Prediction of dynamic trends of landslides in fault zone areas based on time series analysis

The relationship between landslide states and variables exhibits time-varying characteristics due to geological processes. In fault zone areas, landslide monitoring data are often characterized by spatiotemporal discontinuity and high noise interference, which further exacerbate the uncertainty in dynamic trend prediction. Traditional methods overlook the long-term dependencies and cross-scale coupling mechanisms within time series, resulting in insufficient capability to capture critical landslide states. To address this issue, the author developed a method for predicting the dynamic trends of landslides in fault zone areas based on time series analysis. A landslide time series modeling function was established by integrating the variable factors that trigger landslides. An expanded causal convolution was introduced to capture the multi-dimensional variable characteristics within the receptive field of the function, outputting the variable features encompassing the entire landslide time series. The correlation coefficients between the local multi-variable features of the landslide and the landslide state were calculated, and the corresponding relationships were established using the mean correlation coefficients at different time points. By performing deep extreme learning on the variable features within the prediction period, landslide information was mapped and output based on these relationships. In the test results, the relative error of the landslide displacement prediction within the prediction period using the author's method remained stable within 3.0 mm, and the Gaussian distribution level of prediction errors in different deformation zones was close to 0, demonstrating high accuracy.