ABSTRACT
In present study, specific artificial intelligence based modelling techniques, along with nature inspired optimisation algorithms are utilised to model and optimise commercial high density polyethylene (HDPE) plants. Polyethylene is considered the most versatile plastic product with diverse applications in manufacturing products like plastic bottles, milk packets, buckets, cars, suitcases, etc. Commercial HDPE plants need to produce different grades of plastic as per market demand and profit margin. During grade changeover, lots of polymer gets wasted as they don’t conform to quality specs as they are produced during transition from one grade to another. Process parameters of HDPE reactors like temperature, pressure, ethylene flow, catalyst flow, ethylene to hydrogen ratio, etc. need to be adjusted to produce specific grade. As MFI is the most important quality parameter of the final polymer product, plant engineers target a particular MFI to produce a particular grade of polymer. In this study, two AI based modelling techniques, namely Artificial Neural Network (ANN) and Genetic Programming (GP) have been used to model MFI from reactor process parameter data. Historically, 3-years of commercial plant data has been used to develop MFI models. Two modifications were made to both ANN and GP on their basic algorithms to make them suitable to model complex chemical reactors. These modifications ensure that the developed model contains underlying physics and reaction kinetics and not only represents a data-driven blackbox. These modifications accomodate large amounts of industrial data along with their transmitter and process noise and inaccuracies. Commercial plant data and polymer plant operation experience has been utilised to identify the input parameters and train the models. The developed model was found to be very accurate (mean square error less than 1%) at coefficient of determination (R2 = 0.98). Once an adequate and reliable MFI equation is developed, in the next step, genetic algorithm was used to optimise reactor process parameters to achieve target MFI. Usually, MFI is measured by lab analysis which takes 4–6 hours, based on which, production engineers adjust the operating parameters to regulate the MFI. The current model helps the production engineer to get MFI value in real time without waiting for lab analysis. The second advantage of 169this model is that it recommends, in real-time basis, to the production engineer, the value of the process parameters to be kept in the reactor to achieve the target MFI. This model can be deployed in a plant control system (DCS) to indicate MFI of HDPE product in real-time basis. It also recommends the optimal reactor parameters which helps the production engineer to take corrective and preventive action to avoid losses from off-spec polymer products during grade changeover. This real-time application is an excellent example where artificial intelligence and metaheuristic optimisations are applied on commercial polymer plants to increase huge amounts of profit without any investment. The hybrid modelling and optimisation strategies developed in this work are generic and can be extended to any other chemical process and industry where first principle based models don’t exist.
