Robot deformation errors caused by force exhibit strong nonlinearity and significant spatiotemporal variation, making accurate prediction challenging for traditional models. This paper proposes an incremental learning model based on GWO-XGBoost to improve prediction accuracy amid changing workspaces and operating times. First, Grey Wolf Optimization (GWO) tunes XGBoost hyperparameters to build an initial prediction model. XGBoost then maps 6D joint angles, 3D theoretical positions, and 3D force data to deformation errors. An online incremental learning mechanism dynamically updates the model with real-time data to adapt to load variations. Experiments show MSE drops from 0.758 to 0.416 and MAE from 0.587 mm to 0.346 mm under new loads, demonstrating strong fitting and predictive performance for enhancing robot deformation accuracy.