Meteorological-new energy end-to-end power prediction method driven based on meteorological initial field
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-02
Smart Images
Figure CN122136831A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of renewable energy technology and artificial intelligence, and in particular to a meteorological-new energy end-to-end power prediction method based on meteorological initial field driven. Background Technology
[0002] Globally, the penetration rate of renewable energy sources, represented by photovoltaic and wind power, in the power system is growing at an unprecedented rate. However, the output of these energy sources is characterized by significant intermittency, volatility, and uncertainty, posing serious challenges to the stability, safe dispatch, and market-oriented operation of the power grid. Therefore, accurate power forecasting has become a core supporting technology for improving the absorption capacity of new energy sources and ensuring the safe and economical operation of the power grid.
[0003] Currently, traditional short-term power forecasting typically follows a multi-tiered, sequential architecture. The process begins with an initial meteorological field, first generating forecast results through numerical weather prediction (NWP) systems or deep learning-based meteorological models, and then inputting these results into a power forecasting model. Existing technology employs a two-stage forecasting method: "meteorological forecast + power forecast." This method first uses a meteorological model to generate a sequence of future weather forecasts, and then uses these forecasts as input to a new energy forecasting model to predict future power generation.
[0004] The main problem in the two-stage forecasting paradigm lies in the inconsistent optimization objectives among different modules. The core of meteorological model optimization is to achieve accuracy in general meteorological elements at the global or regional level, while the objective of renewable energy power forecasting models is the prediction accuracy of power generation at specific power plants. This results in the entire system being merely a simple concatenation of two independent tasks, without coordinated optimization towards the final objective. Due to this inconsistency in objectives, improvements in meteorological accuracy may not simultaneously lead to improvements in power accuracy. Furthermore, any biases or uncertainties in the meteorological forecasting model can be amplified in the power forecasting stage, especially when weather systems undergo drastic changes, where this cumulative error effect is particularly significant.
[0005] Therefore, there is an urgent need for a meteorological-new energy end-to-end power prediction method driven by the initial meteorological field to improve the above problems. Summary of the Invention
[0006] To address the aforementioned issues, this application proposes a meteorological-new energy end-to-end power prediction method driven by the initial meteorological field, comprising the following steps: S1. Obtain meteorological data from the weather forecast center and analyze the meteorological data to obtain multi-layered meteorological variable data; The input feature set is obtained by aligning multi-level meteorological variable data in time and space and unifying the spatiotemporal resolution; S2. Construct a meteorological feature encoder with a hybrid attention mechanism. Use the meteorological feature encoder to perform local and global interactive modeling on the input feature set to obtain a high-order meteorological implicit representation. The meteorological feature encoder based on the hybrid attention mechanism is a hierarchical architecture based on the improved Vision Transformer, including a meteorological-location embedding unit, a hybrid window unit based on the attention mechanism, and an upsampling spatiotemporal mapping unit. S3. Construct a power prediction decoder based on Swin Transformer. Use the power prediction decoder to fuse high-order meteorological implicit representations with historical power data of power plants to achieve end-to-end power prediction of meteorology and new energy.
[0007] Preferably, the input feature set in S1 is ; in, The length of historical meteorological input. As a meteorological variable dimension, , A geographic grid resolution of 0.25° × 0.25° is used; Meteorological variable dimensions This includes variables of the upper pressure layer and variables of the surface.
[0008] Preferably, the specific content of the meteorological-location embedding unit in S2 includes: With a preset target stride, the convolutional layer slices the geographic grid of the input feature set at the target stride, mapping the input feature set into a high-dimensional feature sequence; By introducing two-dimensional sine and cosine absolute position embedding, spatial location information is injected into the feature vector of a high-dimensional feature sequence. The expression for two-dimensional sine and cosine absolute position embedding is: ; ; in, It is the total dimension of the location encoding. It is the dimension index embedded within the vector. These are position coordinates. It is a component of the even-numbered index dimension. It is a component of the odd-numbered index dimension.
[0009] Preferably, the attention-based hybrid window unit is based on the Vision Transformer to form the backbone network, and is composed of L consecutive feature evolution units stacked together. Each unit consists of multi-scale heterogeneous windows and global attention. The specific content of the multi-scale heterogeneous window includes: The feature sequence is fed into the backbone network composed of Vision Transformer, which transforms the high-dimensional feature sequence into a feature map; The feature map is divided into non-overlapping local windows using anisotropic windows, longitudinal stretching windows, and latitudinal stretching windows, and different windows are dynamically and cyclically called according to the network layer index. Define the window size of the anisotropic window as The size of the radial stretch window is The size of the lateral stretch window is ; For the query matrix within the window Key matrix Sum matrix The expression for calculating window attention is: ; in, For feature dimension, For the introduced relative position offset, where This represents the matrix transpose. This mechanism, by limiting the receptive field, can accurately capture local meteorological topological features, and its computational complexity increases linearly with the sequence length.
[0010] Preferably, it consists of L consecutive stacked feature evolution units, each of which is composed of a multi-scale heterogeneous window and global attention, specifically including: Feature extraction is performed by alternating between three local windows: anisotropic window, meridional stretching window, and latitudinal stretching window. Then, break free from the limitations of the local window and switch to the global self-attention mechanism; The global self-attention mechanism performs a full-domain feature scan across the entire grid, capturing large-scale dynamic evolution patterns such as long-wave disturbances and long-range teleconnections in atmospheric circulation; The global self-attention mechanism computation integrates the Flash Attention operator using a block-based and recomputation strategy. The expression for the Flash Attention operator is: ; in, The output feature matrix calculated for the attention mechanism has the same dimension as the input features and integrates the sequence context information after weighting. For feature dimension, This is the scaling factor, and its value is... .
[0011] Through the cyclic stacking of L feature evolution units and deep nonlinear mapping, a deep feature representation integrating multi-scale spatiotemporal evolution information is finally output. .
[0012] Preferably, the upsampling spatiotemporal mapping unit specifically includes: The feature vector is obtained by performing layer normalization on the local window. This feature vector is then fed into a linear projection layer to achieve feature dimension compression and alignment. The expression is as follows: ; in, This provides a standardized hidden layer feature representation for subsequent upsampling operations. Let be the numerical space to which the feature tensor belongs. The height of the hidden layer feature map. The width of the hidden layer feature map. The number of channels for the hidden layer features. p For feature index identification, For layer normalization operation, This is a linear projection layer used to perform dimensional transformation or spatial mapping on the normalized features to adapt to the subsequent upsampling module.
[0013] Deploy multi-level upsampling blocks. Each level of upsampling block gradually restores the spatial resolution by combining residual neurons with transposed convolutions or interpolation layers to obtain the upsampled output. A higher-order meteorological latent representation is obtained by finely mapping the upsampled output using transposed convolution, expressed as: ; in, As a hidden representation of higher-order meteorology, For activation function, It is a two-dimensional transpose convolution operator. This is the upsampled output.
[0014] Preferably, the specific content of S3, which uses a power prediction decoder to fuse high-order meteorological implicit characterization with historical power data of power plants to achieve end-to-end power prediction of meteorology and new energy sources, includes: S301, the power prediction decoder uses the geographical latitude and longitude coordinates of the power station. By mapping physical coordinates to a meteorological grid index and sampling by slices, higher-order meteorological implicit representations are obtained. Perform local spatial alignment to extract 64 grid points centered on the target power station. Evolutionary perception of meteorological representation at each time step ; S302, the decoder utilizes three-dimensional convolution pairs The meteorological representations, perceived through evolution at each time step, are processed and mapped into high-dimensional block embeddings. ; Simultaneously, one-dimensional convolution is used to sample and project the historical power sequence to generate power feature embeddings. ; A joint feature map is constructed by asymmetrically splicing meteorological block embeddings and power feature embeddings at the channel dimension. To achieve feature projection and alignment between meteorological hidden layer features and one-dimensional historical power sequences; The S303 decoder uses a multi-level Swin Transformer backbone network to perform multi-scale nonlinear mapping on the joint feature map. S304. The decoder halves the resolution of the joint feature map and doubles the number of channels through downsampling blocks, thereby learning the higher-order abstract semantics between evolution-aware meteorological representations and power. The higher-order abstract semantics are then concatenated and stitched with the deep features restored by upsampling at the Nth layer in the channel dimension to obtain the meteorological-power fusion feature tensor. ; S305. At the decoder end, a convolution operator is used to perform cross-channel feature integration and dimensionality reduction on the meteorological-power fusion feature tensor, mapping the high-dimensional latent features to the physical output space, thereby extracting the core representation components that are highly correlated with the evolution trend of power generation. The expression is as follows: ; in, The activation function introduces nonlinear modeling capabilities to the features. It is a core representation component. It is a two-dimensional convolution; The compressed meteorological-power fusion feature tensor is flattened into a one-dimensional feature vector after being randomly deactivated by the Dropout layer and then fed into the linear mapping layer. The linear mapping layer reconstructs the latent space representation into a predicted power sequence of a preset length through fully connected mapping: ; in, This is the predicted future power vector; This is the weight matrix of the linear prediction head. For the bias term of the linear prediction head, The flattened feature vectors, This is a discard layer.
[0015] Preferably, the multi-level Swing Transformer backbone network uses an alternating window configuration; Within each layer, odd-numbered blocks perform standard window attention, while even-numbered blocks perform shifted window attention. No. The input features of the block are Its processing flow follows the expression below: when When the number is odd, the standard window block is: ; ; in, For the first Each block undergoes a window self-attention mechanism. Calculate and perform intermediate feature representation after residual connection. For the first The final output features of each block, It is a multilayer perceptron; when When the number is even, the shift window block: ; ; in, Self-attention for shifted windows; A two-dimensional relative position bias table is introduced when calculating the attention matrix. It compensates for the loss of absolute position and reduces spatial perception, and enhances sensitivity to wind direction in complex terrain.
[0016] in, Let be the side length of the local window. The number of heads in a multi-head attention mechanism.
[0017] The preferred training process for the meteorological-new energy end-to-end power prediction method includes an initial stage, a joint training stage, and a time-history decay training strategy. The specific steps are as follows: In the initial stage, the meteorological feature encoder and the power prediction decoder are trained independently to extract high-quality modal features: In meteorology, this is achieved by minimizing the weighted root mean square error of the hidden layer feature space. To pre-train the encoder; At the power end, the hidden layer features generated by the fixed encoder are used, utilizing the root mean square error. To pre-train the decoder; During the joint training phase, a low-rank adaptive LoRA technique is introduced. The time-course decay training strategy uses the time-course decay function. The weights of the weather forecasting and power forecasting tasks are dynamically adjusted based on the training progress. The time-history decay function expression is as follows: ; in, For the current training iteration round, This is the preset total number of training rounds.
[0018] Preferably, the specific content of introducing low-rank adaptive LoRA technology during the joint training phase includes: (1) Low-rank decomposition logic modeling: Define the pre-trained weight matrix to be fine-tuned as follows: Where D is the input dimension of the pre-trained weight matrix, and K is the output dimension of the pre-trained weight matrix; During fine-tuning, the weight matrix remains frozen and does not participate in gradient updates; instead, a pair of trainable low-rank matrices is introduced. and Increment of constituent parameters : ; in, , and rank ; (2) Reconstruction of the physical mapping process: After introducing LoRA, the original feature mapping process is changed from... The transformation is a forward computation process that includes adaptive corrections, expressed as: ; in, The preset scaling factor, For a constant scaling factor, Let be the rank of the low-rank matrix; (3) Initialization and stability control strategy: matrix Initialize using random Gaussian matrix Initialize to a zero matrix.
[0019] In summary, the meteorological-new energy end-to-end power prediction method based on meteorological initial field driven by this invention has the following advantages compared with traditional technologies: 1. By constructing an end-to-end fully differential mapping architecture that goes directly from the initial meteorological field to the power output, this invention completely breaks the physical barrier of "meteorological forecast error flowing into the power prediction model" in the traditional serial architecture. The gradient can be directly transmitted back from the final power prediction error across modules to the bottom original meteorological feature extraction layer, ensuring that the parameters of the entire model are jointly evolved with the optimal power output as the core objective, and avoiding the deviation generated in the intermediate prediction stage from being nonlinearly amplified in the subsequent stages. 2. The proposed solution employs a time-history decay-based multi-objective balance loss function, guiding the model to focus on learning the physical consistency of atmospheric evolution in the early stages of training, and smoothly switching to extreme optimization for power accuracy in the later stages. This mechanism effectively solves the problem of inconsistent optimization directions between meteorological and power models in traditional architectures due to the independent loss functions, guiding the model to achieve global optimal convergence while maintaining the laws of meteorological dynamics. 3. By utilizing the latitude and longitude asymmetric window attention mechanism, the model can establish accurate physical correspondences for atmospheric dynamic processes with obvious directional characteristics. Through the feature evolution unit composed of local windows and global attention, the model can capture subtle weather features such as local strong convection, while also perceiving the large-scale circulation background of the entire region, which significantly improves the model's prediction robustness under extreme weather or complex boundary conditions. 4. The low-rank fine-tuning technique based on LoRA enables the model to complete targeted feature alignment by only fine-tuning a very small number of parameter increments when facing sites with different geographical environments (such as complex mountainous areas and offshore wind fields). Under the premise of freezing the parameters of the pre-trained meta-model, this strategy realizes the rapid transfer from general atmospheric physics laws to the dynamic characteristics of specific sites, which greatly reduces the consumption of computing resources and data sample requirements when deploying new energy prediction systems on a large scale.
[0020] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0021] Figure 1 This is a framework diagram of a meteorological-new energy end-to-end power prediction method driven by meteorological initial field. Figure 2 A comparison chart of photovoltaic power prediction accuracy under different weather scenarios; Figure 3 A comparison chart showing the accuracy of wind power prediction under different weather scenarios. Detailed Implementation
[0022] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of this application.
[0023] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.
[0024] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.
[0025] In all the examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.
[0026] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0027] Example 1 This embodiment provides a meteorological-renewable energy end-to-end power prediction method driven by the initial meteorological field. Unlike the traditional multi-stage cascaded architecture of "initial meteorological field - weather forecast - power prediction," this scheme constructs an end-to-end fully differential mapping architecture directly from the initial meteorological field to power output. This architecture achieves integrated modeling of weather forecast and power forecast, and by employing a single loss function, it realizes joint optimization of all internal modules, completely eliminating the performance loss and error accumulation caused by inconsistent optimization objectives of multiple modules in the traditional architecture. Figure 1 As shown, it includes the following steps: S1. Obtain meteorological data from the weather forecast center and analyze the meteorological data to obtain multi-layered meteorological variable data.
[0028] The input feature set is obtained by aligning multi-level meteorological variable data in time and space and unifying the spatiotemporal resolution.
[0029] Understandably, this approach models the short-term power prediction task as a fully differential mapping problem from a high-dimensional meteorological dynamic space to a discrete scalar power space. It obtains the global initial meteorological field, aligns all data sources in time and space, and processes them into gridded data with a uniform spatiotemporal resolution as the input feature set of the model.
[0030] Furthermore, the input feature set in S1 is .
[0031] in, The length of historical meteorological input. As a meteorological variable dimension, , A geographic grid resolution of 0.25° × 0.25° is used; Meteorological variable dimensions This includes variables of the upper pressure layer and variables of the surface.
[0032] It is understandable that the variables of the upper pressure layer, including geopotential, specific humidity, wind speed components, temperature and vertical velocity, are distributed across 13 standard pressure layers from 1000 hPa to 50 hPa.
[0033] Surface variables include 10m / 100m wind speed, 2m temperature, total cloud cover, surface air pressure, sea level air pressure, 6-hour cumulative precipitation, and shortwave radiation.
[0034] This scheme designs a composite feature extraction strategy based on alternating triggering of hybrid windows and global attention.
[0035] At the meteorological coding end, anisotropic windows and latitude-longitude asymmetric windows are used to model directional complex atmospheric dynamic processes, and global self-attention is combined to capture long-range correlations, with Flash Attention to achieve operator-level acceleration.
[0036] At the power decoding end, spatially centered sampling is innovatively adopted to achieve the fusion of asymmetric features of the local meteorological field of the power station and the historical power sequence. The shift window mechanism of the Swing Transformer is used to capture high-order abstract semantics, which significantly enhances the model's physical perception ability and prediction robustness under complex geographical topology.
[0037] This scheme adopts a hierarchical architecture based on Vision Transformer as a meteorological feature encoder. Through local and global interactive modeling, it achieves efficient feature extraction and evolution simulation of multidimensional meteorological elements.
[0038] S2. Construct a meteorological feature encoder with a hybrid attention mechanism. The meteorological feature encoder is used to perform local and global interactive modeling of the input feature set to obtain a high-order meteorological implicit representation.
[0039] The meteorological feature encoder based on the hybrid attention mechanism is a hierarchical architecture improved from Vision Transformer, including a meteorological-location embedding unit, a hybrid window unit based on the attention mechanism, and an upsampling spatiotemporal mapping unit.
[0040] First, input the high-dimensional meteorological tensor. Feed to the block embedding layer.
[0041] Furthermore, the specific content of the meteorological-location embedding unit in S2 includes: Patch embedding: With a preset target stride, the convolutional layer slices the geographic grid of the input feature set at the target stride, mapping the input feature set into a high-dimensional feature sequence.
[0042] Physical location awareness: To maintain the spatial topological consistency of meteorological data, a two-dimensional sine and cosine absolute location embedding is introduced. This embedding injects spatial location information into the feature vector of a high-dimensional feature sequence by calculating the frequency response of different geographic grids. The expression for the two-dimensional sine and cosine absolute location embedding is as follows: .
[0043] .
[0044] in, It is the total dimension of the location encoding. It is the dimension index embedded within the vector. These are position coordinates. It is a component of the even-numbered index dimension. It is a component of the odd-numbered index dimension.
[0045] The feature sequence enters the backbone network composed of Vision Transformer. In order to balance computational overhead and remote evolution modeling when processing high-resolution meteorological fields, this application uses L consecutive feature evolution units stacked together. Each unit consists of multi-scale heterogeneous windows and global attention, dividing the feature map into several non-overlapping local windows.
[0046] Furthermore, the attention-based hybrid window unit is constructed based on the Vision Transformer as the backbone network, and uses L units composed of multi-scale heterogeneous windows and global attention.
[0047] The specific content of the multi-scale heterogeneous window includes: The feature sequences are fed into a backbone network composed of Vision Transformers, which transforms the high-dimensional feature sequences into feature maps.
[0048] The feature map is divided into non-overlapping local windows using anisotropic windows, longitudinal stretching windows, and latitudinal stretching windows, and different windows are dynamically and cyclically called according to the network layer index.
[0049] Define the window size of the anisotropic window as The size of the radial stretch window is The size of the lateral stretch window is .
[0050] This can be understood as anisotropic window: a square window is used to capture local symmetrical meteorological topological features, such as local convection systems.
[0051] Asymmetric windows in the warp and weft directions: Warp stretching windows and weft stretching windows with different aspect ratios are used.
[0052] By introducing an asymmetric receptive field, the model can specifically model atmospheric physical processes with obvious directional characteristics, enhancing its ability to capture meteorological systems with specific geographic orientations.
[0053] Define window size as For the query matrix within the window Key matrix Sum matrix The expression for calculating window attention is: .
[0054] in, For feature dimension, For the introduced relative position offset, where This represents the matrix transpose. This mechanism, by limiting the receptive field, can accurately capture local meteorological topological features, and its computational complexity increases linearly with sequence length.
[0055] The backbone network adopts a progressive logic of "local scanning + global alignment". It first uses three local windows alternately for feature extraction, and then switches to a global self-attention mechanism to achieve information fusion and alignment from local to global.
[0056] Global self-attention: Performs a full-domain feature scan across the entire grid to capture large-scale dynamic evolution patterns such as long-wave disturbances and long-range teleconnections in atmospheric circulation.
[0057] Alternating Gains: This “local-local-local-global” cyclical architecture enables the model to achieve complete modeling from local small-scale weather phenomena to global large-scale climate background while maintaining extremely high computational efficiency.
[0058] Furthermore, the L units, each composed of multi-scale heterogeneous windows and global attention, constitute the progressive logic of local scanning and global alignment, including: Feature extraction is performed by alternating between three local windows: anisotropic window, meridional stretch window, and latitudinal stretch window.
[0059] Then, we break free from the local window limitation and switch to a global self-attention mechanism.
[0060] The global self-attention mechanism performs a full-domain feature scan across the entire grid, capturing large-scale dynamic evolution patterns such as long-wave disturbances and long-range teleconnections in atmospheric circulation.
[0061] The global self-attention mechanism computation integrates the Flash Attention operator using a block-based and recomputation strategy. This operator optimizes the data exchange speed between the cache and main memory through the block-based and recomputation strategy.
[0062] The expression for the Flash Attention operator is: .
[0063] in, The output feature matrix calculated for the attention mechanism has the same dimension as the input features and integrates the sequence context information after weighting. For feature dimension, This is the scaling factor, and its value is... .
[0064] Through the cyclic stacking of L feature evolution units and deep nonlinear mapping, a deep feature representation integrating multi-scale spatiotemporal evolution information is finally output. .
[0065] Through kernel fusion, the memory usage is reduced to a minimum when processing ultra-long meteorological feature sequences. This significantly improves computing speed.
[0066] In order to restore the deep hidden features extracted by the Transformer to the original meteorological dynamic space or a specific hidden space, a symmetric decoding sequence with a multi-level residual structure is integrated at the end of the encoder.
[0067] Furthermore, the upsampling spatiotemporal mapping unit specifically includes: Latent space dimensionality reduction projection: Local window dimensionality is normalized to obtain feature vectors, which are then fed into a linear projection layer to achieve feature dimension compression and alignment. The expression is as follows: .
[0068] in, This provides a standardized hidden layer feature representation for subsequent upsampling operations. Let be the numerical space to which the feature tensor belongs. The height of the hidden layer feature map. The width of the hidden layer feature map. The number of channels for the hidden layer features. p For feature index identification, For layer normalization operation, This is a linear projection layer used to perform dimensional transformation or spatial mapping on the normalized features to adapt to the subsequent upsampling module.
[0069] Hierarchical residual upsampling: Deploy multi-level upsampling blocks. Each level of upsampling block gradually restores the spatial resolution by combining residual neurons with transposed convolutions or interpolation layers to obtain the upsampled output.
[0070] Definition of the first The output of the stage is Then its evolutionary logic is, .
[0071] in, This represents a residual block operation, used to mitigate gradient vanishing during feature extraction by using skip connections and to enhance the ability to capture deep features. This represents the upsampling operator, used to gradually restore the spatial resolution of the feature map, achieving feature mapping from low resolution to high resolution.
[0072] This structure ensures the physical continuity of meteorological features during spatial scale reconstruction and prevents the loss of high-frequency information.
[0073] A higher-order meteorological latent representation is obtained by finely mapping the upsampled output using transposed convolution, expressed as: .
[0074] in, As a hidden representation of higher-order meteorology, For activation function, It is a two-dimensional transpose convolution operator. This is the upsampled output.
[0075] Ultimately, a high-order meteorological implicit representation with physical consistency is generated. This provides high-fidelity meteorological feature support for subsequent power prediction decoders.
[0076] The power prediction decoder designed in this scheme adopts a symmetrical U-shaped Swing Transformer architecture, which deeply integrates the high-order meteorological hidden layer features extracted by the encoder with the historical power data of the power plant.
[0077] S3. Construct a power prediction decoder based on Swin Transformer. Use the power prediction decoder to fuse high-order meteorological implicit representations with historical power data of power plants to achieve end-to-end power prediction of meteorology and new energy.
[0078] Furthermore, the specific content of S3's use of a power prediction decoder to fuse high-order meteorological implicit representations with historical power data of power plants to achieve end-to-end power prediction of meteorological and renewable energy sources includes: S301 (Spatial-Centered Sampling) and the power prediction decoder use the geographic latitude and longitude coordinates of the power plant. By mapping physical coordinates to a meteorological grid index and sampling by slices, higher-order meteorological implicit representations are obtained. Local spatial alignment was performed to map the physical coordinates to the meteorological grid index. Through slice sampling, 64 grid points centered on the target power station were extracted. Evolutionary perception of meteorological representation at each time step .
[0079] This step significantly reduces the modeling complexity of the subsequent power prediction module by focusing on the local meteorological evolution environment around the power plant and filtering out irrelevant meteorological noise from afar.
[0080] S302 (Heterogeneous Modal Feature Joint Alignment and Embedding): The decoder achieves feature projection and alignment between the local meteorological field and the one-dimensional historical power sequence through the embedding layer. Meteorological spatiotemporal block slices, the decoder utilizes 3D convolution pairs The meteorological representations, perceived through evolution at each time step, are processed and mapped into high-dimensional block embeddings. .
[0081] The power features are updimensionalized, and a one-dimensional convolution is used to sample and project the historical power sequence to generate a power feature embedding. .
[0082] The joint feature map is constructed by asymmetrically stitching together meteorological block embeddings and power feature embeddings in the channel dimension. This enables the feature projection and alignment of local meteorological fields with one-dimensional historical power sequences.
[0083] The S303 decoder uses a multi-level Swing Transformer backbone network to perform multi-scale nonlinear mapping on the joint feature map.
[0084] S304. The decoder halves the resolution of the joint feature map and doubles the number of channels through downsampling blocks, thereby forcing the model to learn the higher-order abstract semantics between evolutionarily perceived meteorological representations and power. The higher-order abstract semantics are then concatenated with the deep features restored by upsampling at the Nth layer along the channel dimension to obtain the meteorological-power fusion feature tensor. This ensures that the model can capture both deep, large-scale meteorological evolution trends and retain shallow, detailed information that is sensitive to drastic power fluctuations.
[0085] S305, Power Recovery Head: The decoder integrates a projection module with residual restoration capabilities. The decoder uses convolution operators to perform cross-channel feature integration and dimensionality reduction on the meteorological-power fusion feature tensor, mapping high-dimensional latent features to the physical output space. This extracts core representation components highly correlated with the evolution trend of power generation, expressed as: .
[0086] in, The activation function introduces nonlinear modeling capabilities to the features. It is a core representation component. It uses two-dimensional convolution to ensure that the model can fit the complex dynamic mapping between wind speed / radiation and power.
[0087] The compressed meteorological-power fusion feature tensor is flattened into a one-dimensional feature vector after being randomly deactivated by the Dropout layer and then fed into the linear mapping layer.
[0088] The linear mapping layer reconstructs the latent space representation into a predicted power sequence of a preset length through fully connected mapping: .
[0089] in, This is the predicted future power vector; This is the weight matrix of the linear prediction head. For the bias term of the linear prediction head, The flattened feature vectors, As a discard layer, this module realizes the transformation from "meteorological implicit feature space" to "physical scalar power space", and directly outputs the new energy output prediction results with actual physical meaning.
[0090] Furthermore, the multi-level Swin Transformer backbone network employs an alternating window configuration.
[0091] Shift window logic: Within each layer, odd-numbered blocks perform standard window attention, while even-numbered blocks perform shifted window attention.
[0092] No. The input features of the block are Its processing flow follows the expression below: when When the number is odd, the standard window block is: .
[0093] .
[0094] in, For the first Each block undergoes a window self-attention mechanism. Calculate and perform intermediate feature representation after residual connection. For the first The final output features of each block, It is a multilayer perceptron.
[0095] when When the number is even, the shift window block: .
[0096] .
[0097] in, For standard window self-attention, For shifted window self-attention, this alternating design allows the model to maintain local computational efficiency within the window while introducing logical connections between windows through shifting operations, thereby enabling the modeling of broader meteorological flow field evolution patterns.
[0098] A two-dimensional relative position bias table is introduced when calculating the attention matrix. This compensates for the loss of absolute position leading to a decrease in spatial perception and enhances sensitivity to wind direction in complex terrain. Let be the side length of the local window. The number of heads in a multi-head attention mechanism.
[0099] This application adopts a "staged pre-training + efficient parameter fine-tuning" model. First, the decoupled pre-training of the meteorological and power modules is completed sequentially. Then, LoRA (Low-Rank Adaptation) is introduced for end-to-end fine-tuning. An innovative time-history decay-type multi-objective balance loss function is designed to dynamically adjust the weight ratio of meteorological and physical constraints and power prediction objectives as the training process progresses. This guides the model to smoothly transition from "focusing on the physical consistency of atmospheric evolution" to "focusing on the absolute accuracy of power prediction," achieving globally optimal convergence.
[0100] To balance the physical consistency of meteorological dynamics with the absolute accuracy of power prediction, this scheme adopts a joint training strategy of "staged pre-training + efficient parameter fine-tuning". This strategy ensures that the model, based on its understanding of global meteorological evolution, can accurately align with the power output of a specific power plant.
[0101] Furthermore, the training process of the meteorological-new energy end-to-end power prediction method includes an initial stage, a joint training stage, and a time-history decay training strategy. The specific steps are as follows: In the initial stage, the meteorological feature encoder and the power prediction decoder are trained independently to extract high-quality modal features: In meteorology, this is achieved by minimizing the weighted root mean square error of the hidden layer feature space. To pre-train the encoder, loss function Defined as follows, this function directly acts on higher-order meteorological hidden layer characteristics. : .
[0102] in, and The spatial resolution of the hidden layer feature map; These are weighting coefficients that vary with latitude, used to compensate for physical deviations caused by uneven distribution of the Earth's spherical grid. This represents a true meteorological hidden layer. The predicted hidden layer representation generated by the encoder.
[0103] At the power level, the hidden layer features generated by the fixed encoder are used to independently train the prediction decoder using historical power observation data. The goal is to minimize the root mean square error between the predicted power and the measured power. .
[0104] .
[0105] in, This represents the actual measured power generation capacity of the power station. This represents the predicted power output by the decoder.
[0106] During the joint training phase, a low-rank adaptive LoRA technique is introduced.
[0107] The time-course decay training strategy uses the time-course decay function. The weights of the two tasks are dynamically adjusted according to the training progress to ensure a smooth transition of the model from "learning meteorological evolution" to "power accuracy optimization". The expression for the time-history decay function is as follows: .
[0108] in, This is the current training iteration round; This is the preset total number of training rounds.
[0109] Furthermore, during the joint training phase, Low-Rank Adaptation (LoRA) is introduced. This technique allows the model to achieve deep feature alignment between modalities by adding only a very small number of parameters while freezing the pre-trained meta-model parameters. Specifically, this includes: (1) Low-rank decomposition logic modeling: Define the pre-trained weight matrix to be fine-tuned in the meta-model as follows: Where D is the input dimension of the pre-trained weight matrix and K is the output dimension of the pre-trained weight matrix.
[0110] During fine-tuning, the weight matrix remains frozen and does not participate in gradient updates; instead, a pair of trainable low-rank matrices is introduced. and Increment of constituent parameters : .
[0111] in, , and rank .
[0112] (2) Reconstruction of the physical mapping process: After introducing LoRA, the original feature mapping process is changed from... The transformation is a forward computation process that includes adaptive corrections, expressed as: .
[0113] in, The preset scaling factor, For a constant scaling factor, The rank of the low-rank matrix is given by the transformation, which compensates for the prediction bias of the general meteorological model under specific new energy power plant sites by superimposing a low-rank correction term on the original meteorological dynamics mapping.
[0114] (3) Initialization and stability control strategy, to ensure that the fine-tuning process starts smoothly from the pre-trained optimal physical state: matrix Initialize using random Gaussian matrix Initializing to a zero matrix ensures that, at the moment fine-tuning begins... This allows the model to stably retain the physical laws of atmospheric evolution captured by the pre-trained model and gradually converge toward the goal of high-precision power prediction.
[0115] Example 2 The evaluation metric used in this embodiment is the accuracy rate of short-term power forecast for wind farms published by the State Grid Corporation of China. The formula is as follows: ; in for Measured power at time, for Predicted power at time, for The capacity at any given time during startup.
[0116] 1. To verify the predictive performance of the model, this embodiment was tested on photovoltaic and wind power clusters from different manufacturers, and the accuracy was used as the evaluation index. As shown in Tables 1 and 2, a high accuracy of over 90% was achieved.
[0117] Table 1 Comparison of Prediction Accuracy of Wind Power Clusters
[0118] Table 2 Comparison of Prediction Accuracy of Photovoltaic Clusters
[0119] 2. To verify the model's scalability, a comparison was made with the site reporting and main station prediction methods across more than 400 sites in three provinces and eight typical scenarios, such as... Figure 2 and Figure 3 The model outperforms the other two methods in various scenarios, especially in the case of changing weather conditions, and can accurately capture the situation where wind power and photovoltaic output rise or fall sharply, outperforming the second-best method by about 4%.
[0120] 3. To verify the model's robustness across the four seasons, the end-to-end meteorological-new energy method and the current station reporting method were compared, and both achieved the best results, as shown in Table 2. Relying on the model's adaptive perception capability of seasonal climate mode differences, it can deeply understand the meteorological-power coupling law under different seasons, thereby effectively overcoming the performance fluctuations caused by seasonal distribution shifts. Therefore, it can achieve continuous high-precision prediction throughout the year and across seasons.
[0121] Table 3. Annual Cross-Season Forecasting Performance Assessment Results
[0122] RMSE, as the squared mean of the difference between predicted and actual values, can effectively reflect the impact of large prediction errors on overall performance; a lower value indicates higher prediction accuracy. MAE, on the other hand, provides the average level of prediction bias; a lower value indicates stronger model stability. 2 It is used to measure the ability of the prediction model to fit the actual trend of power output change. It is an important standard for measuring explanatory power. The higher the value, the better the prediction fit.
[0123] RMSE, MAE and R 2 The expressions are as follows: ; ; ; in, This represents the average of the actual power.
[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.
Claims
1. A meteorological-new energy end-to-end power prediction method driven by meteorological initial field, characterized in that, Includes the following steps: S1. Obtain meteorological data from the weather forecast center and analyze the meteorological data to obtain multi-layered meteorological variable data; The input feature set is obtained by aligning multi-level meteorological variable data in time and space and unifying the spatiotemporal resolution; S2. Construct a meteorological feature encoder with a hybrid attention mechanism. Use the meteorological feature encoder to perform local and global interactive modeling on the input feature set to obtain a high-order meteorological implicit representation. The meteorological feature encoder based on the hybrid attention mechanism is a hierarchical architecture based on the improved Vision Transformer, including a meteorological-location embedding unit, a hybrid window unit based on the attention mechanism, and an upsampling spatiotemporal mapping unit. S3. Construct a power prediction decoder based on Swin Transformer. Use the power prediction decoder to fuse high-order meteorological implicit representations with historical power data of power plants to achieve end-to-end power prediction of meteorology and new energy.
2. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 1, characterized in that, The input feature set in S1 is ; in, The length of historical meteorological input. As a meteorological variable dimension, , A geographic grid resolution of 0.25° × 0.25° is used; Meteorological variable dimensions This includes variables of the upper pressure layer and variables of the surface.
3. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 2, characterized in that, The specific content of the meteorological-location embedded unit in S2 includes: With a preset target stride, the convolutional layer slices the geographic grid of the input feature set at the target stride, mapping the input feature set into a high-dimensional feature sequence; By introducing two-dimensional sine and cosine absolute position embedding, spatial location information is injected into the feature vector of a high-dimensional feature sequence. The expression for two-dimensional sine and cosine absolute position embedding is: ； ； in, It is the total dimension of the location encoding. It is the dimension index embedded within the vector. These are position coordinates. It is a component of the even-numbered index dimension. It is a component of the odd-numbered index dimension.
4. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 3, characterized in that, The attention-based hybrid window unit is based on the Vision Transformer as the backbone network and consists of L consecutive feature evolution units stacked together. Each unit is composed of multi-scale heterogeneous windows and global attention. The specific content of the multi-scale heterogeneous window includes: The feature sequence is fed into the backbone network composed of Vision Transformer, which transforms the high-dimensional feature sequence into a feature map; The feature map is divided into non-overlapping local windows using anisotropic windows, longitudinal stretching windows, and latitudinal stretching windows, and different windows are dynamically and cyclically called according to the network layer index. Define the window size of the anisotropic window as The size of the radial stretch window is The size of the latitudinal stretch window is ; For the query matrix within the window Key matrix Sum matrix The expression for calculating window attention is: ； in, For feature dimension, For the introduced relative position offset, where This indicates the matrix transpose.
5. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 4, characterized in that, It consists of L consecutive stacked feature evolution units, each of which is composed of a multi-scale heterogeneous window and a global attention, specifically including: Feature extraction is performed by alternating between three local windows: anisotropic window, meridional stretching window, and latitudinal stretching window. Then, break free from the limitations of the local window and switch to the global self-attention mechanism; The global self-attention mechanism performs a full-domain feature scan across the entire grid, capturing the large-scale dynamic evolution of long-wave disturbances and long-range teleconnections in atmospheric circulation; The global self-attention mechanism computation integrates the Flash Attention operator using a block-based and recomputation strategy. The expression for the Flash Attention operator is: ； in, The output feature matrix calculated by the attention mechanism. For feature dimension, This is the scaling factor, and its value is... ; Through the cyclic stacking of L feature evolution units and deep nonlinear mapping, a deep feature representation integrating multi-scale spatiotemporal evolution information is finally output. .
6. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 5, characterized in that, The upsampling spatiotemporal mapping unit specifically includes: The feature vector is obtained by performing layer normalization on the local window. This feature vector is then fed into a linear projection layer to achieve feature dimension compression and alignment. The expression is as follows: ； in, This provides a standardized hidden layer feature representation for subsequent upsampling operations. Let be the numerical space to which the feature tensor belongs. The height of the hidden layer feature map. The width of the hidden layer feature map. The number of channels for the hidden layer features. p For feature index identification, For layer normalization operation, For linear projection layers; Deploy multi-level upsampling blocks. Each level of upsampling block gradually restores the spatial resolution by combining residual neurons with transposed convolutions or interpolation layers to obtain the upsampled output. Renewable energy technology and artificial intelligence employ transposed convolution to perform refined mapping on the upsampled output to obtain a high-order meteorological latent representation, expressed as: ； in, As a hidden representation of higher-order meteorology, For activation function, It is a two-dimensional transpose convolution operator. This is the upsampled output.
7. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 6, characterized in that, The specific content of S3's use of a power prediction decoder to fuse high-order meteorological implicit characterization with historical power plant power data to achieve end-to-end power prediction from meteorology to new energy sources includes: S301, the power prediction decoder uses the geographical latitude and longitude coordinates of the power station. By mapping physical coordinates to a meteorological grid index and sampling by slices, higher-order meteorological implicit representations are obtained. Perform local spatial alignment to extract 64 grid points centered on the target power station. Evolutionary perception of meteorological representation at each time step ; S302, the decoder utilizes three-dimensional convolution pairs The meteorological representations, perceived through evolution at each time step, are processed and mapped into high-dimensional block embeddings. ; Simultaneously, one-dimensional convolution is used to sample and project the historical power sequence to generate power feature embeddings. ; A joint feature map is constructed by asymmetrically splicing meteorological block embeddings and power feature embeddings at the channel dimension. ; The S303 decoder uses a multi-level Swin Transformer backbone network to perform multi-scale nonlinear mapping on the joint feature map. S304. The decoder halves the resolution of the joint feature map and doubles the number of channels through downsampling blocks, thereby learning the higher-order abstract semantics between evolution-aware meteorological representations and power. The higher-order abstract semantics are then concatenated and stitched with the deep features restored by upsampling at the Nth layer in the channel dimension to obtain the meteorological-power fusion feature tensor. ; S305. At the decoder end, a convolution operator is used to perform cross-channel feature integration and dimensionality reduction on the meteorological-power fusion feature tensor, mapping the high-dimensional latent features to the physical output space, thereby extracting the core representation components that are highly correlated with the evolution trend of power generation. The expression is as follows: ； in, The activation function introduces nonlinear modeling capabilities to the features. It is a core representation component. It is a two-dimensional convolution; The compressed meteorological-power fusion feature tensor is flattened into a one-dimensional feature vector after being randomly deactivated by the Dropout layer and then fed into the linear mapping layer. The linear mapping layer reconstructs the latent space representation into a predicted power sequence of a preset length through fully connected mapping: ； in, This is the predicted future power vector; This is the weight matrix of the linear prediction head. For the bias term of the linear prediction head, The flattened feature vectors This is a discard layer.
8. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 7, characterized in that, The multi-level Swing Transformer backbone network uses an alternating window configuration. Within each layer, odd-numbered blocks perform standard window attention, while even-numbered blocks perform shifted window attention. No. The input features of the block are Its processing flow follows the expression below: when When the number is odd, the standard window block is: ； ； in, For the first Each block undergoes a window self-attention mechanism. Calculate and perform intermediate feature representation after residual connection. For the first The final output features of each block, It is a multilayer perceptron; when When the number is even, the shift window block: ； ； in, Self-attention for shifted windows; A two-dimensional relative position bias table is introduced when calculating the attention matrix. It compensates for the loss of absolute position and reduces spatial perception, and enhances sensitivity to wind direction in complex terrain. in, Let be the side length of the local window. The number of heads in a multi-head attention mechanism.
9. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 8, characterized in that, The training process of the meteorological-new energy end-to-end power prediction method includes an initial stage, a joint training stage, and a time-history decay training strategy. The specific steps are as follows: In the initial stage, the meteorological feature encoder and the power prediction decoder are trained independently to extract high-quality modal features: In meteorology, this is achieved by minimizing the weighted root mean square error of the hidden layer feature space. To pre-train the encoder; At the power end, the hidden layer features generated by the encoder are fixed, and the root mean square error is used to pre-train the decoder. During the joint training phase, a low-rank adaptive LoRA technique is introduced. The time-course decay training strategy uses the time-course decay function. The weights of the weather forecasting and power forecasting tasks are dynamically adjusted according to the training progress. The time-history decay function expression is as follows: ； in, For the current training iteration round, This represents the preset total number of training rounds.
10. The meteorological-new energy end-to-end power prediction method based on meteorological initial field driven according to claim 9, characterized in that, The specific content of introducing low-rank adaptive LoRA technology during the joint training phase includes: (1) Low-rank decomposition logic modeling: Define the pre-trained weight matrix to be fine-tuned as follows: Where D is the input dimension of the pre-trained weight matrix, and K is the output dimension of the pre-trained weight matrix; During fine-tuning, the weight matrix remains frozen and does not participate in gradient updates; instead, a pair of trainable low-rank matrices is introduced. and Increment of constituent parameters : ； in, , and rank ; (2) Reconstruction of the physical mapping process: After introducing LoRA, the original feature mapping process is changed from... The transformation is a forward computation process that includes adaptive corrections, expressed as: ； in, The preset scaling factor, For a constant scaling factor, Let be the rank of the low-rank matrix; (3) Initialization and stability control strategy: matrix Initialize using random Gaussian matrix Initialize to a zero matrix.