基于非稳态调和分析和长短时记忆神经网络的河口潮位短期预报混合模型

河口潮汐过程受上游径流、外海潮波等因素影响，动力机制复杂，潮位预报难度大。本文提出了一种基于非稳态调和分析（NS＿TIDE）和长短时记忆（LSTM）神经网络的混合模型，对河口潮位进行12～48 h短期预报。该模型首先对河口实测潮汐数据进行非稳态调和分析，通过与实测资料对比得到分析误差的时序序列，并以此作为LSTM神经网络的输入数据，通过网络学习并预测未来12～48 h潮位预报误差，据此对NS＿TIDE的预测结果进行实时校正。利用该模型对2020年长江口潮位过程进行了预报检验，结果表明混合模型12 h、24 h、36 h和48 h短期水位预报的均方根误差（RMSE）相比NS＿TIDE模型至多分别降低了0.16 m、0.15 m、0.14 m和0.12 m；针对2020年南京站最高水位预测，NS＿TIDE模型预报误差为0.64 m，而混合模型预报误差仅为0.10 m。

Hybrid model for short-term prediction of tide level in estuary based on LSTM and nonstationary harmonic analysis

The tidal process in estuary is affected by comprehensive factors such as river discharge and astronomical tidal, as a result, the estuarine tide level is difficult to predict. In this paper, a hybrid model based on the nonstationary harmonic analysis model (NS_TIDE) and long short-term memory (LSTM) neural network is proposed to predict the estuarine tidal level for 12-48 hours. The hybrid model carries out nonstationary harmonic analysis of the measured tidal level firstly, and obtains the time series of the analysis error by comparing with the measured data. The analysis errors are considered as the input data of the LSTM neural network to predict the tidal prediction error in the next 12 h to 48 h, so as to correct the prediction of NS_TIDE in real time. The hybrid model is used to predict the tide level in the Changjiang Estuary in 2020. The results show that the root mean square error (RMSE) of the short-term water level prediction of the hybrid model at 12 h, 24 h, 36 h, and 48 h is reduced by up to 0.16 m, 0.15 m, 0.14 m, and 0.12 m respectively compared with the NS_TIDE model. For the prediction of the highest water level of Nanjing Station in 2020, the prediction error of the hybrid model is 0.10 m, while that of NS_TIDE is 0.64 m.