Leveraging deep learning methods to enhance hydrological predictions and model interpretability

Abstract

The growing frequency and intensity of extreme weather events, along with the mounting challenges posed by climate change, make hydrological models that incorporate cutting-edge technologies and innovation vital for sustainable water management strategies and efficient flood detection systems.This paper investigates the effectiveness of deep learning models, specifically Temporal Fusion Transformer (TFT), Transformer, and Long Short-Term Memory (LSTM), using a linear model as baseline, for predicting water levels using meteorological data as covariates.Hourly data from three weather stations (Aminteo, Vevi, Zazari) spatially distributed in the catchment were used to forecast water levels, for the next 24 hours, of the Amintas stream.The study employs key evaluation metrics, including Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Nash-Sutcliffe Efficiency (NSE), and Kling-Gupta Efficiency (KGE), to assess model performance.Results reveal satisfactory outcomes across all models, with the Temporal Fusion Transformer (TFT) model demonstrating superior performance in RMSE, MAPE, NSE, and KGE metrics.The discussion underscores the models' ability to outperform linear models by extracting hidden information from weather data.The goal of this paper is to enhance water level prediction capabilities in the face of climate-induced hydrological challenges by leveraging advanced deep learning models and innovative techniques.