Variants of Recurrent Neural Network Models for Real-Time Flood Forecasting in Kelani River Basin, Sri Lanka

Abstract

The rapid advancement in computer technology has supported flood forecasting, especially neural networks (NN), an application of data -driven models. However, predict ion reliability is compromised due to the data manipulation strategies and the length of the predictive horizon, especially the one-month horizon, which is ample for pre-flood management. Therefore, six (06) variants of recurrent neural networks (RNN) such as Long-and Short-Term Model (LSTM ), Gated Recurrent Unit (GRU), Stacked Bidirect ional and Unidirectional LSTM (SBU-LSTM ), SB U-GRU, Convolution Neural Network LSTM (C NN-LSTM ) and CNN-GR U, were developed for the Kelani R iver Basin to validate their applicability in encouraging the accuracy of monthly flood forecasting by adapting a proper data manipulation technique. Initially, climatic, and physiographic factors of the basin, where the social and economic values are grievously interrupted by frequent floods, were gathered for the study. Then, the hydrological and data science cleansing strategies were adapted to enhance the quality of the data. Besides, a Box-Cox transformation was implemented to redistribute the hydrological data into a Gaussian form to remove the significant deviation between higher and lower values. Next, grid analysis was conducted using statistical tools to quantify the performance, while the influence of data handling and model architecture was examined using uncertainty and sensitivity analysis. LSTM, GRU, SB U-LSTM, SBU-GR U, C NN-LSTM , and CNN-GR U expressed nearly 81%, 81%, 83%, 83%, 76%, and 62%, respectively, for the coefficient of determination (R2) which measures how well the forecasted values fit with the actual values. SBU-LSTM and SBU-GRU interpreted similar behavior to LSTM and GRU; however, the pattern was different in CNN-LSTM and CNN-GRU. Specifically, simple variants LSTM and GRU provided satisfactory results for the uncertainty and sensitivity analysis categories.