Regime-aware machine learning for multi-horizon reservoir level forecasting in a regulated hydropower cascade

Abstract

Electricity generation in Albania relies mainly on hydropower from the Drin River cascade, including the Fierza, Koman, and Vau Dejës reservoirs. This study proposes a regime-aware implementation, data-driven framework for predicting reservoir levels at Vau Dejës using a daily dataset spanning 1991 to 2024, integrating hydrological and meteorological variables.To account for seasonal operational conditions, the forecasting is divided into four seasonal regimes: winter inflow control (October to February), seasonal transition (March), spring snowmelt and high inflow (April to May), and summer production management (June to September). Four supervised models are evaluated using rolling-origin cross-validation: Ridge Regression, Support Vector Regression, Gradient Boosting, and Artificial Neural Networks. Results indicate that next-day reservoir dynamics are largely governed by temporal persistence and cascade regulation, resulting in comparable performance between linear and nonlinear models. Ridge Regression achieves the most stable results, with prediction accuracy exceeding 91% and errors within a few centimeters. Lagged reservoir levels dominate short-term predictability, while meteorological variables contribute mainly to regime-specific variability. For longer horizons, a Long Short-Term Memory (LSTM) model captures temporal dependencies, preserving seasonal and interannual variability over multi-year projections.The proposed framework provides an operational decision-support tool for reservoir management, hydropower scheduling, and long-term planning in hydropower-dependent systems.