Integração de modelos hidrológicos baseados em processos físicos e técnicas de aprendizado de máquina

Abstract

In recent years, the incorporation of Artificial Intelligence (AI) techniques in various areas of hydrology has grown dramatically. AI-based models, especially recurrent neural networks such as Long Short-Term Memory (LSTM), have demonstrated superior performance compared to traditional conceptual models in various applied contexts. However, researchers face difficulties in physically interpreting the relationships predicted by these models, as well as limitations in their extrapolation capacity and prediction of untrained variables. Hybrid models emerge as a promising alternative, combining the analytical capacity of AI-based models with the physical constraints of conceptual models. This study investigated hybrid models based on three aspects: (i) the capacity of the data-driven component to abstract the conceptual component in search of better performance indices, (ii) the response of these models in spatial and temporal extrapolation scenarios, and (iii) their ability to simulate hydrological states not present in the input data. To this end, two types of hybrid models based on the GR4J conceptual model were developed and evaluated: one with dynamic parameterization, where an LSTM network estimates time-varying parameters, and another with internal flow substitution by an LSTM network. The results demonstrated that hybrid models may prioritize the data-driven component at the expense of physical representation to increase predictive capacity if they are not adequately constrained or have an incoherent conceptual structure. In spatial extrapolation scenarios, the LSTM network demonstrated superior performance, followed by the hybrid model with dynamic parameterization, while in temporal extrapolation under apparent low data availability, conceptual models outperformed pure LSTM networks and hybrid models. Regarding the simulation of untrained variables, although hybrid models possess this capability, a compromise in the physical coherence of internal states was observed, with both hybrid models showing significantly lower correlations than the GR4J conceptual model when compared with soil moisture estimates derived from remote sensing. These results indicate that the development of hybrid models requires special attention to implementing constraints that preserve the physical coherence of the represented processes, and that gains in predictive performance frequently occur at the expense of hydrological coherence.