[A-HW21] Hydrological modelling to support water resources management and engineering designs

[AHW21-P02] Improving the Performance of a Data-driven Model for Multi-task Streamflow Forecasting

*Jongho Kim¹, Trung Duc Tran¹ (1.University of Ulsan)

Keywords:Data-driven Model, Multi-task forecasting, Input predictor optimization, Hyperparameter optimization

This study systematically outlines a framework for enhancing the performance of data-driven models focused on multi-task streamflow forecasting. Specifically, the framework for four distinct data-driven models—Autoregressive Integrated Moving Average (ARIMA), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Transformer (TRANS)—are developed for multi-task forecasting for the Soyang watershed in South Korea. An approach of using a 'correlation threshold' is introduced for the optimization of the input predictors. Concurrently, hyperparameters are optimized using three robust optimization methods: Bayesian Optimization (BO), Particle Swarm Optimization (PSO), and Gray Wolf Optimization (GWO). The experimental results provide valuable insights into the roles of input predictor optimization, data processing (Wavelet Transform-WT), and hyperparameter optimization. Notably, the 'correlation threshold' method proves effective and straightforward in determining the number of input predictors. Furthermore, the application of WT significantly improves model accuracy across all tasks, resulting in a notable reduction in Nash Sutcliffe Efficiency, Root Mean Square Error, and Peak Error. The GWO method emerges as the most effective in optimizing the hyperparameters of the models. Ultimately, the TRANS model consistently attains the highest accuracy across all tasks. These findings underscore the efficacy of data-driven models in multi-task streamflow forecasting, providing valuable support for managers in implementing proactive measures and formulating optimal decisions for water management strategies.
Acknowledgment: This work was supported by Korea Environment Industry & Technology Institute (KEITI) through Water Management Program for Drought Program funded by Korea Ministry of Environment (2022003610003), and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF- 2022R1A2C2008584).