Method and system for dynamically selecting photovoltaic power generation prediction model

Abstract

The present invention relates to a method for dynamically selecting a photovoltaic power generation prediction model. More specifically, the method for dynamically selecting a photovoltaic power generation prediction model may comprise the steps of: training a dynamic prediction model selector for photovoltaic power generation prediction; and selecting, via the trained dynamic prediction model selector, an optimal prediction model for minimizing a photovoltaic power generation prediction error with respect to weather data and solar information data. The training step includes: (a) a step of learning, with an auto-encoder, weather data and solar information data for a plurality of photovoltaic power plants to extract a plurality of latent variables; (b) a step of classifying the plurality of extracted latent variables into a plurality of clusters on the basis of similarity; (c) a step of training, with the cluster-specific latent variables, a plurality of different prediction models; and (d) a step of dynamically selecting an optimal prediction model for each of the clusters from among the plurality of trained prediction models.

Description

Dynamic selection method and system for solar power generation prediction models

[0001] The present invention relates to a method for dynamically selecting a solar power generation prediction model, and more specifically, to a method for dynamically selecting a solar power generation prediction model capable of dynamically selecting an optimal solar power generation prediction model according to a weather event by learning weather data and solar information data for a plurality of solar power plants using an autoencoder to extract latent variables, and constructing an optimal solar power generation prediction model for each cluster clustered based on the latent variables.

[0002] The Virtual Power Plant (VPP) market integrates and manages power plants distributed across multiple regions, and trades the electricity they produce. Solar power, a renewable energy source, is a key power source in the VPP market, accounting for more than half of the renewable energy supply.

[0003] However, solar power generation suffers from greater fluctuations in power generation and greater unpredictability compared to other power generation methods, raising concerns about the stability of transactions for distributed resources. Transactions in the virtual power plant market are conducted by presenting the next day's power generation volume and delivering the power for that time.

[0004] The power exchange then redistributes the electricity generated during that time and provides it to buyers. Therefore, to ensure the smooth provision of solar-generated power to the virtual power plant market, it is necessary to accurately predict the power generation volume for the next 24 hours.

[0005] Typically, solar power generation is predicted based on meteorological factors. These factors can trigger various weather events (e.g., typhoons, monsoon rains, heavy snow, heat waves, etc.). To accurately predict solar power generation based on these events, it is necessary to select a prediction model that accurately distinguishes the meteorological factors associated with these events and minimizes prediction errors.

[0006] The problem to be solved by the present invention is to provide a method and system for dynamically selecting a solar power generation prediction model that can dynamically select an optimal solar power generation prediction model according to a meteorological event by learning meteorological data and solar information data for multiple solar power plants using an autoencoder to extract latent variables, and constructing an optimal solar power generation prediction model for each cluster based on the latent variables.

[0007] The tasks of the present invention are not limited to the tasks mentioned above, and other tasks not mentioned will be clearly understood by those skilled in the art of the present invention from the description below.

[0008] In order to solve the above-described problem, a method for dynamically selecting a solar power generation prediction model according to one embodiment of the present invention may include a step of learning a dynamic selector for a prediction model for solar power generation prediction, and a step of selecting an optimal prediction model that minimizes a solar power generation prediction error for weather data and solar information data through the learned dynamic selector for a prediction model.

[0009] At this time, the learning step may include (a) a step of learning weather data and solar information data for multiple solar power plants using an autoencoder to extract latent variables, (b) a step of classifying the extracted multiple latent variables into multiple clusters based on similarity, (c) a step of learning multiple different prediction models for the latent variables for each cluster, and (d) a step of dynamically selecting an optimal prediction model for each cluster among the multiple learned prediction models.

[0010] In addition, the learning step may further include a step of collecting weather data and solar information data for the plurality of solar power plants and a step of generating learning data by mixing Gaussian noise into the collected data.

[0011] In addition, the step (a) may include an encoding step for extracting latent variables by dimensionally reducing the generated learning data, and a decoding step for restoring the learning data based on the latent variables.

[0012] In addition, the step (b) may include a first process of randomly selecting k center positions based on the latent variables and generating k clusters, calculating the straight-line distance from each latent variable to the randomly selected center position and including the cluster including the center position with the closest distance, and a second process of setting the position where the sum of the distances to the latent variables within each cluster is minimized as a new center position.

[0013] In addition, steps (a) and (b) are sequentially repeated until mutual conditions are satisfied, and when the restoration error is minimized in step (a) and the center position of each cluster does not change in step (b), the classified cluster can be determined as the final classified cluster.

[0014] In addition, the step (d) may include, for each cluster, a step of sorting the learned multiple prediction models in order of the highest average value of the model performance evaluation indicators; and a step of selecting an ensemble model including the prediction model with the highest average value or a preset number of prediction models in order of the highest average value as the optimal prediction model.

[0015] In addition, a solar power generation prediction model dynamic selection system according to an embodiment of the present invention may include a learning unit that learns a prediction model dynamic selector for solar power generation prediction, and a prediction model dynamic selector that selects an optimal prediction model that minimizes a solar power generation prediction error when weather data and solar information data are input.

[0016] In addition, the dynamic selector for the prediction model may include an autoencoder that learns weather data and solar information data for a plurality of solar power plants to extract latent variables, a clustering unit that classifies the extracted plurality of latent variables into a plurality of clusters based on similarity, and a selection unit that dynamically selects an optimal prediction model for each cluster among a plurality of different prediction models learned for the latent variables for each cluster.

[0017] In addition, the method further includes a collection unit that collects weather data and solar information data for the plurality of solar power plants, and the autoencoder can learn training data generated by mixing Gaussian noise into the collected weather data and solar information data.

[0018] In addition, the autoencoder may include an encoder that reduces the dimensionality of the generated learning data to extract latent variables, and a decoder that restores the learning data based on the latent variables.

[0019] In addition, the clustering unit may include a first process of randomly selecting k center positions based on the latent variables and generating k clusters, calculating a straight-line distance from each latent variable to the randomly selected center position and including the cluster including the center position with the closest distance, and a second process of setting the position where the sum of distances to latent variables within each cluster is minimized as a new center position.

[0020] In addition, the autoencoder and clustering unit perform sequential iterative learning until mutual conditions are satisfied, and when the restoration error of the autoencoder is minimized and the center position of each cluster in the clustering unit does not change, the classified cluster can be determined as the final classified cluster.

[0021] In addition, the selection unit can sort the plurality of learned prediction models in order of the highest average value of the model performance evaluation indicators, and then select the prediction model with the highest average value or an ensemble model including a preset number of prediction models in order of the highest average value as the optimal prediction model.

[0022] Specific details of other embodiments are included in the detailed description and drawings.

[0023] A method and system for dynamically selecting a solar power generation prediction model according to an embodiment of the present invention learns weather data and solar information data for a plurality of solar power plants using an autoencoder to extract latent variables, and constructs an optimal solar power generation prediction model for each cluster based on the latent variables, thereby dynamically selecting and providing an optimal solar power generation prediction model according to a weather event.

[0024] By training data with Gaussian noise added to weather data and solar information data for multiple solar power plants, overfitting can be prevented and latent variable extraction performance can be improved.

[0025] In addition, clusters can be classified according to weather events (e.g., typhoons, monsoon rains, heavy snow, heat waves, etc.) by setting the final cluster and each cluster center point that minimizes the sum of the distances between latent variables within each cluster while minimizing the restoration error of the autoencoder.

[0026] In addition, by selecting the optimal prediction model for each cluster of classified weather events and predicting solar power generation based on weather data and solar information data, the accuracy of power generation prediction can be improved.

[0027] The effects according to the present invention are not limited to those exemplified above, and other effects can be clearly understood by those skilled in the art from the description of the following specification.

[0028] FIG. 1 is a block diagram schematically illustrating a configuration of a solar power generation prediction model dynamic selection system according to an embodiment of the present invention.

[0029] Fig. 2 is a block diagram showing a schematic configuration of the prediction model dynamic selector of Fig. 1.

[0030] FIG. 3 is a flowchart for explaining a method for dynamically selecting a solar power generation prediction model according to one embodiment of the present invention.

[0031] Figure 4 is a flowchart for explaining the learning step (S100) of Figure 3.

[0032] Figure 5 is a diagram for explaining learning data according to one embodiment of the present invention.

[0033] FIG. 6 is a drawing for explaining a cluster according to one embodiment of the present invention.

[0034] Figure 7 is a priority table for explaining the selection of the optimal prediction model of the selection section of Figure 2.

[0035] FIG. 8 is a graph comparing predicted values ​​of a dynamically selected solar power generation prediction model according to one embodiment of the present invention.

[0036] *Explanation of key symbols in the drawing*

[0037] 100: Collection Department

[0038] 200: Learning Department

[0039] 300: Prediction Model Dynamic Selector

[0040] 310: Autoencoder

[0041] 320: Clustering section

[0042] 330: Selection Department

[0043] 400: Storage

[0044] The following merely exemplifies the principles of the invention. Therefore, those skilled in the art will be able to implement the principles of the invention and invent various devices within the scope and spirit of the invention, even if not explicitly described or illustrated herein. Furthermore, all conditional terms and embodiments listed herein are expressly intended, in principle, to facilitate understanding of the invention, and should be understood as being in no way limited to the specifically listed embodiments and conditions.

[0045] Additionally, in the following description, ordinal expressions such as first, second, etc. are intended to describe objects that are equal and independent of each other, and should be understood as having no meaning in terms of main / sub or master / slave.

[0046] The above-described purposes, features and advantages will become clearer through the following detailed description with reference to the attached drawings, so that a person having ordinary skill in the art to which the invention pertains can easily practice the technical idea of ​​the invention.

[0047] The individual features of the various embodiments of the present invention can be partially or wholly combined or combined with each other, and as can be fully understood by those skilled in the art, various technical connections and operations are possible, and each embodiment can be implemented independently of each other or can be implemented together in a related relationship.

[0048] A system and method for dynamically selecting a solar power generation prediction model according to an embodiment of the present invention can increase prediction accuracy by selecting a prediction model that can clearly distinguish meteorological factors according to a meteorological event and minimize prediction errors.

[0049] To this end, a dynamic selection system for a solar power generation prediction model according to an embodiment of the present invention learns weather data and solar information data for multiple solar power plants using an autoencoder to extract latent variables, constructs an optimal solar power generation prediction model for each cluster based on the latent variables, and dynamically selects and provides an optimal solar power generation prediction model according to a weather event.

[0050] Hereinafter, various embodiments of the present invention will be described in detail with reference to the attached drawings.

[0051] FIG. 1 is a block diagram schematically illustrating the configuration of a solar power generation prediction model dynamic selection system according to an embodiment of the present invention. FIG. 2 is a block diagram schematically illustrating the configuration of the prediction model dynamic selector of FIG. 1.

[0052] Referring to FIGS. 1 and 2, a solar power generation prediction model dynamic selection system (hereinafter referred to as the “system”) according to an embodiment of the present invention may include a collection unit (100), a learning unit (200), a prediction model dynamic selector (300), and a storage unit (400). In addition, the prediction model dynamic selector (300) may include an autoencoder (310), a clustering unit (320), and a selection unit (330).

[0053] At this time, each configuration can be implemented as a separate server or a software module, hardware module, or a combination of software and hardware on a single computing device.

[0054] In addition, FIG. 3 is a flowchart for explaining a method for dynamically selecting a solar power generation prediction model according to an embodiment of the present invention.

[0055] Figure 4 is a flowchart for explaining the learning step (S100) of Figure 3.

[0056] Below, the functions and operations of each system configuration of FIGS. 1 and 2 will be described together with the methods of FIGS. 3 and 4.

[0057] First, a dynamic selector (300) for predicting solar power generation can be learned (S100). More specifically, see Fig. 2.

[0058] The collection unit (100) can collect weather data, solar information data, and power plant data, and perform data preprocessing such as scale adjustment and feature extraction.

[0059] For example, referring to FIG. 5, weather data may include cloud (cloud cover), temperature, maximum temperature, minimum temperature, humidity, ground pressure, wind speed, wind direction, rain (precipitation), snow (rainfall), wet bulb temperature, dew point, wind direction, gust speed, relative humidity, visibility, precipitation probability, rain probability, snow probability, ice probability, ice amount, ultraviolet intensity, cloud height, etc. collected through an external weather forecast server, etc.

[0060] Additionally, solar information data may include solar azimuth, apparent altitude, solar irradiance, etc.

[0061] Additionally, power plant data may include power plant ID, latitude, longitude, generation, capacity, etc.

[0062] Data preprocessed through the collection unit (100) can be stored in the storage unit (400), and by specifying a specific point in time, the data before that point in time can be used as learning data, and the data after that point in time can be used as verification data.

[0063] The weather data and solar information data collected from the collection unit (100) can be used for cluster classification according to weather events, and the power generation amount of power plant data together with the weather data and solar information data can be used in learning multiple prediction modules.

[0064] Next, the learning unit (200) can generate learning data by mixing Gaussian noise into the collected weather data and solar information data for multiple solar power plants (S120), and the autoencoder (310) can learn the learning data (S130).

[0065] Specifically, the autoencoder (310) includes an encoder that reduces the dimension of learning data to extract latent variables and a decoder that restores the input learning data based on the latent variables. At this time, the learning unit (200) can extract latent variables using the autoencoder (310).

[0066] Latent variables can be information that contains key elements in a compressed map of meteorological and solar information data for multiple solar power plants.

[0067] Next, the clustering unit (320) can classify the extracted multiple latent variables into multiple clusters based on similarity (S140).

[0068] The k-Means clustering technique is widely used for clustering. However, problems can arise in calculating the Euclidean distance depending on whether the input variables contain singular or outlier values. Therefore, the present invention compensates for singularity by extracting latent variables from the autoencoder and using them as input variables for the clustering unit (320).

[0069] More specifically, the clustering unit (320) may include a first process of randomly selecting k center positions based on latent variables and generating k clusters, and calculating a straight-line distance from each latent variable to the randomly selected center position to include the cluster containing the center position with the closest distance, and a second process of setting the position where the sum of the distances to the latent variables within each cluster is minimized as a new center position.

[0070] At this time, steps S130 and S140 can be sequentially repeated until the mutual condition is satisfied. At this time, the mutual condition is to optimize both the cost function in S130 and S140, that is, when the reconstruction error is minimized in S130 and the center position of each cluster does not change in step S140, in which case the classified cluster can be determined as the final classified cluster.

[0071] If the restoration error of the autoencoder (310) is minimal and the center position of each cluster classified by the clustering unit (320) has not changed (S150, Y), the final classified cluster and the center point of each cluster can be set to determine the final classified cluster (S160).

[0072] Meanwhile, if any of the mutual conditions is not satisfied, steps S130 and S140 can be repeatedly performed and learned (S150, N).

[0073] That is, the autoencoder (310) learning and the change of the center point of the clustering unit (320) can improve performance compared to the k-Means technique as they are learned together.

[0074] At this time, the final number of clusters determined may differ from the initial preset number (k). While classifying into k clusters, only clusters excluding those that do not contain values ​​can be classified as final clusters.

[0075] FIG. 6 is a diagram illustrating clusters according to an embodiment of the present invention. The training data according to an embodiment of the present invention is formed in 31 dimensions (high-dimensional input of the encoder), and for example, the number of clusters is set to 10 (k) and classified into clusters 0 to 9, but the clusters classified based on latent variables are optimally classified into 5 clusters: cluster 2, cluster 3, cluster 4, cluster 5, and cluster 6. Meanwhile, FIG. 6 is displayed in 2 dimensions for diagrammatic purposes.

[0076] Next, the selection unit (330) can dynamically select the optimal prediction model for each cluster, which is finally classified in S160, among multiple different prediction models learned for latent variables for each cluster.

[0077] More specifically, the learning unit (200) can learn multiple different prediction models for the latent variables for each cluster finally classified in S160 (S170). At this time, the prediction models are models generally used for prediction, and for example, multiple different prediction models as in FIG. 7 can be applied and stored in the storage unit (400).

[0078] Figure 7 is a priority table for explaining the selection of the optimal prediction model in the selection section of Figure 2. Figure 7 shows evaluation indicators for the learning results obtained by learning multiple different prediction models for the latent variables for each cluster finally classified in S160.

[0079] At this time, the evaluation indices used can be MAE (Mean Absolute Error), MSE (Mean Squared Error), RMSE (Root Mean Squared Error), R2 (Root Squared), RMSLE (Root Mean Squared Log Error), MAPE (Mean Absolute Percentage Error), and TT (Test Time, Sec).

[0080] Here, MSE, MAE, SMSE, and RMSLE indicate better performance as they are closer to 0, and R2 indicates better performance as they are closer to 1.

[0081] The selection unit (330) can dynamically select the optimal prediction model for each cluster, which is finally classified in S160, from among the multiple learned prediction models. More specifically, for each cluster, the multiple learned prediction models are sorted in descending order of the average values ​​of the model performance evaluation indicators (e.g., FIG. 7), and the prediction model with the highest average value or an ensemble model including a preset number of prediction models in descending order of average values ​​can be selected as the optimal prediction model.

[0082] In this way, the prediction model dynamic selector (300) can be learned through the learning stages of S110 to S180.

[0083] Next, when weather data and solar information data for a certain period are input (S200), the optimal prediction model that minimizes the solar power generation prediction error among multiple solar power generation prediction modules learned for each weather event can be selected through the prediction model dynamic selector (300) learned in S100 (S300).

[0084] According to the embodiment of Fig. 7, the catboost model may be selected as the optimal model, or an ensemble model of catboost, et, and rf may be selected as the optimal model depending on the number of user settings, etc.

[0085] Next, power generation can be predicted using an optimal prediction model for the weather data and solar information data for a certain period entered in step S200 (S400).

[0086] Fig. 8 is a graph comparing the predicted values ​​of a dynamically selected solar power generation prediction model according to an embodiment of the present invention. Fig. 8 is a graph comparing the actual predicted power generation for the input period of weather data and solar information data of five actual power plants, the predicted values ​​based on the LGBM, which is generally widely used for prediction, and the predicted values ​​based on the optimal prediction model (e.g., catboost selected in Fig. 7) according to an embodiment of the present invention. The prediction error rate and incentive acquisition rate were calculated as follows.

[0087]

[0088] As a result, the method and system for dynamically selecting a solar power generation prediction model according to an embodiment of the present invention can learn weather data and solar information data for multiple solar power plants using an autoencoder to extract latent variables, and construct an optimal solar power generation prediction model for each cluster based on the latent variables, thereby dynamically selecting and providing an optimal solar power generation prediction model according to a weather event.

[0089] In addition, by minimizing the reconstruction error of the autoencoder and setting the final cluster and each cluster center point that minimizes the sum of the distances to the latent variables within each cluster, clusters can be classified according to weather events (e.g., typhoons, monsoon rains, heavy snow, heat waves, etc.), and by selecting the optimal prediction model for each classified weather event cluster and predicting solar power generation based on weather data and solar information data, the accuracy of power generation prediction can be improved.

[0090] Although the embodiments of the present invention have been described in more detail with reference to the attached drawings, the present invention is not necessarily limited to these embodiments, and various modifications may be implemented without departing from the technical spirit of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to explain it, and the scope of the technical spirit of the present invention is not limited by these embodiments. Therefore, it should be understood that the embodiments described above are illustrative in all aspects and not restrictive. The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within a scope equivalent thereto should be interpreted as being included in the scope of the rights of the present invention.

Claims

1. A step of learning a dynamic selector of a weather event-specific prediction model for solar power generation prediction; and a step of selecting an optimal prediction model that minimizes a solar power generation prediction error from the learned dynamic selector of the prediction model when weather data and solar information data are input; including; The above learning steps are: (a) A step of extracting latent variables by learning meteorological data and solar information data for multiple solar power plants using an autoencoder; (b) a step of classifying the extracted multiple latent variables into multiple clusters according to weather events based on similarity; (c) a step of training multiple different prediction models for latent variables per cluster; and (d) a step of dynamically selecting the optimal prediction model for each cluster among multiple learned prediction models, The step (b) above includes a first process of selecting k central locations based on the latent variables and generating k clusters, calculating the straight-line distance from the central location for each latent variable and including the cluster containing the central location with the closest distance, and a second process of setting the location where the sum of the distances to the latent variables within each cluster is minimized as a new central location. The above steps (a) and (b) are sequentially repeated until a mutual condition is satisfied, and the mutual condition is when the restoration error is minimized in the above step (a) and the center position of each cluster does not change in the above step (b), and the cluster classified by the satisfaction of the mutual condition is determined as the final classification cluster, and a cluster that does not include a latent variable among the k clusters is excluded from the final classification cluster. A method for dynamically selecting a solar power generation prediction model, wherein the above meteorological data comprises at least one of cloud cover, temperature, maximum temperature, minimum temperature, humidity, ground pressure, wind speed, wind direction, precipitation, rainfall, wet bulb temperature, dew point, wind direction, gust speed, relative humidity, visibility, precipitation probability, rain probability, snow probability, ice probability, ice amount, ultraviolet intensity, and cloud height, and the above solar information data comprises at least one of solar azimuth, apparent altitude, and solar radiation.

2. In paragraph 1, The above learning steps are: A step of collecting weather data and solar information data for the plurality of solar power plants; and A method for dynamically selecting a solar power generation prediction model, further comprising a step of generating training data by mixing Gaussian noise into collected data.

3. In paragraph 2, Step (a) above, An encoding step for extracting latent variables by reducing the dimension of the above-generated learning data; and A method for dynamically selecting a solar power generation prediction model, comprising a decoding step for restoring the learning data based on latent variables.

4. In paragraph 1, Step (d) above, for each cluster, A step of sorting the above-mentioned learned multiple prediction models in order of the highest average value of the model performance evaluation indicators; and A method for dynamically selecting a solar power generation prediction model, comprising a step of selecting an ensemble model including a prediction model with the highest average value or a preset number of prediction models in descending order of average values ​​as the optimal prediction model.

5. A learning unit that learns a dynamic selector for a weather event-specific prediction model for solar power generation prediction; and When meteorological data and solar information data are input, a prediction model dynamic selector is included that selects an optimal prediction model that minimizes solar power generation prediction error; The above prediction model dynamic selector is, An autoencoder that extracts latent variables by learning meteorological data and solar information data for multiple solar power plants; A clustering unit that classifies the extracted multiple latent variables into multiple clusters according to weather events based on similarity; and It includes a selection unit that dynamically selects the optimal prediction model for each cluster among multiple different prediction models learned for latent variables for each cluster; The above clustering unit selects k central positions based on the latent variables and generates k clusters, and includes a first process of calculating a straight-line distance from the central position for each latent variable and including the cluster containing the central position with the closest distance, and a second process of setting the position where the sum of the distances to the latent variables within each cluster is minimized as a new central position. The above autoencoder and the clustering unit are sequentially repeated until a mutual condition is satisfied, and the mutual condition is when the restoration error of the autoencoder is minimized during sequential repetition while the center position of each cluster in the clustering unit does not change, and the clustering unit determines a cluster classified by satisfaction of the mutual condition as a final classification cluster, and a cluster that does not include a latent variable among the k clusters is excluded from the final classification cluster. A system for dynamically selecting a solar power generation prediction model, wherein the above meteorological data comprises at least one of cloud cover, temperature, maximum temperature, minimum temperature, humidity, ground pressure, wind speed, wind direction, precipitation, rainfall, wet bulb temperature, dew point, wind direction, gust speed, relative humidity, visibility, precipitation probability, rain probability, snow probability, ice probability, ice amount, ultraviolet intensity, and cloud height, and the above solar information data comprises at least one of solar azimuth, apparent altitude, and solar radiation.

6. In paragraph 5, Further comprising a collection unit for collecting weather data and solar information data for the above plurality of solar power plants, The above autoencoder is a dynamic selection system for a solar power generation prediction model that learns training data generated by mixing Gaussian noise with the collected weather data and solar information data.

7. In paragraph 6, The above autoencoder is, An encoder that reduces the dimensionality of the above-generated learning data to extract latent variables; and A dynamic selection system for a solar power generation prediction model, comprising a decoder that restores the learning data based on the latent variables.

8. In paragraph 5, The above selection department, A dynamic selection system for a solar power generation prediction model, which sorts the above-mentioned learned multiple prediction models in order of the highest average value of the model performance evaluation indicators, and then selects the prediction model with the highest average value or an ensemble model including a preset number of prediction models in order of the highest average value as the optimal prediction model.

Images