A method for ultra-short-term photovoltaic power combined prediction based on multi-element meteorological data

Abstract

The application provides a kind of based on multivariate weather data's ultra-short-term photovoltaic power combination prediction method, belongs to photovoltaic power generation power prediction technical field, the present application is to solve photovoltaic power generation prediction, consider too many non-key influence factors will increase model training time, and then influence convergence speed and prediction accuracy, and the problem that the importance difference of time series feature is not captured enough, including: obtaining photovoltaic power generation power sequence, and carry out multiscale decomposition and K-L divergence screening;Random weather variable and the to-be-predicted index in photovoltaic power generation power sequence are analyzed in relevance, the weather data most relevant to the to-be-predicted index is selected according to the relevance analysis result;ECA-TCN-BiLSTM combination prediction model is constructed;The photovoltaic power generation power data after decomposition and screening and the selected weather data are input into the ECA-TCN-BiLSTM combination prediction model, the prediction results of each IMF component are superimposed and integrated, and the final ultra-short-term photovoltaic power prediction value is obtained.

Description

Technical Field

[0001] This invention relates to an ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data, belonging to the field of photovoltaic power generation prediction technology. Background Technology

[0002] Solar energy is the most abundant and promising clean energy source among renewable energy sources. Photovoltaic (PV) power generation, as a technology that directly converts solar energy into electricity, is gaining increasing attention due to its safety, efficiency, economic benefits, and environmental friendliness. In recent years, with the rapid development of clean and efficient energy generation technologies, the installed capacity of PV power generation has been continuously increasing. Compared with other renewable energy generation technologies, PV power generation technology is relatively mature and reliable, and is currently the most important form of new energy power generation technology. However, PV power generation is easily affected by seasonal and weather factors, resulting in significant randomness and fluctuations in power output. Connecting PV power generation devices to the power system presents challenges and impacts. Accurate prediction of PV power generation can not only help grid operators better dispatch power generation resources and optimize power system operation, but also reduce the power system's reliance on peak-shaving backup power sources, thereby reducing operating costs. Therefore, PV power generation prediction is a crucial support for the stable operation of modern power systems.

[0003] Based on time scale, photovoltaic (PV) power generation forecasting can be categorized into medium- and long-term forecasts, short-term forecasts, and ultra-short-term forecasts. Medium- and long-term forecasts form the basis for power system planning, while short- and ultra-short-term forecasts are fundamental for dispatching departments to arrange grid operation modes and formulate dispatching strategies. To obtain accurate PV power generation forecasts, researchers have proposed numerous forecasting methods in recent years, broadly classified into statistical methods, physical methods, and deep learning methods, as well as hybrid methods, with hybrid methods being the most prominent. Hybrid model forecasting based on signal decomposition, as a parallel and effective method within hybrid model forecasting, has attracted attention in various fields, including PV power generation forecasting, wind speed forecasting, wind power forecasting, lithium battery life forecasting, and financial asset return forecasting.

[0004] Photovoltaic power generation prediction mainly consists of two parts: data processing and model prediction. Predicting photovoltaic power generation is essentially the analysis and forecasting of non-stationary data sequences, and the accuracy of the model is usually closely related to the extraction and processing methods of time-series features. In data processing, feature construction aims to improve outlier 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 categories. For example, DTW measures the similarity of time-series patterns 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, pre-setting the number of clusters (e.g., K-means), parameter sensitivity (e.g., DBSCAN), and the emphasis on static distribution may weaken the expression of the time-series dependence characteristics of photovoltaic power, leading to biases in pattern extraction. Feature derivation constructs new features through statistical aggregation (such as sliding window mean), frequency domain transformation, or combination of domain knowledge (such as irradiance-temperature ratio). Studies have shown that such methods can deeply mine implicit information in data and improve prediction accuracy in complex scenarios. However, their design heavily relies on prior knowledge, automated derivation is prone to generating invalid features, and some black-box models, while exhibiting excellent performance, suffer from insufficient feature interpretability due to unclear mechanisms, thus limiting the reliability of practical applications. In feature dimensionality reduction, principal component analysis (PCA) preserves high-variance features based on orthogonal transformation, improving computational efficiency and reducing redundant interference. However, limited by linear assumptions, it struggles to capture temporal nonlinear relationships. Signal decomposition algorithms exhibit unique advantages due to their ability to decouple the essential characteristics of time-series signals, making them suitable for scenarios like photovoltaic power, which have strong temporal dependence and non-stationary fluctuation characteristics. For example, VMD, by constructing variational models to constrain the modal center frequencies, can effectively separate the fluctuation patterns of different frequency features; while adaptive noise complete set empirical mode decomposition (CEEMDAN) algorithms significantly improve the mode mixing problem by introducing white noise to assist analysis. However, existing signal decomposition algorithms still face several challenges: First, mode aliasing is difficult to completely avoid in complex fluctuating scenarios (such as sudden changes in photovoltaic power due to cloud cover), leading to cross-interference of signal components at different time scales; Second, the high dependence of algorithms on manually preset parameters (such as VMD requiring the number of decomposition layers and CEEMDAN relying on noise intensity parameters) reduces their adaptability and makes them prone to over-decomposition or under-decomposition due to incorrect parameter settings; Third, endpoint effects and noise sensitivity may introduce boundary distortion or residual noise, especially in the beginning and end segments of time-series signals or in high-noise environments, making the physical interpretability of the decomposition results questionable; Fourth, traditional methods (such as EMD) are still limited by fixed decomposition logic in their 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 mainly falls into three categories: physical models, statistical models, and hybrid models. While the first two types are simple to implement, they are highly dependent on data and have weak generalization ability. With the development of artificial intelligence, models based on machine learning and deep learning, through continuous iterative learning, can more deeply mine data features, thus exhibiting superior algorithmic performance. Some studies have used the XGBoost algorithm, which integrates multiple linear decision trees, to improve model hierarchy and prediction accuracy, but this method is insufficient in handling nonlinear relationships between variables. Another ensemble learning approach uses an architecture combining convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), successfully achieving real-time and accurate prediction of multi-mode photovoltaic power generation. However, LSTM models suffer from slow training speeds and difficulties in parameter tuning, while traditional CNNs, limited by fixed-size convolutional kernels, struggle to effectively capture long-term dependencies. For time-series data processing, temporal convolutional networks (TCNs) exhibit unique advantages; their parallel processing capabilities significantly improve computational efficiency, and their prediction accuracy is comparable to that of recurrent neural network (RNN) algorithms. However, although this network expands its receptive field by dilating convolutions, it is still constrained by the kernel size, limiting its effectiveness in handling long-range correlated features. Other studies have improved the residual structure by converting the serial pattern to a parallel distribution to enhance the model's universality, but such improvements have not yet solved the problem of representing the differences in the importance of different features in time series data. Summary of the Invention

[0006] To address the problem that considering too many non-critical influencing factors in photovoltaic power generation forecasting increases model training time, thereby affecting convergence speed and prediction accuracy, and that it fails to capture the differences in the importance of time series features, this invention proposes an ultra-short-term photovoltaic power combination forecasting method based on multivariate meteorological data.

[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The present invention includes 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;

[0009] Step 2: Conduct correlation analysis on the random meteorological variables and the indicators to be predicted in the photovoltaic power generation series, and select the meteorological data with the highest correlation to the indicators to be predicted based on the correlation analysis results;

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

[0011] Step 4: Input the decomposed and screened photovoltaic power generation data and the selected meteorological data into the ECA-TCN-BiLSTM combined prediction model, 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 the 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 photovoltaic power generation 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 calculations on the photovoltaic power generation data and each IMF component, and select the KL divergence D of the photovoltaic power generation data corresponding to each IMF component. KL For each IMF component, according to the KL divergence D obtained from its sampling KL Sort the values ​​by size and select those that satisfy D. KL The IMF components of <∈ are output, along with the corresponding decomposed and filtered photovoltaic power generation data.

[0016] Furthermore, step 1.2 specifically includes:

[0017] Step 1.2.1: Set the original photovoltaic power generation data sequence as x(t), introduce a dynamic noise adjustment mechanism, and calculate the signal local variance σ. 2 (t), based on the signal local variance σ 2 (t) Adjust the noise ratio k in real time;

[0018] Step 1.2.2: Based on the noise ratio θ k Constructing decomposed sequences Based on the decomposition sequence The k-th order residual r of the photovoltaic power generation data sequence x(t) is calculated. k (t);

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

[0020] Step 1.2.4: When k≤2, apply irradiance-power conversion constraints based on the extracted k-th order IMF component to eliminate mode mixing caused by temperature fluctuations;

[0021] Step 1.2.5: Obtain the energy proportion η based on the L2 norm of the k-th order IMF component. k Repeat steps 1.2.1-1.2.4 for iterative decomposition until the energy percentage η after the current decomposition is reached. k <5%, resulting in a series of IMF components and a residual term;

[0022] Local variance σ 2 The formula for calculating (t) is:

[0023]

[0024] Noise ratio θ k The calculation formula is:

[0025]

[0026] In formulas (1) and (2), Let be the power mean within the sliding window, W be the dynamic noise conditioning window length, and Var(x(t)) be the variance of the time series x(t). The mean of the sequence is... T represents the total time step;

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

[0028]

[0029] decomposition sequence The calculation formula is:

[0030]

[0031] In formula (4), The white noise added for the k-th decomposition, E k (·) represents the kth order component after pre-decomposition of the noise;

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

[0033]

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

[0035] The formula for calculating the k-th order IMF component is:

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

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

[0038]

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

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

[0041]

[0042] In formula (8), η k For energy percentage, ||IMF k ||2 is the L2 norm of the k-th order IMF, 5% of which was determined by Monte Carlo experiments, retaining 95% of the effective energy component.

[0043] Furthermore, step 1.3 specifically includes:

[0044] Step 1.3.1: Perform kernel density estimation on the original photovoltaic power generation data sequence x(t) and its corresponding IMF component to obtain its probability density function;

[0045] Step 1.3.2: For each IMF component, calculate its KL divergence with the original photovoltaic power generation data;

[0046] Step 1.3.3: Uniformly sample z within the interval [zmin, zmax]. j By setting the sampling step size Δz, the KL divergence D of the photovoltaic power generation data corresponding to each IMF component is obtained. KL ;

[0047] Step 1.3.4: For each IMF component, calculate the KL divergence D obtained from its sampling. KL Sort the values ​​by size and select those that satisfy D. KL For the IMF components of <∈, output the corresponding decomposed and filtered photovoltaic power generation data;

[0048] The formula for calculating 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), where K is the kernel function and p i (z) represents the IMF i The kernel density estimation function for the (t) component is commonly the Gaussian kernel function, i.e. hx and hi are the bandwidths, which can be selected using the Silverman rule; σ is the sample standard deviation; and n is the length of the original photovoltaic power generation data sequence.

[0051] The formula for calculating the KL divergence is:

[0052]

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

[0054]

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

[0056] Furthermore, step 2 specifically includes:

[0057] Calculate the random meteorological variable X = [x1, x2, ..., x] in the photovoltaic power generation sequence. n ] and Y = [y1, y2, ..., y n The mutual information MI of [x1, x2, ..., xn] is calculated, where n is the number of samples. n ] and Y = [y1, y2, ..., y n The maximum information coefficient (MIC) is used to select meteorological data that are most correlated with the indicator to be predicted. The closer the MIC values ​​of two variables are to 1, the stronger their correlation.

[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 to 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.

[0060] The ECA-TCN module is a temporal convolutional network with an added efficient channel attention mechanism, including 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 the selected meteorological data, and the output is the extracted attention weights after weighted fusion and the features of the original input data.

[0061] The Temporal Convolutional Network (TCN) consists of multiple TCN residual blocks. Each residual block includes a causal dilated convolutional layer, a normalization layer, a ReLU activation function layer, and a Dropout layer. An ECA module is integrated during the downsampling stage. The ECA module includes an input layer, a convolutional layer with a 1*1*1 kernel, a normalization layer, a quadratic convolutional layer with a 1*1*1 kernel, and an output layer.

[0062] The BiLSTM layer is a bidirectional long short-term memory network, consisting of two LSTM networks, one forward and one backward. The input of the BiLSTM layer is the extracted attention weights extracted after weighted fusion of the output of the ECA-TCN module and the original input data features. 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 specific steps in step 4 for obtaining the final photovoltaic power prediction include:

[0064] The decomposed and screened photovoltaic power generation data and selected meteorological data are input into the ECA-TCN-BiLSTM combined prediction model. The causal dilated convolutional layer in the ECA-TCN module learns the features of the input data. Nonlinear factors are introduced through normalization layer, ReLU activation function layer and Dropout layer. After regularization of the TCN network, it enters the downsampling extreme. The attention weights of the input data are extracted by the ECA module and weighted and fused with the original features of the input data. The weighted and fused features are then input into the BiLSTM network.

[0065] BiLSTM networks capture historical information from input data through a forward LSTM network and obtain future information from input data through a reverse LSTM network, outputting ultra-short-term photovoltaic power prediction values.

[0066] The beneficial effects of this invention are:

[0067] 1. This invention utilizes an improved adaptive noise-complete empirical mode decomposition (DPEC-ICEEMDAN) with dynamic physical constraint entropy termination. This method establishes a local variance-driven dynamic noise injection mechanism, enabling the noise ratio to adaptively adjust with signal fluctuation characteristics. It applies irradiance-power conversion physical constraints to high-frequency mode components, integrating the physical laws of photovoltaic power generation into the signal decomposition process rather than purely mathematical decomposition. This can suppress mode aliasing caused by temperature abrupt changes. Furthermore, it proposes a decomposition termination criterion based on energy ratio, overcoming the shortcomings of traditional methods that rely on manual experience to set the number of decomposition layers.

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

[0069] 3. This invention innovatively designs a joint model that integrates Efficient Channel Attention (ECA), Temporal Convolutional Network (TCN), and Bidirectional Long Short-Term Memory (BiLSTM). This model strengthens key features through the attention mechanism, captures local mutations using TCN, and models global trends using BiLSTM. This significantly improves the accuracy and stability of photovoltaic power generation prediction while drastically reducing model training time. Attached Figure Description

[0070] Figure 1 A flowchart illustrating an ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data provided by this invention;

[0071] Figure 2 This is a schematic diagram of the dilated convolutional layer of the TCN network provided by the present invention;

[0072] Figure 3 This is a schematic diagram of the TVN network residual block provided by the present invention;

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

[0074] Figure 5 A schematic diagram of the cell structure of a conventional LSTM network provided by this invention;

[0075] Figure 6 The bidirectional long short-term memory neural network structure diagram provided by this invention;

[0076] Figure 7 The framework diagram of the ECA-TCN-BiLSTM prediction model provided by this invention. Detailed Implementation

[0077] Combination Figure 1-7 This implementation method 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: Obtain the photovoltaic power sequence;

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

[0080] S2: Data preprocessing;

[0081] Data processing mainly consists of feature selection and photovoltaic power generation sequence decomposition. Photovoltaic power output is affected by various environmental factors. In photovoltaic power generation prediction, considering too many non-critical influencing factors increases model training time, thereby affecting convergence speed and prediction accuracy. To reduce the impact of redundant data on the prediction model, it is necessary to perform correlation analysis on the data to screen out the main factors affecting photovoltaic power and improve calculation speed. Considering the intermittent phenomenon of photovoltaic power curves and the strong randomness and nonlinearity of photovoltaic data, this implementation uses the DPEC-ICEEMDAN algorithm to decompose photovoltaic power into several intrinsic mode functions (IMFs) with different time characteristic scales. This effectively utilizes the local characteristics of photovoltaic power generation and improves the performance of the prediction model.

[0082] S201: DPEC-ICEEMDAN Algorithm

[0083] To address the nonlinearity and non-stationarity of photovoltaic (PV) power generation, the PV power sequence can be decomposed into multiple components with different frequencies. KL divergence is then used to comprehensively measure the difference between the IMF (Integrated Mode Factor) and the reference signal through probability distribution, 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. Existing adaptive noisy complete ensemble empirical mode decomposition (CEEMDAN) overcomes the mode aliasing phenomenon of empirical mode decomposition (EMD), but CEEMDAN is prone to residual noise and spurious modes during signal decomposition. Therefore, Colominas et al. proposed an improved adaptive noisy complete ensemble empirical mode decomposition (ICEEMDAN). Addressing the three major pain points of traditional ICEEMDAN in PV power prediction: fixed noise ratio, making it difficult for a single noise coefficient to adapt to different weather types; high-frequency mode aliasing, with rapid cloud movement causing overlap in high-frequency IMFs; and subjective decomposition termination, relying on manual experience to set the number of decomposition layers, the DPEC-ICEEMDAN algorithm is proposed.

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

[0085] S20101: Multiscale decomposition of photovoltaic power generation data;

[0086] Let the original photovoltaic power sequence be x(t). A dynamic noise adjustment mechanism is introduced to quantify the signal fluctuation intensity near the current moment. The variance is small (power is stable) during sunny weather and large (power abruptly changes) during heavy rain. Based on the signal's local variance σ... 2 (t) Real-time adjustment of noise ratio θ k ,formula:

[0087]

[0088] In formulas (1)-(3): Let be the power mean within the sliding window, W be the dynamic noise conditioning window length, and Var(x(t)) be the variance of the time series x(t). The mean of the sequence is... T represents 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 but enhances mode separation capabilities.

[0089] S20102: Improved mode decomposition process. Construct decomposition sequences. (k-th order decomposition):

[0090]

[0091] In formula (4), The white noise added for the k-th decomposition, Ek (·) indicates that the k-th order component is taken after the noise is pre-decomposed.

[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: Extracting the k-th order IMF component:

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

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

[0098]

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

[0100] S20106: Add an energy entropy termination criterion to stop the decomposition when the following conditions are met:

[0101]

[0102] In formula (8), η k For energy percentage, ||IMF k ||2 is the L2 norm (energy) of the k-th order IMF, 5% of which was determined by the Monte Carlo experiment, retaining 95% of the effective energy component.

[0103] In summary, this invention improves the adaptive noise-complete empirical mode decomposition by using dynamic physical constraint entropy termination, establishes a local variance-driven dynamic noise injection mechanism, and achieves adaptive adjustment of the noise ratio according to signal fluctuation characteristics. It applies irradiance-power conversion physical constraints to high-frequency mode components, integrates the physical laws of photovoltaic power generation into the signal decomposition process rather than purely mathematical decomposition, and can suppress mode aliasing caused by temperature abrupt changes. It also proposes a decomposition termination criterion based on energy ratio, overcoming the shortcomings of traditional methods that rely on manual experience to set the number of decomposition layers.

[0104] S202: Filtering photovoltaic power generation data;

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

[0106] S20201: First, kernel density estimation (KDE) is performed on the original signal x(t) and each IMF component IMFi(t) to obtain its probability density function (PDF):

[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), where K is the kernel function and p i (z) represents the IMF i The kernel density estimation function for the (t) component is commonly the Gaussian kernel function, i.e. hx and hi are the bandwidths, which can be selected using the Silverman rule; σ is the sample standard deviation; and n is the length of the original photovoltaic power generation data sequence.

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

[0110]

[0111] Discretization calculation: uniformly sample z within the interval [zmin, zmax]. j Set the sampling step size Δz, which is approximately:

[0112]

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

[0114] S20203: Final screening of valid IMFs, categorizing IMFs according to D KL Sort by size from smallest to largest, and select the one that satisfies D. KL The components of <∈.

[0115] S203: Correlation analysis;

[0116] When the factors affecting photovoltaic power generation are highly correlated, using all influencing factors as model input introduces redundant information, affecting the model's prediction accuracy and efficiency. Therefore, correlation analysis of the original data features is necessary. This invention selects the MIC method for correlation analysis. MIC overcomes the inconvenience of calculating mutual information for continuous variables, possesses higher accuracy, and with sufficient samples, can detect a wide range of linear and nonlinear relationships between variables, better reflecting the degree of correlation between factors related to photovoltaic power generation.

[0117] The maximal information coefficient (MIC) is a correlation analysis algorithm based on mutual information. It uses a grid partitioning method to measure the strength of the correlation between two variables, whether linear or nonlinear. It is commonly used for feature selection and has good universality, fairness, and symmetry. Let X = [x1, x2, ..., x...]. n ] and Y = [y1, y2, ..., y n Let X and Y be random variables in the dataset, and n be 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 formula for calculating MIC is:

[0121]

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

[0123] S3: Establish the 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 temporal convolutional network with an added efficient channel attention mechanism, including 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 the selected meteorological data, and the output is the extracted attention weights after weighted fusion and the features of the original input data.

[0126] Temporal convolutional networks are an improved convolutional neural network that combines dilated convolution and residual block connections. The difference between dilated convolution and standard convolution is that the former can sample the input data at intervals during the convolution operation, thus being more sensitive to... Figure 2 The sampling rate d is controlled; the residual connection structure avoids the gradient vanishing problem and is a novel network structure. During model training, the output layer's data at time t depends only on the values ​​of the hidden layers at the same time and historical values. When practical applications require capturing long-term dependencies in sequence data, increasing the number of input variables and linearly stacked layers may cause the gradient vanishing problem. To solve this problem, this implementation uses dilated convolution for interval sampling, so that the effective sampling window expands exponentially with the increase of the number of layers, thereby effectively covering the entire time series. The TCN neural network is composed of multiple stacked TCN residual blocks, each residual block combining causal dilated convolution and neural network processing layers, such as... Figure 2 As shown, in this process, the expansion coefficient d is used to adjust the sampling rate. As the layer increases, the value of d also increases. The specific expression for 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 As shown, the TCN neural network is composed of multiple stacked TCN residual blocks. Each residual block combines causal dilated convolution and neural network processing layers, such as... Figure 3 As shown. Therefore, when processing time-series data of photovoltaic power generation, the TCN network fully leverages the advantages of both convolutional neural networks, introduces nonlinear processing through activation functions, and the Dropout module helps alleviate the gradient vanishing problem caused by structural complexity.

[0130] In recent years, incorporating channel attention mechanisms into convolutional blocks to enhance the performance of convolutional networks has received widespread attention. SENet (squeeze-and-inspire networks) has achieved good results in this regard. However, when SENet needs to capture the dependencies of all channels, data compression and dimensionality reduction are necessary to avoid computational burden, which can negatively impact model accuracy. To address these issues, some researchers have proposed an efficient channel attention (ECA) module specifically for convolutional neural networks, which can effectively capture local cross-channel interaction information and allocate more computational resources to important input information. Therefore, this invention introduces a channel attention mechanism to achieve better prediction performance.

[0131] Channel attention mechanism structure with global features, such as Figure 4 As shown, X is the input variable, B is the batch size, H is the time window length, and W is the number of attributes. The ECA module operates as follows: a three-dimensional input X of size B×H×W is transformed into a three-dimensional matrix X of size 1×H×W after being processed by a unit convolution kernel. new After convolution, sequence features are obtained and transformed into a (1×HW) two-dimensional matrix X1 with 1 row and H×W columns. After normalization by the ReLU function, X1 forms a two-dimensional correlation matrix X2 of size 1×HW. The corresponding elements of X1 and X2 are added to form a global feature matrix X3. After a second convolution operation, X3 generates a three-dimensional global feature matrix X4 of size 1×H×W. After processing by a unit convolution block, X4 is added to the corresponding elements of X to form a feature matrix Y with an attention mechanism.

[0132] Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs) designed to solve the vanishing or exploding gradient problem in RNNs. The LSTM model structure is as follows: Figure 5 As shown, an LSTM unit includes a forget gate, an input gate, and an output gate. The forget gate determines which information to discard from the cell state, the input gate updates the cell state, and the output gate determines which information to output from the cell state to the hidden state. The computation process of LSTM is as follows:

[0133]

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

[0135] In photovoltaic (PV) power prediction, fully considering the positive and negative information patterns in PV-related time-series data can effectively improve prediction accuracy. While LSTM effectively solves the long-term dependency problem of RNNs, when using LSTM for PV power prediction, it can only analyze the relationships between factors using historical data and does not consider future information. Bidirectional Long Short-Term Memory (BiLSTM) neural networks consist of two LSTMs, a forward LSTM and a backward LSTM. The forward LSTM captures historical information from the input data, while the backward LSTM obtains future information from the input data. Its structure is as follows: Figure 6 As shown, compared to traditional unidirectional LSTM, BiLSTM overcomes the drawback of insufficient data mining. By referencing historical photovoltaic data and future weather forecasts, using BiLSTM for ultra-short-term photovoltaic power prediction can improve prediction accuracy.

[0136] This invention 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 strengthens key features through an attention mechanism, captures local mutations using a TCN, and models global trends using a BiLSTM, significantly improving the accuracy and stability of photovoltaic power generation prediction while drastically reducing model training time.

[0137] S4: Input the decomposed and filtered data, along with the meteorological data after feature selection, into the combined prediction model for prediction, and then superimpose and integrate the prediction results of each IMF component to obtain the final photovoltaic power prediction value.

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

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope 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: Step 1.1: Obtain the photovoltaic power generation sequence; Step 1.2: Perform multi-scale decomposition on the photovoltaic power generation data sequence in the photovoltaic power generation sequence; Step 1.3: KL divergence screening; In step 1.2, the photovoltaic power generation data is decomposed into multiple scales using the DPEC-ICEEMDAN algorithm 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. Step 1.2 specifically includes: Step 1.2.1: Set the original photovoltaic power generation data sequence as follows A dynamic noise adjustment mechanism is introduced to calculate the local variance of the signal. According to the local variance of the signal Adjust noise ratio in real time ; Step 1.2.2: Based on the noise ratio Constructing decomposed sequences According to the decomposed sequence Calculated photovoltaic power generation data sequence The order residual ; Step 1.2.3: Based on the photovoltaic power generation data sequence The order residual With the order residual Extract the first IMF components of order; Step 1.2.4: When At that time, based on the extracted first Irradiance-power conversion constraints are applied to the first IMF component to eliminate mode aliasing caused by temperature fluctuations; Step 1.2.5: According to the... k Energy percentage of L2 norm acquisition for the first-order IMF component Repeat steps 1.2.1-1.2.4 for iterative decomposition until the energy percentage after the current decomposition is reached. This yields a series of IMF components and a residual term; Local variance The calculation formula is: (1)； Noise ratio The calculation formula is: (2)； In formulas (1) and (2), The average power within the sliding window. For dynamic noise adjustment window length, Time series variance The mean of the sequence is... , This represents the total time step. Time series variance The calculation formula is: (3)； decomposition sequence The calculation formula is: (4)； In formula (4), For the first White noise added during the decomposition To obtain the first value after noise pre-decomposition Order component; Photovoltaic power generation data sequence The order residual The calculation formula is: (5)； In formula (5), This is a local mean calculation. Add the number of times to the noise; No. The formula for calculating the order IMF component is: (6)； The first after applying irradiance-power conversion constraints The expression for the order IMF component is: (7)； In formula (7), For real-time irradiance, Standard irradiance; Energy percentage The calculation formula is: (8)； In formula (8), As a percentage of energy, For the first k The L2 norm of the order IMF, 5%, was determined by Monte Carlo experiments, retaining 95% of the effective energy component; Step 2: Conduct correlation analysis on the random meteorological variables and the indicators to be predicted in the photovoltaic power generation sequence, and select the meteorological data with the highest correlation to the indicators to be predicted based on the correlation analysis results; Step 3: Construct the ECA-TCN-BiLSTM combined prediction model; Step 4: Input the decomposed and screened photovoltaic power generation data and the selected meteorological data into the ECA-TCN-BiLSTM combined prediction model, 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, characterized in that, Step 1 specifically includes: Step 1.1: Obtain the photovoltaic power generation sequence, which 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; Step 1.3: Perform kernel density estimation and discretization calculations on the photovoltaic power generation data and each IMF component, and select the KL divergence of the photovoltaic power generation data corresponding to each IMF component. For each IMF component, the KL divergence obtained from its sampling is used as a basis. Sort the sizes and select those that satisfy the criteria. The IMF components are used to output the corresponding decomposed and filtered photovoltaic power generation data.

3. The ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data according to claim 2, characterized in that, Step 1.3 specifically includes: Step 1.3.1: Process the original photovoltaic power generation data sequence Kernel density estimation is performed on each corresponding IMF component to obtain its probability density function; Step 1.3.2: For each IMF component, calculate its KL divergence with the original photovoltaic power generation data; Step 1.3.3: In the interval Inner uniform sampling Set sampling step size The KL divergence of the photovoltaic power generation data corresponding to each IMF component was obtained. ; Step 1.3.4: For each IMF component, calculate the KL divergence obtained from its sampling. Sort the sizes and select those that satisfy the criteria. The IMF components are used to output the corresponding decomposed and filtered photovoltaic power generation data. The formula for calculating the probability density function is: (9)； (10)； In formulas (9) and (10), Original photovoltaic power generation data sequence x ( t The kernel density estimation function of ). For kernel function, for The kernel density estimation function for the components is commonly the Gaussian kernel function, i.e. , hx and hi For bandwidth, it can be selected via Silverman rules. σ The standard deviation of the sample is 1. The length of the original photovoltaic power generation data sequence; The formula for calculating KL divergence is: (11)； Sampling KL divergence The expression is: (12)； In formula (12), For the signal range, The sampling interval is defined as the divergence; the greater the divergence, the more unique information the IMF components contain.

4. The ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data according to claim 1, characterized in that, Step 2 specifically includes: Calculate random meteorological variables in photovoltaic power generation series and The mutual information MI, in which, The sample size is calculated based on the mutual information (MI). and The maximum information coefficient (MIC) is used to select meteorological data that are most correlated with the indicator to be predicted. The closer the MIC values ​​of two variables are to 1, the stronger their correlation.

5. The ultra-short-term photovoltaic power combination prediction method 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 to 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. The ECA-TCN module is a temporal convolutional network with an added efficient channel attention mechanism, including 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 the selected meteorological data, and the output is the extracted attention weights after weighted fusion and the features of the original input data. The Temporal Convolutional Network (TCN) comprises multiple TCN residual blocks. Each residual block includes a causal dilated convolutional layer, a normalization layer, a ReLU activation function layer, and a Dropout layer. An ECA module is integrated during the downsampling stage. The ECA module includes an input layer and a convolutional kernel of 1. 1 A convolutional layer, a normalization layer, and a convolutional kernel of type 1 are all represented by a number 1. 1 A second convolutional layer and an output layer of 1; The BiLSTM layer is a bidirectional long short-term memory network, consisting of two LSTM networks, one forward and one backward. The input of the BiLSTM layer is the extracted attention weights extracted after weighted fusion of the output of the ECA-TCN module and the original input data features. 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.

6. The ultra-short-term photovoltaic power combination prediction method based on multivariate meteorological data according to claim 5, characterized in that, Step 4, obtaining the final photovoltaic power prediction value, specifically includes: The decomposed and screened photovoltaic power generation data and selected meteorological data are input into the ECA-TCN-BiLSTM combined prediction model. The causal dilated convolutional layer in the ECA-TCN module learns the features of the input data. Nonlinear factors are introduced through normalization layer, ReLU activation function layer and Dropout layer. After regularization of the TCN network, it enters the downsampling extreme. The attention weights of the input data are extracted by the ECA module and weighted and fused with the original features of the input data. The weighted and fused features are then input into the BiLSTM network. The BiLSTM network captures historical information of the input data through a forward LSTM network, obtains future information of the input data through a reverse LSTM network, and integrates the prediction results of each IMF component to output the ultra-short-term photovoltaic power prediction value.

Images