Photovoltaic array short-term power prediction method
By introducing fractional norm and energy entropy filtering mechanisms to optimize the sparse probabilistic attention module, and combining dilated causal convolution to process features, the DESformer model is constructed. This solves the problems of unstable parameter configuration and high computational complexity in photovoltaic power prediction, and achieves more efficient feature extraction and improved prediction accuracy.
Patent Information
- Application Number
- CN202511791476.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-02-27
AI Technical Summary
Existing photovoltaic power prediction methods suffer from unstable parameter configuration, high computational complexity, and severe information loss in mode decomposition and feature optimization, making it difficult to achieve accurate noise reduction and lacking real-time performance, resulting in insufficient prediction accuracy and robustness.
A fractional norm and energy entropy screening mechanism is used to optimize the sparse probabilistic attention module. Combined with dilated causal convolution to process feature dimensions and receptive fields, a DESformer model is constructed, and the number of attention heads is dynamically configured to perform key feature screening and fusion.
It improves the accuracy and real-time performance of photovoltaic power prediction, solves the problems of insufficient feature extraction accuracy and insufficient receptive field in traditional methods, and enhances the robustness and computational efficiency of the model.
Smart Images
Figure CN121579925A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of photovoltaic power prediction and energy management, and particularly relates to a photovoltaic array short-term power prediction method based on a sparsity entropy screening DESformer network. BACKGROUND
[0002] As an important renewable energy, photovoltaic power generation is under the background of smart grid and green energy development, and its short-term power accurate prediction has become the core technology to ensure the stable operation of the power grid and optimize energy scheduling. However, photovoltaic power output is affected by many factors such as climate change and solar radiation, showing strong nonlinearity and volatility, and the aging of monitoring equipment and noise interference further aggravate the prediction challenge. There are three problems in the prior art: first, traditional modal decomposition is mostly a single denoising technology that relies on manual parameter adjustment, which is difficult to balance data smoothing and noise suppression effect, and cannot accurately denoise; second, the data extraction is too comprehensive, and the traditional sparse probability attention mechanism only selects features according to sequence correlation, without quantitative evaluation of the quality of the vector itself, which is difficult to accurately lock the key features, resulting in high computational complexity and insufficient real-time performance; third, when facing multi-modal feature processing, the attention mechanism often relies on more heads, which is difficult to dynamically determine the efficient adaptation of the decomposition layer number, and the traditional distillation method relies on down-sampling compression dimension, which is easy to cause loss of key information, ultimately leading to difficulty in meeting the actual demand for prediction accuracy and robustness.
[0003] In recent years, photovoltaic power prediction methods have gradually developed from traditional probability models and machine learning models to deep learning methods that integrate multi-modal features, but the existing technology that integrates modal decomposition and feature optimization still has limitations: some schemes do not adaptively optimize the modal decomposition parameters, and the denoising effect is unstable; some schemes introduce sparse attention, but do not solve the contradiction between computational complexity and comprehensive feature expression, and do not combine entropy and other indicators to realize dynamic screening of feature quality; at the same time, the traditional multi-head attention mechanism has computational redundancy when fusing features, and the distillation process is easy to lose information, making it difficult to accurately focus on dominant features. Therefore, how to reduce manual configuration of model parameters, accurately screen features using adaptive optimization, and reduce information loss has become the key to improving the accuracy and real-time performance of photovoltaic power prediction. SUMMARY
[0004] The present application aims to overcome the shortcomings of the prior art, and provides a photovoltaic array short-term power prediction method, which combines modal decomposition and parameter adaptive optimization, introduces fractional order entropy screening and expansion factor to construct a DESformer model, and provides a new method for photovoltaic array short-term power prediction. The method dynamically evaluates and optimizes the feature extraction accuracy by fractional norm and energy entropy, introduces an expansion causal convolution to compress the feature dimension and expand the receptive field, dynamically configures the number of attention heads combined with the sparsity index, effectively overcomes the prediction difficulties caused by the complex trend, random fluctuation and noise interference of photovoltaic power, and improves the accuracy, robustness and real-time performance of the model, thereby providing a theoretical basis for the simulation scheduling of the power grid, intelligent operation and maintenance of the photovoltaic base station and fault warning.
[0005] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme:
[0006] A photovoltaic array short-term power prediction method, comprising the following steps,
[0007] Step 1: Real-time acquisition of current and voltage values of the photovoltaic array power generation loop by using current and voltage sensors, calculation of power generation power data as an original data set, and division of the original data set into a training set, a validation set and a test set.
[0008] Step 2: The original data set is pretreated by using a method combining variational modal decomposition and singular spectrum analysis noise reduction, including data smoothing, noise reduction and reconstruction, and using a spider optimization algorithm to adjust the modal number and the penalty factor, to obtain the denoised power generation power data.
[0009] Step 3: By introducing a fractional norm and energy entropy secondary screening mechanism, the sparse probability attention module is optimized to improve the screening accuracy of the key query sequence, a DESformer model is constructed, the sparsity and information concentration of the key query feature are quantified, and the key feature weight is calculated;
[0010] Step 3.1: First, the query vector with strong association with the key vector in the input sequence is preliminarily screened. Assuming that the query vector corresponding to the input sequence is , the key vector is , and the value vector is , wherein B represents the batch size, H represents the number of heads, and D represents the feature dimension. L Q , L K and L V represent the lengths of the query sequence, the key sequence and the value sequence respectively; a sampling matrix is constructed, which has a dimension of [B, H, L Q , f*ln(L K ), D], f is a scaling factor, and the attention score M= is obtained by using the attention score M= quantitative analysis, according to the correlation score, select the top u = f *ln(L Q ) part Q, as Q p .
[0011] Step 3.2: For the Q p after preliminary screening, introduce the fractional norm L q and energy entropy E index for secondary screening; use L q norm to quantify the sparse representation of Q p , and q<1 is more sensitive to sparse features, and E is used to describe the information concentration degree, and the smaller the entropy value is, the higher the information aggregation degree is; by constructing the combined index of fractional norm and energy entropy, the query quality score S(Q i ) is represented as, , wherein D is the feature dimension number of the query Q to be screened, Q i (d) represents the feature value of the i-th query Q corresponding to the d-th dimension, , which is used to quantify the sparsity of each query Q to be screened, , which is a normalization operation, , which is used to quantify the information concentration degree of the query Q to be screened, Norm(·) represents the result min-max normalization, , wherein ε represents an arbitrarily small positive number, β represents the fusion ratio, controls the proportion of L q and E to the quality score, q∈(0,1) is the fractional order, which is determined by multiple experiments.
[0012] Step 3.3: Through the constructed fractional norm and energy entropy double index, the final screened Q F is obtained, which simultaneously considers the sparsity and information concentration degree, which makes up for the limitation of only focusing on Q*K correlation in the initial screening, and is represented as, , wherein the screening length K F = f *In(u), rank(.) represents the ascending screening of Q i according to the quality score S(.), so as to realize the deep improvement of Q quality.
[0013] Step 3.4: Informer selects lnL Qs and Ks through sparse probability sampling, reduces the calculation complexity to O(L*lnL); introduces fractional norm and energy entropy screening for u, and finally obtains K F Qs, assuming f =1, since , , we can get , that is, the number of Qs after sparse entropy screening is ln(lnL Q ); therefore, assuming the query length L Q and the key length LK When all ... .
[0014] Step 4: Subsequently, dilated causal convolution is used to process the key weights. By using the principle of dilation of the convolution kernel layer by layer to introduce the interval, the sequence length is halved under zero padding conditions. This replaces max pooling downsampling to retain information while expanding the model's perception range of the sequence.
[0015] Step 4.1: For the dilated causal convolution of the l-th layer, input feature X l-1 =[B, C, T (l-1) ]∈R, C is the number of channels, T (l-1) Given the length of the input sequence, and the convolution kernel K = [3, C, C] ∈ R, then the output feature X l The calculation is as follows: Where s = k-1 is the position index of the convolution kernel, d l The expansion factor of the l-th layer is determined by the position index of the sequence. This strictly ensures causal convolution.
[0016] Step 4.2: Use dilated causal convolution to process the kernel index positions of the time series to obtain the expanded receptive field; assume the receptive field size of the l-th layer of the dilated causal convolution network is R. l Then its calculation formula is, .
[0017] Step 4.3: Expand the receptive field using an inflation factor. With padding = 0, reduce the sequence dimension using interval sampling. Assuming the initial length T of the input sequence, design a dilated causal convolution with an inflation factor of d. The corresponding output T' is... By using dilation factor interval sampling, the sequence dimension can be halved while enhancing the model's capture of global dependencies, which can effectively replace the hard dimensionality reduction of Maxpool downsampling.
[0018] Step 5: Dimensionality reduction weights are fused with the intermediate feature output of the decoder through multi-head attention. The sparsity of each head is quantified using the fractional norm, and the key heads are focused on for calculation based on the sparsity level with the median as the threshold, thereby achieving weight fusion.
[0019] Step 5.1: Assume the initial number of attention heads is H, and the query vector corresponding to the h-th head ∈ {1,2,...,H} is Q. h =[B, L Q , d Q ], the key vector is K h = [B, L K , d K], where B is batch size, L Q With L K corresponding to the length of query sequence and key sequence respectively, d is the dimension of the single head, d Q = d K . Calculate the normalized similarity matrix W h of each head, denoted as, where, denotes the normalization using L2 norm, and ε is an arbitrarily small positive number.
[0020] Step 5.2: Convert the calculated W h to probability distribution P h by softmax. h After that, build the energy entropy index to calculate the average entropy value S h corresponding to the hth head, which can be expressed as, where, corresponds to , S h is smaller, the sparsity of the hth head is higher, and it is more worthy to be retained.
[0021] Step 5.3: Take the median S m of the entropy values of all heads as the threshold to screen the heads with high sparsity as the effective head index, and the rest as the redundant head index . Finally, for H n , directly reuse the value vector, and use the clipping or zero padding method to ensure alignment with L Q , for H p , calculate the attention weight based on the effective head, and fuse with the value vector to output, to get the fused value After that, a mapping matrix F = [H,H] is also constructed, which is initially a unit matrix, so that the output of the effective head is mapped back to the H-dimensional head space, expressed as where, M p represents the mapping result, F[H p , :] represents the extraction of the corresponding row of the effective head from the mapping matrix, and after transposition, it is multiplied with the output of the effective head, so as to expand the output dimension to H.
[0022] Step 6: Set the prediction step, and after nonlinear transformation by the fully connected layer, output the final prediction value.
[0023] The present application has the following advantages and beneficial effects:
[0024] I. The present application proposes a photovoltaic power prediction core architecture based on sparse entropy screening DESformer network, which innovatively introduces fractional norm and energy entropy dynamic evaluation into the Informer model, optimizes the extraction accuracy of the sparse probability attention mechanism to the query sequence, and replaces the traditional maximum pooling with the dilated causal convolution to realize the halving of the key feature dimension while expanding the receptive field, breaking through the limitations of the traditional model feature extraction accuracy and insufficient receptive field, and significantly improving the capture ability of complex fluctuation characteristics.
[0025] II. The present application constructs a dynamic attention head number configuration and redundant feature screening mechanism, combines the sparse degree index to adaptively adjust the number of attention heads, discards redundant features according to a reasonable retention ratio, realizes efficient weight fusion of key features and decoder output, solves the problems of low feature fusion efficiency, redundant interference of prediction accuracy caused by manual setting of the number of heads, and strengthens the pertinence and effectiveness of feature fusion.
[0026] III. The present application proposes a preprocessing architecture using variational modal decomposition combined with singular spectrum analysis secondary noise reduction, which can effectively deal with the fluctuation complexity of photovoltaic power data, and accurately denoise and reconstruct; at the same time, the spider bee optimization algorithm is introduced to dynamically configure the modal number, penalty factor and other key parameters in modal decomposition, solve the problems of unstable noise reduction effect and poor noise reduction effect caused by manual parameter adjustment, and improve the accuracy of data preprocessing. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 is the flowchart involved in the present application;
[0028] Figure 2 is the original waveform graph of the actual power generation in the present data set;
[0029] Figure 3 is the error iteration graph of adjusting the modal number and penalty factor in the preprocessing of the present application;
[0030] Figure 4 is the waveform graph of the modal decomposition signal and the denoising signal under the optimal parameter adjustment in the preprocessing of the present application;
[0031] Figure 5 is the attention score comparison graph before and after introducing the fractional norm and energy entropy joint index in the present application;
[0032] Figure 6 is the prediction graph of the present application for power generation in 1 step.
[0033] Figure 7 is the prediction graph of the present application for power generation in 8 steps. DETAILED DESCRIPTION
[0034] The application is further illustrated by the following examples. It should be noted that the examples are only used to illustrate and explain the application concept, and various modifications, supplements or replacements of the described specific examples or similar ways can be made by those skilled in the art without departing from the application concept or beyond the scope defined by the claims, which should be considered to fall within the protection scope of the application.
[0035] Referring to Figures 1-6 The application discloses a photovoltaic array short-term power prediction method based on a sparse entropy filter DESformer network, comprising the following steps:
[0036] Step 1: Use a current-voltage sensor to collect the current-voltage values of the photovoltaic array power generation loop, calculate the power generation power data as the original data, and use the data from the Australian Desert Knowledge Solar Centre (M. Zhang, Y. Han, C. Wang, P. Yang, C. Wang, A. S. Zalhaf, Ultra-short-term photovoltaic power prediction based on similar day clustering and temporal convolutional network with bidirectional long short-term memory model: A case study using DKASC data, Applied Energy, (375) (2024) 124085.) in this example, as shown in Figure 2 The data set is collected from photovoltaic stations in various places in Australia, and the overall meteorological and photovoltaic data spans more than 15 years with a sampling interval of 5 minutes per time. The application takes 13000 data points from January 1, 2024 to February 15, 2024 as the data set, and divides them into training set data, validation set data and test set data according to the ratio of 7:2:1 for experiments.
[0037] Step 2: The generated power, as a key indicator of the photovoltaic array's power generation capacity, will be the prediction target, measured in kilowatts. Due to the interference of monitoring equipment and environmental factors on the waveform collection of raw data, this patent uses a combination of variational mode decomposition {X. Zhang, H. Luo, Y. Pei, K. Ma, Mechanism-guided short-term heatload prediction of district heating system based on a hybrid data-driven model, Applied Thermal Engineering, (280) (2025) 128113.} and singular spectrum analysis denoising to smooth, denoise, and reconstruct the original data. The Spider Optimization Algorithm is used to optimize the mode number and penalty factor. The optimization iteration graph and reconstructed signal graph are shown in Figure 3 、 4 The optimized mode number K = 3 and penalty coefficient α = 3000.
[0038] Step 3: Use the improved Informer model {H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, W. Zhang, Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting, Proceedings of the AAAI Conference on Artificial Intelligence, (2021) 11106–11115.}, introduce fractional norm and energy entropy secondary screening mechanism, optimize the sparse probability attention module to screen the key query sequence, construct DESformer model, quantify the sparsity and information concentration of key query features, and then calculate the key feature weight;
[0039] First, the input sequence is screened for query vectors strongly associated with key vectors. Assume that the input sequence corresponds to query vectors , key vectors , and value vectors , where B represents the batch size, H represents the number of heads, and D represents the feature dimension. L Q , L K , and L V represent the lengths of the query sequence, key sequence, and value sequence, respectively; construct a sampling matrix with dimensions [B, H, LQ , f*ln(L K ), D], f is the scaling factor, using attention score M= Q and The correlation strength was quantitatively analyzed, and the top-ranked values were selected based on the correlation scores: u = f * ln(L Q The part Q of ) is considered as Q strongly correlated with K. p .
[0040] For Q after initial screening p Introducing the fractional norm L q Secondary screening using the energy entropy E index; utilizing L q Quantification of Q p The sparse representation capability is enhanced, and it is more sensitive to sparse features when q < 1. E is used to characterize the degree of information concentration; the smaller the entropy value, the higher the degree of information aggregation. By constructing a joint index of fractional norm and energy entropy, the query quality score S(Q) is defined. i ) is represented as, Where D is the number of feature dimensions of the query Q to be screened, and Q i (d) represents the feature value of the d-th dimension corresponding to the i-th query Q to be screened. This is used to quantify the sparsity of each query Q to be screened. For normalization operations, This is used to quantify the information concentration of the query Q to be screened. Norm(·) represents the min-max normalization of the result. Let β represent any small positive number. Through multiple experimental comparisons, β was set to 0.3 as the fusion ratio, controlling L. q The proportion of E in the quality score.
[0041] By constructing a joint index of fractional norm and energy entropy, we obtain Q, which simultaneously considers sparsity and information concentration after final screening. F This overcomes the limitation of the initial screening which only focuses on Q*K correlation, and is expressed as follows: Where, the filtering length K F = f * In(u), rank(.) represents Q according to the quality score S(.). i By performing ascending sorting, a significant improvement in Q quality can be achieved.
[0042] Informer uses sparse probability sampling to select lnL Q and K values for computation, reducing the computational complexity to O(L*lnL); it then introduces fractional norm and energy entropy for u to filter the results, ultimately obtaining K. F There are Q, assuming f = 1, since ,but , can be obtained Q is the number of reserved Q after the sparsity entropy screening, and the formula is ln(lnL Q );thus, assuming the query length L Q and the key length L K are both L, the calculation complexity of Q*K can be optimized to . Figure 5 The change of attention score on sequence distribution after introducing the sparsity index is shown. It can be seen that the subgraph (b) has significantly more regions with high attention scores than the original attention subgraph (a), and the overall sparsity decreases by 45.13%.
[0043] Step 4: Then, the key weight is processed using dilated causal convolution. The principle of introducing interval dilation layer by layer is used, and the sequence length is halved under zero padding conditions. This replaces the maximum pooling downsampling to preserve information while expanding the model's perception range of the sequence.
[0044] For the dilated causal convolution of the l-th layer, the input feature X l-1 =[B, C, T (l-1) ]∈R, C is the number of channels, T (l-1) is the length of the input sequence, the convolution kernel K = [3, C, C] ∈ R, and the calculation of the output feature X l is, where s = k-1 is the position index of the convolution kernel, d l is the dilation factor of the l-th layer, and the position index of the sequence strictly ensures causal convolution.
[0045] Using dilated causal convolution, the index position of the convolution kernel for time series is processed to obtain the expanded receptive field. Assuming that the receptive field size of the l-th layer of the dilated causal convolution network is R l , the calculation formula is, .
[0046] Using the dilation factor to expand the receptive field, under the condition of padding = 0, the sequence dimension is reduced using interval sampling. The initial length of the input sequence T 0 =96 is set, and two layers of dilated causal convolution are designed, with dilation factors d1, d2 being [6,18], and the corresponding output results T 1 , T 2 are, ; Using the dilation factor interval sampling, the sequence dimension is halved while enhancing the model's capture of global dependencies, which can effectively replace the hard dimension reduction of Maxpool downsampling.
[0047] Step 5: The dimension reduction weight is fused with the intermediate feature output in the decoder through multi-head attention, the sparsity of each head is quantified using fractional norm, and the key heads are focused according to the sparsity, with the median as the threshold, and then the weight fusion is realized.
[0048] Assume that the initial number of attention heads is H, and the query vector corresponding to the hth head is Q h = [B, L Q ,d Q ], the key vector is K h = [B, L K , d K ], where B is the batch size, L Q and L K correspond to the lengths of the query sequence and the key sequence respectively, d is the dimension of a single head, and d Q = d K . The normalized similarity matrix W h of each head is calculated, which is represented as, where, represents normalization using L2 norm, and ε is an arbitrarily small positive number.
[0049] After the calculated W h is converted into a probability distribution P h through softmax, an energy entropy index is constructed, and the average entropy value S h corresponding to the hth head is calculated, which can be represented as, where, corresponds to , the smaller S h is, the higher the sparsity of the hth head is, and the more valuable it is to retain.
[0050] The median S m of the entropy values of all heads is used as a threshold to filter the heads with high sparsity as valid head indexes, and the rest are redundant head indexes . For H n , the value vector is directly reused, and clipping or zero padding is used to ensure alignment with L Q . For H p , the attention weight is calculated based on the valid heads, fused with the value vector, and output to obtain the fused value . Then, a mapping matrix F = [H, H] is constructed, which is initially a unit matrix, so that the output of the valid heads is mapped back to the H-dimensional head space, expressed as where M p represents the mapping result, F[H p ,:] represents the extraction of the row corresponding to the valid head from the mapping matrix, and the transpose is multiplied with the output of the valid head, so that the output dimension is expanded to H.
[0051] In the present case, when the initial number of heads is set to 16 and the single-head dimension is 32 for a certain batch, the corresponding valid head indexes H P = [1, 2, 4, 5, 7, 8, 10, 14].
[0052] Step 6: Set the prediction step to be 1 and 8 respectively, and the output is the final prediction value after the weight is nonlinearly transformed by the full connection layer. The final prediction value waveform corresponding to different steps is shown in Figs. Figure 6 and Figure 7 .
[0053] The above is an exemplary description of the application, and it is obvious that the specific implementation of the application is not limited by the above method. Any non-essential improvement or direct application of the inventive concept and technical solution to other occasions is within the protection scope of the application.
Claims
1. A method for short-term power prediction of a photovoltaic array, characterized in that, The method comprises the following steps: Step 1, real-time acquisition of current and voltage values of a photovoltaic array power generation loop by using a current-voltage sensor, calculation of power generation power data as an original data set, and division of the original data set into a training set, a validation set and a test set; Step 2, preprocessing of the original data set by combining variational mode decomposition and singular spectrum analysis noise reduction, wherein the preprocessing includes data smoothing, noise reduction and reconstruction, and adjustment of mode number and penalty factor by using a spider wasp optimization algorithm to obtain denoised power generation power data; Step 3, introduction of a fractional norm and energy entropy secondary screening mechanism to optimize the screening accuracy of a sparse probability attention module for a key query sequence, construction of a DESformer model, quantification of the sparsity and information concentration of key query features, and calculation of key feature weights; Step 4, then, processing of the key weights by using dilated causal convolution, layer-by-layer introduction of the principle of interval dilation by using a convolution kernel, realization of halving of the sequence length under zero padding, replacement of maximum pooling downsampling to retain information while expanding the perception range of the model for the sequence; Step 5, fusion of the reduced dimension weights with intermediate feature outputs in the decoder, quantification of the sparsity of each head by using a fractional norm, focusing on key head calculation according to the sparsity level with the median as the threshold, and then realizing weight fusion; Step 6, setting a prediction step, inputting the fused weights into a fully connected layer after nonlinear transformation, and outputting power generation power prediction values.
2. The photovoltaic array short-term power prediction method according to claim 1, characterized in that, The step 3 comprises: Step 3.1, preliminary screening is performed on the query vectors that are strongly associated with the key vectors in the input sequence, assuming that the input sequence corresponds to the query vector , key vector , value vector , where B represents the batch size, H represents the number of heads, D represents the feature dimension, L Q , L K and L V represent the lengths of the query sequence, the key sequence and the value sequence respectively; a sampling matrix is constructed, with dimensions [B, H, L Q , f*ln(L K ), D], f being a scaling factor, the relevance strength of Q and is quantitatively analyzed by means of the attention score M= , and according to the correlation score, the top u = f*ln(L Q ) parts of Q are selected as the Q p that are strongly associated with K Step 3.2, Q p , the fractional norm L q and energy entropy E index secondary screening; using L q norm quantization Q p sparse representation ability, and q < 1 is more sensitive to sparse features, E is used to depict the information concentration degree, the smaller the entropy value, the higher the information aggregation degree; by constructing the fractional norm and energy entropy joint index, the query quality score S(Q i ) is represented as, (3.2) where D is the feature dimension number of the query Q to be screened, Q i (d) represents the feature value of the dth dimension corresponding to the ith query Q to be screened, is used to quantify the sparsity of each query to be screened, is used for normalization operation, is used to quantify the information concentration of the query Q to be screened, Norm(·) represents the result of min-max normalization, represents an arbitrarily small positive number, β represents the fusion ratio, and L q is the proportion of the quality score, q∈(0,1) is the score order, and is determined by comparison through multiple experiments; Step 3.3, the final screening index Q is obtained by combining the constructed fractional order norm with the energy entropy joint index, which considers sparsity and information concentration after synchronization F , is expressed as, (3.3) Wherein, the filtering length K F = f*ln(u), rank(.) represents Q ranked by quality score S(.). i By performing ascending sorting, a significant improvement in Q quality can be achieved. Step 3.4, Informer filters out lnL Qs by sparse probability sampling, and calculates with K, reducing the computational complexity to O(L*lnL); introduces fractional norm and energy entropy screening for u, and finally obtains K F Qs, assuming f=1, since , , , that is, the number of Qs after sparse entropy screening is ln(lnL Q ); assuming that the query length L Q and the key length L K are both L, the computational complexity of Q*K is optimized to O(L*lnL) by sparse entropy screening mechanism, (3.4)。 3. The photovoltaic array short-term power prediction method according to claim 2, characterized in that, The step 4 comprises: Step 4.1, for the dilated causal convolution of the l-th layer, the input feature X l-1 = [B, C, T (l-1) ] ∈ R, C is the number of channels, T (l -1) is the length of the input sequence, the convolution kernel K = [3, C, C] ∈ R, and the calculation of the output feature X l is: (4.1) where s = k - 1 is the position index of the convolution kernel, d l is the dilation factor for the l-th layer, according to the position index of the sequence , strictly ensures the causal convolution; Step 4.2, processing the convolution kernel index position of the time series by using the dilated causal convolution, to obtain an expanded receptive field; assuming that the receptive field size of the dilated causal convolution network is R l The calculation formula is: (4.2) Step 4.3: expanding the receptive field by using a dilation factor, under the condition of padding = 0, reducing the sequence dimension by using interval sampling, assuming that the initial length of the input sequence is T, designing a dilated causal convolution, the dilation factor is d, and the corresponding output result T' is: (4.3) Interval sampling by using a dilation factor realizes halving of the sequence dimension while enhancing the capture of global dependencies by the model, effectively replacing the hard dimension reduction of Maxpool downsampling.
4. The photovoltaic array short-term power prediction method according to claim 3, characterized in that, The step 5 comprises: Step 5.1, assuming the initial number of attention heads is H, the query vector corresponding to the hth head is Q h = [B, L Q , d Q ] and the key vector is K h = [B, L K , d K ], d is the dimension of a single head, d Q = d K , the normalized similarity matrix W h of each head is calculated, which is represented as: (5.1) wherein, denotes normalization using the L2 norm, and ε is an arbitrarily small positive number; Step 5.2, calculate W h Convert to probability distribution P by softmax h After that, construct the energy entropy index, calculate the average entropy value S corresponding to the hth head h , expressed as: (5.2) wherein, corresponding , S h The smaller, the higher the sparsity of the hth head, the more worthy of retention. Step 5.3, median S of all heads' entropy values m Filtering heads with high sparsity as threshold As valid head index, the rest are redundant head index Finally, for H n Directly reuse the value vector, and use cropping or zero padding to ensure alignment with L Q For H p Then compute attention weight based on valid heads, and fuse with value vector to output, get fused value After that, we also need to construct a mapping matrix F = [H, H] initialized as identity matrix, to map the output of valid heads back to H-dimensional head space, expressed as: (5.3) where M p represents the mapping result, F[H p , :] represents the row corresponding to the effective head extracted from the mapping matrix, and after transposition, matrix multiplication is performed with the effective head output, thereby expanding the output dimension to H.