【摘要】 长短期记忆网络(LSTM)拥有较强的非线性拟合能力，近年来被较多用于基坑变形或沉降预测分析中.针对LSTM存在重要参数确定比较困难的问题，提出采用改进的粒子群算法(IPSO)对LSTM神经网络中的迭代次数、批处理大小、隐含层神经元数量进行寻优.改进的粒子群算法通过引入遗传算法中的变异机制，避免了粒子群算法(PSO)在前期寻优时陷入局部最优，同时利用非线性变化权重和改进学习因子的方法提高了PSO的寻优效率.利用提出的PSO-LSTM方法对实际基坑沉降进行预测分析，并将预测结果与PSO-LSTM、LSTM的预测结果进行对比，发现PSO-LSTM较LSTM、PSO-LSTM的平均百分比误差相对降低了56.47%、11.92%.验证了IPSO-LSTM对基坑地表沉降预测的准确性.更多还原

【Abstract】 The long-term and short-term memory network(LSTM) has strong nonlinear fitting ability, which has been widely used in the prediction and analysis of foundation pit deformation or settlement. However, there is a challenging issue of determining the important parameters in LSTM. In this study, an improved particle swarm optimization algorithm(IPSO) is proposed to optimize the number of iterations, batch size, and the number of hidden layer neurons in LSTM neural network. By introducing the mutation mechanism of genetic algorithm, IPSO avoids the local optimization of PSO in the early stage of the optimization, and improves the optimization efficiency of PSO by using the method of nonlinear weight and improving learning factor. The IPSO-LSTM is used to predict and analyze the actual foundation pit settlement, and a comparison with PSO-LSTM and LSTM is presented. It is shown that the average percentage error of IPSO-LSTM is 56.47% and 11.92% lower than that of LSTM and PSO-LSTM respectively, demonstrating great potential of IPSO-LSTM in accurately predicting the surface settlement of foundation pit.更多还原