Obtaining reliable and high-accuracy prediction results of displacement trends in the long term is crucial for mitigating geohazards. Deep learning can capture dynamic and nonlinear characteristics in long-term time series and is widely used in landslide displacement prediction. However, its black-box attribute prevents decision-makers from understanding the basis of the model output, which limits the application of the model in the final optimization scenario. Thus, a novel “decomposition-prediction-addition-interpretative” framework including variational mode decomposition (VMD), double exponential smoothing (DES), and a long short-term memory network with spatiotemporal attention mechanism (TSAM-LSTM) is proposed. It enables high-precision multi-step prediction and spatiotemporal dimension interpretability analysis. Therein, VMD decomposes the total displacement into the trend, periodic, and random displacement, and the random displacement is further decomposed by VMD. On this basis, DES is used to predict the trend displacement, TSAM-LSTM to predict the periodic and random displacements, and finally, all the predicted values are superimposed to realize the total displacement prediction. The performance of the proposed approach was validated using monitoring data from H8 accumulation, Huangzangsi water conservancy project. The results indicate that the VMD-DES-TSAM-LSTM can achieve satisfactory prediction results. The introduction of TSAM improves the generalization ability of the traditional LSTM and significantly enhances its prediction performance. Meanwhile, TSAM accurately reveals the most relevant temporal and spatial information contained in input data that affects target displacement and visualizes the attention focus during model training, which provides a more favorable basis for model optimization and disaster prevention decision-making.

获得可靠且高精度的长期位移趋势预测结果对于减轻地质灾害至关重要。深度学习可以捕捉长期时间序列中的动态和非线性特征，广泛应用于滑坡位移预测。但其黑箱属性使得决策者无法了解模型输出的依据，从而限制了模型在最终优化场景中的应用。因此，提出了一种新颖的“分解-预测-加法-解释”框架，包括变分模式分解（VMD）、双指数平滑（DES）和具有时空注意机制的长短期记忆网络（TSAM-LSTM）。它能够实现高精度的多步预测和时空维度可解释性分析。其中，VMD将总位移分解为趋势位移、周期位移和随机位移，随机位移又进一步被VMD分解。在此基础上，采用DES预测趋势位移，TSAM-LSTM预测周期性位移和随机位移，最后将所有预测值叠加，实现总位移预测。利用黄藏寺水利枢纽H8堆积的监测数据验证了所提方法的性能。结果表明VMD-DES-TSAM-LSTM能够取得满意的预测结果。TSAM的引入提高了传统LSTM的泛化能力，显着增强了其预测性能。同时，TSAM准确地揭示了输入数据中包含的影响目标位移的最相关的时空信息，并将模型训练过程中的注意力焦点可视化，为模型优化和防灾决策提供了更有利的基础。