基于长期监测数据与LSTM网络的滑坡位移预测

Prediction of Landslide Displacement Based on Long Term Monitoring Data and LSTM Network

滑坡位移变化是危险性的直接表征,位移预测对防灾减灾至关重要。以八字门滑坡为例,基于十年监测数据和神经网络模型(LSTM、RNN)进行滑坡位移预测。用一次移动平均法将总位移分解为趋势项和周期项,趋势项采用三次多项式函数进行分段拟合预测,通过神经网络模型和建立周期项与特征因子的关系并进行预测。其中,周期项特征因子根据位移影响因素初步选取,再通过Pearson相关性分析剔除无关因子。将预测的趋势项、周期项相加即为总位移预测值,对预测值与真实值进行误差分析,绝对误差为10 mm(LSTM)、24 mm(RNN),相关系数R;为0.9715(LSTM)、0.6675(RNN)。结果表明:LSTM在面对长时间序列时表现出更好的预测能力,该预测结果可以为八字门滑坡的防灾减灾工作提供理论参考。

Landslide displacement change is a direct representation of hazard.Displacement prediction is crucial to disaster prevention and mitigation.Taking the Bazimen landslide as an example,landslide displacement prediction was performed based on ten years of monitoring data and neural network models(LSTM,RNN).The total displacement is decomposed into a trend term and a periodic term using the one-time moving average method.The trend term is predicted by segmental fitting using a cubic polynomial function,and the relationship between the periodic term and the characteristic factor is established and predicted by the neural network model.Among them,the characteristic factors of the periodic term are initially selected according to the displacement influencing factors,and then the irrelevant factors are eliminated by Pearson correlation analysis.The predicted trend and period terms are summed to the total displacement prediction,and the error analysis is performed between the predicted and observed values.The absolute errors were 10 mm(LSTM)and 24 mm(RNN).The correlation coefficients R;are 0.9715(LSTM)and 0.6675(RNN).The results show that LSTM performs better in the face of long-time sequences.The prediction results can provide theoretical reference for disaster prevention and mitigation of the Bazimen landslide.