Hydrological simulations and predictions are vital aspects of hydrological change research. Accurate predictions of hydrological factors such as stage and discharge are essential for water resources planning, reservoir dispatching and operation, shipping management and flood control. River discharge forecasting during flood seasons is an important issue in water resources planning and management. To improve the calibration accuracy and stability of the stage–discharge relationship model, the feasibility of integrated algorithms for studying the stage–discharge relationship was explored. A random forest (RF) algorithm based on a neural network (NN) was developed using a framework of integrated algorithms. First, the Levenberg–Marquardt (LM) algorithm was used to optimise the weight updating process of a back-propagation (BP) NN and improve the convergence speed. Then, the LM-BP algorithm was used as a decision tree to build an RF algorithm. The model was tested using hydrological data from Hongqi station on the Dadu River in China in the flood season. Results for the classical model, BP NN model, LM-BP NN model and optimised algorithm model were evaluated based on the mean absolute error (MAE), mean square error (MSE) and mean absolute percentage error (MAPE). The performance indicators showed that the optimised algorithm model (MAE = 3.13 m3/s, MSE = 19.28 m3/s and MAPE = 1.8%) was superior to the other models and showed high accuracy and good stability in flood-season flow forecasting.
Read full abstract