A regional photovoltaic power prediction method of decoupling asymmetric space-time graph fusion Transformer

Abstract

A regional photovoltaic power prediction method decoupled from asymmetric spatiotemporal graph fusion Transformer is proposed. The method comprises the following steps: Step 1: Physically isolating and parallelizing the temporal evolution pattern and spatial topological dependency using a spatiotemporal dual-stream parallel architecture; Step 2: Capturing the long-range temporal evolution patterns of univariate variables using a Patching-level Transformer in the temporal dependency branch, and reducing the computational complexity of long sequences using patching techniques; Step 3: Generating a learnable asymmetric adjacency matrix using a source-target decoupled asymmetric graph learning mechanism in the spatial dependency branch; Step 4: Performing deep mutual querying and feature enhancement of spatiotemporal features before reconstruction using a spatiotemporal bidirectional information interaction module; Step 5: Adopting a prediction-decoding paradigm of feature projection stitching and fusion to map highly collaborative spatiotemporal features to the target prediction space and outputting the final prediction result after inverse normalization.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction, and specifically to a regional photovoltaic power prediction technology that decouples asymmetric spatiotemporal graph fusion Transformer. Background Technology

[0002] Faced with the dual pressures of dwindling fossil fuel reserves and increasingly severe environmental problems, renewable energy is playing an increasingly prominent strategic role in the energy system, with photovoltaic (PV) power generation becoming a core pillar of the global energy transition. With the rapid increase in the proportion of distributed PV installations, regional-level multi-site collaborative power forecasting has become a key technological support for real-time grid dispatching and spot market trading. However, influenced by complex coupling factors such as cloud movement and atmospheric radiation variations, PV power output exhibits extremely strong volatility; particularly in the spatial dimension, for two different PV sites, the power output varies significantly depending on the site. i Arrive at the station j The intensity of the impact is not necessarily equal to that of the site j Arrive at the station i The intensity of the influence is significant. Therefore, it is difficult to fully describe the spatial dependence changes across different historical observation windows using only symmetric distance plots or fixed static plots.

[0003] For photovoltaic (PV) power prediction, current research is increasingly dominated by deep learning, but existing methods still have limitations in handling spatiotemporal features. While convolutional neural networks excel at modeling local spatial relationships, they lack long-range temporal series modeling capabilities; long short-term memory networks, while capable of capturing temporal dynamics, struggle to characterize the dynamic spatial topological relationships between sites. This one-dimensional focus fails to reflect the essential spatiotemporal coupling effect on PV. Furthermore, the serial spatiotemporal combination method is prone to premature mixing and mutual interference of PV power spatiotemporal features, lacking a deep spatiotemporal collaborative interaction mechanism, resulting in poor model robustness under sudden weather scenarios.

[0004] In recent years, graph neural networks and Transformers have gradually become hot topics, but they are constrained by the shortcomings of graph construction mechanisms when dealing with multi-site photovoltaic clusters. Existing studies generally use geographical distance to construct symmetric graph structures, ignoring the unidirectional transmission mechanism of energy propagation in actual meteorological processes. This makes it difficult for models to accurately reproduce the dynamic propagation path and impact timing of cloud shadows. At the same time, most models rely solely on predefined static graphs, failing to achieve the complementary advantages of physical priors and adaptive topology. This makes it difficult to quickly correct prediction results based on prior information under complex weather conditions, limiting the stability of predictions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a regional photovoltaic power prediction method that decouples asymmetric spatiotemporal graphs and fuses Transformers.

[0006] This invention is a method for predicting regional photovoltaic power by decoupling asymmetric spatiotemporal graphs and fusing Transformers, comprising the following steps: Step S1: Obtain historical photovoltaic power time series data of multiple photovoltaic sites in the region, construct input samples according to continuous historical observation windows, and use a preprocessing method based on sample statistics decoupling and supporting reverse reconstruction to perform feature scaling on the input samples and save the sample statistics required for reverse reconstruction. Step S2: Input the same preprocessed historical observation window into the parallel time-dependent branch and spatial-dependent branch; in the time-dependent branch, divide the power time series of each photovoltaic site into multiple overlapping time patches, map each time patch into patch time features, and use the Transformer encoder to extract patch-level long-range time evolution features; in the spatial-dependent branch, generate an explicit sample statistical adjacency matrix based on the Pearson correlation coefficient between photovoltaic sites within the same historical observation window, and generate an implicit learnable asymmetric adjacency matrix using parameter-independent influence source mapping network and influence target mapping network; Step S3: Generate node-by-node pair gating weights based on the explicit sample statistical adjacency matrix and the implicit learnable asymmetric adjacency matrix. Perform edge-by-edge fusion of the two adjacency matrices according to the gating weights to obtain a sample-related asymmetric fused adjacency matrix. Use a graph convolutional network to aggregate the neighborhood information of each photovoltaic site according to the asymmetric fused adjacency matrix. After the graph convolutional network outputs, perform spatial self-attention calculation along the photovoltaic site axis to obtain a spatially enhanced representation that simultaneously contains local directed topological information and global site interaction information. Step S4: Utilize two sets of independent cross-attention networks to perform spatiotemporal information interaction in two directions. In the first cross-attention direction, the pooling time features of each photovoltaic site are used as queries, and the spatial enhancement representations of multiple photovoltaic sites are used as keys and values, enabling the pooling time features of each photovoltaic site to selectively absorb the spatial information of multiple photovoltaic sites. In the second cross-attention direction, the spatial enhancement representations of each photovoltaic site are used as queries, and the patch time features corresponding to that photovoltaic site are used as keys and values, enabling the spatial enhancement representations of each photovoltaic site to selectively absorb its corresponding historical time patch information. The two cross-attention directions respectively employ independent query projection parameters, key projection parameters, and value projection parameters, and combine residual connections and layer normalization operations to obtain enhanced time representations and enhanced spatial representations that maintain the correspondence between sites. Step S5: Map the enhanced temporal representation and enhanced spatial representation to the same target prediction step space to obtain temporal prediction features and spatial prediction features; splice and fuse the temporal prediction features and spatial prediction features to generate multi-site photovoltaic power prediction results at a normalized scale; use the sample statistics obtained and saved from the same historical observation window in step S1 to perform inverse scaling restoration on the multi-site photovoltaic power prediction results and output the regional multi-site photovoltaic power prediction results at the original power scale.

[0007] The advantages of this invention are: 1. By using a Patching-level Transformer and a graph convolutional network to construct parallel temporal and spatial dependency branches, long-range temporal evolution features and spatial topological features are extracted separately before interaction, reducing mutual interference caused by premature mixing of the two types of features.

[0008] 2. A learnable asymmetric adjacency matrix is ​​generated through the source node mapping network and the target node mapping network, enabling the nodes to... i Pointing to node j The connection weights can differ from those of the nodes. j Pointing to node i The connection weights represent the directed dependencies between multiple sites.

[0009] 3. Gated fusion of the Pearson correlation adjacency matrix of the current historical observation window with the learnable asymmetric adjacency matrix is ​​performed so that the final adjacency matrix contains both the statistical correlation of samples and the structural correlation obtained during training.

[0010] 4. Spatiotemporal information interaction is achieved by using cross-attention in two directions: time query space and space query time. This allows time features to selectively absorb spatial information from different sites, and spatial features to selectively absorb historical time information.

[0011] 5. Extensive comparative experiments were conducted on two sets of photovoltaic datasets. The results show that the method of this invention achieves excellent prediction performance compared with existing baseline methods. Attached Figure Description

[0012] Figure 1 This is the overall structure diagram of the decoupled asymmetric spatiotemporal graph fusion Transformer architecture: DAST-GT structure diagram; Figure 2 This is a comparison chart of the model's results in long-term photovoltaic power prediction; Figure 3 It is a visualization of spatiotemporal power propagation and advance prediction; Figure 4 , Figure 5 , Figure 6 as well as Figure 7 This is a performance comparison chart of various ablation variant models under different prediction step sizes; Figure 8 , Figure 9 as well as Figure 10 It is a weight visualization diagram of the asymmetric spatial dependence and spatiotemporal bidirectional information interaction module. Detailed Implementation

[0013] like Figure 1 As shown, this method provides a regional photovoltaic power prediction approach that decouples asymmetric spatiotemporal graph fusion Transformer. The method sets up temporal and spatial dependent branches: the temporal dependent branch utilizes temporal patching and a Transformer encoder to extract long-range temporal evolution features; the spatial dependent branch generates an explicit sample statistical adjacency matrix based on historical observation windows, and generates an implicitly learnable asymmetric adjacency matrix based on learnable node embeddings and parameter-independent influence source mapping networks and influence target mapping networks. The two adjacency matrices are fused through a gating mechanism to obtain a sample-related asymmetric fused adjacency matrix. Based on this, a graph convolutional network, a spatial self-attention module, and a spatiotemporal bidirectional information interaction module are used to achieve the collaborative fusion of temporal and spatial information, generating regional multi-site photovoltaic power prediction results.

[0014] The present invention will be further described in detail below with reference to the embodiments.

[0015] Implementation steps A regional photovoltaic prediction method decoupled from asymmetric spatiotemporal graph fusion Transformer includes the following steps: Step S1: Obtain the historical operation dataset of the photovoltaic power station and preprocess the historical data; The specific method for step S1 is as follows: Historical photovoltaic power time-series data from multiple photovoltaic sites within the region are acquired. Input samples are constructed according to continuous historical observation windows. A preprocessing method based on sample statistics decoupling and supporting reverse reconstruction is used to perform feature scaling on the input samples and save the sample statistics required for reverse reconstruction. Step S2: Build time-dependent branches and space-dependent branches; The specific method for step S2 is as follows: The same preprocessed historical observation window is simultaneously input into parallel time-dependent and spatial-dependent branches. The time-dependent branch is used to extract patch-level long-range temporal evolution features from the power sequences of each photovoltaic site; the spatial-dependent branch is used to extract sample-related asymmetric spatial topological features among multiple photovoltaic sites within the current historical observation window. The two branches operate in parallel during the feature extraction stage, and after completing their respective feature extractions, they exchange information through the spatiotemporal bidirectional information interaction module in step S4.

[0016] Step S3: Obtain spatially enhanced representation; The specific method for step S3 is as follows: Multi-layer graph convolution is performed using the sample-related asymmetric fused adjacency matrix generated in step S2 to aggregate the local neighborhood information of each photovoltaic site. After the graph convolutional network outputs, a spatial self-attention module is used to calculate the global site interaction information along the photovoltaic site axis. The result obtained by the formula simultaneously contains local directed topological information and global site interaction information, while maintaining a one-to-one correspondence with each photovoltaic site.

[0017] Step S4: Utilize the spatiotemporal two-way information interaction module to perform spatiotemporal feature interaction in different directions; The specific method for step S4 is as follows: The first cross-attention direction uses the pooled temporal features of each photovoltaic (PV) site as the query and the spatial augmentation representations of multiple PV sites as the key and value. This direction is used to enable the temporal features of each PV site to selectively absorb the spatial information of multiple PV sites according to the attention weights, and output an augmented temporal representation that maintains the correspondence between PV sites.

[0018] The second cross-attention direction uses the spatial enhancement representation of each photovoltaic (PV) site as the query and the corresponding temporal feature of that PV site as the key and value. This direction is used to selectively absorb the temporal information of each PV site within its historical observation window, outputting an enhanced spatial representation that maintains the correspondence between PV sites.

[0019] The two cross-attention directions each have their own query projection matrix, key projection matrix, and value projection matrix, and they do not share projection parameters. The query objects and query ranges are different in the two directions, and they do not belong to a symmetric structure that simply replicates the same cross-attention calculation in reverse. The cross-attention outputs of the two directions are respectively residual-connected to their query inputs and layer normalized, and then respectively input into independent prediction projection layers.

[0020] Step S5: Prediction mapping, feature fusion, and inverse scaling restoration; The specific method for step S5 is as follows: The enhanced temporal and spatial features are mapped to the target prediction step space through independent linear layers and then concatenated along the feature dimension. Subsequently, the multi-view information is fused using the output mapping layer, and an inverse normalization operation based on RevIN is performed to generate the final future photovoltaic power prediction sequence.

[0021] Step S5 uses MSE, MAE, and MAPE as evaluation indicators.

[0022] The normalization process in step S1 of the method described above uses the following formula for sample statistics decoupling and feature scaling: (1) (2) (3) in, B Indicates batch size, L This indicates the number of time steps in the photovoltaic power sequence within the historical observation window. N Indicates the number of photovoltaic power stations; x b,t,n Indicates the first b The first sample index n The site index is in t The original power value at time; μ b,n and σ b,n These represent the mean and standard deviation of the sample over the time dimension, respectively. A tiny constant to prevent division by zero errors; γ n and β n In order to target the n Learnable affine transformation scaling factor and bias term for each site index; The input data is scaled up to the feature size.

[0023] Step S2 specifically involves: The time-dependent branches perform patch partitioning and patch embedding on the power time series of each photovoltaic site, use the Transformer encoder to generate patch time features that retain the patch axis, and pool the patch time features to generate pooled time features for each photovoltaic site. The spatially dependent branch generates an explicit sample statistical adjacency matrix based on the same historical observation window, and generates an implicit learnable asymmetric adjacency matrix based on learnable node embeddings. The explicit sample statistical adjacency matrix and the implicit learnable asymmetric adjacency matrix are fused by gating weights for each node pair. A graph convolutional network is used to aggregate local neighborhood information according to the fused adjacency matrix, and a spatial self-attention module is used to extract global site interaction information along the photovoltaic site axis to generate spatially enhanced representations of each photovoltaic site. The patch time feature, pooling time feature, and spatial enhancement representation respectively retain the index relationship of the corresponding photovoltaic site and serve as the input of the spatiotemporal bidirectional information interaction module. This allows the spatial dependency branch to serve as the spatial feature extraction step in the complete prediction process, and to work in conjunction with the time dependency branch, the spatiotemporal bidirectional information interaction module, and the prediction mapping module to generate photovoltaic power prediction results.

[0024] The specific operations and calculation processes for the time-dependent branches and space-dependent branches in steps S2 and S3 are as follows: The patch length in the time-dependent branch is set to 16 to achieve a balance between reducing computational overhead and ensuring prediction accuracy. The time-dependent branch uses a sliding window mechanism to patch the sequence. LDivide into lengths P len A patch with a step size of S is then embedded and added to the positional code before being fed into a multi-layer Transformer encoder. The encoder output is then globally aggregated using adaptive average pooling. (4) (5) (6) (7) (8) Among them, P len S and S represent the patch length and step size of the sliding window, respectively. For the floor operation, P num Z represents the total number of patches after splitting; patch W is a patch embedding tensor mapped to feature dimension D. pos L is the position encoding matrix; enc Z represents the total number of encoder layers; (0) Z represents the initial input of the Transformer encoder obtained by adding the patch embedding tensor to the positional encoding. (l-1) Z ´(l) With Z (l) They represent the first l The input features of the layer encoder, the output features of the self-attention sublayer, and the output features of the feedforward neural network sublayer; MSA(·), FFN(·), and Layernorm(·) correspond to the multi-head self-attention mechanism, the feedforward neural network, and the layer normalization operation, respectively; Z (Lenc) For the output tensor of the last encoder layer, the Squeeze(·) operation is used to remove feature axes of size 1; AdaptiveAvgPool1d(·) represents the adaptive average pooling operation performed along the patch dimension. This indicates the pooling time characteristics of each site; The spatially dependent branch generates an implicitly learnable asymmetric adjacency matrix through learnable node embeddings and parameter-independent influence source mapping and influence target mapping networks. It then calculates the Pearson correlation coefficient based on the preprocessed power sequences of each photovoltaic site within the current historical observation window, generating an explicit sample statistical adjacency matrix. Finally, it fuses the two adjacency matrices using gating weights and generates a spatially enhanced representation using a graph convolutional network and a spatial self-attention module. (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (twenty one) (twenty two) (twenty three) (twenty four) in, i and j Indicates a photovoltaic site index; E Represents learnable node embeddings, MLP src (·) and MLP tgt (·) represent the mapping networks between the influencing source and the influencing target, respectively. src and E tgt Let A and B represent the source features and target features respectively, T denote matrix transpose, and A... dyn Represents an implicit asymmetric dynamic adjacency matrix; Represents a node i In the t Preprocessing power values ​​at each time step and Representing nodes respectively i and nodes j The average preprocessed power within the historical observation window Represents a node i With nodes j The Pearson correlation coefficient between them, A stat This represents the explicit statistical adjacency matrix mapped to the [0,1] interval; Concat(·) represents feature concatenation, MLP gate (·) represents the gated weight generation network, W gate A represents the adaptive gated weights, ⊙ represents the Hadamard product, and A final This represents the final adjacency matrix obtained after fusing the explicit statistical graph and the implicit dynamic graph; I N L represents the identity matrix. gcn This indicates the total number of convolutional layers in the graph. W represents the input features of the spatial branch. emb The linear embedding weight matrix represents the spatial input features. D represents the initial spatial characteristics. in Represents an in-degree diagonal matrix. Indicates the first r The weight matrix of the layer graph convolution, where ф(·) represents the activation function. Indicates the first r Layered graph convolution output; W Q W K and W V Q represents the projection matrices of the query, key, and value, respectively. spa K spa and V spa These represent the corresponding query, key, and value characteristics, respectively, d K A represents the feature dimension of the attention head. spa W represents the spatial self-attention output. O H represents the output projection matrix. enh This represents the final spatial augmentation representation.

[0025] The spatiotemporal bidirectional information interaction module in step S4 of the method described above includes a first cross-attention direction and a second cross-attention direction. The two cross-attention directions use independent query projection parameters, key projection parameters, and value projection parameters. In the first cross-attention direction, the pooling time characteristics of each photovoltaic site are used as the query, and the spatial enhancement representations of multiple photovoltaic sites are used as the key and value. This allows the pooling time characteristics of each photovoltaic site to selectively absorb the spatial information of multiple photovoltaic sites according to the attention weights. The calculation formula is as follows: (25) (26) (27) (28) (29) In the second cross-attention direction, the spatial enhancement representation of each photovoltaic site is used as the query, and the corresponding temporal feature of the photovoltaic site is used as the key and value. The spatial enhancement representation of each photovoltaic site selectively absorbs the historical temporal information of that photovoltaic site according to the attention weight. The calculation formula is as follows: (30) (31) (32) (33) (34) in, This indicates the pooling time characteristics of each photovoltaic site; This indicates that the patch time characteristics of the photovoltaic site axis and patch axis are retained; Indicates the first n Patch time characteristics of individual photovoltaic sites; H enh Representation of spatially enhanced characterization; H enh,n Indicates the first n Spatial enhancement characterization of individual photovoltaic sites; , , These represent the query projection matrix, key projection matrix, and value projection matrix in the first cross-attention direction, respectively; Q time , and These represent the query matrix, key matrix, and value matrix corresponding to the first cross-attention direction, respectively. Time cross This represents the cross-attention output calculated using pooled temporal features as queries and spatially augmented representations as keys and values. This represents the enhanced temporal representation obtained after injecting spatial information. , and These represent the query projection matrix, key projection matrix, and value projection matrix in the second cross-attention direction, respectively. K time,n and V time,n They represent the first n The query matrix, key matrix, and value matrix corresponding to each photovoltaic site in the second cross-attention direction; Space cross,n Indicates the first n The cross-attention output of each photovoltaic site is calculated using spatial enhancement representation as query and corresponding patch time features as key and value. Indicates the first n The enhanced spatial representation obtained after injecting time patch information into each photovoltaic site; divided by Used to control the numerical scale of the dot product result, improving the numerical stability of attention calculation.

[0026] According to the method described in claim 1, the multidimensional aggregation and target space mapping process in step S5 is specifically as follows: Enhanced temporal features Spatial features F is obtained by mapping to the target prediction step space through independent linear projection layers. time With F space; Concatenating the features along the feature dimension yields the cascaded feature H. concat Multi-view information fusion is performed using the output mapping layer, and the calculation formula is as follows: (35) (36) (37) (38) (39) in, and These represent the weight matrices of the temporal feature projection layer and the spatial feature projection layer, respectively. and F represents the bias vectors of the temporal feature projection layer and the spatial feature projection layer, respectively; time and F space H represents the temporal and spatial prediction features mapped to the target prediction length space, respectively; concat This represents the cascaded feature obtained by concatenating the two elements; W fusion and b fusion Y represents the weight matrix and bias vector of the final fusion layer, respectively; pred This represents the normalized scaling prediction result output by the fusion layer; Denorm(·) represents the RevIN inverse transform operation performed using the sample statistics and learnable affine parameters saved in step S1. This represents the final photovoltaic power prediction result after restoring to the original power scale.

[0027] The methods described above use the following parameters to evaluate model performance: During training, Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE) were selected as performance evaluation metrics; the calculation formulas are as follows: (40) (41) (42) Among them, y m Represents the true value. This represents the predicted value from the corresponding model.

[0028] The experiment of the present invention The experiment used two publicly available photovoltaic (PV) power datasets: EODP and NREL. The EODP data was sourced from the ods032m dataset published on the Elia open data portal, while the NREL data was sourced from publicly available PV power data from the National Renewable Energy Laboratory (NREL). The datasets were divided into training, validation, and test sets, accounting for 70%, 10%, and 20% respectively, in chronological order. The historical input sequence length was set to 96, predicting PV power for the next 16 or 48 time steps. The data time resolution was 15 minutes, so the 16 and 48 prediction time steps corresponded to the next 4 hours and 12 hours, respectively. The experimental dataset information is shown in Table 1.

[0029] Table 1. Description of Experimental Datasets

[0030] (1) Evaluation metrics and baseline methods: The mean absolute error, mean square error (MSE), and mean absolute percentage error are used to assess the error between the actual photovoltaic power and the predicted results. 1) Mean Absolute Error (MAE): As shown in Formula (40).

[0031] 2) Mean Square Error (MSE): As shown in Formula (41).

[0032] 3) Mean absolute percentage error: as shown in formula (42); Among them, ym and These represent the actual photovoltaic power and the predicted results, respectively; the smaller the predicted values ​​of MAE, MSE, and MAPE, the better the prediction effect.

[0033] (2) It is mainly compared with deep learning methods and other baseline methods, which include: 1) Transformer: Encoder-decoder architecture, self-attention to capture global dependencies, positional encoding to supplement temporal information, and parallel sequence transformation; 2) iTransformer: Treats variables as labels to capture multivariate temporal correlations; 3) PatchTST: Segment time series into subsequences to improve long-term prediction accuracy and has excellent self-supervised performance; 4) TSMixer: A hybrid operation of attention and MLP, which efficiently captures global dependencies and feature interactions, improves inference speed in parallel, and is suitable for multi-temporal prediction. 5) GCN: Graph convolution aggregates neighbor information to capture local spatial dependencies in graph structure; 6) GCN-LSTM: Integrates GCN and LSTM to achieve spatiotemporal joint prediction; 7) FourierGNN: Integrates Fourier transform and GNN, jointly models graph signals in the frequency and spatial domains, and captures global or local dependencies; 8) MTGNN: Designed for multivariate time series prediction, it models the dynamic graph structure of variables to capture spatiotemporal dependencies and variable interactions; 9) DLinear: Decomposes time series into trend / residual components, models time dependence with linear layers, and has a simple and effective prediction method; (3) Experimental results and analysis To intuitively evaluate the model's predictive performance in real-world scenarios, we selected test data from the Brussels site of the EODP dataset for visualization analysis. The experimental data had a time resolution of 15 minutes, and the prediction task was set to predict a future sequence of length 48 using a historical sequence of length 96 as input. Figure 2 The results show the comparison between the model's predictions and the actual values ​​on a continuous 70-day test sample. We can clearly observe that the model achieves a high degree of fit with the actual values ​​over a span of 7,000 time steps. Faced with non-stationary data with strong diurnal periodicity and violent peak oscillations, the model does not exhibit the error drift phenomenon commonly seen in long-sequence predictions. This result verifies the effectiveness of our proposed spatiotemporal dual-stream decoupling architecture. With the help of the Patching mechanism, the model successfully captures long-term time evolution patterns, maintaining a keen perception of local features while accurately grasping the long-term operating trend of the system.

[0034] Figure 3 The results show the three-day spatiotemporal propagation and prediction of West-Flanders as the downstream target site and Brussels and East-Flanders as the upstream sites. In the main plot, the black target curve exhibits a lag of approximately 60 minutes relative to the upstream curve during several key fluctuation events, indicating a stable upstream-to-downstream power propagation relationship. Simultaneously, the 15-minute lead prediction points marked in red closely follow the actual target trajectory, effectively tracking the changing trend near the peak and during the decline phase. The three magnified windows below further validate this phenomenon; the model's predictions are generally superior to the persistent baseline (gray crosses) with small biases (approximately -0.007, -0.018, and -0.016), demonstrating that the method effectively characterizes short-term propagation delays and improves short-term prediction accuracy.

[0035] (4) Comparative experiment The model's performance was compared with that of nine common baseline methods for 4-hour and 12-hour predictions; on both datasets, the method achieved the best prediction results across all evaluation metrics.

[0036] According to Tables 2 and 3, the method of this invention achieves lower prediction errors on most evaluation metrics across both datasets. In the 16-step prediction task on the EODP dataset, the MAE of the method of this invention is reduced from 0.077 for the suboptimal method to 0.068, a relative reduction of approximately 11.69%. Under different datasets and prediction step sizes, some baseline methods remain competitive on individual metrics, indicating that the combination of techniques described in the model can be effectively implemented and improve multi-site photovoltaic power prediction.

[0037] Table 2. Performance comparison of different photovoltaic power prediction methods on the EODP dataset (underlined indicates best).

[0038] Table 3. Performance comparison of different photovoltaic power prediction methods on the NREL dataset (underlined indicates best).

[0039] (5) Ablation test To further investigate the performance of different modules of the proposed DAST-GT method, four variants of the method were designed to verify the impact of each module, and these four variants were compared with the average MAE and MSE of the DAST-GT method on the EODP and NREL datasets.

[0040] The differences between these five variant methods and the original method are as follows: 1) No multi-graph convolution and asymmetric graph: Based on the original model method, remove the Pearson correlation graph or asymmetric graph; 2) No time branches: Based on the original model method, remove time branches; 3) No spatial branches: Based on the original model method, spatial branches are removed; 4) No spatiotemporal bidirectional information interaction mechanism: Based on the original model method, the spatiotemporal bidirectional information interaction module is removed; To better explain the performance of the original model, ablation experiments were visualized and analyzed. For example... Figure 4 , Figure 5 , Figure 6 as well as Figure 7 As shown, experimental results indicate that removing either branch variant 2 or 3 leads to a surge in error, such as the MSE in the NREL dataset increasing from 0.0902 to 0.1138, confirming that the two-stream decoupling architecture is the cornerstone for modeling complex spatiotemporal characteristics; removing multi-graph convolution variant 1 shows that a single physical graph cannot describe dynamic changes, and asymmetric graphs are crucial for capturing the impact of sites; removing the spatiotemporal bidirectional information interaction mechanism variant 4 leads to performance degradation, indicating that simple decoupling will cause information fragmentation, and deep interaction correction is the key to enhancing robustness.

[0041] (6) Experiment on the effectiveness of asymmetric graphs and spatiotemporal two-way information interaction mechanisms To further explore the effectiveness of spatiotemporal cross-attention and asymmetric graphs, a visualization analysis was performed, where Node0, Node1…Node10 correspond to the 11 power plants in the EODP dataset. The results are as follows: Figure 8 , Figure 9 and Figure 10 As shown.

[0042] Figure 8 The asymmetric spatial dependency matrix learned by DAST-GT on real test samples is presented. The results show that the influence between different nodes has a clear directionality. For example, connections such as Node10→Node4, Node4→Node7, and Node2→Node1 have high weights, while their reverse connections have relatively low weights, indicating that the model can learn directed spatial dependencies that are difficult to represent using traditional symmetric distance graphs. From the overall influence strength of each source node, Node10, Node4, and Node2 have high connection weights, indicating that these nodes play a more important role in the propagation of spatial information in the current sample.

[0043] Figure 9 and Figure 10 The spatiotemporal cross-attention distribution of the model is further illustrated. In the attention of temporal features to spatial features, Node1, Node0, and Node9 received higher weights, indicating that the model pays more attention to the spatial information contained in these nodes within the current prediction window. In the attention of spatial features to temporal segments, p-10 and p-9 had the highest average weights, p-3 and p-4 also showed relatively high attention, while p-7 had the lowest weight. This suggests that the model does not rely solely on short-term information near the prediction time, but can dynamically integrate historical features from different time scales based on the input. These visualizations provide intuitive evidence for the asymmetric spatial dependence and multi-scale temporal attention mechanism learned by the model.

[0044] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A regional photovoltaic power prediction method using a decoupled asymmetric spatiotemporal graph fusion Transformer, characterized in that, Includes the following steps: Step S1: Obtain historical photovoltaic power time series data of multiple photovoltaic sites in the region, construct input samples according to continuous historical observation windows, and use a preprocessing method based on sample statistics decoupling and supporting reverse reconstruction to perform feature scaling on the input samples and save the sample statistics required for reverse reconstruction. Step S2: Input the same preprocessed historical observation window into the parallel time-dependent branch and spatial-dependent branch; in the time-dependent branch, divide the power time series of each photovoltaic site into multiple overlapping time patches, map each time patch into patch time features, and use the Transformer encoder to extract patch-level long-range time evolution features. In the spatially dependent branch, an explicit sample statistical adjacency matrix is ​​generated based on the Pearson correlation coefficient between photovoltaic sites within the same historical observation window, and an implicit learnable asymmetric adjacency matrix is ​​generated using an influence source mapping network and an influence target mapping network with mutually independent parameters. Step S3: Generate node-by-node pair gating weights based on the explicit sample statistical adjacency matrix and the implicit learnable asymmetric adjacency matrix. Perform edge-by-edge fusion of the two adjacency matrices according to the gating weights to obtain a sample-related asymmetric fused adjacency matrix. Use a graph convolutional network to aggregate the neighborhood information of each photovoltaic site according to the asymmetric fused adjacency matrix. After the graph convolutional network outputs, perform spatial self-attention calculation along the photovoltaic site axis to obtain a spatially enhanced representation that simultaneously contains local directed topological information and global site interaction information. Step S4: Utilize two sets of independent cross-attention networks to perform spatiotemporal information interaction in two directions. In the first cross-attention direction, the pooling time features of each photovoltaic site are used as queries, and the spatial enhancement representations of multiple photovoltaic sites are used as keys and values, enabling the pooling time features of each photovoltaic site to selectively absorb the spatial information of multiple photovoltaic sites. In the second cross-attention direction, the spatial enhancement representations of each photovoltaic site are used as queries, and the patch time features corresponding to that photovoltaic site are used as keys and values, enabling the spatial enhancement representations of each photovoltaic site to selectively absorb its corresponding historical time patch information. The two cross-attention directions respectively employ independent query projection parameters, key projection parameters, and value projection parameters, and combine residual connections and layer normalization operations to obtain enhanced time representations and enhanced spatial representations that maintain the correspondence between sites. Step S5: Map the enhanced temporal representation and the enhanced spatial representation to the same target prediction step space to obtain temporal prediction features and spatial prediction features; The temporal and spatial prediction features are spliced ​​and fused to generate multi-site photovoltaic power prediction results at a normalized scale. Using the sample statistics obtained and saved from the same historical observation window in step S1, the multi-site photovoltaic power prediction results are reverse-scaled to output the regional multi-site photovoltaic power prediction results at the original power scale.

2. The method according to claim 1, characterized in that, The normalization process in step S1 uses the following formula to decouple sample statistics and scale features: (1) (2) (3) in, B Indicates batch size, L This indicates the number of time steps in the photovoltaic power sequence within the historical observation window. N Indicates the number of photovoltaic power stations; x b,t,n Indicates the first b The first sample index n The site index is in t The original power value at time; μ b,n and σ b,n These represent the mean and standard deviation of the sample over the time dimension, respectively. A tiny constant to prevent division by zero errors; γ n and β n In order to target the n Learnable affine transformation scaling factor and bias term for each site index; The input data is scaled up to the feature size.

3. The method according to claim 1, characterized in that, Step S2 specifically involves: The time-dependent branches perform patch partitioning and patch embedding on the power time series of each photovoltaic site, use the Transformer encoder to generate patch time features that retain the patch axis, and pool the patch time features to generate pooled time features for each photovoltaic site. The spatially dependent branch generates an explicit sample statistical adjacency matrix based on the same historical observation window, and generates an implicit learnable asymmetric adjacency matrix based on learnable node embeddings. The explicit sample statistical adjacency matrix and the implicit learnable asymmetric adjacency matrix are fused by gating weights for each node pair. A graph convolutional network is used to aggregate local neighborhood information according to the fused adjacency matrix, and a spatial self-attention module is used to extract global site interaction information along the photovoltaic site axis to generate spatially enhanced representations of each photovoltaic site. The patch time feature, pooling time feature, and spatial enhancement representation respectively retain the index relationship of the corresponding photovoltaic site and serve as the input of the spatiotemporal bidirectional information interaction module. This allows the spatial dependency branch to serve as the spatial feature extraction step in the complete prediction process, and to work in conjunction with the time dependency branch, the spatiotemporal bidirectional information interaction module, and the prediction mapping module to generate photovoltaic power prediction results.

4. The method according to claim 1, characterized in that, The specific operations and calculation processes for the time-dependent branches and space-dependent branches in steps S2 and S3 are as follows: The patch length in the time-dependent branch is set to 16 to achieve a balance between reducing computational overhead and ensuring prediction accuracy. The time-dependent branch uses a sliding window mechanism to patch the sequence. L Divide into lengths P len A patch with a step size of S is then embedded and added to the positional code before being fed into a multi-layer Transformer encoder. The encoder output is then globally aggregated using adaptive average pooling. (4) (5) (6) (7) (8) Among them, P len S and S represent the patch length and step size of the sliding window, respectively. For the floor operation, P num Z represents the total number of patches after splitting; patch W is a patch embedding tensor mapped to feature dimension D. pos L is the position encoding matrix; enc Z represents the total number of encoder layers; (0) Z represents the initial input of the Transformer encoder obtained by adding the patch embedding tensor to the positional encoding. (l-1) Z ´(l) With Z (l) They represent the first l The input features of the layer encoder, the output features of the self-attention sublayer, and the output features of the feedforward neural network sublayer; MSA(·), FFN(·), and Layernorm(·) correspond to the multi-head self-attention mechanism, the feedforward neural network, and the layer normalization operation, respectively; Z (Lenc) For the output tensor of the last encoder layer, the Squeeze(·) operation is used to remove feature axes of size 1; AdaptiveAvgPool1d(·) represents the adaptive average pooling operation performed along the patch dimension. This indicates the pooling time characteristics of each site; The spatially dependent branch generates an implicitly learnable asymmetric adjacency matrix through learnable node embeddings and parameter-independent influence source mapping and influence target mapping networks. It then calculates the Pearson correlation coefficient based on the preprocessed power sequences of each photovoltaic site within the current historical observation window, generating an explicit sample statistical adjacency matrix. Finally, it fuses the two adjacency matrices using gating weights and generates a spatially enhanced representation using a graph convolutional network and a spatial self-attention module. （9) (10) (11) （12） （13） （14） （15） （16） （17） （18） （19） （20） （21） （22） （23） （24） in, i and j Indicates a photovoltaic site index; E Represents learnable node embeddings, MLP src (·) and MLP tgt (·) represent the mapping networks between the influencing source and the influencing target, respectively. src and E tgt Let A and B represent the source features and target features respectively, T denote matrix transpose, and A... dyn Represents an implicit asymmetric dynamic adjacency matrix; Represents a node i In the t Preprocessing power values ​​at each time step and Representing nodes respectively i and nodes j The average preprocessed power within the historical observation window Represents a node i With nodes j The Pearson correlation coefficient between them, A stat This represents the explicit statistical adjacency matrix mapped to the [0,1] interval; Concat(·) represents feature concatenation, MLP gate (·) represents the gated weight generation network, W gate A represents the adaptive gated weights, ⊙ represents the Hadamard product, and A final This represents the final adjacency matrix obtained by fusing the explicit statistical graph and the implicit dynamic graph; I N L represents the identity matrix. gcn This indicates the total number of convolutional layers in the graph. W represents the input features of the spatial branch. emb The linear embedding weight matrix represents the spatial input features. D represents the initial spatial characteristics. in Represents an in-degree diagonal matrix. Indicates the first r The weight matrix of the layer graph convolution, where ф(·) represents the activation function. Indicates the first r Layered graph convolution output; W Q W K and W V Q represents the projection matrices of the query, key, and value, respectively. spa K spa and V spa These represent the corresponding query, key, and value characteristics, respectively, d K A represents the feature dimension of the attention head. spa W represents the spatial self-attention output. O H represents the output projection matrix. enh This represents the final spatial augmentation representation.

5. The method according to claim 1, characterized in that, The spatiotemporal bidirectional information interaction module in step S4 includes a first cross-attention direction and a second cross-attention direction. The two cross-attention directions use independent query projection parameters, key projection parameters, and value projection parameters. In the first cross-attention direction, the pooling time characteristics of each photovoltaic site are used as the query, and the spatial enhancement representations of multiple photovoltaic sites are used as the key and value. This allows the pooling time characteristics of each photovoltaic site to selectively absorb the spatial information of multiple photovoltaic sites according to the attention weights. The calculation formula is as follows: (25) (26) (27) (28) (29) In the second cross-attention direction, the spatial enhancement representation of each photovoltaic site is used as the query, and the corresponding temporal feature of the photovoltaic site is used as the key and value. The spatial enhancement representation of each photovoltaic site selectively absorbs the historical temporal information of that photovoltaic site according to the attention weight. The calculation formula is as follows: (30) （31） （32） （33） （34） in, This indicates the pooling time characteristics of each photovoltaic site; This indicates that the patch time characteristics of the photovoltaic site axis and patch axis are retained; Indicates the first n Patch time characteristics of individual photovoltaic sites; H enh Representation of spatially enhanced characterization; H enh,n Indicates the first n Spatial enhancement characterization of individual photovoltaic sites; , , These represent the query projection matrix, key projection matrix, and value projection matrix in the first cross-attention direction, respectively; Q time , and These represent the query matrix, key matrix, and value matrix corresponding to the first cross-attention direction, respectively. Time cross This represents the cross-attention output calculated using pooled temporal features as queries and spatially augmented representations as keys and values. This represents the enhanced temporal representation obtained after injecting spatial information. , and These represent the query projection matrix, key projection matrix, and value projection matrix in the second cross-attention direction, respectively. K time,n and V time,n They represent the first n The query matrix, key matrix, and value matrix corresponding to each photovoltaic site in the second cross-attention direction; Space cross,n Indicates the first n The cross-attention output of each photovoltaic site is calculated using spatial enhancement representation as query and corresponding patch time features as key and value. Indicates the first n The enhanced spatial representation obtained after injecting time patch information into each photovoltaic site; divided by Used to control the numerical scale of the dot product result, improving the numerical stability of attention calculation.

6. The method according to claim 1, characterized in that, The multidimensional aggregation and target space mapping process in step S5 is as follows: Enhanced temporal features Spatial features F is obtained by mapping to the target prediction step space through independent linear projection layers. time With F space ; Concatenating the features along the feature dimension yields the cascaded feature H. concat Multi-view information fusion is performed using the output mapping layer, and the calculation formula is as follows: (35) (36) (37) (38) (39) in, and These represent the weight matrices of the temporal feature projection layer and the spatial feature projection layer, respectively. and F represents the bias vectors of the temporal feature projection layer and the spatial feature projection layer, respectively; time and F space H represents the temporal and spatial prediction features mapped to the target prediction length space, respectively; concat This represents the cascaded feature obtained by concatenating the two elements; W fusion and b fusion Y represents the weight matrix and bias vector of the final fusion layer, respectively; pred This represents the normalized scaling prediction result output by the fusion layer; Denorm(·) represents the RevIN inverse transform operation performed using the sample statistics and learnable affine parameters saved in step S1. This represents the final photovoltaic power prediction result after restoring to the original power scale.

7. The method according to claim 6, characterized in that, The parameters for evaluating the model's performance are as follows: During training, Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE) were selected as performance evaluation metrics; the calculation formulas are as follows: (40) (41) (42) Among them, y m Represents the true value. This represents the predicted value from the corresponding model.

Images