Leveraging Machine Learning and TVF-EMD for Improved Streamflow Forecasting

A reliable streamflow prediction is essential for hydrology-related purposes such as flood warnings, water quality management, and emergency readiness. Various machine learning (ML) models have been developed and used for capturing the complex relationships of rainfall, streamflow, and other hydrologic processes. However, there are certain limitations and challenges in their practical applications, such as hyperparameter adjustment and overfitting. This research aims to improve streamflow prediction using different ML models based on daily discharge and precipitation data. To achieve this goal, a new preprocessing data structure is developed in this study to feed ML models for forecasting daily discharge. Specifically, different ML models are selected and combined with a time varying filter based empirical mode decomposition (TVF-EMD) technique to effectively extract relevant features from the datasets in the steps of preprocessing. The models are compared by multicriteria decision-making methods including the MARCOS and ARAS techniques and their performances are assessed using different statistical metrics. The proposed modeling methodology is applied to a watershed in North Dakota to demonstrate its improvement and applicability and showcase its enhanced computing efficiency.