A Weight-Allocation and Privacy-Preserving Federated Learning Approach for the BTM PV Power Forecasting

Distributed ultra-short-term photovoltaic (PV) power forecasting plays a vital role in analyzing behind-the-meter (BTM) solar generation and enhancing grid stability. As federated learning (FL) is increasingly applied to distributed BTM PV forecasting on a large scale, its deployment is confronted with two major challenges, namely performance degradation arising from device heterogeneity and privacy risks associated with global communication. To address these challenges, a novel collaborative method, termed AWA-DDP-FL, is proposed, which integrates adaptive weight assignment (AWA) with reinforcement learning-driven dynamic differential privacy (DDP). Specifically, AWA mitigates the impact of device heterogeneity by strategically excluding underperforming clients and dynamically adjusting local model aggregation weights during training. Meanwhile, DDP enables each distributed PV client to autonomously adjust its privacy budget in real time according to data sensitivity, balancing privacy protection and model performance through dynamic noise injection. Experimental results on real-world PV datasets show that the proposed method obtains normalized root mean square error of 5.110% and normalized mean absolute error of 2.630%, indicating that the proposed algorithm achieves higher accuracy even under noise perturbations.