Enhancing the Applicability of Data-Driven Models for Streamflow Forecasting through Optimal Framework with Exogenous Feature Augmentation

Accurate streamflow forecasting, particularly for peak events, poses a significant challenge in hydrology due to water systems' complex and dynamic nature. This study introduces a comprehensive framework designed to enhance the performance of data-driven models in predicting peak streamflow by employing advanced feature augmentation techniques. We designed a forecasting model using a Long Short-Term Memory (LSTM) network with an attention mechanism, which is particularly suited for capturing long-term dependencies in sequential data. To further refine the model, we integrated four distinct feature augmentation methods: Time Series Classification (TSC) to categorize patterns, Variational Mode Decomposition (VMD) to decompose the time series into different frequency components, Mutual Information (MI) to quantify the dependence of lag-time variables, and a Stacking Model (SM) to combine multiple predictive models as input variables. Simultaneously, we optimized the model’s input features and LSTM hyperparameters using Bayesian optimization, a probabilistic model-based method that efficiently navigates the search space. The model is trained on approximately 1,000 events for the Soyang Dam watershed in South Korea. Evaluation against benchmark machine learning models using metrics such as Peak Error (PE), Nash Sutcliffe Efficiency (NSE), and Kling–Gupta Efficiency (KGE) demonstrated that: (i) the inclusion of feature augmentation significantly improves peak predictability, reducing PE by up to 20%; and (ii) dual optimization of input features and model hyperparameters enhances overall accuracy (with up to a 15% improvement in NSE and KGE) and reduces computation time. This study provides a detailed guide for constructing efficient data-driven models using feature augmentation techniques, offering a robust approach to addressing the challenges of streamflow forecasting.

Key words: Model Framework, Feature Augmentation, Streamflow Forecasting, Input Predictor Optimization, Hyperparameter Optimization.

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).