The accuracy of solar power predictions drasti- cally declines when deep learning models are run to cover climatically heterogeneous areas. The study presents CATHDB Causal-informed Adaptive Transformer with Hierarchical Do- main Bridging that allows effective multi-location solar prediction without retraining on locations. It has three significant innova- tions, including: (1) causal discovery module isolating climatic- invariant physical relationships and climatic-specific spurious correlations, (2) hierarchical domain bridging that gradually transforms models by using climatic-intermediate zones, and (3) climate-informed multi- modal attention schemes that dynam- ically weight meteorological features. Compared with state-of- the-art transfer learning, CATHDB (when used across 15 geo- graphically diverse PV systems across tropical, arid, temperate, and subtropical climates) makes 12.3 and 19.8 percent higher cross-climatic generalization improvements than standard deep learning models do. The structure has a forecasting accuracy of 76 percent when there is extreme weather as compared to 42 percent with conventional Transformers. Analysis of data efficiency indicates that CATHDB can satisfy 95 percent peak performance with 18 percent of the target-domain historical data, which is 68 percent with non-transfer methods. It is a paradigm shift in the climate-aware AI setting an example to a more adaptable renewable energy forecasting system.
Yashika M Gowda (2026) studied this question.