Abstract
River streamflow forecasting is essential for water resources management and flood damage mitigation. Due to their superiority at solving time series problems, several studies have evaluated deep learning models for river streamflow forecasting. However, forecasting small river flow rates, especially during flood events, is a challenge of particular difficulty due to the scarcity of data and the sudden appearance of floods. This work investigates the performance of five deep-learning models for small river streamflow forecasting and particularly focuses on flow forecasting during flood events. The models considered in this work are long short-term memory (LSTM), Gated Recurrent units (GRU), Bidirectional LSTM, 1D convolutional neural networks (1DCNN), and sequential ConvLSTM models. In addition to the standard performance metrics used to evaluate models, this work introduces and applies a new evaluation metric to measure the time effectiveness of models in the early prediction of floods, taking into account the time budget to take early actions. The models are used to forecast the flow rate values over the next 6 hours, 3 hours, and 1 hour. Kyll River, a small river in western Germany, was chosen as a case study because it was subject to several flood events. Results show that all models achieved good performance in overall small river flow forecasting. Results also show that the LSTM-based and the GRU models exhibited poor streamflow forecasting performance (accuracy and time) during normal and extreme flood periods and showed a big decrease in their performance during extreme floods. The findings of this study show that 1DCNN had the highest streamflow forecasting performance, and it reaches impressive forecast accuracy on the entire test set periods and even in normal and extreme flood periods. This suggests that the 1DCNN algorithm can be used as a part of an early flood warning and prediction system in case of both normal and extreme floods.