Model generation device and model generation method

Abstract

To reduce time required for finding out a proper prediction model for predicting the amount of electricity demand in a facility.SOLUTION: A model generation device (10A) that generates a prediction model for predicting the amount of electricity demand in a facility is configured to: derive a feature amount for deriving a predicted value of the amount of electricity demand from time series data that indicates time progression of actual values of the amount of electricity demand; generate a plurality of prediction models based on the feature amount by executing a machine learning algorithm; derive a plurality of index values that indicate quality of each of the plurality of prediction models; and select an appropriate model from among the plurality of prediction models by searching for the prediction model having a better index value based on the plurality of index values according to a predetermined optimization method.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] One aspect of the present invention relates to a model generation device that generates a prediction model for predicting an amount of power demand in a facility. [Background technology]

[0002] In recent years, machine learning has been used to predict various types of energy demand in facilities. For example, Patent Document 1 below discloses a technology for predicting the heat load in a thermal storage utilization system using a neural network. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 9-89348 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 time required to find an appropriate prediction model for predicting the amount of power demand in a facility. [Means for solving the problem]

[0005] A model generation device according to one aspect of the present invention is a model generation device that generates a prediction model for predicting the amount of electricity demand in a facility, deriving feature quantities for deriving a predicted value of the amount of electricity demand from time series data showing the time progression of actual values ​​of the amount of electricity demand, generating a plurality of prediction models based on the feature quantities by executing a machine learning algorithm, deriving a plurality of index values ​​that indicate the quality of each of the plurality of prediction models, and selecting an appropriate model from among the plurality of prediction models by searching for a prediction model with a better index value based on the plurality of index values ​​according to a predetermined optimization method.

[0006] A model generation method according to one aspect of the present invention is a model generation method for generating a prediction model for predicting the amount of electricity demand in a facility, and includes the steps of: deriving feature values ​​for deriving a predicted value of the amount of electricity demand from time series data showing the time progression of actual values ​​of the amount of electricity demand; generating a plurality of prediction models based on the feature values ​​by executing a machine learning algorithm; deriving a plurality of index values ​​indicating the quality of each of the plurality of prediction models; and selecting an appropriate model from the plurality of prediction models by searching for the prediction model having a better index value based on the plurality of index values ​​according to a predetermined optimization method. [Effects of the Invention]

[0007] According to one aspect of the present invention, it is possible to reduce the time required to find an appropriate prediction model for predicting the amount of power demand in a facility. [Brief explanation of the drawings]

[0008] [Figure 1] 1 shows an example of the configuration of an information processing device according to a reference embodiment. [Figure 2] 10 shows an example of an evaluation result table in the reference embodiment. [Figure 3] 1 shows an example of the configuration of an information processing device according to a first embodiment. [Figure 4] 3 shows an outline of the main processing flow of a learning unit in the first embodiment. [Figure 5] This is a diagram that schematically explains the concept of the Nelder-Mead method. [Figure 6] 10 illustrates an example of a process flow for executing optimization in the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] [Reference form] Before describing the information processing device 1A of the first embodiment, an information processing device 1 will be described as a reference embodiment. For convenience 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. For simplicity, descriptions of matters similar to those in known technologies will also be omitted as appropriate.

[0010] Unless otherwise specified, the components and values ​​described in this disclosure are merely examples. Therefore, unless otherwise specified, the positional relationships and connections of the components are not limited to the examples in the drawings.

[0011] (Overview of information processing device 1) 1 shows an example of the configuration of an information processing device 1. The information processing device 1 in the example of FIG. 1 includes a control device 9 and a storage unit 90. The control device 9 comprehensively controls each unit of the information processing device 1. The control device 9 includes a model generation device 10 and a prediction device 20.

[0012] As described below, the model generation device 10 generates a prediction model for predicting the amount of electricity demand in a facility by machine learning. For this reason, the model generation device 10 may be referred to as a learning device. The prediction device 20 derives a predicted value of the amount of electricity demand in the facility using the prediction model previously generated by the model generation device 10.

[0013] The facility according to one aspect of the present invention may be any facility in which the amount of power demand may fluctuate over time, such as an industrial facility such as a factory, a commercial facility such as a shopping mall, or a medical facility such as a hospital.

[0014] The storage unit 90 stores various data and programs used in the processing of the control device 9. In the example of FIG. 1, actual data is stored in the storage unit 90. In this specification, actual data collectively refers to data (more specifically, data structure) that indicates actual values ​​of the amount of power demand in a facility, recorded with a predetermined time resolution. In this specification, unless otherwise contradictory, the term "amount of power demand" refers to the actual values ​​of the amount of power demand.

[0015] The actual data is data in which a time point and an amount of demanded energy are associated, and is therefore an example of time-series data that shows the time transition of the amount of demanded energy. In the example of the reference form, the unit of the amount of demanded energy recorded in the actual data is kWh, and the time resolution of the actual data is one minute. The time point in data with a time resolution of one minute may be expressed in the "year / month / day hour / minute" format ("yyyy / mm / dd / hh / mm" format). In this case, the time point can be used as a key in the data structure.

[0016] (An example of processing in the model generation device 10) The following describes the operation of each unit of the information processing device 1. The processing of the information processing device 1 is roughly divided into processing in the model generation device 10 and processing in the prediction device 20. First, an example of the processing in the model generation device 10 will be described.

[0017] 1 includes a data acquisition unit 11, a feature derivation unit 12, and a learning unit 13. The processing of the model generation device 10 can be broadly divided into a learning phase (a phase in which a predictive model is generated) and a verification phase (a phase in which the generated predictive model is verified). The learning phase may also be referred to as a model generation phase.

[0018] 1, performance data DZ is stored in the storage unit 90. Hereinafter, performance data DZ will be simply abbreviated as DZ. The data acquisition unit 11 acquires DZ. Then, the data acquisition unit 11 divides the acquired DZ into training data and validation data.

[0019] First, the learning phase will be described. Prior to the execution of a machine learning algorithm by the learning unit 13, the feature derivation unit 12 derives features (i.e., explanatory variables) from training data for generating a prediction model.

[0020] The features for generating a prediction model are not particularly limited as long as they are considered to be useful for predicting the amount of power demand. The features derived from the training data may be referred to as training features or model generation evaluation values. As an example, the features can be derived by setting a predetermined lag (a time delay) for the amount of power demand in the training data.

[0021] For example, fluctuations in the amount of power demand within a facility are generally considered to have a short-term cycle of about one day. Therefore, the feature deriving unit 12 may derive a lag feature by setting a first lag for the amount of power demand in the training data. The lag feature can express the short-term cycle of fluctuations in the amount of power demand.

[0022] The first lag may be any value that is considered suitable for deriving a lag feature. In the reference embodiment, the first lag is set in advance. In the reference embodiment, a case where the time length of the first lag is one day is exemplified. Therefore, the feature deriving unit 12 derives the "demand energy amount one day before" corresponding to the demand energy amount at a certain point in time in the training data as a lag feature. To distinguish it from the second lag described below, the first lag may be referred to as a non-rolling lag.

[0023] As an example, the feature derivation unit 12 acquires the "demanded energy amount at 0:00 on 3 / 31 / 2021 (e.g., 400 kWh)" as the "demanded energy amount one day ago" corresponding to the "demanded energy amount at 0:00 on 4 / 1 / 2021 (e.g., 500 kWh)."

[0024] As another example, the feature derivation unit 12 acquires "demanded energy amount at 23:59 on 3 / 31 / 2021 (e.g., 498 kWh)" as "demanded energy amount one day ago" corresponding to "demanded energy amount at 23:59 on 4 / 1 / 2021 (e.g., 530 kWh)."

[0025] Fluctuations in the amount of power demand within a facility may include a period longer than the short-term period described above. For example, the fluctuations may include a relatively long-term period of about one week (7 days). Therefore, the feature deriving unit 12 may derive a rolling feature by setting a second lag for the amount of power demand in the training data. The rolling feature can express the relatively long-term period of fluctuations in the amount of power demand.

[0026] The second lag may be any value that is considered suitable for deriving rolling features. For this reason, the second lag may be referred to as a rolling lag. The second lag may be set to be greater than the first lag described above. In the reference embodiment, the second lag is set in advance. In order to acquire multiple rolling feature components (described later), multiple different second lags may be set. In the reference embodiment, cases where the time length of the second lag is 7 days and 14 days are exemplified.

[0027] First, the feature derivation unit 12 acquires the amount of demand energy at a time point that is the second lag before a certain time point in the training data as a constituent element (rolling feature component) of the rolling feature corresponding to that time point. In the example of the reference embodiment, since the time length of the second lag is 7 days and 14 days, the feature derivation unit 12 acquires the "amount of demand energy 7 days ago" and the "amount of demand energy 14 days ago" corresponding to the certain time point as the rolling feature components.

[0028] As an example, the feature derivation unit 12 acquires the "demanded energy amount at 0:00 on 3 / 25 / 2021 (e.g., 501 kWh)" as the "demanded energy amount 7 days ago" corresponding to the above-mentioned "demanded energy amount at 0:00 on 4 / 1 / 2021 (e.g., 500 kWh)."

[0029] Then, the feature derivation unit 12 acquires the "amount of power demand at 0:00 on 3 / 18 / 2021 (e.g., 503 kWh)" as the "amount of power demand 14 days ago" corresponding to the "amount of power demand at 0:00 on 4 / 1 / 2021."

[0030] Next, the feature derivation unit 12 derives a statistical value of the rolling feature component as a rolling feature. In the reference embodiment, a case where the statistical value is an average value is illustrated. Therefore, the feature derivation unit 12 derives the average value of the "energy demand amount 7 days ago" and the "energy demand amount 14 days ago." The feature derivation unit 12 determines the average value as the rolling feature.

[0031] As an example, the feature derivation unit 12 derives the average value of "demand energy amount at 0:00 on 2021 / 4 / 1" and "demand energy amount at 0:00 on 2021 / 3 / 25" as a rolling feature corresponding to "demand energy amount at 0:00 on 2021 / 4 / 1". In the example of the reference embodiment, the rolling feature is determined as the average value of 501 kWh and 503 kWh, i.e., 502 kWh.

[0032] In the learning phase, the learning unit 13 generates a prediction model based on the features (e.g., lag features and rolling features) derived by the feature derivation unit 12. Specifically, the learning unit 13 generates the prediction model by executing a predetermined machine learning algorithm.

[0033] A prediction model according to an embodiment of the present disclosure may be a trained model that can output a predicted value of the amount of power demand at a time point to be predicted as a response variable. Therefore, the type of machine learning algorithm is not particularly limited as long as it 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.

[0034] After the learning phase is completed, the process proceeds to the verification phase. In the verification phase, the learning unit 13 verifies (evaluates) the prediction model generated in the learning phase. Specifically, the learning unit 13 evaluates the prediction performance of the prediction model.

[0035] In the verification phase, the feature deriving unit 12 derives features from the verification data by performing the same series of processes on the verification data as those performed on the training data. The features derived from the verification data may be referred to as evaluation features.

[0036] As is clear from the above description, in the validation phase, features of the same type as the training features are derived as evaluation features. For example, if lag features and rolling features are derived as training features in the learning phase, lag features and rolling features are derived as evaluation features in the validation phase.

[0037] The learning unit 13 evaluates the prediction model based on the evaluation features. Specifically, the learning unit 13 inputs the evaluation features into the prediction model, causing the prediction model to output a predicted value. Then, the learning unit 13 derives an index value indicating the prediction accuracy of the prediction model based on the predicted value and the true value of the energy demand indicated in the verification data.

[0038] The index value is not particularly limited as long as it is used in the field of machine learning to evaluate prediction accuracy in regression tasks. The index value may be any statistical quantity derived based on the error between a predicted value and a true value. As an example, the error may be the mean absolute error (MAE) between the predicted value and the true value. As another example, the error may be the root mean square error (RMSE) between the predicted value and the true value.

[0039] The index value according to one embodiment of the present invention may be used as a value quantitatively representing the quality of a prediction model. In a 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. In a reference example, the minimum value of the index value is 0% and the maximum value of the index value is 100%.

[0040] As an example, if the index value of a prediction model is equal to or greater than a predetermined threshold, the learning unit 13 may store the prediction model in the storage unit 90. In this specification, a prediction model having an index value equal to or greater than a threshold is referred to as a "valid model." A valid model can be said to be a prediction model that is guaranteed to have a certain level of quality.

[0041] As described above, by storing the valid model in the storage unit 90, the valid model can be used by the prediction device 20 in the prediction phase, which will be described later. Therefore, prediction accuracy in the prediction phase can be ensured.

[0042] (Another example of processing in the model generation device 10) The quality of a predictive model may vary depending on the type of machine learning algorithm used to generate the predictive model. The quality of a predictive model may also vary depending on the hyperparameter set (a set of hyperparameters) applied in a certain machine learning algorithm. Therefore, if an appropriate combination of the type of machine learning algorithm and the hyperparameter set can be found, a higher quality predictive model can be obtained.

[0043] Therefore, as an example, in the reference embodiment, (i) a predetermined number of types of machine learning algorithms to be used for generating a predictive model and (ii) a hyperparameter set to be applied to each of the plurality of types of machine learning algorithms may be determined in advance. In this case, a table showing the correspondence between each machine learning algorithm and each hyperparameter set may be stored in the storage unit 90.

[0044] In this case, in the learning phase, the learning unit 13 may generate multiple prediction models according to the table. For example, the learning unit 13 may comprehensively generate prediction models for all combinations of types of machine learning algorithms and hyperparameter sets.

[0045] As an example, the learning unit 13 generates a plurality of prediction models for each of a plurality of types of machine learning algorithms, using a plurality of hyperparameter sets applied to each of the plurality of types of machine learning algorithms.

[0046] Next, in the verification phase, the learning unit 13 evaluates each of the multiple prediction models generated in the learning phase. Specifically, the learning unit 13 inputs evaluation features into each of the multiple prediction models, causing each of the multiple prediction models to output a predicted value corresponding to the evaluation features. Then, the learning unit 13 derives an index value for each of the multiple prediction models based on the predicted value obtained from each of the multiple prediction models and the true value of the energy demand indicated in the verification data.

[0047] The learning unit 13 selects a best prediction model from among the multiple prediction models based on the derived multiple index values. In the reference embodiment, the learning unit 13 identifies the maximum value (maximum index value) from among the multiple index values. Then, the learning unit 13 selects the prediction model having the maximum index value as the best prediction model.

[0048] In the verification phase, the learning unit 13 may generate a table (evaluation result table) that shows the evaluation results of each of the multiple prediction models. Table TB in FIG. 2 is an example of an evaluation result table. In TB, one index value is recorded so that there is a one-to-one correspondence between one machine learning algorithm and one hyperparameter set (e.g., Para1).

[0049] Figure 2 illustrates four types of machine learning algorithms: NN, LASSO (lasso regression), RIDGE (ridge regression), and SVR. Figure 2 illustrates two hyperparameter sets, Para1 and Para2, from among the multiple hyperparameter sets. Para1 and Para2 each indicate a hyperparameter set applied to a certain machine learning algorithm (e.g., NN).

[0050] In the example in Figure 2, 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 stated above.

[0051] In the example of Figure 2, among the combinations of each of the multiple machine learning algorithms and multiple hyperparameter sets, (machine learning algorithm, hyperparameter set) = (NN, Para1) The maximum index value is obtained in this combination (see the hatched cells in TB). The maximum index value in the example of Figure 2 is 81.6%.

[0052] The learning unit 13 selects the prediction model having the maximum index value as the best prediction model from among the generated prediction models. In the example of Fig. 2, the learning unit 13 selects the prediction model corresponding to the hatched cell in TB as the best prediction model.

[0053] As described above, in the verification phase, the learning unit 13 may select the best prediction model by comprehensively evaluating the multiple prediction models generated in the learning phase. In this way, the learning unit 13 can select the best prediction model from the multiple prediction models by performing a grid search for all combinations of machine learning algorithm types and hyperparameter sets.

[0054] In the example of Fig. 2, the learning unit 13 stores the best prediction model selected as described above in the storage unit 90. This allows the prediction device 20 to use the best prediction model, thereby further improving the prediction accuracy in the prediction phase.

[0055] (An example of processing in the prediction device 20) Next, referring back to Fig. 1, an example of processing in the prediction device 20 following the series of processing (learning phase and verification phase) in the model generation device 10 will be described. The processing in the prediction device 20 may be referred to as a prediction phase. The prediction device 20 in the example of Fig. 1 includes a data acquisition unit 21, a feature derivation unit 22, a prediction calculation unit 23, and a prediction result output unit 24.

[0056] From the viewpoint of distinguishing between the data acquisition unit 11 and the data acquisition unit 21, the data acquisition unit 11 may be referred to as a first data acquisition unit, and the data acquisition unit 21 may be referred to as a second data acquisition unit. Similarly, from the viewpoint of distinguishing between the feature derivation unit 12 and the feature derivation unit 22, the feature derivation unit 12 may be referred to as a first feature derivation unit, and the feature derivation unit 22 may be referred to as a second feature derivation unit.

[0057] In the example of FIG. 1, the memory unit 90 stores actual data DZP used in the prediction phase. In the prediction phase, a demand energy prediction may be performed in real time. Therefore, the actual data DZP may be updated in real time. The data acquisition unit 21 acquires the DZP as input data for the prediction phase. The DZP is assumed to have the same data structure as the DZ.

[0058] As described above, the prediction model generated by the model generation device 10 prior to the prediction phase is stored in the storage unit 90. Therefore, the data acquisition unit 21 acquires the prediction model from the storage unit 90.

[0059] The feature derivation unit 22 derives features in the prediction phase from the DZP by performing the same processing on the DZP as the above-described feature derivation unit 12. The features in the prediction phase may be referred to as prediction features.

[0060] As is clear from the above description, in the prediction phase, features of the same type as the training features and the evaluation features are derived as prediction features. Therefore, in the example of the reference embodiment, lag features and rolling features are derived as prediction features.

[0061] In the example of the reference embodiment, the feature derivation unit 22 derives a lag feature in the prediction phase by setting the above-mentioned first lag for the amount of demand energy in the prediction phase. Therefore, the feature derivation unit 22 acquires, as a lag feature in the prediction phase, the "amount of demand energy one day before" corresponding to the amount of demand energy at a certain point in time in the prediction phase.

[0062] Furthermore, the feature derivation unit 22 derives a rolling feature for the prediction phase by setting the second lag described above for the amount of demand energy in the prediction phase. Therefore, the feature derivation unit 22 acquires, as a rolling feature for the prediction phase, an average value of the "amount of demand energy 7 days ago" and the "amount of demand energy 14 days ago," which correspond to the amount of demand energy at a certain point in time in the prediction phase.

[0063] The prediction calculation unit 23 supplies the prediction features (e.g., lag features and rolling features in the prediction phase) derived by the feature derivation unit 22 to the prediction model as explanatory variables. The prediction model outputs a predicted value of the amount of energy demand corresponding to the explanatory variables as a response variable. In this way, the prediction calculation unit 23 inputs the prediction features to the prediction model as explanatory variables, thereby causing the prediction model to output a predicted value of the amount of energy demand.

[0064] The prediction result output unit 24 generates prediction result data based on the predicted value of the energy demand output by the prediction model. For example, the prediction result output unit 24 may generate data in which each predicted value is associated with each time point as the prediction result data. The prediction result output unit 24 outputs the generated prediction result data.

[0065] The prediction result output unit 24 may convert the predicted value of the energy demand (unit: kWh) into another unit. As an example, the prediction result output unit 24 may convert the predicted value of the energy demand into a predicted value of the power demand (unit: kW). In this case, the prediction result output unit 24 may generate, as prediction result data, data in which each converted predicted value is associated with each time point.

[0066] As described above, according to the reference embodiment, the prediction device 20 can predict the amount of energy demand in a facility using the prediction model generated by the model generation device 10. As an example, by using the above-described valid model, the amount of energy demand can be predicted with high accuracy. Furthermore, by using the above-described best prediction model, the amount of energy demand can be predicted with even higher accuracy.

[0067] [Embodiment 1] In the above-described embodiment, a method for finding the best predictive model through a grid search of all combinations of machine learning algorithm types and hyperparameter sets has been described as an example. However, this method requires comprehensive generation and evaluation of predictive models for all combinations of machine learning algorithm types and hyperparameter sets. This increases the calculation time of the model generation device.

[0068] From the perspective of reducing the calculation time of the model generation device, it is possible to narrow down the number of candidates for each machine learning algorithm and hyperparameter set, generate multiple predictive models, and then select an appropriate predictive model from among the multiple predictive models generated.

[0069] However, for example, if the candidate machine learning algorithms and hyperparameter sets are not narrowed down appropriately, there is a risk that no valid model will be obtained within the range of machine learning algorithms and hyperparameter sets searched. Thus, it can be said that there is room for improvement in the method described in the reference embodiment.

[0070] Therefore, in the first embodiment, a method is described that can find a prediction model (hereinafter referred to as a "better prediction model") that has a better (higher) index value without performing a grid search. According to the method of the first embodiment, for example, it is possible to more reliably obtain an appropriate model while reducing the calculation time of the model generation device. The model generation device 10A described below was newly created by the inventors of the present application based on this idea.

[0071] (Overview of information processing device 1A) FIG. 3 is a block diagram showing an example of the configuration of an information processing device 1A in embodiment 1. FIG. 3 is a diagram paired with the above-mentioned FIG. 1. In embodiment 1, differences from the reference embodiment will be described. The information processing device 1A in the example of FIG. 3 includes a control device 9A instead of the control device 9. The control device 9A includes a model generation device 10A instead of the model generation device 10. The model generation device 10A includes a learning unit 13A instead of the learning unit 13.

[0072] Unlike learning unit 13, learning unit 13A selects a better prediction model from among multiple prediction models based on the index values ​​of each of the multiple prediction models in accordance with a predetermined optimization method other than grid search (for example, the Nelder-Mead method described below).

[0073] In the example of the first embodiment, any one of the multiple types of machine learning algorithms described in the reference embodiment is referred to as the “algorithm of interest.” In the first embodiment, a case where the learning unit 13A generates multiple prediction models for each algorithm of interest is illustrated.

[0074] In the first embodiment, the learning unit 13A performs a search for a plurality of prediction models for each of a plurality of types of machine learning algorithms in accordance with a predetermined optimization method. Specifically, the learning unit 13A advances the search by repeating a learning phase and a verification phase for each of a plurality of types of machine learning algorithms in accordance with a predetermined optimization method. The processing flow in the first embodiment is outlined below.

[0075] (1) In the learning phase, the learning unit 13A generates multiple predictive models corresponding to the algorithm of interest based on the training features acquired from the feature derivation unit 12 by applying multiple hyperparameter settings corresponding to the algorithm of interest.

[0076] (2) In the verification phase, the learning unit 13A acquires index values ​​for each of the multiple prediction models generated in the learning phase based on the verification features acquired from the feature derivation unit 12. Specifically, the learning unit 13A acquires index values ​​for each of the multiple prediction models by inputting the verification features into each of the multiple prediction models.

[0077] Then, the learning unit 13A selects a better prediction model from among the plurality of prediction models based on the index values ​​of each of the plurality of prediction models through a series of processes using a predetermined optimization method, which will be described later with reference to Figures 4 to 6.

[0078] (3) After completing the above steps (1) and (2) for one algorithm of interest, the learning unit 13A selects another machine learning algorithm as a new algorithm of interest. Then, the learning unit 13A executes the above steps (1) and (2) for the new algorithm of interest.

[0079] After completing the above steps (1) to (3) for each of the multiple machine learning algorithms to be searched, the learning unit 13A selects an appropriate model (a prediction model expected to be appropriate) from among the multiple better prediction models selected for each of the multiple machine learning algorithms. As an example, the learning unit 13A selects the prediction model with the largest index value from among the multiple better prediction models as the appropriate model.

[0080] The processes in the first embodiment after the selection of the appropriate model are the same as those in the reference embodiment, except that the "best prediction model" in the reference embodiment is replaced with the "appropriate model."

[0081] (An example of a predetermined optimization method) Next, a specific processing flow by the predetermined optimization method in embodiment 1 will be described with reference to Fig. 4 to Fig. 6. Fig. 4 is a flowchart showing an outline of the main processing flow of the learning unit 13A. Below, as an example, a case where the algorithm of interest is NN will be described.

[0082] 4, in step S1, the learning unit 13A acquires the number of items N in the search parameter list (total number of lists N). The search parameter list refers to a list of parameters (hereinafter referred to as "search parameters") used in searches for multiple prediction models.

[0083] In the example of the first embodiment, the learning unit 13A performs a search for a prediction model according to multiple hyperparameter settings for each of multiple machine learning algorithms. Therefore, the search parameter list in the first embodiment includes only the hyperparameter settings for the algorithm of interest.

[0084] That is, the number of items N in the search parameter list in the first embodiment is equal to the number of items in the hyperparameter settings of the algorithm of interest. For the sake of clarity, the following will exemplify a case where N=2. That is, in the following example, a case where the number of items in the hyperparameter settings of the algorithm of interest is 2 will be described.

[0085] Next, in step S2, the learning unit 13A sets N+1 combinations of random numerical values ​​from the items in the search parameter list. Then, the learning unit 13A sets N+1 lattice points P1 to P2 in the N-dimensional space corresponding to the combinations. N+1 are generated as initial points. The lattice points in this specification refer to points that form a spatial lattice.

[0086] In the first embodiment, since N+1=3, the learning unit 13A calculates P1=(HA1,HB1) P2=(HA2,HB2) P3=(HA3,HB3) These three grid points P1 to P3 are generated as initial points.

[0087] In the description of the first embodiment, an arbitrary natural number j that satisfies 1≦j≦N+1 is used as an index indicating the number of a lattice point. In the example of the first embodiment, 1≦j≦3. Pj is the jth grid point. j is an index indicating the first hyperparameter of the algorithm in question (e.g., batch size of a neural network). j is an index indicating the second hyperparameter of the algorithm in question (e.g., the maximum number of times a NN can learn).

[0088] Next, in step S3, the learning unit 13A calculates the lattice points P1 to P2, which are the initial points set in step S2. N+1 Specifically, the learning unit 13A generates a plurality of prediction models by using a plurality of search parameters corresponding to the lattice points P1 to P N+1 A plurality of prediction models are generated based on the training features using a plurality of search parameters corresponding to

[0089] In the example of the first embodiment, the learning unit 13A calculates the lattice points P1 to P N+1 In this way, the learning unit 13A generates a plurality of prediction models by using a plurality of hyperparameter settings corresponding to the lattice points P1 to P N+1 In the example of the first embodiment, the learning unit 13A generates three prediction models.

[0090] Next, in step S4, the learning unit 13A acquires index values ​​for each of the multiple prediction models generated in step S3 based on the verification features. Specifically, the learning unit 13A acquires index values ​​for each of the multiple prediction models by inputting verification features into each of the multiple prediction models. In the example of embodiment 1, the learning unit 13A acquires three index values.

[0091] In the first embodiment, a function f is set for evaluating an index value (in the example of the first embodiment, the prediction accuracy in %). The function f is an objective function related to the prediction accuracy. As an 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.

[0092] In the example of the first embodiment, the function f is expressed by the following formula (1): f=100-prediction accuracy [%]…(1) The function f expressed by the formula (1) corresponds to one prediction model. In other words, the function f corresponds to one grid point P j Therefore, the lattice point P j If the coordinates of the grid point P are known, j The value of the function f corresponding to f(P j ) can be determined.

[0093] The learning unit 13A uses the lattice points P1 to P2 acquired in step S4 N+1 Using the index values ​​corresponding to N+1 f(P1)~f(P N+1 In the example of the first embodiment, three values ​​f(P1) to f(P3) are determined.

[0094] In the following description, (i) the lattice point P j (ii) obtains an index value of the prediction model based on the validation features; and (iii) calculates a function f from f(P j The series of processes that determine the value of f(P j ) is denoted as "derive".

[0095] Next, in step S5, the learning unit 13A selects the lattice points P1 to P2 as initial points based on the N+1 index values ​​acquired in S4. N+1 In the first embodiment, the learning unit 13A ranks f(P1) to f(P N+1 ) based on the grid points P1 to P N+1 More specifically, the learning unit 13A ranks the three lattice points P1 to P3 based on the three values ​​f(P1) to f(P3).

[0096] In Embodiment 1, the learning unit 13A designates, as the best point B, the point among the three initial points P1 to P3 that has been assigned the best prediction accuracy (the highest prediction accuracy). (m) m is the number of optimization repetitions (number of optimization iterations) in step S6 described below. m is an integer that satisfies 1 ≤ m ≤ q. q is the upper limit value of the number of optimization repetitions. q is set in advance.

[0097] Then, the learning unit 13A designates, as the worst point W, the point among the three initial points P1 to P3 that has been assigned the worst prediction accuracy (the lowest prediction accuracy). (m) Next, the learning unit 13A designates, as the second-worst point BW, the point among the three initial points P1 to P3 that has been assigned the second-worst prediction accuracy (the second-lowest prediction accuracy) after the worst point. (m)

[0098] As an example, consider the case where f(P1) < f(P2) < f(P3). In this case, lattice point P1 is designated as the best point B (m) lattice point P2 is designated as the second-worst point BW (m) and lattice point P3 is designated as the worst point W (m)

[0099] As another example, when f(P2) = f(P3), one of lattice point P2 or lattice point P3 (e.g., lattice point P2) may be designated as the worst point W (m) and the other (e.g., lattice point P3) may be designated as the second-worst point BW (m)

[0100] As yet another example, when f(P1) = f(P2), one of lattice point P1 or lattice point P2 (e.g., lattice point P1) may be designated as the best point B (m) and the other (e.g., lattice point P2) may be designated as the second-worst point BW (m)

[0101] Next, in step S6, the learning unit 13A, after serializing in step S5, for lattice points P1 to P N+1 ​​​​Based on this, optimization is performed using a predetermined optimization method. Details of step S6 will be described later.

[0102] In the first embodiment, the Nelder-Mead method is used as the predetermined optimization method. The Nelder-Mead method is also known as 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.

[0103] The objective function f in the first embodiment is expressed by the above-mentioned formula (1). Therefore, by using the Nelder-Mead method, it is possible to obtain a minimum value (ideally, the smallest value) of f. In other words, by using the Nelder-Mead method, it is possible to obtain a maximum value (ideally, the largest value) of prediction accuracy. In this way, in the first embodiment, it is possible to find a better prediction model by using the Nelder-Mead method.

[0104] Figure 5 is a diagram for explaining the concept of the Nelder-Mead method. For simplicity, Figure 5 illustrates the case of N=2 (two-dimensional case). The X1 axis in Figure 5 corresponds to the first hyperparameter of the algorithm in question, and the X2 axis corresponds to the second hyperparameter of the algorithm in question. Figure 5 shows the B (m) , W (m) , and B.W. (m) The dotted lines in the example of Fig. 5 indicate the contour lines of the prediction accuracy. The contour lines of the prediction accuracy can be interpreted as the contour lines of the objective function.

[0105] As shown in Figure 5, the Nelder-Mead method finds that as the optimization iteration progresses (increasing m), the combination of multiple hyperparameters (i.e., B (m) , W (m) , and B.W. (m) In this way, the Nelder-Mead method can optimize multiple hyperparameters in the algorithm of interest so as to reach a local optimum in prediction accuracy (in the example of the first embodiment, a maximum value of prediction accuracy).

[0106] FIG. 6 is a flowchart illustrating the process flow for executing optimization. FIG. 6 shows the process content of step S6 in FIG. 4 in more detail. FIG. 6 illustrates optimization using the Nelder-Mead method. The process in FIG. 6 starts from a state where m=1.

[0107] First, in step S11, the learning unit 13A 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) Ask for.

[0108] Specifically, the learning unit 13A calculates the following equation (2):

number

[0109] Next, in step S12, the learning unit 13A calculates the line segment W (m) G (m) External division point R that divides (m) Specifically, the learning unit 13A calculates the following equation (3): R (m) =2G (m) -W(m) …(3) As shown above, the external division point R (m) Ask for.

[0110] Next, in step S13, the learning unit 13A uses the above-mentioned function f to calculate f(R (m) ) is derived. Then, the learning unit 13A derives f(R (m) )≦f(B (m) ) …(4) It is determined whether the following condition is satisfied. In the example of the first embodiment, the condition shown in equation (4) is referred to as the first determination condition. If the first determination condition is satisfied, the process proceeds to step S14. If the first determination condition is not satisfied, the process proceeds to step S17, which will be described later.

[0111] If the first determination condition is satisfied, in step S14, the learning unit 13A calculates the line segment W (m) R (m) External division point E, which divides (m) Specifically, the learning unit 13A calculates the following equation (5): E (m) =3G (m) -2W (m) …(5) As shown, the external division point E (m) Ask for.

[0112] Next, in step S15, the learning unit 13A uses the function f to calculate f(E (m) ) is derived. Then, the learning unit 13A derives f(E (m) )≦f(R (m) ) …(6) It is determined whether the following condition is satisfied. In the example of the first embodiment, the condition shown in equation (6) is referred to as the second determination condition. If the second determination condition is satisfied, the process proceeds to step S16. If the second determination condition is not satisfied, the process proceeds to step S18.

[0113] If the second judgment condition is satisfied, in step S16, the learning unit 13A calculates the worst point W (m) is removed from the optimization target, and point E(m) Then, the process proceeds to step S23, which will be described later.

[0114] If it is determined in step S13 that the first determination condition is not satisfied, in step S17, the learning unit 13A f(R (m) )≦f(BW (m) ) …(7) It is determined whether the following condition is satisfied. In the example of the first embodiment, the condition shown in equation (7) is referred to as the third determination condition. If the third determination condition is satisfied, the process proceeds to step S18. If the third determination condition is not satisfied, the process proceeds to step S19.

[0115] If the third judgment condition is satisfied, in step S18, the learning unit 13A calculates the worst point W (m) is removed from the optimization target, and point R (m) is added to the optimization targets, and the process proceeds to step S23.

[0116] If the third determination condition is not satisfied, in step S19, the learning unit 13A (m) G (m) Midpoint S of (m) Specifically, the learning unit 13A calculates the following equation (8): S (m) =(1 / 2)×(G (m) +W (m) ) …(8) As shown, the midpoint S (m) Ask for.

[0117] Next, in step S20, the learning unit 13A uses the function f to calculate f(S (m) ) is derived. Then, the learning unit 13A derives f(S (m) )≦f(W (m) ) …(9) It is determined whether the following condition is satisfied. In the example of the first embodiment, the condition shown in equation (9) is referred to as the fourth determination condition. If the fourth determination condition is satisfied, the process proceeds to step S21. If the fourth determination condition is not satisfied, the process proceeds to step S22.

[0118] If the fourth determination condition is satisfied, in step S21, the learning unit 13A calculates the worst point W (m) is removed from the optimization target, and point S (m) Then, the process proceeds to S43.

[0119] If the fourth determination condition is not satisfied, in step S22, the learning unit 13A calculates the line segment B (m) P j (m) Midpoint P of j (m)’ Specifically, the learning unit 13A calculates the following equation (10): P j (m)’ =(1 / 2)×(P j (m) +B (m) ) …(10) As shown, the midpoint P j (m)’ Then, the learning unit 13A calculates the distance between each point P j (m) At each point P j (m)’ Then, the process proceeds to step S23.

[0120] In step S23, the learning unit 13A calculates the value of each point P based on the function f in the same manner as in step S5. j (m) As a result of the process in step S23, for example, the best point B (m) may be changed.

[0121] Next, in step S24, the learning unit 13A determines whether or not a predetermined setting determination condition is satisfied. If the setting determination condition is satisfied, the process proceeds to step S25. If the setting determination condition is not satisfied, the process returns to step S11. Thus, in the example of FIG. 6, the processes from step S11 to step S23 are repeated until the setting determination condition is satisfied. Note that when returning from step S24 to step S11, the learning unit 13A increments m by 1.

[0122] The set determination conditions in the first embodiment are, for example, the following conditions 1 to 6: (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) Best point B (m) was not changed for more than a given number of iterations; (Condition 5) Best point B (m) The prediction accuracy at 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); It may include at least one of:

[0123] In condition 1, Vm is the j (m) represents the volume of the hyperpolyhedron formed by each initial point, and V0 represents the volume of the hyperpolyhedron formed by each initial point.

[0124] In addition, α in condition 6 is

number

[0125] 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 locally optimal solution of the function f has been found" and "a condition under which further search for the locally optimal solution should be terminated." The "state in which a locally optimal solution of the function f has been found" can also be rephrased as "a state in which a locally optimal solution for prediction accuracy in the example of embodiment 1 has been found."

[0126] Referring again to FIG. 6, if it is determined in step S24 that the set determination condition is satisfied, in step S25, the learning unit 13A determines the best point B (m) In the example of the first embodiment, the learning unit 13A uses the combination of search parameters in the best point B (m) We adopt a combination of hyperparameters in

[0127] Best point B in step S25 (m) is the final best point obtained as a result of the optimization process of FIG. 6. In the example of the first embodiment, the best point B (m) The coordinates of B (m) =(HA better ,HB better ) The learning unit 13A is expressed as HA better and H.B. better The hyperparameter setting items corresponding to the above are selected as better hyperparameter settings for the algorithm of interest.

[0128] In the example of the first embodiment, the learning unit 13A better Select as the first hyperparameter of the featured algorithm (e.g., better batch size for NN), and HB better Select as a better second hyperparameter of the featured algorithm (e.g., a better maximum learning iteration of the NN).

[0129] As described above, according to the process of FIG. 6, better hyperparameter settings are selected for the algorithm of interest. The learning unit 13A determines a prediction model corresponding to the better hyperparameter settings for the algorithm of interest as a better prediction model for the algorithm of interest. In other words, the learning unit 13A determines the best point B (m) The prediction model corresponding to the above is determined as the better prediction model in the algorithm of interest.

[0130] The learning unit 13A in the first embodiment executes the optimization process of FIG. 6 for each of the multiple types of machine learning algorithms, and determines a better prediction model for each of the multiple types of machine learning algorithms.

[0131] Next, the learning unit 13A in the first embodiment selects an appropriate model from among the plurality of better prediction models determined for each of the plurality of types of machine learning algorithms. As an example, the learning unit 13A selects, as the appropriate model, the better prediction model having the highest index value from among the plurality of better prediction models.

[0132] (Additional information about the processing in Figure 6) In the example of FIG. 6, the external division point R (m) The coordinates of the external division point R do not necessarily take integer values. Therefore, if a hyperparameter that requires an integer (e.g., the maximum number of learning iterations in a neural network) is included in the search parameter list, (m) There may be no hyperparameter settings corresponding to each coordinate of f, and it may not be possible to generate a predictive model. (m) ) may not be possible to derive.

[0133] Therefore, the learning unit 13A calculates the external division point R (m) After finding the external division point R (m) The coordinates of the external division point R are all integers. (m) The grid point R corresponding to (m)’ In this case, the learning unit 13A may calculate the lattice point R (m)’ At the external dividing point R (m)may be used instead of

[0134] The method of integerization is arbitrary. For example, integerization may be performed by rounding off the first decimal place of the value of each coordinate. As another example, integerization may be performed by rounding down the first decimal place of the value. Alternatively, integerization may be performed by rounding up the first decimal place of the value.

[0135] By converting the coordinates to integers, the external division point R can be found even when the search parameter list contains hyperparameters that require integers. (m) The grid point R corresponding to (m)’ Using f(R (m)’ ) can be derived. Therefore, the learning unit 13A can be made to execute a series of optimization processes.

[0136] The learning unit 13A calculates the external division point R (m) Among the coordinates of the hyperparameters included in the search parameter list, only the coordinates corresponding to the hyperparameters that need to be integers may be converted to integers. Therefore, the learning unit 13A may maintain the coordinates as they are for the coordinates corresponding to the hyperparameters that do not need to be integers (e.g., learning rate in a NN) among the hyperparameters included in the search parameter list. On the other hand, the learning unit 13A converts the coordinates corresponding to the hyperparameters that need to be integers among the hyperparameters included in the search parameter list into integers.

[0137] In the example of FIG. 6, the external division point E determined by the above equation (5) (m) Similarly, the coordinates of the external division point S determined by the above equation (8) are not necessarily integers. (m) The coordinates of the points do not necessarily take integer values. The midpoint P defined by the above equation (10) j (m)’ The coordinates of the external division point E are not necessarily integers. (m) , midpoint S (m) , midpoint P j (m)’ Regarding the above, the external division point R(m) As in the example above, each coordinate may be converted into an integer by the learning unit 13A.

[0138] (Effects of the First Embodiment) According to the model generating device 10A (particularly, the learning unit 13A) of the first embodiment, an appropriate model can be determined by executing a predetermined optimization method (e.g., the Nelder-Mead method). That is, according to the first embodiment, unlike the reference embodiment, it is possible to efficiently find an appropriate model without requiring a grid search. Thus, according to the first embodiment, the time required to find an appropriate prediction model for predicting the amount of power demand in a facility can be reduced compared to the reference embodiment.

[0139] In the example of the first embodiment, by executing a predetermined optimization method for each of multiple types of machine learning algorithms, the search parameters in the optimization method can be narrowed down to only the hyperparameters of the machine learning algorithm, thereby effectively reducing the time required to find an optimal model based on the optimization method.

[0140] It should be noted that in one aspect of the present invention, the number of types of machine learning algorithms does not have to be two or more. Therefore, for example, in embodiment 1, an appropriate model may be determined by executing a predetermined optimization method for only one type of machine learning algorithm.

[0141] The technology of Patent Document 1 is based on the idea of ​​selectively using multiple prediction models (specifically, multiple neural networks) depending on the situation in order to improve the prediction accuracy of the heat load in a thermal storage utilization system. Therefore, the technology of Patent Document 1 does not assume the use of some kind of optimization method to select one representative prediction model from among multiple prediction models. As is clear from this, the first embodiment is based on an idea that is completely different from the technology of Patent Document 1.

[0142] [Embodiment 2] In the first embodiment, the learning unit 13A performs a search for a predictive model using multiple hyperparameter settings for each of multiple types of machine learning algorithms according to a predetermined optimization method. However, the type of machine learning algorithm can also be a target of the search using the predetermined optimization method. Therefore, the model generation device 10 can also perform a search using multiple hyperparameter settings for multiple types of machine learning algorithms according to the predetermined optimization method.

[0143] Therefore, in the second embodiment, unlike the first embodiment, a case will be exemplified in which the search parameter list further includes the ranking of the machine learning algorithms. The ranking of the machine learning algorithms may be assigned to each of the multiple types of machine learning algorithms according to a predetermined ranking. For example, each of the multiple types of machine learning algorithms may be ranked in advance, such as "rank 1: NN, rank 2: SVR, rank 3: LASSO, ...". In this case, it is preferable that the hyperparameters specific to each machine learning algorithm (e.g., the maximum number of learning times in NN) are fixed.

[0144] The learning unit 13A executes the optimization method described above using the search parameter list in embodiment 2. For example, in the learning phase in embodiment 2, the learning unit 13A generates multiple prediction models by executing multiple types of machine learning algorithms to which multiple hyperparameter settings are respectively applied. Then, in the verification phase in embodiment 2, the learning unit 13A acquires index values ​​for each of the multiple prediction models generated in the learning phase.

[0145] As a result, the learning unit 13A in the second embodiment can find a better prediction model from among a plurality of prediction models obtained by executing a plurality of types of machine learning algorithms to which a plurality of hyperparameter settings are respectively applied. In the second embodiment, the learning unit 13A may determine the found better prediction model as the appropriate model.

[0146] As described above, in the second embodiment, unlike the first embodiment, it is not necessary to search for a predictive model for each of a plurality of types of machine learning algorithms. Therefore, in the second embodiment, the time required to find an appropriate model can be further effectively reduced.

[0147] [Embodiment 3] In the first embodiment, the Nelder-Mead method is exemplified as an optimization method for deriving a minimum value of the function f (in other words, an optimization method for deriving a maximum value of the prediction accuracy). However, 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.

[0148] 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.

[0149] In addition, when solving one-dimensional nonlinear optimization problems, the golden section method can also be used as an optimization method. The golden section method can solve one-dimensional nonlinear optimization problems without relying on differential calculus. With the golden section method, the maximum value of prediction accuracy expressed as a one-dimensional nonlinear function can be derived without relying on differential calculus.

[0150] As will be apparent to those skilled in the art, the optimization method according to one embodiment of the present invention is not particularly limited. However, from the viewpoint of facilitating the implementation of the algorithm, the optimization method 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.

[0151] [Software implementation example] The functions of the information processing devices 1 to 1A (hereinafter referred to as "devices") 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 each part included in the control devices 9 to 9A).

[0152] 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.

[0153] 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.

[0154] 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.

[0155] As described above, 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).

[0156] 〔summary〕 A model generation device according to aspect 1 of the present invention is a model generation device that generates a prediction model for predicting the amount of electricity demand in a facility, deriving features for deriving a predicted value of the amount of electricity demand from time series data showing the time progression of the actual value of the amount of electricity demand, generating a plurality of prediction models based on the features by executing a machine learning algorithm, deriving a plurality of index values ​​indicating the quality of each of the plurality of prediction models, and selecting an appropriate model from among the plurality of prediction models by searching for a prediction model with a better index value based on the plurality of index values ​​according to a predetermined optimization method.

[0157] In the model generation device of aspect 2 of the present invention, in aspect 1, a plurality of predictive models may be generated for each of a plurality of types of machine learning algorithms using a plurality of hyperparameter settings applied to each of the plurality of types of machine learning algorithms, and the appropriate model may be selected by searching for the predictive model having a better index value based on the plurality of index values ​​in accordance with the optimization method.

[0158] The model generation device according to aspect 3 of the present invention may generate a plurality of the predictive models by executing a plurality of types of the machine learning algorithms to which a plurality of hyperparameter settings are respectively applied in the above-mentioned aspect 1.

[0159] In the model generating device according to a fourth aspect of the present invention, in any one of the first to third aspects, the optimization method may be the Nelder-Mead method or the Powell method.

[0160] A model generation method according to aspect 5 of the present invention is a model generation method for generating a predictive model for predicting the amount of electricity demand in a facility, and includes the steps of: deriving features for deriving a predicted value of the amount of electricity demand from time series data showing the time progression of actual values ​​of the amount of electricity demand; generating a plurality of predictive models based on the features by executing a machine learning algorithm; deriving a plurality of index values ​​indicating the quality of each of the plurality of predictive models; and selecting an appropriate model from the plurality of predictive models by searching for the predictive model having a better index value based on the plurality of index values ​​according to a predetermined optimization method.

[0161] [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]

[0162] 1. Information processing equipment 1A Information processing equipment 10 Model generation device 10A Model Generator 11 Data Acquisition Section 12 Feature extraction unit 13 Learning Department 13A Learning Department 20 Prediction Device 21 Data Acquisition Section 22 Feature extraction unit 23 Prediction calculation unit 24 Prediction result output section

Claims

1. A model generation device that generates a prediction model for predicting an amount of power demand in a facility, deriving a feature amount for deriving the predicted value of the amount of power demand from time-series data indicating a time transition of an actual value of the amount of power demand; generating a plurality of predictive models based on the features by executing a machine learning algorithm; deriving a plurality of index values ​​indicative of the quality of each of the plurality of prediction models; A model generation device that selects an appropriate model from among a plurality of the prediction models by searching for the prediction model having a better index value based on a plurality of the index values ​​according to a predetermined optimization method.

2. For each of the plurality of types of machine learning algorithms, generating a plurality of the predictive models using a plurality of hyperparameter settings applied to each of the plurality of types of machine learning algorithms; The model generating device according to claim 1 , wherein the appropriate model is selected by searching for the prediction model having a better index value based on a plurality of the index values ​​according to the optimization technique.

3. The model generation device according to claim 1 , wherein the device generates the plurality of predictive 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. A model generation method for generating a prediction model for predicting an amount of power demand in a facility, comprising: deriving a feature amount for deriving the predicted value of the amount of power demand from time-series data indicating a time transition of an actual value of the amount of power demand; generating a plurality of predictive models based on the features by executing a machine learning algorithm; deriving a plurality of indicators indicative of the quality of each of the plurality of predictive models; and selecting an appropriate model from among the plurality of predictive models by searching for the predictive model having a better index value based on the plurality of index values ​​according to a predetermined optimization technique.

Images