1.0
LT disturbance start Probability Threshold Early detection
0.8
There was increase in probability 30 minutes before LT going bad and 6 hours before rst instance of reboiler instability
One spike in probability at 08:00 on Trip-previous -d ay , which was 8 hours before LT malfunction
0.6
08:00
16:00
0.4
0.2
0.0
06-08 00
06-08 06
06-08 12
06-08 18
06-09 00
06-09 06
06-09 12
06-09 18
06-10 00
06-10 06
Timestamp
Figure 4 Predicted upset probability highlighting early detection capability
entire dataset, the model showed its true power. It gener- ated intermittent spikes in upset probability starting from a day or week ago, well before the incident. Crucially, Figure 4 showed a significant probability spike at 08:00 on the day of upset, around eight hours before the LT began to malfunction, and a continuous increase in probability from 15:30, providing hours of potential warning. The confusion matrix, as shown in Table 1 , indicates the model performance, where it captured all anomalies, 100% correct anomaly prediction and only 0.02% of the normal data were wrongly identified as anomalous. Model 2: Dynamic mass balance and level prediction This model was developed specifically to counteract the pri - mary failure mode of the incident: the unreliable LT without redundancy. It uses first principles of mass balance to pro - vide an independent, calculated value for the column’s liquid inventory and level. • Methodology: The model establishes a system boundary around the C 3 splitter column and calculates a real-time mass balance:
Model accuracy on validation dataset
Pred normal
Pred anomalous
Actual normal
88,060
18
Actual anomalous
0
465
Table 1
• Data preparation and training: Data fetching: Historical data was extracted from the plant’s data historian at a two-minute resolution. Event labelling: The data was labelled based on the known incident. The period from the LT malfunction (16:00 Trip-previous-day) to the HPC trip (07:36 hrs Trip-day) was labelled as ‘bad’ (Class 1). All other periods of normal operation were labelled as ‘good’ (Class 0). Class imbalance: This resulted in a highly imbalanced dataset, with far more ‘good’ data than ‘bad’ data (a ratio of approximately 100:0.7). To prevent the model from being biased towards the majority class, the ‘good’ data was down-sampled by averaging it over 15-minute intervals, bringing the ratio to a more manageable 9:1.
Accumulation = Mass_In - Mass_Out
Where: • Mass_In: Calculated from the feed flow meter.
Train-test split: The bal- anced dataset was split into a 70% training set and a 30% testing set to evaluate the model’s performance on unseen data. • Performance: The logistic regression model demon- strated exceptional perfor- mance, achieving perfect scores (1.0) for precision, recall, and F1 score on both the training and test datasets. This indicated that it could perfectly distinguish between the normal and upset condi- tions present in the historical data (see Figure 4 ). When validated against the
Trendline of mass balance closure
1.15
1.10
1.053
1.05
Mean: 1.011
1.00
0.968
0.95
2stdev 1stdev Mass balance closure
0.90
0.85
2024-01
2024-02
2024-03
2024-04
2024-05
2024-06
2024-07
2024-08
Timestamp
Figure 5 Historical mass balance closure
61
PTQ Q4 2026
www.digitalrefining.com
Powered by FlippingBook