Estimación de caudales en una cuenca tropical mediante modelación neuronal artificial basada en memoria hidrológica

Introduction: Hydrological systems have been severely impacted by the increase in extreme events such as floods and droughts, intensified by climate change. The Tuluá River basin, located in the Cauca Valley of Colombia, faces additional challenges due to deforestation and limited monitoring coverage, complicating efficient water resource management. Objectives: Develop a neural model based on LSTM networks to predict the flow of the Tuluá River from hydrometeorological variables, for the anticipation of extreme scenarios and support in decision-making in regional water management. Methodology: 6,875 daily records for five variables (precipitation, temperature, evaporation, relative humidity, and streamflow) from seven stations were consolidated. To fill missing values, data preprocessing was applied using KNN and Pearson correlation analysis. The LSTM model was configured and trained, and evaluated for the training, test, and validation subsets using standard neural network metrics. Results: The neural model demonstrated high predictive capacity with correlation coefficients R (0.75–0.98). It was evident that the model effectively captures temporal dependencies and nonlinear patterns, in addition to identifying the influence of the ENSO phenomenon on local hydrological dynamics. Conclusions: The LSTM model is an effective tool for estimating flows for the basin under study, facilitating both water management and the prevention of risks associated with climate variability. Future developments can enhance its accuracy and usefulness if they include global climate indices.