Artificial Intelligence-Based Wind and Solar Power Prediction System and Method

Abstract

This invention relates to the field of wind and solar power prediction technology, and discloses a wind and solar power prediction system and method based on artificial intelligence. The system includes: a data acquisition and processing module, a feature extraction module, an intelligent prediction module, a dynamic correction module, a blockchain storage module, and a human-computer interaction module. Before training the wind and solar power prediction model, the original data is cleaned and filled by using an abnormal data detection algorithm based on the 3σ criterion and box plot analysis, and an improved KNN missing data imputation algorithm. This can identify and correct noise and errors in the data, ensure the quality of the generated standardized wind and solar power training sample set, reduce prediction bias, and lay a reliable data foundation for generating high-precision prediction results.

Description

Technical Field

[0001] This invention relates to the field of wind and solar power prediction technology, specifically to a wind and solar power prediction system and method based on artificial intelligence. Background Technology

[0002] With the widespread application of renewable energy and the transformation of the energy structure, wind power and solar power, as important forms of renewable energy, have seen continuous growth in installed capacity globally. However, wind and solar energy are characterized by intermittency and volatility, resulting in unstable power output and posing significant challenges to power system dispatch, operation, and planning. Therefore, accurate wind and solar power forecasting is crucial for improving the reliability and stability of the power system and optimizing energy resource allocation.

[0003] Currently, due to the complex meteorological and operational data involved in the wind and solar power prediction process, outliers and missing values ​​exist in the multi-source data collected during high-precision prediction. If these outliers and missing values ​​are not identified and cleaned, noise will be introduced, resulting in large training biases in the prediction model and making it impossible to guarantee the accuracy of the prediction results.

[0004] Therefore, an artificial intelligence-based wind and solar power prediction system and method are proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based wind and solar power prediction system and method, which solves the problem mentioned in the background technology that the prediction model has large training bias and cannot guarantee the accuracy of the prediction results.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a wind and solar power prediction system and method based on artificial intelligence, the system comprising:

[0007] The data acquisition and processing module acquires wind and solar power operation data and meteorological environment data through the multi-dimensional data acquisition unit, and outputs a standardized wind and solar power training sample set using the spatiotemporal alignment processing unit.

[0008] The feature extraction module receives the standardized wind and solar power training sample set, extracts the wind and solar power influencing factor features through an improved adaptive feature weighting algorithm, and outputs a core feature vector set of wind and solar power.

[0009] The intelligent prediction module receives the core feature vector set of wind and solar power, trains and optimizes it by constructing a deep learning model with a fusion neural network architecture including a spatiotemporal convolution module, an attention mechanism module and a long short-term memory network module, and outputs the wind and solar power prediction result.

[0010] The dynamic correction module receives the wind and solar power prediction results, corrects the results through a prediction error feedback mechanism and a sliding time window, and outputs high-precision wind and solar power prediction data.

[0011] The blockchain evidence storage module receives the high-precision wind and solar power prediction data, generates digital fingerprints through a hash algorithm, and uses smart contracts for distributed storage, outputting an immutable prediction audit trajectory.

[0012] The human-computer interaction module receives the high-precision wind and solar power prediction data and audit trajectory, generates a multi-dimensional display interface through the visualization unit, and outputs a predictive analysis report that can be operated by the user.

[0013] Preferably, the method includes the following steps:

[0014] S1. Collect multi-dimensional operational data and meteorological environment data of wind and solar power fields to generate raw wind and solar power time series datasets and high-dimensional meteorological feature datasets.

[0015] S2. Perform data cleaning and spatiotemporal alignment on the original wind and solar power time series dataset and the high-dimensional meteorological feature dataset to generate a standardized wind and solar power training sample set.

[0016] S3. Based on the improved adaptive feature weighting algorithm, wind and solar power influencing factor features are extracted from the standardized wind and solar power training sample set to generate a core feature vector set of wind and solar power.

[0017] S4. Construct a deep learning prediction model based on the fusion of attention mechanism and spatiotemporal convolutional neural network, and train the deep learning prediction model using the core feature vector set of wind and solar power to generate an initial model for wind and solar power prediction.

[0018] S5. The initial wind and solar power prediction model is optimized for hyperparameters and adaptively adjusted for structure using a multi-objective particle swarm optimization algorithm to generate an optimized wind and solar power prediction model.

[0019] S6. Real-time collection of wind and solar power operation status data and short-term weather forecast data, input into the optimized wind and solar power prediction model to perform wind and solar power prediction calculation, and generate short-term wind and solar power prediction results.

[0020] S7. Based on the prediction error feedback mechanism and sliding time window, the short-term wind and solar power prediction results are dynamically corrected to generate high-precision wind and solar power prediction output data.

[0021] S8. Utilize blockchain technology to perform trusted notarization and distributed storage of the high-precision wind and solar power prediction output data, and generate an unalterable wind and solar power prediction audit trajectory.

[0022] Preferably, the acquisition of multi-dimensional operational data and meteorological environment data of wind and solar power plants in step S1 includes the following steps:

[0023] S11. Collect the operating parameters of wind turbine generator set speed, pitch angle, and yaw angle, and the operating parameters of photovoltaic array string current, voltage, and backplane temperature through the wind and solar power SCADA system to generate the original wind and solar power time series dataset.

[0024] S12. Collect meteorological parameters such as wind speed, wind direction, temperature, humidity, atmospheric pressure, and irradiance within the station area through meteorological monitoring stations, and obtain meteorological characteristic data such as cloud cover, aerosol concentration, and atmospheric transmittance within the region through numerical weather prediction systems to generate a high-dimensional meteorological characteristic dataset.

[0025] S13. Perform timestamp alignment and spatial interpolation on the original wind and solar power time series dataset and the high-dimensional meteorological feature dataset to generate spatiotemporally consistent initial training data for wind and solar power.

[0026] Preferably, the data cleaning and spatiotemporal alignment process in step S2 includes the following steps:

[0027] S21. An anomaly detection algorithm based on the 3σ criterion and box plot analysis is used to identify and process outliers in the initial training data of spatiotemporal consistent wind and solar power, and preprocessed wind and solar power data is generated.

[0028] S22. Apply the improved KNN missing data imputation algorithm to impute the missing values ​​in the preprocessed wind and solar power data to generate a continuous and complete wind and solar power dataset.

[0029] S23. The continuous and complete wind and solar power dataset is standardized using the Z-score standardization algorithm to generate a standardized wind and solar power training sample set.

[0030] Preferably, the feature extraction of wind and solar power influencing factors in step S3 includes the following steps:

[0031] S31. Construct a feature importance evaluation model based on a combination of random forest and XGBoost, and calculate the contribution score of each feature variable in the standardized wind and solar power training sample set.

[0032] S32. An improved adaptive feature weighting algorithm is used to dynamically threshold the contribution score, filter out the influence factor features with a contribution score higher than the threshold, and generate a primary feature subset.

[0033] The calculation formula for the improved adaptive feature weighting algorithm is as follows:

[0034] ;

[0035] in This indicates the threshold for dynamic feature selection. This represents the mean of the contribution scores of all features. This represents the standard deviation of the contribution scores of all features. The adjustment coefficient has a range of [0.5, 2.0].

[0036] S33. Apply principal component analysis and linear discriminant analysis algorithms to perform feature dimensionality reduction and redundancy elimination on the primary feature subset to generate a core feature vector set of wind and solar power.

[0037] Preferably, the construction of a deep learning prediction model based on the fusion of attention mechanism and spatiotemporal convolutional neural network in step S4 includes the following steps:

[0038] S41. Construct a fusion neural network architecture that includes a spatiotemporal convolution module, an attention mechanism module, and a long short-term memory network module;

[0039] S42. The parameter initialization algorithm is used to initialize the parameters of the fusion neural network architecture to generate an untrained prediction model. The parameter initialization algorithm adopts the Xavier initialization method.

[0040] S43. The untrained prediction model is trained by forward and backward propagation using the core feature vector set of wind and solar power. The model parameters are optimized through training to generate an initial model for wind and solar power prediction.

[0041] Preferably, the hyperparameter optimization and adaptive structure adjustment using a multi-objective particle swarm optimization algorithm in step S5 includes the following steps:

[0042] S51. Define a multi-objective optimization function, and simultaneously consider three optimization objectives: prediction accuracy, computational efficiency, and model complexity, to establish a mathematical model for the multi-objective optimization problem of wind and solar power prediction.

[0043] S52. An improved multi-objective particle swarm optimization algorithm is used to search for Pareto optimal solutions for the hyperparameters of the initial model for wind and solar power prediction, including key hyperparameters such as learning rate, number of hidden layer nodes, convolutional kernel size, and number of attention heads.

[0044] The position update formula for the improved multi-objective particle swarm optimization algorithm is:

[0045] ;

[0046] ;

[0047] in Indicates the first The particle in the first Position at the next iteration Indicates the first The particle in the first Speed ​​at the next iteration For inertial weights, and As a learning factor, and A random number in the range [0,1]. Indicates the first The optimal position of each individual particle. Indicates the globally optimal position;

[0048] S53. The model structure is adaptively adjusted through neural network architecture search technology to generate a lightweight network structure that combines depthwise separable convolutions and gated recurrent units, and finally an optimized wind and solar power prediction model is generated.

[0049] Preferably, the real-time acquisition of wind and solar power prediction calculations based on wind and solar farm operation status data and short-term weather forecast data in step S6 includes the following steps:

[0050] S61. Real-time acquisition of SCADA system operation data of wind and solar power plants and measured data of meteorological monitoring stations, and acquisition of high-precision meteorological forecast data for the next 72 hours through numerical weather prediction interface;

[0051] S62. Perform real-time data cleaning and feature engineering on the operational data and weather forecast data to generate a real-time prediction input feature vector;

[0052] S63. Input the real-time prediction input feature vector into the optimized wind and solar power prediction model, perform forward inference calculation, and generate short-term wind and solar power prediction results for the next 72 hours, including power point prediction values ​​and probability prediction intervals.

[0053] Preferably, the dynamic correction based on the prediction error feedback mechanism and sliding time window in step S7 includes the following steps:

[0054] S71. Establish a prediction error correction model based on the combination of Kalman filtering and exponential smoothing to eliminate systematic errors in the short-term wind and solar power prediction results;

[0055] S72. A sliding time window algorithm is used to perform statistical analysis and pattern recognition on recent prediction errors, and the prediction model parameters are dynamically adjusted.

[0056] S73. Apply confidence interval calibration technology to calibrate the probability prediction interval and generate high-precision wind and solar power prediction output data that meets the preset confidence level. The confidence interval calibration technology adopts the quantile regression calibration method.

[0057] Preferably, the trusted evidence storage and distributed storage using blockchain technology in step S8 includes the following steps:

[0058] S81. Use a hash algorithm to extract a digital digest from the high-precision wind and solar power prediction output data to generate a unique data fingerprint;

[0059] S83. Establish a blockchain-based prediction data sharing mechanism that allows authorized institutions to verify the credibility and completeness of prediction results through distributed ledger technology.

[0060] Beneficial effects

[0061] Compared with existing technologies, this invention provides an artificial intelligence-based wind and solar power prediction system and method, which has the following beneficial effects:

[0062] 1. In this invention, before training the wind and solar power prediction model, the original data is cleaned and filled by using an anomaly detection algorithm based on the 3σ criterion and box plot analysis, and an improved KNN missing data filling algorithm. This can identify and correct noise and errors in the data, ensure the quality of the generated standardized wind and solar power training sample set, reduce prediction bias, and lay a reliable data foundation for generating high-precision prediction results.

[0063] 2. In this invention, during the feature engineering process, by introducing an improved adaptive feature weighting algorithm, the core feature vector set of wind and solar power that is strongly correlated with the output of wind and solar power can be automatically selected, redundant information can be eliminated, thereby reducing the complexity of the model, enhancing the generalization ability of the model under different data distributions, and improving the stability and reliability of the prediction results.

[0064] 3. In this invention, during the prediction result output stage, dynamic correction is performed through a prediction error feedback mechanism and a sliding time window, and confidence interval calibration technology is applied. This not only provides accurate power point prediction values, but also generates reliable probability prediction intervals, quantifies prediction uncertainty, provides more comprehensive decision-making information for power grid dispatch and energy management, and improves the practicality and risk response capability of the prediction system. Attached Figure Description

[0065] Figure 1 This is a flowchart of the wind and solar power prediction method based on artificial intelligence according to the present invention;

[0066] Figure 2 This is a schematic diagram of the wind and solar power prediction system based on artificial intelligence according to the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0068] For specific implementation examples, please refer to: Figure 1-2 An artificial intelligence-based wind and solar power prediction system and method, the system including:

[0069] The data acquisition and processing module acquires wind and solar power operation data and meteorological environment data through the multi-dimensional data acquisition unit, and outputs a standardized wind and solar power training sample set using the spatiotemporal alignment processing unit.

[0070] The feature extraction module receives a standardized wind and solar power training sample set, extracts the features of wind and solar power influencing factors through an improved adaptive feature weighting algorithm, and outputs a set of core feature vectors of wind and solar power.

[0071] The intelligent prediction module receives the core feature vector set of wind and solar power, trains and optimizes it by constructing a deep learning model with a fusion neural network architecture that includes a spatiotemporal convolution module, an attention mechanism module and a long short-term memory network module, and outputs the wind and solar power prediction results.

[0072] The dynamic correction module receives wind and solar power prediction results, corrects the results through a prediction error feedback mechanism and a sliding time window, and outputs high-precision wind and solar power prediction data.

[0073] The blockchain evidence storage module receives high-precision wind and solar power prediction data, generates digital fingerprints through a hash algorithm, and uses smart contracts for distributed storage, outputting an immutable prediction audit trajectory.

[0074] The human-computer interaction module receives high-precision wind and solar power prediction data and audit trajectory, generates a multi-dimensional display interface through the visualization unit, and outputs a predictive analysis report that can be operated by the user.

[0075] The method includes the following steps:

[0076] S1. Collect multi-dimensional operational data and meteorological environment data of wind and solar power fields to generate raw wind and solar power time series datasets and high-dimensional meteorological feature datasets.

[0077] S2. Perform data cleaning and spatiotemporal alignment on the original wind and solar power time series dataset and the high-dimensional meteorological feature dataset to generate a standardized wind and solar power training sample set.

[0078] S3. Based on the improved adaptive feature weighting algorithm, wind and solar power influencing factor features are extracted from the standardized wind and solar power training sample set to generate a core feature vector set of wind and solar power.

[0079] S4. Construct a deep learning prediction model based on the fusion of attention mechanism and spatiotemporal convolutional neural network, and train the deep learning prediction model using the core feature vector set of wind and solar power to generate an initial model for wind and solar power prediction.

[0080] S5. The initial model for wind and solar power prediction is optimized for hyperparameters and adaptively adjusted for structure using a multi-objective particle swarm optimization algorithm to generate an optimized wind and solar power prediction model.

[0081] S6. Real-time acquisition of wind and solar power operation status data and short-term weather forecast data, input into the optimized wind and solar power prediction model to perform wind and solar power prediction calculations, and generate short-term wind and solar power prediction results.

[0082] S7. Based on the prediction error feedback mechanism and sliding time window, the short-term wind and solar power prediction results are dynamically corrected to generate high-precision wind and solar power prediction output data.

[0083] S8. Utilize blockchain technology to reliably store and distribute high-precision wind and solar power prediction output data, generating an unalterable wind and solar power prediction audit trajectory.

[0084] The collection of multi-dimensional operational data and meteorological environmental data of wind and solar power plants in S1 includes the following steps:

[0085] S11. Collect the operating parameters of wind turbine generator set speed, pitch angle, and yaw angle, and the operating parameters of photovoltaic array string current, voltage, and backplane temperature through the wind and solar power SCADA system to generate the original wind and solar power time series dataset.

[0086] S12. Collect meteorological parameters such as wind speed, wind direction, temperature, humidity, atmospheric pressure, and irradiance within the station area through meteorological monitoring stations, and obtain meteorological characteristic data such as cloud cover, aerosol concentration, and atmospheric transmittance within the region through numerical weather prediction systems to generate a high-dimensional meteorological characteristic dataset.

[0087] S13. Perform timestamp alignment and spatial interpolation on the original wind and solar power time series dataset and the high-dimensional meteorological feature dataset to generate spatiotemporally consistent initial training data for wind and solar power.

[0088] Data cleaning and spatiotemporal alignment in S2 includes the following steps:

[0089] S21. An outlier detection algorithm based on the 3σ criterion and box plot analysis is used to identify and process outliers in the initial training data of spatiotemporal consistent wind and solar power, generating preprocessed wind and solar power data. The specific operations include the following:

[0090] S211. Calculate the mean of each feature variable in the initial training data for spatiotemporally consistent wind and solar power. and standard deviation ;

[0091] S212, Exceeding the value The data points within the range were initially identified as extreme outliers;

[0092] S213. Calculate the first quartile of each characteristic variable. Third and quartiles Interquartile range:

[0093] ;

[0094] S214, Exceeding the value Data points within a certain range were identified as potential outliers;

[0095] S215. Take the union of extreme outliers and potential outliers, and smooth them using a sliding window mean algorithm to generate preprocessed wind and solar power data. The smoothing process using the sliding window mean algorithm specifically includes:

[0096] S2151, For each outlier data point in the outlier sequence Taking it as the center, take forward and backward... 10 normal data points constitute a length of 1 Sliding window, Indicates the total window length;

[0097] S2152. Calculate the arithmetic mean of all normal data points within this window. Using the arithmetic mean Replace abnormal data points The value;

[0098] Its calculation formula can be expressed as:

[0099] ;

[0100] in For the location of data points The value after smoothing. This represents the actual number of normal data points within the window. Position in the window Data point values, Indicates the length of a single-sided window. Represents the time index in a time series. Indicates a circular index;

[0101] S22. Apply the improved KNN missing data imputation algorithm to impute missing values ​​in the preprocessed wind and solar power data to generate a continuous and complete wind and solar power dataset. The improved KNN missing data imputation algorithm specifically includes:

[0102] S221. Using the sample with missing values ​​as the center, calculate the Euclidean distance between it and other complete samples in the dataset.

[0103] S222. Select the k nearest neighbors of the complete sample based on their distance.

[0104] S223. Calculate the weighted average of the k nearest neighbor samples on the missing feature dimension, where the weights are the normalized values ​​of the inverse distances to the corresponding samples.

[0105] S224. Fill the missing positions with the weighted average value to generate a continuous and complete wind and solar power dataset;

[0106] S23. Standardize the continuous and complete wind and solar power dataset using the Z-score standardization algorithm to generate a standardized wind and solar power training sample set. This includes the following steps:

[0107] S231. Calculate the mean and standard deviation for each feature dimension: For each feature dimension in a continuous and complete wind and solar power dataset... Calculate their mean values ​​respectively. and standard deviation ;

[0108] S232. Standardize using the Z-score formula: For each sample in the dataset, the first... 3D eigenvalues The following formula is used for calculation:

[0109] ;

[0110] in For the first The sample at the th Standardized values ​​on each feature dimension For the index of the feature dimension, For the index of the sample, For the first The mean of each feature dimension, For the first Standard deviation of each feature dimension;

[0111] S233. Generate a standardized dataset: Generate all standardized feature values... Reassemble to generate the final standardized wind and solar power training sample set. .

[0112] The feature extraction of wind and solar power influencing factors in S3 includes the following steps:

[0113] S31. Construct a feature importance evaluation model based on a combination of random forest and XGBoost, and calculate the contribution score of each feature variable in the standardized wind and solar power training sample set.

[0114] S32. An improved adaptive feature weighting algorithm is used to dynamically threshold the contribution score, filter out the influence factor features with a contribution score higher than the threshold, and generate a primary feature subset.

[0115] The calculation formula for the improved adaptive feature weighting algorithm is as follows:

[0116] ;

[0117] in This indicates the threshold for dynamic feature selection. This represents the mean of the contribution scores of all features. This represents the standard deviation of the contribution scores of all features. The adjustment coefficient has a range of [0.5, 2.0].

[0118] The specific steps are as follows:

[0119] S321. Calculate the statistics of feature contribution scores: Based on the feature contribution scores obtained in step S31, calculate the mean of all scores. and standard deviation ;

[0120] S322. Determine the dynamic threshold: Based on the formula of the improved adaptive feature weighting algorithm, set the adjustment coefficient. The value is selected empirically within the range of [0.5, 2.0], and the dynamic threshold is calculated. This threshold is used to distinguish high-contribution features, where The value of the threshold affects its level. The larger the value, the higher the threshold and the more stringent the feature selection.

[0121] S323. Compare contribution scores with thresholds: Compare the contribution score of each feature with a dynamic threshold one by one. When the feature score is greater than If the value is positive, the feature is considered a high-contribution influence factor; otherwise, it is considered a low-contribution feature.

[0122] S324. Generate a primary feature subset: retain all features with scores higher than the threshold and combine them into a primary feature subset. This subset contains features that are strongly correlated with wind and solar power output and removes redundant information with low contribution.

[0123] S33. Apply principal component analysis and linear discriminant analysis algorithms to perform feature dimensionality reduction and redundancy elimination on the primary feature subset to generate a core feature vector set of wind and solar power. This includes the following steps:

[0124] S331. Data standardization preprocessing: Standardize the primary feature subset according to the feature dimension;

[0125] S332, Principal Component Analysis for Redundancy Removal:

[0126] Calculate the covariance matrix of the standardized data;

[0127] Perform eigenvalue decomposition on the covariance matrix to solve for its eigenvalues ​​and eigenvectors;

[0128] Arrange the eigenvalues ​​in descending order of size and calculate their cumulative variance contribution rate;

[0129] Based on the preset cumulative variance contribution rate threshold, select the top... The eigenvectors corresponding to the largest eigenvalues ​​form the projection matrix;

[0130] The original data is projected onto the selected principal component direction to obtain the dimension-reduced and redundant feature matrix;

[0131] S333, Linear Discriminant Analysis Optimization of Separability:

[0132] The samples were divided into multiple categories according to their corresponding wind and solar power output levels;

[0133] Calculate the intra-class and inter-class scatter matrices of the features after dimensionality reduction;

[0134] Solving the generalized eigenvalue problem formed by the two matrices above, selecting the first... The eigenvectors corresponding to the largest generalized eigenvalues ​​constitute the linear discriminant projection matrix;

[0135] S334. Feature Fusion and Output: The redundant feature matrix is ​​projected twice through the linear discriminant projection matrix to generate the final core feature vector set. The dimension of the core feature vector set is smaller than that of the original primary feature set, and the redundancy between features is eliminated, while the separability between different power level categories is enhanced.

[0136] The steps involved in building a deep learning prediction model based on the fusion of attention mechanism and spatiotemporal convolutional neural network in S4 are as follows:

[0137] S41. Construct a fusion neural network architecture that includes a spatiotemporal convolution module, an attention mechanism module, and a long short-term memory network module. The spatiotemporal convolution module extracts spatial features, the attention mechanism module weights important time step features, and the long short-term memory network module captures long-term temporal dependencies.

[0138] S42. The parameter initialization algorithm is used to initialize the parameters of the fused neural network architecture, generating an untrained prediction model. The parameter initialization algorithm adopts the Xavier initialization method, specifically including:

[0139] S421. For the weight matrix of the fully connected layer in a fusion neural network architecture, from a uniform distribution... Initialization is performed using random sampling;

[0140] S422, among which , This represents the number of neurons in the previous layer. This represents the number of neurons in the current layer.

[0141] S43. Use the core feature vector set of wind and solar power to train the untrained prediction model through forward and backward propagation. Optimize the model parameters through training to generate an initial model for wind and solar power prediction.

[0142] S5 employs a multi-objective particle swarm optimization algorithm for hyperparameter optimization and adaptive structure adjustment, including the following steps:

[0143] S51. Define a multi-objective optimization function, and simultaneously consider three optimization objectives: prediction accuracy, computational efficiency, and model complexity, to establish a mathematical model for the multi-objective optimization problem of wind and solar power prediction.

[0144] S52. An improved multi-objective particle swarm optimization algorithm is used to search for Pareto optimal solutions for the hyperparameters of the initial model for wind and solar power prediction, including key hyperparameters such as learning rate, number of hidden layer nodes, convolutional kernel size, and number of attention heads.

[0145] The improved position update formula for the multi-objective particle swarm optimization algorithm is:

[0146] ;

[0147] ;

[0148] in Indicates the first The particle in the first Position at the next iteration Indicates the first The particle in the first Speed ​​at the next iteration For inertial weights, and As a learning factor, and A random number in the range [0,1]. Indicates the first The optimal position of each individual particle. Indicates the globally optimal position;

[0149] S53. The model structure is adaptively adjusted using neural network architecture search technology to generate a lightweight network structure combining depthwise separable convolutional units and gated recurrent units, ultimately generating an optimized wind and solar power prediction model. The neural network architecture search technology employs a controller network based on reinforcement learning, specifically including:

[0150] S531, The controller network is implemented using a recurrent neural network (RNN) to generate the structural encoding of the deep learning prediction model;

[0151] S532, the structure encoding specifies the convolution kernel size, the number of hidden layer nodes, and whether to use gated recurrent units (GRUs);

[0152] S533. During training, the parameters of the controller network are updated using the proximal policy optimization algorithm to maximize the reward signal of the initial model for wind and solar power prediction on the validation set. The proximal policy optimization algorithm updates the parameters of the controller network by maximizing the expected reward. Its core objective function is as follows:

[0153] ;

[0154] in The probability ratio of the strategies. For the sample The estimate of the dominance function, This represents the empirical expectation for the current batch of samples. This is a hyperparameter.

[0155] The real-time acquisition of wind and solar power prediction calculations using wind and solar farm operation status data and short-term weather forecast data in S6 includes the following steps:

[0156] S61. Real-time acquisition of SCADA system operation data of wind and solar power plants and measured data of meteorological monitoring stations, and acquisition of high-precision meteorological forecast data for the next 72 hours through numerical weather prediction interface;

[0157] S62. Perform real-time data cleaning and feature engineering on operational data and weather forecast data to generate real-time prediction input feature vectors. This includes the following steps:

[0158] S621. Data validity verification: Verify the validity of each data point acquired in real time. To determine validity, when the data point value exceeds its physical possible range. If the data point is invalid, then mark it as an invalid value. Indicates the first The physical lower and upper bounds corresponding to each data point;

[0159] S622, Fixed Threshold Anomaly Filtering: For data points that pass the validity check, compare them with a predefined reasonable operating range. When comparing, Then, based on the valid data points immediately before and after it, linear interpolation is used for replacement, and the formula is:

[0160] ;

[0161] in These are the data point values ​​after anomaly filtering and interpolation correction. For the first real-time data stream The value of each data point;

[0162] S623, Simple Moving Average Smoothing: To suppress high-frequency measurement noise, a window size of [size missing] is applied to continuous numerical sequence data. The simple moving average filter is processed using the following formula:

[0163] ;

[0164] in The values ​​are smoothed. This is the window size for the moving average filter. To start from the current point The step size index for backtracking to the past;

[0165] S624. Real-time Feature Construction: Based on the cleaned instantaneous data, construct real-time features to characterize the dynamic changes of the system, mainly including:

[0166] Change rate characteristics: Calculating the short-term change rate of key parameters:

[0167] ;

[0168] in For data points The rate of change characteristic calculated at the location;

[0169] Mean ratio feature: Calculates the ratio of the current instantaneous value to the mean value over a short-term window.

[0170] ;

[0171] in For data points The characteristic of the mean ratio calculated at this point, This represents the average value over a short-term time window in the past.

[0172] S625, Vectorization: Combine all feature values ​​processed in the above steps in a predefined order to generate the final real-time prediction input feature vector. .

[0173] S63. Input the real-time predicted input feature vector into the optimized wind and solar power prediction model, perform forward inference calculation, and generate short-term wind and solar power prediction results for the next 72 hours, including power point prediction values ​​and probability prediction intervals.

[0174] The dynamic correction based on the prediction error feedback mechanism and sliding time window in S7 includes the following steps:

[0175] S71. Establish a prediction error correction model based on a combination of Kalman filtering and exponential smoothing to eliminate systematic errors in short-term wind and solar power prediction results. The establishment of the prediction error correction model specifically includes:

[0176] S711. Establish the system state equation and observation equation, and treat the prediction error as the system state;

[0177] S712. The latest observed prediction error is fused with the error estimate from the previous time step using Kalman gain to obtain the optimal estimate of the error at the current time step.

[0178] S713. Apply the exponential smoothing method to further smooth the optimal estimate to obtain the system error value used for correction.

[0179] S714. The system extreme error value is added as a compensation term to the short-term wind and solar power prediction results to eliminate system errors.

[0180] S72. The sliding time window algorithm is used to perform statistical analysis and pattern recognition on recent prediction errors, and the prediction model parameters are dynamically adjusted. This includes the following steps:

[0181] S721, Set a fixed-length sliding time window The window slides forward as new error data arrives, always maintaining a focus on the most recent data. Error data points Analysis;

[0182] in An index representing the latest current time point. For time indexing The prediction error value at the location;

[0183] S722. Within the sliding window, perform real-time statistical analysis on the prediction error and calculate a series of statistical characteristics, including:

[0184] Mean error:

[0185] ;

[0186] in The sample mean of the error sequence is... This indicates that from the current moment... The step size index for backtracking to the past;

[0187] Error standard deviation:

[0188] ;

[0189] in is the sample standard deviation of the error sequence;

[0190] Error trend slope: The slope of the linear trend of error over time within the fitted window using the least squares method. Used to determine whether the error exhibits a pattern of continuous increase or decrease;

[0191] S723. Calculate the statistical characteristic quantities. , , With the preset threshold , , Comparison and pattern recognition:

[0192] when If so, a systematic bias pattern is identified;

[0193] when Then, a pattern of increased uncertainty is identified;

[0194] when Then, the identification will show a continuous divergence and convergence pattern of errors;

[0195] S724. Based on the identified pattern, dynamically adjust the parameters of the prediction model:

[0196] When a systematic bias pattern is identified, the model output correction mechanism is activated, and the final predicted power value is dynamically corrected. Revised to To eliminate deviation;

[0197] When an increasing uncertainty pattern is identified, the bandwidth of the probability output interval of the prediction model is adaptively increased;

[0198] When a pattern of continuous error divergence is identified, a warning of model performance degradation is generated, and an incremental learning process is triggered to fine-tune the model using the latest data in the window.

[0199] S73. Apply confidence interval calibration technology to calibrate the probability prediction interval, generating high-precision wind and solar power prediction output data that meets the preset confidence level. The confidence interval calibration technology adopts the quantile regression calibration method, specifically including:

[0200] S731. Calculate the prediction bias of the probability prediction interval at different quantiles on the validation set;

[0201] S732. Construct a mapping function of quantile deviation with respect to predicted quantiles;

[0202] S733. Use a mapping function to calibrate the quantile points of future predictions so that the coverage of the calibrated probability prediction interval reaches the preset confidence level.

[0203] The use of blockchain technology for trusted evidence storage and distributed storage in S8 includes the following steps:

[0204] S81. Utilize a hash algorithm to extract digital digests from the high-precision wind and solar power prediction output data, generating unique data fingerprints. The hash algorithm used is SHA-256, specifically including:

[0205] S811, serialize the high-precision wind and solar power prediction output data into a byte stream;

[0206] S812. Apply the SHA-256 hash function to the byte stream to generate a 256-bit, 32-byte hash value;

[0207] S813. Use this hash value as a unique data fingerprint;

[0208] S82. By using smart contracts to write data fingerprints, timestamps, and prediction model version information together into the blockchain network, an immutable wind and solar power prediction audit record is generated. Writing this data into the blockchain network via smart contracts includes:

[0209] S821. Construct a smart contract that contains a publicly accessible function, storeHash;

[0210] S822, Call the storeHash function to send the data fingerprint, current block timestamp, and prediction model version information as transaction data to the blockchain network;

[0211] S823. After the nodes in the blockchain network verify the transaction through consensus, they package it into a new block to form an immutable audit record.

[0212] S83. Establish a blockchain-based prediction data sharing mechanism that allows authorized institutions to verify the credibility and completeness of prediction results through distributed ledger technology.

[0213] The operation steps of the AI-based wind and solar power prediction system and method are as follows:

[0214] Step 1: Multi-source data acquisition and standardization processing

[0215] The system collects multi-dimensional operational data such as wind turbine speed, pitch angle, photovoltaic string current, and voltage through the wind and solar power SCADA system, generating a raw wind and solar power time-series dataset. Simultaneously, it collects and acquires meteorological environmental data on wind speed, wind direction, irradiance, and cloud cover within the station and surrounding area through meteorological monitoring stations and numerical weather prediction systems, generating a high-dimensional meteorological feature dataset. Subsequently, the two datasets undergo timestamp alignment and spatial interpolation to form spatiotemporally consistent initial training data for wind and solar power. Finally, an outlier detection algorithm based on the 3σ criterion and box plot analysis, along with an improved KNN missing data imputation algorithm, is used to clean the initial data. This data is then processed using the Z-score normalization algorithm to ultimately generate a high-quality, standardized wind and solar power training sample set.

[0216] Step 2: Adaptive Feature Extraction and Dimensionality Reduction

[0217] The system extracts features of wind and solar power influencing factors from a standardized wind and solar power training sample set based on an improved adaptive feature weighting algorithm. First, a feature importance assessment model is constructed to calculate the contribution score of each feature variable. Then, an adaptive feature weighting algorithm is used to dynamically calculate and select a threshold based on the mean and standard deviation of the contribution scores, filtering out influencing factor features with contribution scores higher than the threshold to generate a primary feature subset. Finally, principal component analysis is applied to perform feature dimensionality reduction and redundancy elimination on this primary feature subset, generating a core feature vector set of wind and solar power that best represents the variation patterns of wind and solar power.

[0218] Step 3: Building and Training a Fusion Deep Learning Model

[0219] The system constructs a deep learning prediction model based on the fusion of an attention mechanism and a spatiotemporal convolutional neural network. This model integrates a spatiotemporal convolutional module, an attention mechanism module, and a long short-term memory network module, which are used to extract spatial features, weighted important time step features, and capture long-term temporal dependencies, respectively. The model is trained using the previously generated core feature vector set of wind and solar power, and the model parameters are optimized through forward and backward propagation processes to generate an initial model for wind and solar power prediction.

[0220] Step 4: Multi-objective optimization and model lightweighting

[0221] A multi-objective particle swarm optimization algorithm is employed to optimize the hyperparameters and adaptively adjust the structure of the initial wind and solar power prediction model. This algorithm simultaneously considers three optimization objectives: prediction accuracy, computational efficiency, and model complexity. Pareto optimal solutions are searched for key hyperparameters such as the learning rate and the number of hidden layer nodes. Simultaneously, a neural network architecture search technique is used to adaptively adjust the model structure, generating a lightweight network structure combining depthwise separable convolutional layers and gated recurrent units, ultimately yielding an optimized wind and solar power prediction model.

[0222] Step 5: Real-time prediction and dynamic correction

[0223] In the prediction application phase, the system collects real-time data on the operational status of wind and solar power farms and short-term weather forecasts. After undergoing the same real-time data cleaning and feature engineering processes, a real-time prediction input feature vector is generated. This vector is then input into the optimized wind and solar power prediction model for forward inference calculations, generating future short-term wind and solar power prediction results, including predicted power point values ​​and probability prediction intervals. Subsequently, based on a prediction error feedback mechanism and a sliding time window, the prediction results are dynamically corrected. A prediction error correction model is established to eliminate systematic errors, and confidence interval calibration technology is applied to calibrate the probability prediction intervals, ultimately generating high-precision wind and solar power prediction output data.

[0224] Step Six: Trusted Evidence Storage and Traceability

[0225] To ensure the credibility and auditability of the prediction results, the system utilizes blockchain technology to reliably store the high-precision wind and solar power prediction output data. A unique data fingerprint is generated through a hash algorithm, and the data fingerprint and timestamp information are written into the blockchain network via a smart contract, generating an immutable wind and solar power prediction audit trail. This provides authorized institutions with the ability to verify the credibility and integrity of the prediction results.

[0226] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0227] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind and solar power prediction system based on artificial intelligence, characterized in that, The system includes: The data acquisition and processing module acquires wind and solar power operation data and meteorological environment data through the multi-dimensional data acquisition unit, and outputs a standardized wind and solar power training sample set using the spatiotemporal alignment processing unit. The feature extraction module receives the standardized wind and solar power training sample set, extracts the wind and solar power influencing factor features through an improved adaptive feature weighting algorithm, and outputs a core feature vector set of wind and solar power, including the following steps: A feature importance evaluation model based on a combination of random forest and XGBoost is constructed to calculate the contribution score of each feature variable in the standardized wind and solar power training sample set. An improved adaptive feature weighting algorithm is used to dynamically threshold the contribution scores, filter out the influence factor features with contribution scores higher than the threshold, and generate a primary feature subset. The calculation formula for the improved adaptive feature weighting algorithm is as follows: ； in This indicates the threshold for dynamic feature selection. This represents the mean of the contribution scores of all features. This represents the standard deviation of the contribution scores of all features. The adjustment coefficient has a range of [0.5, 2.0]. Principal component analysis and linear discriminant analysis algorithms are applied to perform feature dimensionality reduction and redundancy elimination on the primary feature subset to generate a core feature vector set of wind and solar power. The intelligent prediction module receives the core feature vector set of wind and solar power, trains and optimizes it using a deep learning model that integrates an attention mechanism and a spatiotemporal convolutional neural network, and outputs the wind and solar power prediction result, including the following steps: Construct a fusion neural network architecture that includes a spatiotemporal convolution module, an attention mechanism module, and a long short-term memory network module; The parameters of the fusion neural network architecture are initialized using a parameter initialization algorithm to generate an untrained prediction model. The untrained prediction model is trained by forward and backward propagation using the core feature vector set of wind and solar power. The model parameters are optimized through training to generate an initial model for wind and solar power prediction. The dynamic correction module receives the wind and solar power prediction results, corrects the results through a prediction error feedback mechanism and a sliding time window, and outputs high-precision wind and solar power prediction data. The blockchain evidence storage module receives the high-precision wind and solar power prediction data, generates digital fingerprints through a hash algorithm, and uses smart contracts for distributed storage, outputting an immutable prediction audit trajectory. The human-computer interaction module receives the high-precision wind and solar power prediction data and audit trajectory, generates a multi-dimensional display interface through the visualization unit, and outputs a predictive analysis report that can be operated by the user.

2. An artificial intelligence-based wind and solar power prediction method, used to implement the artificial intelligence-based wind and solar power prediction system as described in claim 1, characterized in that: The method includes the following steps: S1. Collect multi-dimensional operational data and meteorological environment data of wind and solar power fields to generate raw wind and solar power time series datasets and high-dimensional meteorological feature datasets. S2. Perform data cleaning and spatiotemporal alignment on the original wind and solar power time series dataset and the high-dimensional meteorological feature dataset to generate a standardized wind and solar power training sample set. S3. Based on the improved adaptive feature weighting algorithm, wind and solar power influencing factor features are extracted from the standardized wind and solar power training sample set to generate a core feature vector set of wind and solar power. S4. Construct a deep learning prediction model based on the fusion of attention mechanism and spatiotemporal convolutional neural network, and train the deep learning prediction model using the core feature vector set of wind and solar power to generate an initial model for wind and solar power prediction. S5. The initial wind and solar power prediction model is optimized for hyperparameters and adaptively adjusted for structure using a multi-objective particle swarm optimization algorithm to generate an optimized wind and solar power prediction model. S6. Real-time collection of wind and solar power operation status data and short-term weather forecast data, input into the optimized wind and solar power prediction model to perform wind and solar power prediction calculation, and generate short-term wind and solar power prediction results. S7. Based on the prediction error feedback mechanism and sliding time window, the short-term wind and solar power prediction results are dynamically corrected to generate high-precision wind and solar power prediction output data. S8. Utilize blockchain technology to perform trusted notarization and distributed storage of the high-precision wind and solar power prediction output data, and generate an unalterable wind and solar power prediction audit trajectory.

3. The wind and solar power prediction method based on artificial intelligence according to claim 2, characterized in that: The collection of multi-dimensional operational data and meteorological environmental data of wind and solar power plants in S1 includes the following steps: S11. Collect the operating parameters of wind turbine generator set speed, pitch angle, and yaw angle, and the operating parameters of photovoltaic array string current, voltage, and backplane temperature through the wind and solar power SCADA system to generate the original wind and solar power time series dataset. S12. Collect meteorological parameters such as wind speed, wind direction, temperature, humidity, atmospheric pressure, and irradiance within the station area through meteorological monitoring stations, and obtain meteorological characteristic data such as cloud cover, aerosol concentration, and atmospheric transmittance within the region through numerical weather prediction systems to generate a high-dimensional meteorological characteristic dataset. S13. Perform timestamp alignment and spatial interpolation on the original wind and solar power time series dataset and the high-dimensional meteorological feature dataset to generate spatiotemporally consistent initial training data for wind and solar power.

4. The wind and solar power prediction method based on artificial intelligence according to claim 2, characterized in that: The data cleaning and spatiotemporal alignment process in S2 includes the following steps: S21. An anomaly detection algorithm based on the 3σ criterion and box plot analysis is used to identify and process outliers in the initial training data of spatiotemporal consistent wind and solar power, and preprocessed wind and solar power data is generated. S22. Apply the improved KNN missing data imputation algorithm to impute the missing values ​​in the preprocessed wind and solar power data to generate a continuous and complete wind and solar power dataset. S23. The continuous and complete wind and solar power dataset is standardized using the Z-score standardization algorithm to generate a standardized wind and solar power training sample set.

5. The wind and solar power prediction method based on artificial intelligence according to claim 2, characterized in that: The S5 process employs a multi-objective particle swarm optimization algorithm for hyperparameter optimization and adaptive structural adjustment, including the following steps: S51. Define a multi-objective optimization function, and simultaneously consider three optimization objectives: prediction accuracy, computational efficiency, and model complexity, to establish a mathematical model for the multi-objective optimization problem of wind and solar power prediction. S52. An improved multi-objective particle swarm optimization algorithm is used to search for Pareto optimal solutions for the hyperparameters of the initial model for wind and solar power prediction, including key hyperparameters such as learning rate, number of hidden layer nodes, convolutional kernel size, and number of attention heads. The position update formula for the improved multi-objective particle swarm optimization algorithm is: ； ； in Indicates the first The particle in the first Position at the next iteration Indicates the first The particle in the first Speed ​​at the next iteration For inertial weights, and As a learning factor, and A random number in the range [0,1]. Indicates the first The optimal position of each individual particle. Indicates the globally optimal position; S53. The model structure is adaptively adjusted through neural network architecture search technology to generate a lightweight network structure that combines depthwise separable convolutions and gated recurrent units, and finally an optimized wind and solar power prediction model is generated.

6. The wind and solar power prediction method based on artificial intelligence according to claim 2, characterized in that: The real-time acquisition of wind and solar power prediction calculations based on wind and solar farm operation status data and short-term weather forecast data in S6 includes the following steps: S61. Real-time acquisition of SCADA system operation data of wind and solar power plants and measured data of meteorological monitoring stations, and acquisition of high-precision meteorological forecast data for the next 72 hours through numerical weather prediction interface; S62. Perform real-time data cleaning and feature engineering on the operational data and weather forecast data to generate a real-time prediction input feature vector; S63. Input the real-time prediction input feature vector into the optimized wind and solar power prediction model, perform forward inference calculation, and generate short-term wind and solar power prediction results for the next 72 hours, including power point prediction values ​​and probability prediction intervals.

7. The wind and solar power prediction method based on artificial intelligence according to claim 2, characterized in that: The dynamic correction based on the prediction error feedback mechanism and sliding time window in S7 includes the following steps: S71. Establish a prediction error correction model based on the combination of Kalman filtering and exponential smoothing to eliminate systematic errors in the short-term wind and solar power prediction results; S72. A sliding time window algorithm is used to perform statistical analysis and pattern recognition on recent prediction errors, and the prediction model parameters are dynamically adjusted. S73. Apply confidence interval calibration technology to calibrate the probability prediction interval and generate high-precision wind and solar power prediction output data that meets the preset confidence level.

8. The wind and solar power prediction method based on artificial intelligence according to claim 2, characterized in that: The S8 method for trusted evidence storage and distributed storage using blockchain technology includes the following steps: S81. Use a hash algorithm to extract a digital digest from the high-precision wind and solar power prediction output data to generate a unique data fingerprint; S82. Write the data fingerprint, timestamp, and prediction model version information together into the blockchain network through a smart contract to generate an immutable wind and solar power prediction audit record. S83. Establish a blockchain-based prediction data sharing mechanism that allows authorized institutions to verify the credibility and completeness of prediction results through distributed ledger technology.

Images