A photovoltaic power prediction method and system based on multi-strategy optimization MVMD

Abstract

The application discloses a photovoltaic power prediction method and system based on a multi-strategy optimized MVMD, and the method comprises the following steps: performing AP clustering according to seasons to construct a similar day mode library, and matching a to-be-predicted day to extract a training data set; adopting a multi-strategy improved Guanhuo pig optimization algorithm to adaptively optimize the decomposition layer number and the penalty factor of the MVMD with the minimum average envelope entropy as the target; synchronously decomposing the power-weather sequence by using the optimal MVMD to obtain mode components corresponding to central frequencies one by one; dividing the power components into a low-frequency trend set and a high-frequency fluctuation set according to the daily change base frequency, and inputting the low-frequency trend set and the high-frequency fluctuation set into an LSTM-KAN subnetwork and a CNN-KAN subnetwork respectively for prediction; reconstructing and outputting an initial prediction value; training a residual error model by using the first-order change rate of the weather to output a dynamic error compensation quantity, and superimposing the dynamic error compensation quantity to obtain a final prediction result. The application solves the problems of the existing technology, such as no physical correspondence of mode decomposition, dependence of hyperparameters on artificial setting, prediction lag under weather mutation and the like, and significantly improves the photovoltaic power prediction precision and the generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of power automation technology, and in particular relates to a photovoltaic power prediction method and system based on multi-strategy optimization MVMD. Background Technology

[0002] Photovoltaic power generation is characterized by strong randomness, fluctuation, and intermittency. Its output characteristics are directly affected by meteorological factors such as solar irradiance, ambient temperature, and humidity. Especially under sudden weather changes, photovoltaic output can experience sharp increases and decreases, posing a severe challenge to the safe and stable operation of the power grid and power dispatch. Therefore, improving the accuracy of photovoltaic power forecasting is a core prerequisite for smoothing out output fluctuations and optimizing dispatch plans.

[0003] Currently, data-driven methods based on deep learning have become the mainstream research direction in the field of photovoltaic power prediction. In response to the non-stationary and highly volatile characteristics of photovoltaic power sequences, the industry widely adopts the technical framework of "decomposition-prediction-reconstruction". That is, the original non-stationary sequence is decomposed into multiple stationary intrinsic mode components by signal decomposition methods such as variational mode decomposition (VMD) to reduce the difficulty of sequence prediction. Then, intelligent optimization algorithms are combined to optimize the decomposition parameters and model hyperparameters. Finally, deep learning models such as recurrent neural networks and temporal convolutional networks are used to complete the prediction of each component and the reconstruction of the results.

[0004] Chinese invention patent application CN121124032A proposes a photovoltaic power generation prediction method based on VMD-ISSA-BiGRU. This method decomposes the photovoltaic power sequence into eight intrinsic mode components using VMD, employs a strategy-improved Sparrow Search Algorithm (ISSA), optimizes the hyperparameters of the BiGRU-Transformer model, utilizes BiGRU to capture local dependencies of time-series features, and uses Transformer to capture global long-term dependencies. However, this method only performs VMD decomposition on a single photovoltaic power sequence, failing to decompose the power sequence into a meteorological sequence. The decomposed mode components lack physical correspondence, failing to fully exploit the strong coupling relationship between meteorological factors and photovoltaic output, and easily affecting prediction accuracy in scenarios of sudden weather changes. Summary of the Invention

[0005] This invention provides a photovoltaic power prediction method and system based on multi-strategy optimized MVMD, which is used to solve the problems of prediction lag and poor physical interpretability under meteorological changes.

[0006] In a first aspect, the present invention provides a photovoltaic power prediction method based on multi-strategy optimized MVMD, comprising: Historical photovoltaic power data and corresponding historical meteorological data are obtained. Grey relational analysis is used to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data, and strong correlated meteorological factors are screened out. Cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homologous sequence within the similar day cluster as the model training dataset. The improved hog optimization algorithm using a multi-strategy approach aims to minimize the average envelope entropy of the decomposed modal components. It adaptively optimizes the number of decomposition layers and the penalty factor of the multivariate variational mode decomposition to obtain the optimal combination of hyperparameters. Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, the training dataset of the model is simultaneously decomposed to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies. According to the daily variation fundamental frequency division and determination rules, the power intrinsic mode function components are divided into a low-frequency trend set and a high-frequency fluctuation set; The low-frequency trend set is input into a pre-constructed LSTM-KAN sub-network for prediction to obtain the low-frequency trend prediction value. The high-frequency fluctuation set is concatenated with the meteorological intrinsic mode function components of the corresponding center frequency and then input into a pre-constructed CNN-KAN sub-network for prediction to obtain the high-frequency fluctuation prediction value. The temporal hidden state features output by the LSTM-KAN subnetwork and the spatial features output by the CNN-KAN subnetwork are mapped in the same dimension and weighted and concatenated to reconstruct the initial photovoltaic power prediction value. Using the first-order rate of change of the strongly correlated meteorological factors as input, a residual prediction model is trained, and a dynamic error compensation amount is output. The dynamic error compensation amount is then superimposed with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0007] Secondly, the present invention provides a photovoltaic power prediction system based on multi-strategy optimized MVMD, comprising: The acquisition module is configured to acquire historical photovoltaic power data and historical meteorological data corresponding to the historical photovoltaic power data, and to use grey relational analysis to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data, and to screen out strongly correlated meteorological factors. The matching module is configured to cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homology sequence within the similar day cluster as the model training dataset. The first decomposition module is configured to use the improved Crowned Hog optimization algorithm with multiple strategies. The optimization objective is to minimize the average envelope entropy of the decomposed modal components. The module adaptively optimizes the number of decomposition layers and the penalty factor of the multivariate variational mode decomposition to obtain the optimal hyperparameter combination. The second decomposition module is configured to perform synchronous decomposition on the model training dataset according to the multivariate variational mode decomposition model configured based on the optimal hyperparameter combination, to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies. The partitioning module is configured to divide the power intrinsic mode function components into a low-frequency trend set and a high-frequency fluctuation set according to the daily variation fundamental frequency partitioning judgment rule; The prediction module is configured to input the low-frequency trend set into a pre-constructed LSTM-KAN sub-network for prediction to obtain low-frequency trend prediction values, and to concatenate the high-frequency fluctuation set with the meteorological intrinsic mode function components of the corresponding center frequency and input them into a pre-constructed CNN-KAN sub-network for prediction to obtain high-frequency fluctuation prediction values. The splicing module is configured to perform same-dimensional mapping and weighted splicing of the temporal hidden state features output by the LSTM-KAN sub-network and the spatial features output by the CNN-KAN sub-network to reconstruct the initial photovoltaic power prediction value. The output module is configured to take the first-order rate of change of the strongly correlated meteorological factors as input, train the residual prediction model, output the dynamic error compensation amount, and superimpose the dynamic error compensation amount with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the photovoltaic power prediction method based on multi-strategy optimized MVMD according to any embodiment of the present invention.

[0009] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the photovoltaic power prediction method based on multi-strategy optimized MVMD according to any embodiment of the present invention.

[0010] The photovoltaic power prediction method and system based on multi-strategy optimization MVMD in this application have the following beneficial effects: By filtering strongly correlated meteorological factors using grey relational analysis and constructing a similar day pattern library using seasonal AP clustering, accurate matching of meteorological patterns for the day to be predicted was achieved, significantly improving the model's generalization ability in different seasons and meteorological scenarios, and solving the problem of poor model adaptability caused by training with full data. The MCPO algorithm, which improves the MVMD parameters through a multi-strategy approach, employs four major improvement strategies: Circle chaotic mapping, adaptive Gaussian random walk, dynamic sine and cosine update, and nonlinear convergence factor. These strategies solve the problems of the standard algorithm being prone to getting trapped in local optima and having insufficient optimization accuracy, and thus obtain better decomposition hyperparameters. The MVMD was used to simultaneously decompose the power series and the meteorological series, ensuring that the frequencies of each modal component after decomposition corresponded one-to-one, enhancing the physical correspondence between power and meteorology, and providing high-quality input for subsequent differential modeling. The LSTM-KAN network is designed to address the long-term temporal dependence of low-frequency trend components, and the CNN-KAN network is designed to address the local mutation characteristics of high-frequency fluctuation components. This fully leverages the stronger nonlinear fitting capability of the KAN network compared to traditional fully connected layers. At the same time, the adaptive weighted fusion of temporal and spatial features is achieved through a gated spatiotemporal fusion module, thereby improving prediction accuracy. By employing a two-stage training strategy, the LightGBM residual model is trained using the first-order change rate of meteorological factors as input. This effectively compensates for the dynamic error of the main model under meteorological change scenarios, solves the prediction lag problem, and further improves prediction accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart of a photovoltaic power prediction method based on multi-strategy optimization MVMD provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a photovoltaic power prediction system based on multi-strategy optimized MVMD, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0014] Please see Figure 1 The diagram shows a flowchart of a photovoltaic power prediction method based on multi-strategy optimization MVMD according to this application.

[0015] like Figure 1 As shown, the photovoltaic power prediction method based on multi-strategy optimization MVMD specifically includes the following steps: Step S101: Obtain historical photovoltaic power data and historical meteorological data corresponding to the historical photovoltaic power data; use grey relational analysis to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data; and screen out strongly correlated meteorological factors.

[0016] In this step, the photovoltaic output power data of the photovoltaic power station is obtained every 15 minutes. The corresponding meteorological observation data includes six dimensions: total horizontal irradiance, direct normal irradiance, ambient temperature, relative humidity, wind speed, and wind direction.

[0017] Outliers are removed using the interquartile range (IQR) method: the upper quartile Q3 and lower quartile Q1 of each sequence are calculated, and the interquartile range IQR = Q3 - Q1 is set. The outlier threshold is set to [Q1 - 1.5IQR, Q3 + 1.5IQR]. Samples outside the range are identified as outliers, and linear interpolation of adjacent normal samples is used to replace the outliers.

[0018] Min-max normalization was used to map all data to the [0,1] interval. The normalized photovoltaic power sequence was used as the reference sequence, and each meteorological sequence was used as a comparison sequence. The resolution coefficient ρ=0.5, and the grey relational degree was calculated. In this embodiment, the calculated values ​​were: total horizontal irradiance 0.92, direct normal irradiance 0.89, ambient temperature 0.78, relative humidity 0.65, wind speed 0.52, and wind direction 0.31. A threshold of 0.6 was set, and five strongly correlated meteorological factors (N=5) were selected: total horizontal irradiance, direct normal irradiance, ambient temperature, relative humidity, and wind speed.

[0019] Step S102: Cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homologous sequence within the similar day cluster as the model training dataset.

[0020] In this step, the data is divided into four calendar months: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). Historical photovoltaic power data and historical meteorological data are then further divided into spring, summer, autumn, and winter subsets based on the calendar month. For each seasonal subset, the daily series of strongly correlated meteorological factors selected within the seasonal subset are used as clustering samples. A sample similarity matrix is ​​constructed using negative Euclidean distance, expressed as follows: , In the formula, For the sample With sample Similarity value between them , The samples are two independent samples, each of which is a vector of strongly correlated meteorological factors in a daily series. Based on the sample similarity matrix, the nearest neighbor propagation clustering algorithm is used iteratively to update the responsibility matrix and availability matrix. The update formula is as follows: , , In the formula, For the sample at the t-th iteration Select Sample The responsibility value as a cluster center For the sample at the t-th iteration With sample Similarity between them For the sample at the t-th iteration With sample Similarity between them For the sample at the t-th iteration Suitable as a sample Availability value of cluster centers For the sample As the responsibility value of its own cluster center For the sample at the t-th iteration Select Sample The responsibility value as a cluster center; Introducing damping factor The responsibility matrix and availability matrix are updated smoothly using the following expression: , , In the formula, , These are the responsibility value and availability value after the (t+1)th iteration, respectively. , These are the responsibility matrix and availability matrix for the (t+1)th iteration, respectively. , These are the responsibility matrix and availability matrix for the t-th iteration, respectively; Repeat the iterations until the matrix converges or the maximum number of iterations is reached, which will satisfy... samples Selected as the cluster center; A similar day pattern library is established for each season. The similar day pattern library stores the historical photovoltaic power-meteorological homology sequences of each cluster center and the corresponding clusters under each season.

[0021] It should be noted that the daily series of strongly correlated meteorological factors for the day to be predicted are obtained; Determine the target season to which the day to be predicted belongs, and obtain a target similar day pattern library corresponding to the target season; The Euclidean distance between the daily series of the strongly correlated meteorological factors and the cluster centers in the target similarity daily pattern library is calculated using the following formula: , In the formula, For the first Cluster centers, This refers to the daily series of strongly correlated meteorological factors for the forecast of the day to be predicted. For sequence length, For the first The cluster centers at the in The value of each sampling point The first in the daily series of strongly correlated meteorological factors The values ​​of each sampling point; Match the day to be predicted to the similar day cluster corresponding to the cluster center with the smallest Euclidean distance; All historical photovoltaic power-meteorological homologous sequences within similar daily clusters were extracted as the model training dataset.

[0022] Step S103: Using the improved multi-strategy hog optimization algorithm, with the goal of minimizing the average envelope entropy of the decomposed modal components, the decomposition level and penalty factor of the multivariate variational mode decomposition are adaptively optimized to obtain the optimal hyperparameter combination.

[0023] In this step, the algorithm parameters are initialized, and the number of decomposition layers is set. With penalty factor The optimization interval; An initial population uniformly distributed within the interval [0,1] is generated using the Circle chaotic mapping population initialization strategy. The Circle mapping formula is as follows: , In the formula, For the first The position vector of each individual in the population For the first The position vector of each individual in the population, corresponding to the decomposition level. With penalty factor Parameter combinations, , For mapping coefficients, This is the modulo operation; The individual position is updated using an adaptive Gaussian random walk visual defense strategy, with the following formula: , , In the formula, , They are respectively the (t+1)th and tth iterations. The position vector of each individual in the population Let be the optimal position of the individual in the population during the t-th iteration. The mean is variance is Gaussian distribution, A random number within the interval [0,1]. For adaptive iterative weights, , These represent the maximum and minimum weights, respectively, and t is the current iteration number. This represents the maximum number of iterations. The individual position is updated using a dynamic sine and cosine physical attack defense strategy, with the following formula: , , , , In the formula, As an adaptive weighting factor, As a dynamic search factor, , A random number within the interval [0,1]. These are parameters for inelastic collision forces. This is a search direction control parameter, with a value of 1 or -1. For defense coefficient, This is the collision scaling factor, used to control the average force of inelastic collisions. The strength of the effect on position updates is a constant between (0,1). The average force of inelastic collisions is calculated using the standard porcupine optimization algorithm. A global-local equilibrium strategy for nonlinear convergence factors is adopted, with the convergence factor... The update formula is: , In the formula, Let be the convergence factor for the t-th iteration. The initial value of the convergence factor is... Where p is the attenuation coefficient and p is the nonlinear adjustment exponent; The fitness function is defined as minimizing the average envelope entropy of the decomposed modal components, and the formula is as follows: , In the formula, To decompose the number of layers With penalty factor When used as input parameters, the average envelope entropy value obtained from multivariate variational mode decomposition is used, where M is the number of homologous sequences, including photovoltaic power sequences and strongly correlated meteorological sequences. The m-th homologous sequence after multivariate variational mode decomposition is the... Normalized envelope signal of each intrinsic mode component; In each iteration, the fitness value under the current hyperparameter combination is calculated, the population is updated according to the position update mechanism of the improved hog optimization algorithm, and the historical best solution is recorded. When the preset maximum number of iterations is reached, output the decomposition level with the smallest fitness value. With penalty factor The combination is taken as the optimal hyperparameter combination.

[0024] Step S104: Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, the model training dataset is synchronously decomposed to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies.

[0025] In this step, the historical photovoltaic power sequences in the model training dataset are aligned with each strongly correlated meteorological sequence by time to form a multivariate time series matrix. ,in, It is a power sequence; For the first A strongly correlated meteorological sequence, For time sampling points; Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, variational mode decomposition is performed on the multivariate time series matrix to construct a constrained variational problem. The objective of the constrained variational problem is to minimize the sum of the estimated bandwidths of each modal component, and the constraint condition is that the sum of all components is equal to the original sequence. Introducing a penalty factor and Lagrange multipliers This transforms the constrained variational problem into an augmented Lagrangian function. The alternating direction multiplier method is used to alternately update the center frequency and amplitude of each eigenmode function component in the frequency domain until the convergence condition is met, resulting in a set of K eigenmode functions with the same frequency. Output K IMF groups with the same frequency. Each group contains the power IMF component and the IMF components of each meteorological factor corresponding to the same center frequency. The expression is: , In the formula, This is the k-th power intrinsic mode function component; For the first The k-th intrinsic mode function component of a meteorological factor This represents the number of decomposition layers.

[0026] Step S105: According to the daily variation fundamental frequency division judgment rule, the power intrinsic mode function components are divided into a low-frequency trend set and a high-frequency fluctuation set.

[0027] In this step, the fundamental frequency of the daily variation in photovoltaic power is calculated. And based on the daily variation frequency of the photovoltaic power Set frequency threshold ,in, For threshold coefficient, This represents the total number of sampling points corresponding to the daily photovoltaic power generation duration. Perform a Fast Fourier Transform on each power eigenmode function component to calculate its center frequency, expressed as: , , In the formula, For the Fast Fourier Transform operator, For the frequency domain amplitude spectrum, The angular frequency corresponding to the maximum amplitude. The center frequency of the power eigenmode function component; center frequency The power IMF components are divided into low-frequency trend sets; center frequency The power IMF components are divided into a high-frequency fluctuation set.

[0028] Step S106: Input the low-frequency trend set into the pre-constructed LSTM-KAN sub-network for prediction to obtain the low-frequency trend prediction value; and concatenate the high-frequency fluctuation set with the meteorological intrinsic mode function components of the corresponding center frequency and input them into the pre-constructed CNN-KAN sub-network for prediction to obtain the high-frequency fluctuation prediction value.

[0029] In this step, the temporal hidden state features output by the LSTM-KAN (Long Short-Term Memory - Kolmogorov-Arnold Network) subnetwork are respectively... and the spatial features output by the CNN-KAN (Convolutional Neural Network - Kolmogorov-Arnold Network) subnetwork. Inputting independent Kolmogorov-Arnold mapping layers, nonlinear dimension alignment is achieved through spline functions to obtain temporal feature vectors with consistent dimensions. With spatial feature vectors The expression is: , , In the formula, and It is a learnable Kolmogorov-Arnold mapping layer containing an adjustable B-spline mesh; The adaptive weights of the temporal features and spatial features are calculated using a gating mechanism, as shown in the formula: , , In the formula, The weight matrix is ​​a learnable matrix with dimensions D×D. For bias terms, It is a sigmoid activation function, and the output value is between (0,1). For adaptive weights of time-series features, Adaptive weights for spatial features; The temporal features and spatial features are weighted and concatenated according to the adaptive weights to obtain a fused feature vector. The formula is: , In the formula, For feature splicing operations; The fused feature vector is input into the Kolmogorov-Arnold output layer to complete nonlinear reconstruction and output the initial photovoltaic power prediction value. The expression is: , In the formula, The Kolmogorov-Arnold output layer contains a learnable spline activation function, and the output is a scalar prediction.

[0030] Step S107: Perform same-dimensional mapping and weighted concatenation on the temporal hidden state features output by the LSTM-KAN subnetwork and the spatial features output by the CNN-KAN subnetwork to reconstruct the initial photovoltaic power prediction value.

[0031] In this step, the low-frequency trend set is... The power IMF component is input into the Long Short-Term Memory (LSTM) network layer to extract long-term temporal trend features, and outputs temporal hidden state features. The expression is: , In the formula, It is a long short-term memory network layer, containing forget gates, input gates, and output gates; The temporal hidden state features output from the Long Short-Term Memory (LSTM) network layer are input into the Kolmogorov-Arnold network layer, and nonlinearly mapped using a learnable B-spline activation function to output low-frequency trend prediction values. The expression is: , In the formula, Kolmogorov-Arnold network layer for low-frequency prediction; The high-frequency fluctuation set The power IMF component in the matrix is ​​concatenated with the meteorological IMF component at the same center frequency along the channel dimension to construct a multi-channel time-series feature matrix. The expression is: , In the formula, This is a channel-level splicing operation. For the k-th power intrinsic mode function component, For the first The k-th intrinsic mode function component of a meteorological factor The number of meteorological factors. This indicates that for a set of high-frequency fluctuations Each eigenmode in will be all All IMF components corresponding to each meteorological factor were included in the splicing; By employing multiple one-dimensional convolutional layers and pooling layers, local meteorological abrupt change features and abrupt change correlation features are extracted from multi-channel features, outputting spatial abrupt change features. The expression is: , In the formula, It is a one-dimensional convolutional layer. It is a linear rectification activation function; This is the max pooling layer; The spatial mutation features output from the convolutional neural network layer are input into the Kolmogorov-Arnold network layer to perform nonlinear fitting of the high-frequency fluctuation features and output high-frequency fluctuation prediction values. The expression is: , In the formula, This is a Kolmogorov-Arnold network layer used for high-frequency prediction.

[0032] Step S108: Using the first-order change rate of the strongly correlated meteorological factors as input, train the residual prediction model, output the dynamic error compensation amount, and superimpose the dynamic error compensation amount with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0033] In this step, the first-order rate of change of the strongly correlated meteorological factors is calculated using the following formula: , In the formula, For the first Strongly correlated meteorological factors The first rate of change at time t, For the first Strongly correlated meteorological factors The normalized value at time 10:00 For the first Strongly correlated meteorological factors The normalized value at time 10:00 For discrete-time indexing; The first-order rates of change of all strongly correlated meteorological factors are aligned by time to form a first-order rate of change sequence. The first-order rate of change sequence is used as input, and the prediction residuals at the corresponding time are used as training targets to train a residual prediction model. The prediction residual of the residual prediction model is the difference between the measured power and the initial photovoltaic power prediction value. The first-order rate of change of the forecast meteorological data for the day to be predicted is input into the trained residual prediction model, and the dynamic error compensation amount is output. The dynamic error compensation amount is superimposed on the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0034] In summary, the method in this application constructs a similar daily pattern library by performing AP clustering in different seasons and extracts training datasets by matching the days to be predicted; it adopts a multi-strategy improved crown porcupine optimization algorithm to adaptively optimize the number of decomposition layers and penalty factor of MVMD with the goal of minimizing the average envelope entropy; it uses the optimal MVMD to synchronously decompose the power-meteorological sequence to obtain modal components with one-to-one correspondence of center frequencies; it divides the power components into low-frequency trend and high-frequency fluctuation sets according to the daily variation fundamental frequency, and inputs them into LSTM-KAN sub-networks and CNN-KAN sub-networks for prediction, respectively; it reconstructs the output of the initial prediction value; it trains the residual model with the first-order meteorological change rate to output the dynamic error compensation amount, and superimposes it to obtain the final prediction result; it solves the problems of no physical correspondence in modal decomposition, hyperparameter dependence on manual setting, and prediction lag under meteorological changes in the prior art, and significantly improves the accuracy and generalization ability of photovoltaic power prediction.

[0035] Please see Figure 2 The diagram shows a structural block diagram of a photovoltaic power prediction system based on multi-strategy optimized MVMD according to this application.

[0036] like Figure 2 As shown, the photovoltaic power prediction system 200 includes an acquisition module 210, a matching module 220, a first decomposition module 230, a second decomposition module 240, a division module 250, a prediction module 260, a splicing module 270, and an output module 280.

[0037] The acquisition module 210 is configured to acquire historical photovoltaic power data and corresponding historical meteorological data, and use grey relational analysis to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data to screen out strongly correlated meteorological factors. The matching module 220 is configured to cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, and match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homologous sequences within the similar day clusters as the model training dataset. The first decomposition module 230 is configured to use a multi-strategy improved crown porcupine optimization algorithm, with the minimum average envelope entropy of the decomposed modal components as the optimization objective, adaptively optimize the number of decomposition layers and the penalty factor of the multivariate variational mode decomposition to obtain the optimal hyperparameter combination. The second decomposition module 240 is configured to simultaneously decompose the model training dataset according to the multivariate variational mode decomposition model configured according to the optimal hyperparameter combination to obtain several sets of power corresponding to the center frequency. The system includes: a partitioning module 250, configured to partition the power intrinsic mode function components into a low-frequency trend set and a high-frequency fluctuation set based on the diurnal variation fundamental frequency partitioning rule; a prediction module 260, configured to input the low-frequency trend set into a pre-constructed LSTM-KAN sub-network for prediction to obtain a low-frequency trend prediction value, and to concatenate the high-frequency fluctuation set with the meteorological intrinsic mode function components of the corresponding center frequency and input them into a pre-constructed CNN-KAN sub-network for prediction to obtain a high-frequency fluctuation prediction value; a concatenation module 270, configured to perform same-dimensional mapping and weighted concatenation of the temporal hidden state features output by the LSTM-KAN sub-network and the spatial features output by the CNN-KAN sub-network to reconstruct and output an initial photovoltaic power prediction value; and an output module 280, configured to train a residual prediction model using the first-order change rate of the strongly correlated meteorological factors as input, output a dynamic error compensation amount, and superimpose the dynamic error compensation amount with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0038] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0039] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the photovoltaic power prediction method based on multi-strategy optimization MVMD in any of the above method embodiments. In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows: Historical photovoltaic power data and corresponding historical meteorological data are obtained. Grey relational analysis is used to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data, and strong correlated meteorological factors are screened out. Cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homologous sequence within the similar day cluster as the model training dataset. The improved hog optimization algorithm using a multi-strategy approach aims to minimize the average envelope entropy of the decomposed modal components. It adaptively optimizes the number of decomposition layers and the penalty factor of the multivariate variational mode decomposition to obtain the optimal combination of hyperparameters. Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, the training dataset of the model is simultaneously decomposed to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies. According to the daily variation fundamental frequency division and determination rules, the power intrinsic mode function components are divided into a low-frequency trend set and a high-frequency fluctuation set; The low-frequency trend set is input into a pre-constructed LSTM-KAN sub-network for prediction to obtain the low-frequency trend prediction value. The high-frequency fluctuation set is concatenated with the meteorological intrinsic mode function components of the corresponding center frequency and then input into a pre-constructed CNN-KAN sub-network for prediction to obtain the high-frequency fluctuation prediction value. The temporal hidden state features output by the LSTM-KAN subnetwork and the spatial features output by the CNN-KAN subnetwork are mapped in the same dimension and weighted and concatenated to reconstruct the initial photovoltaic power prediction value. Using the first-order rate of change of the strongly correlated meteorological factors as input, a residual prediction model is trained, and a dynamic error compensation amount is output. The dynamic error compensation amount is then superimposed with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0040] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the photovoltaic power prediction system based on multi-strategy optimized MVMD, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected to the photovoltaic power prediction system based on multi-strategy optimized MVMD via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the photovoltaic power prediction method based on multi-strategy optimization (MVMD) as described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the photovoltaic power prediction system based on multi-strategy optimization (MVMD). The output device 340 may include a display screen or other display device.

[0042] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0043] In one implementation, the above-described electronic device is applied to a photovoltaic power prediction system based on multi-strategy optimization (MVMD) for use as a client, and includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: Historical photovoltaic power data and corresponding historical meteorological data are obtained. Grey relational analysis is used to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data, and strong correlated meteorological factors are screened out. Cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homologous sequence within the similar day cluster as the model training dataset. The improved hog optimization algorithm using a multi-strategy approach aims to minimize the average envelope entropy of the decomposed modal components. It adaptively optimizes the number of decomposition layers and the penalty factor of the multivariate variational mode decomposition to obtain the optimal combination of hyperparameters. Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, the training dataset of the model is simultaneously decomposed to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies. According to the daily variation fundamental frequency division and determination rules, the power intrinsic mode function components are divided into a low-frequency trend set and a high-frequency fluctuation set; The low-frequency trend set is input into a pre-constructed LSTM-KAN sub-network for prediction to obtain the low-frequency trend prediction value. The high-frequency fluctuation set is concatenated with the meteorological intrinsic mode function components of the corresponding center frequency and then input into a pre-constructed CNN-KAN sub-network for prediction to obtain the high-frequency fluctuation prediction value. The temporal hidden state features output by the LSTM-KAN subnetwork and the spatial features output by the CNN-KAN subnetwork are mapped in the same dimension and weighted and concatenated to reconstruct the initial photovoltaic power prediction value. Using the first-order rate of change of the strongly correlated meteorological factors as input, a residual prediction model is trained, and a dynamic error compensation amount is output. The dynamic error compensation amount is then superimposed with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

[0044] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic power prediction method based on multi-strategy optimization MVMD, characterized in that, include: Historical photovoltaic power data and corresponding historical meteorological data are obtained. Grey relational analysis is used to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data, and strong correlated meteorological factors are screened out. Cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homologous sequence within the similar day cluster as the model training dataset. An improved multi-strategy optimization algorithm for the Crowned Porcupine is used, with the goal of minimizing the average envelope entropy of the decomposed modal components. It adaptively optimizes the number of decomposition levels and the penalty factor in the multivariate variational mode decomposition to obtain the optimal hyperparameter combination, specifically including: Initialize algorithm parameters and set the number of decomposition layers. With penalty factor The optimization interval; An initial population uniformly distributed within the interval [0,1] is generated using the Circle chaotic mapping population initialization strategy. The Circle mapping formula is as follows: ， In the formula, For the first The position vector of each individual in the population For the first The position vector of each individual in the population, corresponding to the decomposition level. With penalty factor Parameter combinations, , For mapping coefficients, This is the modulo operation; The individual position is updated using an adaptive Gaussian random walk visual defense strategy, with the following formula: ， ， In the formula, , They are respectively the (t+1)th and tth iterations. The position vector of each individual in the population Let this be the optimal position of the individual in the population during the t-th iteration. The mean is variance is Gaussian distribution, A random number within the interval [0,1]. For adaptive iterative weights, , These represent the maximum and minimum weights, respectively, and t is the current iteration number. This represents the maximum number of iterations. The individual position is updated using a dynamic sine and cosine physical attack defense strategy, with the following formula: ， ， ， ， In the formula, As an adaptive weighting factor, As a dynamic search factor, , A random number within the interval [0,1]. These are parameters for inelastic collision forces. This is a search direction control parameter, with a value of 1 or -1. For defense coefficient, This is the collision scaling factor, used to control the average force of inelastic collisions. The strength of the effect on position updates is a constant between (0,1). The average force of inelastic collisions is calculated using the standard porcupine optimization algorithm. A global-local equilibrium strategy for nonlinear convergence factors is adopted, with the convergence factor... The update formula is: ， In the formula, Let be the convergence factor for the t-th iteration. The initial value of the convergence factor is... Where p is the attenuation coefficient and p is the nonlinear adjustment exponent; The fitness function is defined as minimizing the average envelope entropy of the decomposed modal components, and the formula is as follows: ， In the formula, To decompose the number of layers With penalty factor When used as input parameters, the average envelope entropy value obtained from multivariate variational mode decomposition is used, where M is the number of homologous sequences, including photovoltaic power sequences and strongly correlated meteorological sequences. The m-th homologous sequence after multivariate variational mode decomposition is the... Normalized envelope signal of each intrinsic mode component; In each iteration, the fitness value under the current hyperparameter combination is calculated, the population is updated according to the position update mechanism of the improved hog optimization algorithm, and the historical best solution is recorded. When the preset maximum number of iterations is reached, output the decomposition level with the smallest fitness value. With penalty factor The combination is taken as the optimal hyperparameter combination; Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, the training dataset of the model is simultaneously decomposed to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies. According to the daily variation fundamental frequency division and determination rules, the power intrinsic mode function components are divided into a low-frequency trend set and a high-frequency fluctuation set; The low-frequency trend set is input into a pre-constructed LSTM-KAN sub-network for prediction to obtain the low-frequency trend prediction value. The high-frequency fluctuation set is concatenated with the meteorological intrinsic mode function components of the corresponding center frequency and then input into a pre-constructed CNN-KAN sub-network for prediction to obtain the high-frequency fluctuation prediction value. The temporal hidden state features output by the LSTM-KAN subnetwork and the spatial features output by the CNN-KAN subnetwork are mapped in the same dimension and weighted and concatenated to reconstruct the initial photovoltaic power prediction value. Using the first-order rate of change of the strongly correlated meteorological factors as input, a residual prediction model is trained, and a dynamic error compensation amount is output. The dynamic error compensation amount is then superimposed with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

2. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The step of clustering the data of the strongly correlated meteorological factors according to seasonal type to construct a similar day pattern library includes: Historical photovoltaic power data and historical meteorological data are divided into spring subset, summer subset, autumn subset, and winter subset according to the Gregorian calendar month; For each seasonal subset, the daily series of strongly correlated meteorological factors selected within the seasonal subset are used as clustering samples. A sample similarity matrix is ​​constructed using negative Euclidean distance, expressed as follows: ， In the formula, For the sample With sample Similarity value between them , The samples are two independent samples, each of which is a vector of strongly correlated meteorological factors in a daily series. Based on the sample similarity matrix, the nearest neighbor propagation clustering algorithm is used iteratively to update the responsibility matrix and availability matrix. The update formula is as follows: ， ， In the formula, For the sample at the t-th iteration Select Sample The responsibility value as a cluster center For the sample at the t-th iteration With sample Similarity between them For the sample at the t-th iteration With sample Similarity between them For the sample at the t-th iteration Suitable as a sample Availability value of cluster centers For the sample As the responsibility value of its own cluster center For the sample at the t-th iteration Select Sample The responsibility value as a cluster center; Introducing damping factor The responsibility matrix and availability matrix are updated smoothly using the following expression: ， ， In the formula, , These are the responsibility value and availability value after the (t+1)th iteration, respectively. , These are the responsibility matrix and availability matrix for the (t+1)th iteration, respectively. , These are the responsibility matrix and availability matrix for the t-th iteration, respectively; Repeat the iterations until the matrix converges or the maximum number of iterations is reached, which will satisfy... samples Selected as the cluster center; A similar day pattern library is established for each season. The similar day pattern library stores the historical photovoltaic power-meteorological homology sequences of each cluster center and the corresponding clusters under each season.

3. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The step of matching the forecast meteorological data for the day to be predicted to the corresponding similar day cluster in the similar day model library, and extracting the historical photovoltaic power-meteorological homologous sequence within the similar day cluster as the model training dataset includes: Obtain the daily series of strongly correlated meteorological factors for the day to be predicted; Determine the target season to which the day to be predicted belongs, and obtain a target similar day pattern library corresponding to the target season; The Euclidean distance between the daily series of the strongly correlated meteorological factors and the cluster centers in the target similarity daily pattern library is calculated using the following formula: ， In the formula, For the first Cluster centers, This refers to the daily series of strongly correlated meteorological factors for the forecast of the day to be predicted. For sequence length, For the first The cluster centers at the in The value of each sampling point The first in the daily series of strongly correlated meteorological factors The values ​​of each sampling point; Match the day to be predicted to the similar day cluster corresponding to the cluster center with the smallest Euclidean distance; All historical photovoltaic power-meteorological homologous sequences within similar daily clusters were extracted as the model training dataset.

4. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The synchronous decomposition of the model training dataset to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequency includes: The historical photovoltaic power sequences in the model training dataset are aligned with each strongly correlated meteorological sequence by time to form a multivariate time series matrix. ,in, It is a power sequence; For the first A strongly correlated meteorological sequence, For time sampling points; Based on the multivariate variational mode decomposition model configured with the optimal hyperparameter combination, variational mode decomposition is performed on the multivariate time series matrix to construct a constrained variational problem. The objective of the constrained variational problem is to minimize the sum of the estimated bandwidths of each modal component, and the constraint condition is that the sum of all components is equal to the original sequence. Introducing a penalty factor and Lagrange multipliers This transforms the constrained variational problem into an augmented Lagrangian function. The alternating direction multiplier method is used to alternately update the center frequency and amplitude of each eigenmode function component in the frequency domain until the convergence condition is met, resulting in a set of K eigenmode functions with the same frequency. Output K sets of intrinsic mode functions (IMFs) at the same frequency. Each set contains the power IMF component corresponding to the same center frequency and the IMF components of various meteorological factors. The expression is as follows: ， In the formula, This is the k-th power intrinsic mode function component; For the first The k-th intrinsic mode function component of a meteorological factor This represents the number of decomposition layers.

5. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The step of dividing the power intrinsic mode function components into a low-frequency trend set and a high-frequency fluctuation set according to the daily variation fundamental frequency division judgment rule includes: Calculate the fundamental frequency of daily variation in photovoltaic power = And based on the daily variation frequency of the photovoltaic power Set frequency threshold · ,in, For threshold coefficient, This represents the total number of sampling points corresponding to the daily photovoltaic power generation duration. Perform a Fast Fourier Transform on each power eigenmode function component to calculate its center frequency, expressed as: ， F ， In the formula, For the Fast Fourier Transform operator, For the frequency domain amplitude spectrum, The angular frequency corresponding to the maximum amplitude. The center frequency of the power eigenmode function component; center frequency The power IMF components are divided into low-frequency trend sets; center frequency The power IMF components are divided into a high-frequency fluctuation set.

6. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The process of mapping and weighting the temporal hidden state features output by the LSTM-KAN subnetwork to the spatial features output by the CNN-KAN subnetwork to reconstruct the initial photovoltaic power prediction value includes: The temporal hidden state features output by the LSTM-KAN sub-network respectively and the spatial features output by the CNN-KAN subnetwork Inputting independent Kolmogorov-Arnold mapping layers, nonlinear dimension alignment is achieved through spline functions to obtain temporal feature vectors with consistent dimensions. With spatial feature vectors The expression is: ， ， In the formula, and It is a learnable Kolmogorov-Arnold mapping layer containing an adjustable B-spline mesh; The adaptive weights of the temporal features and spatial features are calculated using a gating mechanism, as shown in the formula: , ， In the formula, The weight matrix is ​​a learnable matrix with dimensions D×D. For bias terms, It is a sigmoid activation function, and the output value is between (0,1). For adaptive weights of time-series features, Adaptive weights for spatial features; The temporal features and spatial features are weighted and concatenated according to the adaptive weights to obtain a fused feature vector. The formula is: ， In the formula, For feature splicing operations; The fused feature vector is input into the Kolmogorov-Arnold output layer to complete nonlinear reconstruction and output the initial photovoltaic power prediction value. The expression is: ， In the formula, The Kolmogorov-Arnold output layer contains a learnable spline activation function, and its output is a scalar prediction.

7. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The steps of inputting the low-frequency trend set into a pre-constructed LSTM-KAN sub-network for prediction to obtain low-frequency trend prediction values, and concatenating the high-frequency fluctuation set with the meteorological intrinsic mode function components of the corresponding center frequency and inputting the concatenation into a pre-constructed CNN-KAN sub-network for prediction to obtain high-frequency fluctuation prediction values ​​include: The low-frequency trend set The power IMF component is input into the Long Short-Term Memory (LSTM) network layer to extract long-term temporal trend features, and outputs temporal hidden state features. The expression is: ， In the formula, It is a long short-term memory network layer, containing forget gates, input gates, and output gates; The temporal hidden state features output from the Long Short-Term Memory (LSTM) network layer are input into the Kolmogorov-Arnold network layer, and nonlinearly mapped using a learnable B-spline activation function to output low-frequency trend prediction values. The expression is: ， In the formula, Kolmogorov-Arnold network layer for low-frequency prediction; The high-frequency fluctuation set The power IMF component in the matrix is ​​concatenated with the meteorological IMF component at the same center frequency along the channel dimension to construct a multi-channel time-series feature matrix. The expression is: ， In the formula, This is a channel-level splicing operation. For the k-th power intrinsic mode function component, For the first The k-th intrinsic mode function component of a meteorological factor The number of meteorological factors. This indicates that for a set of high-frequency fluctuations Each eigenmode in will be all All IMF components corresponding to each meteorological factor were included in the splicing; By employing multiple one-dimensional convolutional layers and pooling layers, local meteorological abrupt change features and abrupt change correlation features are extracted from multi-channel features, outputting spatial abrupt change features. The expression is: ， In the formula, It is a one-dimensional convolutional layer. It is a linear rectification activation function; This is the max pooling layer; The spatial mutation features output from the convolutional neural network layer are input into the Kolmogorov-Arnold network layer to perform nonlinear fitting of the high-frequency fluctuation features and output high-frequency fluctuation prediction values. The expression is: ， In the formula, This is a Kolmogorov-Arnold network layer used for high-frequency prediction.

8. The photovoltaic power prediction method based on multi-strategy optimization MVMD according to claim 1, characterized in that, The process of training a residual prediction model using the first-order rate of change of the strongly correlated meteorological factors as input, outputting a dynamic error compensation amount, and then superimposing the dynamic error compensation amount with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result includes: The first-order rate of change of the strongly correlated meteorological factors is calculated using the following formula: ， In the formula, For the first Strongly correlated meteorological factors The first rate of change at time t, For the first Strongly correlated meteorological factors The normalized value at time 10:00 For the first Strongly correlated meteorological factors The normalized value at time 10:00 For discrete-time indexing; The first-order rates of change of all strongly correlated meteorological factors are aligned by time to form a first-order rate of change sequence. The first-order rate of change sequence is used as input, and the prediction residuals at the corresponding time are used as training targets to train a residual prediction model. The prediction residual of the residual prediction model is the difference between the measured power and the initial photovoltaic power prediction value. The first-order rate of change of the forecast meteorological data for the day to be predicted is input into the trained residual prediction model, and the dynamic error compensation amount is output. The dynamic error compensation amount is superimposed on the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

9. A photovoltaic power prediction system based on multi-strategy optimization MVMD, characterized in that, include: The acquisition module is configured to acquire historical photovoltaic power data and historical meteorological data corresponding to the historical photovoltaic power data, and to use grey relational analysis to analyze the correlation between the historical photovoltaic power data and the multidimensional meteorological factors in the historical meteorological data, and to screen out strongly correlated meteorological factors. The matching module is configured to cluster the data of the strongly correlated meteorological factors according to the seasonal type, construct a similar day pattern library, match the forecast meteorological data of the day to be predicted to the corresponding similar day cluster in the similar day pattern library, and extract the historical photovoltaic power-meteorological homology sequence within the similar day cluster as the model training dataset. The first decomposition module is configured to utilize a multi-strategy improved hog optimization algorithm. With the goal of minimizing the average envelope entropy of the decomposed modal components, it adaptively optimizes the number of decomposition layers and the penalty factor in the multivariate variational mode decomposition to obtain the optimal hyperparameter combination. Specifically, it includes: Initialize algorithm parameters and set the number of decomposition layers. With penalty factor The optimization interval; An initial population uniformly distributed within the interval [0,1] is generated using the Circle chaotic mapping population initialization strategy. The Circle mapping formula is as follows: ， In the formula, For the first The position vector of each individual in the population For the first The position vector of each individual in the population, corresponding to the decomposition level. With penalty factor Parameter combinations, , For mapping coefficients, This is the modulo operation; The individual position is updated using an adaptive Gaussian random walk visual defense strategy, with the following formula: ， ， In the formula, , They are respectively the (t+1)th and tth iterations. The position vector of each individual in the population Let this be the optimal position of the individual in the population during the t-th iteration. The mean is variance is Gaussian distribution, A random number within the interval [0,1]. For adaptive iterative weights, , These represent the maximum and minimum weights, respectively, and t is the current iteration number. This represents the maximum number of iterations. The individual position is updated using a dynamic sine and cosine physical attack defense strategy, with the following formula: ， ， ， ， In the formula, As an adaptive weighting factor, As a dynamic search factor, , A random number within the interval [0,1]. These are parameters for inelastic collision forces. This is a search direction control parameter, with a value of 1 or -1. For defense coefficient, This is the collision scaling factor, used to control the average force of inelastic collisions. The strength of the effect on position updates is a constant between (0,1). The average force of inelastic collisions is calculated using the standard porcupine optimization algorithm. A global-local equilibrium strategy for nonlinear convergence factors is adopted, with the convergence factor... The update formula is: ， In the formula, Let be the convergence factor for the t-th iteration. The initial value of the convergence factor is... Where p is the attenuation coefficient and p is the nonlinear adjustment exponent; The fitness function is defined as minimizing the average envelope entropy of the decomposed modal components, and the formula is as follows: ， In the formula, To decompose the number of layers With penalty factor When used as input parameters, the average envelope entropy value obtained from multivariate variational mode decomposition is used, where M is the number of homologous sequences, including photovoltaic power sequences and strongly correlated meteorological sequences. The m-th homologous sequence after multivariate variational mode decomposition is the... Normalized envelope signal of each intrinsic mode component; In each iteration, the fitness value under the current hyperparameter combination is calculated, the population is updated according to the position update mechanism of the improved hog optimization algorithm, and the historical best solution is recorded. When the preset maximum number of iterations is reached, output the decomposition level with the smallest fitness value. With penalty factor The combination is taken as the optimal hyperparameter combination; The second decomposition module is configured to perform synchronous decomposition on the model training dataset according to the multivariate variational mode decomposition model configured based on the optimal hyperparameter combination, to obtain several sets of power intrinsic mode function components and meteorological intrinsic mode function components corresponding to the center frequencies. The partitioning module is configured to divide the power intrinsic mode function components into a low-frequency trend set and a high-frequency fluctuation set according to the daily variation fundamental frequency partitioning judgment rule; The prediction module is configured to input the low-frequency trend set into a pre-constructed LSTM-KAN sub-network for prediction to obtain low-frequency trend prediction values, and to concatenate the high-frequency fluctuation set with the meteorological intrinsic mode function components of the corresponding center frequency and input them into a pre-constructed CNN-KAN sub-network for prediction to obtain high-frequency fluctuation prediction values. The splicing module is configured to perform same-dimensional mapping and weighted splicing of the temporal hidden state features output by the LSTM-KAN sub-network and the spatial features output by the CNN-KAN sub-network to reconstruct the initial photovoltaic power prediction value. The output module is configured to take the first-order rate of change of the strongly correlated meteorological factors as input, train the residual prediction model, output the dynamic error compensation amount, and superimpose the dynamic error compensation amount with the initial photovoltaic power prediction value to obtain the final photovoltaic power prediction result.

Images