公路边坡失稳引发的滑坡等灾害问题对人们的生活影响重大，建立滑坡预测模型对地质灾害防治工作具有重大意义。以福建省南平市公路边坡为例，提出基于贝叶斯优化算法（Bayesian optimization algorithm, BOA）的长短期记忆网络（long short term memory, LSTM）模型，探究土压力变化对边坡稳定性的影响。该模型综合考虑了多个滑坡影响因素的相关性，尤其是各个影响因素之间的时间滞后性，并对滞后时间进行计算。对LSTM模型4个超参数（时间步长、隐藏元数量、迭代次数和批大小）进行自动搜索和寻优，解决了用LSTM模型建立滑坡预测模型时需要手动调参的问题。为了验证该模型的准确性和有效性，将该模型与循环神经网络（recurrent neural network, RNN）、LSTM模型进行对比，以南平市公路边坡土压力的监测数据集作为对照。结果表明：相对于其他两种模型，BOA-LSTM模型的均方误差（mean square error, MSE）、平均绝对误差（mean absolute error, MAE）和平均绝对百分比误差（mean absolute percentage error, MAPE）分别降低了52.4%和59.9%，28.8%和30.1%，30.2%和29.9%，预测精度从95%左右提高到96.56%，判定系数也更加接近于1，说明该模型可以更加准确地预测土压力的变化，为边坡变形稳定性分析提供有效的数据支撑。

Landslides and other disaster problems caused by highway slope instability have a significant impact on people's lives, so the establishment of landslide prediction models is of great significance to the prevention and control of geological disasters. Long short term memory(LSTM) model based on Bayesian optimization algorithm (BOA) WAs proposed to investigate the effect of soil pressure changes on slope stability. The highway slope in Nanping City, Fujian Province was taken as an example. The correlation of multiple landslide influencing factors were integrated in the model, especially the time lag between each influencing factor, and the lag time was calculated. The four hyperparameters of the LSTM model (time step, number of hidden elements, number of iterations, and batch size) were automatically searched and optimized, which solves the problem of needing to manually tune the parameters when building landslide prediction models with the LSTM model. To verify the accuracy and effectiveness of the model, the model was compared with recurrent neural network(RNN) and LSTM models, and the monitoring data set of soil pressure on highway embankments in Nanping City was used as a control. The results show that compared with the other two models, the mean square error(MSE), mean absolute error(MAE) and mean absolute percentage error(MAPE) of the BOA-LSTM model are reduced by 52.4% and 59.9%, 28.8% and 30.1%, and 30.2% and 29.9% respectively, and the prediction accuracy is improved from about 95% to 96.56%, and the coefficient of determination is also closer to 1, indicating that the model can more accurately predict the earth pressure change and provide effective data support for slope deformation and stability analysis.