Heat Exchanger Performance Prediction Modeling using NARX-type Neural Networks

The Crude Preheat Train (CPT) in a petroleum refinery recovers waste heat from product streams
to preheat the crude oil. Due to high fouling nature of the fluids that flow through the exchangers, the
performance deteriorates significantly over time as less heat can be transferred through the fouling layers.
Prediction of the performance for optimal scheduling of the CPT operations requires a reasonably accurate
mathematical model. There are no guidelines for selecting relevant input variables and correct functional
forms for building theoretical models for such nonlinear systems. Neural Network (NN) offers the flexibility
to model complex and nonlinear systems with good prediction capabilities. In this paper, prediction models
using two different types of NNs are developed and compared for a heat exchanger to predict the change in the
outlet temperatures over time. The data required for model building were collected from plant historian in a
refinery. The data were processed for removal of outliers through Principal Component Analysis (PCA) and
the important input variables (predictors) were selected using Projection to Latent Structures (PLS). A
nonlinear auto-regression with exogenous inputs (NARX) type neural network model demonstrates its superior
prediction capabilities with a root mean square error of less than 2.5 oC in the outlet temperatures and
possesses a correct directional change index of more than 90%.