PTQ Q4 2026 Issue

Reboiler tube side parameters – ∆P, ∆T, Flow

∆T started decreasing after 21:30 and started uctuating

∆T started increasing after 21:30 and started uctuating

∆P ∆T Flow

06-09 18

06-09 19

06-09 20

06-09 21

06-09 22

06-09 23

06-10 00

Timestamp

Figure 3 Onset of thermosyphon loss

reboiler Reb-A: As the liquid level in the column bottom dropped, the static head available to drive the thermosi- phon flow to the reboilers diminished. At around 21:30, the process parameters for Reb-A began to show signs of distress. The temperature difference (ΔT) across the tube side started to fluctuate wildly and decrease (see Figure 3 ), while the pressure difference (ΔP) increased, indicat - ing a loss of stable heat exchange and thermosiphon loss. Thermosiphon loss on the shell side reduced heat transfer, causing incomplete condensation in the tube side, which resulted in two-phase flow. Two-phase flow has higher fric - tion, which increases tube-side pressure drop. This can be interpreted by the Lockhart-Martinelli relation:2 ΔP two_phase = ϕ L 2 . ΔP liquid_only where ϕ L 2 increases as vapour fraction increases.

propagated to the overhead system. The flow to the HPC became highly erratic and began to drop significantly. At 07:36 on the next day (trip-day), the suction flow fell below the low-low trip point, and the HPC tripped to protect itself from damage, bringing the entire unit to shut down. The investigation confirmed that the entire cascade was triggered by the initial, undetected malfunction of the LT. Development of the predictive models To address the vulnerabilities exposed by the incident, a multi-pronged AI/ML strategy was devised. The goal was not just to predict the final HPC trip but to detect the earli - est deviations from normal operation, providing a hierarchy of alerts as discussed in the following three models: Model 1: Overall process upset prediction The flagship model was designed to act as a general early warning system for the entire C3 splitter unit. It uses logistic regression, a robust and interpretable classification algo - rithm, to predict the probability of a process upset. • Methodology: The model takes a wide range of process variables (tags) as inputs, including flows, temperatures, pressures, and controller outputs from across the C3 split- ter system. The core of the logistic regression model is the logistic function (or sigmoid function), which transforms a linear combination of these input features into a prob- ability score between 0 and 1 using the logistic transfor - mation as: P(z) = 1 Where z = β 0 + β 1 x 1 + ⋯ + β n x n 1+ⅇ -(z) Here P(z) is the probability of an upset, χ i represents the input process variables, and β i are the weights learned by the model during training. These weights signify the con - tribution of each variable to the probability of an upset. The goal of training logistic regression is to find the values of β i that minimise this cost function L( β ). This is done using gradient descent optimisation algorithms:

As shell-side boiling becomes unstable: • Vapour fraction fluctuates. • ϕ L 2 changes. • ΔP oscillates.

This explains the initial ΔP rise and later large oscillations.  Loss of thermosiphon in reboiler Reb-B: Shortly after Reb-A failed, reboiler Reb-B also began to experience ther- mosiphon loss, starting around 22:15.  HPC disturbance and trip: With both reboilers failing to provide consistent boil-up, the heat and material balance of the entire column was severely disrupted. This instability The goal was not just to predict the final HPC trip but to detect the earliest deviations from normal operation, providing a hierarchy of alerts

PTQ Q4 2026