With the development of large-scale power grid infrastructure, higher requirements have been imposed on the insulation performance of high-voltage cross-linked polyethylene (XLPE) cables to ensure the safe and stable operation of electric power transmission systems. Insulation eccentricity is a critical quality indicator that directly affects the reliability and safety of power transmission. To address the issues of adjustment lag and reliance on manual experience in production, this paper proposes a real-time prediction method for cable insulation eccentricity. First, a feature selection approach combining Fast Correlation-Based Filter (FCBF) and Principal Component Analysis (PCA) is employed to extract key multi-source process parameters, forming a low-dimensional time-series dataset. Subsequently, an improved Transformer model incorporating a Cross-Variable Attention Mechanism (CVAM) is introduced, which enhances the model’s ability to capture long-term temporal dependencies and dynamic inter-variable relationships. Finally, a hybrid multi-step sliding window strategy is applied for real-time forecasting. Experimental results demonstrate that the proposed feature selection method significantly improves prediction accuracy and model generalization. The enhanced Transformer model effectively addresses spatiotemporal inconsistencies in industrial data, showing strong performance in complex time-series prediction tasks.