Recalibration of Neural Networks Using Transfer Learning for Streamflow Forecasting

Abstract Streamflow forecasting is a critical component in the design of resilient civil infrastructure, especially when considering extreme events related to river flows, such as flooding. While physics-based simulation models have traditionally been used for streamflow forecasting, their accuracy is often hindered by simplifications and assumptions made in the analysis. On the other hand, data-driven machine learning models leverage historical data from monitoring states, providing an alternative approach to streamflow prediction. However, purely data-driven models can be limited by the availability of monitoring data, which can lead to less accurate predictions. To address these challenges, this paper presents a novel hybrid modeling technique for streamflow forecasting by integrating the strengths of both physics-based and data-driven approaches. The key idea lies in the application of transfer learning to recalibrate neural networks. A long short-term memory (LSTM) model is first constructed using simulation data generated from a physics-based streamflow-forecasting model. This initial LSTM model is then fine-tuned using gauged monitoring data, transforming it into a recalibrated LSTM specially designed for streamflow forecasting. The results of this case study demonstrate the effectiveness of neural network recalibration using transfer learning, which significantly enhances the accuracy of LSTM-based streamflow forecasting and could greatly facilitate the design of resilient civil infrastructures in the face of climate change.