In this article we address over-excavation induced localized deformation in tunnels by proposing a real-time spatio-temporal prediction model that couples multi-layer LSTM for local temporal dynamics with Transformer self-attention to capture global cross-variable dependencies among multi-source environmental and displacement data. Key hyperparameters (learning rate, attention heads, hidden size, regularization) are jointly tuned via an Alpha Evolution (AE) global optimizer. On four CEC2025 benchmarks, AE converged faster and reached lower fitness than RIME, GWO, and SSA. Using 1,500 in-situ monitoring samples, the AE-optimized model reduced MAE by 32.9% and RMSE by 28.5%, and increased by 3.2% relative to its untuned counterpart, outperforming standalone Transformer, LSTM, and BP baselines. Case studies across varied lithologies and over-excavation regimes demonstrate stable accuracy and robustness, supporting short-term deformation forecasting and real-time early warning in complex geotechnical settings