value based on the multiple index values in accordance with a predetermined optimization method. The learning device 10 outputs the appropriate learning model as a trained model 11. In the first embodiment, as an example, a case will be described in which the learning device 10 searches for a learning model (referred to as a better learning model) having a better index value for each of multiple types of machine learning algorithms (NN, SVR, LASSO, RIDGE, etc.). Specifically, the learning device 10 searches for a better learning model from multiple learning models generated by applying multiple hyperparameter settings applied to a certain machine learning algorithm (referred to as a machine learning algorithm of interest). Below, a series of processes performed by the learning device 10 when executing a certain machine learning algorithm will be described.
[0042] Similar to the learning device 10s in the reference embodiment, the learning device 10 first acquires a DZ and divides the DZ into training data and validation data. Next, the learning device 10 searches for a learning model for each of a plurality of machine learning algorithms according to a predetermined optimization method. Specifically, the learning device 10 searches for a learning model by repeating a learning phase and a validation phase according to the predetermined optimization method.
[0043] In the learning phase, the learning device 10s generates multiple learning models by applying multiple hyperparameter settings of a target machine learning algorithm based on the training data. The target machine learning algorithm is one of the multiple predetermined types of machine learning algorithms described in the embodiment.
[0044] In the verification phase, the learning device 10 derives index values for each of the multiple learning models generated in the learning phase using the verification data. The learning device 10 finds a better learning model based on the index values from multiple learning models obtained by applying multiple hyperparameter settings for multiple types of machine learning algorithms, according to a predetermined optimization method. The specific processing flow of the predetermined optimization method will be described later with reference to Figures 7 to 9.
[0045] Next, the learning device 10 selects an appropriate learning model from among the better learning models found for each of the multiple types of machine learning algorithms. For example, the learning device 10 selects, from among all the better learning models, the learning model having the maximum index value (as an example of a statistical value) as the appropriate learning model. The subsequent processing in the first embodiment is the same as that in the reference embodiment, except that the term "best learning model" in the reference embodiment is replaced with "appropriate learning model."
[0046] (An example of a predetermined optimization method) A specific processing flow using a predetermined optimization method will be described below with reference to Figures 7 to 9. Figure 7 is a flowchart that outlines the main processing flow of the learning device 10. Figure 7 shows more specifically the processing content of S15 in Figure 6. Below, a case where NN (an example of a notable machine learning algorithm) is executed as the machine learning algorithm will be described.
[0047] First, the learning device 10 acquires the number N of items in the search parameter list (total number N of lists) (S21). The search parameter list is a list of parameters (hereinafter referred to as search parameters) used to search for a learning model. In the first embodiment, the learning device 10 searches for a learning model using multiple hyperparameter settings for each of multiple machine learning algorithms, and therefore the search parameter list includes only hyperparameters. That is, the number N of items in the search parameter list is equal to the number of items in the hyperparameter settings. For simplicity, the following example illustrates a case where N=2. In other words, the example illustrates a case where the number of items in the hyperparameter settings is 2.
[0048] Next, the learning device 10 sets N+1 combinations of random numerical values from the items in the search parameter list. Then, the learning device 10 sets N+1 lattice points P1 to P2 in the N-dimensional space corresponding to the combinations. N+1 are generated as initial points (S22). In this specification, the lattice points refer to points that form a spatial lattice. In the first embodiment, since N+1=3, the learning device 10 generates P1=(HA1,HB1) P2=(HA2,HB2) P3=(HA3,HB3) These three grid points P1 to P3 are generated as initial points. j (j=1,2,3) is the index indicating the first hyperparameter of the NN (e.g., batch size), and HB j (j=1,2,3) is an index indicating the second hyperparameter of the NN (e.g., the maximum number of learning iterations).
[0049] Next, the learning device 10 sets the lattice points P1 to P2 in S22 based on the training data. N+1 In this way, the learning device 10 generates a plurality of learning models by using a plurality of search parameters (a plurality of hyperparameter settings) corresponding to each of the lattice points P1 to P N+1 Generate a learning model corresponding to
[0050] Next, the learning device 10 acquires the index values of each of the plurality of learning models generated in S23 using the verification data (S24). N+1 The function f(P1)~f(P N+1 ) is obtained (derived).
[0051] The function f may be set as a function (so-called objective function) for evaluating an index value (e.g., prediction accuracy). For example, the function f may be set as a monotonically decreasing function of the prediction accuracy. In this case, the value of the function f decreases as the prediction accuracy increases. In the example of the first embodiment, the function f is f=100-prediction accuracy [%]…(1) is set as:
[0052] Below, (i) based on the above training data, the lattice point P j (ii) using the validation data, obtain index values for the learning model; and (iii) calculating f(P j ) is simply called "f(P j ) is denoted as "derive".
[0053] Next, the learning device 10 calculates the index values (in other words, functions f(P1) to f(P N+1 )) based on each grid point P1 to P N+1 In the first embodiment, the learning device 10 ranks the three initial points to which the best prediction accuracy (highest prediction accuracy) is assigned as the best point B. (m) m is the number of optimization iterations (optimization iteration count) in S26 described below. m is an integer that satisfies 1≦m≦q. q is the upper limit of the optimization iteration count. q is set in advance. Then, the learning device 10 determines the point assigned the worst prediction accuracy (lowest prediction accuracy) among the three initial points as the worst point W (m)Next, the learning device 10 determines the point among the three initial points to which the second worst prediction accuracy (the second lowest prediction accuracy) is assigned as the second worst point BW (m) For example, if f(P1)≦f(P2)≦f(P3), the grid point P1 is the best point B (m) , lattice point P2 is the second worst point BW (m) , lattice point P3 is the worst point W (m) is.
[0054] Next, the learning device 10 calculates the rank of each of the ordered lattice points P1 to P N+1 Based on this, optimization is performed using a predetermined optimization method (S26). The details of the process in S26 will be described later.
[0055] In the first embodiment, an example is shown in which the Nelder-Mead method is used as the predetermined optimization method. The Nelder-Mead method is also called the downhill simplex method, the hyperpolyhedron method, or the amoeba method. The Nelder-Mead method can solve multidimensional nonlinear optimization problems without relying on differential calculus. For example, the Nelder-Mead method can derive a locally optimal solution of an objective function without relying on differential calculus. Therefore, the Nelder-Mead method can obtain an extreme value (maximum or minimum value) of the objective function without relying on differential calculus. In the first embodiment, the objective function is f(P j ) is expressed as equation (1). Therefore, in embodiment 1, by using the Nelder-Mead method, it is possible to obtain a local minimum (ideally the smallest value) of the objective function. In other words, by using the Nelder-Mead method, it is possible to obtain a local maximum (ideally the largest value) of the prediction accuracy. In other words, by using the Nelder-Mead method, it is possible to find a better learning model.
[0056] Figure 8 is a diagram for explaining the concept of the Nelder-Mead method. For simplicity, Figure 8 shows an example in which N = 2 (two-dimensional case). Figure 8 shows the relationship between B and m = 1, 2, q-1, and q. (m) , W (m) , and B.W.(m) The dotted lines in the example of Fig. 8 indicate the contour lines of the prediction accuracy. The contour lines of the prediction accuracy can also be interpreted as the contour lines of the objective function.
[0057] As shown in Figure 8, according to the Nelder-Mead method, as the optimization iteration process progresses (increases in m), the combination of hyperparameters (B (m) , W (m) , and B.W. (m) In other words, the Nelder-Mead method allows each hyperparameter to be optimized to reach a local optimum (maximum value of prediction accuracy).
[0058] Fig. 9 is a flowchart showing the flow of the process for executing optimization. Fig. 9 shows more specifically the process content of S26 in Fig. 7. Fig. 9 illustrates optimization using the Nelder-Mead method.
[0059] First, in S31, the learning device 10 calculates the worst point W (m) Each point other than (collectively, each point P j (m) ) and each point P j (m) The center of gravity of the hyperpolyhedron G (m) Specifically, the learning device 10 calculates
number
[0060] Next, in S32, the learning device 10 calculates the line segment W (m) G (m) External division point R that divides (m) Specifically, the learning device 10 calculates R (m) =2G (m) -W (m) …(3) As shown above, the external division point R (m) Ask for.
[0061] Next, in S33, the learning device 10 calculates f(R (m) ) is derived. Then, the learning device 10 f(R (m) )≦f(B (m) ) …(4) It is determined whether or not the following condition (for convenience, referred to as the first determination condition) is satisfied. If the first determination condition is satisfied, the process proceeds to S34. If the first determination condition is not satisfied, the process proceeds to S37.
[0062] If the first judgment condition is satisfied, in S34, the learning device 10 calculates the line segment W (m) R (m) External division point E, which divides (m) Specifically, the learning device 10 calculates E (m) =3G (m) -2W (m) …(5) As shown, the external division point E (m) Ask for.
[0063] Next, in S35, the learning device 10 calculates f(E (m) ) is derived. Then, the learning device 10 f(E (m) )≦f(R (m) ) …(6) It is determined whether or not the following condition (for convenience, referred to as the second determination condition) is satisfied. If the second determination condition is satisfied, the process proceeds to S36. If the second determination condition is not satisfied, the process proceeds to S38.
[0064] If the second judgment condition is satisfied, in S36, the learning device 10 calculates the worst point W (m) is removed from the optimization target, and point E (m) Then, the process proceeds to step S43, which will be described later.
[0065] If the first determination condition is not satisfied, in S37, the learning device 10 f(R (m) )≦f(BW (m) ) …(7) It is determined whether or not the following condition (for convenience, referred to as the third determination condition) is satisfied. If the third determination condition is satisfied, the process proceeds to S38. If the third determination condition is not satisfied, the process proceeds to S39.
[0066] If the third judgment condition is satisfied, in S38, the learning device 10 calculates the worst point W (m) is removed from the optimization target, and point R (m) Then, the process proceeds to S43.
[0067] If the third judgment condition is not satisfied, in S39, the learning device 10 (m) G (m) Midpoint S of (m) Specifically, the learning device 10 calculates S (m) =(1 / 2)×(G (m) +W (m) ) …(8) As shown, the midpoint S (m) Ask for.
[0068] Next, in S40, the learning device 10 calculates f(S (m) ) is derived. Then, the learning device 10 f(S (m) )≦f(W (m) ) …(9) It is determined whether or not the following condition (for convenience, referred to as the fourth determination condition) is satisfied. If the fourth determination condition is satisfied, the process proceeds to S41. If the fourth determination condition is not satisfied, the process proceeds to S42.
[0069] If the fourth judgment condition is satisfied, in S41, the learning device 10 calculates the worst point W (m) is removed from the optimization target, and point S (m) Then, the process proceeds to S43.
[0070] If the fourth judgment condition is not satisfied, in S42, the learning device 10 determines whether the line segment B (m) P j (m) Midpoint P j (m)’ Specifically, the learning device 10 calculates P j (m)’ =(1 / 2)×(P j (m) +B (m) ) …(10) As shown, the midpoint P j (m)’ Then, the learning device 10 calculates the j (m) At each point P j (m)’ Then, proceed to S43.
[0071] In S43, the learning device 10 calculates the value of each point P based on the function f in the same manner as in S25 described above. j (m) Rank the following.
[0072] Next, in S44, the learning device 10 determines whether a predetermined setting judgment condition is satisfied. If the setting judgment condition is satisfied, the process proceeds to S45. If the setting judgment condition is not satisfied, the process returns to S31. In this way, the learning device 10 repeats each process from S31 to S43 until the setting judgment condition is satisfied. Note that when returning from step S44 to step S31, the learning device 10 increments m by 1.
[0073] The set determination conditions in the first embodiment may include, for example, at least one of the following conditions 1 to 6: j (m) V0 is the volume of the hyperpolyhedron formed by each initial point. Also, α in the following condition 6 is
number
[0074] (Condition 1) Vm / V0 (the ratio of Vm to V0) is equal to or less than a predetermined threshold (volume ratio threshold); (Condition 2) m reaches the upper limit q; (Condition 3) The calculation time for the optimization process has reached a predetermined time; (Condition 4) Point B (m) was not changed for more than a given number of iterations; (Condition 5) Point B (m) The prediction accuracy of is equal to or greater than a predetermined threshold (prediction accuracy threshold); (Condition 6) α is equal to or less than a predetermined threshold (gradient threshold); However, as will be apparent to those skilled in the art, the set judgment condition is not limited to the above example. The set judgment condition according to one aspect of the present invention may be set to represent at least one of "a state in which a local optimum solution of the function f (i.e., the prediction accuracy in the example of the first embodiment) has been found" and "a condition under which further search for the local optimum solution should be terminated."
[0075] If the set judgment condition is satisfied, in S45, the learning device 10 determines the best point B (m) As an example, the best point B in S45 is adopted. (m) The coordinates of B (m) =(HA better ,HB better ) In this case, the learning device 10 uses HA better ,HB better The learning device 10 then selects, as a better hyperparameter setting, the hyperparameter setting item corresponding to the above. Then, based on the training data, the learning device 10 finds, as a better learning model, a learning model obtained by applying the above better hyperparameter setting of the NN.
[0076] The learning device 10 executes the above-described optimization method for each of the multiple types of machine learning algorithms to find a better learning model. Then, the learning device 10 selects an appropriate learning model (e.g., the better learning model having the highest index value) from the better learning models found for each of the multiple types of machine learning algorithms.
[0077] (Supplement to the Nelder-Mead method) In the above-mentioned Nelder-Mead method, the external division point R obtained according to equation (3) (m) The coordinates of are not necessarily integers. Therefore, if the search parameter list includes hyperparameters that must be integers (e.g., the number of training steps, the number of network layers, or the number of nodes), the external division point R (m) There is a possibility that there is no hyperparameter set corresponding to each coordinate of , and a learning model cannot be generated (i.e., the function f cannot be derived).
[0078] Therefore, the learning device 10 calculates the external division point R (m) After finding the external division point R (m) All the coordinates of are converted to integers (the first decimal place of each coordinate value is rounded up, down, or rounded up), and the external division point R (m) The grid point R corresponding to (m)’ Thereafter, the learning device 10 calculates the lattice points R (m)’ External division point R (m) In this way, even when a hyperparameter that must be an integer is included in the search parameter list, the learning device 10 can find the external division point R (m) The grid point R corresponding to (m)’ The function f can be derived using the external division point R (m) By converting all of the coordinates of the vectors into integers as described above, the configuration of the learning device 10 can be simplified.
[0079] The learning device 10 determines the external division point R (m)Among the coordinates of the search parameter list, only the coordinates corresponding to the hyperparameters that must be integers may be converted to integers. Specifically, the learning device 10 does not convert the coordinates corresponding to the hyperparameters (e.g., learning rate, coefficient of regularization term, etc.) that do not need to be integers among the hyperparameters included in the search parameter list to integers. On the other hand, the learning device 10 converts the coordinates corresponding to the hyperparameters that must be integers among the hyperparameters included in the search parameter list to integers.
[0080] In addition, the external division point E (m) , midpoint S (m) , midpoint P j (m)’ Regarding the external division point R (m) Similarly, each coordinate is converted to an integer.
[0081] (Effects of the First Embodiment) According to the learning device 10 of the first embodiment, an appropriate learning model can be found by executing a predetermined optimization method (e.g., the Nelder-Mead method). Therefore, unlike the reference embodiment, an appropriate learning model can be found efficiently without searching all possible hyperparameter setting patterns. Therefore, the learning time required to find an appropriate learning model can be reduced.
[0082] Furthermore, by executing a predetermined optimization method for each of multiple types of machine learning algorithms, the search parameters in the predetermined optimization method can be limited to the hyperparameters of the machine learning algorithms, thereby shortening the processing time required to search for a learning model in the predetermined optimization method.
[0083] [Variation 1] In the first embodiment, the learning device 10 searches for a learning model using multiple hyperparameter settings for each of multiple types of machine learning algorithms according to a predetermined optimization method. However, the learning device 10 may also search for a learning model using multiple hyperparameter settings for multiple types of machine learning algorithms according to a predetermined optimization method. In other words, multiple types of machine learning algorithms may also be optimized according to a predetermined optimization method.
[0084] Specifically, the learning device 10 searches for a learning model by repeating a learning phase and a verification phase according to a predetermined optimization method. In the learning phase, the learning device 10 generates multiple learning models by executing multiple types of machine learning algorithms to which multiple hyperparameter settings are respectively applied. In the verification phase, the learning device 10 derives index values for each of the multiple learning models generated in the learning phase using the verification data.
[0085] In this case, in the predetermined optimization method, the search parameter list includes the ranking of the machine learning algorithms and hyperparameters. That is, compared to the first embodiment, the ranking of the machine learning algorithms is added as a search parameter. Here, the ranking of the machine learning algorithms is assigned to multiple types of machine learning algorithms according to a predetermined ranking. For example, the ranking is assigned to multiple types of machine learning algorithms as follows: 1. NN, 2. LASSO, 3. RIDGE, etc. In this case, it is preferable that the hyperparameters specific to each machine learning algorithm (e.g., the number of layers in the NN) are fixed.
[0086] The learning device 10 optimizes these search parameters according to the predetermined optimization method described above. As a result, the learning device 10 finds a better learning model from among multiple learning models obtained by executing multiple types of machine learning algorithms to which multiple hyperparameter settings are respectively applied. In the case of Modification 1, the appropriate learning model is the better learning model.
[0087] In Modification 1, a learning model is searched for by optimizing multiple types of machine learning algorithms according to a predetermined optimization method. Therefore, it is not necessary to search for a learning model for each of the multiple types of machine learning algorithms, and the learning time required to find an appropriate learning model can be further reduced.
[0088] [Variation 2] In the first embodiment, the Nelder-Mead method is used as an optimization method, but as will be apparent to those skilled in the art, other methods may be adopted as the optimization method according to one aspect of the present invention.
[0089] Another example of an optimization method is the Powell method. Like the Nelder-Mead method, the Powell method can solve multidimensional nonlinear optimization problems without relying on differential calculus. Therefore, the Powell method can be used instead of the Nelder-Mead method to derive the maximum value of prediction accuracy.
[0090] Furthermore, when solving one-dimensional nonlinear optimization problems, the golden section method can also be used as an optimization technique. The golden section method can solve one-dimensional nonlinear optimization problems without relying on differential calculus. The golden section method can derive the maximum value of prediction accuracy expressed as a one-dimensional nonlinear function without relying on differential calculus.
[0091] As will be apparent to those skilled in the art, the optimization method according to an embodiment of the present invention is not particularly limited. However, from the viewpoint of facilitating implementation of the algorithm, the optimization method according to an embodiment of the present invention is preferably an optimization method that does not rely on differential calculus, i.e., derivative-free optimization (DFO). The Nelder-Mead algorithm, the Powell algorithm, and the golden section method are all examples of DFO.
[0092] [Software implementation example] The functions of the model generation device 1 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly the learning device 10).
[0093] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.
[0094] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.
[0095] In addition, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of one aspect of the present invention. In addition, the functions of the control blocks can be realized by, for example, a quantum computer.
[0096] As is clear from the above explanations, each process described in each of the above embodiments can be executed by AI (Artificial Intelligence). In this case, the AI may run on the control device or on another device (for example, an edge computer or a cloud server).
[0097] [Additional Notes] One aspect of the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of one aspect of the present invention. [Explanation of symbols]
[0098] 1. 1s model generation device 2. Power generation estimation device 10, 10s learning device (learning section) 11, 11s trained model 12 Solar radiation measuring device (acquisition unit) 13 Temperature measuring device (acquisition part) 14 RTC (Acquisition Department) 15 Power generation measuring device (acquisition unit)
Claims
1. an acquisition unit that acquires feature quantities for estimating the amount of power generated by the solar panel; a learning unit that executes a machine learning algorithm to generate a plurality of learning models based on the feature amounts, and derives a plurality of index values that indicate the quality of each of the plurality of generated learning models; A model generation device in which the learning unit searches for a learning model having a better index value based on the multiple index values in accordance with a predetermined optimization method, and selects the learning model having the better index value as an appropriate learning model.
2. The learning unit, for each of the plurality of types of machine learning algorithms, generating the plurality of learning models using a plurality of hyperparameter settings applied to each of the plurality of types of machine learning algorithms; 2. The model generation device according to claim 1, wherein the learning model having a better index value is selected as the appropriate learning model by searching for the learning model having a better index value based on the plurality of index values according to the predetermined optimization method.
3. The model generation device according to claim 1 , wherein the learning unit generates the plurality of learning models by executing a plurality of types of the machine learning algorithms to which a plurality of hyperparameter settings are respectively applied.
4. 4. The model generating device according to claim 1, wherein the optimization method is the Nelder-Mead method or the Powell method.
5. an acquisition step of acquiring feature quantities for estimating the amount of power generated by the solar panel; a learning process of generating a plurality of learning models based on the feature quantities by executing a machine learning algorithm, and deriving a plurality of index values indicating the quality of each of the plurality of generated learning models; A model generation method in which, in the learning process, a learning model having a better index value is searched for based on the multiple index values according to a predetermined optimization method, and the learning model having the better index value is selected as the appropriate learning model.
Citation Information
Patent Citations
Solar radiation amount prediction method, photovoltaic power generation output prediction method and system
JP2013164286A