基于气象初始场驱动的气象-新能源端到端功率预测方法
By constructing an end-to-end meteorological-new energy power prediction method, and using a hybrid attention mechanism and Swin Transformer to directly fuse meteorological features with historical power plant data, the problem of inconsistent objectives in traditional methods is solved, high-precision power prediction is achieved, and the robustness and efficiency of the model under extreme weather and complex geographical conditions are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-17
AI Technical Summary
In traditional meteorological-new energy power prediction methods, the optimization objectives of meteorological models and new energy power prediction models are inconsistent. This means that improvements in meteorological accuracy may not necessarily lead to improvements in power accuracy, and the errors are amplified in the power prediction process, especially when the weather system changes drastically.
An end-to-end power prediction method based on meteorological initial field is adopted. By constructing a meteorological feature encoder with a hybrid attention mechanism and a power prediction decoder with a Swin Transformer, the meteorological features and historical power data of the power plant are directly fused. Low-rank adaptive LoRA technology and time history decay training strategy are adopted to ensure that the model converges to the global optimum.
It significantly improves the robustness of predictions under extreme weather and complex geographical conditions, reduces the consumption of computing resources and the demand for data samples, and enables the direct cross-module transmission of weather forecast errors back to the underlying meteorological feature extraction layer, avoiding the nonlinear amplification of errors and improving the accuracy of power prediction.
Smart Images
Figure CN122136831B_ABST