System and method for predicting photovoltaic power generation amount
Patent Information
- Application Number
- PCT/KR2025/003003
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-03-06
- Filing Date
- 2025-03-06
- Publication Date
- 2025-10-02
AI Technical Summary
Solar power generation in the Virtual Power Plant market is highly unpredictable and fluctuates significantly, posing challenges for accurate 24-hour power generation predictions, which are crucial for smooth transactions and stability.
A hybrid prediction system combining Light Gradient Boosting Machine (LGBM) with a similarity-based model to predict solar power generation, using past data to correct predictions outside a set confidence interval, incorporating weather, solar, and power plant data.
Enhances prediction accuracy by correcting outliers, reducing errors, and improving the reliability of solar power generation forecasts, thereby supporting stable transactions in the Virtual Power Plant market.
Smart Images
Figure KR2025003003_02102025_PF_FP_ABST
Abstract
Description
Solar power generation prediction system and method
[0001] The present invention relates to a solar power generation prediction system, and more specifically, to a solar power generation prediction system that learns past data based on LGBM (Light Gradient Bossing Model) to predict solar power generation, and corrects the predicted value based on a similarity-based prediction model when the predicted solar power generation falls outside a set confidence interval.
[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] The problem to be solved by the present invention is to provide a solar power generation prediction system and method that learns past data based on LGBM (Light Gradient Bossing Model) to predict solar power generation, and corrects the predicted value based on a similarity-based prediction model when the predicted solar power generation falls outside a set confidence interval.
[0006] 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.
[0007] In order to solve the problem described above, in one embodiment of the present invention, a solar power generation prediction system may include an LGBM prediction unit that predicts a first prediction value for prediction target data based on an LGBM prediction model generated by learning past data for predicting solar power generation, a similarity-based prediction unit that calculates a similarity between the prediction target data and the past data and predicts a second prediction value for the prediction target data based on a preset number of past data in descending order of similarity, and a final prediction unit that outputs the predicted first prediction value as a final power generation prediction value when the predicted first prediction value is included in a set confidence interval, and outputs the second prediction value as a final power generation prediction value when the predicted first prediction value is outside the set confidence interval.
[0008] At this time, the past data and predicted target data may include weather data, solar information data, and power plant data.
[0009] In addition, the similarity-based prediction unit may include a similarity calculation unit that generates an input variable vector including cloud cover, humidity, wind speed, wind direction, precipitation, and temperature for each of the prediction target data and the past data, and calculates a Euclidean distance between the input variable vector of the prediction target data and the input variable vector of each past data.
[0010] In addition, the similarity-based prediction unit may further include a confidence interval setting unit that extracts m (m is a natural number) past data in order of high similarity and sets a confidence interval of the distribution of solar power generation of the extracted past data.
[0011] In addition, the similarity-based prediction unit may further include a second prediction unit that extracts k past data (k is a natural number smaller than m) in order of high similarity among the m past data, and calculates a second prediction value for the prediction target data by weighting the past power generation corresponding to the extracted past data.
[0012] In addition, the similarity-based prediction unit may further include a clustering unit that distinguishes a cluster including the prediction target data through a classification model that classifies the past data into multiple clusters based on weather data.
[0013] At this time, the similarity determination unit can calculate the similarity based on past data corresponding to the same cluster, same time, and same power plant as the prediction target data.
[0014] In addition, a method for predicting solar power generation according to an embodiment of the present invention may include a step of predicting a first prediction value for prediction target data based on an LGBM prediction model generated by learning past data for predicting solar power generation, a step of calculating a similarity between the prediction target data and the past data, a step of outputting the predicted first prediction value as a final power generation prediction value if the predicted first prediction value is included in a set confidence interval, and a step of predicting a second prediction value based on the similarity if the predicted first prediction value is outside the set confidence interval, and outputting the predicted second prediction value as a final power generation prediction value. The past data and the prediction target data may include weather data, solar information data, and power plant data.
[0015] In addition, the step of calculating the similarity may include a step of generating an input variable vector including cloud cover, humidity, wind speed, wind direction, precipitation, and temperature for each of the prediction target data and the past data, and a step of calculating a Euclidean distance between the input variable vector of the prediction target data and the input variable vector of each past data.
[0016] In addition, the above confidence interval can be set based on the distribution of solar power generation of m past data (m is a natural number) in order of high similarity.
[0017] In addition, the second predicted value can be calculated by extracting k past data (k is a natural number smaller than m) with high similarity among the m past data, and calculating a weighted average of past power generation corresponding to the extracted past data.
[0018] In addition, the method further includes a step of distinguishing a cluster including the prediction target data through a classification model that classifies the past data into multiple clusters based on weather data, and the step of calculating the similarity may calculate the similarity based on past data corresponding to the same cluster, same time, and same power plant as the prediction target data.
[0019] Specific details of other embodiments are included in the detailed description and drawings.
[0020] A solar power generation prediction system and method according to an embodiment of the present invention predicts solar power generation by learning past data based on LGBM (Light Gradient Bossing Model), and when the predicted solar power generation falls outside the confidence interval, the predicted value can be corrected based on a similarity-based prediction model.
[0021] In particular, the similarity-based prediction model can further improve prediction accuracy by predicting power generation based on similar past data under the same conditions (same cluster, same time, and same power plant) as new input data (data to be predicted).
[0022] 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.
[0023] FIG. 1 is a block diagram schematically illustrating a configuration of a solar power generation prediction system according to an embodiment of the present invention.
[0024] Fig. 2 is a block diagram showing a schematic configuration of the LGBM prediction unit of Fig. 1.
[0025] Fig. 3 is a block diagram showing a schematic configuration of the similarity-based prediction unit of Fig. 1.
[0026] Figure 4 is an example of a data table according to one embodiment of the present invention.
[0027] Figure 5 is a graph for explaining the setting of a confidence interval according to one embodiment of the present invention.
[0028] FIG. 6 is a graph comparing a first predicted value (LGBM predicted value), a second predicted value (similarity-based predicted value) and actual power generation outside the confidence interval according to one embodiment of the present invention.
[0029] Figure 7 is a flowchart for explaining a method for predicting solar power generation according to one embodiment of the present invention.
[0030] *Explanation of key symbols in the drawing*
[0031] 100: Data collection and preprocessing unit
[0032] 200: Hybrid Prediction Unit
[0033] 210: LGBM Prediction Department
[0034] 211: LGBM Learning Department 212: LGBM Prediction Model
[0035] 220: Similarity-based prediction unit
[0036] 221: Clustering section 222: Similarity determination section
[0037] 223: Confidence interval setting section 224: Second prediction section
[0038] 230: Final Prediction Department
[0039] 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's concepts and should be understood as being in no way limiting to the specifically enumerated embodiments and conditions.
[0040] 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.
[0041] 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.
[0042] 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.
[0043] A solar power generation prediction system according to one embodiment of the present invention can predict solar power generation based on a LightGBM regression model (LGBM).
[0044] LGBM uses a histogram-based algorithm to group continuous feature values into bins, which allows for faster learning and training, reduced memory usage, and higher accuracy of results, making it used in prediction models.
[0045] However, solar power generation varies significantly depending on cloud cover and precipitation, and also varies over time. The LGBM model trains to reduce these residuals, but it falls short of accurately predicting all of these diverse conditions.
[0046] Furthermore, power generation can vary from power plant to power plant, making it difficult to predict and distinguish these numerous conditions using a regression model. To address this issue, a solar power generation prediction system according to one embodiment of the present invention utilizes a hybrid model combining the LGBM prediction model and a similarity-based prediction model to predict solar power generation.
[0047] Hereinafter, various embodiments of the present invention will be described in detail with reference to the attached drawings.
[0048] FIG. 1 is a block diagram schematically illustrating the configuration of a solar power generation prediction system according to an embodiment of the present invention. FIG. 2 is a block diagram schematically illustrating the configuration of the LGBM prediction unit of FIG. 1. FIG. 3 is a block diagram schematically illustrating the configuration of the similarity-based prediction unit of FIG. 1.
[0049] Referring to FIGS. 1 to 3, a solar power generation prediction system (hereinafter referred to as “system”) according to an embodiment of the present invention may include a data collection and preprocessing unit (100), a hybrid prediction unit (200), and the hybrid prediction unit (200) may include an LGBM prediction unit (210), a similarity-based prediction unit (220), and a final prediction unit (230).
[0050] In addition, the LGBM prediction unit (210) may include an LGBM learning unit (211) and an LGBM prediction model (212), and the similarity-based prediction unit (220) may include a clustering unit (221), a similarity judgment unit (222), a confidence interval setting unit (223), and a second prediction unit (224).
[0051] The data collection and preprocessing unit (100) can collect weather data, solar information data, and power plant data, and perform data preprocessing such as scale adjustment and feature extraction. At this time, a data table (e.g., FIG. 4) having weather data, power plant data, and solar information data as columns for each hour from all power plants over the entire period can be created and utilized.
[0052] Here, weather data may include 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 ray intensity, cloud height, etc. collected through an external weather forecast server, etc.
[0053] Additionally, solar information data may include solar azimuth, apparent altitude, solar irradiance, etc.
[0054] Additionally, power plant data may include power plant ID, latitude, longitude, generation, capacity, etc.
[0055] Additionally, solar irradiance was calculated using data such as cloud cover, humidity, and temperature. Columns containing data to be used for forecasting were selected and subjected to maximum-minimum normalization. Furthermore, power generation was divided by capacity and normalized to a value between 0 and 1.
[0056] Data preprocessed through the data collection and preprocessing unit (100) can be used as learning data before a specific point in time by specifying that point in time, and data after that point in time can be used as verification data.
[0057] The LGBM prediction unit (210) can predict a first prediction value, which is a prediction value for the prediction target data, based on the LGBM prediction model (212) generated by learning past data (learning data) for predicting solar power generation.
[0058] At this time, the LGBM prediction model (212) can be constructed by learning learning data through the LGBM learning unit (211). That is, the LGBM learning unit (211) can learn learning data, i.e., previously collected and preprocessed weather data, solar information data, and power plant data, based on the LGBM learning algorithm to generate the LGBM prediction model (212) that predicts solar power generation.
[0059] The similarity-based prediction unit (220) is a similarity-based forecast module (SBFM), which calculates the similarity between the prediction target data and past data, and can predict a second prediction value, which is a power generation prediction value for the prediction target data, based on a preset number of past data in order of high similarity.
[0060] The clustering unit (221) can classify clusters containing prediction target data using a classification model learned from past data. K-means clustering can be performed.
[0061] For example, nine weather data, namely cloud cover, humidity, wind speed, wind direction, rain (precipitation), temperature, dew point, ultraviolet intensity, precipitation probability, zenith angle, apparent altitude, and solar irradiance, which are utilized data from weather forecast data and solar information data, can be normalized to the maximum and minimum, and the elbow point can be searched for in the range of 1 to 15 for the number of clusters to determine the optimal number of clusters (k).
[0062] The clustering unit (221) builds a classification model through learning to classify past data into clusters with an optimal number of clusters, and can distinguish a cluster containing the prediction target data based on the classification model.
[0063] The similarity judgment unit (222) can calculate the similarity based on past data corresponding to the same cluster, same time, and same power plant as the prediction target data.
[0064] Specifically, the similarity judgment unit (222) can determine the similarity by generating an input variable vector including cloud cover, humidity, wind speed, wind direction, precipitation, and temperature for each of the prediction target data and the past data, and calculating the Euclidean distance between the input variable vector of the prediction target data and the input variable vector of each past data. At this time, it can be determined that the closer the Euclidean distance, the higher the similarity.
[0065] The confidence interval setting unit (223) extracts m (where m is a natural number) historical data in order of high similarity and can set a confidence interval for the distribution of solar power generation of the extracted historical data (for example, FIG. 6). In one embodiment of the present invention, m is set to 40 and the confidence interval is set to 95%, but m and the confidence interval can be changed based on the designer's experience.
[0066] The second prediction unit (224) can extract k pieces of past data (k is a natural number smaller than m) with high similarity among m pieces of past data, and calculate a weighted mean of past power generation corresponding to the extracted past data to produce a second prediction value for the prediction target data. At this time, the second prediction unit (224) can produce the second prediction value when the first prediction value is outside the confidence interval.
[0067] The second prediction unit (224) can further increase prediction accuracy by using past data with high similarity among past data under similar conditions (same cluster, same time, and same power plant) to the prediction target data.
[0068] Specifically, the second predicted value ( ) can be calculated as shown in the following mathematical formula 1.
[0069] [Mathematical Formula 1]
[0070]
[0071] Here, the weight , d is the distance , x can be the input variable vector of the prediction target data, i can be the data index, v can be the input variable vector of past data, and g can be the past power generation.
[0072] The final prediction unit (230) outputs the predicted first prediction value as the final power generation prediction value if the first prediction value predicted by the LGBM prediction unit (211) is included in the set confidence interval, and if the predicted first prediction value is outside the set confidence interval, only the first prediction value can be corrected with the second prediction value predicted by the similarity-based prediction unit (220) and output as the final power generation prediction value.
[0073] A hybrid prediction model (200) according to one embodiment of the present invention was trained and verified based on learning data and verification data as shown in [Table 1] below.
[0074] [Table 1]
[0075]
[0076] Specifically, we used meteorological data provided by Openweathermap and accuweather, collecting hourly data from 20 power plants from October 1, 2020, to November 13, 2023. The total number of data points was 296,342. Assuming a clear day with no clouds, we calculated solar information data, such as solar azimuth and apparent elevation, using latitude and longitude. Furthermore, we calculated solar radiation using empirical formulas. Data from October 1, 2023, onward, were used for verification.
[0077] The validation data predicted the power generation of 20 power plants from 2023-10-01 to 2023-11-13. The time required was 33 seconds for LightGBM and 1,376 seconds for SBFM. Using a classification model (kmeans cluster) trained on past data, the validation data were classified by predicting which cluster they belong to. The distances between the validation data and past data in the same cluster were calculated and sorted in ascending order. The power generation was predicted by taking a weighted average of the seven (k) data points with the closest distances. Only six variables were used to calculate the distances: cloud cover, humidity, wind speed, wind direction, rain, and temperature.
[0078] For example, the meteorological variable table for a power plant in the verification data can be displayed as in Fig. 5. Here, 40 (m) past data with similar conditions to the input data were sequentially selected to check the distribution of power generation. Here, similar conditions mean that among data in the same time and same cluster, the distance from the input data is close. In addition, as in Fig. 5, the confidence interval of the distribution was set to 95%, and if the power generation predicted by the LGBM model (212) was outside this interval, the power generation was replaced with the second predicted value of the similarity-based prediction unit (220, SBFM).
[0079] As a result of the verification, among the 19,608 first prediction values of the LGBM verification data, 633 first prediction values were corrected to the second prediction values of SBFM. In Fig. 6, 100 of the 633 are shown as a scatter plot to compare the actual power generation, predicted values, and confidence intervals together. The gray area represents the confidence interval, and cases where the LGBM predicted value (first predicted value) is outside the confidence interval can be compared with the actual power generation, and the first predicted values that are outside the confidence interval are corrected to the second predicted value (SBFM predicted value).
[0080] As shown below, the predicted power generation and actual power generation were compared to determine the error. Since settlement payments are not available when the predicted power generation is less than 10%, only cases exceeding 10% were calculated. The error is calculated by dividing the absolute value of the difference between the actual power generation and the final predicted power generation by the capacity. Table 2 below shows the predicted power generation results.
[0081] [Table 2]
[0082]
[0083] Here, the probability that an incorrect prediction differing by more than 8% from the actual value was outside the confidence interval was approximately 1.98%, and the error rate was reduced and the prediction was more accurate when using the hybrid prediction model (LGBM+SBFM) according to an embodiment of the present invention than when using only the LGBM prediction model to predict the power generation amount, and more profit could be obtained in terms of settlement amount.
[0084] FIG. 7 is a flowchart illustrating a method for predicting solar power generation according to an embodiment of the present invention. The method for predicting solar power generation of FIG. 7 can be performed based on the systems of FIGS. 1 to 3. Accordingly, the method of FIG. 7 can be described with reference to FIGS. 1 to 6.
[0085] First, as shown in Table 1, weather data, solar energy information data, and power plant data are collected and preprocessed to be used for learning and verification.
[0086] Specifically, data preprocessing, such as scaling and feature extraction, can be performed. At this time, a data table (e.g., Figure 4) containing hourly weather data, power plant data, and solar information data as columns for all power plants over the entire period can be created and utilized.
[0087] In addition, the amount of solar radiation was calculated using data such as clouds, humidity, and temperature, and the column containing the data to be used for prediction was selected to perform maximum-minimum normalization, and the amount of power generation was divided by capacity and normalized to a value between 0 and 1 for use.
[0088] Preprocessed data, for example, can be used to predict solar power generation based on the LGBM prediction model (212) and similarity-based prediction model (220) generated by specifying a specific point in time as in Table 1, using data before that point as learning data, and using data thereafter as verification data.
[0089] Based on the LGBM prediction model (212), the first prediction value for the prediction target data can be predicted (S100).
[0090] Next, similarity can be calculated (S200). Similarity can be calculated based on past data corresponding to the same cluster, time, and power plant as the target data. At this time, a classification model trained on past data can be used to classify the cluster containing the target data. K-means clustering can be performed.
[0091] Specifically, for each of the target data and past data, input variable vectors including cloud cover, humidity, wind speed, wind direction, precipitation, and temperature are generated, and the similarity can be determined by calculating the Euclidean distance between the input variable vector of the target data and the input variable vector of each past data. At this time, the closer the Euclidean distance, the higher the similarity can be determined.
[0092] At this time, m past data (m is a natural number) with high similarity can be extracted, and a confidence interval for the distribution of solar power generation of the extracted past data can be set.
[0093] Next, the final power generation prediction value can be determined by determining whether the first prediction value (LGBM prediction value) is included in the confidence interval. Specifically, if the first prediction value is included in the set confidence interval (S300, Y), the predicted first prediction value is output as the final power generation prediction value (S400), and if the predicted first prediction value is outside the set confidence interval (S300, N), only the first prediction values that are outside the confidence interval can be corrected by the similarity-based prediction unit (220) with the second prediction value (SBFM prediction value) predicted and output as the final power generation prediction value (S500).
[0094] Here, the second predicted value can be calculated by extracting k past data (k is a natural number smaller than m) in order of high similarity among m past data having similar conditions (cluster, same time, same power plant) to the prediction target data, and calculating a weighted mean of the past power generation corresponding to the extracted k past data.
[0095] In this way, the solar power generation prediction system according to one embodiment of the present invention can further increase prediction accuracy by predicting solar power generation using a hybrid model of the LGBM prediction model and the similarity-based prediction model.
[0096] 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 claims below, 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. An LGBM prediction unit that predicts a first prediction value for the prediction target data based on an LGBM prediction model generated by learning past data for predicting solar power generation; A similarity-based prediction unit that calculates the similarity between the prediction target data and the past data, and predicts a second prediction value for the prediction target data based on a preset number of past data in order of high similarity; and A final prediction unit is included that outputs the predicted first predicted value as a final power generation predicted value if the predicted first predicted value is included in a set confidence interval, and outputs the second predicted value as a final power generation predicted value if the predicted first predicted value is outside the set confidence interval. The above similarity-based prediction unit is, A clustering unit that distinguishes a cluster that includes the prediction target data through a classification model that classifies the above past data into multiple clusters based on weather data; and It includes a similarity calculation unit that calculates similarity based on past data corresponding to the same cluster, same time, and same power plant as the above prediction target data. A solar power generation prediction system, wherein the above past data and predicted target data include weather data, solar information data, and power plant data.
2. In paragraph 1, The above similarity calculation unit is, A solar power generation prediction system that generates an input variable vector including cloud cover, humidity, wind speed, wind direction, precipitation, and temperature for each of the above prediction target data and the above past data, and calculates the Euclidean distance between the input variable vector of the prediction target data and the input variable vector of each past data.
3. In paragraph 1, The above similarity-based prediction unit is, A solar power generation prediction system further comprising a confidence interval setting unit that extracts m (where m is a natural number) pieces of past data in order of high similarity and sets a confidence interval of the distribution of solar power generation of the extracted past data.
4. In paragraph 3, The above similarity-based prediction unit is, A solar power generation prediction system further comprising a second prediction unit that extracts k pieces of past data (k is a natural number smaller than m) in order of high similarity among the m pieces of past data, and calculates a second prediction value for the prediction target data by weighting the past power generation corresponding to the extracted past data.
5. A step of predicting a first predicted value for the prediction target data based on the LGBM prediction model created by learning past data for predicting solar power generation; A step of distinguishing a cluster including the prediction target data through a classification model that classifies the above past data into multiple clusters based on weather data; A step of calculating the similarity between the above prediction target data and the above past data; A step of outputting the predicted first predicted value as a final power generation predicted value if the predicted first predicted value is included in the set confidence interval; and If the predicted first predicted value is outside the set confidence interval, a step of predicting a second predicted value based on the similarity and outputting the predicted second predicted value as the final power generation predicted value is included. The step of calculating the above similarity is: Similarity is calculated based on past data corresponding to the same cluster, same time, and same power plant as the above prediction target data. A method for predicting solar power generation, wherein the above past data and predicted target data include weather data, solar information data, and power plant data.
6. In paragraph 5, The step of calculating the above similarity is: A step of generating an input variable vector including cloud cover, humidity, wind speed, wind direction, precipitation, and temperature for each of the above prediction target data and the above past data; and A method for predicting solar power generation, comprising: a step of calculating a Euclidean distance between an input variable vector of prediction target data and an input variable vector of each past data.
7. In paragraph 5, The above confidence interval is, A method for predicting solar power generation, which extracts m (where m is a natural number) pieces of past data in order of high similarity and is established based on the distribution of solar power generation of the extracted past data.
8. In paragraph 7, The above second prediction value is, A method for predicting solar power generation, wherein k pieces of past data (k is a natural number smaller than m) with high similarity are extracted from the m pieces of past data, and the past power generation corresponding to the extracted past data is weighted and averaged.