高精度的水位预测能为防洪决策、水资源管理等提供重要的调度依据，减少洪旱灾害损失。为提高预报精度，提出一种基于小波分析的NARX神经网络模型（DWT-NARX），综合考虑洪泽湖入湖流量、出湖流量、周边用水、前期水位等因素，对洪泽湖日水位进行预报，并与BP神经网络、NARX神经网络模型进行比较。结果表明，三种模型在短历时预报中均取得了较好的模拟预测效果。当预见期为1或2天时，Nash-Sutcliffe效率系数均大于0.9，合格率大于85%；当预见期超过3 d，NARX模型在水位变幅较大的时段预测结果变差，BP模型出现严重的震荡现象，NARX和DWT-NARX模型结果均优于BP神经网络，DWT-NARX在整体上结果最优。研究成果可为洪泽湖的水位预报提供一定的参考价值。

Reliable water level forecasting is essential for flood prevention decision-making and water resources management,which can effectively reduce the loss of flood and droughts disasters.In order to improve the accuracy of forecasting,a nonlinear autoregressive with exogenous inputs neural network (NARX) model based on wavelet analysis (DWT-NARX) was proposed and compared with BP neural network,and NARX neural network model.The daily inflow,outflow,water utilization and the previous daily water level of Hongze Lake were considered to forecast the water level of Hongze Lake.The results indicated that three models achieved good simulation results with higher accuracy when the leading time was short,such as 1 or 2 days.The results exhibited that Nash-Sutcliffe coefficient was higher than 0.9,and the qualified rates surpass was not less than 85%.When the prediction period was further increased to 3 days,the NARX model showed poor prediction and the water level changed greatly,while BP model suggest severe oscillations.In overall performance,the NARX and DWT-NARX models showed superiority in comparison of BP neural network,while DWT-NARX yields the best performance among all other models.The research results can provide a certain reference value for the water level forecast of Hongze Lake.

国家重点研发计划（2016YFC0402705）；国家自然科学基金(51679061；41130639)