Comparison of mathematical models, artificial neural networks and adaptive neuro-fuzzy inference system (ANFIS) in prediction of instantaneous drying curves of potato slices in a microwave dryer

The main objective of this research was to compare of accuracy of three widely used simulation methods including mathematical thin0layer models, artificial neural networks (ANN), and adaptive neuro-fuzzy inference system (ANFIS) in estimation of instantaneous moisture ratio of microwave power dried potato slices. To predict the moisture ratio, seven mathematical models were used. Furthermore, based on the experimental data, microwave power, samples thickness and process time, and the moisture ratio were considered as inputs and output of artificial neural networks and adaptive neuro-fuzzy inference system, respectively. Designing of neural networks was performed based on multi-layer feed-forward back-propagation (MFFBP) and cascade forward back-propagation (CFBP) structures, linear (Lin), sigmoid hyperbolic tangent (Tan) and logarithmic (Log) threshold functions, and Levenberg-Marquardt (LM) and Bayesian Regularization (BR) training algorithms. For simulation by adaptive neuro-fuzzy inference system, Takagi-Sugeno fuzzy system was selected, structure of the fuzzy inference system (FIS) created by grid partitioning method, and the membership functions in fuzzy logic toolbar of MATLAB software used. Among the studied modeling methods, Midilli model, CFBP network with 3-10-10-1 topology, LM training algorithm and Tan-Tan-Lin function, and ANFIS model with sigmoid membership function in input and 3×3×4 fuzzy rules were found as the best models. Based on the obtained results, all the three modeling methods were capable t estimate the instantaneous moisture ratio with desirable accuracy. However, showing the coefficient of determination of 0.9997 and mean square error of 4.53×10-5, ANFIS model had the better performance in estimation of the experimental data.