PTQ Q2 2024 Issue

1150

1100 1150

1050

1050

1000

950

950

37 4 7 101316 22 25 28 31 34 40 19 1 850

900

52 43 46 49

1

3

5

7

9

11

13

15

25 17 19 21 23 27

33 29 31

Run length

Run length

Actual TMT

Predicted TMT

Actual TMT

Forecasted TMT

Figure 1 Actual vs predicted TMT

Figure 2 Scenario 1

processed by the AI model, and optimal values that need to be set in the real process are recommended. A key point is that the optimal values will never violate the acceptable operational ranges of suggested critical parameters. AI algorithms equipped with real-time data analysis capabilities play a crucial role in predicting and proactively addressing potential furnace issues. By dynamically adjust- ing parameters and ensuring adaptive control, the AI system prevents runway TMT and mitigates stress on critical com- ponents. This dynamic optimisation lays the groundwork for extended operational runs, minimising the interruptions caused by unplanned downtime. AI/digital twins symbiosis AI-driven optimisation plays a key role in ensuring stable fur- nace operation with reduced delta pressure and dilution steam, as well as an optimised feed ratio. This leads to a more con- trolled furnace temperature, contributing to operational stead- iness. The nuanced control provided by AI not only enhances operational efficiency but also contributes significantly to the overall stability of the ethylene production process. The discussion takes an intriguing turn with the introduc- tion of advanced process modelling, including the concept of digital twins. This innovative approach enables virtual test- ing and scenario optimisation, allowing operators to explore various conditions without impacting the physical furnace. The symbiosis of AI and digital twins not only optimises cur- rent operational parameters but also lays the groundwork for future advancements in ethylene production processes. However, challenges persist, and historical data serves as both a hurdle and a guiding light. The initial phase of AI

model building includes critical process parameters identifi - cation, preliminary investigations, pattern discovery, anom- aly spotting, hypothesis testing, and establishing correlations between process parameters and run length by considering TMT as the primary target. AI leverages historical data to train the run length optimi- sation model. Coking in the radiation coil is inevitable at high cracking temperatures. As TMT reaches maximum threshold values, the furnace undergoes decoking, compromising run length. Furnace run length is directly linked to coke forma- tion rate, with operational parameters such as TMT, dilution steam ratio, firing rate, feed composition, feed rate, wall and floor burners in operation, excess oxygen %, coil pressure ratio (CPR), coil outlet pressure and venturi ratio determining the end of furnace runs. In response to these intricacies, the furnace is operated, and parameters are controlled to minimise the rate of increase in TMT throughout the run. This proactive approach mitigates the rise in TMT, ultimately contributing to longer furnace run lengths and improved operational efficiency. TMT prediction model The CokeNil TMT prediction model has demonstrated a com- mendable accuracy rate of 98.2%, underscoring the efficacy of our AI system. This high level of precision substantiates the model’s reliability and underscores its potential impact in practical applications. It is imperative to recognise that model success extends beyond accuracy alone, encompass- ing aspects such as generalisation to unseen data and robust performance across diverse scenarios. Furthermore, a holistic assessment that includes metrics like precision, recall, and F1 score contributes to a more com- prehensive evaluation of the model’s efficacy. The observed effectiveness of the base model in predicting TMT for opti- mal parameter recommendation holds promise for positively influencing associated processes and systems. Continuous monitoring, evaluation, and potential retraining with new data are pivotal for sustaining and improving model performance. Figure 1 shows the results of the TMT prediction model. Whenever a new furnace run starts, the CokeNil model captures critical parameters, which are the base for the model, and generates optimal values for subsequent furnace runs. The generated optimal values are typically in the acceptable operational range, which the plant operator can incorporate into the actual process and see the TMT improvement.

1100 1150

1050

1000

950

900

1

4

7

19 10 16 22 13

46 25 31 37 43 49 28 34 40 Run length

52

55 61

58 64

67

70

Actual TMT

Forecasted TMT

Figure 3 Scenario 2 where furnace run length is extended by 18 days

16

PTQ Q2 2024

www.digitalrefining.com

Powered by