水闸深基坑开挖变形具有明显的非线性和非稳定性特征，基于此，引入极点对称模态分解算法（extreme-point symmetric mode decomposition method，ESMD）对水闸深基坑开挖变形原型监测序列进行多模态分解，并基于模糊熵（fuzzy entropy, FE）理论对各分解分量进行模糊多模态相空间重构，从而有效甄别水闸基坑变形不同时间尺度有效物理特征。构建基于人工水母搜索算法（artificial jellyfish search optimizer，AJSO）优化的长短期记忆（long short-term memory，LSTM）人工神经网络模型，以重构后的各重构子序列为基础进行优化训练，并把训练后的各预测模态分量合并，实现对水闸基坑开挖变形动态预测和分析。以张家港市十一圩江边枢纽改建工程基坑开挖变形监测为例，采用上述方法对该枢纽工程基坑开挖过程变形进行预测和分析。结果表明：基于ESMD-FE-AJSO-LSTM算法的水闸深基坑变形预测方法能够有效预测基坑开挖变形非线性特征，相比传统LSTM、循环神经网络（recurrent neural network, RNN）和支持向量机（support vector machine，SVM）等算法具有更高的预测精度和稳定性，为实现对基坑开挖安全性态实时科学诊断和分析提供技术参考。

The excavation and deformation of deep foundation pits for sluice gates were influenced by various factors, including engineering and hydrogeological conditions, spatial dimensions of the foundation pit, type of support structure, and excavation stage. Additionally, random environmental factors such as vibrations from construction machinery, loads from surrounding traffic, and weather conditions play a role. The excavation-induced deformation of these foundation pits exhibited significant nonlinearity and instability. The deformation monitoring data acquired from the excavation site of the foundation pit consist of a series of multimodal sequences across various temporal dimensions. Scientifically identifying key data features in different dimensions and subsequently modelling and predicting them in a targeted manner holds significant importance. The extreme-point symmetric mode decomposition method (ESMD) was employed for the prototype monitoring sequences of deformation during the deep excavation of a sluice foundation pit, involving multimodal decomposition. The deformation monitoring data of the sluice foundation pit were decomposed into several distinct subsequence components, intrinsic mode functions (IMFs), and trend components (Res), each exhibiting unique features. Fuzzy entropy (FE) theory was utilized, and fuzzy multimodal phase space reconstruction was applied to each modal subsequence component and trend component, resulting in multiple reconstructed subsequence components. Physically significant features of sluice foundation pit deformation at various time scales were effectively discerned through this process. Subsequently, a model was constructed based on the artificial jellyfish search optimizer (AJSO)-optimized long short-term memory (LSTM) artificial neural network. The optimization involved training on the reconstructed subsequences, yielding an AJSO-LSTM optimized model for each reconstructed subsequence. Finally, using the optimized AJSO-LSTM models, dynamic predictions were made for each reconstructed subsequence at fixed time intervals. The predicted results for each reconstructed subsequence were synthesized to obtain the overall prediction of foundation pit deformation. To evaluate the prediction accuracy of the ESMD-FE-AJSO-LSTM model for foundation pit deformation, multiple accuracy evaluation metrics were introduced.Taking the excavation deformation monitoring of the Eleven Weir Riverbank Hub Reconstruction Project in Zhangjiagang city as an example, the methods described above are employed to predict and analyse the excavation-induced deformations in the hub project. The results indicate that this approach can effectively forecast the nonlinear characteristics of excavation-induced deformations. The multidimensional feature scale components obtained through the ESMD algorithm exhibit distinct physical oscillation characteristics. Simultaneously, the calculation results of the variance contribution rate for each mode indicate that the short-term fluctuation in the deformation of the sluice foundation pit is mainly dominated by the high-frequency modes IMF1 and IMF2, while the long-term fluctuation is primarily governed by the trend component Res. The consistency between these decomposition results and on-site observations demonstrates that the ESMD method is effective at identifying the physical characteristics of excavation-induced deformation at different time scales. The proposed ESMD-FE-AJSO-LSTM method achieves an overall deformation prediction accuracy ranging from 97.63% to 99.52%. The prediction results generally fall between those of the AJSO-LSTM, LSTM, RNN, and SVM algorithms, indicating that the ESMD-FE-AJSO-LSTM model presented has higher predictive accuracy. The predicted residuals of the ESMD-FE-AJSO-LSTM method fluctuate near the zero-value mean and exhibit an overall normal distribution. This finding suggested that the proposed method has better predictive stability and robustness than the other four models, indicating its practical value in scientific and engineering applications.