Real-time error correction of multiple-hour-ahead flash flood forecasting based on the sliding runoff-rain data and deep learning models

Abstract

Real-time error correction of flood forecasting is a key method for improving forecast accuracy. However, due to the rapid and unpredictable rise of flash flood discharge and the limited availability of analysis data on short temporal scales, developing robust forecast error correction methods remains a challenge. In this study, we employed a physically-based distributed hydrological model combined with deep learning techniques to develop a real-time error correction method for continuous flash flood forecasting based on sliding runoff-rain data. Taking a typical mountainous river in southwestern China as the study area, we established three input schemes: “sliding runoff data only” (Scheme 1), “hydrological model outputs and sliding runoff data” (Scheme 2), and “hydrological model outputs and sliding runoff-rain data” (Scheme 3). We compared the real-time correction performance of three deep learning models with different architectures—CNN, LSTM, and Transformer—under different input schemes. The results indicate that: 1) LSTM performed the best and most consistently according to the three main evaluation metrics. Although the Transformer showed performance fluctuations, it demonstrated great potential in long forecast correction times, where the correlation between feature inputs and target values is relatively weak. 2) After adding sliding cumulative maximum precipitation data, CNN performance improved significantly, especially in correcting multi-peak floods. 3) The length of the forecast correction time has a significant impact on correction performance. When the forecast correction time approximates the basin’s lag time of runoff concentration, the corrected results have reached a relatively reliable level and entered a more stable phase. This method effectively improves the accuracy of real-time flash flood multiple-hour-ahead forecasting and could provide reliable references for disaster management authorities.