The Use of Regression and Neural Network Modelling in Production Monitoring of an Industrial Enterprise

Abstract

Regression and neural network model for predicting the required dose of coagulant, depending on the quality of river water supplied for water treatment, are considered, their comparative analysis is carried out. For modelling and forecasting, statistical data collected for the period from 2005 to nowadays. Regression models were built on the true values of the factors (water quality indicators) and on their first differences to eliminate the trend in the series. For the models built on the true values, the statistical significance, was confirmed, high values of the coefficient of the determination were obtained, the values of the approximation errors were 22–25 %. In neural network modelling, networks of the multilayer perception were used. Generalization error on the test set when using other type of networks (RBF-networks, Elman networks), was significant above. Computational experiments have shown that, in general, the accuracy of neural network models is higher than regression ones. The average learning error was 7–9 %, the error on the test set increases to 12–16 %. The reliability of the forecast is increased by training the network on more recent data and using a larger set of facts. An increase in the influence of indicators of permanganate oxidability and colour of the initial river water on the dose of reagents with a simultaneous decrease in the degree of influence of alkalinity over the last five-year period was revealed. This confirms the need to periodically update data for building models. Selected models recommended for implementation in industrial monitoring of water treatment technology at the enterprise.