New energy station meteorological data reconstruction method and system based on diffusion model

Abstract

The invention discloses a new energy station meteorological data reconstruction method and system based on a diffusion model, and belongs to the technical field of meteorological data processing and new energy. The method comprises the following steps: constructing a multi-modal condition input set; inputting the multi-modal input set into a designed space-time diffusion model, and outputting a preliminarily generated high-resolution meteorological data graph; inputting the space-time diffusion model again, embedding physical constraints in the space-time diffusion model, and outputting a high-resolution meteorological estimation map; training the space-time diffusion model by adopting a two-stage training strategy to obtain a diffusion generation model; and inputting low-resolution meteorological data acquired in real time into the diffusion generation model, and outputting a target high-resolution meteorological data graph to complete reconstruction. According to the invention, through collaborative innovation of multi-modal data fusion, physical constraint embedding and an adaptive training strategy, high-precision generation and physical consistency guarantee of the meteorological field are realized.

Description

Technical Field

[0001] This invention relates to the intersection of meteorological data processing and new energy technologies, and in particular to a method and system for reconstructing meteorological data from new energy power stations based on a diffusion model. Background Technology

[0002] With the rapid increase in the proportion of photovoltaic power generation in new power systems, high-resolution reconstruction of meteorological data has become a key link in improving the absorption capacity of new energy sources and the stability of the power grid. Traditional numerical weather prediction (NWP) is limited by the computational complexity of physical models, and its spatial resolution is usually only at the kilometer level, making it difficult to capture the local micro-meteorological characteristics of photovoltaic power plants (such as cloud gaps at the hundred-meter level and abrupt changes in irradiance gradients). While deep learning-based super-resolution methods can improve data resolution, they face the risk of distortion of physical laws: for example, the irradiance field reconstructed by generative adversarial networks may violate the laws of atmospheric radiative transfer, causing the predicted photovoltaic output to exceed the theoretical limit of the modules.

[0003] Existing methods generally lack deep coupling with the operation data of new energy power plants. The minute-level fluctuations of photovoltaic power output curves can be used to invert local meteorological changes (such as a sudden drop in irradiance caused by cloud cover). However, traditional time series models are unable to establish a two-way coupling relationship between "power output and meteorology" due to insufficient nonlinear mapping capabilities, resulting in the meteorological reconstruction results being out of sync with the actual operating status of the power plants.

[0004] Therefore, how to solve the problem of reconstructing high-resolution meteorological data from new energy power plants is a current challenge. Summary of the Invention

[0005] To address the above problems, this invention provides a method and system for reconstructing meteorological data from new energy power stations based on a diffusion model. Through the collaborative innovation of multimodal data fusion, physical constraint embedding, and adaptive training strategies, it achieves high-precision generation of meteorological fields and ensures physical consistency.

[0006] The technical solution of this invention is: a method for reconstructing meteorological data from new energy power stations based on a diffusion model, comprising the following steps:

[0007] S1. Construct a multimodal conditional input set, including low-resolution meteorological data, time-series power output data of new energy power plants, and geographical feature data;

[0008] S2. Input the multimodal input set into the designed spatiotemporal diffusion model and output the initially generated high-resolution meteorological data map;

[0009] S3. Input the initially generated high-resolution meteorological data map, the power output time series data of new energy power stations and the geographical feature data into the spatiotemporal diffusion model again, embed physical constraints in the spatiotemporal diffusion model, and output a high-resolution meteorological estimation map.

[0010] S4. Based on the high-resolution meteorological estimation map, a two-stage training strategy is adopted to train the spatiotemporal diffusion model and obtain the diffusion generation model.

[0011] S5. Input the real-time acquired low-resolution meteorological data into the diffusion generation model, output the target high-resolution meteorological data map, and complete the reconstruction.

[0012] In step S2, the spatiotemporal diffusion model adopts a U-Net++ network structure and embeds a spatiotemporal attention module;

[0013] The spatiotemporal attention module calculates the following:

[0014]

[0015] In the formula, Q is the query vector, K is the key vector, V is the value vector, and d k Let K be the feature dimension, and T be the transpose of vector K.

[0016] In step S3, the physical constraints include energy conservation constraints, photovoltaic power output efficiency constraints, and fluid dynamic constraints.

[0017] Energy conservation constraint: Calculating the reconstructed irradiance field G rec Energy balance equation residuals:

[0018]

[0019] In the formula, L energy Let G be the energy conservation constraint loss function. rec For the irradiance field, Let be the partial derivative of the irradiance field with respect to time. The irradiance energy velocity vector, It is a divergence operator;

[0020] Photovoltaic power output efficiency constraint: based on real-time power output P real Reverse-constrained irradiance and temperature field:

[0021] L PV =‖P real -ηAG rec (1-0.05(T rec -25))‖ 2 ;

[0022] In the formula, L PV Let P be the photovoltaic power output efficiency constraint loss function.real For real-time power output, A is the effective area of ​​the photovoltaic module, and T is... rec Let η be the temperature field and η be the photovoltaic conversion efficiency.

[0023] Fluid dynamic constraints: on wind speed field Apply Navier-Stokes regularization:

[0024]

[0025] In the formula, L NS The fluid dynamics constraint loss function, This refers to the wind speed field.

[0026] In step S4, the two-stage training strategy includes a pre-training stage and a fine-tuning stage.

[0027] in,

[0028] The pre-training phase uses historical high-resolution radar data to train the spatiotemporal diffusion model, with the optimization objective being:

[0029]

[0030] In the formula, L pre For the pre-training loss function, Let ε be the expected value of the prediction error. θ Let x be the noise generation function. t The sample is a noisy sample, and ε is the actual noise.

[0031] The fine-tuning phase involves optimization using a joint physical constraint loss function, where the physical constraint loss function is:

[0032] L total =L diff +aL energy +bL PV +cL NS , ;

[0033] In the formula, L total Let L be the total loss function. diff The difference between the high-resolution meteorological data map and the actual value is represented by a, b, and c, which are dynamically adjustable coefficients.

[0034] In step S5, inference processing is performed by combining the accelerated sampling algorithm with the physical correction module, specifically as follows:

[0035] A denoising diffusion implicit model is used to accelerate the sampling algorithm and compress the number of sampling steps;

[0036] Perform physical correction after each sampling step:

[0037]

[0038] In the formula, z′ represents the data after physical correction. Let L be the gradient value of z. phys α represents the physical constraint loss, and α is the correction step size.

[0039] It also includes,

[0040] S6. Take the target high-resolution meteorological data map as input and perform post-processing operations, including super-resolution reconstruction and spatiotemporal consistency filtering.

[0041] In step S6, post-processing includes:

[0042] A physics-guided super-resolution network generates a 100m resolution field from 1km resolution data.

[0043] The temporal optical flow method is used to correct the cloud motion trajectory, and the displacement vector between adjacent frames is calculated:

[0044] Δx=argmin∑‖I t (x)-I t+1 (x+Δx)‖ 2 ;

[0045] In the formula, Δx is the displacement vector between adjacent frames, I t (x) represents the value at position x of the image in frame t; t+1 (x+Δx) is the value of the (t+1)th frame image at position x+Δx.

[0046] In step S1, the geographic feature data includes:

[0047] Elevation model data: 30m resolution digital elevation map from Space Shuttle radar topographic mapping;

[0048] Surface reflectance: calculated based on Landsat 8 satellite data;

[0049] Obstacle distribution: 0.5m resolution raster map of the three-dimensional outline of buildings surrounding the station.

[0050] A meteorological data reconstruction system for new energy power stations based on a diffusion model includes:

[0051] The construction module is used to build a multimodal conditional input set, including low-resolution meteorological data, time-series power output data of new energy power plants, and geographic feature data;

[0052] The data graph module is used to input the multimodal input set into the designed spatiotemporal diffusion model and output a preliminary high-resolution meteorological data graph.

[0053] The embedding module is used to input the initially generated high-resolution meteorological data map, the power output time series data of new energy power stations and geographical feature data back into the spatiotemporal diffusion model, embed physical constraints in the spatiotemporal diffusion model, and output a high-resolution meteorological estimation map.

[0054] The training module is used to train the spatiotemporal diffusion model based on the high-resolution meteorological estimation map using a two-stage training strategy to obtain the diffusion generation model.

[0055] The output module is used to input real-time acquired low-resolution meteorological data into the diffusion generation model and output a high-resolution meteorological data map of the target to complete the reconstruction.

[0056] The present invention has the following technical effects:

[0057] 1) Achieve high-precision meteorological element reconstruction: By introducing a spatiotemporal diffusion generation mechanism, it is possible to reconstruct a high-precision meteorological field with a spatial resolution of hundreds of meters and a temporal resolution of minutes based on low-resolution meteorological observation data, which significantly improves the spatial details and evolution continuity of the data, and provides more refined meteorological support for the prediction and scheduling of new energy power generation.

[0058] 2) Ensure the physical consistency of the meteorological field: embed physical constraints such as energy conservation equation, photovoltaic power output efficiency equation and Navier-Stokes equation during the diffusion model generation process to realize physical correlation modeling and dynamic balance control between meteorological variables, avoid physical inconsistencies and variable imbalances that are prone to occur in traditional deep learning methods, and improve the credibility and application reliability of the results.

[0059] 3) Enhance the ability to invert the actual power output of power plants: Using the power output curve of power plants as an auxiliary input, guide the diffusion process to converge in an interpretable direction, effectively improve the coupling between the meteorological field reconstruction results and the actual changes in new energy power, and thus realize intelligent inversion of meteorological data for power application scenarios.

[0060] 4) Improve training efficiency and generation speed: Adopt a two-stage training strategy of "pre-training + physical joint optimization" so that the model learns the spatiotemporal structure first and then integrates physical laws, resulting in high optimization efficiency; Combine the accelerated sampling algorithm and the physical correction mechanism at each step to improve the generation speed and stability of the inference stage and meet the real-time or near-real-time business requirements. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0062] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0063] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0064] like Figure 1 As shown, this invention provides a method for reconstructing meteorological data from new energy power stations based on a diffusion model, comprising the following steps:

[0065] S1. Construct a multimodal conditional input set, including low-resolution meteorological data (i.e., meteorological data with a spatial resolution of 25km or more), time-series data of power output from new energy power plants (such as photovoltaic power and wind turbine speed), and geographic feature data (such as surface reflectance).

[0066] S2. Input the multimodal input set into the designed spatiotemporal diffusion model (i.e., use the diffusion model to diffuse spatial range data and temporal data), and output the initially generated high-resolution meteorological data map (i.e. meteorological estimation data with a spatial resolution of less than 3km).

[0067] S3. Input the initially generated high-resolution meteorological data map, the power output time series data of the new energy power station and the geographical feature data into the spatiotemporal diffusion model again, embed physical constraints in the spatiotemporal diffusion model, and output a high-resolution meteorological estimation map (i.e., meteorological estimation data with a spatial resolution of less than 3km after considering physical constraints).

[0068] S4. Based on the high-resolution meteorological estimation map, a two-stage training strategy (i.e., progressive training through pre-training and fine-tuning) is adopted to train the spatiotemporal diffusion model and obtain the diffusion generation model.

[0069] S5. Input the real-time acquired low-resolution meteorological data into the diffusion generation model, output the target high-resolution meteorological data map (i.e., meteorological estimation data with a spatial resolution of less than 3km), and complete the reconstruction.

[0070] The principles behind each step are explained below:

[0071] In S1, the inputs first include low-resolution meteorological data, time-series power output data from renewable energy plants, and geographic feature data. These multi-source data undergo preprocessing, including format standardization, time alignment, spatial mapping, and normalization, to construct a standardized multimodal conditional input set. This input set serves as the foundation for subsequent model calculations, providing comprehensive prior conditions for high-resolution reconstruction of meteorological information.

[0072] In step S2, the multimodal conditional input set generated in step S1 is input into the designed spatiotemporal diffusion model. This model adopts a U-Net++ network structure and embeds a spatiotemporal attention module to extract and model the spatiotemporal dependency features of meteorological data during diffusion and de-diffusion. This spatiotemporal feature is then used to initially generate a high-resolution meteorological data map, providing a basic data representation for physical constraint guidance.

[0073] In step S3, the high-resolution meteorological data map output from step S2, along with the time-series power output data and geographical feature data of the original new energy power plants, are input into the model again. By introducing energy conservation, photovoltaic power output efficiency, and fluid dynamics constraints into the diffusion generation process, the generation process is guided to ensure physical consistency. The output is a high-resolution meteorological estimation map that satisfies physical laws, improving the reliability and interpretability of the model output.

[0074] In step S4, the high-resolution meteorological data map that satisfies physical laws, generated in step S3, is used as the prediction result. Historical high-resolution measured meteorological data is combined with this data as supervisory labels and input into the training framework for two-stage training. The pre-training stage optimizes the basic model parameters, while the fine-tuning stage introduces a physical constraint loss function to further correct generation errors. The output is a fully trained and optimized diffusion generation model, which can be used for meteorological reconstruction inference in real-world scenarios.

[0075] In S5, real-time acquired low-resolution meteorological data is input into the trained diffusion generation model, and rapid inference processing is performed using an accelerated sampling algorithm and a physical correction module. The output is a high-resolution meteorological data map with spatial resolution at the hundred-meter level and temporal resolution at the hour level, enabling accurate reconstruction of the microscale meteorological conditions of new energy power plants.

[0076] It also includes S6, post-processing optimization, and result output:

[0077] The target high-resolution meteorological data generated in step S5 is used as input, and further post-processing operations are performed, including super-resolution image thinning and spatiotemporal consistency filtering, to eliminate local jumps and boundary artifacts in the generated image. The final output is structured meteorological data with optimized quality, which can be directly applied to operational scenarios such as new energy power prediction and scheduling optimization.

[0078] The specific solution of the present invention is as follows:

[0079] S1. Multi-source data fusion and spatiotemporal feature construction.

[0080] This invention integrates low-resolution meteorological data (such as satellite cloud images and numerical weather prediction), time-series data of power generation from renewable energy plants (such as photovoltaic power and wind turbine speed), and geographic feature data, wherein...

[0081] Geographic feature data includes:

[0082] Elevation model data: Shuttle Radar Topography Mission (SRTM) 30m resolution digital elevation map;

[0083] Surface reflectance: calculated based on Landsat 8 satellite data;

[0084] Obstacle distribution: 0.5m resolution raster map of the three-dimensional outline of buildings surrounding the station.

[0085] New energy power stations include:

[0086] Photovoltaic power plants: collect data on module temperature, inverter efficiency, and output curves;

[0087] Wind farm: Collects wind turbine speed, yaw angle, and power output;

[0088] Energy storage power stations: Optimize charging and discharging strategies by combining weather forecasts.

[0089] The above multi-source time series data are aligned by Dynamic Time Wrapping (DTW), and Spatial Pyramid Pooling (SPP) is introduced to extract terrain-related irradiance distribution patterns.

[0090] S2. Design of the spatiotemporal diffusion model.

[0091] A spatiotemporal diffusion model based on U-Net++ is constructed, including:

[0092] A. Gaussian noise is gradually added to high-resolution meteorological data during the forward diffusion process, with a time step of T = 1000;

[0093] B. The inverse denoising network adopts a 3D U-Net++ structure, containing 5 encoder layers and 5 decoder layers. Each layer embeds a spatiotemporal attention module. Its calculation process is as follows:

[0094]

[0095] In the formula, Attention(Q,K,V) is the calculation formula for the attention mechanism. Q, K, and V are generated by convolution from the input features, where Q is the query vector, K is the key vector, V is the value vector, and d k Let T be the feature dimension, and T represent the transpose of vector K.

[0096] C. The decoder is connected to a physical constraint projection layer, which performs Navier-Stokes equation residual minimization calculations on the generated data.

[0097] S3. Physical constraint collaborative optimization mechanism.

[0098] Embedding triple physical constraints in diffusion training:

[0099] Energy dynamic balance: Reconstructing the irradiance field G using the radiative transfer equation rec The energy residual, and the loss function is:

[0100]

[0101] In the formula, L energy Let G be the energy conservation constraint loss function. rec For the irradiance field, Let be the partial derivative of the irradiance field with respect to time. The irradiance energy velocity vector, It is a divergence operator.

[0102] Photovoltaic power output efficiency constraints: Irradiance generation is optimized by inverse optimization of the photovoltaic panel efficiency equation, with the loss term being:

[0103] L PV =‖P real -ηAG rec (1-0.05(T rec -25))‖ 2 ;

[0104] In the formula, L PV Let P be the photovoltaic power output efficiency constraint loss function. real For real-time power output, A is the effective area of ​​the photovoltaic module, and T is... rec Let η be the temperature field and η be the photovoltaic conversion efficiency.

[0105] Fluid dynamic constraints: on the generated wind speed field The Navier-Stokes equations are applied to suppress non-physical turbulence modes. The constraint terms are weighted and summed to incorporate diffusion losses, forming a dual optimization mechanism of "data-driven generation + physical law correction".

[0106]

[0107] In the formula, L NS The fluid dynamics constraint loss function, This refers to the wind speed field.

[0108] S4. Two-stage progressive training strategy.

[0109] Phase 1 (Pre-training): The spatiotemporal diffusion model is trained on a historical high-resolution radar dataset. A noise scheduling strategy is used to gradually increase the weights of physical constraints, allowing the model to initially grasp the generation patterns of meteorological fields. The optimization objective is:

[0110]

[0111] In the formula, L pre For the pre-training loss function, Let ε be the expected value of the prediction error. θ Let x be the noise generation function. t The sample is a noisy sample, and ε is the actual noise.

[0112] Phase Two (Fine-tuning): Optimization is performed using the joint physical constraint loss function, where the physical constraint loss function is:

[0113] L total =L diff +aL energy +bL PV +cL NS ;

[0114] In the formula, L total Let L be the total loss function. diff The difference between the high-resolution meteorological data map and the actual value is represented by a, b, and c, which are dynamically adjustable coefficients.

[0115] S5. Real-time refactoring.

[0116] The sampling algorithm is accelerated by using Denoising Diffusion Implicit Models (DDIM), which compresses the number of sampling steps to 50.

[0117] Perform physical correction after each sampling step:

[0118]

[0119] In the formula, z′ represents the data after physical correction. Let L be the gradient value of z. phys =aL energy +bL PV +cL NS The physical constraint loss is represented by α = 0.01, which is the correction step size.

[0120] S6. Post-processing optimizes the generated results, including super-resolution reconstruction and spatiotemporal consistency filtering.

[0121] Post-processing includes:

[0122] A physics-guided super-resolution network generates a 100m resolution field from 1km resolution data.

[0123] The temporal optical flow method is used to correct the cloud motion trajectory, and the displacement vector between adjacent frames is calculated:

[0124] Δx=argmin∑‖I t (x)-It+1 (x+Δx)‖ 2 ;

[0125] In the formula, Δx is the displacement vector between adjacent frames, I t (x) represents the value at position x of the image in frame t; t+1 (x+Δx) is the value of the (t+1)th frame image at position x+Δx.

[0126] This invention also provides a meteorological data reconstruction system for new energy power stations based on a diffusion model, comprising:

[0127] The construction module is used to build a multimodal conditional input set, including low-resolution meteorological data, time-series power output data of new energy power plants, and geographic feature data;

[0128] The geographic feature data in the construction module includes:

[0129] Elevation model data: 30m resolution digital elevation map from Space Shuttle radar topographic mapping;

[0130] Surface reflectance: calculated based on Landsat 8 satellite data;

[0131] Obstacle distribution: 0.5m resolution raster map of the three-dimensional outline of buildings surrounding the station.

[0132] The data graph module is used to input the multimodal input set into the designed spatiotemporal diffusion model and output a preliminary high-resolution meteorological data graph.

[0133] The spatiotemporal diffusion model in the data graph module adopts the U-Net++ network structure and embeds a spatiotemporal attention module;

[0134] The spatiotemporal attention module calculates the following:

[0135]

[0136] In the formula, Q is the query vector, K is the key vector, V is the value vector, and d k Let K be the feature dimension, and T be the transpose of vector K.

[0137] The embedding module is used to input the initially generated high-resolution meteorological data map, the power output time series data of new energy power stations and geographical feature data back into the spatiotemporal diffusion model, embed physical constraints in the spatiotemporal diffusion model, and output a high-resolution meteorological estimation map.

[0138] The physical constraints in the embedded module include energy conservation constraints, photovoltaic power output efficiency constraints, and fluid dynamics constraints.

[0139] Energy conservation constraint: Calculating the reconstructed irradiance field G recEnergy balance equation residuals:

[0140]

[0141] In the formula, L energy Let G be the energy conservation constraint loss function. rec For the irradiance field, Let be the partial derivative of the irradiance field with respect to time. The irradiance energy velocity vector, It is a divergence operator;

[0142] Photovoltaic power output efficiency constraint: based on real-time power output P real Reverse-constrained irradiance and temperature field:

[0143] L PV =‖P real -ηAG rec (1-0.05(T rec -25))‖ 2 ;

[0144] In the formula, L PV Let P be the photovoltaic power output efficiency constraint loss function. real For real-time power output, A is the effective area of ​​the photovoltaic module, and T is... rec Let η be the temperature field and η be the photovoltaic conversion efficiency.

[0145] Fluid dynamic constraints: on wind speed field Apply Navier-Stokes regularization:

[0146]

[0147] In the formula, L NS The fluid dynamics constraint loss function, This refers to the wind speed field.

[0148] The training module is used to train the spatiotemporal diffusion model based on the high-resolution meteorological estimation map using a two-stage training strategy to obtain the diffusion generation model.

[0149] The two-stage training strategy in the training module includes a pre-training stage and a fine-tuning stage.

[0150] in,

[0151] The pre-training phase uses historical high-resolution radar data to train the spatiotemporal diffusion model, with the optimization objective being:

[0152]

[0153] In the formula, L pre For the pre-training loss function, Let ε be the expected value of the prediction error. θ Let x be the noise generation function. t The sample is a noisy sample, and ε is the actual noise.

[0154] The fine-tuning phase involves optimization using a joint physical constraint loss function, where the physical constraint loss function is:

[0155] L total =L diff +aL energy +bL PV +cL NS ;

[0156] In the formula, L total Let L be the total loss function. diff The difference between the high-resolution meteorological data map and the actual value is represented by a, b, and c, which are dynamically adjustable coefficients.

[0157] The output module is used to input real-time acquired low-resolution meteorological data into the diffusion generation model and output a high-resolution meteorological data map of the target to complete the reconstruction.

[0158] The output module combines an accelerated sampling algorithm with a physical correction module for inference processing, specifically:

[0159] A denoising diffusion implicit model is used to accelerate the sampling algorithm and compress the number of sampling steps;

[0160] Perform physical correction after each sampling step:

[0161]

[0162] In the formula, z′ represents the data after physical correction. Let L be the gradient value of z. phys α represents the physical constraint loss, and α is the correction step size.

[0163] This invention achieves high-precision generation and physical consistency assurance of meteorological fields through the collaborative innovation of multimodal data fusion, physical constraint embedding, and adaptive training strategies.

[0164] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for reconstructing meteorological data from new energy power stations based on a diffusion model, characterized in that, Includes the following steps: S1. Construct a multimodal conditional input set, including low-resolution meteorological data, time-series power output data of new energy power plants, and geographical feature data; S2. Input the multimodal input set into the designed spatiotemporal diffusion model and output the initially generated high-resolution meteorological data map; S3. Input the initially generated high-resolution meteorological data map, the power output time series data of new energy power stations and the geographical feature data into the spatiotemporal diffusion model again, embed physical constraints in the spatiotemporal diffusion model, and output a high-resolution meteorological estimation map. S4. Based on the high-resolution meteorological estimation map, a two-stage training strategy is adopted to train the spatiotemporal diffusion model and obtain the diffusion generation model. S5. Input the real-time acquired low-resolution meteorological data into the diffusion generation model, output the target high-resolution meteorological data map, and complete the reconstruction.

2. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 1, characterized in that, In step S2, the spatiotemporal diffusion model adopts a U-Net++ network structure and embeds a spatiotemporal attention module; The spatiotemporal attention module calculates the following: In the formula, Q is the query vector, K is the key vector, V is the value vector, and d k Let K be the feature dimension, and T be the transpose of vector K.

3. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 1, characterized in that, In step S3, the physical constraints include energy conservation constraints, photovoltaic power output efficiency constraints, and fluid dynamic constraints.

4. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 3, characterized in that, Energy conservation constraint: Calculating the reconstructed irradiance field G rec Energy balance equation residuals: In the formula, L energy Let G be the energy conservation constraint loss function. rec For the irradiance field, Let be the partial derivative of the irradiance field with respect to time. The irradiance energy velocity vector, It is a divergence operator; Photovoltaic power output efficiency constraint: based on real-time power output P real Reverse-constrained irradiance and temperature field: L PV =‖P real -ηAG rec (1-0.05(T rec -25))‖ 2 ; In the formula, L PV Let P be the photovoltaic power output efficiency constraint loss function. real For real-time power output, A is the effective area of ​​the photovoltaic module, and T is... rec Let η be the temperature field and η be the photovoltaic conversion efficiency. Fluid dynamic constraints: on wind speed field Apply Navier-Stokes regularization: In the formula, L NS The fluid dynamics constraint loss function, This refers to the wind speed field.

5. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 1, characterized in that, In step S4, the two-stage training strategy includes a pre-training stage and a fine-tuning stage. in, The pre-training phase uses historical high-resolution radar data to train the spatiotemporal diffusion model, with the optimization objective being: In the formula, L pre For the pre-training loss function, Let ε be the expected value of the prediction error. θ Let x be the noise generation function. t The sample is a noisy sample, and ε is the actual noise. The fine-tuning phase involves optimization using a joint physical constraint loss function, where the physical constraint loss function is: L total =L diff +aL energy +bL PV +cL NS ; In the formula, L total Let L be the total loss function. diff The difference between the high-resolution meteorological data map and the actual value is represented by a, b, and c, which are dynamically adjustable coefficients.

6. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 1, characterized in that, In step S5, inference processing is performed by combining the accelerated sampling algorithm with the physical correction module, specifically as follows: A denoising diffusion implicit model is used to accelerate the sampling algorithm and compress the number of sampling steps; Perform physical correction after each sampling step: In the formula, z′ represents the data after physical correction. Let L be the gradient value of z. phys α represents the physical constraint loss, and α is the correction step size.

7. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 1, characterized in that, It also includes, S6. Take the target high-resolution meteorological data map as input and perform post-processing operations, including super-resolution reconstruction and spatiotemporal consistency filtering.

8. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 7, characterized in that, In step S6, post-processing includes: A physics-guided super-resolution network generates a 100m resolution field from 1km resolution data. The temporal optical flow method is used to correct the cloud motion trajectory, and the displacement vector between adjacent frames is calculated: Δx=argmin∑‖I t (x)-I t+1 (x+Δx)‖ 2 ; In the formula, Δx is the displacement vector between adjacent frames, I t (x) represents the value at position x of the image in frame t; t+1 (x+Δx) is the value of the (t+1)th frame image at position x+Δx.

9. The method for reconstructing meteorological data from new energy power stations based on a diffusion model according to claim 1, characterized in that, In step S1, the geographic feature data includes: Elevation model data: 30m resolution digital elevation map from Space Shuttle radar topographic mapping; Surface reflectance: calculated based on Landsat 8 satellite data; Obstacle distribution: 0.5m resolution raster map of the three-dimensional outline of buildings surrounding the station.

10. A meteorological data reconstruction system for new energy power stations based on a diffusion model, characterized in that, include: The construction module is used to build a multimodal conditional input set, including low-resolution meteorological data, time-series power output data of new energy power plants, and geographic feature data; The data graph module is used to input the multimodal input set into the designed spatiotemporal diffusion model and output a preliminary high-resolution meteorological data graph. The embedding module is used to input the initially generated high-resolution meteorological data map, the power output time series data of new energy power stations and geographical feature data back into the spatiotemporal diffusion model, embed physical constraints in the spatiotemporal diffusion model, and output a high-resolution meteorological estimation map. The training module is used to train the spatiotemporal diffusion model based on the high-resolution meteorological estimation map using a two-stage training strategy to obtain the diffusion generation model. The output module is used to input real-time acquired low-resolution meteorological data into the diffusion generation model and output a high-resolution meteorological data map of the target to complete the reconstruction.

Images