Dynamic Graph Weather Model for Wind-Solar Power Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing meteorological models, particularly those used for new-energy power prediction, suffer from low accuracy due to reliance on traditional numerical weather prediction (NWP) models that are not customized for new-energy needs, have limited updating frequency, and utilize static graph structures that fail to dynamically simulate changing meteorological conditions, leading to suboptimal prediction of critical variables like hub-height wind and irradiance.
Innovation Solution
A new-energy meteorological large model is constructed using a multilayer dynamic graph-neural-network structure with a wind-solar-node attention mechanism, incorporating ERA5 data and dynamic graph adjustments based on atmospheric pressure layers, and employing a node attention mechanism to enhance prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional NWP models are used for meteorological prediction, then the prediction can be obtained, but the computational resource consumption is large and the prediction accuracy is limited due to complex physical processes
Solution Approach 1:
The patent replaces the traditional physical-based NWP models with a data-driven meteorological large model. Instead of relying on complex physical equations and numerical calculations, the system uses a neural network architecture that processes historical meteorological data to generate predictions. This substitution of mechanical/physical computation with data-driven computational approaches significantly reduces computational resource requirements while improving prediction accuracy for new-energy-specific variables.
2Productivity
If existing meteorological large models are used, then computational efficiency is improved, but the prediction accuracy for critical new-energy variables such as hub-height wind and irradiance remains insufficient
Solution Approach 1:
The patent applies local quality by customizing the meteorological large model specifically for new-energy applications. Instead of using a generic model, the system incorporates specialized components including a wind-solar-node attention mechanism that selectively focuses on nodes relevant to wind and solar power prediction, and a dynamic graph structure that adapts to specific meteorological conditions. This localized optimization ensures high prediction accuracy for critical variables like hub-height wind and irradiance while maintaining computational efficiency.
Solution Approach 2:
The patent introduces a dynamic graph structure that can adaptively reconfigure based on meteorological conditions and prediction requirements. The graph structure dynamically adjusts connections and node activations to focus computational resources on the most relevant variables and spatial locations, thereby improving prediction accuracy for new-energy-specific parameters while maintaining high computational efficiency through selective processing.
3Device complexity
If static graph structure is used in graph neural networks, then the model structure is simple, but the ability to simulate dynamic meteorological conditions is limited
Solution Approach 1:
The patent transforms the static graph structure into a dynamic one that can adaptively reconfigure based on meteorological conditions. The dynamic graph structure allows the model to adjust connections and node activations in real-time, enabling accurate simulation of evolving weather patterns and meteorological dynamics. This dynamic adaptation capability significantly improves simulation accuracy while the modular design keeps the overall structure manageable.
Data Source
Figure 1
AI summary
A method and apparatus for constructing a new-energy meteorological large model, and a power predicting method are provided by the present application. The method includes, based on ERAS data, constructing datasets for training the new-energy meteorological large model; constructing a graph-neural-network large model that has a wind-solar node attention mechanism and a dynamic graph structure; applying the constructed input dataset and output dataset into the graph-neural-network large model, and performing training on the graph-neural-network large model to obtain the new-energy meteorological large model; and by using the new-energy meteorological large model, obtaining new-energy meteorological-element-prediction data, and inputting the station new-energy meteorological-prediction data into the meteorology-power post-correction model to obtain a predicted power. In the present application, the accuracy of prediction on wind-solar meteorological variables by meteorological large models is increased, and it is of great significance for the improvement of the power prediction service.