Evaluation of machine learning techniques to enact energy consumption control of compressed air generation in production plants

Industrial energy management is an important topic of discussion nowadays for both economic and sustainability reasons. A monitoring and control system able to guarantee the practice of a real-time control is a key point in enacting an effective management of energy consumption in a complex organization. In this context, the ISO 50000 family of standards suggest the application of different types of energy performance indicators (EnPIs), in a range of varying complexity: from simple absolute values of energy consumption, to statistical models, to engineering models. The evolution of machine learning techniques falls between the statistical and the engineering models, depending on the volume of data and the human involvement required for building a model. Therefore, the value of the present work is to explore the use of these tools, already consolidated in other fields, but not yet adequately assessed for energy performance control. In particular, the generation and distribution of compressed air is among the biggest uses of energy in production plants. This work starts with the application of the classical statistical approach and then proceeds to compare two different machine learning techniques, artificial neural networks and support vector machines, for the creation of energy performance indicators. The analysis begins comparing the feasibility of application, implementation complexity, data and level of human interaction required, making use of the results of a real application to a compressed air generation unit in a production plant. The comparison was then carried out using various performance indicators (R-squared, Mean Squared Error, Mean Absolute Percentage Error) as well as a graphical inspection of the resulting control charts produced with the different models. The work demonstrates the applicability of machine learning techniques in this specific context, proving them as an efficient compromise between the complexity and accuracy of statistical and engineering models.