24AI Machine Learning for Drainage Flow Prediction

Drainage flow studies were usually conducted for complex terrains, such as for mountainous areas. SAVANT (stable atmospheric variability and transport) collected drainage flow and atmospheric data at a site of shallow terrain. Based on this data, machine learning was applied to predict atmospheric conditions that can inform possible drainage flows at a shallow terrain based on the standards in research conducted by Bhimireddy et al. (2023).

Six models were trained using SAVANT 5-minute atmospheric data collected in September, October, and early November 2018. Then, these models were applied to forecast drainage-flow meteorological parameters and were evaluated for data in mid-to-late November 2018. Inputs to the models were 24 hours historical atmospheric data with 5-minute frequencies, and the predicted data were future three hours data with 5-minute frequency (wind speed, wind direction, friction velocity, surface heat flux, rate of wind direction change, and temperature change with height).

The research found models with high robustness and accuracy. Also found during the research was that the graph structure to express the dependency between the meteorological parameters dynamically can effectively improve the accuracy of the models.