Drift-Aware Day-Ahead Price Forecasting in MIBEL with Temporal Fusion Transformers and PEFT Adapters

Carlos Miguel Rodrigues Sequeira

Key information

Authors:

Carlos Miguel Rodrigues Sequeira (Carlos Miguel Rodrigues Sequeira)

Supervisors:

Chrysoula Zerva (Chrysoula Zerva); Rui Fuentecilla Maia Ferreira Neves (Rui Fuentecilla Maia Ferreira Neves)

Published in

November 27, 2025

Abstract

Accurate and interpretable day-ahead forecasts are operationally critical for commitment, hedging, and risk control in MIBEL. However, they must also remain robust to market shocks and policy shifts. We study day-ahead electricity price forecasting in Spain’s MIBEL under a strict pre-issuance information set, predicting 24 hourly prices over 2018-2025. Inputs combine provider generation forecasts with engineered features that encode market logic (net-load; gas marginal-cost proxy). We compare a transparent LASSO-ARX (LEAR) and a feed-forward DNN with a Temporal Fusion Transformer (TFT), evaluated under four maintenance regimes, including parameter-efficient adapters on a frozen backbone. Evaluation is leak-free and walk-forward, using mean absolute error (MAE) as the primary point accuracy metric, Diebold–Mariano tests for conditional predictive ability, and quantile losses, Continuous Ranked Probability Score (CRPS), and interval coverage for probabilistic quality. We show that fortnightly-recalibrated TFT with daily adapters achieves the best point accuracy (MAE 10.33 C/MWh), improving MAE by 23.5% vs LEAR and 11.6% vs DNN. Its quantile-trained variant attains the lowest CRPS, though with under-dispersion (I50~0,35; I90~0,70), while a residual-quantile LEAR wrapper is closer to nominal coverage. Adapter updates are lightweight– median of 15-20 seconds per day– and thus a practical and efficient solution. Overall, this study features a deployment-faithful pipeline for MIBEL day-ahead prices, a backbone-adapter maintenance design that improves accuracy over baselines at low daily cost, and interpretable TFT diagnostics that reveal feature weighting and temporal focus.