<p>Forecasting Earth surface displacements is critical for monitoring tectonic processes, mitigating geo-hazards, and ensuring infrastructure resilience. This study systematically evaluates advanced deep learning architectures for GNSS displacement prediction by integrating meteorological drivers, temporal encodings, and spatial information from neighboring stations. Daily displacement time series from 19 continuous GNSS stations in California were combined with ERA5 reanalysis meteorological data. A range of models was evaluated, including Long Short-Term Memory (LSTM), Stacked LSTM, Bidirectional LSTM (Bi-LSTM), Attention-based LSTM, CNN-LSTM, Regularized LSTM, and a Temporal Convolutional Network (TCN). Model performance was assessed using RMSE, MAE, and R² under two scenarios: a baseline configuration using meteorological and spatial features with global scaling, and an enhanced configuration incorporating temporal encodings and station-specific normalization. Results demonstrate that integrating temporal features and local normalization substantially improves predictive accuracy. Among the evaluated models, Bi-LSTM, Attention-LSTM, and Stacked-LSTM achieved the best performance, with R² values exceeding 0.8, while the baseline LSTM consistently underperformed. Classical models, including persistence and ARIMA, also showed competitive performance, reflecting strong temporal autocorrelation and linear dependencies; however, they were unable to capture the full complexity of the signal. In contrast, deep learning architectures provided more robust and accurate forecasts by effectively modeling nonlinear spatiotemporal interactions and environmental influences. Overall, the findings highlight that accurate GNSS displacement forecasting depends not only on model architecture but also on the effective integration of spatial, temporal, and environmental information, demonstrating the value of data-driven approaches as a complement to traditional geophysical modeling.</p>