Federated learning-enhanced non-IID solar power prediction using multi-layer transformers

The increasing integration of photovoltaic (PV) systems into power grids requires accurate solar power forecasting to ensure stability, driven by renewable energy growth and concerns about data privacy under regulations such as General Data Protection Regulation (GDPR). Traditional centralized machine learning struggles with heterogeneous, distributed datasets, motivating the adoption of federated learning for privacy-preserving solutions. This work introduces the SolarTransformer architecture, a novel 2-layer Transformer with multi-head attention, optimized via federated learning (FL) using FedProx across 7 heterogeneous clients, and a performance-optimized variant, SolarTransformer-Enhanced, with 100,673 parameters. The decentralized approach trains on non-independent and identically distributed (non-IID) data, sharing only parameter updates, and employs client-side anomaly detection, real-time ECMWF weather integration, post-training structured pruning + INT8 quantization ( > 5.8 × compression), and incremental federated learning. The results show SolarTransformer (compressed) achieves an average R 2 of 0.9283 after weather integration, RMSE of 0.0836, and SMAPE of 38.39% with only 68,225 parameters and 69 KB storage, while SolarTransformer-Enhanced reaches R 2 0.9303, RMSE 0.0804, though less efficient for edge deployment. The proposed framework significantly outperforms conventional centralized and recent federated learning baselines in predictive precision, robustness, and adaptability. This work establishes a scalable, privacy-preserving, edge-ready forecasting solution that enables real-time grid management and lifelong adaptation in distributed PV systems, with the compressed SolarTransformer favored for efficient deployment on resource-constrained IoT devices. • Novel federated 2-layer Transformer for privacy-preserving solar forecasting. • Real-time ECMWF weather integration boosts R² to 0.9423 on non-IID data. • Client-side anomaly detection + compression enables edge deployment (<70 KB). • Incremental federated learning sustains accuracy over a full year. • Outperforms recent FL and centralized baselines with >5.8× efficiency gain.