MODELING AND PREDICTION OF YEAR-BASED CHEMICAL ENGINEERING PLANT COST INDEX DATA USING FEEDFORWARD BACKPROPAGATION NEURAL NETWORK (FFBPNN) AND REGRESSION

This study presents a comparative analysis of Single Input Single Output (SISO) Feedforward Backpropagation Neural Network (FFBPNN) and conventional regression models for predicting year-based Chemical Engineering Plant Cost Index (CEPCI) data spanning 64 years from 1960 to 2023. The primary aim is to evaluate the predictive performance of both approaches using the coefficient of determination (R²) and mean squared error (MSE) as key performance indicators. In the regression modelling phase, a variety of functions, linear, polynomial (up to sixth order), logarithmic, exponential, logistic, and power models, were examined. The best regression performance was obtained for hexic polynomial (n=6), yielding an R² of 0.9760, indicating strong, but not optimal predictive accuracy. In contrast, the FFBPNN architecture demonstrated superior performance with a model tuned by varying the data split ratio, number of neurons in the hidden layer, and activation functions for hidden and output layers. The best configuration was found to be data split ratio of 0.90 (training) - 0.05 (validation) - 0.05 (testing), 30 neurons in the hidden layer, tansig activation in the hidden layer, and purelin in the output layer. The training was carried out using the Levenberg-Marquardt algorithm with gradient descent with momentum-based learning over 10 epochs. This configuration achieved an R² of 0.9997 and an MSE of 10.80, significantly outperforming the regression models. The findings confirm that FFBPNN provides a more robust and accurate framework for modelling complex, nonlinear trends in year-based CEPCI data, offering substantial advantages over traditional regression methods in both precision and generalizability.