Short-term flood probability density forecasting using a conceptual hydrological model with machine learning techniques

Abstract

Making accurate and reliable probability density forecasts of flood processes is fundamentally challenging for machine learning techniques, especially when prediction targets are outside the range of training data. Conceptual hydrological models can reduce rainfall-runoff modelling errors with efficient quasi-physical mechanisms. The Monotone Composite Quantile Regression Neural Network (MCQRNN) is used for the first time to make probability density forecasts of flood processes and serves as a benchmark model, whereas it confronts the drawbacks of overfitting and biased-prediction. Here we propose an integrated model (i.e. XAJ-MCQRNN) that incorporates Xinanjiang conceptual model (XAJ) and MCQRNN to overcome the phenomena of error propagation and accumulation encountered in multi-step-ahead flood probability density forecasts. We consider flood forecasts as a function of rainfall factors and runoff data. The models are evaluated by long-term (2009–2015) 3-hour streamflow series of the Jianxi River catchment in China and rainfall products of the European Centre for Medium-Range Weather Forecasts. Results demonstrated that the proposed XAJ-MCQRNN model can not only outperform the MCQRNN model but also prominently enhance the accuracy and reliability of multi-step-ahead probability density forecasts of flood process. Regarding short-term forecasts in testing stages at four horizons, the XAJ-MCQRNN model achieved higher Nash-Sutcliffe Efficiency but lower Root Mean Square Error values, while improving Coverage Ratio and Relative Bandwidth values in comparison to the MCQRNN model. Consequently, the improvement can benefit the mitigation of the impacts associated with uncertainties of extreme flood and rainfall events as well as promote the accuracy and reliability of flood forecasting and early warning.