Hybrid Deep Learning Time-Lagged Precipitation-Runoff Model

Hydrological models are pivotal tools for comprehending and managing water resources in various applications, including urban and hydrological planning. Nevertheless, these models often grapple with challenges due to the unavailability and sensitivity of specific data required for calibration. This paper aims to explore machine learning algorithms as a potential alternative to traditional hydrological models for flow estimation, focusing on the case study of the dam Três Marias-MG in Brazil. Hybrid models merging spatial data processing with deep learning were investigated to determine the optimal data configuration for enhanced learning. The study area was partitioned into 91 quadrants, with distances from each quadrant’s centroid to the flow point calculated. With an estimated velocity of 1 m/s, the lag time, denoting the time each quadrant takes to contribute to the flow, was established. This information facilitated the adjustment of the input precipitation data vector to align with the flow contribution output. The model achieved an average error of approximately 4.4% and a relative peak flow of 19.7% compared with 5.1% and 40% with the regular model, making it a robust approach to flow estimation.