Landslides along China’s Three Gorges Reservoir now face a new computational adversary: a hybrid deep learning model called GRU–TimeMixerKAN that its developers report consistently outperformed nine rival architectures in forecasting landslide displacement. In a study published in AI in Civil Engineering, researchers Bingheng Li, Hong Zheng, and Xun Zhang trained the model on monthly GPS monitoring records from three reactivated reservoir landslides—Shuping, Baishuihe, and Bazimen—and found it achieved the lowest root mean square error and mean absolute error in every case, alongside the highest coefficient of determination among the models compared.
The Three Gorges Reservoir Area presents a notoriously difficult forecasting environment. Since impoundment began in 2003, landslide activity has intensified, reactivating ancient slope failures such as the Qianjiangping landslide in 2003 and the Gongjiafang landslide in 2008. Reservoir landslides in this region typically display step-like deformation, driven by the coupled effects of seasonal water-level fluctuations and extreme rainfall. These patterns are marked by pronounced nonlinearity, hysteresis effects, and diverse spatiotemporal responses, which have long frustrated both physically based models, which are sensitive to geological parameters and generalize poorly, and conventional data-driven techniques, which rely on static regression and struggle to capture temporal dynamics.
The new architecture combines three complementary components. A Gated Recurrent Unit, or GRU, serves as the sequential memory backbone, recursively aggregating historical observations into a compact hidden representation that encodes both short-term fluctuations and longer-term lag effects. A TimeMixer module, originally proposed by Google DeepMind in 2023, then decouples temporal dependency modeling from feature interaction modeling, allowing the network to treat smooth trends, periodic seasonal oscillations, and abrupt pulses as distinct patterns rather than conflating them. Finally, a Kolmogorov–Arnold Network, or KAN, introduced by MIT researchers in 2024, replaces traditional linear weight matrices with learnable nonlinear combinations of univariate functions, enhancing the model’s capacity to capture high-order interactions between hydrological drivers and displacement response.
Before training, the team applied a differencing-based detrending technique to both displacement and reservoir water-level series. Rather than decomposing cumulative displacement into trend and periodic terms—a conventional approach the authors note rests on the implicit and potentially biasing assumption that the trend component is independent of external stimuli—differencing computes the change between consecutive observations, eliminating long-term variability while stabilizing the series mean. Predicted increments are later reconstructed to absolute displacement values by cumulatively adding them to the last observed displacement, preserving physical interpretability. Sliding-window sampling converts the continuous monitoring record into supervised learning samples, and Z-score standardization places all features on a comparable scale.
The study drew on a 72-month record spanning January 2007 to December 2012, comprising monthly rainfall, reservoir water level, and displacement measurements. Eighty percent of the data trained the models, with the remaining 20 percent—October 2011 through December 2012—reserved for testing. At the Shuping landslide, a large accumulation landslide in Zigui County roughly 47 kilometers downstream of the Three Gorges Dam, station ZG85 provided the representative record. The Baishuihe landslide, also in Zigui County and about 56 kilometers upstream of the dam, contributed data from station ZG118, while the Bazimen landslide on the Xiangxi River in Badong County supplied records from station ZG111. All three sites used Trimble GPS receivers with a planar accuracy of 5 millimeters plus 1 part per million.
Training employed the Adam optimizer with weight decay set to 1e−4 and gradient clipping capped at a maximum norm of 1.0. The Huber loss function, which applies squared-error penalties for small deviations and absolute-error penalties for large ones, served as the optimization criterion, offering robustness to outliers in the heavily disturbed displacement series. An early stopping mechanism halted training if validation loss failed to improve over 20 consecutive epochs, and the Optuna framework automated hyperparameter search using a Tree-structured Parzen Estimator sampler with median-based pruning. Candidate settings spanned batch sizes from 9 to 72, learning rates from 1e-3 to 5e-4, hidden dimensions from 8 to 128, and dropout rates from 0.1 to 0.5.
The benchmark comparison covered nine architectures: GRU and its attention-augmented variant, LSTM and an LSTM–TimeMixerKAN hybrid, TimeMixer, TimeMixerKAN, CNN, MLP, and a support vector machine. At Shuping, GRU–TimeMixerKAN recorded an RMSE of 29.85, an MAE of 20.54, and an R² of 0.853, fitting both stable deformation phases and the sharp displacement increase in mid-2012. TimeMixerKAN followed with an RMSE of 34.60 and R² of 0.803, while the standard GRU achieved an R² of 0.778. TimeMixer posted the lowest MAE of 19.47 but a moderate RMSE of 37.41. MLP and CNN fared worst, with R² values of 0.418 and 0.276 respectively, and SVM reached an R² of 0.346.
Results at Baishuihe followed a similar pattern. GRU–TimeMixerKAN attained the lowest RMSE at 8.53 and lowest MAE at 6.66, with the highest R² at 0.763, accurately reproducing sharp increases and declines across multiple periods. GRU–Attention and GRU performed respectably with R² values of 0.711 and 0.704, while CNN ranked last at 0.364. At Bazimen, where step-like deformation was most pronounced, the proposed model achieved an RMSE of 13.65, an MAE of 10.97, and an R² of 0.895, capturing peak magnitudes and trend directions during the extreme displacement phases around 2012. LSTM–TimeMixerKAN ranked second there with an R² of 0.743, and CNN and GRU–Attention fell to the bottom with R² values of 0.328.
Computational efficiency also favored the new design. Training times remained on the order of seconds across all three cases, comparable to or lower than several recurrent and hybrid baselines, and faster than LSTM–TimeMixerKAN and GRU–Attention in the Shuping and Baishuihe cases. Because Optuna-based hyperparameter optimization runs offline, deployment requires only the final tuned model for inference, keeping the operational burden minimal for near-real-time early warning applications.
The authors note that their framework currently relies on large-scale environmental variables—precipitation and reservoir water levels—and on individual monitoring points, limiting its capacity to capture spatial interactions across a deformation field. They identify several priorities for future work: incorporating seismic disturbances, groundwater fluctuations, and subsurface moisture; integrating distributed sensor networks and remote sensing data such as InSAR; exploring transfer learning for newly instrumented sites with limited records; and combining physics-informed constraints with the data-driven architecture. They also suggest that stable performance requires training data spanning multiple deformation cycles and at least one complete annual hydrological period, since underrepresentation of key deformation stages or seasonal cycles degrades model performance.
Subject of Research: A hybrid deep learning model for forecasting displacement of reservoir-induced landslides in the Three Gorges Reservoir Area.
Article Title: A novel GRU–TimeMixerKAN hybrid model for robust landslide displacement forecasting: a case study of the Three Gorges Reservoir Area
Article References: Li, B., Zheng, H., & Zhang, X. (2026). A novel GRU–TimeMixerKAN hybrid model for robust landslide displacement forecasting: a case study of the Three Gorges Reservoir Area. AI in Civil Engineering, 5(1), Article 9. https://doi.org/10.1007/s43503-026-00091-z
Image Credits: AI Generated
DOI: 10.1007/s43503-026-00091-z
Keywords: landslide forecasting, deep learning, GRU, TimeMixer, Kolmogorov–Arnold Network, Three Gorges Reservoir, displacement prediction, reservoir water level, rainfall triggers, early warning, time series modeling, Optuna
News Source: Violet Maxwell. (October 30, 2026). Hybrid AI model forecasts landslide displacement in Three Gorges Reservoir. Scienmag.
