A generative pre-trained transformer-based decision focused framework for wind-solar joint learning on distributed network

With the rapid development of renewable energy, improving forecasting quality has become increasingly important. However, in distribution system operation, the ultimate goal is not prediction accuracy, but cost-effective day-ahead dispatch under wind-solar uncertainty. Decision-Focused Learning (DFL) enhances the decision quality of wind-solar joint forecasting by deeply integrating prediction with optimization. To bridge this gap, this paper proposes GPT-DFL, a two-stage decision-focused learning framework that synergistically integrates a GPT-based forecasting backbone with an Input Convex Neural Network(ICNN)-based decision-aware refinement module. In the first stage, the forecasting-focused method is used to fine-tune GPT, enabling more accurate modeling and prediction of wind and solar power generation. In the second stage, decision-focused fine-tuning is performed using the bias-correlation layer. Decision-Focused Learning is formulated as a constrained optimization problem that maintains the proximity of the decision-enhanced model to the original predictive model within a predefined trust region. The proposed framework is validated on the IEEE 33-bus and IEEE 141-bus systems. The results show cost reductions of 2.36% and 2.42% compared with the best-performing baseline, demonstrating the effectiveness of the proposed method.