Drift analysis on neural network model of heat exchanger fouling

Neural Networks (NN) provide a good platform for modeling complex and poorly understood systems in many different fields. Due to the empirical nature of NN, it is typically valid only for small operating windows. As the process drifts, the prediction accuracy of such models deteriorates very much rendering the models unfit. An on-line mechanism to follow the drift in the process is necessary in order to retrain the NN models. Information Criteria have been reported to be used for the selection of relevant input variables and determination of optimal NN model structures. This paper proposes the use of information criteria for tracking the model prediction accuracy and provides an algorithm for retraining the model. A heat exchanger in a refinery Crude Preheat Train (CPT) has been used as a case study. The operational problems of heat exchangers in CPT are compounded by the varying nature of crude blends and the complex fouling phenomenon. Fouling develops slowly and therefore the drift in the process occurs on a slower scale. The performance of a NN fouling model, developed using industrial data is investigated for drift. Model performance at different operating conditions is evaluated and it has been shown that drifts do occur in the process. An algorithm for retraining NN model has been proposed