This research proposes a Bayesian MPC framework for continuous chemical plant operations facing load variability and non-stationary time delays. The process model is built in a grey-box manner and supplemented with Gaussian Process-based residual learning to capture the model-plant mismatch and its uncertainties. Delays are modeled as time-varying variables through probabilistic estimation (multiple-model/particle-based), which are then integrated into a delay-aware predictor so that state propagation considers the delay distribution over the horizon. State estimation is performed using Unscented Kalman Filter (UKF) or Ensemble Kalman Filter (EnKF), while control decisions are derived from economic MPC with chance constraints to ensure that the risk of constraint violations remains below a predefined threshold. To make it suitable for real-time application on industrial platforms (cycle time ~2 s), we employ adaptive move-blocking, warm-start, and real-time iteration. Evaluation against three benchmarks—coordinated PID, deterministic MPC, and robust MPC—shows consistent performance improvement in scenarios with ±30% throughput change and non-stationary delays ranging from 2 to 8 minutes. Quantitatively, the proposed approach reduces daily economic costs by approximately 6.4% compared to deterministic MPC, decreases energy consumption, reduces off-spec rate to approximately 1.1%, minimizes constraint violations to ≈2 occurrences per 24 hours, and shortens settling time for grade changes to approximately 19 minutes. Ablation studies confirm the complementary contributions of residual learning, delay-aware predictors, and chance constraints to risk and cost reduction. These results confirm the readiness of implementing Bayesian MPC in modern DCS/SCADA for more reliable and cost-effective plant-wide operations.