From LMP to eLMP: An accurate transfer strategy for electricity price prediction based on learning ensemble

Extended Locational Marginal Pricing (eLMP) integrates start-up costs into market clearing, enhancing participant incentives compared to classical Locational Marginal Pricing (LMP). Despite eLMP’s operational complexity, we propose a two-stage transfer learning strategy that overcomes perceived prediction challenges. First, an LMP predictor is trained by decomposing historical data into distinct modes via Variational Mode Decomposition (VMD). Three customized deep learning architectures and an improved loss function optimize training for each mode, with predictions aggregated nonlinearly through a learning ensemble. Second, the pre-trained model transfers to eLMP by decomposing new pricing data into analogous modes, testing mode compatibility, and selectively fine-tuning mismatched modes alongside the ensemble. Experiments on three Midcontinent ISO (MISO) sites demonstrate that adjusting at most one mode achieves more than 95% prediction accuracy across datasets, surpassing direct eLMP forecasting by more than 50% in accuracy with minimal computational overhead. Theoretical error bounds align with empirical results, confirming near-optimal transfer efficiency. This strategy bridges LMP and eLMP forecasting frameworks, proving that eLMP’s complexity can be practically addressed without exhaustive retraining. By reusing mode-specific knowledge from LMP and enabling targeted fine-tuning, our approach reduces deployment costs while maintaining high generalizability, offering a scalable solution for future electricity market designs.