Enhancing hydrological extremes prediction accuracy: Integrating diverse loss functions in Transformer models

Abstract

In this hydrological study, we developed a Transformer-based model to forecast urban river discharges and predict flood peaks, crucial for flood mitigation in urban areas prone to inundation. Utilizing daily precipitation data from 63 meteorological stations and flow data from hydrological stations, we established a correlation using the Random Forest method to determine the lag time between precipitation and flow. The model, enhanced with alternative loss functions – Weighted MSE Loss (WMSE), Huber Loss (Hloss), and Quantile Loss (Qloss) – instead of traditional Mean Squared Error (MSE), aims to project daily flow rates for seven days. Our findings indicate that Hloss significantly reduces absolute errors in peak value predictions, while WMSE improves linear correlation in forecasting. The accuracy remains stable for the initial four days, with a decrease from the fifth day. This approach, integrating diverse loss functions, presents a novel method for accurately predicting river discharges, offering vital insights for proactive flood management.