Neural network based iterative prediction of multivariable processes

It is commonly required to predict process inputs given the desired process outputs for multivariable systems. However, the desired process output set used for prediction may not necessarily be realistic to the process. Consequently the prediction will be inaccurate due to anomalies in the prediction input data. With developing processes it is often unknown whether a set of data presented for prediction is feasible or not until the prediction results are applied to the process. This paper presents a method using feedforward backpropagation neural networks to firstly adjust the data presented for prediction to be realistic to the process and secondly, to implement an iterative process to quickly converge to a prediction. The innovation is in the iterations being processed through a combination of separate backwards and forwards neural networks. By implementing this prediction method, expensive and time consuming process verification runs can be reduced through improved accuracy of desktop studies.