Real-time rainfall-runoff prediction using light gradient boosting machine coupled with singular spectrum analysis

Abstract Urban rainfall-runoff prediction is an effective method for flood mitigation. However, it is difficult to realize real-time and accurate prediction due to the strong nonlinearity and fluctuation of urban rainfall. This study develops a data-driven model by integrating singular spectrum analysis (SSA) and a light gradient boosting machine (LightGBM) to achieve the high-accuracy, real-time prediction of regional urban runoff. First, SSA decomposes and reconstructs runoff series into trend terms, fluctuating terms, and noise components. Then, a LightGBM is used to simulate the trend and fluctuating terms. Finally, a comparative analysis of the performance of the SSA–LightGBM model under different time steps and prediction lead times is performed, and the peak prediction accuracy of the SSA–LightGBM model under independent rainfall-runoff events is investigated. Evaluation results show that the Nash–Sutcliffe efficiency coefficient of the SSA–LightGBM model remains greater than 0.9, and the error of the peak discharge is within 18%, which significantly outperforms the LightGBM and long short-term memory (LSTM) models under different conditions. Moreover, the computation time of the SSA–LightGBM model (10.000 s) is much shorter than that of the LSTM model (301.128 s), which can meet the calculation time requirement of real-time updating model and prediction. Overall, the SSA–LightGBM model is suitable for urban rainfall-runoff real-time prediction with low computation time and high accuracy.