Water level prediction based on SBGS model and post-processing error correction

In order to effectively prevent the risk of lower reaches inundation caused by the operation of the Three Gorges-Gezhouba cascade reservoirs, this paper takes the water level of Yanzhiba as the research object, constructs a SBGS deep learning model based on Squeeze-and-Excitation Block, GRU and Selfattention coupling, and forms a dual-stage prediction framework of “deep learning+error correction” by combining autoregressive, Gaussian process regression and random forest post-processing methods, and realizes the water level prediction in the 24-hour forecasting. Using hourly water level data from 2023 to 2024 demonstrates: (1)The SBGS model has a mean absolute error of 0.08m and a pass rate of 70.1%, exhibiting higher accuracy than the single GRU model; (2)Three post-processing methods of autoregression, Gaussian process regression, and random forest can improve the prediction accuracy of the original model. Random forest post-processing shows the highest precision, reducing the mean absolute error to 0.05m and increasing the pass rate to 81.9%; (3)After excluding the period of time when there is a large deviation between the Gezhouba dam outflow plan and the actual outflow, the accuracy of the original model prediction results and the three post-processing results can be improved, which is more realistically reflecting the performance of the model. The dual-stage prediction framework has high prediction accuracy, which provides a new idea for the operation prediction of water level in the lower reaches of the reservoir, and can be used as a scientific basis for preventing the risk of flooding in the lower reaches of Gezhouba Dam and optimizing the operation of cascade reservoirs.