Enhanced streamflow forecasting using attention-based neural network models: a comparative study in MOPEX basins

To mitigate the adverse effects of floods, hydrologists are increasingly turning to artificial intelligence methodologies to enhance streamflow forecasting capabilities. Drawing inspiration from the efficacy of the Long Short-Term Memory (LSTM) model in capturing temporal dynamics and dependencies within data set, we have employed LSTM for predicting sequential flow rates utilizing collected data sets. Recognizing that not all data set contribute equally to accurate flood forecasts, it becomes imperative to discern and prioritize the relevant variables. Conventional LSTM models often fall short in effectively identifying and ranking informative factors. To overcome this limitation, we introduce an Attention LSTM (ALSTM) model tailored for streamflow forecasting, adept at identifying and capturing critical factors within the time series dataset. Leveraging data set sourced from the United States Geological Survey (USGS), our proposed model exhibits notable performance enhancements. By integrating an attention mechanism during the pre-processing stage, the ALSTM model showcases its ability to generate precise long-term forecasts across most of the basins. Utilizing a continuous 33-year streamflow data set (1970–2003), our proposed model surpasses conventional time series approaches in streamflow forecasting accuracy.