Transient Thermal Response and Simulation-Driven Machine Learning Prediction for Temperature in a Petroleum Coke Vertical Shaft Calciner

Transient temperature evolution in a petroleum coke vertical shaft calciner (VSC) helps assess responses to operating disturbances, but transient multiphysics simulations are costly. A transient two-dimensional multiphysics model representing one pot and the flues on one side of a 24-pot, eight-layer co-current VSC, with prescribed flue temperatures, was developed to investigate step changes in flue temperature and per-pot discharge rate. Raising the upper-flue temperature by 200 K extended the high-temperature calcination zone from 1.3 to 1.9 m and shifted it upward. Increasing the discharge rate from 95 to 115 kg/h shortened this zone from 1.3 to 0.3 m and shifted it downward. The furnace wall approached a quasi-steady state within 10 h, whereas the coke bed required up to 40 h. Using 4.59 × 105 numerically generated temperature values from 510 pot-centerline locations under nine flue-temperature disturbance conditions, a Temporal Convolutional Network–Bidirectional Long Short-Term Memory–Convolutional Neural Network–Attention (TCLCA) surrogate model was developed. TCLCA achieved an R2 of 0.997 in recursive prediction while reducing computational time by approximately 94%. The framework enables rapid prediction of simulated centerline temperature evolution.