CN-122000875-A - Photovoltaic power prediction method with self-adaptive depth modeling and super-parameter optimization

CN122000875ACN 122000875 ACN122000875 ACN 122000875ACN-122000875-A

Abstract

According to the photovoltaic power prediction method for self-adaptive depth modeling and super-parameter optimization, firstly, characteristic screening is carried out on meteorological factors through the pearson correlation coefficient and the maximum information coefficient, meteorological characteristics closely related to the photovoltaic power are selected, and the model is ensured to capture the linear and nonlinear relation between the photovoltaic power and the meteorological factors. And adopting a multivariate variation modal decomposition technology to perform collaborative modal decomposition on the meteorological data and the historical power data, suppressing noise and extracting multi-scale characteristic information. Parallel two-way long-short-term memory network and iTransformer subnetwork are designed for capturing short-term time sequence characteristics and long-range dependency relationship respectively and carrying out interactive fusion at a mode level. And performing global optimization on the super parameters of the model by using secretary bird optimization algorithm so as to further improve the prediction performance. The method can remarkably improve the prediction precision in the photovoltaic power prediction, and has better generalization capability and robustness compared with the traditional method.

Inventors

XIAO FAN
XIAO JINGLIANG
ZHONG BINBIN
CHEN HONGWEI
PENG MINGWEI
Yang Yangchun
WU YINXUAN
CHEN RUILIANG
ZHONG JINGJING
LIU YU

Assignees

中国能源建设集团浙江省电力设计院有限公司

Dates

Publication Date: 20260508
Application Date: 20260115

Claims (8)

1. A photovoltaic power prediction method with self-adaptive depth modeling and super-parameter optimization is characterized by comprising the following steps, S1, preprocessing and feature screening are carried out on original meteorological data and photovoltaic power data: S2, adopting a Multivariate Variation Modal Decomposition (MVMD) technology to conduct collaborative eigenmode decomposition on meteorological data and historical power data, suppressing noise and aliasing in the data, and extracting feature information of different scales; s3, designing a two-way long-short-term memory network (BiLSTM) and a iTransformer parallel subnet, respectively capturing short-term time sequence memory and long-range dependency relationships, simultaneously processing cross-modal interaction, and fusing at a modal level mapping layer; s4, introducing secretary bird an optimization algorithm (SBOA) to globally optimize the analysis model and the super parameters of the depth network.
2. The photovoltaic power prediction method of adaptive depth modeling and super-parametric optimization as claimed in claim 1, wherein two feature screening methods are used in S1: the pearson correlation coefficient is calculated to calculate the linear correlation between the meteorological factors and the historical power data, and the characteristics closely related to the photovoltaic power are further screened out; the Maximum Information Coefficient (MIC) is used for evaluating the nonlinear correlation of the multisource meteorological factors and screening out key features most related to the photovoltaic power; The first round of screening screens out the characteristic with strong linear correlation with photovoltaic power through the pearson correlation coefficient; the second round of screening uses the Maximum Information Coefficient (MIC) to further screen out features with strong nonlinear correlation with photovoltaic power, thereby ensuring that the selected features can capture both linear and nonlinear relationships.
3. The photovoltaic power prediction method of adaptive depth modeling and super-parametric optimization as claimed in claim 2, wherein S1 comprises the steps of: s11, data preprocessing and feature screening: All data are uniformly converted into fixed time resolution, and time stamp standardization and alignment are carried out to ensure the consistency of the sequence; carrying out explicit marking and retaining time information aiming at the missing value of the photovoltaic power data, and simultaneously, identifying and correcting the abnormal value to ensure the reliability of the data; S12, pelson correlation coefficient screening For each meteorological factor, the pearson correlation coefficient with photovoltaic power is calculated as: (1) in the formula, And Sample values of meteorological factors and photovoltaic power respectively, And Is the average value thereof; Selecting related features, namely selecting the first N features with strong linear relation with the photovoltaic power according to the calculated related coefficient value, and selecting to enter the next round of screening if the related coefficient is higher than the threshold value; s13, screening a Maximum Information Coefficient (MIC): the MIC method is introduced to measure the correlation between meteorological variables and photovoltaic power sequences, screen out the characteristics related to the target prediction height, calculate the mutual information And (3) with The formula of (2) is as follows (2) (3) Wherein: And The number of grids divided longitudinally and transversely respectively; Representation pair And Is a two-dimensional continuous integral of (a); 、 Respectively is And Is a boundary probability distribution of (1); for samples on a grid And Is a joint probability distribution in (a); Is 0.55 or 0.6 power of the corresponding data; combining the features screened by the pearson correlation coefficient and the MIC to obtain a candidate feature set, and if the same feature is selected in two rounds of screening, reserving the feature; s14 feature normalization After feature screening is completed, the screened features are normalized, and the specific operation is as follows: MinMax normalization, scaling the value of each feature to a range of [0,1], with the formula: (4) Wherein: Is the original characteristic value of the image, and the image is the original characteristic value, Is the minimum value of the feature and, The maximum value of the characteristic is the characteristic value after MinMax normalization, and the range of the characteristic value is ensured to be between 0 and 1.
4. The photovoltaic power prediction method of adaptive depth modeling and hyper-parameter optimization according to claim 3, wherein the modal decomposition in step S2 comprises: MVMD feature extraction and modal decomposition in photovoltaic power prediction S21 The MVMD mode decomposition method is as follows: In inputting the original sequence data comprising Data vector of individual channels Among them, the original sequence is expressed as In MVMD, the number of modes is set to be Decomposing an input signal into The aligned modal components satisfy Wherein each modal component is represented as ; Modulating the oscillating signal with multiple elements Performing Hilbert change, and recording the processed result as The center frequencies of all modal components of the signal are then set Adding exponential term and adjusting to corresponding frequency baseband, calculating Hilbert transform Gradient function Norms, establishing optimization constraints: (5) Wherein: Is a channel Is the first of (2) A modal component; The bandwidth constraint of the high-frequency oscillation is restrained; By adding a secondary punishment term of a reconstruction error and introducing a Lagrange multiplier term, punishment is carried out on constraint violation terms and strict satisfaction of constraint conditions is ensured, so that the original constrained variation problem is converted into an equivalent unconstrained optimization problem; Its corresponding augmented lagrangian function is expressed as: (6) Wherein: the Lagrangian multiplier corresponding to the constraint is introduced; s22, for efficiently solving the augmented Lagrangian function, updating parameters by using an alternate direction multiplier method, converting the parameters into a plurality of sub-optimization problems, and then solving the sub-optimization problems; Decomposing the signal frequency band after the whole steps to finally obtain the number of channels And modal component number Multiply the products by several narrowband eigenmode function (INTRINSIC MODE FUNCTION, IMF) components.
5. The photovoltaic power prediction method of adaptive depth modeling and super-parametric optimization as claimed in claim 4, wherein the parallel depth modeling process of step S3 is as follows: In order to consider both short-term memory and long-range dependence on a single prediction channel, a serial coding-prediction architecture is adopted; s31, a two-way long-short-term memory network (Bilstm) modeling process: BiLSTM the basic structural units and algorithms are derived from LSTM, which will input the final output result And the current variable First, the cell state is determined Is used for calculating input variables through a forgetting gate to obtain forgetting factors Next, input variables And Determining which new information of the cell state needs to be updated by sigmoid activation function of the input gate, and creating new cell candidate state Through the inlet door Activating a function; The specific calculation process is shown in the formulas (7) to (12); (7) (8) (9) (10) (11) (12) In the middle of , , , , , , , Is a weight matrix of the corresponding gates of the network , , , Is a bias matrix; And Respectively representing sigmoid and tanh activation functions; By bi-directional time series feature extraction BiLSTM has superior performance to LSTM, by combining bi-directional LSTM, the output result of BiLSTM is calculated as follows: (13) In the middle of And Output results of the forward LSTM and the backward LSTM are represented respectively, And Is a weight matrix; S32: iTransfomer modeling process: The self-attention mechanism in the multi-head attention layer is a core mechanism of a transducer series model, calculated attention scores are compared through calculating the attention of loads and environment variables, and then key information of multi-variable load prediction is selected in (6) - (7); (14) (15) Wherein, the 、 And The method comprises the steps of calculating a parameter matrix of Q, K and a parameter matrix of V vectors, wherein Q, K and V correspond to a query vector, a key vector and a value vector respectively, comprehensively extracting characteristic representation of each time sequence by calculating a dot product K between a query value and the key value, multiplying the characteristic representation by the value vector after exponential normalization, and obtaining attention expression A (Q, K and V); The gradient of softmax was eliminated; the scaling calculation is achieved by dividing Q.K7 by d, where d is the number of dimensions of K, the resulting attention score reveals the correlation between the multiple load and the environmental variable; for independent mapping, subspaces of n input variable sets are created, and different attention expression forms are established for parallel operation (16) (17) Wherein i represents the i-th attention head; A calculation expression representing an ith attention header; 、 And A query vector, a keyword vector, and a value vector, respectively, of the ith attention header of the input vector X; A weight matrix representing a multi-head attention merging computation expression; Is the number of attention headers created, Is a multi-head attention merging calculation expression; the output layer is a projection module which consists of MLP, the module carries out nonlinear mapping output through multivariable marks independently processed by the encoder module, and selects photovoltaic power generation, wind power generation and user power load prediction sequences for output; s33, outputting by a full-connection layer fusion module The outputs of BiLSTM and iTransformer are spliced (spliced) into a longer eigenvector before the output layer 、 This is achieved by stitching the output eigenvectors of the two subnetworks in the characteristic dimension, as represented by: (18) Spliced feature vector The full connection layer calculates a new characteristic representation through a weighted sum bias term, and then generates a final output through a nonlinear activation function, and the full connection layer calculates the formula: (19) Is a matrix of weights that are to be used, Is a bias term that is used to determine, Is the function of the activation and, Is the final prediction result after calculation of the full connection layer.
6. The photovoltaic power prediction method of adaptive depth modeling and superparameter optimization of claim 5, wherein the superparameter optimization process of step S4 is as follows: S41, initializing population: initializing the setting of a problem, wherein the setting comprises the dimension, the upper and lower bounds, the d size and the iteration times of a solution space; The dimension of the solution space; , Upper and lower bounds of the problem; Population size, which represents the number of solutions; maximum iteration times; The current iteration times; upon initializing the population, each individual Representing one potential solution: (20) Represent the first A solution for each dimension; the fitness of each individual is calculated by an objective function: (21) Representing the corresponding fitness; Optimal solution Is the optimal solution selected by comparing fitness of all individuals: (22) Wherein: representing solving variables that minimize the objective function; s42, exploration phase: The goal of the exploration phase is to search the solution space extensively to avoid trapping in locally optimal solutions. The stage is divided into three time periods, namely an initial stage, a middle stage and a later stage, and each stage uses different updating strategies; s43, development stage (Exploitation): the aim of the development stage is to refine and optimize the known optimal solution, so as to further improve the quality of the solution; S44, verifying and final modeling Validating the trained model by using validation set to obtain final TimesNet-BiLSTM network, inputting test set into final model, using And analyzing evaluation indexes such as Mean Absolute Percentage Error (MAPE), mean Absolute Error (MAE) and Root Mean Square Error (RMSE), and testing the prediction effect of the final model.
7. The method for photovoltaic power prediction with adaptive depth modeling and hyper-parameter optimization according to claim 6, wherein S42 comprises the steps of, The initial exploration stage is that : At this stage secretary bird updates its position by random motion, the position update formula is as follows: (23) in the formula, Representing the location of the individual after the update, This is the first Individual at the first Current position in dimension And These two are randomly generated solutions or the locations of the individuals, which represent the locations of the two solutions randomly selected from the search space, Is an adjustment factor, controls the search step; According to the fitness, the position is updated only when the fitness of the new position is better than the current solution: (24) in the formula, Representing the location of the individual after the update, Indicating the fitness of the new solution, Representing an old solution; Mid-term exploration phase : In the mid-term stage secretary bird begins to focus on the vicinity of the current optimal solution for optimization, and the update formula is as follows: (25) in the formula, Is the current optimal solution and is then, Is a random factor, controlling the randomness of the location update; When updating the position, the method also judges whether to accept the new solution according to the adaptability: (26) Later exploration stage : In the later stage secretary bird will further optimize the known optimal solution, converge to the optimal solution by local search, the location update formula is as follows: (27) in the formula, Is a random step size used to adjust the search process; also, it is decided whether to update the position according to the fitness: (28)。
8. The method for adaptive depth modeling and hyper-parameter optimization photovoltaic power prediction according to claim 7, wherein S43 comprises the steps of, Development stage (Exploitation): At this stage, the algorithm will perform a local search, selecting two strategies: a first policy is to be applied to the first policy, ; When random number Less than 0.5, secretary bird may perform local optimization based on the current optimal solution: (29) A second policy is provided for the second policy, ; When random number When the position is greater than or equal to 0.5, secretary bird can adjust based on random positions: (30) Each individual judges whether to select a new position according to the fitness value after updating, if the fitness of the new position is better, the current solution is updated, and the optimal solution is stored: W(31) finally, when the maximum iteration number T is reached, outputting an optimal solution.

Description

Photovoltaic power prediction method with self-adaptive depth modeling and super-parameter optimization Technical Field The invention relates to the technical field of photovoltaic power prediction, in particular to a photovoltaic power prediction method with self-adaptive depth modeling and super-parameter optimization. Background Photovoltaic power prediction is used as an important energy management technology and has wide application in the fields of power system scheduling, load prediction, energy transaction and the like. With the global increasing demand for clean energy, photovoltaic power generation has become one of the fastest growing renewable energy sources worldwide. However, the output of photovoltaic power is greatly affected by meteorological factors, including irradiation intensity, temperature, humidity, wind speed, etc., which are often time-sequential, non-stationary and multi-scale, resulting in photovoltaic power prediction facing challenges. At present, common photovoltaic power prediction methods mainly comprise a statistical model, a machine learning algorithm, a deep learning method and the like. Traditional statistical methods, such as time series analysis and regression models, although capable of providing a certain predictive effect, often fail to meet the requirements of practical applications in the face of complex weather changes and highly uncertain environments. Machine learning methods (e.g., support vector machines, random forests) can deal with certain non-linearity problems, but often rely on manual feature engineering, which is limited in its performance under large-scale, rapidly changing meteorological data. The deep learning method (such as LSTM, biLSTM, CNN) has become a mainstream technology in photovoltaic power prediction due to strong learning ability, but under the condition of high data redundancy and noise interference, the problems of low prediction accuracy, poor model generalization ability and the like still exist. Along with the expansion of the photovoltaic power generation scale, the traditional method is difficult to predict stably and efficiently when processing high-dimensional and multi-source data and complex weather changes. Therefore, it is highly desirable to provide a more accurate, robust and adaptive photovoltaic power prediction method to cope with the processing requirements of multi-source meteorological data, cross-scale problems and large-scale data, and further improve the scheduling efficiency and energy management capability of the photovoltaic power generation system. Disclosure of Invention The invention provides a photovoltaic power prediction method based on feature screening, modal decomposition, parallel depth modeling and super-parameter optimization, and aims to solve the problem that the accuracy and the robustness of the existing prediction method are insufficient under complex weather conditions. In order to achieve the above purpose, the present invention adopts the following technical scheme: A photovoltaic power prediction method with adaptive depth modeling and super-parameter optimization is implemented by computer equipment, Step S1, data preprocessing and feature screening: two feature screening methods were used: And A, calculating the linear correlation between the meteorological factors and the historical power data, and further screening out the characteristics closely related to the photovoltaic power. And B, maximum Information Coefficient (MIC) is used for evaluating the nonlinear correlation of the multisource meteorological factors and screening out key features most relevant to the photovoltaic power. The first round of screening screens out features with strong linear correlation with photovoltaic power through pearson correlation coefficients, and the correlation coefficient threshold is usually set to be 0.5. The second round of screening uses the Maximum Information Coefficient (MIC) to further screen out features with strong nonlinear correlation with photovoltaic power, thereby ensuring that the selected features can capture both linear and nonlinear relationships. And S2, performing modal decomposition, namely performing collaborative eigenmode decomposition on meteorological data and historical power data by adopting a Multivariate Variation Modal Decomposition (MVMD) technology, inhibiting noise and aliasing in the data, and extracting characteristic information of different scales. And step S3, parallel deep modeling, namely designing a two-way long-short-term memory network (BiLSTM) and a iTransformer parallel sub-network, respectively capturing short-term time sequence memory and long-range dependency relationships, simultaneously processing cross-modal interaction, and fusing at a modal level mapping layer. And S4, super-parameter optimization, namely introducing secretary bird an optimization algorithm (SBOA) to perform global optimization on the super-parameters of the analysis model and