Forecasting day-ahead hydropower bids in the colombian electricity market: A two-stage machine learning framework

Forecasting day-ahead energy prices in Colombia is challenging, as bids are shaped not only by hydrology and demand but also by strategic bidding under evolving regulatory and economic conditions. We develop a Machine Learning (ML) framework to forecast day-ahead bids of individual hydroelectric plants in the National Interconnected System (SIN), using plant-specific models that capture operational and strategic heterogeneity. To our knowledge, this is the first study to develop plant-specific ML models for Colombian hydro bidding. The framework employs a two-stage architecture: (i) a classifier predicts whether a plant will bid at the regulatory minimum, and (ii) a regression model estimates the actual bid otherwise. Using daily data from XM (2016–2025), we benchmark multiple algorithms—including XGBoost, Random Forest, Gradient Boosting, MLP, and stacked ensembles. Across plants, the framework achieves average F1-scores above 0.90 for classification, and regression errors in the range MAE = 1.5–3.0 COP/kWh (normalized by regulatory floor). Tercile classification accuracy exceeds 70% for key strategic plants such as Guavio and San Carlos, indicating strong ordinal predictive skill even under volatile conditions. The models perform particularly well for plants with stable bidding patterns, while highly intermittent plants (e.g. Porce II & III) remain harder to anticipate. These results demonstrate that ML methods can capture mid-term gradients and relative positioning of competitor bids, providing actionable insights for market operators and traders. Future extensions should integrate contextual data—portfolio composition, regulatory signals, or sentiment indicators—to improve forecast robustness in behaviorally complex markets. • Per-plant two-stage ML models for Colombia’s hydro bids. • Operational validation with stress-period test (Oct 2024). • Demonstration that ordinal accuracy (terciles) is more relevant than pointwise error.