ABSTRACT Diagram showing precipitation, temperature, and lagged inflow as inputs to RF, MLP, and LSTM models for daily and monthly streamflow forecasting, with best performance varying by basin and poorer accuracy for high flows. This study evaluates three machine learning methods: Random Forest (RF), Long Short-Term Memory networks (LSTM), and Multilayer Perceptron Artificial Neural Networks, to forecast streamflow at daily and monthly scales for the Betania, Quimbo, Hidrosogamoso, and Urrá hydropower basins in Colombia. Model skill was assessed using two input configurations: (1) temperature and rainfall with multiple time delays, and (2) the same predictors including past inflow. Results indicate that no single model, configuration, or temporal resolution performs best across all basins. In some catchments, meteorological variables alone are sufficient, while in others, antecedent conditions represented by previous inflow are essential to improve accuracy. Monthly aggregation can enhance peak-flow representation but may increase overfitting, especially in basins with limited data; thus, daily data often remain more informative. During El Niño -Southern Oscillation (ENSO) events, forecasting skill is satisfactory for normal and low flows, whereas performance deteriorates for high-flow conditions. Overall, model performance tends to be better during El Niño than La Niña conditions. These findings support improved strategic energy planning in hydropower-dependent systems, where accurate inflow forecasts are critical for reliability.
Pulgarín-Morales et al. (Fri,) studied this question.