Research on deep learning flood forecasting model coupled with interval probability calculation method

Abstract

[Objective] Accurate and reliable flood forecasting is one of the important non-engineering measures for flood control and disaster reduction. Improving the accuracy and reliability of inflow flood forecasting for the Guxian Reservoir using deep learning technologies is of great importance for flood control and dispatch decision-making. [Methods] A deep learning-based flood process probabilistic forecasting model, Transformer-Bootstrap, was constructed by coupling the output layer of the Transformer model with the Bootstrap interval prediction method. Using the observed rainfall-runoff data from the Lushi Hydrological Station, which controls the catchment area of the Guxian Reservoir for the period 1990—2016, and applying the Chapman filtering method for baseflow separation, 49 hourly flood event datasets were obtained. Of these, 39 flood events from 1990 to 2010 were used for training, while 10 flood events from 2011 to 2016 were used for validation. The model's performance was evaluated using metrics such as Nash-Sutcliffe Efficiency(NSE), Root Mean Square Error(RMSE), bias, and Coefficient of Determination(R²). [Results] The result showed that for forecast lead times of 1~6 hours, the NSE for the training period decreased from 0.98 to 0.92, while the NSE for the validation period decreased from 0.89 to 0.55. Forecast errors showed an increasing trend, with RMSE and bias for the training period increasing from 31.54 m³/s to 64.04 m³/s and from 9.08% to 24.05%, respectively, while RMSE and bias for the validation period increased from 11.01 m³/s to 20.57 m³/s and from 3.90% to 11.36%, respectively. The model began to improve after approximately 75 iterations and only approached convergence after more than 175 iterations. During the validation period, the PIPC value of the forecast decreased from 89.5% to 61.6%, and the PINAW value increased from 0.008 to 0.076. Particularly, as the forecast lead time increased to 4~6 hours, the bias between the forecast and observed discharge became larger, leading to poorer forecast performance. [Conclusion] The Transformer-Bootstrap model demonstrates good performance for short lead times(1~3 hours), with NSE values exceeding 0.80 for both the training and validation periods. The probabilistic forecast coverage for the Transformer-Bootstrap model generally exceeds or is close to the 90% confidence level interval, and the probabilistic forecast result are reasonably reliable. However, as the lead time increases, the forecast accuracy decreases, and forecast errors increase. Additionally, the model requires more iterative calculations to converge, with the forecast accuracy and stability during the training period superior to those during the validation period. A key scientific issue for future research is how to improve the accuracy and robustness of deep learning flood forecasting models for longer lead times. The findings provide technical support for flood control and disaster reduction in the Guxian Reservoir and parts of the middle reaches of the Yellow River.

Keywords

flood forecasting / deep learning / probabilistic prediction / Bootstrap interval prediction method / Transformer model / Guxian Reservoir / lead-time / flood control and disaster reduction