Aiming at the lack of multi-step prediction methods for soil landslide displacement and the issue of significant prediction errors over extended time horizons, this paper proposes a multi-step landslide displacement prediction method based on a parallel model with a multi-level attention mechanism. The method employs a multi-input multi-output prediction strategy. Utilizing a Transformer encoder branch incorporating a multi-head attention mechanism and a bidirectional gated recurrent unit (BiGRU) branch optimized with a global attention mechanism (GAM), the two parallel network branches process historical landslide monitoring data. The landslide feature information extracted by the parallel networks is then fused via a cross attention mechanism (CAM), subsequently outputting the predicted multi-step displacement values. Experimental results demonstrate that the multi-level attention mechanism model achieves a mean absolute error (MAE) of 2.17 mm, a root mean square error (RMSE) of 3.05 mm, and a coefficient of determination (R2) of 0.968 9 in multi-step landslide displacement prediction. Compared to other models, it yields the lowest errors and the optimal R2 result. The model exhibits more stable prediction performance over long time horizons, facilitating the early anticipation of landslide development trends. This provides crucial technical support for landslide prevention and mitigation.