A Framework for Streamflow Short-term Forecasting and Long-term Projection based on Long Short-Term Memory and Bayesian Optimization

It is essential to ensure the accuracy and reliability of short-term and long-term time-series forecasts in various fields (especially in the water resources) because short-term forecasting is the key to early warning when the extreme events occur, and long-term projection provides important information for water resource management. This study proposed a deep learning model based on a long short-term memory (LSTM) network that is structured for short-term forecasting and long-term projection, respectively and optimized using a Bayesian optimization method. Specifically, this study includes: (i) introducing the LSTM structures that differentiate short-term forecasting and long-term projection; (ii) designing a modelling framework optimized for not only input predictors but also hyperparameters; (iii) assessing the accuracy and efficiency of short-term forecasting and long-term projection for different structures within the framework. Using historical observations and future information downscaled from the Coupled Model Inter-comparison Project Phase 6 (CMIP6) for the Han river watershed, we show that the LSTM models are optimized quickly and with high accuracy (R= 0.9-0.95 and NSE=0.83-0.85). Overall, future streamflow for the period 2020-2010 increases by 10% to 50%, similar to the increase in CMIP6 precipitation. Reliable forecasts and projection of streamflow obtained from this study will help policy makers and operators in reservoir operation, planning and management.