Machine learning modeling structures and framework for short-term forecasting and long-term projection of Streamflow

Reliable short-term forecasting and long-term projection of streamflow are essential. However, few research models for machine learning structures systematized for short- and long-term time-series predictions and frameworks for simultaneously optimizing hyperparameters are available. In this study, to improve the accuracy degradation phenomenon that occurs especially in long-term forecasting and maximize its efficiency, four long-short term memory (LSTM) model structures (msLSTM, srLSTM, spLSTM, and mjLSTM) with time-invariant or time-variant features are clearly systematized for the first time. We then compare three optimization methods and newly present a dual-loop framework for simultaneously optimizing various hyperparameters that are independent of each other. Finally, using a proposed model structure and dual-optimization framework, future changes in daily streamflow are evaluated. From the results obtained, we can conclude that: (i) the srLSTM structure is appropriate for short-term forecasting, specifically for predicting one to five steps ahead, while the mjLSTM structure is recommended for long-term projection; (ii) Bayesian optimization offers superior accuracy and efficiency compared with a Grid or Random search, and it is preferred for the design of the dual-loop optimization framework required to maximize model performance; and (iii) under the SSP5-8.5 climate change scenario, annual streamflow is projected to increase by approximately 10-20%, and monthly precipitation and streamflow are also expected to increase by at least 10% to a maximum of 40%. The comprehensive model structure and framework proposed in this study offer a promising solution for addressing both short-term forecasting and long-term projection problems using machine learning models.