Modeling of a continuous food process with neural networks

Abstract

Three neural networks were constructed and trained to provide both next step prediction and multi-step prediction of a snack food continuous frying operation. The three neural models were a feedforward sigmoidal network (FFN), a radial basis function (RBF) network, and a time lagged recurrent neural network (TLRNN). As a benchmark for comparison, a linear autoregressive model with exogenous inputs (ARX) was constructed. The models were all designed to provide multi-input/multi-output (MIMO) prediction. For next step prediction, the ARX model provided the best prediction. For multi-step prediction, the TLRNN model performed best. In addition to constructing the models, a variety of techniques for evaluating neural applications is presented, and a methodolgy for applying neural networks for time series modeling is proposed.