Full-automatic photovoltaic power prediction model training method, system and device and storage medium
By adopting a fully automated pipeline photovoltaic power prediction model training method, the problems of subjectivity, single mechanism and static hyperparameter in photovoltaic power prediction are solved, realizing efficient and accurate photovoltaic power prediction, which is applicable to photovoltaic power plants of various sizes and types.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU SHENGNENG TECH CO LTD
- Filing Date
- 2026-01-21
- Publication Date
- 2026-04-21
AI Technical Summary
Existing photovoltaic power prediction technologies suffer from contradictions between subjectivity and completeness in practical engineering applications, a trade-off between the fitting ability and generalization performance of single-mechanism prediction models, and the lag of static hyperparameter optimization mechanisms in dynamic time-varying environments, resulting in low prediction accuracy and high model maintenance costs.
A fully automated pipeline photovoltaic power prediction model training method is adopted, which automates the entire process of data acquisition, preprocessing, feature extraction, feature selection and model training. It combines multiple machine learning and deep learning algorithms and uses MinMax normalization, maximum information coefficient (MIC) and differential evolution algorithm to optimize model performance, thereby achieving the model's adaptability and generalization ability.
It improves the adaptability, accuracy, and efficiency of photovoltaic power prediction, reduces human intervention, lowers model maintenance costs, and achieves efficient and accurate photovoltaic power prediction.
Smart Images
Figure CN121906418A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photovoltaic power generation technology, and in particular relates to a fully automated photovoltaic power prediction model training method, system, device and storage medium. Background Technology
[0002] Photovoltaic power forecasting is the foundation for decision-making regarding the grid connection of photovoltaic power generation systems and the coordinated control of power generation, grid, load, and energy storage. Its forecasting accuracy directly determines the safety margin of grid dispatching, the trading revenue of the electricity spot market, and the photovoltaic absorption rate. However, existing photovoltaic power forecasting technologies still face many challenges in practical engineering applications, mainly concentrated in the following three aspects: 1. Traditional manual feature engineering suffers from a contradiction between subjectivity and completeness: existing data preprocessing and feature extraction heavily rely on experts' prior experience. On the one hand, subjectively set threshold cleaning strategies are too rigid, making it difficult to distinguish between "power surges caused by cloud cover" and "data acquisition noise," easily misjudging genuine power ramp signals as outliers and removing them, leading to data distortion. On the other hand, feature mining suffers from a lack of dimensionality. Since manual feature extraction is usually limited to statistical indicators (such as mean and variance), it often ignores the periodic fluctuations of meteorological data in the frequency domain and the spatial correlation between multiple stations, resulting in an "island effect" in the information input to the model, failing to capture the deep coupling relationships under complex meteorological conditions.
[0003] 2. Single-mechanism prediction models face a trade-off between "fitting ability and generalization performance": Currently, most mainstream methods employ single machine learning (such as XGBoost) or deep learning (such as LSTM) models, but both paths have inherent flaws. First, shallow machine learning models are insufficient in capturing temporal dependencies. Although they are computationally efficient under stable weather conditions, they cannot effectively remember long-term historical states in scenarios of continuous rain or sudden weather changes, leading to significant lags in prediction results. Second, deep neural network models are prone to overfitting, especially when there is limited sample data from photovoltaic power plants. The models are highly susceptible to fitting specific noise in the training set, resulting in a significant decrease in generalization reasoning ability under unknown weather conditions, and the prediction accuracy in actual operation is far lower than the theoretical test value.
[0004] 3. Static hyperparameter optimization mechanisms lag behind dynamic time-varying environments: The operating environment of a photovoltaic power plant is a typical nonlinear time-varying system (affected by factors such as seasonal changes, component aging, and dust accumulation). However, existing model training often employs "offline one-time training" or static parameter tuning based on grid search. This open-loop parameter setting method suffers from severe "time lag," failing to adjust model weights or hyperparameters in real time based on the latest prediction errors. Furthermore, grid search is extremely inefficient in high-dimensional parameter spaces, often getting stuck in local optima, making it difficult to find a balance between "training efficiency" and "global optimum." This leads to a gradual decline in model performance after long-term operation, making it difficult to meet the high standards of the power grid for ultra-short-term predictions. Summary of the Invention
[0005] To address the aforementioned technical problems and the challenges faced by existing photovoltaic power prediction methods in ultra-short-term forecasting, such as model complexity and cumbersome manual intervention in data processing, this invention provides a fully automated photovoltaic power prediction model training method. This fully automated pipeline-based photovoltaic power prediction model training method automates the entire process from data acquisition to model output, improving the predictive adaptability, accuracy, stability, and efficiency, reducing reliance on manual intervention, and achieving efficient and accurate photovoltaic power prediction. It is applicable to photovoltaic power plants of various sizes and types. Specifically, it includes: Achieve end-to-end automation: From data acquisition, preprocessing, feature extraction, and feature selection to model training and optimization, the entire process is automated, reducing manual operations.
[0006] Improve prediction accuracy: By combining a variety of advanced machine learning and deep learning algorithms, the model performance is optimized to improve the accuracy and stability of predictions.
[0007] Enhance model adaptability: enable the model to adapt to different photovoltaic power generation environments and conditions, and improve the model's generalization ability and adaptability.
[0008] Simplify model maintenance: By implementing systematic processes and automated data processing and model training methods, reduce model maintenance costs and improve application convenience.
[0009] The objective of this invention is specifically achieved through the following technical solutions: In a first aspect, the present invention discloses a fully automated photovoltaic power prediction model training method, comprising: Step 1: Perform outlier processing on the acquired meteorological source data and historical monitoring data from the base station to obtain preprocessed data; Step 2: Extract the time domain features, frequency domain features, spatial features, offset features, and difference features of the preprocessed data to obtain feature-derived data; Step 3: Use the MinMax normalization method to normalize all feature-derived data to obtain standardized data; Step 4: Use the maximum information coefficient (MIC) to perform feature selection on the standardized data, and select the standardized data with the preset quantitative feature that has the highest MIC value as the dataset. Step 5: Use the prediction model as the base model for photovoltaic power prediction; use the five-fold cross-validation method to randomly and evenly divide the dataset into five parts, use one part as the validation set and the other four parts as the training set to train the base model; in each five-fold cross-validation process, use the differential evolution algorithm to calculate the single optimal weighting coefficient of the base model; statistically analyze the weighting results obtained from the five cross-validations to determine the globally optimal weighting coefficient, and obtain the trained photovoltaic power prediction model for outputting photovoltaic power prediction values.
[0010] Secondly, this invention discloses a fully automated photovoltaic power prediction model training system, comprising: The data preprocessing module is used to process outliers in the acquired meteorological source data and historical monitoring data from the base station to obtain preprocessed data. The feature derivation module is used to extract time-domain features, frequency-domain features, spatial features, offset features, and difference features from the preprocessed data to obtain feature-derived data. The data standardization module is used to normalize all feature-derived data using the MinMax normalization method to obtain standardized data. The feature selection module uses the maximum information coefficient (MIC) to select features from standardized data, choosing the standardized data with the highest preset number of features as the dataset. The model training and optimization module is used to use the prediction model as the base model for photovoltaic power prediction. The dataset is randomly and uniformly divided into five parts using the five-fold cross-validation method. One part is used as the validation set and the other four parts are used as the training set to train the base model. In each five-fold cross-validation process, the differential evolution algorithm is used to calculate the single optimal weighting coefficient of the base model. The weighting results obtained from the five cross-validations are statistically analyzed to determine the globally optimal weighting coefficient, resulting in the trained photovoltaic power prediction model, which can be used to output photovoltaic power prediction values.
[0011] Thirdly, the present invention discloses a fully automated photovoltaic power prediction model training device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the fully automated photovoltaic power prediction model training method described in the first aspect.
[0012] Fourthly, the present invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the fully automated photovoltaic power prediction model training method described in the first aspect.
[0013] The beneficial effects of this invention are: The technical solution disclosed in this invention aims to improve the adaptability, prediction efficiency, and accuracy of photovoltaic power prediction models, reduce manual intervention, and achieve personalized photovoltaic power prediction. Specific effects include: 1. Efficient Data Processing: Through automated multi-level outlier processing, a combination of methods is used to handle outliers in the data, including thresholding, sliding windowing, finite difference, and 3-sigma methods. This ensures data quality and reliability, lays a solid foundation for subsequent analysis and feature extraction, improves the efficiency and accuracy of data processing, and reduces manual operations.
[0014] 2. Multidimensional feature extraction: Integrating time domain, frequency domain, dimensionality reduction, offset and difference features, it comprehensively extracts effective information from the data and improves the accuracy of the prediction model.
[0015] 3. Intelligent Feature Selection: Feature selection is performed using the maximum information coefficient (MIC) to ensure that the selected features contribute the most to the prediction results, thereby improving the accuracy and stability of the prediction model.
[0016] 4. Flexible model selection and optimization: Users can choose to use specific models according to their needs. It supports three models: LightGBM, XGBoost and CNN-LSTM, which increases the flexibility and adaptability of model training. Through five-fold cross-validation and differential evolution algorithm, the model is weighted and optimized to improve the model's adaptability and generalization ability.
[0017] 5. Improve prediction accuracy: By combining multiple advanced machine learning and deep learning algorithms and using differential evolution algorithm for weighted optimization, the model performance is optimized, and the accuracy and stability of prediction are improved.
[0018] 6. Enhance model adaptability: By utilizing multiple data sources and feature selection methods, the model can adapt to different photovoltaic power generation environments and conditions, thereby improving the model's generalization ability and adaptability.
[0019] 7. Achieve full-process automation: This invention automates the entire process from data acquisition, outlier handling, feature derivation, data standardization, feature selection to model training and prediction output, reducing manual operations, lowering labor costs and operational difficulty, and improving prediction efficiency and accuracy.
[0020] 8. Simplify model maintenance: Reduce model maintenance costs and improve application convenience through systematic processes and automated data processing and model training methods. Attached Figure Description
[0021] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0022] Figure 1 This is a schematic diagram of a fully automated photovoltaic power prediction model training method provided in an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of the model prediction results provided in an embodiment of the present invention. Detailed Implementation
[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0025] Example 1 like Figure 1 As shown, Embodiment 1 of the present invention provides a fully automated photovoltaic power prediction model training method. This method achieves an end-to-end closed loop from raw data input to high-precision prediction model output through fully automated pipeline processing. It includes: Step 1: Perform outlier processing on the acquired meteorological source data and historical monitoring data from the base station to obtain preprocessed data; Step one employs a serial cleaning strategy to clean multi-level adaptive data, with the system accessing meteorological source data and historical monitoring data from base stations. The meteorological source data covers all-dimensional meteorological elements, including total radiation, direct radiation, diffuse radiation, ambient temperature, relative humidity, wind speed, wind direction, atmospheric pressure, rainfall, and cloud cover, providing a data foundation for capturing complex weather changes. The historical monitoring data from base stations accurately collects the historical power output of the photovoltaic power station, as well as on-site measured total radiation, direct radiation, diffuse radiation, and inverter operating status, in order to thoroughly remove noise and retain effective features.
[0026] Step 2: Extract the time domain features, frequency domain features, spatial features, offset features, and difference features of the preprocessed data to obtain feature-derived data; Step two involved multi-dimensional feature derivation. To uncover deeper patterns in the data, the system automatically performed feature engineering. Time-domain and frequency-domain characteristics: statistical values such as mean, variance, and kurtosis of statistical data; at the same time, the energy spectral density of meteorological data in the frequency domain is extracted through Fast Fourier Transform (FFT) to capture the periodic variation pattern of weather.
[0027] Spatial Feature Dimensionality Reduction: Considering the high correlation between different meteorological factors (such as temperature and irradiance), direct input can lead to multicollinearity. This embodiment uses Principal Component Analysis (PCA) to reduce the dimensionality of high-dimensional meteorological data and extract the spatial features with the highest contribution rate.
[0028] Temporal logic features: Construct offset features and differential features to capture the time lag effect of photovoltaic power (such as the delay effect caused by cloud movement) and rate of change features.
[0029] Step 3: Use the MinMax normalization method to normalize all feature-derived data to obtain standardized data; Step 4: Use the maximum information coefficient (MIC) to perform feature selection on the standardized data, and select the standardized data with the preset quantitative feature that has the highest MIC value as the dataset. In steps three and four, intelligent feature optimization based on MIC (Maximum Information Coefficient) introduces the MIC after data standardization (normalization to the 0-1 range) to address feature redundancy. Unlike the traditional Pearson correlation coefficient, which only measures linear relationships, MIC quantifies the complex nonlinear correlation between feature variables and target power through a grid partitioning method (refer to steps S41-S43). The system automatically calculates and sorts the MIC values of all derived features, selecting the top K features (e.g., the top 20) as input to the final model, thus ensuring sufficient information while reducing computational dimensionality.
[0030] Step 5: Use the prediction model as the base model for photovoltaic power prediction; use the five-fold cross-validation method to randomly and evenly divide the dataset into five parts, use one part as the validation set and the other four parts as the training set to train the base model; in each five-fold cross-validation process, use the differential evolution algorithm to calculate the single optimal weighting coefficient of the base model; statistically analyze the weighting results obtained from the five cross-validations to determine the globally optimal weighting coefficient, and obtain the trained photovoltaic power prediction model for outputting photovoltaic power prediction values.
[0031] Step five involves training a base model based on differential evolution to obtain a trained photovoltaic power prediction model for outputting photovoltaic power prediction values. This is the core optimization step of this invention. A prediction model library containing LightGBM, XGBoost, and CNN-LSTM was constructed as the base model.
[0032] Five-fold cross-validation: The dataset is divided into five parts, which are used alternately as the training set and the validation set to ensure the objectivity of the model evaluation.
[0033] Differential Evolutionary Algorithm: In each training fold, the system does not rely on manually set fixed weights, but instead runs a differential evolutionary algorithm.
[0034] Initialization (refer to step S51): Generate a random population of weight vectors (e.g., [w1, w2, w3], and sum(w)=1).
[0035] Evolutionary iteration (refer to steps S52-S55): Generate new weight combinations through mutation and crossover operations, and test their prediction errors (e.g., RMSE) on the validation set. If the new weights have smaller errors, the new weights are retained through "survival of the fittest".
[0036] After multiple iterations, the algorithm will automatically find the single-step optimal weight for that data fold.
[0037] Global strategy: After training, the system statistically analyzes the optimal weights obtained from five cross-validations, calculates their arithmetic mean, and determines it as the globally optimal weighting coefficient.
[0038] Ultimately, depending on the actual situation, if multiple prediction models are selected, the system uses the global coefficient to fuse the outputs of multiple prediction models as the output of the base model, thereby achieving high-precision prediction of photovoltaic power; if a single prediction model is selected, the system uses the global coefficient to select the output of a single prediction model as the output of the base model, thereby achieving high-precision prediction of photovoltaic power.
[0039] In step one, the methods for handling outliers include: S11, based on the threshold method, the abnormal values of the acquired meteorological source data and the historical monitoring data of the base station are filtered according to the preset interval to obtain the threshold data; Among them, the physical threshold filtering in S11 is as follows: First, based on the installed capacity of the photovoltaic power station and the local meteorological limit records, a reasonable physical range is set to directly filter out obvious out-of-bounds data (such as negative power and irradiance exceeding the theoretical maximum value).
[0040] S12, The sliding window method is used to calculate the difference results of the threshold data based on a fixed window size and sliding step size; S13, Use the difference method to perform difference processing on the difference results and identify the abrupt change points in the difference results; Among them, dead value and jump detection in S12-S13, such as using the sliding window method to calculate the difference sequence of data. If the data does not change for a long time (the difference is 0), it is judged as a "dead value" caused by equipment failure; if the difference result shows a non-physical rapid jump, it is marked as a suspected anomaly.
[0041] S14. Applying the 3-sigma principle, abrupt changes exceeding three standard deviations in the difference results are removed to obtain preprocessed data.
[0042] Among them, the statistical elimination in S14 is as follows: applying the 3-sigma principle to the data after the above processing, calculating the mean and standard deviation, and eliminating outliers that deviate from the mean by more than three times the standard deviation, finally obtaining clean preprocessed data.
[0043] The meteorological data obtained in step one includes: total radiation data, direct radiation data, diffuse radiation data, ambient temperature data, relative humidity data, wind speed data, wind direction data, atmospheric pressure data, rainfall data, and cloud cover data. Historical monitoring data of the base station includes: historical power generation data of photovoltaic power station output, measured total radiation data, measured direct radiation data, measured diffuse radiation data, and inverter operating status data.
[0044] In step two, the methods for extracting time-domain features, frequency-domain features, spatial features, offset features, and difference features from the preprocessed data to obtain feature-derived data include: S21, Obtain the statistical characteristics of the preprocessed data as time-domain features, wherein the statistical characteristics include the mean, variance, maximum value, minimum value, kurtosis and skewness of the preprocessed data; S22, Fourier transform is used to extract the frequency components and power spectrum features in the frequency domain of the preprocessed data as frequency domain features; S23, Principal Component Analysis (PCA) is used to reduce the dimensionality of high-dimensional data in the preprocessed data and extract the dimensionality-reduced features as spatial features. S24, shift the time series data of the preprocessed data forward or backward by a preset time step to create an offset feature; S25, calculate the difference features of the time series data of the preprocessed data, capture the changing trend and fluctuation characteristics of the preprocessed data, and obtain the difference features; S26, the time-domain features, frequency-domain features, spatial features, offset features, and differential features are combined to form feature-derived data.
[0045] In step three, the MinMax normalization method is used to normalize all feature-derived data to obtain standardized data. The MinMax normalization method is used to normalize all feature-derived data to the [0, 1] interval to obtain standardized data.
[0046] In step four, feature selection is performed on the standardized data using the Maximum Information Coefficient (MIC). The standardized data with the highest MIC value among the preset quantitative features is selected as the dataset. S41, calculate the mutual information value between each feature variable and the target variable photovoltaic power under each grid division covering the standardized data, and take the maximum value of the ratio of the mutual information value to the logarithm of the number of grid divisions as the maximum information coefficient MIC of the current feature variable; S42, sort the MICs of all feature variables from largest to smallest and construct the MIC sequence; S43, sequentially extract a preset number of standardized data corresponding to the feature variables in the MIC sequence as a dataset.
[0047] In step five, the prediction model includes one or more of the following: LightGBM prediction model, XGBoost prediction model, and CNN-LSTM prediction model.
[0048] In step five, the method for calculating the single-optimal weighting coefficients of the basic model using the differential evolution algorithm is as follows: S51, Initialize the population: Randomly generate a population of N original individuals, each original individual representing a set of weight vectors of the basic model, and the sum of all elements in the weight vector is 1, where N is the number of original individuals in the population; S52, Mutation Operation: Perform mutation operation on the original individuals in the population, and generate mutation vectors by randomly selecting different original individuals in the population; S53, Crossover operation: Cross the mutation vector with the original individual vector to generate the experimental vector; S54, Selection Operation: Substitute the weight vector corresponding to the experimental vector into the base model and calculate the prediction error on the current validation set; if the prediction error of the experimental vector is less than the prediction error of the original individual, then select the experimental vector to enter the next generation of the population; otherwise, retain the original individual. S55, iterate until the preset number of generations is reached, and output the individual with the highest fitness, i.e. the smallest prediction error, as the single-best weighting coefficient for this cross-validation.
[0049] In step five, the method for determining the globally optimal weighting coefficients is as follows: Record the single-best weighting coefficients obtained after each training iteration in the five-fold cross-validation, and statistically analyze the five best weighting coefficients to obtain five sets of weight vectors; The arithmetic mean of the five weight vectors is calculated to obtain the global average weight vector, which is then used as the global optimal weighting coefficient of the trained photovoltaic power prediction model.
[0050] Example 2 Embodiment 2 of the present invention provides a fully automated photovoltaic power prediction model training system, comprising: The data preprocessing module is used to process outliers in the acquired meteorological source data and historical monitoring data from the base station to obtain preprocessed data. The feature derivation module is used to extract time-domain features, frequency-domain features, spatial features, offset features, and difference features from the preprocessed data to obtain feature-derived data. The data standardization module is used to normalize all feature-derived data using the MinMax normalization method to obtain standardized data. The feature selection module uses the maximum information coefficient (MIC) to select features from standardized data, choosing the standardized data with the highest preset number of features as the dataset. The model training and optimization module is used to use the prediction model as the base model for photovoltaic power prediction. The dataset is randomly and uniformly divided into five parts using the five-fold cross-validation method. One part is used as the validation set and the other four parts are used as the training set to train the base model. In each five-fold cross-validation process, the differential evolution algorithm is used to calculate the single optimal weighting coefficient of the base model. The weighting results obtained from the five cross-validations are statistically analyzed to determine the globally optimal weighting coefficient, resulting in the trained photovoltaic power prediction model, which can be used to output photovoltaic power prediction values.
[0051] Each module works in concert through software instructions to implement the above fully automated process on general-purpose computer hardware (memory and processor).
[0052] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the aforementioned method embodiment one, and will not be repeated here.
[0053] Example 3 Embodiment 3 of the present invention provides a fully automated photovoltaic power prediction model training device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the fully automated photovoltaic power prediction model training method corresponding to Embodiment 1 of the aforementioned method.
[0054] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the fully automated photovoltaic power prediction model training method as described in Embodiment 1 of the preceding method.
[0055] Verification experiment: To verify the effectiveness of the technical solution of the present invention, a verification experiment is provided for illustration: Example 1: Photovoltaic power generation project: In a photovoltaic power generation project, applying the technical solution of this invention, users only need to select meteorological source data for the target area and historical monitoring data of the specified photovoltaic base station, and input necessary configuration information, such as installed capacity. After receiving the input data, the system begins to automatically process the data, hiding all complex intermediate processing steps, and realizing end-to-end model training and prediction.
[0056] In practical applications, the system platform first acquires meteorological source data (including radiation, temperature, humidity, wind speed and direction, air pressure, rainfall, and cloud cover) for the target area and historical monitoring data from designated base stations (including historical power, on-site measured radiation values, and inverter status). It automatically performs data preprocessing, combining thresholding, sliding windowing, differential methods, and 3-sigma methods for outlier handling. Then, it extracts time-domain features, frequency-domain features, and spatial features extracted through principal component analysis (PCA), performing data standardization and feature selection. The platform allows users to choose whether to use LightGBM, XGBoost, and CNN-LSTM prediction models, and automatically performs five-fold cross-validation and differential evolution algorithms to calculate the optimal weighting coefficients. Finally, it outputs the photovoltaic power prediction value through a weighted combination of models.
[0057] like Figure 2 The image shown is a screenshot of the actual operation of this embodiment at a photovoltaic power station from June 14 to June 17, 2024.
[0058] Combination Figure 2 Through detailed data analysis, this invention solves the following problems in the prior art and achieves significant results: 1. The model addresses the issue of weak generalization ability under complex operating conditions: As shown in the figure, this period encompasses various meteorological scenarios, including sunny, cloudy, and abrupt weather changes (such as the typical fluctuating weather observed on June 16th). Faced with complex conditions where meteorological data fluctuates drastically or where only predicted irradiance is used as input, the ensemble model proposed in this invention does not exhibit overfitting or significant prediction distortion, demonstrating extremely strong generalization ability. Quantitative indicators show that in fully automatic operation mode, the prediction accuracy (C) reaches 90.37%, the pass rate (Q) reaches 95.31%, the correlation coefficient (r) is 0.87, and the root mean square error (RMSE) is controlled at 9.63%. This fully demonstrates that the model can adapt to different meteorological conditions and maintain a high level of prediction accuracy during long-term continuous operation without human intervention.
[0059] 2. The effectiveness of the differential evolution algorithm for optimization was verified: The mean bias (ME) in the figure is only -0.04, and the mean absolute error (MAE) is 0.05. Such low bias values indicate that the globally optimal weighting coefficients obtained by the differential evolution algorithm successfully found the best balance among the three basic models: LightGBM, XGBoost, and CNN-LSTM. The algorithm effectively utilizes the advantages of each model, eliminates the systematic bias that may exist in a single model, and verifies the scientific validity and effectiveness of the weighted optimization strategy proposed in this invention in a statistical sense.
[0060] 3. Achieved end-to-end automation and low-cost operation and maintenance: This verification covered the entire cycle from stable weather to complex and fluctuating weather. When facing different weather pattern transitions, the system automatically completed the entire process from data cleaning and feature extraction to model inference without manual parameter readjustment or intervention in feature engineering. This demonstrates that the invention has strong environmental adaptability, can significantly reduce the workload of operation and maintenance personnel, and achieve low-cost, automated operation and maintenance of photovoltaic power prediction.
[0061] In this project, model training was fully automated, and the prediction accuracy for ultra-short-term, short-term, and medium-term predictions all met national standards. The entire process required no manual intervention, greatly improving work efficiency and algorithm stability.
[0062] The beneficial effects of this verification experiment are as follows: Fully automated processing flow: This invention automates the entire process from data acquisition, outlier handling, feature derivation, data standardization, feature selection to model training and prediction output, greatly reducing manual intervention and improving prediction efficiency and accuracy.
[0063] Multi-level outlier handling: A combination of methods is used to handle outliers in the data, including thresholding, sliding windowing, differencing, and 3-sigma methods, to ensure data quality and reliability and lay a solid foundation for subsequent analysis.
[0064] Multidimensional feature extraction: By integrating time domain, frequency domain, dimensionality reduction, offset, and difference features, effective information in the data is extracted comprehensively, improving the accuracy of the prediction model.
[0065] Intelligent feature selection: Features are selected using the maximum information coefficient (MIC) to ensure that the selected features contribute the most to the prediction results, thereby improving the accuracy and stability of the prediction model.
[0066] Flexible model selection and optimization: Supports three models: LightGBM, XGBoost, and CNN-LSTM. Through five-fold cross-validation and differential evolution algorithm, the model is weighted and optimized to improve its adaptability and generalization ability.
[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A fully automated photovoltaic power prediction model training method, characterized in that, include: Step 1: Perform outlier processing on the acquired meteorological source data and historical monitoring data from the base station to obtain preprocessed data; Step 2: Extract the time domain features, frequency domain features, spatial features, offset features, and difference features of the preprocessed data to obtain feature-derived data; Step 3: Use the MinMax normalization method to normalize all feature-derived data to obtain standardized data; Step 4: Use the maximum information coefficient (MIC) to perform feature selection on the standardized data, and select the standardized data with the preset quantitative feature that has the highest MIC value as the dataset. Step 5: Use the prediction model as the base model for photovoltaic power prediction; use the five-fold cross-validation method to randomly and evenly divide the dataset into five parts, use one part as the validation set and the other four parts as the training set to train the base model; in each five-fold cross-validation process, use the differential evolution algorithm to calculate the single optimal weighting coefficient of the base model; statistically analyze the weighting results obtained from the five cross-validations to determine the globally optimal weighting coefficient, and obtain the trained photovoltaic power prediction model for outputting photovoltaic power prediction values.
2. The fully automated photovoltaic power prediction model training method as described in claim 1, characterized in that, In step one, the methods for handling outliers include: S11, based on the threshold method, the abnormal values of the acquired meteorological source data and the historical monitoring data of the base station are filtered according to the preset interval to obtain the threshold data; S12, The sliding window method is used to calculate the difference results of the threshold data based on a fixed window size and sliding step size; S13, Use the difference method to perform difference processing on the difference results and identify the abrupt change points in the difference results; S14. Applying the 3-sigma principle, abrupt changes exceeding three standard deviations in the difference results are removed to obtain preprocessed data.
3. The fully automated photovoltaic power prediction model training method as described in claim 1 or 2, characterized in that, The meteorological data obtained in step one includes: total radiation data, direct radiation data, diffuse radiation data, ambient temperature data, relative humidity data, wind speed data, wind direction data, atmospheric pressure data, rainfall data, and cloud cover data. Historical monitoring data of the base station includes: historical power generation data of photovoltaic power station output, measured total radiation data, measured direct radiation data, measured diffuse radiation data, and inverter operating status data.
4. The fully automated photovoltaic power prediction model training method as described in claim 3, characterized in that, In step two, the methods for extracting time-domain features, frequency-domain features, spatial features, offset features, and difference features from the preprocessed data to obtain feature-derived data include: S21, Obtain the statistical characteristics of the preprocessed data as time-domain features, wherein the statistical characteristics include the mean, variance, maximum value, minimum value, kurtosis and skewness of the preprocessed data; S22, Fourier transform is used to extract the frequency components and power spectrum features in the frequency domain of the preprocessed data as frequency domain features; S23, Principal Component Analysis (PCA) is used to reduce the dimensionality of high-dimensional data in the preprocessed data and extract the dimensionality-reduced features as spatial features. S24, shift the time series data of the preprocessed data forward or backward by a preset time step to create an offset feature; S25, calculate the difference features of the time series data of the preprocessed data, capture the changing trend and fluctuation characteristics of the preprocessed data, and obtain the difference features; S26, the time-domain features, frequency-domain features, spatial features, offset features, and differential features are combined to form feature-derived data.
5. The fully automated photovoltaic power prediction model training method as described in claim 4, characterized in that, In step three, the MinMax normalization method is used to normalize all feature-derived data to obtain standardized data. The MinMax normalization method is used to normalize all feature-derived data to the [0, 1] interval to obtain standardized data.
6. The fully automated photovoltaic power prediction model training method as described in claim 5, characterized in that, In step four, feature selection is performed on the standardized data using the Maximum Information Coefficient (MIC). The standardized data with the highest MIC value among the preset quantitative features is selected as the dataset. S41, calculate the mutual information value between each feature variable and the target variable photovoltaic power under each grid division covering the standardized data, and take the maximum value of the ratio of the mutual information value to the logarithm of the number of grid divisions as the maximum information coefficient MIC of the current feature variable; S42, sort the MICs of all feature variables from largest to smallest and construct the MIC sequence; S43, sequentially extract a preset number of standardized data corresponding to the feature variables in the MIC sequence as a dataset.
7. The fully automated photovoltaic power prediction model training method as described in claim 6, characterized in that, In step five, the prediction model includes one or more of the following: LightGBM prediction model, XGBoost prediction model, and CNN-LSTM prediction model.
8. The fully automated photovoltaic power prediction model training method as described in claim 7, characterized in that, In step five, the method for calculating the single-optimal weighting coefficients of the basic model using the differential evolution algorithm is as follows: S51, Initialize the population: Randomly generate a population of N original individuals, each original individual representing a set of weight vectors of the basic model, and the sum of all elements in the weight vector is 1, where N is the number of original individuals in the population; S52, Mutation Operation: Perform mutation operation on the original individuals in the population, and generate mutation vectors by randomly selecting different original individuals in the population; S53, Crossover operation: Cross the mutation vector with the original individual vector to generate the experimental vector; S54, Selection Operation: Substitute the weight vector corresponding to the experimental vector into the base model and calculate the prediction error on the current validation set; if the prediction error of the experimental vector is less than the prediction error of the original individual, then select the experimental vector to enter the next generation of the population; otherwise, retain the original individual. S55, iterate until the preset number of generations is reached, and output the individual with the highest fitness, i.e. the smallest prediction error, as the single-best weighting coefficient for this cross-validation.
9. The fully automated photovoltaic power prediction model training method as described in claim 8, characterized in that, In step five, the method for determining the globally optimal weighting coefficients is as follows: Record the single-best weighting coefficients obtained after each training iteration in the five-fold cross-validation, and statistically analyze the five best weighting coefficients to obtain five sets of weight vectors; The arithmetic mean of the five weight vectors is calculated to obtain the global average weight vector, which is then used as the global optimal weighting coefficient of the trained photovoltaic power prediction model.
10. A fully automated photovoltaic power prediction model training system, characterized in that, include: The data preprocessing module is used to process outliers in the acquired meteorological source data and historical monitoring data from the base station to obtain preprocessed data. The feature derivation module is used to extract time-domain features, frequency-domain features, spatial features, offset features, and difference features from the preprocessed data to obtain feature-derived data. The data standardization module is used to normalize all feature-derived data using the MinMax normalization method to obtain standardized data. The feature selection module uses the maximum information coefficient (MIC) to select features from standardized data, choosing the standardized data with the highest preset number of features as the dataset. The model training and optimization module is used to use the prediction model as the base model for photovoltaic power prediction. The dataset is randomly and uniformly divided into five parts using the five-fold cross-validation method. One part is used as the validation set and the other four parts are used as the training set to train the base model. In each five-fold cross-validation process, the differential evolution algorithm is used to calculate the single optimal weighting coefficient of the base model. The weighting results obtained from the five cross-validations are statistically analyzed to determine the globally optimal weighting coefficient, resulting in the trained photovoltaic power prediction model, which can be used to output photovoltaic power prediction values.
11. A fully automated photovoltaic power prediction model training device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a fully automated photovoltaic power prediction model training method as described in any one of claims 1 to 9.
12. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of an automated photovoltaic power prediction model training method as described in any one of claims 1 to 9.