A method for predicting top-of-foundation-pit deformation was proposed by optimizing the control parameters of variational mode decomposition (VMD) using particle swarm optimization (PSO) and combining echo state networks (ESN) with a self-attention mechanism. Firstly, the PSO was used to optimize the control parameters of VMD, decomposing the deformation sequence data of the foundation pit top into different intrinsic mode functions (IMF). According to the different frequency characteristics, the deformation sequence was decomposed into seasonal, trend, and random components. The self-attention mechanism was combined with ESN to model the reconstructed long-term sequence data, and different input data lengths are compared and selected to determine the optimal input length, thereby improving prediction accuracy. The proposed method was validated using deformation monitoring data from a foundation pit in Guangzhou. Experimental results show that with an input step of 3, the method achieves an MSE （mean square error）of 0.62 and an R2 of 0.986, significantly improving the accuracy and stability of top-of-foundation-pit deformation prediction.