AI new energy meteorological large model system

Abstract

The invention relates to the technical field of data processing, in particular to an AI new energy meteorological large model system which comprises a multi-source data integration module, a data governance and quality control module, an intelligent modeling and training module, a business collaboration and decision closed-loop module and an integrated service support module. The multi-source data integration module is used for acquiring a multi-dimensional heterogeneous data source; the data governance and quality control module processes the multivariate heterogeneous data and constructs a meteorological energy fusion data cube; the intelligent modeling and training module is used for constructing a new energy scene-oriented meteorological prediction large model based on a deep learning architecture fusing VDT and TX-Diffusion mechanisms; and the service collaboration and decision closed-loop module deeply fuses a weather forecast result output by the weather forecast large model with a service rule, and generates a service operation suggestion or instruction through multi-objective optimization. According to the method, high-sensitivity perception of the meteorological non-stationary process and end-to-end conversion from a prediction result to an executable scheduling instruction are realized, and the new energy meteorological prediction precision, physical consistency and service availability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an AI-powered new energy meteorological large-scale model system. Background Technology

[0002] The New Energy Meteorological Big Data Model System is an intelligent system that provides weather forecasting for the new energy industry. It addresses the high dependence of the new energy sector (such as wind power and photovoltaics) on meteorological data by integrating meteorological science, big data analysis, and domain expertise.

[0003] Defects and shortcomings of existing technology:

[0004] 1) Traditional machine learning models (such as support vector machines) have limited ability to handle multi-source, heterogeneous, and high-dimensional data (satellite, radar, observation station, and power plant operation data), lack semantic understanding, and rely on expert experience for feature engineering. These models are often static, exhibiting a sharp drop in generalization performance in the face of new data sites (scarce data), new climate models, or extreme weather events (few historical samples), requiring frequent manual retraining and parameter tuning. The model architecture lacks sufficient representational learning capabilities, lacks a deep understanding of atmospheric physical laws and spatiotemporal correlations, and does not possess the ability to continuously learn autonomously and adapt to changes, leading to biased prediction results, slow inference speed, and difficulty in meeting the comprehensive meteorological intelligence needs of new energy sources.

[0005] 2) The system is isolated, and forecasts are disconnected from business decisions, failing to create a closed-loop value chain. The meteorological forecasting system has low integration with downstream business systems such as power grid coordination, forming data silos. Forecast results are only used as raw data outputs, failing to deeply integrate business knowledge and directly transform into actionable optimization decision suggestions (such as dispatch instructions and risk warnings), resulting in a break in the overall system value chain. Summary of the Invention

[0006] This invention provides an AI-powered new energy meteorological big data model system that achieves highly sensitive knowledge of non-stationary meteorological processes and end-to-end conversion of prediction results into executable scheduling instructions, significantly improving the accuracy, physical consistency, and operational availability of new energy meteorological forecasts.

[0007] To achieve the objectives of this invention, the technical solution adopted is: an AI-powered new energy meteorological big data model system, comprising a multi-source data integration module, a data governance and quality control module, an intelligent modeling and training module, a business collaboration and decision-making closed-loop module, and an integrated service support module;

[0008] Multi-source data integration module: Acquires multi-dimensional heterogeneous data sources including: meteorological data, energy data, and geographic auxiliary data;

[0009] Data governance and quality control module: processes diverse and heterogeneous data to construct a meteorological and energy fusion data cube;

[0010] Intelligent modeling and training module: Based on a deep learning architecture that integrates VDT and TX-Diffusion mechanisms, a large-scale meteorological forecasting model for new energy scenarios is constructed.

[0011] Business collaboration and decision-making closed-loop module: Deeply integrates the weather forecast results output by the large weather forecast model with business rules, and generates business operation suggestions or instructions through multi-objective optimization;

[0012] Integrated service support module: Opens up the capabilities of large-scale meteorological forecasting models to the public via API interfaces.

[0013] As an optimized solution of the present invention, the meteorological data includes meteorological system observation data, station meteorological observation data and background field data; the energy data includes new energy station SCADA data, power grid dispatching instructions and power grid topology and constraint information; the geographic auxiliary data includes new energy station coordinates and terrain elevation data.

[0014] As an optimized solution of the present invention, the processing of multi-source heterogeneous data includes data standardization, temporal / spatial dimension alignment, outlier removal and missing value filling, and the final fusion is achieved through multi-source feature fusion and structured organization.

[0015] As an optimized solution of the present invention, the intelligent modeling and training module divides the meteorological and energy fusion data cube into patches containing multiple time steps and local spatial regions through spatiotemporal joint cube embedding, projects them into high-dimensional embedding sequences, and adopts a U-Net-like backbone network with a bottleneck layer composed of VDT. By modeling the differential signals of adjacent time steps, the perception of meteorological dynamic processes is enhanced.

[0016] The TX-Diffusion Model is introduced as an end-to-end inference framework: after the backbone network generates the initial prediction, multi-step iterative denoising is performed through chained Earth Transformer Blocks and DConv, supporting dual modes of one-step cascaded prediction and multi-step fine-tuning prediction.

[0017] The entire TX-Diffusion chain is trained jointly with the backbone network. The loss function is applied to the denoised output at each step, and a custom composite loss function is adopted, which includes physical constraint terms, business-sensitive weighting terms, and extreme event enhancement terms.

[0018] A two-stage transfer learning strategy is adopted: first, a general Earth system dynamic model is pre-trained on global reanalysis data; then, small-sample fine-tuning is performed on regional or field-level fusion data to generate a large-scale meteorological forecasting model.

[0019] As an optimized solution of the present invention, Cube Embedding encodes the meteorological and energy fusion data cube into a high-dimensional spatiotemporal sequence, specifically including: inputting the meteorological and energy fusion data cube, obtaining the total number of patches through spatiotemporal segmentation, flattening each patch, projecting all patches together using a learnable weight matrix, and embedding them; assigning a unique spatiotemporal index to each patch, using learnable three-dimensional position encoding, and jointly encoding the time sequence and spatial coordinates to obtain a high-dimensional spatiotemporal sequence.

[0020] As an optimization scheme of the present invention, the U-Net-like backbone network feature extraction adopts an encoder-decoder structure. The encoder contains several levels of Down Blocks, which compress the spatial dimension and increase the channel dimension layer by layer, and retain the features of each level for skip connections. The bottleneck layer adopts VDT, which explicitly models the differential signal of adjacent time steps on the deepest high semantic features. The decoder upsamples layer by layer through Up Blocks, and fuses the skip connection features of the corresponding encoder level, and combines deep convolution DConv to restore spatial details and optimize boundaries, and finally outputs the initial prediction result.

[0021] As an optimized solution of the present invention, the two-stage transfer learning strategy specifically includes:

[0022] 1) Construct a global pre-trained dataset, define a multi-step autoregressive prediction task, adopt a basic loss function that includes physical constraints, and perform end-to-end pre-training on the backbone architecture that integrates VDT and TX-Diffusion to obtain a general Earth system dynamic model.

[0023] 2) Construct a local fine-tuning dataset, employ efficient parameter fine-tuning, freeze non-critical layers in Cube Embedding, Down / Up Blocks, and TX-Diffusion, and fine-tune only the attention projection matrix, the last 1-2 EarthTransformer Blocks, and the output header in VDT; introduce LoRA to avoid catastrophic forgetting and overfitting, and use a composite loss function that includes physical constraint terms, business-sensitive weighting terms, and extreme event enhancement terms in the fine-tuning process, and use the validation value of the composite loss function itself as the early stopping criterion, finally generating a large-scale weather forecasting model.

[0024] As an optimized solution of the present invention, the business collaboration and decision-making closed-loop module receives the forecast results output by the meteorological forecasting large model, performs spatiotemporal alignment and unit standardization, generates structured meteorological time series data objects for business nodes, loads a pre-configured business rule template library, and each rule includes triggering conditions, hard constraints, optimization objectives, and action templates; for each business node and time period, it dynamically matches all activated rules and constructs a decision variable space accordingly; it extracts multi-objective functions, assigns weights to the soft objectives corresponding to each rule according to the business strategy, and uses a weighted sum method to fuse them into a single objective function; it integrates all hard constraints to form a constrained optimization problem; it calls a lightweight solver to obtain the optimal decision and generates structured business instructions.

[0025] As an optimization of this invention, the trained large-scale weather forecasting model is exported as an ONNX or TorchScript inference optimization format and encapsulated into a standardized inference function that includes preprocessing, model inference, and postprocessing; a RESTful interface is provided, supporting synchronous / asynchronous calls, with clearly defined request parameters and response structures, and output supporting NetCDF, GeoJSON, and CSV formats.

[0026] The present invention has positive effects: 1) The present invention accurately captures the dynamic evolution of meteorological processes, introduces differential signal modeling through VDT, explicitly learns the rate of change of variables (such as wind speed acceleration) in adjacent time steps, focuses on the dynamic change area through the attention mechanism, and significantly improves the sensitivity to non-stationary weather processes (thunderstorms). Compared with the traditional method of modeling only with original values, the enhanced dynamic features reduce the short-term (0-6h) prediction error by 39.6%.

[0027] 2) This invention takes into account both global semantics and local details. The encoder extracts large-scale weather system features (such as the westerly wind belt) step by step; the decoder restores micro-scale effects such as local terrain and shading through skip connections; DConv (dynamic convolution) optimizes the boundary in upsampling, effectively restoring the fine irradiance distribution of the wake area of ​​wind farms and the edge of photovoltaic power stations, so as to restore the fine physical phenomena that are crucial to the meteorological forecast of new energy with high fidelity, and significantly reduce the prediction bias caused by spatial smoothing or boundary blurring.

[0028] 1) This invention introduces VDT into the bottleneck layer of the backbone network to explicitly model the first-order difference signal of adjacent time steps to capture non-stationary processes; at the same time, it embeds a constraint loss term in the training to ensure that the prediction results are both accurate and in line with the laws of atmospheric physics.

[0029] 4) This invention automatically transforms forecast results into structured meteorological sequences for specific stations through a business collaboration and decision-making closed-loop module, dynamically matches preset business rules (including triggering conditions, hard constraints, and optimization objectives), constructs and solves multi-objective optimization problems, generates compliant and executable scheduling instructions, seamlessly connects to the business network, ends the data silo dilemma of forecasting being the end point, and forms a complete value chain. Attached Figure Description

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0031] Figure 1 This is a schematic diagram of the principle of the present invention. Detailed Implementation

[0032] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 This invention discloses an AI-powered new energy meteorological big data model system, which includes a multi-source data integration module, a data governance and quality control module, an intelligent modeling and training module, a business collaboration and decision-making closed-loop module, and an integrated service support module.

[0034] Multi-source data integration module: Acquires multi-dimensional heterogeneous data sources including: meteorological data, energy data, and geographic auxiliary data;

[0035] Meteorological data includes meteorological system observation data, station meteorological observation data, and background field data; among which, meteorological system observation data includes surface meteorological observation data, upper-air sounding data, meteorological radar data, meteorological satellite remote sensing data, and marine meteorological observation data.

[0036] Meteorological observation data from wind farms and photovoltaic power plants are collected. Wind farms acquire parameters such as wind speed, wind direction, air temperature, atmospheric pressure, relative humidity, precipitation, and visibility using meteorological towers or lidar / sonic radar (LiDAR) / sonic radar (SODAR). Photovoltaic power plants collect data including total horizontal irradiance, out-of-plane irradiance, diffuse irradiance, ambient temperature, module temperature, wind speed, wind direction, precipitation, and visibility.

[0037] Background field data in meteorology typically refers to reference data describing the basic state of the atmosphere or ocean, used for analyzing anomalies or trends. This includes climate mean data and global gridded data, with global gridded data including sea surface temperature and land temperature.

[0038] Energy data includes SCADA data from new energy power plants, power grid dispatch instructions, and power grid topology and constraint information; geographic auxiliary data includes coordinates of new energy power plants and terrain elevation data.

[0039] Data governance and quality control module: Processes diverse and heterogeneous data to construct a meteorological and energy fusion data cube; the processing of diverse and heterogeneous data includes data standardization, temporal / spatial dimension alignment, outlier removal and missing value filling, and finally achieves fusion through multi-source feature fusion and structured organization.

[0040] 1. Data standardization processing: unify physical units (e.g., wind speed converted to m / s); standardize variable naming (following IEC standards); map enumerated values ​​(e.g., equipment status, weather phenomenon codes) to a unified dictionary.

[0041] 2. Spatiotemporal alignment includes time alignment and spatial alignment.

[0042] Time alignment: Using a 15-minute granularity, high-frequency data is downsampled and low-frequency data is interpolated;

[0043] Spatial alignment: Grid data is matched to the coordinates of business entities (such as stations) through bilinear interpolation or nearest neighbor matching, and the WGS84 coordinate system is used uniformly.

[0044] 3. Outlier detection and removal: Based on physical boundaries (e.g., wind speed > 80 m / s), statistical rules (Z-score > 3), and temporal consistency (mutation detection), anomalies are identified, invalid values ​​are automatically marked or removed, and quality labels are retained for traceability.

[0045] 4. Intelligent imputation of missing values, with confidence scores attached to the imputation results.

[0046] Short-term missing data: use linear interpolation or moving average;

[0047] Long-term or systemic missing: Utilize multi-source coordination (e.g., use background field wind speed to correct for site missing).

[0048] The data governance and quality control module endows data with semantic, spatiotemporal, and high-value attributes, supporting business upgrades from passive response to proactive anticipation.

[0049] The intelligent modeling and training module, based on a deep learning architecture integrating VDT and TX-Diffusion mechanisms, constructs a large-scale meteorological forecasting model for new energy scenarios. This module uses spatiotemporal joint cube embedding to divide the meteorological and energy fusion data cube into patches containing multiple time steps and local spatial regions, projecting them into a high-dimensional embedding sequence. This achieves end-to-end multi-source fusion, improving feature representation capabilities by an order of magnitude and completely eliminating reliance on expert experience.

[0050] A U-Net-like backbone network is adopted, with the bottleneck layer consisting of VDT. The perception of meteorological dynamics is enhanced by modeling the differential signals of adjacent time steps. Cube Embedding encodes the meteorological and energy fusion data cube into a high-dimensional spatiotemporal sequence. Specifically, it includes: inputting the meteorological and energy fusion data cube, obtaining the total number of patches through spatiotemporal segmentation, flattening each patch, projecting all patches using a learnable weight matrix, and embedding them together; assigning a unique spatiotemporal index to each patch, using learnable 3D position encoding, and jointly encoding the temporal order and spatial coordinates to obtain the high-dimensional spatiotemporal sequence.

[0051] The U-Net-like backbone network feature extraction adopts an encoder-decoder structure. The encoder contains several levels of DownBlock, which compresses the spatial dimension and increases the channel dimension layer by layer, and retains the features of each level for skip connections. The bottleneck layer adopts VDT, which explicitly models the differential signal of adjacent time steps on the deepest high semantic features. The decoder upsamples layer by layer through UpBlock and fuses the skip connection features of the corresponding encoder level. It combines deep convolution DConv to restore spatial details and optimize boundaries, and finally outputs the initial prediction result.

[0052] Step 1: Input data—meteorological and energy fusion data cube;

[0053] Input tensor: X∈R T×H×W×C Where T is the time step, C is the fusion channel (from the data governance module, such as temperature), H is the number of grid points in the latitude direction, and W is the number of grid points in the longitude direction.

[0054] Step 2: Spatiotemporal Joint Cube Embedding; Set the patch size (the basic unit for local segmentation), perform non-overlapping segmentation (including temporal and spatial segmentation) on the input cube, and obtain the total number of patches. Flatten each patch, project it using the learnable weight matrix We, and concatenate all patches for embedding and null position encoding injection.

[0055] Each patch is assigned a unique spatiotemporal index, using a learnable 3D positional encoding E. pos By jointly encoding the temporal order and spatial coordinates, the final sequence Z is obtained. in =Z+E pos Among them, Z in Z is the final input sequence (used as input to the encoder), and Z is the sequence of patch embedding vectors.

[0056] Step 3: Feature extraction from the U-Net-like backbone network;

[0057] The encoder uses multi-stage downsampling, with a total of 3 downblock stages. Each stage contains:

[0058] 3D Depthwise Conv: Capture local spatiotemporal patterns;

[0059] LayerNorm+GeLU: Normalization and Nonlinear Activation;

[0060] Patch Merging: Merges adjacent 2×2×1 patches, halves the spatial resolution, and doubles the channel dimension;

[0061] The outputs at each level are saved as jump connection features: Level 1: F1, Level 2: F2, Level 3: F3.

[0062] Bottleneck layer - Vision Differential Transformer (VDT)

[0063] Input: Deepest feature F3;

[0064] Differential signal construction: Calculate the first-order difference along the time dimension (implied in the patch index), and concatenate the original features with the differential features to obtain the differential enhancement feature F. diff Then, VDT attention calculation is performed, and the global representation F after VDT attention enhancement is output. vdt The decoder involves multi-stage upsampling and feature fusion, and the final backbone network outputs Y. backbon .

[0065] Specifically, Cube Embedding achieves end-to-end multi-source fusion encoding, directly using the fusion cube (containing meteorological and energy variables) from the data governance module as input, eliminating the need for manual feature design. Through unified patch segmentation, physically aligned data from different sources are naturally embedded into the same semantic space. Learnable 3D location encoding automatically captures the joint semantics of when, where, and what variables, replacing the need for manually constructed spatiotemporal lag features. This eliminates expert dependence and achieves fully automated mapping from raw data to high-dimensional representations.

[0066] Step 4: Output to the TX-Diffusion inference framework;

[0067] The backbone network output serves as the initial noise estimate or conditional guide for TX-Diffusion;

[0068] Subsequently, multi-step denoising and fine-tuning are performed through the Earth Transformer Block chain to generate the final high-precision prediction.

[0069] The TX-Diffusion Model is introduced as an end-to-end inference framework: after the backbone network generates the initial prediction, multi-step iterative denoising is performed through chained Earth Transformer Blocks and DConv, supporting dual modes of one-step cascaded prediction and multi-step fine-tuning prediction.

[0070] The entire TX-Diffusion chain is jointly trained with the backbone network (consisting of K concatenated Earth Transformer Blocks (ETBs), each integrating a dynamic convolutional (DConv) module to perform multi-step denoising). The loss function is applied to the denoised output at each step, employing a custom composite loss function that includes physical constraint terms, business-sensitive weighting terms, and extreme event enhancement terms.

[0071]

[0072] Where: k = 0 corresponds to the main output Y backbon k (diffusion step number) ≥ 1 corresponds to the output of the diffusion step, with weight λ. k It can be set to increment, with a loss at each step. It consists of three items:

[0073]

[0074] in: The basic regression term is α, where α is the weight of the physical constraint term. For physical constraints, β is the business-sensitive weighting factor. γ represents the weight of the business-sensitive weighting item, and γ represents the weight of the extreme event enhancement item. Enhancements for extreme events.

[0075] The one-step cascaded mode directly outputs the initial prediction of the backbone network, skipping all denoising steps; the multi-step fine-tuning mode only enables K-step Earth Transformer Block denoising when high accuracy is required (such as day-ahead planning or extreme weather). In low-latency scenarios such as real-time scheduling and AGC control, the one-step mode can keep the inference latency within 500ms; it avoids the inherent latency of traditional diffusion models that must perform dozens of denoising steps, enabling fine-grained inference on demand and significantly improving data processing speed.

[0076] A two-stage transfer learning strategy is adopted: first, a general Earth system dynamic model is pre-trained on global reanalysis data; then, small-sample fine-tuning is performed on regional or field-level fused data to generate a large-scale meteorological forecasting model. The two-stage transfer learning strategy specifically includes:

[0077] 1) Construct a global pre-trained dataset, define a multi-step autoregressive prediction task, adopt a basic loss function that includes physical constraints, and perform end-to-end pre-training on the backbone architecture that integrates VDT and TX-Diffusion to obtain a general Earth system dynamic model.

[0078] 2) A local fine-tuning dataset was constructed, employing efficient parameter fine-tuning. Non-critical layers in Cube Embedding, Down / Up Blocks, and TX-Diffusion were frozen, with only the attention projection matrix, the last 1-2 EarthTransformer Blocks, and the output header in the VDT being fine-tuned. LoRA was introduced to avoid catastrophic forgetting and overfitting. The fine-tuning process used a composite loss function including physical constraint terms, business-sensitive weighting terms, and extreme event enhancement terms. The validation value of the composite loss function itself was used as the early stopping criterion, ultimately generating a large-scale weather forecasting model. When TX-Diffusion detected a sudden increase in prediction residuals, it automatically switched to a "multi-step fine-tuning" mode for real-time correction.

[0079] The composite loss early stopping mechanism ensures stability: using the composite loss of the validation set as the criterion, it prevents overfitting during fine-tuning and ensures long-term robustness.

[0080] Two-stage transfer learning: Global pre-training: Learning universal dynamics of the Earth system (such as frontogenesis and convection triggering conditions) on reanalysis data such as ERA5;

[0081] Regional LoRA fine-tuning: Activating the low-rank adapter with only a small amount of local data (<1 month) to quickly inject local climate features;

[0082] TX-Diffusion supports uncertainty-aware generation, multi-step denoising process output, and automatically expands the prediction range for extreme events to avoid overconfidence;

[0083] The composite loss function enhances learning from extreme samples, explicitly increasing the gradient weights for low-probability, high-impact events. It reduces the cold-start error of new data stations by 40%, significantly improves the reliability of extreme weather predictions, and enables multi-location adaptation with a single pre-training. It overcomes the drawback of static models where generalization performance drops sharply under new data stations, new climate models, or extreme weather conditions.

[0084] The business collaboration and decision-making closed-loop module deeply integrates the meteorological forecast results output by the large meteorological forecasting model with business rules. Through multi-objective optimization, it generates business operation suggestions or instructions. The module receives the forecast results output by the large meteorological forecasting model, performs spatiotemporal alignment and unit standardization, generates structured meteorological time series data objects for business nodes, loads a pre-configured business rule template library, and each rule includes trigger conditions, hard constraints, optimization objectives, and action templates. For each business node and time period, it dynamically matches all activated rules and constructs a decision variable space accordingly. It extracts multi-objective functions, assigns weights to the soft objectives corresponding to each rule according to business strategies, and uses a weighted sum method to fuse them into a single-objective function. It integrates all hard constraints to form a constrained optimization problem, calls a lightweight solver to obtain the optimal decision, and generates structured business instructions.

[0085] The business collaboration and decision-making closed-loop module automatically maps the grid forecasts output by the meteorological big data model to specific business nodes and converts them into the physical units required by the business system, eliminating the integration barrier that "grid data cannot be recognized by the scheduling system" and eradicating the problem of data silos.

[0086] A pre-configured business rule template library: Each rule encapsulates trigger conditions (such as "wind speed > 15m / s"), hard constraints (such as "climb rate ≤ 0.1MW / min"), optimization objectives (such as "minimize wind curtailment"), and action templates (such as "issue power curtailment instructions"). Rules are dynamically bound to meteorological data, and business knowledge such as scheduling procedures and safety specifications are explicitly encoded and embedded into the decision-making process, making meteorological forecasts a dynamic variable that drives business logic and achieving a deep integration of meteorological intelligence and business semantics.

[0087] By dynamically constructing a decision variable space (such as the curtailment ratio) through multi-objective optimization, weighting it according to business strategies (day-ahead / real-time / emergency), and merging it into a single objective function, all hard constraints (equipment limits, safety boundaries) are strictly integrated; constrained optimization problems are quickly solved to generate optimal solutions; and executable, compliant, and optimized business instructions (such as risk warnings and equipment protection actions) are directly output, realizing forecasting as decision-making and decision-making as execution, significantly shortening the response chain. This enables meteorological forecasting to support autonomous, safe, and economical operation under high-proportion renewable energy access, improving the overall system efficiency by more than 30% (measured reduction in wind curtailment rate, lower standby costs, and extended equipment lifespan).

[0088] The integrated service support module exposes the capabilities of large-scale meteorological forecasting models to the public via API interfaces. It exports trained meteorological forecasting models to ONNX or TorchScript inference optimization formats and encapsulates them into standardized inference functions encompassing preprocessing, model inference, and post-processing. A RESTful interface is provided, supporting synchronous / asynchronous calls, with clearly defined request parameters and response structures. Output supports NetCDF, GeoJSON, and CSV formats. Users do not need to concern themselves with model training and maintenance details; they can simply call the interface as needed to obtain high-precision weather forecasts, extreme event warnings, and other results, achieving on-demand intelligent meteorological services. Customized access is supported across multiple scenarios, scales, and frequencies.

[0089] Six new energy wind farms in a certain area were selected to evaluate the wind speed prediction results of their AI new energy meteorological big data model system, examining the prediction performance for day D+1 of September 2025 (i.e., one day in advance). Comparative data included forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) and the forecasts from this system (AI new energy meteorological big data model system), with root mean square error (RMSE) used as the evaluation metric.

[0090] Table 1 Forecast of New Energy Wind Farms

[0091]

[0092] As shown in Table 1, the RMSE of the AI ​​New Energy Meteorological Big Data Model System was lower than that of the EC forecast for all six wind farms, indicating that its wind speed prediction results are closer to the actual observations and have higher prediction accuracy. The AI ​​New Energy Meteorological Big Data Model System performed better.

[0093] The above-described 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-powered new energy meteorological big data model system, characterized in that: It includes a multi-source data integration module, a data governance and quality control module, an intelligent modeling and training module, a business collaboration and decision-making closed-loop module, and an integrated service support module; Multi-source data integration module: Acquires multi-dimensional heterogeneous data sources including: meteorological data, energy data, and geographic auxiliary data; Data governance and quality control module: processes diverse and heterogeneous data to construct a meteorological and energy fusion data cube; Intelligent modeling and training module: Based on a deep learning architecture that integrates VDT and TX-Diffusion mechanisms, a large-scale meteorological forecasting model for new energy scenarios is constructed. Business collaboration and decision-making closed-loop module: Deeply integrates the weather forecast results output by the large weather forecast model with business rules, and generates business operation suggestions or instructions through multi-objective optimization; Integrated service support module: Opens up the capabilities of large-scale meteorological forecasting models to the public via API interfaces.

2. The AI-powered new energy meteorological big data model system according to claim 1, characterized in that: Meteorological data includes meteorological system observation data, station meteorological observation data, and background field data; energy data includes SCADA data of new energy power stations, power grid dispatch instructions, and power grid topology and constraint information; geographic auxiliary data includes coordinates of new energy power stations and terrain elevation data.

3. The AI-powered new energy meteorological big data model system according to claim 2, characterized in that: Processing diverse and heterogeneous data includes data standardization, temporal / spatial dimension alignment, outlier removal, and missing value filling. The final fusion is achieved through multi-source feature fusion and structured organization.

4. The AI-powered new energy meteorological big data model system according to claim 3, characterized in that: The intelligent modeling and training module divides the meteorological and energy fusion data cube into patches containing multiple time steps and local spatial regions through spatiotemporal joint cube embedding, and projects them into a high-dimensional embedding sequence. It adopts a U-Net-like backbone network with a bottleneck layer composed of VDT, and enhances the perception of meteorological dynamic processes by modeling the differential signals of adjacent time steps. The TX-Diffusion Model is introduced as an end-to-end inference framework: after the backbone network generates the initial prediction, multi-step iterative denoising is performed through chained Earth Transformer Blocks and DConv, supporting dual modes of one-step cascaded prediction and multi-step fine-tuning prediction. The entire TX-Diffusion chain is trained jointly with the backbone network. The loss function is applied to the denoised output at each step, and a custom composite loss function is adopted, which includes physical constraint terms, business-sensitive weighting terms, and extreme event enhancement terms. A two-stage transfer learning strategy is adopted: first, a general Earth system dynamic model is pre-trained on global reanalysis data; then, small-sample fine-tuning is performed on regional or field-level fusion data to generate a large-scale meteorological forecasting model.

5. The AI-powered new energy meteorological big data model system according to claim 4, characterized in that: Cube Embedding encodes the meteorological and energy fusion data cube into a high-dimensional spatiotemporal sequence. Specifically, it includes: inputting the meteorological and energy fusion data cube, obtaining the total number of patches through spatiotemporal segmentation, flattening each patch, projecting all patches using a learnable weight matrix, and embedding them together; assigning a unique spatiotemporal index to each patch, using learnable 3D position encoding, and jointly encoding the temporal order and spatial coordinates to obtain a high-dimensional spatiotemporal sequence.

6. The AI-powered new energy meteorological big data model system according to claim 5, characterized in that: The U-Net-like backbone network feature extraction adopts an encoder-decoder structure. The encoder contains several levels of down blocks, which compress the spatial dimension and increase the channel dimension layer by layer, and retain the features of each level for skip connections. The bottleneck layer adopts VDT, which explicitly models the differential signal of adjacent time steps on the deepest high semantic features. The decoder upsamples step by step through Up Blocks and fuses the skip connection features of the corresponding encoder levels. It then combines deep convolution (DConv) to restore spatial details and optimize boundaries, finally outputting the initial prediction results.

7. The AI-powered new energy meteorological big data model system according to claim 6, characterized in that: The two-stage transfer learning strategy specifically includes: 1) Construct a global pre-trained dataset, define a multi-step autoregressive prediction task, adopt a basic loss function that includes physical constraints, and perform end-to-end pre-training on the backbone architecture that integrates VDT and TX-Diffusion to obtain a general Earth system dynamic model. 2) Construct a local fine-tuning dataset, employ efficient parameter fine-tuning, freeze non-critical layers in Cube Embedding, Down / Up Blocks, and TX-Diffusion, and fine-tune only the attention projection matrix, the last 1-2 Earth TransformerBlocks, and the output header in VDT; introduce LoRA to avoid catastrophic forgetting and overfitting, and use a composite loss function that includes physical constraint terms, business-sensitive weighting terms, and extreme event enhancement terms in the fine-tuning process, and use the validation value of the composite loss function itself as the early stopping criterion, finally generating a large-scale weather forecasting model.

8. The AI-powered new energy meteorological big data model system according to claim 7, characterized in that: The business collaboration and decision-making closed-loop module receives forecast results from the large meteorological forecasting model, performs spatiotemporal alignment and unit standardization, generates structured meteorological time series data objects for business nodes, loads a pre-configured business rule template library, and each rule includes trigger conditions, hard constraints, optimization objectives, and action templates. For each business node and time period, it dynamically matches all activated rules and constructs a decision variable space accordingly. It extracts multi-objective functions, assigns weights to the soft objectives corresponding to each rule according to business strategies, and uses a weighted sum method to fuse them into a single objective function. It integrates all hard constraints to form a constrained optimization problem, calls a lightweight solver to obtain the optimal decision, and generates structured business instructions.

9. The AI-powered new energy meteorological big data model system according to claim 8, characterized in that: Export the trained weather forecasting model to an ONNX or TorchScript inference optimization format and encapsulate it into a standardized inference function that includes preprocessing, model inference, and postprocessing; provide a RESTful interface that supports synchronous / asynchronous calls, defines clear request parameters and response structures, and supports NetCDF, GeoJSON, and CSV formats for output.

Images