PatchXFormer: Bridging the Data Gap in Tropical Solar Power Forecasting for Equitable Renewable Energy

Accurate solar power forecasting is critical for grid stability and energy planning. Yet the majority of machine learning research in this domain focuses on datasets from temperate, resource-rich regions, leaving tropical and developing countries severely underrepresented. We address this inequity by introducing PatchXFormer, a novel Transformer-based architecture that combines frequency-enhanced attention, cross-attention for integrating meteorological features, and adaptive normalization to achieve robust solar power forecasting across diverse geographic conditions. Crucially, we contribute three new real-world datasets collected from solar installations in Sri Lanka, a tropical region are critically underserved by existing benchmarks, alongside a temperate-climate dataset from Kansas, USA. Experiments show that PatchXFormer outperforms seven state-of-the-art baselines (including PatchTST, iTransformer, NLinear, and LSTM) across all four forecasting horizons (24 h to 7.5 days), achieving up to a 4.7% lower MSE than the best competing model on the primary tropical dataset. Cross-geographic evaluation reveals that models trained on more variable tropical data generalize substantially better to temperate regions than vice versa, offering actionable guidance for practitioners deploying forecasting systems in data-scarce locations.