Comparative analysis of neural network techniques for short-term water level forecasting in a flooding scenario

Flood disasters are one of the most devastating natural hazards. Every year they cause numerous deaths and extensive damage to properties and economic systems in different parts of the world. Consequently, flood prediction is a fundamental research topic in the hydrology field. Researches have attempted to address this problem by using various techniques, ranging from model-driven to data-driven approaches. However, the complex and dynamic nature of the flooding phenomena makes its prediction a challenging task. Nowadays, the improvements in computing power have impulsed the application of neural network models in the flood prediction problem, showing a potential success. This work explores five neural network techniques to predict the water levels in Doboj, Bosnia and Herzegovina (B&H). The employed methods include a Feedforward Neural Network (FNN), a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM), a Gated Recurrent Unit (GRU), and a Temporal Fusion Transformer (TFT). Unlike most of the state-of-the-art studies, this work exploits exogenous inputs to make predictions. We feed the models with multiple input variables recorded at various hydrological and meteorological stations. In addition, this study adopts the flood index (IF), an index based on the Effective Precipitation (PE) used to quantify floods. The results indicate that the CNN model exhibited the best performance, closely followed by GRU and LSTM. In detail, the coefficient of determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) obtained with the CNN model were 0.951, 16.603 cm, and 9.245 cm, respectively, for one day ahead prediction. In contrast, the worst-performing models were the FNN and the TFT. Despite being capable of predicting typical water levels, these networks showed limitations in the flooding scenario. Overall, the results of this study demonstrated the outstanding capabilities of neural network models to predict normal water levels. Furthermore, this work evidences the potential use of CNN, GRU, and LSTM to forecast the high water levels presented during a flood disaster and contribute in this way to flooding risk mitigation.