Ultra-short-term photovoltaic power combined prediction method based on multivariate meteorological data

Abstract

The invention provides an ultra-short-term photovoltaic power combined prediction method based on multivariate meteorological data, belongs to the technical field of photovoltaic power generation power prediction, and aims to solve the problem that in photovoltaic power generation prediction, model training time can be increased by considering excessive non-key influence factors, and then convergence speed and prediction precision are affected. The method comprises the following steps: acquiring a photovoltaic power generation power sequence, and performing multi-scale decomposition and K-L divergence screening; performing correlation analysis on the random meteorological variables in the photovoltaic power generation power sequence and the to-be-predicted indexes, and selecting meteorological data having the maximum correlation with the to-be-predicted indexes according to a correlation analysis result; an ECA-TCN-BiLSTM combined prediction model is constructed; and inputting the decomposed and screened photovoltaic power generation power data and the selected meteorological data into an ECA-TCN-BiLSTM combined prediction model, and superposing and integrating prediction results of the IMF components to obtain a final ultra-short-term photovoltaic power prediction value.

Description

Technical Field

[0001] The invention relates to an ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data, and belongs to the technical field of photovoltaic power generation prediction. Background Art

[0002] Solar energy is the most abundant and promising clean energy source among renewable energy sources. Photovoltaic power generation, a technology that directly converts solar energy into electricity, is gaining increasing attention for its safety, efficiency, cost-effectiveness, and environmental friendliness. In recent years, with the rapid development of clean and efficient energy generation technologies, the installed capacity of photovoltaic power generation has continued to increase. Compared to other renewable energy generation technologies, photovoltaic power generation is more mature and reliable, and is currently the most predominant form of new energy generation technology. However, photovoltaic power generation is susceptible to seasonal and weather factors, resulting in significant randomness and volatility. Integrating photovoltaic power generation devices into power systems can bring challenges and impacts. Improving the accuracy of photovoltaic power generation forecasts can not only help grid operators better dispatch generation resources and optimize power system operations, but also reduce the power system's reliance on peak-shaving backup power sources, thereby reducing operating costs. Therefore, photovoltaic power generation forecasts are crucial for the stable operation of modern power systems.

[0003] Based on the time scale, photovoltaic power generation forecasts can be divided into medium- and long-term forecasts, short-term forecasts, and ultra-short-term forecasts. Medium- and long-term forecasts are the foundation for power system planning, while short-term and ultra-short-term forecasts are the basis for dispatching departments to arrange grid operation modes and formulate dispatching strategies. To obtain accurate photovoltaic power generation forecasts, researchers have proposed many forecasting methods in recent years, which can be roughly divided into three types: statistical methods, physical methods, deep learning methods, and hybrid methods. Hybrid methods are the most prominent. As a parallel and effective method in hybrid model prediction, the hybrid model prediction method based on signal decomposition has attracted attention in various fields, including photovoltaic power generation forecasting, wind speed forecasting, wind power forecasting, lithium battery life forecasting, and financial asset return forecasting.

[0004] Photovoltaic power generation portfolio forecasting primarily involves data processing and model prediction. Photovoltaic power generation forecasting involves analyzing and predicting nonstationary data sequences. The accuracy of its model is often closely related to the extraction and processing methods of time series features. Feature construction in data processing aims to improve anomalous features or amplify effective features. Feature construction typically focuses on three aspects: feature dimensionality reduction, feature clustering, and feature derivation. Feature clustering relies on algorithms such as K-means, hierarchical clustering, DBSCAN, and Dynamic Time Warping (DTW) to classify data. For example, DTW measures the similarity of time series morphology to identify similar fluctuation patterns. Its advantages lie in the efficiency of K-means, the noise tolerance of DBSCAN, and the adaptability of DTW to time series alignment. However, the pre-set number of clusters (e.g., K-means), parameter sensitivity (e.g., DBSCAN), and the emphasis on static distributions can weaken the representation of the time-dependent characteristics of photovoltaic power, leading to biased pattern extraction. Feature derivation constructs new features through statistical aggregation (such as sliding window mean), frequency domain transformation, or domain knowledge integration (such as irradiance-temperature ratio). Studies have shown that such methods can deeply mine implicit data information and improve prediction accuracy in complex scenarios. However, their design relies heavily on prior knowledge, and automated derivation can easily generate invalid features. Furthermore, some black-box models, while performing well, lack interpretability due to unclear mechanisms, limiting their reliability in practical applications. In feature dimensionality reduction, principal component analysis (PCA) preserves high-variance features based on orthogonal transformations, improving computational efficiency and reducing redundant interference. However, due to its linearity assumption, it struggles to capture nonlinear temporal relationships. Signal decomposition algorithms, with their ability to decouple the essential characteristics of time series signals, offer unique advantages, making them suitable for scenarios such as photovoltaic power, which exhibit strong temporal dependencies and nonstationary fluctuations. For example, VMD, by constructing a variational model to constrain modal center frequencies, can effectively separate the fluctuation patterns of different frequency features. Algorithms like the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) significantly improve modal aliasing by introducing white noise as an auxiliary analysis. However, existing signal decomposition algorithms still face many challenges: First, the problem of modal aliasing is difficult to completely avoid in complex fluctuation scenarios (such as sudden changes in photovoltaic power due to cloud obstruction), resulting in cross-interference of signal components at different time scales; second, the algorithm's high dependence on manually preset parameters (such as VMD requires specifying the number of decomposition layers, and CEEMDAN relies on noise intensity parameters) reduces its adaptability and is prone to over- or under-decomposition due to parameter missetting; third, endpoint effects and noise sensitivity may introduce boundary distortion or residual noise, especially in the first and last segments of time series signals or in high-noise environments, making the physical interpretability of the decomposition results easily questionable; fourth, traditional methods (such as EMD) are still limited to fixed decomposition logic for the analysis of nonlinear dynamic characteristics, making it difficult to dynamically adapt to the cross-scale coupling characteristics of photovoltaic power.

[0005] Existing research on solar power prediction models is primarily categorized into three main categories: physical models, statistical models, and hybrid models. While the first two models are simple to implement, they are highly data-dependent and have weak generalization capabilities. With the advancement of artificial intelligence (AI), models based on machine learning and deep learning are able to more deeply exploit data features through continuous iterative learning, thereby demonstrating superior algorithmic performance. Some studies have used the XGBoost algorithm, which fuses multiple linear decision trees, to improve model hierarchy and prediction accuracy. However, this approach struggles with handling nonlinear relationships between variables. Another ensemble learning approach, combining a convolutional neural network (CNN) with a long short-term memory network (LSTM), has successfully achieved accurate real-time prediction of multi-mode photovoltaic power generation. However, the LSTM model suffers from slow training speed and difficulty in parameter tuning. Furthermore, traditional CNNs are limited by fixed-size convolution kernels, making it difficult to effectively capture long-term dependencies. Temporal convolutional networks (TCNs) offer unique advantages for processing time series data. Their parallel processing capabilities significantly improve computational efficiency, achieving prediction accuracy comparable to recurrent neural network (RNN) algorithms. However, while this network expands the receptive field through dilated convolution, it is still constrained by the size of the convolution kernel, limiting its effectiveness in processing long-range correlation features. Other studies have improved the residual structure, shifting from a serial distribution to a parallel distribution to enhance model universality, but such improvements have not yet addressed the problem of representing the varying importance of different features in time series. Summary of the Invention

[0006] In order to solve the problems in photovoltaic power generation prediction that considering too many non-critical influencing factors will increase model training time, thereby affecting the convergence speed and prediction accuracy, and insufficiently capturing the importance differences of time series features, the present invention proposes an ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data.

[0007] The technical solution adopted by the present invention to solve the above problems is: the present invention comprises the following steps:

[0008] Step 1: Obtain the photovoltaic power generation sequence, and perform multi-scale decomposition and KL divergence screening on the photovoltaic power generation data sequence in the photovoltaic power generation sequence;

[0009] Step 2: Perform correlation analysis on the random meteorological variables in the photovoltaic power generation sequence and the indicator to be predicted. Select the meteorological data with the greatest correlation with the indicator to be predicted based on the correlation analysis results.

[0010] Step 3: Build the ECA-TCN-BiLSTM combined prediction model;

[0011] Step 4: Input the decomposed and filtered photovoltaic power generation data and the selected meteorological data into the ECA-TCN-BiLSTM combined prediction model, superimpose and integrate the prediction results of each IMF component to obtain the final ultra-short-term photovoltaic power prediction value.

[0012] Furthermore, step 1 specifically includes:

[0013] Step 1.1: Obtain photovoltaic power generation sequence, which includes photovoltaic power generation data and random meteorological variable data;

[0014] Step 1.2: Use the DPEC-ICEEMDAN algorithm to perform multi-scale decomposition on the PV power data to obtain a series of IMF components and a residual term. The DPEC-ICEEMDAN algorithm is an improved adaptive noise complete empirical mode decomposition method based on dynamic physical constraint entropy termination.

[0015] Step 1.3: Perform kernel density estimation and discretization calculation on the photovoltaic power generation data and each IMF component, and filter out the KL divergence D of the photovoltaic power generation data corresponding to each IMF component. KL , for each IMF component according to its sampled KL divergence D KL Sort by size and select the one that satisfies D KL <∈ IMF component, output corresponding decomposed and filtered photovoltaic power data.

[0016] Furthermore, step 1.2 specifically includes:

[0017] Step 1.2.1: Set the original photovoltaic power data sequence as x(t), introduce a dynamic noise adjustment mechanism, and calculate the signal local variance σ 2 (t), according to the local variance σ of the signal 2 (t) Real-time adjustment of the noise ratio k;

[0018] Step 1.2.2: According to the noise ratio θ k Constructing a decomposition sequence According to the decomposition sequence Calculate the k-th order residual r of the photovoltaic power generation data sequence x(t) k (t);

[0019] Step 1.2.3: According to the k-th order residual r of the photovoltaic power generation data sequence x(t) k (t) and the k-1th order residual r k-1 (t) extracting the k-th order IMF component;

[0020] Step 1.2.4: When k≤2, the irradiance-power conversion constraint is applied based on the extracted k-th order IMF component, eliminating the modal aliasing caused by temperature fluctuations;

[0021] Step 1.2.5: Obtain the energy ratio η according to the L2 norm of the k-th order IMF component k Repeat steps 1.2.1-1.2.4 for iterative decomposition until the current decomposition energy ratio η k <5%, obtaining a series of IMF components and a residual term;

[0022] The calculation formula of local variance σ 2 (t) is:

[0023]

[0024] The calculation formula of noise ratio θ k is:

[0025]

[0026] In formulas (1) and (2), is the power mean in the sliding window, W is the dynamic noise adjustment window length, Var(x(t)) is the variance of the time series x(t), is the mean of the sequence, i.e. T is the total time step;

[0027] The calculation formula of the variance Var(x(t)) of the time series x(t) is:

[0028]

[0029] The calculation formula of the decomposed sequence is:

[0030]

[0031] In formula (4), is the white noise added in the k-th decomposition, E k (·) is the k-th component after pre-decomposition of noise;

[0032] The calculation formula of the k-th residual r k (t) of the photovoltaic power generation data sequence x(t) is:

[0033]

[0034] In formula (5), M(·) is the local mean operation, and N is the number of noise addition times;

[0035] The calculation formula of the k-th IMF component is:

[0036] IMF k (t) = r k-1 (t)-r k (t) (6);

[0037] The expression of the k-th order IMF component after applying the irradiance-power conversion constraint is:

[0038]

[0039] In formula (7), G(t) is the real-time irradiance, G STC is the standard irradiance;

[0040] Energy proportion η k The calculation formula is:

[0041]

[0042] In formula (8), η k is the energy ratio, ||IMF k ||2 is the L2 norm of the k-th order IMF, 5% is determined by Monte Carlo experiments, and 95% of the effective energy components are retained.

[0043] Furthermore, step 1.3 specifically includes:

[0044] Step 1.3.1: Perform kernel density estimation on the original photovoltaic power data sequence x(t) and each of its corresponding IMF components to obtain their probability density functions;

[0045] Step 1.3.2: For each IMF component, calculate the KL divergence between it and the original PV power data;

[0046] Step 1.3.3: Sample z uniformly in the interval [zmin,zmax] j , set the sampling step Δz, and obtain the KL divergence D of the photovoltaic power data corresponding to each IMF component KL ;

[0047] Step 1.3.4: For each IMF component, calculate the KL divergence D obtained by sampling. KL Sort by size and select the one that satisfies D KL <∈IMF component, output the corresponding decomposed and filtered photovoltaic power data;

[0048] The calculation formula of the probability density function is:

[0049]

[0050] In formulas (9) and (10), p x(z) is the kernel density estimation function of the original photovoltaic power generation data sequence x(t), K is the kernel function, p i (z) is the IMF i (t) component kernel density estimation function, the commonly used kernel density function is the Gaussian kernel function, that is hx and hi are bandwidths, which can be selected by the Silverman rule, σ is the sample standard deviation, and n is the length of the original photovoltaic power data sequence;

[0051] The calculation formula of KL divergence is:

[0052]

[0053] KL divergence D obtained by sampling KL The expression is:

[0054]

[0055] In formula (12), zmin and zmax are the signal ranges, Δz is the sampling interval, and the larger the divergence, the more unique information the IMF component contains.

[0056] Furthermore, step 2 specifically includes:

[0057] Calculate the random meteorological variables X=[x1,x2,…,x n ] and Y=[y1,y2,…,y n ], where n is the number of samples, and the mutual information MI is calculated based on X = [x1, x2, ..., x n ] and Y=[y1,y2,…,y n ] is the maximum information coefficient MIC. According to the maximum information coefficient MIC, the meteorological data with the greatest correlation with the indicator to be predicted is selected. If the MIC value between the two variables is closer to 1, the stronger the correlation is.

[0058] Furthermore, the ECA-TCN-BiLSTM combined prediction model constructed in step 3 includes an input layer, n ECA-TCN modules, n BiLSTM layers, and an output layer connected in sequence;

[0059] The input of the ECA-TCN-BiLSTM combined prediction model is the decomposed and filtered photovoltaic power generation data and selected meteorological data, and the output is the photovoltaic power prediction value;

[0060] The ECA-TCN module is a temporal convolutional network with an efficient channel attention mechanism. It includes an efficient channel attention mechanism layer and a temporal convolutional network. The input of the ECA-TCN module is the decomposed and filtered photovoltaic power generation data and selected meteorological data, and the output is the weighted fusion of the extracted attention weights and the original input data features.

[0061] The temporal convolutional network (TCN) consists of multiple TCN residual blocks. Each residual block includes a causal dilated convolution layer, a normalization layer, a ReLU activation function layer, and a Dropout layer. In the downsampling stage, an ECA module is integrated. The ECA module includes an input layer, a convolution layer with a convolution kernel of 1*1*1, a normalization layer, a secondary convolution layer with a convolution kernel of 1*1*1, and an output layer.

[0062] The BiLSTM layer is a bidirectional long short-term memory network, consisting of two LSTM networks, forward and reverse. The input of the BiLSTM layer is the attention weight extracted after weighted fusion of the output of the ECA-TCN module and the original input data features, and the output is the photovoltaic power prediction value. The LSTM network is a long short-term memory network, including a forget gate, an input gate, and an output gate. The forget gate is used to determine the information discarded from the cell state, the input gate is used to update the cell state, and the output gate is used to determine the information input from the cell state to the hidden state.

[0063] Furthermore, the acquisition of the final photovoltaic power prediction value in step 4 specifically includes:

[0064] The decomposed and filtered photovoltaic power generation data and selected meteorological data are input into the ECA-TCN-BiLSTM combined prediction model. The causal dilated convolution layer in the ECA-TCN module learns the input data features. Nonlinear factors are introduced through the normalization layer, ReLU activation function layer, and Dropout layer. The TCN network is regularized and then enters the extreme downsampling. The ECA module extracts the attention weight of the input data and performs a weighted fusion with the original features of the input data. The weighted fusion features are then input into the BiLSTM network.

[0065] The BiLSTM network captures the historical information of the input data through the forward LSTM network, obtains the future information of the input data through the backward LSTM network, and outputs the ultra-short-term photovoltaic power prediction value.

[0066] The beneficial effects of the present invention are:

[0067] 1. The present invention adopts an improved adaptive noise complete empirical mode decomposition with dynamic physical constraint entropy termination (DPEC-ICEEMDAN). This method establishes a dynamic noise injection mechanism driven by local variance, realizes adaptive adjustment of the noise ratio with the signal fluctuation characteristics, imposes physical constraints of irradiance-power conversion on high-frequency modal components, and integrates the physical laws of photovoltaic power generation into the signal decomposition process instead of pure mathematical decomposition. It can suppress modal aliasing caused by temperature mutations and proposes a decomposition termination criterion based on energy proportion, overcoming the defect of traditional methods that rely on manual experience to set the number of decomposition layers.

[0068] 2. This paper combines the DPEC-ICEEMDAN algorithm with KL divergence to remove high-frequency noise from the original data and improve data quality.

[0069] 3. This paper innovatively designs a joint model that integrates an efficient channel attention mechanism (ECA), a temporal convolutional network (TCN), and a bidirectional long short-term memory network (BiLSTM). This model uses the attention mechanism to enhance key features, the TCN to capture local mutations, and the BiLSTM to model global trends. This significantly improves the accuracy and stability of photovoltaic power generation predictions while significantly reducing model training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A schematic flow chart of an ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data provided by the present invention;

[0071] Figure 2 Schematic diagram of the dilated convolutional layer of the TCN network provided by the present invention;

[0072] Figure 3 Schematic diagram of the TVN network residual block provided by the present invention;

[0073] Figure 4 A schematic diagram of the structure of the efficient channel attention mechanism module provided by the present invention;

[0074] Figure 5 Schematic diagram of the cell structure of the traditional LSTM network provided by the present invention;

[0075] Figure 6 A diagram showing the structure of a bidirectional long short-term memory neural network provided by the present invention;

[0076] Figure 7 This is a framework diagram of the ECA-TCN-BiLSTM prediction model provided by the present invention. DETAILED DESCRIPTION

[0077] Combine Figure 1-7 This embodiment is described as follows. Figure 1 As shown, the steps of the ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data described in this embodiment include:

[0078] S1: Get photovoltaic power sequence;

[0079] In this embodiment, the photovoltaic power sequence obtained includes photovoltaic power generation data and random meteorological variable data, wherein the original photovoltaic power sequence is x(t), and the random meteorological variable data includes X=[x1, x2,…, x n ] and Y=[y1,y2,…,y n ]wait.

[0080] S2: data preprocessing;

[0081] Data processing is mainly divided into feature selection and photovoltaic power generation series decomposition. Photovoltaic power generation output is affected by a variety of environmental factors. In photovoltaic power generation prediction, considering too many non-critical influencing factors will increase model training time, thereby affecting the convergence speed and prediction accuracy. In order to reduce the impact of redundant data on the prediction model, it is necessary to perform correlation analysis on the data and screen out the main factors affecting photovoltaic power to improve the calculation speed. In view of the intermittent phenomenon of photovoltaic power curves and the strong randomness and nonlinearity of photovoltaic data, this embodiment uses the DPEC-ICEEMDAN algorithm to decompose photovoltaic power into several intrinsic mode components (IMF) with different time characteristic scales. It can effectively utilize the local characteristics of photovoltaic power generation and improve the performance of the prediction model.

[0082] S201: DPEC-ICEEMDAN algorithm

[0083] To address the nonlinearity and nonstationarity of photovoltaic power generation, the photovoltaic power generation series can be decomposed into multiple components with different frequencies. The KL divergence is then used to comprehensively measure the difference between the IMF and the reference signal through probability distribution, thereby removing noise and capturing more complex statistical features (such as skewness and kurtosis). This method can better extract signal features, reduce data volatility, and improve prediction accuracy. The existing adaptive complete ensemble empirical mode decomposition (CEEMDAN) overcomes the modal aliasing problem inherent in empirical mode decomposition (EMD). However, CEEMDAN is prone to generating residual noise and pseudo-modes when decomposing the signal. Therefore, Colomina et al. proposed the improved adaptive complete ensemble empirical mode decomposition (ICEEMDAN). To address the three major drawbacks of the traditional ICEEMDAN in photovoltaic power forecasting: fixed noise ratios and a single noise coefficient that is difficult to adapt to different weather conditions; high-frequency modal aliasing, where rapid cloud movement causes overlap of high-frequency IMFs; and subjective decomposition termination, relying on manual experience to determine the number of decomposition levels. The DPEC-ICEEMDAN algorithm was proposed.

[0084] The main decomposition steps of DPEC-ICEEMDAN are as follows:

[0085] S20101: Multi-scale decomposition of photovoltaic power data;

[0086] Assume that the original photovoltaic power sequence is x(t), and introduce a dynamic noise adjustment mechanism to quantify the fluctuation intensity of the signal near the current moment. The variance is small on sunny days (power is stable), and large on rainy days (power changes suddenly). According to the local variance of the signal σ 2 (t) Real-time adjustment of noise ratio θ k ,formula:

[0087]

[0088] In formulas (1)-(3): is the power mean in the sliding window, W is the dynamic noise adjustment window length, Var(x(t)) is the variance of the time series x(t), is the mean of the sequence, that is T is the total time step. Sunny days (low variance) reduce noise injection and prevent excessive decomposition of high-frequency components. Heavy rain (high variance) increases noise and enhances modal separation capabilities.

[0089] S20102: Improve the modal decomposition process. Construct decomposition sequence (kth order decomposition):

[0090]

[0091] In formula (4), is the white noise added to the kth decomposition, Ek (·) indicates that the k-th order component is obtained after pre-decomposition of the noise.

[0092] S20103: Calculate the k-th order residual r k (t):

[0093]

[0094] In formula (5), M(·) is the local mean operation, and N is the number of noise additions.

[0095] S20104: Extract the k-th order IMF component:

[0096] IMF k (t) = r k-1 (t)-r k (t) (6);

[0097] S20105: Apply irradiance-to-power conversion constraints to high-frequency IMFs (only for k ≤ 2) to eliminate modal aliasing caused by temperature fluctuations:

[0098]

[0099] In formula (7), G(t) is the real-time irradiance, G STC is the standard irradiance.

[0100] S20106: Added energy entropy termination criteria to stop decomposition when the following conditions are met:

[0101]

[0102] In formula (8), η k is the energy ratio, ||IMF k ||2 is the L2 norm (energy) of the k-th order IMF, 5% is determined by Monte Carlo experiments, and 95% of the effective energy components are retained.

[0103] In summary, the present invention improves the adaptive noise complete empirical mode decomposition with dynamic physical constraint entropy termination, establishes a local variance-driven dynamic noise injection mechanism, realizes adaptive adjustment of the noise ratio with the signal fluctuation characteristics, imposes irradiance-power conversion physical constraints on high-frequency modal components, and integrates the physical laws of photovoltaic power generation into the signal decomposition process rather than pure mathematical decomposition. It can suppress modal aliasing caused by temperature mutations, proposes a decomposition termination criterion based on energy proportion, and overcomes the defect of traditional methods that rely on manual experience to set the number of decomposition layers.

[0104] S202: Photovoltaic power generation data screening;

[0105] This implementation uses DPEC-ICEEMDAN to decompose the original signal, generating a series of IMFs and a residual term. DPEC-ICEEMDAN reduces modal aliasing and improves decomposition stability through adaptive noise addition and iterative optimization. A KL divergence filtering mechanism is introduced to automatically remove high-frequency noise components with low correlation with the original signal while retaining valid IMF components with high correlation, enabling more refined feature extraction.

[0106] S20201: First, perform kernel density estimation (KDE) on the original signal x(t) and each IMF component IMFi(t) to obtain their probability density functions (PDFs):

[0107]

[0108] In formulas (9) and (10), p x (z) is the kernel density estimation function of the original photovoltaic power generation data sequence x(t), K is the kernel function, p i (z) is the IMF i (t) component kernel density estimation function, the commonly used kernel density function is the Gaussian kernel function, that is hx and hi are bandwidths, which can be selected by the Silverman rule, σ is the sample standard deviation, and n is the length of the original photovoltaic power data sequence;

[0109] S20202: Then, for each IMF component, calculate its KL divergence with the original signal:

[0110]

[0111] Discrete calculation: uniformly sample z in the interval [zmin,zmax] j , set the sampling step Δz, which is approximately:

[0112]

[0113] In formula (12), zmin and zmax are the signal ranges, Δz is the sampling interval, and the larger the divergence, the more unique information the IMF component contains.

[0114] S20203: Finally, filter out the effective IMF and sort the IMF according to D KL Sort from small to large and select the one that satisfies D KL <∈component.

[0115] S203: Correlation analysis;

[0116] When there is a large correlation between the factors affecting photovoltaic power generation, using all the influencing factors as model input will introduce redundant information, affecting the model prediction accuracy and efficiency. Therefore, it is necessary to perform correlation analysis on the original data features. The present invention selects the MIC method to perform correlation analysis. MIC overcomes the disadvantage that mutual information is inconvenient to calculate continuous variables and has higher accuracy. When there are enough samples, MIC can detect a wide range of linear and nonlinear relationships between variables, and can better reflect the degree of correlation between factors related to photovoltaic power generation.

[0117] The maximum information coefficient (MIC) is a correlation analysis algorithm based on mutual information. It uses a grid partitioning method to measure the degree of correlation between two variables, the strength of linear or nonlinearity, and is often used for feature selection. It has good universality, fairness, and symmetry. Let X = [x1, x2, ..., x n ] and Y=[y1,y2,…,y n ] are random variables in the data set, n is the number of samples, then the MI between X and Y is:

[0118]

[0119] In formula (13), p(x,y) is the joint probability density between X and Y; p(x) and p(y) are the marginal probability densities between X and Y, respectively.

[0120] MIC overcomes the difficulty of MI in calculating the joint probability density function of continuous variables and can find the correlation between two variables to the greatest extent. The calculation formula of MIC is:

[0121]

[0122] In formula (14), B is the number of samples; N is the sample variable; and I(x,y) is the MI between x and y. The closer the MIC value between two variables is to 1, the stronger their correlation.

[0123] S3: Establish an ECA-TCN-BiLSTM prediction model;

[0124] like Figure 7 As shown, the ECA-TCN-BiLSTM prediction model established in this embodiment includes an input layer, n ECA-TCN modules, n BiLSTM layers and an output layer connected in sequence. The input of the ECA-TCN-BiLSTM combined prediction model is the decomposed and filtered photovoltaic power generation data and the selected meteorological data, and the output is the photovoltaic power prediction value;

[0125] The ECA-TCN module is a time convolution network with an efficient channel attention mechanism, including an efficient channel attention mechanism layer and a time convolution network. The input of the ECA-TCN module is the photovoltaic power generation data and the selected meteorological data after decomposition and screening, and the output is the extracted attention weight after weighted fusion and the original input data features.

[0126] The time convolution network is an improved convolutional neural network that combines dilated convolution and residual block connection. The difference between dilated convolution and standard convolution is that the former can perform interval sampling on the input data during convolution operation, which is controlled by the Figure 2 sampling rate d. The residual connection structure can avoid the problem of gradient disappearance, and is a new type of network structure. In the model training process, the data of the output layer at time t only depends on the values of the network hidden layer at the same time and the historical values. When the actual application needs to capture the long-term dependence of the sequence data, increasing the input variables and the number of linear stacked layers may cause the problem of gradient disappearance. To solve this problem, the dilated convolution is used for interval sampling in the present embodiment, so that the effective sampling window expands exponentially with the increase of the number of layers, thereby effectively covering the entire time sequence. The TCN neural network is stacked by multiple TCN residual blocks, each of which combines a causal dilated convolution and a neural network processing layer, as shown in Figure 2 . In this process, the dilated coefficient d is used to adjust the sampling rate, and the value of d also increases with the increase of the level, and the specific expression of d is:

[0127] d=2 n-1 (15);

[0128] In formula (15), n is the number of dilated convolution layers.

[0129] Structure of TCN network Figure 3 The TCN neural network is stacked by multiple TCN residual blocks, each of which combines a causal dilated convolution and a neural network processing layer, as shown in Figure 3 . Therefore, when processing the time series data of photovoltaic power generation, the TCN network fully utilizes the advantages of the two types of convolutional neural networks, and introduces non-linear processing through the activation function, and the Dropout module helps to alleviate the problem of gradient disappearance caused by the complexity of the structure.

[0130] In recent years, the method of integrating channel attention mechanism into convolution block to enhance the performance of convolution network has attracted widespread attention. Among them, SENet (squeeze-and-inspire networks) achieves good use. When SENet needs to capture the dependency relationship of all channels, in order to avoid the burden of calculation, the data needs to be compressed and dimensionally processed, but the model accuracy will be affected to a certain extent. In view of the above problems, some researches propose an efficient channel attention (ECA) module specially for convolutional neural network, which can effectively capture local cross-channel interaction information and allocate more operation resources to important input information. Therefore, the channel attention mechanism is introduced in the present application to achieve better prediction performance.

[0131] The channel attention mechanism structure with global features is as shown in Figure 4 , wherein X is an input variable, B is a batch size parameter, H is a time window length, and W is an attribute number. The ECA module operating mechanism is as follows: a three-dimensional input X with a size of BxHxW is processed by a unit convolution kernel, and is converted into a three-dimensional matrix X with a size of 1xHxW new ; after convolution operation, the sequence feature is obtained, which is converted into a two-dimensional matrix X1 with a row number of 1 and a column number of HxW (1xHW); X1 is normalized by ReLU function to form a two-dimensional correlation matrix X2 with a size of 1xHW; X1 and X2 corresponding elements are added to constitute a global feature matrix X3; X3 is processed by a second convolution operation to generate a three-dimensional global feature X4 with a size of 1xHxW; X4 is processed by a unit convolution block and added to the corresponding elements in X to form a feature matrix Y with attention mechanism.

[0132] Long short-term memory network (LSTM) is a variant of recurrent neural network (RNN), which aims to solve the problem of gradient disappearance or explosion in RNN. The structure of LSTM model is as shown in Figure 5 , which includes a forget gate, an input gate and an output gate. The forget gate decides which information to discard from the cell state, the input gate is used to update the cell state, and the output gate decides which information to output from the cell state to the hidden state. The calculation process of LSTM is as follows:

[0133]

[0134] In formula (16), x t is the current input; h t-1 , h t are the hidden states at the previous moment and the current moment; C t-1 , C t are the cell states at the previous moment and the current moment; W f , W i , W c , W oare the weight matrices of the forget gate, input gate, cell state, and output gate respectively; b f 、b i 、b c 、b o is the bias corresponding to the above modules; σ is the sigmoid activation function.

[0135] In the prediction of photovoltaic power, fully considering the positive and negative information patterns of photovoltaic related time series data can effectively improve the prediction accuracy. Although LSTM effectively solves the long-term dependency problem of RNN, when using LSTM for photovoltaic power prediction, it can only use historical data to analyze the relationship between various factors, without considering future information. The bidirectional long short-term memory neural network (BiLSTM) consists of two LSTMs, forward and reverse. The forward LSTM can capture the historical information of the input data, and the reverse LSTM can obtain the future information of the input data. Its structure is as follows Figure 6 Compared to traditional unidirectional LSTMs, BiLSTMs overcome the drawback of insufficient data mining information. Using BiLSTMs to perform ultra-short-term PV power forecasts, combined with historical PV data and future weather forecasts, can improve prediction accuracy.

[0136] This paper innovatively designs a joint model that integrates an efficient channel attention mechanism, a temporal convolutional network, and a bidirectional long short-term memory network. This model uses the attention mechanism to enhance key features, employs a TCN to capture local mutations, and employs a BiLSTM to model global trends. This significantly improves the accuracy and stability of photovoltaic power prediction while significantly reducing model training time.

[0137] S4: The decomposed and filtered data and the feature-selected meteorological data are input into the combined prediction model for prediction, and the prediction results of each IMF component are superimposed and integrated to obtain the final photovoltaic power prediction value.

[0138] The construction of the ECA-TCN model is based on the stacking of multiple ECA-TCN residual modules. In the basic TCN residual module, the TCN convolution layer learns the input features, and introduces nonlinear factors and regularizes the network through the WeightNorm, ReLU and Dropout layers. The improved TCN residual block integrates the ECA module in the downsampling stage, and finally multiplies the attention weights extracted by the ECA module with the weighted fusion of the original features to achieve the output. The output of the ECA-TCN module serves as the input of the BiLSTM model, because the TCN module is good at capturing local mutations, while the BiLSTM module is good at modeling global trends. The parallel computing capability of TCN and the sequence modeling capability of LSTM balance the prediction accuracy and inference speed. The specific structure is as follows: Figure 7 shown.

[0139] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for ultra-short-term photovoltaic power combination prediction based on multivariate meteorological data, characterized in that: include: Step 1: Obtain a photovoltaic power generation sequence, and perform multi-scale decomposition and KL divergence screening on the photovoltaic power generation data sequence in the photovoltaic power generation sequence; Step 2: Perform correlation analysis on the random meteorological variables in the photovoltaic power generation sequence and the indicators to be predicted, and select the meteorological data with the greatest correlation with the indicators to be predicted based on the correlation analysis results; Step 3: Build the ECA-TCN-BiLSTM combined prediction model; Step 4: Input the decomposed and filtered photovoltaic power generation data and the selected meteorological data into the ECA-TCN-BiLSTM combined prediction model, superimpose and integrate the prediction results of each IMF component to obtain the final ultra-short-term photovoltaic power prediction value.

2. The ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data according to claim 1 is characterized in that: Step 1 specifically includes: Step 1.1: Acquire a photovoltaic power generation sequence, wherein the photovoltaic power generation sequence includes photovoltaic power generation data and random meteorological variable data; Step 1.2: Use the DPEC-ICEEMDAN algorithm to perform multi-scale decomposition on the photovoltaic power generation data to obtain a series of IMF components and a residual term, wherein the DPEC-ICEEMDAN algorithm is an improved adaptive noise complete empirical mode decomposition method based on dynamic physical constraint entropy termination; Step 1.3: Perform kernel density estimation and discretization calculation on the photovoltaic power generation data and each IMF component, and filter out the KL divergence D of the photovoltaic power generation data corresponding to each IMF component. KL , for each IMF component according to its sampled KL divergence D KL Sort by size and select the one that satisfies D KL <∈ IMF component, output corresponding decomposed and filtered photovoltaic power data.

3. The ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data according to claim 2 is characterized in that: Step 1.2 specifically includes: Step 1.2.1: Set the original photovoltaic power data sequence as x(t), introduce a dynamic noise adjustment mechanism, and calculate the signal local variance σ 2 (t), according to the local variance σ of the signal 2 (t) Real-time adjustment of noise ratio θ k ; Step 1.2.2: According to the noise ratio θ k Constructing a decomposition sequence According to the decomposition sequence Calculate the k-th order residual r of the photovoltaic power generation data sequence x(t) k (t); Step 1.2.3: According to the k-th order residual r of the photovoltaic power generation data sequence x(t) k (t) and the k-1th order residual r k-1 (t) extracting the k-th order IMF component; Step 1.2.4: When k ≤ 2, apply irradiance-to-power conversion constraints based on the extracted k-th-order IMF components to eliminate modal aliasing caused by temperature fluctuations. Step 1.2.5: Obtain the energy fraction η based on the L2 norm of the k-th order IMF component k Repeat steps 1.2.1 to 1.2.4 for iterative decomposition until the energy proportion after current decomposition is η k <5%, a series of IMF components and a residual term are obtained; Local variance σ 2 The calculation formula for (t) is: Noise ratio θ k The calculation formula is: In formulas (1) and (2), is the power mean in the sliding window, W is the dynamic noise adjustment window length, Var(x(t)) is the variance of the time series x(t), is the mean of the sequence, that is T is the total time step; The calculation formula for the variance Var(x(t)) of the time series x(t) is: decomposition sequence The calculation formula is: In formula (4), is the white noise added to the kth decomposition, E k (·) is the k-th order component after pre-decomposition of the noise; The k-th order residual r of the photovoltaic power generation data sequence x(t) k The calculation formula for (t) is: In formula (5), M(·) is the local mean operation, and N is the number of noise additions; The calculation formula of the k-th order IMF component is: IMF k (t)=r k-1 (t)-r k (t) (6); The expression of the k-th order IMF component after applying the irradiance-power conversion constraint is: In formula (7), G(t) is the real-time irradiance, G STC is the standard irradiance; Energy proportion η k The calculation formula is: In formula (8), η k is the energy ratio, ||IMF k ||2 is the L2 norm of the k-th order IMF, 5% is determined by Monte Carlo experiments, and 95% of the effective energy components are retained.

4. The method for ultra-short-term photovoltaic power combination prediction based on multivariate meteorological data according to claim 2, characterized in that: Step 1.3 specifically includes: Step 1.3.1: Perform kernel density estimation on the original photovoltaic power data sequence x(t) and each of its corresponding IMF components to obtain their probability density functions; Step 1.3.2: For each IMF component, calculate the KL divergence between it and the original PV power data; Step 1.3.3: Sample z uniformly in the interval [zmin,zmax] j , set the sampling step Δz, and obtain the KL divergence D of the photovoltaic power data corresponding to each IMF component KL ; Step 1.3.4: For each IMF component, calculate the KL divergence D obtained by sampling. KL Sort by size and select the one that satisfies D KL <∈IMF component, output the corresponding decomposed and filtered photovoltaic power data; The calculation formula of the probability density function is: In formulas (9) and (10), p x (z) is the kernel density estimation function of the original photovoltaic power generation data sequence x(t), K is the kernel function, p i (z) is the IMF i (t) component kernel density estimation function, the commonly used kernel density function is the Gaussian kernel function, that is hx and hi are bandwidths, which can be selected by the Silverman rule, σ is the sample standard deviation, and n is the length of the original photovoltaic power data sequence; The calculation formula of KL divergence is: KL divergence D obtained by sampling KL The expression is: In formula (12), zmin and zmax are the signal ranges, Δx is the sampling interval, and the larger the divergence, the more unique information the IMF component contains.

5. The ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data according to claim 1 is characterized in that: Step 2 specifically includes: Calculate the random meteorological variables X=[x1,x2,…,x n ] and Y=[y1,y2,…,y n ], where n is the number of samples, and the mutual information MI is calculated based on X=[x1,x2,…,x n ] and Y=[y1,y2,…,y n ] is the maximum information coefficient MIC. According to the maximum information coefficient MIC, the meteorological data with the greatest correlation with the indicator to be predicted is selected. If the MIC value between the two variables is closer to 1, the stronger the correlation is.

6. The method for ultra-short-term photovoltaic power combination prediction based on multivariate meteorological data according to claim 1, characterized in that: The ECA-TCN-BiLSTM combined prediction model constructed in step 3 includes an input layer, n ECA-TCN modules, n BiLSTM layers, and an output layer connected in sequence; The input of the ECA-TCN-BiLSTM combined prediction model is the decomposed and filtered photovoltaic power generation data and selected meteorological data, and the output is the photovoltaic power prediction value; The ECA-TCN module is a temporal convolutional network with an efficient channel attention mechanism. It includes an efficient channel attention mechanism layer and a temporal convolutional network. The input of the ECA-TCN module is the decomposed and filtered photovoltaic power generation data and selected meteorological data, and the output is the weighted fusion of the extracted attention weights and the original input data features. The temporal convolutional network (TCN) consists of multiple TCN residual blocks. Each residual block includes a causal dilated convolution layer, a normalization layer, a ReLU activation function layer, and a Dropout layer. In the downsampling stage, an ECA module is integrated. The ECA module includes an input layer, a convolution layer with a convolution kernel of 1*1*1, a normalization layer, a secondary convolution layer with a convolution kernel of 1*1*1, and an output layer. The BiLSTM layer is a bidirectional long short-term memory network, consisting of two LSTM networks, forward and reverse. The input of the BiLSTM layer is the attention weight extracted after weighted fusion of the output of the ECA-TCN module and the original input data features, and the output is the photovoltaic power prediction value. The LSTM network is a long short-term memory network, including a forget gate, an input gate, and an output gate. The forget gate is used to determine the information discarded from the cell state, the input gate is used to update the cell state, and the output gate is used to determine the information input from the cell state to the hidden state.

7. The method for ultra-short-term photovoltaic power combination prediction based on multivariate meteorological data according to claim 6, characterized in that: The final photovoltaic power prediction value in step 4 is obtained by: The decomposed and filtered photovoltaic power generation data and selected meteorological data are input into the ECA-TCN-BiLSTM combined prediction model. The causal dilated convolution layer in the ECA-TCN module learns the input data features. Nonlinear factors are introduced through the normalization layer, ReLU activation function layer, and Dropout layer. The TCN network is regularized and then enters the extreme downsampling. The ECA module extracts the attention weight of the input data and performs a weighted fusion with the original features of the input data. The weighted fusion features are then input into the BiLSTM network. The BiLSTM network captures historical information of input data through the forward LSTM network, obtains future information of input data through the backward LSTM network, and superimposes and integrates the prediction results of each IMF component to output an ultra-short-term photovoltaic power prediction value.

Images