A power temperature element downscaling method and system based on physical loss guided diffusion model

Abstract

This invention discloses a method and system for downscaling power temperature elements based on a physical loss-guided diffusion model. The method includes: acquiring low-resolution NWP data and micro-topographic data, performing spatial alignment and standardization; inputting the preprocessed data into a pre-trained conditional diffusion downscaling model, outputting a kilometer-level high-resolution temperature field; this model adopts a U-Net architecture conditional diffusion probability model, optimized during the training phase through a joint loss function, the joint loss being a weighted sum of data losses from the generated and observed fields and physical losses calculated based on the heat advection-diffusion control equations; inputting the generated high-resolution temperature field into the downstream power business system, and adjusting the physical loss weight coefficients inversely according to the downstream power prediction error to fine-tune the model online. The system is used to implement the method. This invention ensures thermodynamic consistency while improving resolution and achieves adaptive optimization through a business closed loop.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of power meteorological forecasting and artificial intelligence, specifically to a method and system for downscaling power temperature elements based on a physical loss-guided diffusion model. Background Technology

[0002] Against the backdrop of the macro-strategy of building a new power system, new energy sources, represented by wind power and photovoltaics, have become the main contributors to new installed capacity. The high proportion of new energy power generation connected to the grid naturally exhibits volatility and intermittency, and its output level is highly dependent on meteorological conditions. This makes the safety margin of the entire power system closely coupled with the ability to accurately forecast weather.

[0003] Current mainstream numerical weather prediction (NWP) models and most AI meteorological models typically have a spatial resolution of 10 to 30 kilometers, exhibiting a significant "resolution gap." This makes it difficult to meet the micro-meteorological environment characterization requirements of new power systems for transmission line corridors at the meter level or new energy power stations at the kilometer level. To bridge this gap, existing downscaling techniques face bottlenecks: traditional dynamic downscaling (such as nested WRF) involves enormous computational loads and long processing times, failing to meet the minute-level real-time requirements of power dispatch; while AI statistical downscaling based on computer vision (such as simple super-resolution networks) lacks physical constraints, resulting in refined meteorological fields that may violate fundamental laws of atmospheric dynamics and thermodynamics, making them unsuitable for serious engineering scenarios such as calculating conductor current carrying capacity and assessing icing risk. Summary of the Invention

[0004] The technical problem this invention aims to solve is that it cannot meet the minute-level real-time requirements of power dispatching, and the generated refined meteorological field may violate the basic laws of atmospheric dynamics and thermodynamics, making it unsuitable for serious engineering scenarios such as calculating conductor current carrying capacity and assessing icing risk. To address these technical problems, this invention provides a method and system for downscaling power temperature elements based on a physical loss-guided diffusion model, comprising the following steps: S1. Data Input: Acquire real-time low-resolution numerical weather forecast (NWP) data and 30m resolution micro-topographic data for the target area. The micro-topographic data includes digital elevation model (DEM) data, slope data, and aspect data. S2. Model Inference: The NWP data and micro-topography data are used as conditional inputs to a pre-trained conditional diffusion downscaling model, which outputs a kilometer-level high-resolution power temperature element field. The pre-trained conditional denoising diffusion downscaling model adopts a U-Net architecture conditional denoising diffusion probability model. During the reverse denoising process, the shallow features extracted by the encoder are passed to the decoder through the skip connections of the U-Net to recover the spatial high-frequency details of the temperature field. Its training phase is achieved by calculating the data loss between the generated high-resolution temperature field and the real high-resolution observation data. Based on the two-dimensional heat advection-diffusion control equations, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field through the automatic differentiation mechanism of a deep learning framework. ; through joint loss function Optimize model parameters, where Dynamic hyperparameters to balance physical constraints and data fitting weights; S3. Business Application and Closed-Loop Optimization: Input the high-resolution power temperature element field into the downstream power business system; obtain the new energy power prediction error fed back by the downstream power business system; adjust the weight coefficient of physical loss in the joint loss function in reverse based on the power prediction error; and fine-tune or re-optimize the pre-trained conditional diffusion downscaling model online using recent historical data based on the adjusted weight coefficient.

[0005] Further, in step S2, the expression for the two-dimensional heat advection-diffusion control equation is: ; in, The temperature field generated for the downscaling model. The background wind speed component is the input, derived from NWP data. Where is the thermal diffusivity, For system source and sink items, the The horizontal zonal and meridional wind speed components are derived from low-resolution NWP data. Spatially upsampled using bilinear interpolation or spline interpolation algorithms, they are spatially aligned with the high-resolution target grid output by the downscaling network to characterize the advection transport effect of heat. The thermal diffusivity is set as a learnable tensor parameter in the downscaling network during model training. Physical inversion and dynamic updates are performed using a gradient backpropagation mechanism based on data loss and physical loss to adaptively characterize the heat dissipation capability of local atmospheric turbulence; As a system source and sink term, its construction method is to couple the micro-topographic data with the surface radiation heat balance model, calculate the surface net radiation affected by the elevation, slope and aspect of the micro-topography, specifically including the solar shortwave radiation heat source absorbed by the surface, the longwave radiation heat sink emitted by the surface, and the related local sensible heat and latent heat flux, which are used to characterize the non-advectional and non-diffusion local heat forcing under complex terrain.

[0006] Furthermore, in step S2, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field through the automatic differentiation mechanism of the deep learning framework. The specific method is as follows: The generated temperature field is calculated using the automatic differentiation module of a deep learning framework. Relative to time t and spatial coordinates The partial derivatives of the equation are substituted into the governing equation to obtain the residual vector, and the mean square value of the residual vector is calculated. The formula for calculating the residual vector is as follows: .

[0007] Furthermore, the training phase of the pre-trained conditional diffusion downscaling model also includes: A conditional denoising diffusion probability model using the U-Net architecture is employed. During the reverse denoising process, shallow features extracted by the encoder are passed to the decoder through skip connections to restore the spatial high-frequency details of the temperature field. The acquired multi-source meteorological data and micro-topographic data are spatially aligned and standardized before being input into the model for training.

[0008] Furthermore, the downstream power business system mentioned in step S3 includes at least one of the following: a new energy power prediction system, a transmission line dynamic capacity expansion system, and a line icing early warning system. The high-resolution power temperature field is used to support at least one of the following power services: new energy power generation forecasting, transmission line current carrying capacity calculation, and line icing risk assessment, meeting the micro-meteorological forecasting needs at the power equipment level.

[0009] Furthermore, the specific rule for adjusting the weighting coefficient λ based on the power prediction error in step S3 is as follows: The root mean square error of power prediction output from the downstream renewable energy power prediction system is used as a feedback signal. When the power prediction deviation exceeds a preset threshold under specific meteorological modes such as cold waves, high temperatures, or severe convection, the value of λ is automatically increased to strengthen the guiding role of physical constraints in model generation. The "specific meteorological modes" specifically refer to long-tailed or extreme weather scenarios that cause drastic fluctuations in renewable energy power generation and have scarce historical samples, including: Cold wave and drastic temperature drop pattern: With the rapid southward movement of cold air, the wind speed gradient and wind direction change drastically (causing a "step" change or cut-off in wind power), which may also cause micro-meteorological icing on transmission lines or wind turbine blades.

[0010] Sustained high temperatures and heatwave pattern: Under the influence of the subtropical high pressure, accompanied by intense solar radiation and stable weather with low wind speeds, extreme environmental high temperatures can cause a non-linear decline in the photoelectric conversion efficiency of photovoltaic modules.

[0011] Severe convection and short-duration sudden weather patterns: such as squall lines, thunderstorms, or localized heavy precipitation. These types of weather are highly localized in time and space, leading to instantaneous blocking of shortwave radiation (sudden drop in photovoltaic power) and a dramatic increase in near-surface wind shear.

[0012] Meteorological transition model for complex micro-topography: targeting complex mountainous terrain, including sudden disruption of the inversion layer, dissipation of the valley cold lake effect, or nonlinear wind speed surge caused by the mountain funneling effect.

[0013] Furthermore, the specific method for online fine-tuning or re-optimization in step S3 is as follows: The adjusted λ value was selected, and the fine-tuning sample set was constructed using the historical low-resolution NWP data, micro-topographic data and corresponding ground observation temperature data of the most recent 7 days. The pre-trained model parameters were used as the initial weights, and the parameters were iteratively updated for 10 to 20 epochs based on the adjusted joint loss function to obtain the optimized model adapted to the specific meteorological scenario.

[0014] Furthermore, the pre-trained conditional diffusion downscaling model described in step S2 transmits the shallow features of micro-topography edges and slope changes extracted by the encoder through the skip connections of U-Net, and outputs a high-resolution power and temperature element field with a resolution of 1 kilometer or 100 meters to meet the needs of capturing micro-meteorological features under complex terrain.

[0015] Furthermore, the micro-topographic data obtained in step S1 also includes 30m resolution digital elevation model (DEM) data, and the resolution of the NWP data is on the order of 10km to 30km.

[0016] This invention also provides a power temperature factor downscaling system based on a physical loss-guided diffusion model, comprising: The inference data input module is used to acquire real-time low-resolution numerical weather forecast (NWP) data and 30m resolution micro-topographic data of the target area. The micro-topographic data includes digital elevation model (DEM) data, slope data, and aspect data. The downscaling inference module takes the NWP data and micro-topography data as input conditions and feeds them into a pre-trained conditional diffusion downscaling model, outputting a kilometer-level high-resolution power temperature element field. The pre-trained conditional diffusion downscaling model uses U-Net as its backbone network, and its training phase is achieved by calculating the data loss between the generated high-resolution temperature field and the actual high-resolution observation data. Based on the two-dimensional heat advection-diffusion control equations, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field through the automatic differentiation mechanism of a deep learning framework. ; through joint loss function Optimize model parameters, where Dynamic hyperparameters to balance physical constraints and data fitting weights; The closed-loop optimization module is used to input the high-resolution power temperature element field into the downstream power business system; obtain the new energy power prediction error fed back by the downstream power business system; adjust the weight coefficient of physical loss in the joint loss function based on the power prediction error; and fine-tune or re-optimize the pre-trained conditional diffusion downscaling model online using recent historical data based on the adjusted weight coefficient.

[0017] Compared with existing technologies, the advantages of this invention are: it achieves a deep integration of physical mechanisms and artificial intelligence, and solves the problem of physical violations caused by purely data-driven AI models. By introducing a physical loss function, the model is forced to learn a mapping relationship that conforms to the laws of energy conservation and atmospheric thermodynamics, significantly improving the reliability of the model under complex terrain and extreme weather conditions.

[0018] It breaks through the resolution bottleneck of weather forecasts in power applications: It can accurately downscale weather forecasts from 10 to 30 kilometers to the 1-kilometer or even 100-meter level, and can accurately capture micro-meteorological features such as temperature inversion layers and mountain funneling effects, meeting the high-resolution requirements of equipment.

[0019] A closed-loop optimization mechanism for meteorological and power business has been established: the output downscaled high-resolution physical field is directly connected to power business such as new energy power prediction, and the error feedback of downstream business drives the dynamic optimization of the underlying meteorological downscaled model, creating direct economic benefits. Attached Figure Description

[0020] Figure 1 The present invention provides an overall flowchart of a method for downscaling electrical temperature elements based on a physical loss-guided diffusion model.

[0021] Figure 2 The flowchart illustrates a method for downscaling electrical temperature elements based on a physical loss-guided diffusion model, as provided in this embodiment of the invention.

[0022] Figure 3 A schematic diagram of the conditional diffusion model architecture of the fused physical information neural network (PINNs) provided in an embodiment of the present invention.

[0023] Figure 4 A schematic diagram of the "meteorological-electricity" dual verification and closed-loop optimization process provided in the embodiments of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0025] The overall method flowchart of the present invention is as follows: Figure 1 As shown, it includes: S1. Data Input: Acquire real-time low-resolution numerical weather forecast (NWP) data and 30m resolution micro-topographic data for the target area. The micro-topographic data includes digital elevation model (DEM) data, slope data, and aspect data. S2. Model Inference: The NWP data and micro-topography data are used as conditional inputs to a pre-trained conditional diffusion downscaling model, which outputs a kilometer-level high-resolution power temperature element field. The pre-trained conditional denoising diffusion downscaling model adopts a U-Net architecture conditional denoising diffusion probability model. During the reverse denoising process, the shallow features extracted by the encoder are passed to the decoder through the skip connections of the U-Net to recover the spatial high-frequency details of the temperature field. Its training phase is achieved by calculating the data loss between the generated high-resolution temperature field and the real high-resolution observation data. Based on the two-dimensional heat advection-diffusion control equations, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field through the automatic differentiation mechanism of a deep learning framework. ; through joint loss function Optimize model parameters, where Dynamic hyperparameters to balance physical constraints and data fitting weights; S3. Business Application and Closed-Loop Optimization: Input the high-resolution power temperature element field into the downstream power business system; obtain the new energy power prediction error fed back by the downstream power business system; adjust the weight coefficient of physical loss in the joint loss function in reverse based on the power prediction error; and fine-tune or re-optimize the pre-trained conditional diffusion downscaling model online using recent historical data based on the adjusted weight coefficient.

[0026] The technical solution adopted in this embodiment is as follows: Figure 2 As shown, it includes the following steps: Step S1: Acquire multi-source heterogeneous data. Collect numerical weather forecasts with a resolution of 10km-30km (such as the ERA5 reanalysis dataset) as background field input, and simultaneously acquire micro-topographic data such as 30m resolution elevation (DEM) and slope of the area where the power equipment is located, and perform spatial alignment and standardization processing; Step S2: Construct a downscaling network. Build a conditional denoising diffusion probability model (DDPM) based on the U-Net architecture. Input the low-resolution features obtained in Step S1 as conditional variables into the model to guide the reverse generation process of recovering high-resolution images from random Gaussian noise; Step S3: Calculate the joint loss function. This is a core innovative step. It involves calculating the mean squared error between the generated high-resolution image and the label. Using the automatic differentiation engine built into the deep learning framework, the physical residuals are calculated by substituting the spatiotemporal partial derivatives of the generated temperature field into the heat advection-diffusion partial differential equation. The two are then weighted and merged. Step S4: Model optimization training. Guided by the joint loss function, the network weights are updated using optimizers such as AdamW. This process forces the network not only to "mimic" the pixel distribution of high-resolution data, but also to "obey" the laws of thermodynamics. Step S5: Business Application. Deploy the trained model in the power dispatching system to downscale the incoming coarse-grid meteorological data in real time, and push the results to systems such as wind and solar power prediction or dynamic line capacity expansion.

[0027] like Figure 3 As shown, the core network architecture of this invention integrates the diffusion process with the constraints of physical information neural networks (PINNs).

[0028] In step t of the reverse denoising process, the U-Net network outputs the predicted high-resolution temperature field. At this point, an Auto-Diff Module is introduced. Since all operators in the network are differentiable, the system directly performs time and spatial denoising on the network output. Differentiation of coordinates.

[0029] Let the wind field provided by NWP be The thermal diffusivity is The constructed physical residual is:

[0030] Joint loss during final backpropagation The convolution kernel parameters in U-Net are updated by being jointly computed by the data loss network endpoints and the physical residual network endpoints.

[0031] like Figure 4As shown, the high-resolution power temperature data generated by downscaling not only undergoes simple meteorological index verification (such as RMSE and anomaly correlation coefficient ACC), but is also directly input into the downstream renewable energy forecasting system. The downstream renewable energy power forecast results are compared with the actual grid-connected power. If the power forecast deviation is large under a specific meteorological pattern (such as a cold wave), the system will automatically adjust the forecast for that scenario. Dynamic hyperparameters This guides the adaptive retraining of the next round of diffusion model.

[0032] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0033] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for downscaling electrical temperature elements based on a physical loss-guided diffusion model, characterized in that, include: S1. Data Input: Acquire real-time low-resolution numerical weather forecast (NWP) data and 30m resolution micro-topographic data for the target area. The micro-topographic data includes digital elevation model (DEM) data, slope data, and aspect data. S2. Model Inference: The NWP data and micro-topography data are used as conditional inputs and fed into a pre-trained conditional diffusion downscaling model to output a high-resolution electric temperature element field at the kilometer level. The pre-trained conditional denoising diffusion downscaling model adopts a conditional denoising diffusion probability model with a U-Net architecture. In the reverse denoising process, the shallow features extracted by the encoder are passed to the decoder through the skip connections of U-Net to restore the spatial high-frequency details of the temperature field. The training phase is achieved by calculating the data loss between the generated high-resolution temperature field and the actual high-resolution observation data. ; Based on the two-dimensional heat advection-diffusion control equation, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field through the automatic differentiation mechanism of a deep learning framework. ; through joint loss function Optimize model parameters, where Dynamic hyperparameters to balance physical constraints and data fitting weights; S3. Business Application and Closed-Loop Optimization: Input the high-resolution power temperature element field into the downstream power business system; obtain the new energy power prediction error fed back by the downstream power business system; adjust the weight coefficient of physical loss in the joint loss function in reverse based on the power prediction error; and fine-tune or re-optimize the pre-trained conditional diffusion downscaling model online using recent historical data based on the adjusted weight coefficient.

2. The method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 1, characterized in that, In step S2, the expression for the two-dimensional heat advection-diffusion control equation is: ； in, The temperature field generated for the downscaling model. The background wind speed component is the input, derived from NWP data. Where is the thermal diffusivity, For system source and sink items.

3. The method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 2, characterized in that, In step S2, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field using the automatic differentiation mechanism of the deep learning framework. The specific method is as follows: The generated temperature field is calculated using the automatic differentiation module of a deep learning framework. Relative to time t and spatial coordinates Substituting the partial derivatives into the governing equations yields the residual vector, and the mean square value of the residual vector is calculated. The formula for calculating the residual vector is as follows: 。 4. The method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 1, characterized in that, The training phase of the pre-trained conditional diffusion downscaling model also includes: A conditional denoising diffusion probability model using the U-Net architecture is employed. During the reverse denoising process, shallow features extracted by the encoder are passed to the decoder through skip connections to restore the spatial high-frequency details of the temperature field. The acquired multi-source meteorological data and micro-topographic data are spatially aligned and standardized before being input into the model for training.

5. The method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 1, characterized in that, The downstream power business system mentioned in step S3 includes at least one of the following: a new energy power prediction system, a transmission line dynamic capacity expansion system, and a line icing early warning system. The high-resolution power temperature field is used to support at least one of the following power services: new energy power generation forecasting, transmission line current carrying capacity calculation, and line icing risk assessment, meeting the micro-meteorological forecasting needs at the power equipment level.

6. The method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 5, characterized in that, The specific rule for adjusting the weighting coefficient λ based on the power prediction error in step S3 is as follows: The root mean square error of power prediction output by the downstream new energy power prediction system is used as a feedback signal; when the power prediction deviation exceeds the preset threshold under specific meteorological modes such as cold wave, high temperature or strong convection, the value of λ is automatically increased to strengthen the guiding role of physical constraints on model generation.

7. A method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 6, characterized in that, The specific method for online fine-tuning or re-optimization in step S3 is as follows: The adjusted λ value was selected, and the fine-tuning sample set was constructed using the historical low-resolution NWP data, micro-topographic data and corresponding ground observation temperature data of the most recent 7 days. The pre-trained model parameters were used as the initial weights, and the parameters were iteratively updated for 10 to 20 epochs based on the adjusted joint loss function to obtain the optimized model adapted to the specific meteorological scenario.

8. The method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 1, characterized in that, The pre-trained conditional diffusion downscaling model described in step S2 transmits the shallow features of micro-topography edges and slope changes extracted by the encoder through the skip connections of U-Net, and outputs a high-resolution power and temperature element field with a resolution of 1 kilometer or 100 meters to meet the needs of capturing micro-meteorological features under complex terrain.

9. A method for downscaling electrical temperature elements based on a physical loss-guided diffusion model according to claim 1, characterized in that, The micro-topographic data obtained in step S1 also includes 30m resolution digital elevation model (DEM) data, and the resolution of the NWP data is on the order of 10km to 30km.

10. A power temperature element downscaling system based on a physical loss-guided diffusion model, characterized in that, include: The inference data input module is used to acquire real-time low-resolution numerical weather forecast (NWP) data and 30m resolution micro-topographic data of the target area. The micro-topographic data includes digital elevation model (DEM) data, slope data, and aspect data. The downscaling inference module takes the NWP data and micro-topography data as input conditions and feeds them into a pre-trained conditional diffusion downscaling model, outputting a kilometer-level high-resolution power temperature element field. The pre-trained conditional diffusion downscaling model uses U-Net as its backbone network, and its training phase is achieved by calculating the data loss between the generated high-resolution temperature field and the actual high-resolution observation data. Based on the two-dimensional heat advection-diffusion control equations, the physical loss is obtained by calculating the physical equation residuals of the generated temperature field through the automatic differentiation mechanism of a deep learning framework. ; through joint loss function Optimize model parameters, where Dynamic hyperparameters to balance physical constraints and data fitting weights; The closed-loop optimization module is used to input the high-resolution power temperature element field into the downstream power business system; obtain the new energy power prediction error fed back by the downstream power business system; adjust the weight coefficient of physical loss in the joint loss function based on the power prediction error; and fine-tune or re-optimize the pre-trained conditional diffusion downscaling model online using recent historical data based on the adjusted weight coefficient.

Images