Explanation-Enhanced Federated Fine-Tuning of Large Language Models for Photovoltaic Power Forecasting

Accurate photovoltaic (PV) power forecasting is essential for the reliable operation of modern power systems. While federated learning (FL) preserves data privacy across geographically distributed PV stations, it suffers from performance degradation under non-independent and identically distributed data. Large language models (LLMs), known for their strong generalization and few-shot learning capabilities, offer a promising solution to this challenge. However, full-model fine-tuning of LLMs incurs considerable communication and computation costs in FL settings. To address this bottleneck, we propose a parameter-efficient fine-tuning framework using low-rank adaptation (LoRA) modules. While reducing overhead, conventional federated LoRA introduces aggregation noise due to the separate averaging of low-rank factors. To mitigate this, we introduce an explanation-enhanced unilateral LoRA paradigm that selectively aggregates generalizable modules while retaining task-specific ones locally. This design reduces noise, accelerates convergence, and improves interpretability. Extensive experiments on real-world PV datasets, using GPT-2 Small and GPT-2 XLarge, validate the effectiveness of the proposed approach in improving prediction accuracy and training efficiency under heterogeneous data conditions.