Machine Learning Can Maximize Efficiency in an Industrial Process

Abstract

- Continuous industrial manufacturing processes are generally controlled by a set of continuous control variables. The process usually produces a steady flow of output material, such as cement, food, milk, chemicals, and sugar or electrical power in a power plant. Control variables may be electrical or fossil power, cooling or heating, lubrication, pressure, etc. The process responds with a given flow of output measured in tons per day or power expressed in Megawatt. Dividing the input power response yields a variable proportional to the degree of efficiency of the process, which is a very important parameter in most cases. To understand, analyze or predict the process, in a first step, we will approximate the empirical response values by a smooth function, mapping the space of controls onto the interval [0,100%], using Machine Learning Techniques and Multivariate Statistics such as Tensor Flow or Generalized Linear Models (GLMs), respectively. Both approaches provide suitable approximation measures. In a second step, the process will be optimized within a given set of constraints concerning the control variables. This step will be illustrated by GLMs only due to their lack of overfitting and their continuous differentiability properties. This way, optimal set points, and sensitivity coefficients will be given.