Regional distributed photovoltaic power prediction method and system based on multivariate data cross-modal fusion

Abstract

The invention relates to a regional distributed photovoltaic power prediction method and system based on multivariate data cross-modal fusion, and belongs to the technical field of photovoltaic power prediction. Firstly, adaptive extraction of meteorological data spatial features and power data spatial heterogeneity is realized through a dynamic perception convolution kernel of AG-CNN, and a high-quality spatial basis is provided for cross-modal fusion; secondly, utilizing a two-dimensional dynamic attention mechanism of a DR-Transform to deeply mine cross-modal association of'historical power time sequence-future weather driving ', adapting to fusion requirements in different scenes through dynamic weight distribution, and capturing association evolution in a long period; thirdly, designing a cross-modal feature bridging module; and finally, verifying the robustness of the model in a typical scene, and ensuring that the model meets the precision requirement of multi-scale scheduling of the power grid.

Description

TECHNICAL FIELD

[0001] The application relates to a regional distributed photovoltaic power prediction method and system based on multi-element data cross-modal fusion, and belongs to the technical field of photovoltaic power prediction. BACKGROUND

[0002] Current distributed photovoltaic power prediction research has formed multiple technical paths, but the existing methods are limited in both "power-weather cross-modal fusion" and "long-time sequence dependence modeling", which always restricts the improvement of prediction accuracy. From the physical model, it is based on the conversion mechanism of photovoltaic components, which can reflect the direct influence of weather factors on power, and has certain reliability in short-term prediction, but this kind of model depends on fixed component parameters and high-resolution weather data, which is difficult to describe the long-period cumulative effect of weather factors (such as cloud cover and temperature), and lacks dynamic cross-modal adaptation of power and weather data - only through simple correlation between the two by mechanism formula, which cannot meet the fusion needs in the case of weather mutation or data heterogeneity.

[0003] The limitation of statistical model is concentrated in the shallowization of cross-modal correlation and the insufficient nonlinear adaptation. This kind of model establishes correlation by digging the law of historical data, which can realize basic prediction when the data distribution is stable, but it cannot capture the nonlinear relationship between weather and power, and the cross-modal information utilization is limited to shallow statistical correlation (such as simple linear regression); when the modal difference between weather data and power data increases (such as the sharp drop of irradiance under rainy weather leading to the intensification of power fluctuation), the model will quickly distort the correlation between "weather driving-power response", and the accuracy will decrease sharply.

[0004] These limitations are not isolated, but are intertwined and progressive around the two core problems of cross-modal fusion and long-time sequence dependence, which together constitute the modeling difficulty of distributed photovoltaic power prediction - the former determines whether weather and power data can be effectively coordinated, the latter affects whether this coordination relationship can be stably transmitted in the time dimension, and the coupling of the two further increases the modeling complexity.

[0005] The above challenges are superimposed, and finally point to the top-level architecture bottleneck of the lack of time-space-cross-modal collaborative modeling capability. On the one hand, most models separate spatial correlation and cross-modal fusion, for example, some spatio-temporal models only capture the spatial correlation of power at the station level, without considering the spatial distribution of weather; on the other hand, traditional time sequence architecture is difficult to bear long-period cross-modal correlation transmission: LSTM memory unit decays with time, which is easy to lose early key weather features, and Transformer attention is easy to scatter under long sequence, which is difficult to focus on the core period of "weather trend-power trend". This lack of coordination makes it difficult for existing models to balance regional differences and long-time coupling, which seriously restricts the generalization ability and robustness of the prediction model under complex weather conditions. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a regional distributed photovoltaic power prediction method and system based on multi-element data cross-modal fusion, which takes distributed photovoltaic accurate power prediction as the core target, focuses on the synergistic breakthrough of power-weather cross-modal deep fusion and long-time sequence dependence accurate capture, combines the technical characteristics of adaptive spatial convolution (AG-CNN) and dynamic correlation Transformer (DR-Transformer), and constructs an end-to-end combined prediction framework. First, through the dynamic perception convolution kernel of AG-CNN, the spatial characteristics of meteorological data and the spatial heterogeneity of power data are adaptively extracted, which not only preserves the continuous structure of meteorological data, but also conforms to the characteristics of distributed photovoltaic "many points and wide surfaces", providing a high-quality spatial basis for cross-modal fusion. Secondly, the dual-dimensional dynamic attention mechanism of DR-Transformer is used to deeply mine the cross-modal correlation of "historical power time sequence-future weather driving", and through dynamic weight distribution, the fusion requirements in different scenarios are adapted, and the correlation evolution in a long period is captured. Thirdly, a cross-modal feature bridging module is designed to solve the heterogeneity of convolution features output by AG-CNN and sequence features input by DR-Transformer in dimension and semantics, realizing seamless collaboration of spatial features and cross-modal time sequence features. Finally, the robustness of the model is verified in typical scenarios to ensure that it meets the accuracy requirements of power grid multi-scale scheduling. The present application aims to break through the dual limitations of existing models in cross-modal fusion and long-time sequence modeling, provide a more reliable technical solution for distributed photovoltaic power prediction, and support the safe and economic operation of high-penetration photovoltaic grid-connected power grids, with both theoretical depth and engineering practical value.

[0007] The technical scheme of the present application is as follows: The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion comprises the following steps: (1) Spatiotemporal data gridding consistency processing, collecting original meteorological data and historical power data, and performing data standardization, and then taking a unified geographic coordinate as a reference to realize the spatial correlation of meteorological data and historical power data; (2) AG-CNN spatial feature based on dynamic perception convolution for adaptive extraction of data, AG-CNN relies on a unified basic spatial reference to realize fine local modeling through dynamic feature perception and cross-regional information fusion, and its core is to construct a feature extraction mechanism with spatial consistency and flexible adaptability, which can not only ensure the natural correlation of meteorological and power data in the spatial dimension, but also accurately capture the feature mode differences of different regions; (3) Cross-modal feature bridging module design: Through a three-level processing flow of adaptive projection-feature reparameterization-spatiotemporal fusion gating, the high-dimensional local features output by AG-CNN are transformed into a spatiotemporal token sequence that can be parsed by DR-Transformer, while retaining key spatial details and cross-regional correlation information; (4) DR-Transformer time-series correlation modeling with dual-dimensional dynamic attention. DR-Transformer addresses the core limitations of existing models such as "insufficient capture of long-term time-series dependencies" and "loose coupling of cross-modal features". It takes the spatial dimension + temporal dimension dynamic attention mechanism as the core, deeply integrates the long-term time-series dependency of historical power with the driving role of future meteorological data, realizes the accurate coupling of meteorological-power cross-modal spatiotemporal features, and obtains the total power prediction result of distributed photovoltaic system. Its design goal is to solve the problems of "transmission of meteorological uncertainty" and "attenuation of long-period time-series correlation", so that the model can capture the long-term evolution law of historical power and accurately respond to the dynamic driving of future meteorological conditions.

[0008] According to a preferred embodiment of the present invention, in step (1), the meteorological data includes solar irradiance, temperature, and humidity, etc. The data standardization process processes meteorological data and historical power data separately to eliminate the influence of different characteristic units. The Z-score standardization method is used, and the formula is as follows: (1) Where x represents the original data, μ represents the data mean, and σ represents the standard deviation. This is the standardized data.

[0009] Spatial correlation: Combining the coverage of photovoltaic arrays and meteorological monitoring points, and based on the spatial scale of meteorological data, the geographical location of photovoltaic panels is matched with the corresponding meteorological regions, so that power data and meteorological data form a gridded spatial correlation. At the same time, the data is time-aligned according to the time axis of prediction granularity, and abnormal data is removed, forming a spatiotemporally aligned input structure, which lays the foundation for subsequent models to capture spatiotemporal correlations.

[0010] According to a preferred embodiment of the present invention, in step (2), the core design of AG-CNN is dynamic perceptual convolutional kernel and adaptive region fusion: with grid cells as the basic spatial index, the convolutional kernel of AG-CNN adapts in real time to the station distribution density and feature fluctuation pattern in different grid cells through learnable parameters. For example, in grid cells with dense stations, the kernel function automatically focuses on local details (irradiance gradient, cloud shadow edge and other subtle features) in the cell; in grid cells with sparse stations, the kernel function flexibly extends to adjacent grid cells to capture a wider range of meteorological trends (such as cloud movement at the regional scale). This dynamic adjustment mechanism enables the model to naturally fit the actual spatial feature distribution and avoid mechanical dependence on fixed grid boundaries. The convolutional kernel weights of AG-CNN are constructed collaboratively by a Gaussian function offset by spatial relative positions and learnable parameters. (2) Where (i, j) represents the spatial relative position offset, such as the offset of the surrounding area centered on the target area, i is the horizontal spatial relative position offset, j is the vertical spatial relative position offset, and σ g θ is the Gaussian kernel bandwidth, used to control the decay rate of the spatial influence range; s(i,j) is the feature similarity between the offset region and the target region, including information such as the consistency of power fluctuation patterns and irradiance change trends; θ is a learnable parameter, α=5, used to balance the influence of spatial location correlation and feature pattern similarity. When performing convolution operations on the feature map, the output features will adaptively fuse information from the target region and surrounding related regions. (3) Among them, y x，y To output the feature values ​​of the target region in the feature map, x x+i，y+j For features of surrounding related regions in the input feature map, such as irradiance, temperature, and average power within the region, the model can maintain natural spatial continuity and capture feature relationships between different regions through dynamic weights w(i,j), providing a feature foundation that combines spatial consistency and local detail for subsequent global modeling.

[0011] The AG-CNN network architecture employs a multi-layer convolutional block cascade design: the first layer captures instantaneous irradiance abrupt changes within a region using a dynamic kernel; the second layer extracts cross-regional meteorological propagation patterns using a dynamic kernel; and the third layer fuses features from the first two layers through convolution, enhancing key details. Each convolutional layer undergoes batch normalization and activation processing to enhance feature discriminative power and mitigate the gradient vanishing problem. This design allows the convolutional kernel to dynamically adjust its activation intensity based on regional features as it slides, effectively addressing the difficulty of adapting fixed kernels to complex spatial distributions.

[0012] According to a preferred embodiment of the present invention, in step (3), the cross-modal feature bridging module (CMFB) adapts the AG-CNN features to the DR-Transformer input. The feature map output by AG-CNN is first flattened into a one-dimensional sequence according to the region index, and then the channel dimension is mapped to the hidden dimension of DR-Transformer through linear projection to form an initial token sequence. To balance the preservation of local features and the adaptability of global modeling, CMFB dynamically adjusts the feature weights through a gating mechanism: for regions with high photovoltaic capacity and large power fluctuations, the feature weights are strengthened; for regions without sites or with stable features, noise interference is weakened. At the same time, the temporal features of historical power are introduced, and the spatial features and temporal context are explicitly fused through the spatiotemporal gating unit (STGU) to generate a token sequence that carries both regional details and temporal correlation, ensuring a smooth transition of features from the convolutional feature map to the Transformer sequence.

[0013] The feature map F output by AG-CNN is a four-dimensional tensor: (4) Where B is the training batch size, C is the number of feature channels, and H×W corresponds to the number of region divisions; First, the two-dimensional region structure is transformed into a one-dimensional sequence through spatial flattening, and the channel dimension is adapted to the hidden dimension of the Transformer through learnable linear transformation: (5) in, The flattened sequence features are represented by N = H × W, where N is the total number of regions. For the projection matrix, Here, D is the bias term and D is the hidden dimension of the Transformer. This step compresses the local features of each region into tokens of fixed dimensions, forming a sequence of length N, which provides the basic unit for global modeling. Since the optimization objectives of AG-CNN and DR-Transformer differ, a reparameterization layer is introduced to dynamically adjust the feature distribution to alleviate training bias: (6) in, For element-wise gating coefficients, It is the Sigmoid activation function. , As learnable parameters, the gating mechanism achieves dynamic alignment of feature distributions between the two models by strengthening key features and suppressing noise; To display the temporal features of fused historical power and the spatial features of AG-CNN, a spatiotemporal gating unit (STGU) is introduced, and the correlation weights between the two are dynamically calibrated through cross-modal attention: Let the historical power sequence be , among which, T h Let R be the historical time step, B be the training batch size, N be the total number of regions, and D be the Transformer hidden dimension. Then the gated fusion process is as follows: (7) in, F is the dynamic fusion coefficient. in The output features of the Space-Time Gated Unit (STGU) carry both spatial details and temporal correlations. The STGU is a space-time gated unit used to fuse spatial and temporal features. α For the cross-modal attention weight matrix, b α For cross-modal attention bias term, F reparam The reparameterized spatial features output by equation (6), As a cross-modal attention mechanism, it achieves semantic-level calibration of spatial and temporal features by learning the association weights between them.

[0014] Final output Directly used as input to the DR-Transformer encoder, each token retains both the local spatial details extracted by AG-CNN and incorporates the temporal context of historical power, providing a high-quality feature foundation for the subsequent dynamic attention mechanism to accurately capture the dual dependence of "spatial correlation-temporal evolution".

[0015] According to a preferred embodiment of the present invention, in step (4), the DR-Transformer serves as the core of global spatiotemporal fusion and realizes cross-modal correlation modeling of historical power and future weather through an encoder-decoder structure; The encoder receives the sequence output from the cross-modal feature bridging module, processes historical power and spatial features, and captures implicit correlations and long-term temporal dependencies between regions through dynamic spatial attention. It employs a multi-layered stacked structure: the bottom layer introduces a dynamic window partitioning strategy, strengthening the temporal correlation of local regions through attention calculation within the window; the upper layer uses global attention to capture long-distance temporal dependencies. Each layer contains multi-head dynamic spatial attention and a feedforward neural network, where the attention weights are modulated by both the relative position of the regions and the similarity of power features. For example, for regions with high historical power correlation coefficients, the attention ratio between them is automatically increased.

[0016] The decoder processes meteorological data for future time steps. Under the constraint of the masked self-attention mechanism, it dynamically couples meteorological features with historical features output by the encoder through cross-attention. It adopts a multi-layer stacked structure, with each layer consisting of multi-head masked self-attention, multi-head cross-attention, and a feedforward neural network: masked self-attention ensures the temporal causality of the meteorological sequence; cross-attention adjusts the weights through dynamic correlation coefficients, for example, strengthening the attention ratio of irradiance during noon and increasing the driving weight of cloud cover during cloudy periods. After mapping the encoder and decoder features to the same dimension through a linear projection layer, cross-attention achieves a precise correlation between historical power trends and future meteorological drivers, outputting features that fuse global spatiotemporal information. Finally, the output layer converts the fused features into regional power prediction values ​​for multiple future time steps through a fully connected layer. Then, based on the photovoltaic installed capacity of each region, the values ​​are weighted and summed to obtain the total power prediction result of the distributed photovoltaic system, completing the entire process from input data to prediction output.

[0017] According to a preferred embodiment of the present invention, the core design of DR-Transformer focuses on a dynamic attention mechanism in both spatial and temporal dimensions: in the spatial dimension, attention weights are adaptively adjusted based on the similarity of regional features and relative positional relationships to avoid mechanical dependence on fixed spatial units; in the temporal dimension, the evolution of historical power and the driving effect of future weather are dynamically coupled through a cross-attention mechanism to achieve deep fusion of cross-modal features. Its design focus is not on following a fixed structure, but on how to make the correlation modeling of spatial and temporal features more consistent with the actual fluctuation patterns of distributed photovoltaics through the dynamic adjustment of the attention mechanism. To address the discrete distribution characteristics of distributed photovoltaic (PV) sites, the DR-Transformer encoder captures implicit correlations between regions through dynamic spatial attention. The attention weight calculation considers both relative spatial distance and feature pattern similarity. The formula for calculating the dynamic spatial attention weight is as follows: (8) Where q, k, and v are the query, key, and value vectors, respectively, and d k For feature dimension, β(r) is the feature similarity coefficient (calculated based on the cosine similarity of the regional power sequence, with a value of 0-1), and β(r) is the relative spatial distance coefficient, which is generated based on the Gaussian decay function and reflects the spatial correlation between regions. Through the coordinated modulation of the two, the attention weight can not only reflect the natural continuity in space, but also focus on the related regions with similar features, avoiding the correlation fragmentation caused by fixed spatial division. The DR-Transformer encoder dynamically fuses future meteorological data with historical features from the encoder output, achieving cross-modal correlation through masked self-attention and cross-attention. Future meteorological data has a natural time-series dependency, but directly using self-attention may lead to future information being focused on. Masked self-attention forces the model to follow temporal causality through the lower triangular matrix, allowing only the current time step to focus on historical and concurrent meteorological characteristics. Let the future weather sequence be the input to the decoder. T represents the number of future prediction steps, D represents the dimension of the meteorological feature vector, R represents the real number field, and m t Given the meteorological feature vector at step t, a query is generated through linear projection. ,key ,value , , , , where is the meteorological feature projection matrix, and h is the number of attention heads; For time step t, the similarity of the mask self-attention is calculated as follows: (9) in, For the time step that is being noticed, This is a scaling factor to prevent the softmax gradient from vanishing due to excessively large inner product values. Introducing a lower triangular mask matrix Suppressing future information through masking: (10) In the formula, This represents the t-th row and the th element in the mask matrix. The elements of the column, MaskedSim is the attention similarity score after masking; Final The masked self-attention output MaskedAttn(t) of the step is: (11) in, The output projection matrix is ​​given by v, which represents the value vector. softmax(·) is the normalization exponential function, and Concat(·) represents the multi-head attention concatenation operation. After concatenation and projection, the meteorological features containing time-dependent characteristics are obtained from the multi-head attention results. .

[0018] According to a preferred embodiment of the present invention, in distributed photovoltaic data, the contribution of different regions to the total power varies significantly, and the power prediction needs to meet the requirements of overall fitting accuracy and spatiotemporal correlation. A single loss function is difficult to take into account these requirements. Therefore, a multi-objective loss function is designed to collaboratively optimize the model parameters, which not only retains the role of spatial benchmark in integrating discrete data, but also enhances the model's ability to capture dynamic meteorological-power correlation.

[0019] As the core foundation of the loss function, it is necessary to prioritize ensuring the model's fitting accuracy to the overall power level. Considering that photovoltaic power prediction is a continuous value regression problem, and that the penalty for larger errors is more stringent in practical applications, the squared error penalty characteristic of MSE can amplify the impact of extreme biases, forcing the model to focus on reducing large power errors. At the same time, its continuous differentiability is compatible with the model's gradient descent optimization mechanism, ensuring the stability of the training process. Therefore, the method constructs a multi-objective loss function oriented towards regional characteristics and cross-modal constraints: The mean squared error (MSE) is used as the basic loss term, and the formula is as follows: (12) Where N is the total number of regions. To predict the time step, and Let be the predicted power and the actual power of the i-th region at time t, respectively; AG-CNN extracts local features through dynamic kernels and cross-regional fusion, but its output regional features may suffer from fragmented correlations between adjacent regions due to the randomness of dynamic neighborhoods. The core of the regional feature consistency loss is to force the model to respect the implicit correlations between regions in physical space, ensuring that feature extraction not only focuses on the local details of a single region, but also captures the collaborative changes in adjacent regions caused by common meteorological disturbances. (13) Where, N i For the dynamic neighborhood of region i, w i，j The feature similarity weights for regions i and j are as follows: the higher the similarity, the greater the weight. This term compensates for the spatial correlation fragmentation problem caused by fixed grid division by constraining the consistency of power differences between adjacent regions. The core of DR-Transformer is to fuse historical power time-series features with future meteorological driving features through cross-attention. However, the model may exhibit prediction bias due to an unreasonable allocation of attention weights, resulting in a "disconnect between meteorological and power trends." The goal of cross-modal time-series correlation loss is to force the model to learn the physical causal relationship between meteorological factors and power changes, ensuring that cross-modal correlations in the time-series dimension conform to objective laws, rather than merely fitting the surface correlation of data. It regulates the model's correlation learning from the physical mechanism level through dual constraints: trend consistency + reasonable attention weights. (14) in, As a trend consistency loss function, it is used to constrain the physical rationality of the direction of change. β is the weighting coefficient of the trend consistency loss, used to balance the strength of trend constraint and attention weight constraint. (15) in, To predict the time-series variation of power, This represents the temporal variation of meteorological factors. For cross-entropy loss, a high penalty is imposed for situations where the power trend is opposite to the meteorological trend; As a cross-modal attention loss function, it is used to constrain the objective correlation of attention weights: (16) in, In the DR-Transformer cross-attention, meteorological features affect the region. exist Attention weight at any moment This is the true correlation coefficient between meteorology and power calculated based on historical data.

[0020] The overall loss function is the weighted sum of the three losses mentioned above: (17).

[0021] A regional distributed photovoltaic power prediction system based on multi-data cross-modal fusion includes: The data processing module is used for spatiotemporal data gridding consistency processing. It collects raw meteorological data and historical power data, performs data standardization, and then uses unified geographic coordinates as a benchmark to realize the spatial correlation between meteorological data and historical power data. The extraction module adaptively extracts data based on the spatial features of AG-CNN with dynamic perceptual convolution. AG-CNN relies on a unified basic spatial benchmark and achieves local modeling through dynamic feature perception and cross-regional information fusion. The design module is used for the design of cross-modal feature bridging module. Through a three-level processing flow of adaptive projection-feature reparameterization-spatiotemporal fusion gating, the high-dimensional local features output by AG-CNN are transformed into a spatiotemporal token sequence that can be parsed by DR-Transformer, while retaining key spatial details and cross-regional correlation information. The prediction module is used to incorporate the DR-Transformer temporal correlation modeling with dual-dimensional dynamic attention. The DR-Transformer takes the spatial dimension + temporal dimension dynamic attention mechanism as its core, deeply integrates the long-term temporal dependence of historical power and the driving role of future meteorological data, realizes the accurate coupling of meteorological and power cross-modal spatiotemporal features, and obtains the total power prediction result of the distributed photovoltaic system.

[0022] The beneficial effects of this invention are as follows: 1. This invention focuses on "precise coupling of cross-modal spatiotemporal features" and "deep capture of long-term time-series dependencies" as its core objectives. It relies on an implicit spatial benchmark (gridized meteorological-power correlation units) to associate meteorological data with photovoltaic locations. It combines AG-CNN's adaptive extraction capability of "local meteorological-power spatial features" with DR-Transformer's cross-modal fusion advantage of "long-term historical power data - future meteorological drivers" to deeply explore the dynamic coupling relationship between "meteorological disturbances and power fluctuations." This approach ensures spatial consistency of data through the implicit spatial benchmark while overcoming the limitations of fixed structures through dynamic mechanisms, adapting to the characteristics of distributed photovoltaic systems—"numerous points, wide distribution, and significant susceptibility to micro-meteorological disturbances"—and specifically addresses the challenges faced by existing models.

[0023] 2. The dynamic region modeling of AG-CNN in this invention provides a refined spatial foundation for global fusion. The cross-modal feature bridging module solves the heterogeneous adaptation problem between convolutional features and Transformer sequences, while DR-Transformer enables in-depth mining of cross-modal spatiotemporal correlations. The synergy of these three components allows the model to retain the power heterogeneity of different regions while accurately capturing the dynamic correlation between meteorological drivers and power time series, maintaining stable prediction performance even in complex meteorological scenarios such as cloudy and rainy weather. Attached Figure Description

[0024] Figure 1 This is a flowchart of the overall AG-CNN process of the present invention; Figure 2 This is a schematic diagram of the DR-Transformer part of the present invention; Figure 3 This is a diagram of the overall architecture of the DR-Transformer of the present invention; Figure 4 This is a flowchart of the cross-modal feature bridging module (CMFB) of the present invention; Figure 5 This is a flowchart of the method of the present invention; Figure 6 This is a diagram showing the relative location distribution of photovoltaic power stations according to an embodiment of the present invention; Figure 7 This is a comparison chart of power prediction results for different models in a photovoltaic power station according to an embodiment of the present invention. Figure 7 The middle (A) graph is a comparison of the power prediction results of different models in station A; Figure 7 The middle (B) graph is a comparison of the power prediction results of different models in station B; Figure 7 The middle (C) graph is a comparison of the power prediction results of different models in station C; Figure 7 The middle (D) graph is a comparison of the power prediction results of different models in station D; Figure 7 The middle (E) is a comparison chart of the power prediction results of different models in station E; Figure 8 This is a comparison chart of error indices of different prediction models for various stations according to embodiments of the present invention, wherein, Figure 8 (A) is a comparison chart of error indices for model M1; Figure 8 (B) is a comparison chart of the error indices of model M2; Figure 8 The middle (C) section is a comparison chart of the error indices of model M3; Figure 8 The middle (D) graph is a comparison of the error indices of model M4; Figure 9 Box plots comparing the prediction effects of different models at various stations according to embodiments of the present invention, wherein, Figure 9 (A) is a box plot comparing the prediction effects of different models for station A; Figure 9 (B) is a box plot comparing the prediction effects of different models for station B; Figure 9 The middle (C) is a box plot comparing the prediction effects of different models for station C; Figure 9 The middle (D) is a box plot comparing the prediction effects of different models for station D; Figure 9 The middle (E) is a box plot comparing the prediction effects of different models for station E. Detailed Implementation

[0025] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0026] Example 1: like Figures 1-5 As shown, this embodiment provides a regional distributed photovoltaic power prediction method based on multi-data cross-modal fusion, the steps of which are as follows: (1) Spatiotemporal data gridding consistency processing: collect the original meteorological data and historical power data, and standardize the data. Then, based on the unified geographic coordinates, realize the spatial correlation between meteorological data and historical power data. Meteorological data includes solar irradiance, temperature, and humidity; The data standardization process processes meteorological data and historical power data separately to eliminate the influence of different characteristic units. The Z-score standardization method is used, and the formula is as follows: (1) Where x represents the original data, μ represents the data mean, and σ represents the standard deviation. This is the standardized data.

[0027] Spatial correlation: Combining the coverage of photovoltaic arrays and meteorological monitoring points, and based on the spatial scale of meteorological data, the geographical location of photovoltaic panels is matched with the corresponding meteorological regions, so that power data and meteorological data form a gridded spatial correlation. At the same time, the data is time-aligned according to the time axis of prediction granularity, and abnormal data is removed, forming a spatiotemporally aligned input structure, which lays the foundation for subsequent models to capture spatiotemporal correlations.

[0028] (2) Based on the spatial features of AG-CNN with dynamic perception convolution, the data is adaptively extracted. AG-CNN relies on a unified basic spatial benchmark and achieves refined local modeling through dynamic feature perception and cross-regional information fusion. Its core is to build a feature extraction mechanism that has both spatial consistency and flexible adaptability. It can ensure the natural correlation between meteorological and power data in the spatial dimension and accurately capture the feature pattern differences in different regions. The core design of AG-CNN is dynamic perceptive convolutional kernels and adaptive region fusion: using grid cells as the basic spatial index, the convolutional kernels of AG-CNN adapt in real time to the station distribution density and feature fluctuation patterns within different grid cells through learnable parameters. For example, in grid cells with dense stations, the kernel function automatically focuses on local details within the cell (subtle features such as irradiance gradient and cloud shadow edges); in grid cells with sparse stations, the kernel function flexibly extends to adjacent grid cells to capture a wider range of meteorological trends (such as cloud movement at the regional scale). This dynamic adjustment mechanism allows the model to naturally fit the actual spatial feature distribution and avoid mechanical dependence on fixed grid boundaries. The convolutional kernel weights of AG-CNN are constructed collaboratively by a Gaussian function offset by spatial relative positions and learnable parameters. (2) Where (i, j) represents the spatial relative position offset, such as the offset of the surrounding area centered on the target area, i is the horizontal spatial relative position offset, j is the vertical spatial relative position offset, and σ g θ is the Gaussian kernel bandwidth, used to control the decay rate of the spatial influence range; s(i,j) is the feature similarity between the offset region and the target region, including information such as the consistency of power fluctuation patterns and irradiance change trends; θ is a learnable parameter, α=5, used to balance the influence of spatial location correlation and feature pattern similarity. When performing convolution operations on the feature map, the output features will adaptively fuse information from the target region and surrounding related regions. (3) Among them, y x，y To output the feature values ​​of the target region in the feature map, x x+i，y+jFor features of surrounding related regions in the input feature map, such as irradiance, temperature, and average power within the region, the model can maintain natural spatial continuity and capture feature relationships between different regions through dynamic weights w(i,j), providing a feature foundation that combines spatial consistency and local detail for subsequent global modeling.

[0029] (3) Cross-modal feature bridging module design: Through a three-level processing flow of adaptive projection-feature reparameterization-spatiotemporal fusion gating, the high-dimensional local features output by AG-CNN are transformed into a spatiotemporal token sequence that can be parsed by DR-Transformer, while retaining key spatial details and cross-regional correlation information; The Cross-Modal Feature Bridging Module (CMFB) adapts AG-CNN features to the input of the DR-Transformer. The feature map output by AG-CNN is first flattened into a one-dimensional sequence according to the region index, and then the channel dimension is mapped to the hidden dimension of the DR-Transformer through linear projection to form the initial token sequence. To balance the preservation of local features with the adaptability of global modeling, CMFB dynamically adjusts the feature weights through a gating mechanism: for regions with high photovoltaic capacity and large power fluctuations, the feature weights are strengthened; for regions without sites or with stable features, noise interference is weakened. At the same time, the temporal features of historical power are introduced, and the spatial features and temporal context are explicitly fused through the Spatiotemporal Gating Unit (STGU) to generate a token sequence that carries both regional details and temporal correlation, ensuring a smooth transition of features from the convolutional feature map to the Transformer sequence.

[0030] The feature map F output by AG-CNN is a four-dimensional tensor: (4) Where B is the training batch size, C is the number of feature channels, and H×W corresponds to the number of region divisions; First, the two-dimensional region structure is transformed into a one-dimensional sequence through spatial flattening, and the channel dimension is adapted to the hidden dimension of the Transformer through learnable linear transformation: (5) in, The flattened sequence features are represented by N = H × W, where N is the total number of regions. For the projection matrix, Here, D is the bias term and D is the hidden dimension of the Transformer. This step compresses the local features of each region into tokens of fixed dimensions, forming a sequence of length N, which provides the basic unit for global modeling. Since the optimization objectives of AG-CNN and DR-Transformer differ, a reparameterization layer is introduced to dynamically adjust the feature distribution to alleviate training bias: (6) in, For element-wise gating coefficients, It is the Sigmoid activation function. , As learnable parameters, the gating mechanism achieves dynamic alignment of feature distributions between the two models by strengthening key features and suppressing noise; To display the temporal features of fused historical power and the spatial features of AG-CNN, a spatiotemporal gating unit (STGU) is introduced, and the correlation weights between the two are dynamically calibrated through cross-modal attention: Let the historical power sequence be , among which, T h Let R be the historical time step, B be the training batch size, N be the total number of regions, and D be the Transformer hidden dimension. Then the gated fusion process is as follows: (7) in, F is the dynamic fusion coefficient. in The output features of the Space-Time Gated Unit (STGU) carry both spatial details and temporal correlations. The STGU is a space-time gated unit used to fuse spatial and temporal features. α For the cross-modal attention weight matrix, b α For cross-modal attention bias term, F reparam The reparameterized spatial features output by equation (6), As a cross-modal attention mechanism, it achieves semantic-level calibration of spatial and temporal features by learning the association weights between them.

[0031] Final output Directly used as input to the DR-Transformer encoder, each token retains both the local spatial details extracted by AG-CNN and incorporates the temporal context of historical power, providing a high-quality feature foundation for the subsequent dynamic attention mechanism to accurately capture the dual dependence of "spatial correlation-temporal evolution".

[0032] (4) DR-Transformer, as the core of global spatiotemporal fusion, realizes cross-modal correlation modeling of historical power and future weather through encoder-decoder structure; The encoder receives the sequence output from the cross-modal feature bridging module, processes historical power and spatial features, and captures implicit correlations and long-term temporal dependencies between regions through dynamic spatial attention. It employs a multi-layered stacked structure: the bottom layer introduces a dynamic window partitioning strategy, strengthening the temporal correlation of local regions through attention calculation within the window; the upper layer uses global attention to capture long-distance temporal dependencies. Each layer contains multi-head dynamic spatial attention and a feedforward neural network, where the attention weights are modulated by both the relative position of the regions and the similarity of power features. For example, for regions with high historical power correlation coefficients, the attention ratio between them is automatically increased.

[0033] The decoder processes meteorological data for future time steps. Under the constraint of the masked self-attention mechanism, it dynamically couples meteorological features with historical features output by the encoder through cross-attention. It adopts a multi-layer stacked structure, with each layer consisting of multi-head masked self-attention, multi-head cross-attention, and a feedforward neural network: masked self-attention ensures the temporal causality of the meteorological sequence; cross-attention adjusts the weights through dynamic correlation coefficients, for example, strengthening the attention ratio of irradiance during noon and increasing the driving weight of cloud cover during cloudy periods. After mapping the encoder and decoder features to the same dimension through a linear projection layer, cross-attention achieves a precise correlation between historical power trends and future meteorological drivers, outputting features that fuse global spatiotemporal information. Finally, the output layer converts the fused features into regional power prediction values ​​for multiple future time steps through a fully connected layer. Then, based on the photovoltaic installed capacity of each region, the values ​​are weighted and summed to obtain the total power prediction result of the distributed photovoltaic system, completing the entire process from input data to prediction output.

[0034] The core design of DR-Transformer focuses on a dynamic attention mechanism in both spatial and temporal dimensions: In the spatial dimension, attention weights are adaptively adjusted based on the similarity of regional features and their relative positional relationships to avoid mechanical dependence on fixed spatial units; in the temporal dimension, the evolution of historical power and the driving force of future weather are dynamically coupled through a cross-attention mechanism to achieve deep fusion of cross-modal features. Its design point is not about following a fixed structure, but about how to make the correlation modeling of spatial and temporal features more in line with the actual fluctuation patterns of distributed photovoltaics through the dynamic adjustment of the attention mechanism. To address the discrete distribution characteristics of distributed photovoltaic (PV) sites, the DR-Transformer encoder captures implicit correlations between regions through dynamic spatial attention. The attention weight calculation considers both relative spatial distance and feature pattern similarity. The formula for calculating the dynamic spatial attention weight is as follows: (8) Where q, k, and v are the query, key, and value vectors, respectively, and d k For feature dimension, β(r) is the feature similarity coefficient (calculated based on the cosine similarity of the regional power sequence, with a value of 0-1), and β(r) is the relative spatial distance coefficient, which is generated based on the Gaussian decay function and reflects the spatial correlation between regions. Through the coordinated modulation of the two, the attention weight can not only reflect the natural continuity in space, but also focus on the related regions with similar features, avoiding the correlation fragmentation caused by fixed spatial division. The DR-Transformer encoder dynamically fuses future meteorological data with historical features from the encoder output, achieving cross-modal correlation through masked self-attention and cross-attention. Future meteorological data has a natural time-series dependency, but directly using self-attention may lead to future information being focused on. Masked self-attention forces the model to follow temporal causality through the lower triangular matrix, allowing only the current time step to focus on historical and concurrent meteorological characteristics. Let the future weather sequence be the input to the decoder. T represents the number of future prediction steps, D represents the dimension of the meteorological feature vector, R represents the real number field, and m t Given the meteorological feature vector at step t, a query is generated through linear projection. ,key ,value , , , , where is the meteorological feature projection matrix, and h is the number of attention heads; For time step t, the similarity of the mask self-attention is calculated as follows: (9) in, For the time step that is being noticed, This is a scaling factor to prevent the softmax gradient from vanishing due to excessively large inner product values. Introducing a lower triangular mask matrix Suppressing future information through masking: (10) Final The masked self-attention output of the step is: (11) in, To output the projection matrix, the multi-head attention results are concatenated and projected to obtain meteorological features with temporal dependencies. .

[0035] In distributed photovoltaic data, the contribution of different regions to the total power varies significantly, and power prediction needs to meet the requirements of overall fitting accuracy and spatiotemporal correlation. A single loss function is difficult to satisfy these requirements. Therefore, a multi-objective loss function is designed to collaboratively optimize the model parameters, which not only retains the role of spatial benchmark in integrating discrete data, but also enhances the model's ability to capture dynamic weather-power correlation.

[0036] As the core foundation of the loss function, it is necessary to prioritize ensuring the model's fitting accuracy to the overall power level. Considering that photovoltaic power prediction is a continuous value regression problem, and that the penalty for larger errors is more stringent in practical applications, the squared error penalty characteristic of MSE can amplify the impact of extreme biases, forcing the model to focus on reducing large power errors. At the same time, its continuous differentiability is compatible with the model's gradient descent optimization mechanism, ensuring the stability of the training process. Therefore, the method constructs a multi-objective loss function oriented towards regional characteristics and cross-modal constraints: The mean squared error (MSE) is used as the basic loss term, and the formula is as follows: (12) Where N is the total number of regions. To predict the time step, and Let be the predicted power and the actual power of the i-th region at time t, respectively; AG-CNN extracts local features through dynamic kernels and cross-regional fusion, but its output regional features may suffer from fragmented correlations between adjacent regions due to the randomness of dynamic neighborhoods. The core of the regional feature consistency loss is to force the model to respect the implicit correlations between regions in physical space, ensuring that feature extraction not only focuses on the local details of a single region, but also captures the collaborative changes in adjacent regions caused by common meteorological disturbances. (13) Where, N i For the dynamic neighborhood of region i, w i，j The feature similarity weights for regions i and j are as follows: the higher the similarity, the greater the weight. This term compensates for the spatial correlation fragmentation problem caused by fixed grid division by constraining the consistency of power differences between adjacent regions. The core of DR-Transformer is to fuse historical power time-series features with future meteorological driving features through cross-attention. However, the model may exhibit prediction bias due to an unreasonable allocation of attention weights, resulting in a "disconnect between meteorological and power trends." The goal of cross-modal time-series correlation loss is to force the model to learn the physical causal relationship between meteorological factors and power changes, ensuring that cross-modal correlations in the time-series dimension conform to objective laws, rather than merely fitting the surface correlation of data. It regulates the model's correlation learning from the physical mechanism level through dual constraints: trend consistency + reasonable attention weights. (14) in, As a trend consistency loss function, it is used to constrain the physical rationality of the direction of change. β is the weighting coefficient of the trend consistency loss, used to balance the strength of trend constraint and attention weight constraint. (15) in, To predict the time-series variation of power, This represents the temporal variation of meteorological factors. For cross-entropy loss, a high penalty is imposed for situations where the power trend is opposite to the meteorological trend; As a cross-modal attention loss function, it is used to constrain the objective correlation of attention weights: (16) in, In the DR-Transformer cross-attention, meteorological features affect the region. exist Attention weight at any moment This is the true correlation coefficient between meteorology and power calculated based on historical data.

[0037] The overall loss function is the weighted sum of the three losses mentioned above: (17) Example Analysis: This embodiment features targeted configuration of manual parameters. AG-CNN employs a 3-layer convolutional block cascade structure. The first layer captures instantaneous features within a region using a dynamic kernel; the second layer extracts cross-regional meteorological propagation patterns using a dynamic kernel, outputting 64 feature channels; the third layer fuses the features from the first two layers through convolution, outputting 128 feature channels. Each convolutional layer is processed with batch normalization and the LeakyReLU activation function (slope set to 0.01). The DR-Transformer's encoder and decoder both use a multi-layer stacked structure, with a hidden dimension of 128, an 8-head multi-head self-attention mechanism, and a feedforward neural network dimension of 512. The encoder introduces a dynamic window partitioning strategy with a 4×4 window size; the linear transformation matrix for cross-attention in the decoder has a dimension of 128×128. In the Cross-Modal Feature Bridging Module (CMFB), the adaptive projection matrix has a dimension of 128×128, and the initial value of the bias term is 0.01; the cross-modal attention weight matrix of the Spatiotemporal Gated Unit (STGU) has a dimension of 128×128, the temporal feature embedding dimension is 64, and the loss function parameters are as follows. =0.3, =0.5, =0.1.

[0038] 3.1 Dataset and Experimental Environment Configuration; To verify the advantages of the proposed AG-CNN and DR-Transformer combined model in "power-meteorological cross-modal deep fusion" and "accurate modeling of long time-series dependencies", five distributed photovoltaic power stations (AEs) in a certain region were selected, and the relative locations of the stations are shown in the figure. Figure 6 Experiments were conducted using actual data. The dataset covers all 15-minute data points for a given year, including power generation data from various power plants and meteorological data (solar irradiance, cloud cover, temperature, etc.) for the corresponding regions. The data was Z-score normalized and spatiotemporally aligned.

[0039] The experiment was implemented using Python 3.9, with a hardware configuration of one GTX 2060 graphics card, 16GB of memory, and a 6-core Intel(R) Core (TM) i7-10750H CPU @ 2.60GHz.

[0040] To systematically verify the comprehensive performance of this embodiment (M4) in distributed photovoltaic power prediction tasks, three representative prediction models were selected as a benchmark for comparison. The BP neural network (M1), as a classic feedforward neural network model, represents a basic power prediction method; the TCN-Transformer model (M2), combining a temporal convolutional network and the Transformer architecture, is a commonly used model in long-sequence prediction; and the GCN-LSTM model (M3), through the combination of a graph convolutional network and a long short-term memory network, can simultaneously handle spatial and temporal dependencies. Through comparative analysis with these three models, the predictive ability and adaptability of the proposed model in complex meteorological scenarios can be evaluated from multiple dimensions, and its performance in power-meteorological cross-modal fusion and long-sequence dependency modeling can be systematically examined.

[0041] 3.2 Predictive performance comparison and error analysis; To visually compare the differences in power fitting performance of the four models in real-world application scenarios, we selected five random days of measured data covering normal and fluctuating scenarios from five distributed photovoltaic power stations of AE. We then conducted a comparative analysis of the prediction results of the M1 (BP neural network), M2 (TCN-Transformer), M3 (GCN-LSTM), and the M4 (AG-CNN+DR-Transformer) model proposed in this embodiment.

[0042] based on Figure 7Comparative analysis of the power prediction curves from various power plants reveals that the M4 model (AG-CNN+DR-Transformer) proposed in this embodiment exhibits significant advantages in curve fitting accuracy, dynamic response speed, and adaptability to complex scenarios. Its prediction curves closely track actual power changes, especially during the midday period in fluctuating scenarios. The M4 model not only accurately captures the fluctuation amplitude but also achieves precise temporal synchronization. In contrast, the M1 model (BP neural network) shows a significant smoothing bias at power abrupt changes, the M2 model (TCN-Transformer), while able to grasp the trend, suffers from phase lag, and the M3 model (GCN-LSTM) exhibits overshoot in peak tracking. This visualization confirms the effectiveness of the AG-CNN module in enhancing local feature extraction through dynamic convolutional kernels and the DR-Transformer module in deepening temporal correlation through cross-modal attention, jointly ensuring the model's prediction robustness across multiple distributed photovoltaic scenarios.

[0043] To more intuitively quantify and analyze the prediction performance of the proposed model, this paper uses two indicators, normalized root mean square error (NRMSE) and normalized mean absolute error (NMAE), to evaluate the prediction accuracy of the model.

[0044] (18) (19) Where n is the total number of samples. and P represents the actual power and the predicted power of the j-th sample, respectively. N The rated power is used as the reference for normalization.

[0045] Table 1: Error Indicators of Each Prediction Model

[0046] Table 1 quantifies the differences in prediction accuracy of the four models across five sites under normal and fluctuating scenarios using two core indicators: Normalized Root Mean Square Error (NRMSE) and Normalized Mean Absolute Error (NMAE). Figure 8The numerical results are visualized in the form of bar charts, and the two mutually corroborate each other, highlighting the significant advantages of the M4 model in this paper. In normal scenarios, the NRMSE and NMAE of M4 are as low as 0.0821-0.1069 and 0.0697-0.1241, respectively, which are 35%-40% lower than M1 on average, 18%-25% lower than M2 on average, and 10%-15% lower than M3 on average. In fluctuating scenarios, M4 has the smallest error increase, with NRMSE and NMAE remaining at 0.1006-0.1269 and 0.0897-0.1341, respectively, still significantly better than M1 which relies on static structures, M2 which fuses shallow layers across modalities, and M3 which is limited by fixed graph structures and convolutional kernels.

[0047] Figure 8 Further analysis revealed the spatial distribution and stability differences of errors in each model. The M4 histogram was generally low and fluctuated gently, with NRMSE and NMAE fluctuations of only 0.0248 and 0.0544 for all stations. Even at sparsely populated stations D and E or frequently disturbed meteorological station C, it maintained a low error level, demonstrating the synergistic effectiveness of AG-CNN dynamic spatial modeling and DR-Transformer cross-modal long-term fusion. In contrast, the M1 histogram was generally high and fluctuated the most, while M2 and M3 showed significant error peaks in areas with long-term meteorological changes or sparse stations. Their limitations contrasted sharply with the innovative mechanism of M4, confirming M4's core advantages in adapting to the spatial heterogeneity of distributed photovoltaic systems and complex meteorological scenarios.

[0048] To further illustrate the prediction performance of this embodiment, the prediction results of the proposed model and the comparative model are compared and analyzed in different ways. Figure 9 The distribution characteristics of the absolute prediction errors of each model are clearly presented: M1 (BP neural network) has a wide box and dense extreme values. Due to regional distribution differences and insufficient modeling of modal correlation, the error dispersion is extremely high. Although the box of M2 (TCN-Transformer) is narrower than that of M1, there are still many outliers, reflecting that its shallow cross-modal fusion leads to limited stability. M3 (GCN-LSTM) has a low median. The combination of spatial and temporal modeling alleviates some error fluctuations, but the fixed structure still leaves room for optimization in dynamic scenes. M4 (AG-CNN+DR-Transformer) performs the best. The box is extremely narrow and there are no extreme outliers. Its dynamic spatial feature extraction of AG-CNN and cross-modal long temporal fusion of DR-Transformer work together to effectively suppress error dispersion and maintain stable performance in different regions.

[0049] Example 2: This embodiment provides a regional distributed photovoltaic power prediction system based on multi-data cross-modal fusion, including: The data processing module is used for spatiotemporal data gridding consistency processing. It collects raw meteorological data and historical power data, performs data standardization, and then uses unified geographic coordinates as a benchmark to realize the spatial correlation between meteorological data and historical power data. The extraction module adaptively extracts data based on the spatial features of AG-CNN with dynamic perceptual convolution. AG-CNN relies on a unified basic spatial benchmark and achieves local modeling through dynamic feature perception and cross-regional information fusion. The design module is used for the design of cross-modal feature bridging module. Through a three-level processing flow of adaptive projection-feature reparameterization-spatiotemporal fusion gating, the high-dimensional local features output by AG-CNN are transformed into a spatiotemporal token sequence that can be parsed by DR-Transformer, while retaining key spatial details and cross-regional correlation information. The prediction module is used to incorporate the DR-Transformer temporal correlation modeling with dual-dimensional dynamic attention. The DR-Transformer takes the spatial dimension + temporal dimension dynamic attention mechanism as its core, deeply integrates the long-term temporal dependence of historical power and the driving role of future meteorological data, realizes the accurate coupling of meteorological and power cross-modal spatiotemporal features, and obtains the total power prediction result of the distributed photovoltaic system.

Claims

1. A regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion, characterized in that, The steps are: (1) Spatiotemporal data grid consistency processing, collecting original meteorological data and historical power data, and performing data standardization, then taking the unified geographic coordinates as the reference to realize the spatial correlation of meteorological data and historical power data; (2) AG-CNN spatial feature based on dynamic perception convolution for adaptive data extraction, AG-CNN relies on a unified basic spatial reference, and realizes local modeling through dynamic feature perception and cross-region information fusion; (3) Cross-modal feature bridge module design, through the three-level processing process of adaptive projection-feature reparameterization-spatiotemporal fusion gating, the high-dimensional local features output by AG-CNN are converted into spatiotemporal token sequences that can be analyzed by DR-Transformer, while retaining key spatial details and cross-region association information; (4) DR-Transformer time sequence correlation modeling with dual-dimensional dynamic attention, DR-Transformer takes the spatial dimension + time sequence dimension dynamic attention mechanism as the core, deeply fuses the long time sequence dependence of historical power and the driving effect of future meteorological data, realizes the precise coupling of meteorological-power cross-modal spatiotemporal features, and obtains the total power prediction result of the distributed photovoltaic system.

2. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 1, characterized in that, In step (1), the meteorological data includes solar irradiance, temperature and humidity; The data standardization process processes meteorological data and historical power data respectively, eliminates the influence of different characteristic dimensions, and adopts Z-score standardization method, the formula is as follows: （1） where x is the original data, μ is the data mean, and σ is the standard deviation, is the standardized data; Spatial correlation: combined with the coverage of photovoltaic array and meteorological monitoring point, based on the spatial scale of meteorological data, the geographical position of photovoltaic panel is matched with the corresponding meteorological area, so that the power data and meteorological data form a grid spatial correlation, at the same time, the data is time aligned according to the time axis of the prediction granularity, and the abnormal data is eliminated, forming a spatiotemporal double-dimensional aligned input structure.

3. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 2, characterized in that, In step (2), taking the grid unit as the basic space index, the convolution kernel of AG-CNN is adapted to the site distribution density and feature fluctuation mode in different grid units through learnable parameters; The convolution kernel weight of AG-CNN is constructed by the spatial relative position offset Gaussian function and the learnable parameter: （2） wherein (i, j) represents a spatial relative position offset, i is a horizontal direction spatial relative position offset, j is a vertical direction spatial relative position offset, σ g is a Gaussian kernel bandwidth, s(i, j) is a feature similarity between the offset region and the target region, θ is a learnable parameter, and α = 5 is used to balance the influence of spatial position correlation and feature mode similarity. When performing a convolution operation on the feature map, the output feature adaptively fuses the information of the target region and the surrounding associated region: （3） wherein y x，y is a feature value of the target region in the output feature map, and x x+i，y+j is a feature of the peripheral associated region in the input feature map.

4. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 1, characterized in that, The network structure of AG-CNN adopts multi-layer convolution block series design: the first layer captures the instantaneous irradiance mutation in the region through dynamic kernel; the second layer extracts the meteorological propagation law across regions through dynamic kernel; the third layer fuses the features of the previous two layers through convolution, and strengthens the key details, and each layer after convolution is processed by batch normalization and activation.

5. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 4, characterized in that, In step (3), the feature map F output by AG-CNN is a four-dimensional tensor: （4） Where B is the training batch size, C is the number of feature channels, and HxW corresponds to the number of regional divisions; Firstly, the two-dimensional regional structure is converted into a one-dimensional sequence through spatial flattening operation, and the channel dimension is adapted to the Transformer hidden dimension through learnable linear transformation: （5） wherein, is the flattened sequence feature, N = H x W is the total number of regions, is the projection matrix, is the bias term, D is the Transformer hidden dimension; Since there is a difference between the optimization objectives of AG-CNN and DR-Transformer, a reparameterization layer is introduced to dynamically adjust the feature distribution to alleviate the training deviation: （6） wherein, is an element-wise gating coefficient, is a Sigmoid activation function, , is a learnable parameter, the gating mechanism realizes dynamic alignment of the feature distributions of the two models by reinforcing key features and suppressing noise; To show the temporal characteristics of the fusion history power and the spatial characteristics of the AG-CNN, a spatio-temporal gating unit is introduced to dynamically calibrate the correlation weight between the two; Let the historical power sequence be where T h is the historical time step, R is the real domain, B is the training batch size, N is the total number of regions, and D is the Transformer hidden dimension. The gating fusion process is as follows: （7） wherein, is a dynamic fusion coefficient, F in is an output feature of the STGU, which carries both spatial details and temporal correlations, STGU is a spatio-temporal gating unit for fusing spatial and temporal features, W α is a cross-modal attention weight matrix, b α is a cross-modal attention bias term, F reparam is a re-parameterized spatial feature output by equation (6), is a cross-modal attention mechanism that learns the correlation weights between spatial and temporal features to align their semantics. Final output directly as input to the DR-Transformer encoder.

6. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 5, characterized in that, In step (4), the DR-Transformer realizes the cross-modal correlation modeling of the historical power and the future weather through an encoder-decoder structure; The encoder receives the sequence output by the cross-modal feature bridge module, processes the historical power and the spatial features, and captures the implicit correlation and long-term temporal dependence between regions through dynamic spatial attention. The encoder adopts a multi-layer stacking structure: the bottom layer introduces a dynamic window division strategy to strengthen the temporal correlation of local regions through window attention; The upper layer captures long-distance temporal dependence through global attention. Each layer contains multi-head dynamic spatial attention and a feedforward neural network. The attention weight is modulated by the relative position of the region and the similarity of the power features; The decoder processes the future time step weather data. Under the constraint of the mask self-attention mechanism, the weather features are dynamically coupled with the historical features output by the encoder through cross-attention. The decoder adopts a multi-layer stacking structure, each layer of which is composed of multi-head mask self-attention, multi-head cross-attention, and a feedforward neural network: the mask self-attention ensures the temporal causality of the weather sequence; the cross-attention adjusts the weight through a dynamic correlation coefficient. After the encoder and decoder features are mapped to the same dimension through a linear projection layer, the cross-attention realizes the precise correlation between the historical power trend and the future weather driving, and outputs the features that fuse global spatial and temporal information; Finally, the output layer converts the fused features into regional-level power prediction values for multiple future time steps through a fully connected layer, and then weights and aggregates the regional power prediction values according to the installed capacity of the photovoltaic system to obtain the total power prediction result of the distributed photovoltaic system, completing the whole process from input data to prediction output.

7. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 6, characterized in that, DR-Transformer focuses on the dynamic attention mechanism in the spatial and temporal dimensions: in the spatial dimension, the attention weight is adaptively adjusted by the similarity of the regional features and the relative position relationship; In the temporal dimension, the evolution law of the historical power is dynamically coupled with the driving effect of the future weather through the cross-attention mechanism; To address the discrete distribution characteristics of the distributed photovoltaic station, the encoder of the DR-Transformer captures the implicit correlation between regions through dynamic spatial attention. The attention weight calculation considers both the relative spatial distance and the feature pattern similarity. The calculation formula of the dynamic spatial attention weight is as follows: （8） where q, k, v are query, key, value vectors respectively, d k is the feature dimension, is the feature similarity coefficient, and β(r) is the relative spatial distance coefficient. Through the synergistic modulation of the two, the attention weight can not only reflect the natural continuity in space, but also focus on the feature-similar associated regions. The encoder of the DR-Transformer dynamically fuses the future weather data with the historical features output by the encoder through mask self-attention and cross-attention to realize cross-modal correlation; The mask self-attention enforces the model to follow the temporal causality through a lower triangular matrix, allowing only the current time step to focus on the historical and contemporaneous weather features. Let the future weather sequence input to the decoder be , T is the number of future prediction steps, D is the dimension of the weather feature vector, R represents the real number domain, m t is the weather feature vector at the t-th step, the query , key , value , , , is generated by linear projection, h is the number of attention heads; For the t-th time step, the similarity calculation of the mask self-attention is as follows: （9） wherein, is the time step of interest, is a scaling factor; Introducing a lower triangular mask matrix Suppressing future information by a mask （10） In the formula, denotes the element in the t-th row and the column of the mask matrix, and MaskedSim is the attention similarity score after masking. Final step The masked self-attention output MaskedAttn(t) of the step is: （11） where, is the output projection matrix, v represents the value vector, softmax(·) is the normalized exponential function, Concat(·) represents the multi-head attention splicing operation, and after splicing and projection of the multi-head attention result, the meteorological features containing time sequence dependence are obtained .

8. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 7, characterized in that, In the method, a multi-objective loss function is constructed to address the regional characteristics and cross-modal constraints.

9. The regional distributed photovoltaic power prediction method based on multi-element data cross-modal fusion according to claim 8, characterized in that, The multi-objective loss function uses mean square error as the basic loss term, and the formula is as follows: （12） wherein N is the total number of regions, is the prediction time step, and are the predicted and real power of the i-th region at time t, respectively. AG-CNN extracts local features through dynamic kernels and cross-region fusion, but the output region features are fragmented due to the randomness of dynamic neighborhoods, and the core of the loss of region feature consistency is to force the model to respect the implicit correlation between regions in physical space, ensuring that feature extraction not only focuses on the local details of a single region, but also captures the collaborative changes of adjacent regions due to common meteorological disturbances: （13） where N i is the dynamic neighborhood of region i, w i，j is the feature similarity weight of region i, j, the higher the similarity, the greater the weight; The core of DR-Transformer is to fuse historical power time series features and future meteorological driving features through cross-attention, and the goal of cross-modal temporal correlation loss is to force the model to learn the physical causal relationship between meteorological factors and power changes, ensuring that cross-modal correlation in the time dimension conforms to objective laws, rather than just fitting surface correlations. Data, it regulates the model's correlation learning from the physical mechanism level through double constraints: （14） wherein, As a physical rationality to constrain the change direction, the trend consistency loss function is used, and β is the weight coefficient of the trend consistency loss, which is used to balance the strength of the trend constraint and the attention weight constraint: （15） wherein, is a time series variation of the power, is a time series variation of the weather factor, is a cross-entropy loss, giving a high penalty for cases where the power trend and the weather trend are opposite. As a cross-modal attention loss function to constrain the objective relevance of attention weights: （16） wherein, is the weather feature to region cross attention in the DR-Transformer, In the attention weight at time t, is the weather-power true correlation coefficient calculated based on historical data; The overall loss function is the weighted sum of the three losses mentioned above: （17）。 10. A regional distributed photovoltaic power prediction system based on multi-element data cross-modal fusion, characterized in that, Including: Data processing module, for spatial-temporal data grid consistency processing, collecting original meteorological data and historical power data, and performing data standardization, then taking unified geographic coordinates as the basis, realizing the spatial correlation of meteorological data and historical power data; Extraction module, based on AG-CNN spatial feature extraction of dynamic perception convolution, AG-CNN relies on a unified basic spatial reference to realize local modeling through dynamic feature perception and cross-region information fusion; Design module, for cross-modal feature bridge module design, through the three-level processing process of adaptive projection-feature reparameterization-time-space fusion gate, the high-dimensional local features output by AG-CNN are converted into time-space token sequences that can be analyzed by DR-Transformer, while retaining key spatial details and cross-region correlation information; Prediction module, for DR-Transformer time correlation modeling with dual-dimensional dynamic attention, DR-Transformer takes spatial dimension + time dimension dynamic attention mechanism as the core, deeply fuses the long-time sequence dependence of historical power and the driving effect of future meteorological data, realizes the precise coupling of meteorological-power cross-modal spatio-temporal features, and gets the total power prediction result of distributed photovoltaic system.

Images