|BRUSAFERRI ALESSANDRO||Cycle: XXXIII |
Section: Computer Science and Engineering
Tutor: DANIEL FLORIAN
Advisor: MATTEUCCI MATTEO Major Research topic
:Machine learning for optimization of energy intensive industrial processesAbstract:
Flexibility and efficiency are fundamental challenges to maintain the competitiveness of European industry in the global market. On the one hand, modern production systems must be able to operate in dynamic contexts, characterized by a strong variation in demand and by the request for highly differentiated products. In addition, production costs, resource consumption and environmental impact are the subject of increasing attention for their overall reduction. In this scenario, minimization of energy costs represents a fundamental target to be addressed to contain operating costs; indeed it would impact on the final price of the product allowing for a greater competitiveness on the market. However, technologies available today for factory automation and management (e.g., Manufacturing Execution System - MES) are essentially dedicated to achieving production targets while energy consumption is considered separately through dedicated software such as the Energy Management Systems (EMS). Energy performance is therefore assessed ex-post - as a consequence of productivity-oriented process management policies - limiting the operating margins in favor of maximum efficiency. To move forward from the current, inefficient, separate managements of productivity and energy consumption, we need a new generation of technologies for the integrated management of the productive asset which support the identification of the best strategy to be pursued in modern production lines. The intelligent integration of the factory into the Smart Grid represents an additional fundamental requirement to be addressed. To date, factories interact with the energy system essentially according to a consumption logic: power is absorbed according to production needs and costs are charged - depending on the contract - typically on fixed price and/or time band. Looking ahead, the consumption profiles of energy-intensive processes must be oriented towards the timed availability from the supply network as much as possible, thus favoring a growing integration of renewable sources. New factory management systems must natively support the operation in the energy market, e.g., by integrating forecast price of energy within the optimization strategy to be pursued. A development in this direction would greatly extend the flexibility (and the consequent availability) compared to the current Demand Response programs on pre-set load, which must obviously be very low to avoid disturbing production. At the same time, the power profile purchased on the energy market (e.g., the day-ahead market) can be considered as a constraint to be respected, adapting appropriately the management strategy of the process in order to avoid deviations. Factories could therefore improve operating margins by reducing energy costs, on the one hand reconfiguring production based on hourly price (e.g., day-ahead market), on the other hand actively contributing to the creation of a market with reduced costs. The research will focus on the challenges above by investigating the development of a framework based on machine learning and optimization techniques for energy-aware production process optimization. The research will exploit the increased availability of huge amount of process data which has resulted from the widespread investments in production system digitalization (within the Industry4.0 framework). The adoption of machine learning approaches for time series forecast is also foreseen, including both regression and classification techniques for day-ahead energy market prediction. Recent models for time series predictions will be investigated, such as deep recurrent neural networks (e.g., LSTM based architectures) and sequence to sequence models. Application scenarios will focus on energy intensive production processes such as those from the iron & steel manufacturing sector.