New energy meteorological prediction method and system based on multi-source fusion conditional diffusion and physical constraint

Abstract

The application discloses a new energy weather forecast method and system based on multi-source fusion condition diffusion and physical constraints, and belongs to the technical field of new energy weather forecast. The method comprises the following steps: collecting and spatially aligning heterogeneous data in real time; extracting multi-modal features in parallel; adaptively fusing channels and associating time-space decoupling; extrapolating state deduction by autoregressive dynamics; generating low-dimensional hidden space accelerated denoising; and intercepting and mapping forced physical mechanism priori defense lines. The application can efficiently fuse and spatially align the time-space heterogeneous observation live, satellite cloud map and two-dimensional plane radar echo data, overcome the error accumulation of long sequence extrapolation of traditional models under the premise of ensuring the second-level reasoning timeliness, and output high-time-space resolution short-term weather forecast results fully complying with physical priori constraints.

Description

Technical Field

[0001] This invention relates to the field of new energy meteorological forecasting technology. Specifically, it relates to a new energy meteorological forecasting method and system based on multi-source fusion conditional diffusion and physical constraints, which is applicable to the prediction of short-term wind speed and solar irradiance in wind power and photovoltaic power generation scenarios. Background Technology

[0002] With the large-scale grid connection of renewable energy sources such as wind and solar power, the power grid system faces increasingly stringent short-term power generation deviation assessments. The accuracy of short-term weather forecasts for the next 0 to 6 hours, especially the forecasts of wind speed and surface solar irradiance, is directly related to the safety of power grid dispatch and the power generation efficiency of renewable energy power plants.

[0003] Existing numerical weather prediction (NWP) methods suffer from bottlenecks such as time lag and insufficient resolution in the short-term phase. Traditional methods based on statistical extrapolation, while computationally fast, cannot effectively capture abrupt changes in meteorological elements caused by extreme weather events. In recent years, deep learning models have been applied to meteorological forecasting, but they still have the following technical limitations: 1) The inconsistent spatiotemporal resolution of multi-source meteorological data makes fusion difficult. The temporal frequency and spatial scale of the three types of data—observational data, satellite cloud images, and radar echoes—differ significantly, and existing methods lack a fusion architecture that can simultaneously process these three heterogeneous data and achieve dynamic adaptive alignment; 2) The computational complexity of traditional spatiotemporal self-attention mechanisms is too high. For a meteorological field with time step T and spatial grid number H×W, the computational complexity of spatiotemporal self-attention is O((T·H·W)²). When the spatial resolution reaches 1km, the existing GPU memory cannot support this amount of computation. 3) Existing diffusion models are mostly used for single time series or image data, and lack the ability to process multi-source spatiotemporal heterogeneous features; when performing long sequence extrapolation, the accumulation of errors in the method based on cyclic iterative networks leads to forecast ambiguity. 4) Regression models based on mean squared error loss tend to output smoothed average values, and are not good at predicting low-probability events such as extreme wind speeds and sudden changes in irradiance. 5) The prediction results output by deep learning generative models may violate basic physical laws, such as producing outliers like negative wind speeds or positive nighttime irradiance. Existing methods lack effective physical hard constraint mechanisms at the output end.

[0004] Therefore, there is an urgent need for a high-resolution short-term weather forecasting method that can efficiently integrate multi-source heterogeneous meteorological data and output high-resolution short-term weather forecasts that conform to physical laws while ensuring the timeliness of inference. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a new energy meteorological forecasting method and system based on multi-source fusion conditional diffusion and physical constraints. This method can efficiently fuse spatially aligned and spatiotemporally heterogeneous observation data, satellite cloud images, and two-dimensional planar radar echo data. While ensuring second-level inference timeliness, it overcomes the error accumulation of long-sequence extrapolation in traditional models and outputs high spatiotemporal resolution short-term meteorological forecast results that fully comply with physical prior constraints.

[0006] The technical solution adopted in this invention is: A new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints includes the following steps: Acquire multi-source meteorological data for the target area, and align the multi-source meteorological data to a uniform spatiotemporal resolution to obtain multi-source aligned data; wherein, the multi-source meteorological data includes real-time observation data, satellite cloud image data, and radar echo data; High-dimensional features are extracted from each modality of the multi-source aligned data using multiple dedicated encoders to obtain high-dimensional hidden state features corresponding to each modality; wherein the multiple dedicated encoders include an observation real-time encoder, a satellite encoder, and a radar encoder; After spatial alignment of the high-dimensional hidden state features corresponding to each modality, the input is into the cross-attention fusion layer for adaptive weighted fusion to obtain multi-source fused hidden state features. The multi-source fusion latent state features are input into the spatiotemporal decoupling Transformer, and spatiotemporal correlation features are obtained by alternately executing temporal global self-attention calculation and spatial moving window self-attention calculation. Based on the aforementioned spatiotemporal correlation features, the latent conditional field features for multiple future prediction time steps are generated by performing autoregressive iterative deduction in the latent space through recurrent neural network units. Using the hidden conditional field features as guiding conditions, the Gaussian random noise sequence is iteratively denoised through a conditional diffusion model to generate a pre-prediction field. The preceding prediction field is processed by a forced physical constraint layer. The wind speed feasible region projection operator in the forced physical constraint layer maps the out-of-bounds value of the wind speed channel to a preset safe range. The day and night gate switch in the forced physical constraint layer performs zeroing processing on the irradiance channel according to the solar altitude angle, and outputs the final wind speed prediction result and irradiance prediction result for multiple future prediction time steps.

[0007] Furthermore, the observation real-time encoder includes a spatiotemporally separated three-dimensional convolutional network. The spatiotemporally separated three-dimensional convolutional network uses a one-dimensional convolutional kernel in the time dimension and a two-dimensional convolutional kernel in the spatial dimension to perform decoupled convolution operations on the observation real-time data, and outputs a compact spatiotemporal background field feature map through three-dimensional residual blocks and multi-level convolution downsampling.

[0008] Furthermore, the satellite encoder is built on the SwinTransformer architecture, alternately using standard window attention and moving window attention to perform self-attention calculations on the satellite cloud image data to extract multi-scale cloud image features; the satellite encoder also generates a daily periodic feature vector from the hour of the current moment through an embedded daily periodic encoder, and the calculation formula for the daily periodic feature vector is:

[0009] In the formula, Let be the diurnal eigenvector at time t; The current hour; is the learnable projection weight matrix; sin and cos are trigonometric functions; T is the matrix transpose.

[0010] Furthermore, the radar encoder employs a spatiotemporal convolutional layer to perform multi-step spatiotemporal joint extraction on the radar echo data to obtain a planar reflectivity evolution feature map; wherein, the radar encoder only performs spatial convolution and temporal convolution operations on the horizontal plane dimension of the radar echo data, without configuring an attention operator in the vertical height dimension; the spatiotemporal convolutional layer captures the diffusion velocity, advection displacement, and morphological evolution features of clouds and precipitation systems on the horizontal plane through convolutional blocks and one-dimensional temporal convolution.

[0011] Furthermore, a modal gating mechanism is introduced into the cross-attention fusion layer. The modal gating mechanism adaptively learns the fusion weights of each modality based on the features of the current time and spatial location. The fusion weights are obtained by nonlinearly mapping the high-dimensional hidden state features of each modality through a multilayer perceptron composed of two fully connected layers, and then normalizing them through the Softmax function to obtain the spatial point-by-point fusion weights of each modality. The cross-attention fusion layer uses the high-dimensional hidden state features output by the observation real-time encoder as the query vector, and the high-dimensional hidden state features output by the satellite encoder and radar encoder as the key vector and value vector, respectively, to perform cross-attention calculation. The fusion weights are then used to adaptively weight and sum the outputs of multiple attention heads to obtain the multi-source fused hidden state features.

[0012] Furthermore, the temporal global self-attention calculation compresses all historical moment features of each spatial location into a query matrix, a key matrix, and a value matrix. The core calculation formula is as follows:

[0013] In the formula, These are the query matrix, key matrix, and value matrix, respectively, representing the time dimension. This is the normalization function; Let T be the number of hidden features of the vector; T is the transpose of the matrix.

[0014] Furthermore, the recurrent neural network unit is a ConvLSTM unit. The ConvLSTM unit uses the spatiotemporal correlation features as the initial hidden state and combines them with the future time position encoding to perform autoregressive iteration, generating the hidden conditional field features step by step. The formula for calculating the future time location code is as follows:

[0015] In the formula, D represents the prediction duration corresponding to the current extrapolation step, D is the dimension of the input feature channels of ConvLSTM, and k is the channel index; It is a temporal location encoding vector used to quantify temporal periodic priors of different prediction durations and distinguish between short-term and medium-to-long-term meteorological evolution differences; After encoding, the spatiotemporal correlation features are concatenated and input into a ConvLSTM unit to constrain the temporal evolution pattern of stepwise extrapolation; The iteration of the ConvLSTM unit consists of five sets of convolution operations: input gate, forget gate, output gate, cell state, and hidden state. The core iterative formula is as follows:

[0016] In the formula, , , , , These represent the input gate, forget gate, output gate, cell state, and hidden state at time t, respectively. The input features at time t are the combined features of the spatiotemporal correlation features at the current time and the live broadcast field. and The hidden state and cell state at time (t-1); is the Sigmoid activation function, which maps the input to the (0, 1) interval; tanh is the hyperbolic tangent activation function, which maps the input to the (-1, 1) interval; This is a two-dimensional convolution operation; For Hadamah accumulation; , , , , , , , These are the weight matrices from input features to the input gate, from the hidden state at the previous time step to the input gate, from input features to the forget gate, from the hidden state at the previous time step to the forget gate, from input features to the output gate, from the hidden state at the previous time step to the output gate, from input features to the candidate cell state, and from the hidden state at the previous time step to the candidate cell state, respectively. , , , These are the bias vectors for the input gate, forget gate, output gate, and candidate cell state, respectively.

[0017] Furthermore, the feasible region projection operator maps the original wind speed prediction value of the preceding prediction field to the interval between a preset lower wind speed limit and a preset upper wind speed limit using a Sigmoid function. The calculation formula for the feasible region projection operator is as follows:

[0018] In the formula, These are the mapped wind speed predictions; The original wind speed values ​​in the forecast field; For the Sigmoid function; and These are the upper and lower limits of the predicted wind speed, respectively. The day / night gating switch first applies a non-negative constraint to the original irradiance prediction value of the preceding prediction field using the Softplus function, and then performs nighttime zeroing processing using an indicative function based on the solar altitude angle at the current prediction time. The calculation formula for the day / night gating switch is as follows:

[0019] In the formula, This is the final predicted irradiance value; The original irradiance value in the prediction field; Softplus is the non-negative activation function; Let be the solar altitude angle at time t; This is an indicative function; it takes the value 1 when the solar altitude angle is greater than or equal to 0°, and 0 otherwise.

[0020] Furthermore, the method employs a random modality discarding strategy during model training. In each training iteration, at least one modality of data from the observed real-time data, satellite cloud image data, and radar echo data is randomly discarded with a preset probability. When a modality of data is discarded, the encoder input corresponding to that modality is set to a zero tensor to enhance the robustness of the model when some modal data is missing. The conditional diffusion model uses the DDIM deterministic sampling algorithm for denoising iteration.

[0021] Another technical solution adopted by the present invention is: A new energy meteorological forecasting system based on multi-source fusion conditional diffusion and physical constraints includes: The multi-source input parallel coding module includes: The data acquisition and processing unit is used to acquire multi-source meteorological data of the target area and align the multi-source meteorological data to a uniform spatiotemporal resolution to obtain multi-source aligned data; wherein, the multi-source meteorological data includes real-time observation data, satellite cloud image data and radar echo data; The multi-source feature extraction unit is used to extract high-dimensional features from each modality data in the multi-source aligned data through multiple dedicated encoders to obtain the high-dimensional hidden state features corresponding to each modality. The dynamic feature fusion module is used to spatially align the high-dimensional hidden state features corresponding to each modality and then input them into the cross-attention fusion layer for adaptive weighted fusion to obtain multi-source fused hidden state features. The spatiotemporal correlation modeling module is used to input the multi-source fusion latent state features into the spatiotemporal decoupling Transformer, and obtain the spatiotemporal correlation features by alternately executing temporal global self-attention calculation and spatial moving window self-attention calculation; The dynamic state deduction and autoregressive expansion module is used to generate hidden conditional field features for multiple future prediction time steps by performing autoregressive iterative deduction in the latent space through recurrent neural network units based on the spatiotemporal correlation features. The conditional diffusion generation decoding module is used to use the hidden conditional field features as guiding conditions to iteratively denoise the Gaussian random noise sequence through the conditional diffusion model to generate a pre-prediction field. The physical constraint and output module is used to process the pre-prediction field through a forced physical constraint layer. The wind speed feasible region projection operator in the forced physical constraint layer maps the out-of-bounds value of the wind speed channel to a preset safe range. The day and night gate switch in the forced physical constraint layer performs zeroing processing on the irradiance channel according to the solar altitude angle, and outputs the final wind speed prediction result and irradiance prediction result for multiple future prediction time steps.

[0022] Compared with the prior art, the present invention has the following advantages: 1) This invention extracts high-dimensional features from real-time observations, satellite cloud images, and radar echo data by constructing multiple dedicated encoders, and introduces a modal gating mechanism in the cross-attention fusion layer to achieve adaptive weighted fusion, thus solving the problem of data fragmentation and dimensional mismatch caused by inconsistent observation frequencies and spatial grid scales in multi-source meteorological data. Compared with existing prediction methods that rely on only a single data source, this invention integrates multimodal data to provide the model with a more comprehensive three-dimensional meteorological observation background. 2) This invention decouples the traditional all-time-space self-attention into global attention in the time dimension and moving window self-attention in the spatial dimension, thereby reducing the computational complexity of spatiotemporal correlation modeling, solving the problems of memory overflow and computational explosion in long-time-series high-resolution grid scenarios, and greatly reducing memory usage and computational overhead during multi-source data training and inference. 3) This invention solves the problems of forecast ambiguity and exponential error accumulation caused by traditional deep learning models directly performing multi-step prediction extrapolation at the pixel level by performing nonlinear extrapolation of meteorological evolution laws in a low-dimensional latent space using ConvLSTM units and combining the generation mechanism of conditional diffusion models. Compared with existing pixel-level extrapolation methods, it effectively breaks the error divergence chain and ensures the stability of long-term extrapolation. 4) This invention utilizes a conditional diffusion model to gradually recover and generate a preliminary prediction field from a Gaussian random noise sequence. Combined with DDIM deterministic sampling, the number of inference steps is compressed to 10 to 20 steps, solving the problem of smoothing meteorological extreme values ​​caused by direct prediction of values ​​in traditional regression models. It can both finely extrapolate the details of meteorological evolution and has excellent uncertainty estimation capabilities. 5) This invention solves the model illusion problem of outliers that violate physical laws in the output of deep learning generative models by setting a wind speed feasible domain projection operator and a day / night gate switch at the end of the model output, and implementing hard mathematical projection and geometric astronomical logic gate interception on the pre-prediction field; compared with the existing methods that only perform soft constraint penalties in the loss function, this invention provides an absolute physical safety line and directly guarantees the industrial applicability of the output results. 6) By introducing a random modality discarding mechanism during the training phase, this invention enhances the model's fault tolerance and robust prediction capabilities when some meteorological data is not transmitted in a timely manner or is abnormally missing in actual production environments, and has good industrial-grade production implementation capabilities. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of the new energy meteorological forecasting method in this invention; Figure 2 This is a diagram of the architecture of the new energy meteorological forecasting system in this invention; Figure 3 A heatmap comparing the predicted results with the measured values ​​provided in the embodiments of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1

[0026] This embodiment details the overall architecture of a new energy meteorological forecasting system based on multi-source fusion conditional diffusion and physical constraints. The system adopts a joint architecture of "multimodal encoder - spatiotemporal self-attention association - hidden state autoregressive expansion - conditional diffusion decoder", and the structure of each core module is as follows: like Figures 1 to 2 As shown, the system includes a multi-source input parallel encoding module, a dynamic feature fusion module, a spatiotemporal correlation modeling module, a dynamic state deduction and autoregressive expansion module, a conditional diffusion generation and decoding module, and a physical constraint and output module; among which, the multi-source input parallel encoding module includes a data acquisition and processing unit and a multi-source feature extraction unit.

[0027] During system operation, the data acquisition and processing unit in the multi-source input parallel encoding module first synchronously collects real-time observation data, satellite cloud image data, and radar echo data of the target area. The real-time observation data consists of numerical grid data with a 60-minute time resolution, including elements such as wind speed and irradiance. The satellite cloud image data is a multi-channel time-series image with a 10-minute time resolution. The radar echo data is a two-dimensional planar reflectivity image with a 6-minute time resolution. Specific acquisition parameters are shown in Table 1. Table 1. Acquisition parameters for real-time observation data, satellite cloud image data, and radar echo data of the target area.

[0028] The data acquisition and processing unit performs alignment operations on three types of data: real-time observation data, satellite cloud image data, and radar echo data. 1) Time alignment: Using 10 minutes as the unified time resolution, the observed data is upsampled from 60 minutes to 10 minutes through linear interpolation (12 time steps are generated in 2 hours), and the radar echo data is downsampled from 6 minutes to 10 minutes through thinning sampling (2 hours are reduced from 20 time steps to 12 time steps). The satellite cloud image data (original 10-minute resolution, 12 time steps) remains unchanged. 2) Spatial alignment: The original spatial resolution of all three types of data is 1km×1km, so no spatial interpolation is required. Spatial coordinate matching ensures that 64×64 grid points cover the same area.

[0029] Through the above operations, the three types of data—observational data, satellite cloud image data, and radar echo data—are aligned to a unified spatiotemporal resolution to obtain multi-source aligned data, which provides the basic data format for subsequent parallel coding.

[0030] In this embodiment, the input and output specifications of the data acquisition and processing unit are shown in Table 2: Table 2 Input / output specifications of the data acquisition and processing unit

[0031] In this embodiment, real-time observation data provides a large-scale meteorological background field, satellite cloud image data provides information on cloud spatial distribution and texture, and radar echo data provides the horizontal motion characteristics of the precipitation system. The three types of data have significant differences in time frequency and spatial resolution. Initial alignment is achieved through spatial interpolation and timestamp matching to solve the problem of data discontinuity and dimensional mismatch caused by inconsistent observation frequency and spatial grid scale of multi-source meteorological equipment.

[0032] The multi-source feature extraction unit in the multi-source input parallel coding module extracts high-dimensional latent state features of three modes through its included observational encoder, satellite encoder, and radar encoder, respectively, to capture the initial dynamic patterns of various meteorological elements. The input and output specifications of each encoder are shown in Table 3. Table 3 Input / output specifications of each encoder

[0033] In this embodiment, the observation real-time encoder uses spatiotemporal separation three-dimensional convolution to extract the spatiotemporal evolution law of the large-scale background field. Specifically, it uses a one-dimensional convolution kernel in the time dimension and a two-dimensional convolution kernel in the spatial dimension to perform decoupled convolution operation on the observation real-time data. Then, through three-dimensional residual blocks and multi-level convolution downsampling with a stride of 2, the spatial resolution of the input feature map is compressed to 1 / 2, 1 / 4 and 1 / 8 of the original resolution, and the number of channels is expanded to 64, 128 and 256, respectively. Finally, a compact spatiotemporal background field feature map is output.

[0034] The satellite encoder utilizes the self-attention mechanism of the SwingTransformer to capture the texture details and motion trajectories of cloud formations, and embeds a diurnal periodicity encoding unit to enhance the physical periodicity of the time dimension. Specifically, based on the SwingTransformer architecture, it alternately uses standard window attention and moving window attention to perform self-attention calculations on satellite cloud image data to extract multi-scale cloud image features. The satellite encoder also uses the embedded diurnal periodicity encoding unit to generate a diurnal periodicity feature vector from the current hour. The formula for calculating the diurnal periodicity feature vector is as follows:

[0035] In the formula, Let be the diurnal eigenvector at time t; The current hour; is the learnable projection weight matrix; sin and cos are trigonometric functions; T is the matrix transpose.

[0036] The radar encoder uses a spatiotemporal convolutional layer to perform multi-step spatiotemporal joint extraction on radar echo data to obtain a planar reflectivity evolution feature map. Spatial convolution and temporal convolution operations are performed only in the horizontal plane dimension of the radar echo data, without configuring an attention operator in the vertical height dimension. The spatiotemporal convolutional layer captures the diffusion velocity, advection displacement and morphological evolution features of clouds and precipitation systems on the horizontal plane through convolutional blocks and one-dimensional temporal convolution.

[0037] In this embodiment, dedicated encoders are designed to extract features based on the differences in physical characteristics of the three types of data. The high-dimensional hidden state features output by the three encoders retain the physical characteristics of their respective modes, solving the technical problem that traditional general neural networks cannot adapt to their specific physical laws when extracting multiple meteorological elements. In addition, the invalid vertical attention calculation in radar data processing is eliminated, which greatly reduces the computational dimension.

[0038] The dynamic feature fusion module aligns the high-dimensional latent state features of the three modalities through spatial bilinear interpolation, then performs dimensionality reduction through channel concatenation and 1×1 convolution, and finally performs adaptive weighted fusion through a cross-attention fusion layer to generate multi-source fused latent states. This fusion layer introduces a modality gating mechanism, whereby the network adaptively learns the contribution weights of each modality based on the current time and spatial location. Specifically, the fusion weights are obtained by nonlinearly mapping the high-dimensional latent state features of each modality through a multilayer perceptron composed of two fully connected layers, and then normalizing them using the Softmax function to obtain the spatial point-by-point fusion weights of each modality. The cross-attention fusion layer uses the high-dimensional latent state features output by the observation real-time encoder as the query vector, and the high-dimensional latent state features output by the satellite encoder and radar encoder as the key vector and value vector, respectively, to perform cross-attention calculation, and then performs adaptive weighted summation of the outputs of multiple attention heads through the fusion weights.

[0039] In this embodiment, the generated multi-source fusion latent state features include not only large-scale background field information, but also cloud details and precipitation system motion information; adaptive gating fusion can automatically adjust the contribution ratio of each modality according to different scenarios, avoiding the defect that fixed fusion weights cannot adapt to complex and variable meteorological scenarios, and can still ensure the effectiveness of fusion features when the quality of multi-source data fluctuates.

[0040] The spatiotemporal correlation modeling module processes multi-source fusion latent state features through a spatiotemporal decoupling Transformer, alternately executing temporal global self-attention and spatial moving window self-attention calculations to obtain spatiotemporal correlation features. Temporal attention enables each spatial location to independently capture long-term dependencies along the time axis, while spatial attention independently performs correlation calculations on the spatial regions within the window at each time step. The temporal global self-attention calculation compresses all historical moment features of each spatial location into a query matrix, a key matrix, and a value matrix. The core calculation formula is as follows:

[0041] In the formula, These are the query matrix, key matrix, and value matrix, respectively, representing the time dimension. This is the normalization function; Let T be the number of hidden features of the vector; T is the transpose of the matrix.

[0042] In this embodiment, the decoupling design of global time and local space solves the problems of memory overflow and computational explosion caused by traditional all-time and all-space attention mechanisms in long-time high-resolution mesh scenarios, and significantly reduces computational complexity.

[0043] The dynamic state deduction and autoregressive expansion module takes the spatiotemporal correlation features as the initial hidden state, and performs autoregressive iterative deduction by combining ConvLSTM units with future time position encoding to generate hidden conditional field features for multiple future prediction time steps. The formula for calculating the future time location code is as follows:

[0044] In the formula, D represents the prediction duration corresponding to the current extrapolation step, D is the dimension of the input feature channels of ConvLSTM, and k is the channel index; It is a temporal location encoding vector used to quantify temporal periodic priors of different prediction durations and distinguish between short-term and medium-to-long-term meteorological evolution differences; After encoding, the spatiotemporal correlation features are concatenated and input into a ConvLSTM unit to constrain the temporal evolution pattern of stepwise extrapolation; The iteration of the ConvLSTM unit consists of five sets of convolution operations: input gate, forget gate, output gate, cell state, and hidden state. The core iterative formula is as follows:

[0045] In the formula, , , , , These represent the input gate, forget gate, output gate, cell state, and hidden state at time t, respectively. The input features at time t are the combined features of the spatiotemporal correlation features at the current time and the live broadcast field. and The hidden state and cell state at time (t-1); is the Sigmoid activation function, which maps the input to the (0, 1) interval; tanh is the hyperbolic tangent activation function, which maps the input to the (-1, 1) interval; This is a two-dimensional convolution operation; For Hadamah accumulation; , , , , , , , These are the weight matrices from input features to the input gate, from the hidden state at the previous time step to the input gate, from input features to the forget gate, from the hidden state at the previous time step to the forget gate, from input features to the output gate, from the hidden state at the previous time step to the output gate, from input features to the candidate cell state, and from the hidden state at the previous time step to the candidate cell state, respectively. , , , These are the bias vectors for the input gate, forget gate, output gate, and candidate cell state, respectively.

[0046] In this embodiment, the ConvLSTM unit performs autoregressive dynamic extrapolation in a low-dimensional latent space. This design solves the problems of image blurring and exponential error accumulation caused by traditional deep learning models performing multi-step prediction extrapolation at the pixel level. It performs nonlinear extrapolation of meteorological evolution laws in a low-dimensional latent space, effectively breaking the error divergence chain and avoiding error accumulation in pixel-level extrapolation.

[0047] The conditional diffusion generation decoding module uses hidden conditional field features as a guiding condition, inputs Gaussian random noise into the conditional diffusion decoder, and generates a pre-prediction field through multi-step denoising iteration, including the original wind speed prediction value and the original irradiance prediction value; the conditional diffusion decoder adopts the U-Net architecture.

[0048] In this embodiment, the forward process of the conditional diffusion model gradually adds noise to the target ground value, and the backward process learns to recover the target from the noise, which solves the problem of smoothing meteorological extreme values ​​caused by direct prediction of values ​​in traditional regression models, and ensures the authenticity of physical texture and the reliability of probability distribution of the generated results.

[0049] The physical constraint and output module applies hard physical constraints to the preceding prediction field using the wind speed feasible region projection operator and a day / night gating switch, outputting the final wind speed and irradiance prediction results for multiple future prediction time steps. The number of prediction time steps ranges from 6 to 36, corresponding to prediction windows of 1 to 6 hours in the future. Specifically, the wind speed feasible region projection operator maps the original wind speed prediction values ​​of the preceding prediction field to the interval between a preset lower wind speed limit and a preset upper wind speed limit using the Sigmoid function. The calculation formula for the wind speed feasible region projection operator is as follows:

[0050] In the formula, These are the mapped wind speed predictions; The original wind speed forecast value in the preceding forecast field; For the Sigmoid function; and These are the upper and lower limits of the predicted wind speed, respectively. The day / night gating switch first applies a non-negative constraint to the original irradiance prediction value of the preceding prediction field using the Softplus function, and then performs nighttime zeroing processing using an indicative function based on the solar altitude angle at the current prediction time; the calculation formula for the day / night gating switch is as follows:

[0051] In the formula, This is the final predicted irradiance value; The original irradiance value in the prediction field; Softplus is the non-negative activation function; Let be the solar altitude angle at time t; This is an indicative function; it takes the value 1 when the solar altitude angle is greater than or equal to 0°, and 0 otherwise.

[0052] In this embodiment, the forced physical constraint layer performs hard interception and mapping on the pre-prediction field. The wind speed channel maps out-of-bounds values ​​to a preset safe range through Sigmoid projection. The irradiance channel uses a day and night gating switch to intercept at night based on the solar altitude angle, completely eliminating the model illusion problem of negative wind speed and high irradiance at night that violates physical laws, such as the output of the deep learning model. Compared with the existing methods that only perform soft constraint penalties in the loss function, it provides an absolute physical safety defense.

[0053] Example 2

[0054] This embodiment details the training process of the new energy meteorological prediction model in Embodiment 1.

[0055] The training data consisted of historical meteorological observations of the target area (a 64km × 64km region around 33°N, 119°E) from January 1, 2023 to December 31, 2025. Each sample used multi-source time-series observations from the past 2 hours (12 time steps, one time step every 10 minutes) as input, and real-time wind speed and irradiance observed at 1km resolution for the corresponding next hour (6 time steps, one time step every 10 minutes) as ground truth labels. Approximately 1,000,000 samples were generated and divided into training, validation, and test sets in an 8:1:1 ratio.

[0056] The total loss function is obtained by weighted summation of three parts: noise prediction mean square error loss, spatial gradient smoothing loss, and physical feasible region regularization loss.

[0057] During model training, training samples are fed into the model in batches. From multi-source encoding, feature fusion, spatiotemporal correlation modeling to autoregressive inference, prediction results are finally generated through conditional diffusion and forced physical constraints.

[0058] During training, a random modality discarding strategy is adopted. In each training iteration, at least one modality of the observed real-time data, satellite cloud image data and radar echo data is randomly discarded with a preset probability of 0.1 to 0.3. Preferably, the preset probability is 0.2. When a modality of data is discarded, the encoder input corresponding to that modality is set to a zero tensor to enhance the robustness of the model when some modal data is missing.

[0059] Training employs the PyTorch deep learning framework, utilizing full-shard data parallelism to shard model parameters across a multi-GPU cluster. This, combined with BF16 mixed-precision training, exponential moving average weight maintenance, and gradient norm pruning with an upper limit of 1.0, ensures the convergence stability of large model training.

[0060] Repeat the iterative training process described above, calculating the prediction error on the validation set after each epoch (approximately 62,500 iterations). When the mean root mean square error on the validation set no longer decreases for 10 consecutive epochs, training is stopped, and the optimal weights are saved for production inference. In this embodiment, the model converges around the 50th epoch.

[0061] After training, the model performs inference on a single NVIDIA A100 GPU. The conditional diffusion model uses the DDIM deterministic sampling algorithm for denoising iteration, with the number of denoising steps set to 10 to 20, preferably 15. At this number of denoising steps, the inference time for a single prediction does not exceed 2 seconds, meeting the operational timeliness requirements of rolling forecast updates for new energy power plants (updated every 10 minutes).

[0062] Example 3

[0063] This embodiment details the process of performing a complete typical case prediction using the method in Embodiment 1.

[0064] A specific province was selected, with latitude and longitude ranging from 31.3°N to 34.39°N and longitude ranging from 116.86°E to 121.21°E, covering an area of ​​64km × 64km. The time range is from April 1, 2026 to April 30, 2026.

[0065] The forecast starts at 14:00 on April 6, 2026 (UTC+8), and the forecast targets are wind speed and irradiance for the next hour (14:00 to 15:00, with one hour step every 10 minutes). The key hyperparameters of the model are configured as follows: 128 hidden channels, 6 spatiotemporal Transformer coding layers, 6 future prediction steps, corresponding to an extrapolation window of 6 hours.

[0066] The data acquisition and processing unit in the multi-source input parallel coding module synchronously collected three types of data from this area from 12:00 to 14:00 on April 6, 2026 (the past 2 hours): ① Observational data: 12:00, 13:00, and 14:00 (generated into 12 10-minute interval time steps through linear interpolation); ② Satellite cloud image data: 12 time slots from 12:00 to 14:00 (native 10-minute resolution); ③ Radar echo data: 20 time intervals from 12:00 to 14:00 (sampling was reduced to 12 10-minute intervals).

[0067] The three types of data are spatially bilinearly interpolated to a uniform 1km×1km grid (64×64 grid points), and temporally aligned to a 10-minute resolution (12 time steps), outputting multi-source aligned data: F 实况raw : B×12×2×64×64 (B=1, the same below); F 卫星raw :B×12×3×64×64; F 雷达raw :B×12×1×64×64.

[0068] Multi-source aligned data are input into a multi-source feature extraction unit, and high-dimensional features are extracted through an observation real-time encoder, a satellite encoder, and a radar encoder, outputting high-dimensional latent state features: F 实况 : B×12×256×8×8; F 卫星 : B×12×256×8×8; F 雷达 :B×12×256×8×8.

[0069] After spatial alignment, the high-dimensional hidden state features are input into the dynamic feature fusion module and then into the cross-attention fusion layer for adaptive weighted fusion, outputting the multi-source fused hidden state features F. fused :B×12×256×8×8.

[0070] The spatiotemporal correlation modeling module processes multi-source fusion latent state features through a spatiotemporal decoupling Transformer, alternately performing temporal global self-attention and spatial moving window self-attention calculations, and outputs spatiotemporal correlation features F. st :B×12×256×8×8.

[0071] The dynamic state deduction and autoregressive expansion module performs autoregressive iterative deduction in the latent space using ConvLSTM units, with the initial hidden state H0 = st Initial cell state C0: B×12×256×8×8, initialized to zero tensor; Future time position encoding TE t : B×64×8×8; Output hidden conditional field feature Cond: B×6×256×8×8.

[0072] The conditional diffusion generation and decoding module uses hidden conditional field features to guide the conditions, and iteratively denoises Gaussian random noise through a conditional diffusion model. Inputs include hidden conditional field features Cond: B×6×256×8×8; Gaussian random noise Z: B×6×2×64×64; and output the pre-prediction field X. pred : B×6×2×64×64; Channel 1 is the original wind speed prediction value V raw The second channel represents the original predicted irradiance value I. raw .

[0073] The physical constraint and output module processes the preceding prediction field through a forced physical constraint layer, resulting in the original wind speed prediction value V with a value range of (-∞, +∞). raw Input the feasible region projection operator for wind speed, and output the final predicted wind speed value V. pred : B×6×1×64×64, range [0, 30] m / s; original irradiance prediction value I in range (-∞, +∞) raw and the solar altitude angle α at each predicted time t Input the day / night gate switch, output the final predicted irradiance value I. pred : B×6×1×64×64, range [0, +∞), and all values ​​are zero at night.

[0074] The physical constraints and output module outputs the final predicted wind speed value V. pred (B×6×1×64×64, range [0, 30] m / s) and the final predicted irradiance value I pred (B×6×1×64×64, value range [0, +∞) and all values ​​return to zero at night), corresponding to the prediction of the next 6 time steps (one time step every 10 minutes in the next hour).

[0075] In this embodiment, real-time collected observation data, satellite cloud image data, and radar echo data are input into the model. After alignment and fusion, the features are recursively output by ConvLSTM to provide the preceding hidden field features for the next 6 time steps, which are used as guiding conditions input into the conditional diffusion module. The DDIM fast sampler is used to perform 20 steps of reverse denoising generation, with the computation time controlled within seconds. The generated grid is finally filtered out by the wind speed feasible domain projection operator and day / night gating switch to remove negative wind speeds and nighttime irradiance anomalies, generating a deterministic 1km×1km high-definition meteorological extrapolation grid.

[0076] like Figure 3 As shown, the comparison between the predicted results and the measured values ​​at 15:00 on April 6, 2026 (the predicted endpoint) is presented. It can be seen from the figure that the actual wind speed value is highly consistent with the overall trend of the heat map of the predicted area, and the model can track the fluctuations in wind speed; the actual irradiance value matches well with the heat map of the predicted area, accurately reflecting the rising and falling trends and fluctuations in irradiance over time.

[0077] Table 4 shows the root mean square error statistics between predicted and measured values ​​from April 1, 2026 to April 30, 2026. Table 4. Statistical Table of Experimental Errors

[0078] As shown in Table 4, the overall average root mean square error of wind speed prediction is 0.389, and the overall average root mean square error of irradiance prediction is 56.142, indicating that the present invention has high prediction accuracy over the entire time period.

[0079] This invention combines three dedicated encoders—an observational encoder, a satellite encoder, and a radar encoder—for short-term weather forecasting. It integrates a spatiotemporally decoupled Transformer with ConvLSTM latent space inference to reduce computational complexity. A hard physical constraint layer is set at the end of the model output, employing a wind speed feasible domain projection operator and a day / night gating switch to implement hard interception, ensuring that each frame's prediction fully conforms to physical laws. Through end-to-end collaboration among various modules in the system, a complete mapping from raw, multi-source heterogeneous meteorological data to high-resolution prediction results that satisfy physical laws is achieved while ensuring inference timeliness. Compared to traditional single-source data-driven models, this invention reduces prediction bias under extreme weather conditions, achieves minute-level inference speeds to meet the operational timeliness requirements of rolling forecast updates for new energy power plants, and provides refined meteorological input support for power forecasting of wind and solar power plants with a 1km high-resolution output.

[0080] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the principles and essence of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints, characterized in that, Includes the following steps: Acquire multi-source meteorological data for the target area, and align the multi-source meteorological data to a uniform spatiotemporal resolution to obtain multi-source aligned data; wherein, the multi-source meteorological data includes real-time observation data, satellite cloud image data, and radar echo data; High-dimensional features are extracted from each modality of the multi-source aligned data using multiple dedicated encoders to obtain high-dimensional hidden state features corresponding to each modality; wherein the multiple dedicated encoders include an observation real-time encoder, a satellite encoder, and a radar encoder; After spatial alignment of the high-dimensional hidden state features corresponding to each modality, the input is into the cross-attention fusion layer for adaptive weighted fusion to obtain multi-source fused hidden state features. The multi-source fusion latent state features are input into the spatiotemporal decoupling Transformer, and spatiotemporal correlation features are obtained by alternately executing temporal global self-attention calculation and spatial moving window self-attention calculation. Based on the aforementioned spatiotemporal correlation features, the latent conditional field features for multiple future prediction time steps are generated by performing autoregressive iterative deduction in the latent space through recurrent neural network units. Using the hidden conditional field features as guiding conditions, the Gaussian random noise sequence is iteratively denoised through a conditional diffusion model to generate a pre-prediction field. The preceding prediction field is processed by a forced physical constraint layer. The wind speed feasible region projection operator in the forced physical constraint layer maps the out-of-bounds value of the wind speed channel to a preset safe range. The day and night gate switch in the forced physical constraint layer performs zeroing processing on the irradiance channel according to the solar altitude angle, and outputs the final wind speed prediction result and irradiance prediction result for multiple future prediction time steps.

2. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The observation data encoder includes a spatiotemporally separated 3D convolutional network. The spatiotemporally separated 3D convolutional network uses a one-dimensional convolutional kernel in the time dimension and a two-dimensional convolutional kernel in the spatial dimension to perform decoupled convolution operations on the observation data, and outputs a compact spatiotemporal background field feature map through 3D residual blocks and multi-level convolution downsampling.

3. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The satellite encoder is built on the SwingTransformer architecture and alternately uses standard window attention and moving window attention to perform self-attention calculations on the satellite cloud image data to extract multi-scale cloud image features. The satellite encoder also generates a daily periodic feature vector from the current hour using an embedded daily periodic encoder. The formula for calculating the daily periodic feature vector is as follows: In the formula, Let be the diurnal eigenvector at time t; The current hour; is the learnable projection weight matrix; sin and cos are trigonometric functions; T is the matrix transpose.

4. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The radar encoder uses a spatiotemporal convolutional layer to perform multi-step spatiotemporal joint extraction on the radar echo data to obtain a planar reflectivity evolution feature map. The radar encoder only performs spatial convolution and temporal convolution operations on the horizontal plane dimension of the radar echo data, without configuring an attention operator in the vertical height dimension. The spatiotemporal convolutional layer captures the diffusion velocity, advection displacement, and morphological evolution features of clouds and precipitation systems on the horizontal plane through convolutional blocks and one-dimensional temporal convolution.

5. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The cross-attention fusion layer introduces a modality gating mechanism, which adaptively learns the fusion weights of each modality based on the features of the current time and spatial location. The fusion weights are obtained by nonlinearly mapping the high-dimensional hidden state features of each modality through a multilayer perceptron composed of two fully connected layers, and then normalizing them using the Softmax function to obtain the spatial point-by-point fusion weights of each modality. The cross-attention fusion layer uses the high-dimensional hidden state features output by the observation real-time encoder as the query vector, and the high-dimensional hidden state features output by the satellite encoder and radar encoder as the key vector and value vector, respectively, to perform cross-attention calculation. The fusion weights are then used to adaptively weight and sum the outputs of multiple attention heads to obtain the multi-source fused hidden state features.

6. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The temporal global self-attention calculation compresses all historical time features of each spatial location into a query matrix, a key matrix, and a value matrix. The core calculation formula is as follows: In the formula, These are the query matrix, key matrix, and value matrix, respectively, representing the time dimension. This is the normalization function; Let T be the number of hidden features of the vector; T is the transpose of the matrix.

7. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The recurrent neural network unit is a ConvLSTM unit. The ConvLSTM unit uses the spatiotemporal correlation features as the initial hidden state, combines the future time position encoding to perform autoregressive iteration, and generates the hidden conditional field features step by step. The formula for calculating the future time location code is as follows: In the formula, D represents the prediction duration corresponding to the current extrapolation step, D is the dimension of the input feature channels of ConvLSTM, and k is the channel index; It is a temporal location encoding vector used to quantify the temporal periodic priors of different prediction durations and distinguish the differences between short-term and medium-to-long-term meteorological evolution; After encoding, the spatiotemporal correlation features are concatenated and input into a ConvLSTM unit to constrain the temporal evolution pattern of stepwise extrapolation; The iteration of the ConvLSTM unit consists of five sets of convolution operations: input gate, forget gate, output gate, cell state, and hidden state. The core iterative formula is as follows: In the formula, , , , , These represent the input gate, forget gate, output gate, cell state, and hidden state at time t, respectively. The input features at time t are the combined features of the spatiotemporal correlation features at the current time and the live broadcast field. and The hidden state and cell state at time (t-1); is the Sigmoid activation function, which maps the input to the interval (0, 1); tanh is the hyperbolic tangent activation function, which maps the input to the interval (-1, 1); This is a two-dimensional convolution operation; For Hadamah accumulation; , , , , , , , These are the weight matrices from input features to the input gate, from the hidden state at the previous time step to the input gate, from input features to the forget gate, from the hidden state at the previous time step to the forget gate, from input features to the output gate, from the hidden state at the previous time step to the output gate, from input features to the candidate cell state, and from the hidden state at the previous time step to the candidate cell state, respectively. , , , These are the bias vectors for the input gate, forget gate, output gate, and candidate cell state, respectively.

8. The new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The feasible region projection operator maps the original wind speed prediction values ​​of the preceding prediction field to the interval between a preset lower wind speed limit and a preset upper wind speed limit using the Sigmoid function. The calculation formula for the feasible region projection operator is as follows: In the formula, These are the mapped wind speed predictions; The original wind speed values ​​in the forecast field; For the Sigmoid function; and These are the upper and lower limits of the predicted wind speed, respectively. The day / night gating switch first applies a non-negative constraint to the original irradiance prediction value of the preceding prediction field using the Softplus function, and then performs nighttime zeroing processing using an indicative function based on the solar altitude angle at the current prediction time. The calculation formula for the day / night gating switch is as follows: In the formula, This is the final predicted irradiance value; This refers to the original irradiance value in the preceding prediction field; Softplus is a non-negative activation function; Let be the solar altitude angle at time t; This is an indicative function; it takes the value 1 when the solar altitude angle is greater than or equal to 0°, and 0 otherwise.

9. A new energy meteorological forecasting method based on multi-source fusion conditional diffusion and physical constraints according to claim 1, characterized in that, The method employs a random modality discarding strategy during model training. In each training iteration, at least one modality of data from the observed real-time data, satellite cloud image data, and radar echo data is randomly discarded with a preset probability. When a modality of data is discarded, the encoder input corresponding to that modality is set to a zero tensor to enhance the robustness of the model when some modal data is missing. The conditional diffusion model uses the DDIM deterministic sampling algorithm for denoising iteration.

10. A new energy meteorological forecasting system based on multi-source fusion conditional diffusion and physical constraints, characterized in that, include: The multi-source input parallel coding module includes: The data acquisition and processing unit is used to acquire multi-source meteorological data of the target area and align the multi-source meteorological data to a uniform spatiotemporal resolution to obtain multi-source aligned data; wherein, the multi-source meteorological data includes real-time observation data, satellite cloud image data and radar echo data; The multi-source feature extraction unit is used to extract high-dimensional features from each modality data in the multi-source aligned data through multiple dedicated encoders to obtain the high-dimensional hidden state features corresponding to each modality. The dynamic feature fusion module is used to spatially align the high-dimensional hidden state features corresponding to each modality and then input them into the cross-attention fusion layer for adaptive weighted fusion to obtain multi-source fused hidden state features. The spatiotemporal correlation modeling module is used to input the multi-source fusion latent state features into the spatiotemporal decoupling Transformer, and obtain the spatiotemporal correlation features by alternately executing temporal global self-attention calculation and spatial moving window self-attention calculation; The dynamic state deduction and autoregressive expansion module is used to generate hidden conditional field features for multiple future prediction time steps by performing autoregressive iterative deduction in the latent space through recurrent neural network units based on the spatiotemporal correlation features. The conditional diffusion generation decoding module is used to use the hidden conditional field features as guiding conditions to iteratively denoise the Gaussian random noise sequence through the conditional diffusion model to generate a pre-prediction field. The physical constraint and output module is used to process the pre-prediction field through a forced physical constraint layer. The wind speed feasible region projection operator in the forced physical constraint layer maps the out-of-bounds value of the wind speed channel to a preset safe range. The day and night gate switch in the forced physical constraint layer performs zeroing processing on the irradiance channel according to the solar altitude angle, and outputs the final wind speed prediction result and irradiance prediction result for multiple future prediction time steps.

Images