Date of Award

25-1-2024

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.S.Jayalalitha

Keywords

Machine Learning Technique, Adaptive MPC, Heat Exchanger Fouling, Dual Extended Kalman Filter, Linear Parametric Varying System

Abstract

Fouling is an undesirable and inevitable phenomenon that affects Heat Exchanger (HE) dynamics, increases maintenance cost (>50%), and loss of production. Conventional fouling prediction and control techniques demand an in-depth understanding of the HE-specific fouling dynamics. Hence, machine learning emerges as an ideal tool for learning the fouling process from the available measurement dataset.

The proposed work aims to design a machine learning-driven adaptive control technique for an industrial HE, subjected to varying fouling conditions. Naphtha cooler, an industrial HE utilized in petroleum refineries, employs cooling water to lower Naphtha temperature, but the high impurities in the recycled cooling water leads to fouling. In this work, a Linear Parameter-Varying (LPV) model with Fouling Resistance (FR) as the drive parameter is developed using the data acquired from the Naphtha cooler at different fouling conditions. LPV model shows a model fit percentage of 89.30%, which is higher than other standard models.

Random Forest (RF) integrated Long-term Short-term Model (LSTM) is developed to predict FR. Experimental investigation shows good performance with the coefficient of determination ( ) of 0.9770. Robust analysis indicates the proposed RF-LSTM to predict FR accurately even with noisy measurements. Dual Extended Kalman Filter (DEKF) is modified to include RF-LSTM as a guiding input for FR estimation. Multiobjective Genetic Algorithm (GA) is used to tune the weights for the modified FR prediction. Five different fouling conditions are used to evaluate the FR estimation performance and show an improvement of FR estimation by 38.49% on average.

A novel Iterative Quality Weighted Interpolation (IQWI) is proposed to derive the model parameters from the LPV. It is based on model quality to determine the optimal model parameters and a 24.44% improvement in model parameter estimation is observed. Model parameters from LPV are used to define an accurate prediction model and are used by adaptive MPC to effectively control the HE. Comparative analysis illustrates the reliable performance of the adaptive MPC under FR variations. Adaptive tuning of MPC parameters like horizons, weights, and constraints based on the HE fouling condition will be investigated in the future.

Download

Share

COinS