Beyond forecast accuracy: A statistical and financial evaluation of machine learning models for small hydropower forecasting

This study develops and evaluates a machine learning-based forecasting framework for improving the day-ahead production estimates of a run-of-river hydropower plant operating under Turkish electricity market conditions. The analysis is based on hourly operational data and compares operator-based heuristic forecasts with four alternative modeling approaches: Multilayer Perceptron (MLP), Random Forest (RF), standard XGBoost, and a cost-aware XGBoost (CA-XGBoost) formulation. To reflect real-world operational constraints, the framework relies on a parsimonious feature structure composed of the operator’s original forecast and a proxy variable derived from the upstream plant’s generation data with a 16-hour hydraulic lag. In this way, the study aims to provide a practical forecasting approach for data-constrained hydropower settings where detailed meteorological inputs may not be readily available. The models are evaluated through both statistical and financial criteria. Statistical performance is assessed using MAE, RMSE, and (R2), while financial performance is examined through imbalance-related market costs calculated using hourly Market Clearing Price and System Marginal Price data under the asymmetric settlement structure of the Turkish electricity market. The results show that Random Forest achieved the best overall economic performance, yielding the lowest total imbalance cost and the lowest MAE, whereas MLP produced the best RMSE and (R2) values. By contrast, the CA-XGBoost approach model did not outperform the benchmark machine learning models, although it did alter the directional structure of forecast errors. The findings reveal that improvements in statistical accuracy do not necessarily imply superior financial performance. Accordingly, the study demonstrates the importance of evaluating hydropower forecasting models not only in terms of conventional error metrics but also with respect to their economic consequences under market-based imbalance pricing. From a practical perspective, the CA- XGBoost proxy-based framework offers a lightweight and potentially scalable decision-support approach for small hydropower plants operating under limited data availability.