Temporal convolutional network algorithm for streamflow predictions in a subtropical river

Abstract

Rainfall-runoff models have a high degree of uncertainty and stochasticity, and the relationship between them is non-linear. Conventional hydrology streamflow prediction models are mostly built for specific watersheds and specific prediction scales, which are poorly promoted and applied. Therefore, in some scenarios, data-driven machine learning predictive models are replacing traditional physical models. Long short-term memory (LSTM) network is a machine learning algorithm for predicting time series and has been applied in the field of streamflow prediction. Temporal convolutional network (TCN) is another machine learning algorithm that is gaining popularity in the field of time series forecasting. LSTM and TCN were implemented in this study to analyse the hourly streamflow prediction for the Nerang River at the Numinbah gauging site, and the predictive accuracy of the models on the test dataset was calculated based on the historical data of the study area. According to the results of the analysis, the TCN model achieved better performance for the hourly streamflow prediction with a coefficient of determination (R2) of 0.9837 and Nash-Sutcliffe efficiency (NSE) of 0.9829 in the best scenario and the lag time for hourly streamflow generation is about three hours in the study area. Additionally, the maximum predicted lead time is six hours in the study area on the TCN model.