Exploring the applicability and reliability of machine learning tools in streamflow forecasting

Abstract

Recent advancements in machine learning, particularly Long Short-Term Memory (LSTM) networks, have demonstrated remarkable success in hydrological modelling, often outperforming traditional hydrological models. This thesis provides a comprehensive analysis of LSTM networks for streamflow forecasting under various conditions. First, the impact of incorporating historical streamflow data as an input was evaluated, demonstrating significant improvements in prediction accuracy across diverse catchments. While LSTM outperformed the persistence for one-day-ahead forecasts, accuracy decreased for longer lead times for both models. The effect of noisy precipitation inputs was subsequently investigated, revealing that while noise generally reduces performance, LSTMs trained with noisy data exhibit resilience. Basin sensitivity to precipitation noise varied and correlated with catchment attributes. Lastly, interpolation and extrapolation under stationary and non-stationary climate scenarios were examined. LSTM performed remarkably well under stationary conditions but showed biases when predicting under changing precipitation regimes, highlighting challenges in extrapolation. In conclusion, the thesis summarizes the key findings, addresses its limitations, and suggests avenues for future research, such as incorporating forecasted forcing data and developing hybrid models to improve the robustness of LSTM-based streamflow forecasting.