A photovoltaic cluster power prediction method based on meteorological large model progressive cross-modal migration

Abstract

The application provides a photovoltaic cluster power prediction method based on meteorological large model progressive cross-modal migration, and relates to the technical field of photovoltaic cluster power prediction. The method comprises the following steps: taking a meteorological large model as a backbone, performing self-supervised pre-training on global meteorological reanalysis data to obtain general meteorological characteristics; performing domain self-adaptive fine-tuning on the meteorological backbone by using target area meteorological data to obtain a regional meteorological backbone adapted to regional climate characteristics; obtaining the historical power sequence, installed capacity, geographical position and solar geometric information of photovoltaic power stations in the target area, and sequentially performing multi-station spatio-temporal feature extraction and installed capacity enhancement, meteorological and power cross-modal coupling, and future power sequence generation based on solar geometric information constraint to obtain the normalized power prediction value of each power station in the future prediction period. The application can realize effective migration of meteorological large model knowledge to photovoltaic cluster power prediction task, and effectively improve the prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of photovoltaic cluster power prediction technology, and in particular to a photovoltaic cluster power prediction method based on a gradual cross-modal migration of a large meteorological model. Background Technology

[0002] As photovoltaic (PV) penetration continues to increase, cloud movement, abrupt changes in irradiance, and regional weather processes directly cause fluctuations in PV power output. Different power plants are influenced by shared weather patterns but also constrained by local meteorological conditions and geographical variations, resulting in significant temporal and spatial dependence and correlation in PV cluster power output. Furthermore, PV power prediction errors exhibit spatiotemporal dependence depending on the power plant's location, prediction time, and operating conditions. Therefore, PV cluster prediction requires not only utilizing historical power information but also effectively characterizing the regional meteorological field and its spatiotemporal evolution.

[0003] Existing photovoltaic (PV) power prediction methods typically use numerical weather prediction (NWP) as the primary meteorological input, combined with historical power data, and employ machine learning or deep learning models to establish a mapping relationship between meteorological conditions and PV output. Related research has also improved prediction performance through fusion of physics and data, probabilistic prediction, and the fusion of satellite imagery and solar position information. In the field of wind power and power plant cluster prediction, methods such as multi-task learning, fusion of homogeneous and heterogeneous information interaction, and dynamic graph neural networks have been used to characterize multivariate temporal relationships and dynamic dependencies between power plants, providing methodological references for PV cluster prediction. However, most existing methods still rely on a two-stage process of "meteorological prediction—power prediction." Front-end meteorological biases are easily propagated backward, and the power prediction objective is difficult to use to constrain meteorological representation learning in the reverse direction. This leads to amplitude bias and phase lag in models under scenarios such as cloud movement, sudden changes in irradiance, and power ramp-up.

[0004] To overcome the limitations of traditional two-stage forecasting methods, some research has begun to introduce large-scale models into time series forecasting tasks. Existing methods mainly fall into three categories: general time series foundation models, language model transfer methods, and meteorological foundation models. General time series foundation models learn transferable statistical laws through large-scale, multi-domain time series pre-training. For example, Chronos quantizes continuous numerical sequences into discrete tokens for probabilistic prediction, while TimesFM uses a decoder-only structure to learn general forecasting patterns across datasets. Lag-Llama, Moirai, and MOMENT improve the model's generalization ability from the perspectives of probabilistic modeling, unified multi-frequency training, and masked pre-training, respectively. Language model transfer methods utilize the sequence modeling capabilities of pre-trained language models, conducting time series forecasting through numerical sequence texturing, model adaptation, and representation reprogramming. Representative methods include LLMTime, GPT4TS, and Time-LLM. However, the first two types of methods mainly learn general time series statistical laws or language representations, making it difficult to directly characterize the driving effect of meteorological physical fields on photovoltaic power output. Therefore, they may have certain limitations in power change scenarios dominated by meteorological processes.

[0005] Compared to general time-series and language-based large models, the representations learned by meteorological large models are more aligned with the information required for photovoltaic power prediction. The Pangoal Meteorological Model (PGM) employs a 3D neural network to jointly model atmospheric states across different pressure layers and spatial dimensions, while GraphCast utilizes graph neural networks to characterize spatial relationships within the global meteorological field. FourCastNet, ClimaX, and FuXi enhance meteorological field prediction capabilities from the perspectives of frequency domain modeling, unified representation of heterogeneous meteorological variables, and multi-terminal cascade prediction, respectively. The representations obtained from the basic meteorological model have a more direct connection to cloud evolution, irradiance changes, and photovoltaic power fluctuations, and are more aligned with the meteorological driving mechanisms of photovoltaic power prediction compared to general time-series and language-based model transfer methods. Transferring these models to photovoltaic power prediction tasks is expected to compensate for the shortcomings of traditional NWP methods and general time-series and language-based large models in characterizing meteorological drivers.

[0006] While large-scale meteorological models provide a more meteorological-driven representation basis for photovoltaic (PV) power prediction, their application to PV cluster power prediction still faces three challenges: First, the traditional two-stage NWP process suffers from meteorological error propagation and separation of meteorological and power objectives; how to achieve synergistic optimization between meteorological representation learning and power prediction objectives remains to be addressed. Second, large-scale meteorological models typically predict global or large-scale meteorological fields, with large model scales and high variable dimensionality, differing from PV cluster prediction scenarios in spatial scale, climate distribution, and mission objectives, making direct application difficult to adapt to local power prediction needs. Third, large-scale meteorological models learn continuous meteorological field representations, while PV prediction objectives are discrete multi-power station power sequences, resulting in significant modal differences in spatial structure, variable semantics, and output format. Furthermore, PV power is constrained by the sun's trajectory, requiring the prediction structure to balance cross-modal coupling capability and physical consistency. Summary of the Invention

[0007] This application provides a photovoltaic cluster power prediction method based on a progressive cross-modal transfer using a large meteorological model. Using a large meteorological model as the backbone, it first progressively adapts global meteorological characteristics to the climate features of the target region. Then, in the power prediction stage, it establishes cross-modal coupling between multi-scale meteorological fields and multi-site power sequences, and introduces solar geometric prior constraints for future power sequence modeling, realizing the transfer of meteorological knowledge to the photovoltaic cluster power prediction task. The objectives of this application include: 1) A three-stage meteorological big model progressive task transfer is proposed. Through global meteorological representation learning, regional meteorological domain adaptive fine-tuning, and meteorological-to-power cross-modal transfer, meteorological knowledge is gradually adapted to the local photovoltaic cluster prediction task.

[0008] 2) A cross-level meteorological power heterogeneous representation coupling mechanism is proposed to aggregate multi-scale meteorological field information layer by layer, so as to realize the structured interaction between continuous meteorological representation and discrete multi-site power representation.

[0009] 3) A bidirectional Mamba sequence modeling structure with enhanced solar geometric priors is proposed, which incorporates solar position information into future power prediction and enhances the model's ability to perceive the intraday power generation patterns of photovoltaics.

[0010] To achieve the above objectives, this application adopts the following technical solution: This application provides a photovoltaic cluster power prediction method based on a progressive cross-modal migration of a large meteorological model, the method comprising: Global meteorological reanalysis data is acquired, and a large meteorological model is used as the meteorological backbone. The meteorological backbone is then subjected to global meteorological self-supervised pre-training to obtain a general meteorological characterization. Acquire regional meteorological data for the target area, use the general meteorological characterization as the initial model parameters of the meteorological backbone, and perform regional meteorological domain adaptive fine-tuning on the meteorological backbone to obtain a regional meteorological backbone adapted to the climate characteristics of the target area. The process involves acquiring historical power sequences and installed capacity of multiple photovoltaic (PV) power plants within the target area, along with their geographical location information and solar geometric information. Multi-site spatiotemporal representations are extracted based on the historical power sequences and geographical location information, and enhanced according to the installed capacity to obtain site representations that integrate installed capacity information. Based on the regional meteorological backbone, a cross-modal relationship is established between the continuous meteorological field and the discrete multi-site power sequences, using the site representations with integrated installed capacity information as the query, to obtain the final site representations that integrate multi-scale meteorological information. Based on the final site representations and the solar geometric information, normalized power prediction values ​​for each PV power plant are generated for the future prediction period.

[0011] Furthermore, the aforementioned meteorological backbone personnel undergo global meteorological self-monitoring pre-training, including: A spatial window consistent with the target area scale is cropped from the global meteorological reanalysis data. Using historical meteorological fields as input, future meteorological conditions are predicted, and the meteorological backbone is trained by combining variable-weighted prediction loss with a spatial gradient smoothing term. Define spatial gradient smoothing term As shown in the formula below: (1) in, For meteorological grids ( i , j The predicted value vector at position ) i , j For spatial grid indexing; For meteorological grids ( i+ 1, j The predicted value vector at () The global pre-training loss function is constructed as shown in the following formula: (2) in, For the first v Predicted values ​​of several meteorological variables To correspond to the actual value, For the first v The penalty weights of each weather variable, The weighted prediction loss for variables in the upper atmosphere is used. The balance coefficient for the spatial gradient smoothing term; For spatial gradient smoothing, The loss function is the global meteorological self-supervised pre-training function.

[0012] The meteorological backbone is trained by minimizing the global pre-training loss function until the global pre-training loss function converges, and the trained meteorological backbone is obtained. The model parameters of the trained meteorological backbone constitute the general meteorological representation.

[0013] Furthermore, the meteorological backbone is subjected to regional meteorological domain adaptive fine-tuning, including: The meteorological backbone is initialized using the general meteorological representation, and fine-tuned using regional meteorological data of the target area. The global pre-training loss function is then adjusted accordingly. Replace with dynamic weights The formula for calculating the dynamic weight is: (3) (4) in, The zenith angle of the sun. For daytime masking, , The weights for day and night variables are respectively. This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise; therefore, when hour, Indicates daytime; otherwise, it indicates nighttime. Using the replaced loss function as the optimization objective, the shallow feature extraction parameters are frozen in the early stage of training, the deep encoding module and decoding module are updated, and then the remaining parameters are gradually unfrozen to fine-tune the meteorological backbone so that the meteorological feature distribution of the meteorological backbone is adapted from the global scale to the climate features of the target region, thus obtaining the regional meteorological backbone.

[0014] Furthermore, the geographical location information of multiple photovoltaic power stations within the target area is obtained, including: Obtain the latitude and longitude coordinates of each photovoltaic power station, and establish graph connections based on the geographical distance between the stations, with each photovoltaic power station as a graph node. The site is calculated using the following formula. i With the site j spherical distance between : (5) (6) in, R The average radius of the Earth and Sites i and j latitude and longitude and These represent the difference in latitude and longitude between the two stations, respectively. This is an intermediate variable in the calculation of the Haversine spherical distance; According to the site i With the site j spherical distance between For each site, select K neighboring sites, construct an undirected graph through symmetry processing, use the median of effective edge distances as the scaling parameter, map geographical distances to continuous edge weights using a Gaussian kernel, and obtain a weighted adjacency matrix after normalization. : (7) in, Here, KNN is the adjacency indicator function, and KNN is the set of K nearest neighbors obtained after symmetrization. e It is a natural constant. k Summing index for neighboring sites, d ik For the site i With the site k The spherical distance between them This is the distance scale parameter, and its value is the median of the effective connected edge distances; The geographic location information includes the latitude and longitude coordinates of each photovoltaic power station and the weighted adjacency matrix.

[0015] Further, multi-site spatiotemporal representations are extracted based on the historical power sequence and the geographical location information, and the multi-site spatiotemporal representations are enhanced based on the installed capacity to obtain site representations that integrate installed capacity information, including: The latitude and longitude coordinates of each station are encoded into absolute position codes, and after linear projection, initial power features are added to obtain initial station features; Using the initial site features as input, temporal dynamic features within each site are extracted through temporal convolution to obtain intermediate features after temporal convolution; spatial association features of neighboring sites are aggregated through graph messaging to obtain site features output in this round of encoding; the temporal convolution and graph messaging are executed alternately for two rounds to obtain a multi-site spatiotemporal representation; The spatiotemporal representations of the multiple sites are weighted and aggregated along the time dimension to obtain the site-level representation of each site; the site-level representations are enhanced according to the installed capacity of each site to obtain the site representation that integrates the installed capacity information.

[0016] Furthermore, temporal dynamic features within each site are extracted through temporal convolution to obtain intermediate features after temporal convolution, including: In the lIn round coding, the time series data of each station are subjected to one-dimensional convolution and then nonlinear transformation through a feedforward network, as shown in the following formula: (8) (9) in, Represents one-dimensional convolution. For GELU activation function, For convolution weights, , and For learnable linear transformation matrices, For the first l The input features of round coding These are intermediate features after temporal convolution. These are intermediate features output by the feedforward network.

[0017] Furthermore, by aggregating the spatial association features of neighboring sites through graph messaging, the site features output in this round of encoding are obtained, including: The weighted adjacency matrix is ​​used to aggregate neighbor site features, and nonlinear fusion is completed through a feedforward network, as shown in the following formula: (10) (11) (12) in, For neighboring sites j The intermediate features output by the feedforward network, For the site i The neighbor messages obtained through aggregation The message transformation matrix, and For feedforward network parameters, The intermediate features after fusing neighbor messages, For the first l Station characteristics output by round coding.

[0018] Furthermore, the spatiotemporal representations of the multiple sites are weighted and converged along the time dimension to obtain site-level representations for each site. These site-level representations are then enhanced based on the installed capacity of each site to obtain site representations that integrate installed capacity information, including: After two rounds of spatiotemporal encoding, a learnable attention pooling method is used to converge the time dimension, and the attention score is calculated using the following formula: (13) in, For the site after the second round of coding iAt any moment t Features For the site i At any moment t Attention score This is the learnable attention vector used to compute the temporal attention score; Based on the attention score, the site-level representation is calculated using the following formula: (14) in, For the site i Site-level representation, t For the current moment in the historical power sequence, T The number of time steps in the historical power series. For the summation in the attention normalization process, For the site i exist Attention score; Based on the site-level representation, the site representation with integrated installed capacity information is calculated using the following formula: (15) in, This is a normalized scalar representation of installed capacity. For the projection matrix, This is a site representation that integrates installed capacity information.

[0019] Furthermore, based on the regional meteorological backbone, and using the station representation of the fused installed capacity scale information as a query, a cross-modal relationship is established between the continuous meteorological field and the discrete multi-station power sequence to obtain the final station representation of the fused multi-scale meteorological information, including: Obtain the current meteorological feature sequence , with site i The site representation, which integrates installed capacity information, is used as the query, while meteorological features serve as both keys and values. The query vector, key matrix, and value matrix are calculated using the following formulas: (16) in, For learnable projection matrices, For the site i The query vector, K , V These are the key matrix and value matrix of meteorological characteristics, respectively; According to the spherical distance The geographic distance attention bias is calculated using the following formula: (17) in, For learnable positive scalars, For the site i With meteorological grid j Geographical distance between them leads to attention bias; Based on the query vector, key matrix, and geographic distance attention bias, the attention score is calculated using the following formula: (18) in, k j Meteorological grid j The key vector, v j Meteorological grid j The value vector, d k The dimension of the key vector. s ij For the site i Meteorological grid j Attention score; Based on the attention score, the normalized weights are calculated using the following formula: (19) in, To normalize the weights, This is a normalized exponential function used to convert the attention scores corresponding to each meteorological grid point into normalized weights that sum to 1; The site is calculated using the following formula based on the normalized weights and the value matrix. i In the l Meteorological backbone level, meteorological flow r Enhanced representation under : (20) in, l For meteorological backbone hierarchical index, r Indicates surface or upper-air meteorological features, For the first l The number of meteorological grid points at each level j For meteorological grid indexing; Each meteorological backbone level performs the cross-attention function on surface meteorological features and upper-air meteorological features respectively to obtain enhanced surface characterization. and high-altitude enhanced characterization The surface enhancement characterization and the upper-air enhancement characterization are fused using the following formula: (twenty one) in, For the first lEach level of gating parameter is used to adjust the relative contribution of surface and upper-air meteorological information at that level. Enhance the representation of the merged sites; The regional meteorological backbone comprises four layers, from Layer 1 to Layer 4. Station representations are updated progressively in the order of Layer 1, Layer 2, Layer 3, and Layer 4. Layer 1 consists of high-resolution coded features, Layers 2 and 3 are downsampled low-resolution deep coded features, and Layer 4 is upsampled decoded features. Using the station representations as queries, cross-attention and gating fusion is performed with the surface meteorological features and upper-air meteorological features of Layer 1, respectively, to obtain the updated station representations for Layer 1. Before entering Layer 2, the updated station representations of Layer 1 undergo dimensionality reduction projection to match Layer 4. The feature dimensions of Layer 2 and Layer 3 are used to obtain the dimensionality-reduced site representation. Using the dimensionality-reduced site representation as a query, the cross-attention and gating fusion is performed with the surface meteorological features and upper-air meteorological features of Layer 2 respectively to obtain the updated site representation of Layer 2. Using the updated site representation of Layer 2 as a query, the cross-attention and gating fusion is performed with the surface meteorological features and upper-air meteorological features of Layer 3 respectively to obtain the updated site representation of Layer 3. Before entering Layer 4, the updated site representation of Layer 3 is subjected to dimensionality-up projection, and the updated site representation of Layer 1 is introduced through skip connections to obtain the dimensionality-up projected site representation. The upgraded site representations and Layer 4 features are subjected to cross-attention, resulting in Layer 4 surface flow cross-attention enhancement results. and high-altitude flow across attention enhancement results The incremental representation of Layer 4 is calculated using the following formula: (twenty two) in, For Layer 4 surface-high altitude fusion gating, This represents the incremental representation of Layer 4; Using residual gating parameters The Layer4 incremental representation is interpolated into the upgraded station representation, and the final station representation, which integrates multi-scale meteorological information, is calculated using the following formula: (twenty three) in, To provide the final station representation that integrates multi-scale meteorological information, The site representation is projected in a higher dimension and updated from Layer 1 before entering Layer 4.

[0020] Further, based on the final site characterization and the solar geometry information, normalized power prediction values ​​for each photovoltaic power station are generated for the future prediction period, including: Characterize the final site Expand along the prediction time dimension and overlay the learnable temporal location codes corresponding to the prediction steps. The initial predicted time-series features are calculated using the following formula: (twenty four) in, H To predict the number of time steps, For the site n The final site representation, For the site n In the h Initial prediction time series characteristics at each prediction time point; Based on the station's latitude and longitude and the predicted time, the 6-dimensional solar geometric features are calculated using the following formula. : (25) in, For the site n In the h The solar altitude angle at the predicted moment. The azimuth of the sun. To illuminate the upper realm of clear skies, Use a daytime mask; Based on the initial predicted time-series features and the solar geometric features, the predicted time-series features after incorporating solar geometric information are calculated using the following formula: (26) in, For auxiliary feature projection matrix; Based on the predicted time-series characteristics after incorporating solar geometric information, sequence modeling is performed using bidirectional Mamba, where the continuous state-space parameters are discretized using the zero-order preservation method, and the discretization formula is as follows: (27) (28) in, To enable learning step size, It is the identity matrix. , For continuous state-space parameters, , These are the discretized state transition matrix and input matrix; Based on the discretized state-space parameters, the first...t Hidden state and output of the step: (29) (30) in, and The first t Step and the first t -1 step hidden state For the first t Step input, Depend on Dynamically generated For skip connectivity coefficients, For the first t Step output; Bidirectional scanning is used, with the positive branch along 1 to... H The states are updated sequentially in each direction. The reverse branch scans the flipped sequence and restores the original order. Based on the outputs of the forward and reverse scans, the final fused features are calculated using the following formula: (31) in, This is the output of the forward scan. This is the output of the reverse scan. To fuse the projection matrix, LN is the layer normalization. For the site In the The final fusion features at each prediction time; Based on the final fusion characteristics, the normalized power prediction value is calculated using the following formula: (32) in, For the site n In the h Normalized power prediction values ​​at each prediction time. For the site n In the h The final fusion features at each prediction time point, , This is the weight matrix of the feedforward network.

[0021] The photovoltaic cluster power prediction method based on progressive cross-modal migration using a large meteorological model provided in this application has at least the following beneficial effects: This application addresses the problems of meteorological error propagation and the separation of meteorological and power targets in traditional two-stage forecasting methods. Through progressive cross-modal transfer of a large meteorological model, it enables synergistic optimization of meteorological representation learning and power prediction objectives. To address the mismatch between the global scale of the large meteorological model and the local power prediction scenario of photovoltaic clusters, this application employs regional meteorological domain adaptive fine-tuning. This allows the meteorological backbone to gradually adapt from general meteorological field predictions to the climate characteristics of the target region, thereby bridging the scale difference between the global meteorological model and the local power prediction task. To address the modal difference between continuous meteorological field representations and discrete multi-site power sequences, this application utilizes a cross-level meteorological and power heterogeneous representation coupling mechanism. This mechanism progressively injects multi-scale meteorological knowledge into multi-site power representations and combines cross-attention mechanisms and geographical distance bias to achieve structured interaction between the meteorological field and the power field. Furthermore, by introducing solar geometric priors to physically constrain the prediction process, this application enhances the model's ability to perceive intraday photovoltaic power generation patterns, ensuring high prediction accuracy and stability under complex weather scenarios such as cloud movement, sudden irradiance changes, and power ramp-ups. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1 A flowchart illustrating a photovoltaic cluster power prediction method based on a progressive cross-modal migration model using a large meteorological model, provided for embodiments of this application; Figure 2 A structural diagram of a multi-site spatiotemporal representation encoder based on a geographic map provided in an embodiment of this application; Figure 3 A structural diagram of the cross-level meteorological and power heterogeneous characterization coupling mechanism provided in the embodiments of this application; Figure 4 A diagram illustrating the bidirectional Mamba power sequence modeling structure with enhanced solar geometry priors provided in this application embodiment; Figure 5 A comparison chart of photovoltaic cluster power prediction curves between the proposed method and different benchmark models provided in the embodiments of this application; Figure 6 A comparison chart of photovoltaic cluster power prediction curves for different component ablation models provided in the embodiments of this application.

[0024] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0026] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0027] For photovoltaic (PV) cluster power prediction, existing methods mainly rely on numerical weather prediction to drive deep learning models for statistical fitting. This separation of meteorological forecasting and power modeling easily leads to the propagation of meteorological errors to the power prediction stage. Furthermore, the limited explicit meteorological variables cannot fully represent the multi-scale spatiotemporal correlations in regional weather processes, limiting the model's ability to characterize the power fluctuation patterns of PV clusters. Therefore, this application provides a PV cluster power prediction method based on a progressive cross-modal transfer model using a large meteorological model. This method adapts regional meteorological representations, gradually adapting the meteorological backbone from general meteorological field predictions to the climatic characteristics of the target PV cluster region. By coupling cross-level meteorological and heterogeneous power representations, multi-scale meteorological knowledge is injected into the power representations of multiple sites, and solar geometric information is integrated to enhance the model's ability to perceive intraday PV power generation patterns. Experiments based on PV power plant data from a specific region show that, compared to the traditional two-stage combined method, the proposed method reduces NRMSE and NMAE by an average of 6.28% and 16.97%, respectively, validating its effectiveness.

[0028] like Figure 1 As shown, this method uses the Pangoal Meteorological Model as the meteorological backbone. Its three-dimensional Earth-specific transformer jointly models the atmospheric state in the dimensions of pressure layer, latitude and longitude. It also combines the U-Net encoding and decoding structure to extract multi-scale meteorological features. The method is progressively advanced according to the path of "global meteorological knowledge learning - regional meteorological representation adaptation - power prediction task transfer", including three stages: global meteorological self-supervised pre-training, regional meteorological domain adaptive fine-tuning, and cross-modal multi-scale task transfer from meteorology to power.

[0029] The first phase, based on the global ERA5, enables meteorological backbones to obtain a universal multivariate meteorological evolution characterization.

[0030] The second stage uses data from the target area to adjust the distribution of meteorological characteristics, shifting the model from a global scale to the target area.

[0031] The third stage will transfer regional meteorological characteristics to photovoltaic cluster power prediction tasks, achieving further adaptation from meteorological prediction to power prediction.

[0032] Specifically, the photovoltaic cluster power prediction method based on the gradual cross-modal migration of the meteorological big model includes the following steps S10-S30, which correspond to the three stages mentioned above.

[0033] S10: Obtain global meteorological reanalysis data, use a large meteorological model as the meteorological backbone, conduct global meteorological self-supervised pre-training on the meteorological backbone, and obtain general meteorological characterization.

[0034] Step S10 corresponds to the first stage. The first stage involves conducting self-supervised meteorological pre-training on global ERA5 reanalysis data. In this embodiment, a spatial window consistent with the target region's scale is cropped from the global meteorological field. Using historical meteorological fields as input, future meteorological conditions are predicted, allowing the model to learn the coupling relationships between multiple variables such as temperature, humidity, wind field, and radiation under different regions and weather conditions. This stage is not targeted at a specific photovoltaic scenario but rather establishes a transferable, general meteorological representation through a broad climate distribution, providing parameter initialization for subsequent regional adaptation. Considering the differences in numerical range, prediction difficulty, and photovoltaic driving effects among different meteorological variables, a variable-weighted prediction loss is used for training. Simultaneously, a spatial gradient smoothing term is introduced to constrain changes between adjacent grid points, reducing unreasonable spatial abrupt changes while maintaining local gradient characteristics in the prediction results. (1) (2) in, For meteorological grids ( i , j The predicted value vector at position ) i , j For spatial grid indexing; For meteorological grids ( i+ 1, j The predicted value vector at position ) For the first v Predicted values ​​of several meteorological variables To correspond to the actual value, For the first v The penalty weights of each weather variable, The weighted prediction loss for variables in the upper atmosphere is used. The balance coefficient for the spatial gradient smoothing term; For spatial gradient smoothing, The loss function is the global meteorological self-supervised pre-training function.

[0035] In this embodiment =0.1. Through the above training, meteorological backbone personnel acquire general spatiotemporal evolution knowledge of multivariable meteorological fields, and provide a foundation for learning meteorological characteristics of target areas.

[0036] S20: Obtain regional meteorological data for the target area, use general meteorological characteristics as initial model parameters for the meteorological backbone, and perform regional meteorological domain adaptive fine-tuning on the meteorological backbone to obtain a regional meteorological backbone adapted to the climate characteristics of the target area.

[0037] Step S20 corresponds to the second stage. The second stage uses global pre-trained weights as initialization and uses ERA5 data of the target area to perform domain adaptive fine-tuning of the meteorological backbone. Since the global meteorological data covers a wide range of climate types and spatial distributions, its feature distribution still differs from that of the target area. Therefore, it is necessary to further adjust the model's representation of the cloud cover, radiation, and temperature and humidity variation patterns in the target area. In order to avoid excessive disturbance of general meteorological knowledge due to the limited scale of regional data, the shallow feature extraction parameters are frozen in the early stage of training, and only the deep encoding and decoding modules are updated, so that the model prioritizes the adjustment of high-level meteorological semantics; then the remaining parameters are gradually unfrozen so that the shallow local features and deep regional representations can be adapted to the target climate distribution in a coordinated manner. In view of the problem that a large number of low-value samples in the nighttime radiation variable may weaken the daytime prediction weight, the fixed variables in equation (2) are... Replace with dynamic weights : (3) (4) in, The zenith angle of the sun. For daytime masking, , The weights for day and night variables are respectively. This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise; therefore, when hour, This indicates daytime; otherwise, it indicates nighttime. This strategy reduces the dominant role of low-radiation samples at night in the loss function, allowing the model to focus more on daytime radiation and cloud cover changes. It also promotes the gradual adaptation of the meteorological backbone from a globally universal representation to the climate characteristics of the target region, providing a regionalized representation basis for subsequent meteorological-to-power tasks.

[0038] S30: Obtain historical power sequences and installed capacity of multiple photovoltaic power plants within the target area, as well as the geographical location information and solar geometric information of the photovoltaic power plants; extract multi-site spatiotemporal representations based on historical power sequences and geographical location information, and enhance the multi-site spatiotemporal representations based on installed capacity to obtain site representations that integrate installed capacity information; based on regional meteorological backbone, establish cross-modal relationships between continuous meteorological fields and discrete multi-site power sequences using the site representations that integrate installed capacity information as queries to obtain the final site representations that integrate multi-scale meteorological information; generate normalized power prediction values ​​for each photovoltaic power plant in the future prediction period based on the final site representations and solar geometric information.

[0039] Step S30 corresponds to the third stage. Based on the regional meteorological backbone, the third stage establishes cross-modal connections between continuous meteorological fields and discrete multi-site power sequences. This stage consists of three modules: multi-site spatiotemporal representation encoding, cross-level meteorological and power heterogeneous representation coupling, and BiMamba power prediction. These modules are used to extract spatiotemporal features on the power side, aggregate multi-scale meteorological information, and model future power sequences, respectively.

[0040] Within a photovoltaic cluster, each power station is affected by shared weather processes and local environmental factors, resulting in power sequences exhibiting both intra-station temporal dependence and inter-station spatial correlation. To address this, this embodiment constructs a multi-site spatiotemporal representation encoder based on a geographic map, alternately employing temporal convolution and graph message passing to extract intra-site dynamics and inter-site collaborative features, as shown in the structure below. Figure 2 As shown.

[0041] Using photovoltaic power plants as graph nodes, graph connections are established based on the geographical distance between the sites. The Haversine formula is used to calculate the number of sites. With the site The spherical distance between them: (5) (6) in, R The average radius of the Earth and Sites i and j latitude and longitude and These represent the difference in latitude and longitude between the two stations, respectively. This is an intermediate variable used in the calculation of Haversine spherical distance.

[0042] Select each station based on the distance between stations. K A neighboring site is identified, and an undirected graph is constructed through symmetry processing. Using the median of the effective edge distance as the scale parameter, a Gaussian kernel is used to map geographical distances to continuous edge weights, which are then normalized to obtain a weighted adjacency matrix. : (7) in, Here, KNN is the adjacency indicator function, and KNN is the set of K nearest neighbors obtained after symmetrization. e It is a natural constant. k Summing index for neighboring sites, d ik For the site i With the site k The spherical distance between them is the distance scale parameter, which takes the value of the median distance of effective connected edges. Edge weights are calculated only for stations within the nearest neighbor range. This weighted adjacency matrix preserves the local graph topology while adjusting the information contribution of different neighboring stations based on geographical distance.

[0043] In addition to the relative distances between stations, the model also needs to retain the absolute geographical locations of the stations. (The last part, " Put the stations...", appears to be an error and is left untranslated.) i The latitude and longitude codes are And after linear projection, the initial power characteristics are added. Let... N The historical power sequence length of the site is T The initial features are obtained after linear mapping. The encoder then alternately performs two rounds of temporal convolution and graph messaging, enabling the site representation to progressively interact between intra-site temporal dynamics and inter-site spatial associations.

[0044] In the l In round coding, a one-dimensional convolution is first performed on the time series data of each station to extract intraday variations and local power fluctuations. Let the input of this round be... The temporal convolution operation is defined as follows: (8) The convolutional features are then subjected to a nonlinear transformation via a feedforward network: (9) in, Represents one-dimensional convolution. For GELU activation function, For convolution weights, , and For learnable linear transformation matrices, For the first l The input features of round coding These are intermediate features after temporal convolution. These are intermediate features output by the feedforward network. Convolutions are used to extract local temporal patterns, and residual connections are used to preserve the original temporal information and stabilize feature updates.

[0045] After extracting temporal features, the features of neighboring sites are aggregated using a weighted adjacency matrix, and the spatial association between sites is injected into the current site representation. (10) After linear transformation, the aggregated message is added to the current site features in the form of residuals, and then nonlinear fusion is completed through a feedforward network. (11) (12) in The normalized adjacency weights defined in equation (7) For neighboring sites j The intermediate features output by the feedforward network, For the site i The neighbor messages obtained through aggregation The message transformation matrix, and For feedforward network parameters, The intermediate features after fusing neighbor messages, For the first l The station features are output by round-encoding. Temporal convolution and graph message passing are executed alternately for two rounds, enabling the model to extract the power dynamics of a single station while incorporating the cooperative change information of neighboring power stations.

[0046] After two rounds of spatiotemporal encoding, a learnable attention pooling method is used to converge the time dimension, allowing the model to assign corresponding weights to different time points. The attention score is: (13) in, For the site after the second round of coding i At any moment t Features For the site i At any moment t Attention score This is a learnable attention vector used to compute the temporal attention score.

[0047] After softmax normalization, the features at each time step are weighted and summed to obtain the station-level representation: (14) in, For the site i Site-level representation, t For the current moment in the historical power sequence, T The number of time steps in the historical power series. For the summation in the attention normalization process, For the site i exist Attention score.

[0048] in This is a learnable attention vector. Considering the differences in installed capacity among different power plants, the normalized installed capacity is linearly projected and added to the site representation: (15) in, This is the normalized scalar of installed capacity. For the projection matrix, This is a site representation that integrates installed capacity information.

[0049] The final site characterization includes historical power temporal dynamics, inter-site spatial correlation, absolute geographical location and installed capacity information, and serves as the power-side input for subsequent cross-modal coupling.

[0050] The features of different levels of meteorological backbone have different spatial resolutions and semantic ranges. Directly splicing meteorological fields and power features not only makes it difficult to establish the correspondence between meteorological grid points and discrete power stations, but also makes it difficult to fully utilize the complementary information of meteorological representations at different levels. Layer 1, located at the high-resolution encoding end, mainly retains fine-grained information such as local cloud cover distribution and radiation gradient. Layers 2 and 3 are downsampled to describe the weather background over a larger spatial area. Layer 4, located at the decoding end, recovers local spatial details after upsampling. To simultaneously utilize local variations and large-scale weather processes, this embodiment uses station representations as queries, progressively aggregating meteorological information across four levels. It also combines geographical distance bias and surface-upper-level dual-flow gating to achieve coupling between multi-scale meteorological fields and multi-station power representations, as shown in the structure below. Figure 3 As shown.

[0051] Let the current meteorological characteristic sequence be... ,in L This represents the number of meteorological grid points at this scale. d As a feature dimension. Based on site i The representation As a query, the meteorological feature serves as both the key and value, and is obtained through linear projection: (16) in, For learnable projection matrices, For the site i The query vector consists of K and V, which are the key and value matrices of the meteorological features, respectively. Conventional cross-attention methods assign weights solely based on feature similarity, which may cause stations to focus on distant meteorological grid points. To incorporate the geographical relationship between stations and meteorological grid points, the spherical distance between them is used. Mapped to attention bias: (17) in, For learnable positive scalars, For the site i With meteorological grid j Geographical distance between them leads to attention bias.

[0052] set up k j and v j Grid points j The key vector and value vector, d k If the key vector dimension is used, then the attention score, normalized weights, and enhanced site representation are as follows: (18) (19) (20) in, s ij For the site i Meteorological grid j Attention score To normalize the weights, This is a normalized exponential function used to convert the attention scores corresponding to each meteorological grid point into normalized weights that sum to 1. l For meteorological backbone hierarchical index, r Indicates surface or upper-air meteorological features, For the first l The number of meteorological grid points at each level j For meteorological grid indexes.

[0053] Thus, each station can aggregate meteorological information relevant to itself from a continuous meteorological field under the combined constraints of feature correlation and geographical proximity.

[0054] Surface meteorological variables primarily describe near-surface radiation, cloud cover, and temperature and humidity conditions, while upper-air meteorological variables reflect circulation, humidity, and cloud water distribution at different pressure levels. These two types of information are complementary. Therefore, the aforementioned cross-attention method is applied to both surface and upper-air meteorological features at each scale to obtain enhanced surface and upper-air representations, which are then fused using learnable gating. (twenty one) in, For the first l Each level of gating parameter is used to adjust the relative contribution of surface meteorological information and upper-air meteorological information at that level. For the first lEnhanced site representation after fusion of multiple levels.

[0055] The regional meteorological backbone comprises four layers, from Layer 1 to Layer 4, with station representations updated sequentially in the order of Layer 1, Layer 2, Layer 3, and Layer 4. Layer 1 consists of high-resolution coded features, Layers 2 and 3 are downsampled low-resolution deep coded features, and Layer 4 is upsampled decoded features. Each layer includes both surface meteorological features and upper-air meteorological features.

[0056] The input to Layer 1 is the site representation of the fused installed capacity information obtained by equation (15). Using the site representation of the fused installed capacity information as the query, the surface meteorological features and upper-air meteorological features of Layer 1 are used as keys and values, respectively. Cross-attention is performed according to equations (16) to (20) to obtain the surface enhancement representation and upper-air enhancement representation of Layer 1. The surface enhancement representation and upper-air enhancement representation are then gated and fused according to equation (21) to obtain the updated site representation of Layer 1. The updated site representation of Layer 1 incorporates local cloud cover, radiation variation and near-surface meteorological information from the high-resolution meteorological features.

[0057] The input to Layer 2 is the site representation obtained by dimensionality reduction projection of the updated site representation of Layer 1. The dimensionality reduction projection is used to make the updated site representation of Layer 1 have the same feature dimension as the meteorological features of Layer 2 and Layer 3. Using the dimensionality-reduced site representation as the query, the surface meteorological features and upper-air meteorological features of Layer 2 are used as keys and values, respectively. Cross-attention is performed according to equations (16) to (20) to obtain the surface enhancement representation and upper-air enhancement representation of Layer 2. Then, gated fusion is performed according to equation (21) to obtain the updated site representation of Layer 2. The updated site representation of Layer 2 retains the local meteorological information of Layer 1 and further integrates the meteorological information of a larger spatial range represented by Layer 2.

[0058] The input to Layer 3 is the updated site representation of Layer 2. Since Layer 2 and Layer 3 have the same feature dimensions, no further dimensional transformation is performed on the updated site representation of Layer 2. Using the updated site representation of Layer 2 as the query, the surface meteorological features and upper-air meteorological features of Layer 3 are used as keys and values, respectively. Cross-attention is performed according to equations (16) to (20) to obtain the enhanced surface representation and enhanced upper-air representation of Layer 3. Then, gating fusion is performed according to equation (21) to obtain the updated site representation of Layer 3. The updated site representation of Layer 3 is further fused with the deep regional weather background information represented by Layer 3.

[0059] Layer 4's input is the upgraded site representation obtained by projecting the updated site representation from Layer 3 into Layer 1. Specifically, before entering Layer 4, the updated site representation from Layer 3 is projected into Layer 1 to make its feature dimensions consistent with the meteorological feature dimensions of Layer 4. Then, the updated site representation from Layer 1 is introduced through a skip connection to obtain the upgraded site representation, denoted as […]. Through the aforementioned upscaling projection and skip connections, the input station representation of Layer 4 retains the meteorological information of the deep regions of Layer 2 and Layer 3 while reintroducing high-resolution information related to local cloud cover, radiation variation, and near-surface meteorological conditions from Layer 1.

[0060] The upgraded site representation For querying, the surface meteorological features and upper-air meteorological features of Layer 4 are used as keys and values, respectively. Cross-attention is performed according to equations (16) to (20) to obtain the Layer 4 surface flow cross-attention enhancement characterization and upper-air flow cross-attention enhancement characterization, respectively. The Layer 4 incremental characterization is then calculated using the following formula: (twenty two) in, For Layer 4 surface-high altitude fusion gating, This is an incremental representation of Layer 4.

[0061] Then, residual gating parameters were used. Integrate Layer 4 incremental representations into existing site representations using interpolation: (twenty three) in, To provide the final station representation that integrates multi-scale meteorological information, To introduce the site representation updated from Layer 1 before entering Layer 4, through dimensional projection. To control the access amplitude of features at the decoding end, this design enables the model to progressively utilize the local meteorological details recovered from Layer 4 while preserving the multi-scale representations of the first three layers. Final station representation. It also includes local meteorological textures, large-scale weather backgrounds, and historical power information from multiple stations, which are used as conditional inputs to the power prediction head.

[0062] After fusing multi-scale meteorological information, the site-level representation needs to be converted into a future power sequence. The intraday variation of photovoltaic power is closely related to the solar trajectory, and there is also a continuous evolutionary relationship between the predicted times. Therefore, this embodiment uses the final site representation as a condition, introduces a solar geometric prior, and employs bidirectional Mamba modeling to model the future power sequence, as shown in the structure... Figure 4 As shown. First, the site is characterized. Expand along the prediction time dimension and overlay the learnable temporal location codes corresponding to the prediction steps. The initial predicted time-series features are obtained: (twenty four) in, H To predict the number of time steps, For the site n The final site representation, For the site n In the h Initial prediction time series characteristics at each prediction time point.

[0063] To enable the model to perceive the solar position conditions at different stations and prediction times, the solar altitude angle, solar azimuth angle, upper limit of clear-sky irradiance, and daytime conditions are calculated based on the station's latitude and longitude and the prediction time, thus constructing a 6-dimensional solar geometric feature: (25) in, For the site n In the h The solar altitude angle at the predicted moment. The azimuth of the sun. To illuminate the upper realm of clear skies, As a daytime mask, the solar geometric features are linearly projected and then added to the predicted temporal features: (26) in, This is the auxiliary feature projection matrix.

[0064] By incorporating solar position and clear-sky irradiance conditions at each prediction time, the predicted time-series characteristics can distinguish intraday power generation conditions at different sites and at different times. (The last sentence appears to be incomplete and possibly refers to a separate topic: "sites...") n In the future HThe fused prediction temporal features at each prediction time point are arranged in chronological order and input into the forward scan branch and the reverse scan branch, respectively. The forward scan branch runs from the first prediction time point to the second prediction time point. H The fused prediction temporal features are processed sequentially according to the direction of each prediction time, and the first... t Step input for The reverse scan branch first flips the fused prediction temporal features in chronological order, and then processes them sequentially. t Step input for .

[0065] Each Mamba module projects the input into two paths: one path, after depthwise separable convolution and SiLU activation, enters a selective state space scan to extract local variation patterns and long-range temporal dependencies; the other path serves as a gating signal to selectively filter the state space scan results. The continuous state space parameters are discretized using a zero-order preserve method as follows: (27) (28) in, To enable learning step size, It is the identity matrix. , For continuous state-space parameters, , These are the discretized state transition matrix and input matrix.

[0066] Based on the discretization parameters, the forward scan branch and the reverse scan branch are updated using the following formulas: Determine the hidden state of each step and calculate the corresponding output: (29) (30) in, and The first t Step and the first t -1 step hidden state For the first t Step input, For the first t Step output, From the current input Dynamically generated For jump connection coefficients. Inputting relevant state parameters enables the model to adjust state updates according to power and meteorological characteristics at different times, which is beneficial for characterizing non-stationary processes such as uphill, sudden drop, and peak changes.

[0067] The forward scan branch generates the output for each prediction time according to the original prediction time sequence, and stores the stations... n In the h Output corresponding to each prediction time Recorded as The reverse scan branch scans the fused prediction temporal features after the time sequence has been flipped, and restores the output to the original prediction time sequence, including the site... n In the h Output corresponding to each prediction time Recorded as .therefore and These represent the forward scan output and the reverse scan output at the same station and the same prediction time, respectively.

[0068] The forward and reverse scan outputs are concatenated along the feature dimension and then fused using linear mapping and layer normalization. (31) in, This is the output of the forward scan. This is the output of the reverse scan. To fuse the projection matrix, LN is the layer normalization. This is the final fusion feature.

[0069] The final fused features are mapped to normalized power prediction values ​​through a two-layer feedforward network and an activation function: (32) in, For the site n In the h Normalized power prediction values ​​at each prediction time. For the site n In the h The final fusion features at each prediction time point, , This is the weight matrix of the feedforward network. Through the above process, the model completes the joint mapping from multi-scale meteorological fields, historical power at multiple sites, and solar geometry conditions to the power sequence for the next 24 hours, realizing the effective transfer of knowledge from large meteorological models to photovoltaic cluster power prediction tasks.

[0070] To further illustrate the feasibility and progressiveness of the method in this application, an exemplary case study will be provided below.

[0071] The analysis samples used in this embodiment involve global ERA5 reanalysis data, target region ERA5 reanalysis data, and data from 48 photovoltaic power plants in the target region. The photovoltaic data includes the historical power and latitude / longitude of each power plant, with a total installed capacity of 1310MW. ERA5 includes the surface layer and the upper atmosphere. The surface layer contains 7 near-surface variables, and the upper atmosphere contains four pressure layers (850, 700, 500, and 200 hPa), each containing 6 variables. The data configuration for the three-stage progressive training is shown in Table 1.

[0072] Table 1 Data configuration for three-stage progressive training

[0073] All experiments were implemented using PyTorch on an NVIDIA L20 GPU with 48 GB of VRAM and CUDA version 12.8. Training employed BF16 mixed precision and the AdamW optimizer, with three training epochs of 200, 100, and 30, and initial learning rates of [missing information]. , and The weight decays are respectively , and The equivalent batch sizes are 192, 32, and 64, respectively. Preheating and cosine annealing learning rate scheduling are used in each stage, with phases two and three employing gradual thawing. The model uses the ERA5 weather field and historical power data from the past 96 hours as input to predict photovoltaic power for the next 24 hours.

[0074] This embodiment uses normalized root mean square error (NRMSE), normalized mean absolute error (NMAE), and correlation coefficient (R) as evaluation indicators, which are defined as follows: (33) (34) (35) in n For the number of test samples, This is the actual power. To predict power, This is the average actual power. To predict the average power, This represents the total installed capacity. NRMSE, NMAE, and R in the table are the average values ​​of the evaluation indicators for each time point from 1 to 24 hours in the future. The forecast curve only shows the comparison between the actual power and the forecast power for the 24th hour in the future.

[0075] To evaluate the proposed method, this embodiment sets up two types of experiments: 1) Experiment I compares the proposed method with classical time series models, basic time series models, and language models.

[0076] 2) Experiment II analyzes the role of multi-scale meteorological coupling and solar geometric priors through component ablation.

[0077] Experiment I: Model Comparison Experiment To verify the effectiveness of the large-scale meteorological model in task transfer, this experiment compares the classic deep time series model LSTM with ModernTCN, the large-scale time series model Chronos-Bolt with TimesFM, and the large-scale language model GPT4TS with Time-LLM. As shown in Table 2, the method proposed in this application leads in three aspects: NRMSE=0.0575, NMAE=0.0301, and R=0.9463.

[0078] Table 2 Model Comparison Results

[0079] Different baseline models employ various approaches to photovoltaic power prediction, including recurrent networks, temporal convolution, general temporal pre-training, and language model transfer. LSTM and ModernTCN emphasize time-dependent modeling, Chronos-Bolt and TimesFM primarily utilize general temporal pre-training knowledge, while GPT4TS and Time-LLM leverage the sequence representation capabilities of language models. In contrast, the proposed method further utilizes regional meteorological representations and their cross-level interactions with power state to more fully explore the correlation between meteorological evolution and photovoltaic power changes, thus achieving superior results across various evaluation metrics.

[0080] The proposed method introduces multi-scale meteorological field representation through large-scale meteorological model transfer, and combines historical power and solar geometry information, enabling the model to simultaneously perceive weather changes and intraday power generation patterns. Figure 5 It is evident that its prediction curve is closer to the true value during the climbing, peak, and decay phases, indicating that meteorological driving characteristics help improve the prediction of amplitude and timing of changes during key periods.

[0081] To further analyze the role of multi-scale meteorological coupling and solar geometric prior, this embodiment constructs four ablation variants based on the complete model. Shallow only retains only Layer 1 shallow coupling; Deep encoder only retains only Layer 2–3 deep coupling; Decoder only retains only Layer 4 decoder coupling; and solar geometric prior is removed, while other settings remain consistent.

[0082] like Figure 6As shown, the three single-scale variants all exhibit varying degrees of amplitude deviation in the peak segment. After removing the solar geometric prior, the phase and amplitude errors near sunrise ramp and sunset decay are more pronounced.

[0083] Table 3 Component ablation results

[0084] As shown in Table 3, the NRMSE and NMAE of the three single-scale variants are all higher than those of the complete model, indicating that a single layer cannot fully describe meteorological processes at different spatial scales. Among the three single-scale models, the Decoder-only variant achieves the lowest NRMSE and NMAE, suggesting that the high-resolution local features recovered by upsampling at the decoder end help characterize cloud cover and radiation variations. The relatively high R-value of shallow-layer coupling indicates that shallow local textures also play an important role in power variation trends. However, shallow features lack large-scale weather background, while deep features lose some local details; therefore, a four-scale layer-by-layer coupling approach is needed to achieve complementarity.

[0085] Removing solar geometric priors improved both NRMSE and NMAE, with a greater performance degradation than the three single-scale variants. Solar elevation angle, azimuth angle, clear-sky irradiance upper bound, and daytime mask provided explicit intraday temporal constraints for the prediction head, thereby reducing the temporal location and amplitude biases during sunrise, peak, and sunset phases.

[0086] The above results indicate that multi-scale meteorological coupling is responsible for characterizing weather drivers, while solar geometric priors are responsible for supplementing intraday physical constraints.

[0087] The experimental results of the photovoltaic clusters described above demonstrate that the proposed method exhibits good predictive performance in various comparative experiments. Further ablation experiments show that regional meteorological domain adaptation improves the model's adaptability to local climate distribution, cross-level meteorological-power coupling enhances the correlation modeling between the meteorological field and multi-site power, and solar geometric priors and bidirectional scanning improve the characterization of intra-diurnal photovoltaic variations. These results indicate that progressive transfer can effectively transform general meteorological knowledge into a task representation for photovoltaic cluster prediction.

[0088] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A photovoltaic cluster power prediction method based on a progressive cross-modal migration of a large meteorological model, characterized in that, The method includes: Global meteorological reanalysis data is acquired, and a large meteorological model is used as the meteorological backbone. The meteorological backbone is then subjected to global meteorological self-supervised pre-training to obtain a general meteorological characterization. Acquire regional meteorological data for the target area, use the general meteorological characterization as the initial model parameters of the meteorological backbone, and perform regional meteorological domain adaptive fine-tuning on the meteorological backbone to obtain a regional meteorological backbone adapted to the climate characteristics of the target area. The process involves acquiring historical power sequences and installed capacity of multiple photovoltaic (PV) power plants within the target area, along with their geographical location information and solar geometric information. Multi-site spatiotemporal representations are extracted based on the historical power sequences and geographical location information, and enhanced according to the installed capacity to obtain site representations that integrate installed capacity information. Based on the regional meteorological backbone, a cross-modal relationship is established between the continuous meteorological field and the discrete multi-site power sequences, using the site representations with integrated installed capacity information as the query, to obtain the final site representations that integrate multi-scale meteorological information. Based on the final site representations and the solar geometric information, normalized power prediction values ​​for each PV power plant are generated for the future prediction period.

2. The method according to claim 1, characterized in that, The aforementioned meteorological backbone personnel underwent global meteorological self-monitoring pre-training, including: A spatial window consistent with the target area scale is cropped from the global meteorological reanalysis data. Using historical meteorological fields as input, future meteorological conditions are predicted, and the meteorological backbone is trained by combining variable-weighted prediction loss with a spatial gradient smoothing term. Define spatial gradient smoothing term As shown in the formula below: (1) in, For meteorological grids ( i , j The predicted value vector at position ) i , j For spatial grid indexing; For meteorological grids ( i+ 1, j The predicted value vector at () The global pre-training loss function is constructed as shown in the following formula: (2) in, For the first v Predicted values ​​of several meteorological variables To correspond to the actual value, For the first v The penalty weights of each weather variable, The weighted prediction loss for variables in the upper atmosphere is used. The balance coefficient for the spatial gradient smoothing term; For spatial gradient smoothing, The loss function is the global meteorological self-supervised pre-training function. The meteorological backbone is trained by minimizing the global pre-training loss function until the global pre-training loss function converges, and the trained meteorological backbone is obtained. The model parameters of the trained meteorological backbone constitute the general meteorological representation.

3. The method according to claim 2, characterized in that, The meteorological backbone is subjected to regional meteorological domain adaptive fine-tuning, including: The meteorological backbone is initialized using the general meteorological representation, and fine-tuned using regional meteorological data of the target area. The global pre-training loss function is then adjusted accordingly. Replace with dynamic weights The formula for calculating the dynamic weight is: (3) (4) in, The zenith angle of the sun. For daytime masking, , The weights for day and night variables are respectively. This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise; therefore, when hour, Indicates daytime; otherwise, it indicates nighttime. Using the replaced loss function as the optimization objective, the shallow feature extraction parameters are frozen in the early stage of training, the deep encoding module and decoding module are updated, and then the remaining parameters are gradually unfrozen to fine-tune the meteorological backbone so that the meteorological feature distribution of the meteorological backbone is adapted from the global scale to the climate features of the target region, thus obtaining the regional meteorological backbone.

4. The method according to claim 1, characterized in that, Obtaining the geographical location information of multiple photovoltaic power stations within the target area, including: Obtain the latitude and longitude coordinates of each photovoltaic power station, and establish graph connections based on the geographical distance between the stations, with each photovoltaic power station as a graph node. The site is calculated using the following formula. i With the site j spherical distance between : (5) (6) in, R The average radius of the Earth and Sites i and j latitude and longitude and These represent the difference in latitude and longitude between the two stations, respectively. This is an intermediate variable in the calculation of the Haversine spherical distance; According to the site i With the site j spherical distance between For each site, select K neighboring sites, construct an undirected graph through symmetry processing, use the median of effective edge distances as the scaling parameter, map geographical distances to continuous edge weights using a Gaussian kernel, and obtain a weighted adjacency matrix after normalization. : (7) in, Here, KNN is the adjacency indicator function, and KNN is the set of K nearest neighbors obtained after symmetrization. e It is a natural constant. k Summing index for neighboring sites, d ik For the site i With the site k The spherical distance between them This is the distance scale parameter, and its value is the median of the effective connected edge distances; The geographic location information includes the latitude and longitude coordinates of each photovoltaic power station and the weighted adjacency matrix.

5. The method according to claim 4, characterized in that, Based on the historical power sequence and the geographical location information, multi-site spatiotemporal representations are extracted, and the multi-site spatiotemporal representations are enhanced according to the installed capacity to obtain site representations that integrate installed capacity information, including: The latitude and longitude coordinates of each station are encoded into absolute position codes, and after linear projection, initial power features are added to obtain initial station features; Using the initial site features as input, temporal dynamic features within each site are extracted through temporal convolution to obtain intermediate features after temporal convolution; spatial association features of neighboring sites are aggregated through graph messaging to obtain site features output in this round of encoding; the temporal convolution and graph messaging are executed alternately for two rounds to obtain a multi-site spatiotemporal representation; The spatiotemporal representations of the multiple sites are weighted and aggregated along the time dimension to obtain the site-level representation of each site; the site-level representations are enhanced according to the installed capacity of each site to obtain the site representation that integrates the installed capacity information.

6. The method according to claim 5, characterized in that, Temporal dynamic features within each site are extracted using temporal convolution, resulting in intermediate features after temporal convolution, including: In the l In round coding, the time series data of each station are subjected to one-dimensional convolution and then nonlinear transformation through a feedforward network, as shown in the following formula: (8) (9) in, Represents one-dimensional convolution. For GELU activation function, For convolution weights, , and For learnable linear transformation matrices, For the first l The input features of round coding These are intermediate features after temporal convolution. These are intermediate features output by the feedforward network.

7. The method according to claim 5, characterized in that, By aggregating the spatial association features of neighboring sites through graph messaging, the site features output in this round of encoding are obtained, including: The weighted adjacency matrix is ​​used to aggregate neighbor site features, and nonlinear fusion is completed through a feedforward network, as shown in the following formula: (10) (11) (12) in, For neighboring sites j The intermediate features output by the feedforward network, For the site i The neighbor messages obtained through aggregation The message transformation matrix, and For feedforward network parameters, The intermediate features after fusing neighbor messages, For the first l Station characteristics output by round coding.

8. The method according to claim 5, characterized in that, The spatiotemporal representations of the multiple sites are weighted and converged along the time dimension to obtain a site-level representation for each site. Then, the site-level representations are enhanced based on the installed capacity of each site to obtain a site representation that integrates installed capacity information, including: After two rounds of spatiotemporal encoding, a learnable attention pooling method is used to converge the time dimension, and the attention score is calculated using the following formula: (13) in, For the site after the second round of coding i At any moment t Features For the site i At any moment t Attention score This is the learnable attention vector used to compute the temporal attention score; Based on the attention score, the site-level representation is calculated using the following formula: (14) in, For the site i Site-level representation, t For the current moment in the historical power sequence, T The number of time steps in the historical power series. For the summation in the attention normalization process, For the site i exist Attention score; Based on the site-level representation, the site representation with integrated installed capacity information is calculated using the following formula: (15) in, This is a normalized scalar representation of installed capacity. For the projection matrix, This is a site representation that integrates installed capacity information.

9. The method according to claim 5, characterized in that, Based on the regional meteorological backbone, and using the site representation of the fused installed capacity information as a query, a cross-modal relationship is established between the continuous meteorological field and the discrete multi-site power sequence, resulting in the final site representation of the fused multi-scale meteorological information, including: Obtain the current meteorological feature sequence , with site i The site representation, which integrates installed capacity information, is used as the query, while meteorological features serve as both keys and values. The query vector, key matrix, and value matrix are calculated using the following formulas: (16) in, For learnable projection matrices, For the site i The query vector, K , V These are the key matrix and value matrix of meteorological characteristics, respectively; According to the spherical distance The geographic distance attention bias is calculated using the following formula: (17) in, For learnable positive scalars, For the site i With meteorological grid j Geographical distance between them leads to attention bias; Based on the query vector, key matrix, and geographic distance attention bias, the attention score is calculated using the following formula: (18) in, k j Meteorological grid j The key vector, v j Meteorological grid j The value vector, d k The dimension of the key vector. s ij For the site i Meteorological grid j Attention score; Based on the attention score, the normalized weights are calculated using the following formula: (19) in, To normalize the weights, This is a normalized exponential function used to convert the attention scores corresponding to each meteorological grid point into normalized weights that sum to 1; The site is calculated using the following formula based on the normalized weights and the value matrix. i In the l Meteorological backbone level, meteorological flow r Enhanced representation under : (20) in, l For meteorological backbone hierarchical index, r Indicates surface or upper-air meteorological features, For the first l The number of meteorological grid points at each level j For meteorological grid indexing; Each meteorological backbone level performs the cross-attention function on surface meteorological features and upper-air meteorological features respectively to obtain enhanced surface characterization. and high-altitude enhanced characterization The surface enhancement characterization and the upper-air enhancement characterization are fused using the following formula: (21) in, For the first l Each level of gating parameter is used to adjust the relative contribution of surface and upper-air meteorological information at that level. Enhance the representation of the merged sites; The regional meteorological backbone comprises four layers, from Layer 1 to Layer 4. Station representations are updated progressively in the order of Layer 1, Layer 2, Layer 3, and Layer 4. Layer 1 consists of high-resolution coded features, Layers 2 and 3 are downsampled low-resolution deep coded features, and Layer 4 is upsampled decoded features. Using the station representations as queries, cross-attention and gating fusion is performed with the surface meteorological features and upper-air meteorological features of Layer 1, respectively, to obtain the updated station representations for Layer 1. Before entering Layer 2, the updated station representations of Layer 1 undergo dimensionality reduction projection to match Layer 4. The feature dimensions of Layer 2 and Layer 3 are used to obtain the dimensionality-reduced site representation. Using the dimensionality-reduced site representation as a query, the cross-attention and gating fusion is performed with the surface meteorological features and upper-air meteorological features of Layer 2 respectively to obtain the updated site representation of Layer 2. Using the updated site representation of Layer 2 as a query, the cross-attention and gating fusion is performed with the surface meteorological features and upper-air meteorological features of Layer 3 respectively to obtain the updated site representation of Layer 3. Before entering Layer 4, the updated site representation of Layer 3 is subjected to dimensionality-up projection, and the updated site representation of Layer 1 is introduced through skip connections to obtain the dimensionality-up projected site representation. The upgraded site representations and Layer 4 features are subjected to cross-attention, resulting in Layer 4 surface flow cross-attention enhancement results. and high-altitude flow across attention enhancement results The incremental representation of Layer 4 is calculated using the following formula: (22) in, For Layer 4 surface-high altitude fusion gating, This represents the incremental representation of Layer 4; Using residual gating parameters The Layer4 incremental representation is interpolated into the upgraded station representation, and the final station representation, which integrates multi-scale meteorological information, is calculated using the following formula: (23) in, To provide the final station representation that integrates multi-scale meteorological information, The site representation is projected in a higher dimension and updated from Layer 1 before entering Layer 4.

10. The method according to claim 9, characterized in that, Based on the final site characterization and the solar geometry information, normalized power prediction values ​​for each photovoltaic power station are generated for the future prediction period, including: Characterize the final site Expand along the prediction time dimension and overlay the learnable temporal location codes corresponding to the prediction steps. The initial predicted time-series features are calculated using the following formula: (24) in, H To predict the number of time steps, For the site n The final site representation, For the site n In the h Initial prediction time series characteristics at each prediction time point; Based on the station's latitude and longitude and the predicted time, the 6-dimensional solar geometric features are calculated using the following formula. : (25) in, For the site n In the h The solar altitude angle at the predicted moment. The azimuth of the sun. To illuminate the upper realm of clear skies, Use a daytime mask; Based on the initial predicted time-series features and the solar geometric features, the predicted time-series features after incorporating solar geometric information are calculated using the following formula: (26) in, For auxiliary feature projection matrix; Based on the predicted time-series characteristics after incorporating solar geometric information, sequence modeling is performed using bidirectional Mamba, where the continuous state-space parameters are discretized using the zero-order preservation method, and the discretization formula is as follows: (27) (28) in, To enable learning step size, It is the identity matrix. , For continuous state-space parameters, , These are the discretized state transition matrix and input matrix; Based on the discretized state-space parameters, the first... t Hidden state and output of the step: (29) (30) in, and The first t Step and the first t -1 step hidden state For the first t Step input, Depend on Dynamically generated For skip connectivity coefficients, For the first t Step output; Bidirectional scanning is used, with the positive branch along 1 to... H The states are updated sequentially in each direction. The reverse branch scans the flipped sequence and restores the original order. Based on the outputs of the forward and reverse scans, the final fused features are calculated using the following formula: (31) in, This is the output of the forward scan. This is the output of the reverse scan. To fuse the projection matrix, LN is the layer normalization. For the site In the The final fusion features at each prediction time; Based on the final fusion characteristics, the normalized power prediction value is calculated using the following formula: (32) in, For the site n In the h Normalized power prediction values ​​at each prediction time. For the site n In the h The final fusion features at each prediction time point, , This is the weight matrix of the feedforward network.

Images