基于时间序列与时间卷积网络的滑坡位移预测

滑坡位移预测作为滑坡监测预警的重要组成部分，对滑坡灾害的防治具有重要意义。目前，滑坡位移预测大多集中在循环架构的神经网络模型上，其存在梯度爆炸、消失问题等问题。为此，提出了一种基于时间序列与时间卷积网络(time convolution network, TCN)的滑坡位移预测模型。首先，该模型通过移动平均法将滑坡位移分解为趋势项位移和周期项位移。其次，采用Holt线性趋势模型预测趋势项位移，并建立时间卷积网络预测周期项位移。最后，将趋势项位移和周期项位移叠加，实现滑坡位移的预测。将该模型用于八字门滑坡的观测研究，结果表明：该模型相较于循环架构的神经网络模型能更有效地提取时序特征，预测精度更高。将基于TCN的滑坡位移预测模型应用于滑坡位移预测具有广阔的应用前景。

A model for predicting landslide displacement based on time series and time convolution network

As an important part of landslide monitoring and early warning, landslide displacement prediction is of great significance to landslide disaster prevention. At present, the prediction of landslide displacement is mostly concentrated on the neural network model of circular structure, which has the problems of gradient explosion and disappearance. Therefore, a landslide displacement prediction model based on time series and time convolution network ( TCN ) is proposed. The model decomposes landslide displacement into trend displacement and periodic displacement by moving average method. Secondly, Holt linear trend model is used to predict trend term displacement, and time convolution network is established to predict periodic term displacement. Finally, the trend displacement and periodic displacement are superimposed to predict landslide displacement. The model is applied to the observation and study of Bazimen landslide. The results show that the model is more effective in extracting time series characteristics and higher prediction accuracy than the neural network model with cyclic architecture. The application of TCN-based landslide displacement prediction model to landslide displacement prediction has broad application prospects.