Abstract:
Displacement prediction is vital to slope failure monitoring and early warning. We took a slope in Wenzhou Belt Highway as a case and proposed a novel multivariate slope displacement prediction method based on graph deep learning and Beidou monitoring. First, we modeled the graph structure of the entire displacement monitoring system based on the spatial distance of Beidou high-precision monitoring points and built the adjacency matrix of graph nodes. Then, we performed time series data processing, including denoising, interpolation, and normalization on Beidou high-precision displacement, accumulative rainfall, groundwater table, and soil moisture content, and applied multivariate spatial-temporal correlativity analysis. The analysis reveals that the displacement is mainly controlled by the two consecutive months' rainfall, groundwater table, and soil moisture content in the Grade Ⅲ slope. Lastly, we introduced the state-of-the-art graph deep learning GTS(Graph for Time Series) model into the slope displacement prediction and built the GTS-BDS model specifically for Beidou high-precision slope displacement monitoring. The GTS-BDS model outperforms others such as LSTM and T-GCN in Root Mean Square Error(RMSE, 0.301), Mean Absolute Error(MAE, 0.154), and Mean Absolute Percentage Error(MAPE, 3.5%)evaluation metrics with a prediction horizon of 1 hour. The displacement prediction method proposed in this study takes full advantage of the spatial topology between monitoring points and temporal features of the monitoring data, thus improving the accuracy and reliability of displacement prediction from the perspective of the entire displacement monitoring system. The new method may have a favorable application in slope failure early warning.