Data-driven predictor and soft-sensor models of a cement grate cooler based on neural network and effective dynamics

Abstract

In cement factories, a cement cooler is one of the important parts which plays a fundamental role in quantity and quality of produced cement. Grate cooler is a complex structure system, has the characteristics of nonlinearity, hysteresis, non-stationary, and it's hard to establish an accurate mathematical model to describe. Although some approximation models of grate coolers were presented by mass and energy balance theories, they still haven't worked out (purely in exceptional case). Also with the advancement of science and technology, a lot of hardware sensors have produced to measure variables in the cement production process, but nevertheless many variables have still monitored Through laboratory analyses. These are expensive, inadequate for online supervision and involve considerable delays, so they caused many problems for designing a cooler controller. For solving these problems, first of all, we preprocessed raw data of a “Volga-75” grate cooler and selected effective input dynamics by FS method. Then we built some nonlinear predictor models of a real grate cooler by using nonlinear identification technique (MLP neural network and LM training function). After that, for maximizing heat exchange and recover heat from the grate cooler we estimated difficult-to-measure variables by easy-to-measure variables, so-called soft sensor models which are computer programs and important cheap tools for the industrial process. At the end, a comparison between these two models is done. These models are helpful to study the behavior of a cement cooler in various conditions and to design controllers.