Answers are grounded in this paper and cite their sources.
GNSS monitoring data for highway slopes exhibit strong non-stationarity and high-frequency noise, making traditional deep learning models prone to error accumulation in long-sequence forecasting. This paper proposes a highway slope displacement forecasting model based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and a weighted ensemble of Bidirectional Long Short-Term Memory (Bi-LSTM) networks. OC-SVM and Simple Moving Average (SMA) are employed for outlier removal and noise reduction, while CEEMDAN decomposes the displacement sequence into a multi-channel feature matrix to suppress environmental interference. A Bi-LSTM incorporating a self-attention mechanism and Global Average Pooling (GAP) is then developed, and an inverse-error-based homogeneous weighted ensemble strategy is adopted for prediction. The model is validated using three-dimensional GNSS displacement data from a highway slope in Qingxiu District, Nanning, Guangxi. A total of 9000 E-W displacement samples (2-min interval) are used for evaluation. As the forecasting horizon increases from 1 to 12 h, prediction accuracy gradually decreases (R2 = 0.9577, 0.9122, and 0.8717) but consistently outperforms baseline models. When applied without retraining to the N-S and U displacement components recorded at the same GNSS monitoring station, the model also achieved R² values above 0.86. The proposed model effectively mitigates long-sequence forecasting divergence and provides technical support for dynamic highway slope monitoring.
No takes yet. Share an insight, caveat, or question.
Rao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: