In daily operations, IT operations centers (IT-OC) within contact centers manage numerous incident tickets triggered by service disruptions across IT subsystems. Such incidents are handled reactively, leading to delayed responses and operational inefficiencies. This study proposes an improved incident forecasting framework by applying predictive maintenance (PdM) with deep learning models such as long short-term memory (LSTM), gated recurrent units (GRU), and transformers to shift from reactive ticket handling to proactive failure anticipation. Considering over 39,000 real-world tickets enriched with priority and impact attributes, this study compares the forecasting performance and computational efficiency of the three models under various data balancing strategies. The results show that LSTM-SMOTE achieves the highest F1 score (80.34%) and AUC (89.97%), while GRU-SMOTE yields a competitive F1 score (80.00%) and AUC (89.27%) with faster training and a smaller model. A transformer model, despite its theoretical ability to capture long-term dependencies via self-attention, exhibits limited performance under class-imbalanced conditions. To ensure real-world applicability, a benefit cost ratio (BCR) is determined, incorporating the training time, the amount of GPU memory used, and model size. The GRU-SMOTE model achieves a higher BCR (5.02) than does LSTM-SMOTE (4.66), indicating better computational efficiency. This multicriteria evaluation highlights GRU as a balanced alternative for real-time forecasting in resource-constrained environments. The proposed approach has the potential to support proactive failure detection, SLA compliance, and optimized IT resource allocation, promoting scalable and intelligent IT-OC operations in contact center infrastructures. Findings are based on multi-season production data from a single provider; external validation across additional organizations is left for future work. © 2013 IEEE.