Neural networks for small scale ORC optimization
Alessandro Massimiani, Laura Palagi, Enrico Sciubba, Lorenzo Tocci
Session: Session 2A: System Optimization (1)
Session starts: Wednesday 13 September, 14:20
Presentation starts: 15:20
Room: Building 27 - Lecture room 01
Alessandro Massimiani ()
Laura Palagi ()
Enrico Sciubba ()
Lorenzo Tocci ()
Organic Rankine Cycle (ORC) technology is considered as a cost effective technology to produce electricity out of low grade thermal energy sources. Despite of the potential market available, ORCs for small scale applications still find it difficult to create their own market. One of the reasons is undoubtedly their excessive specific market price [$/kW] which leads to a high payback period. Another reason is that SMEs (Small - Medium Enterprises), which are by far the customers that may profit the most from small scale ORCs, are not sufficiently aware of the potential savings this technology could lead to. As it is often the case, at the smaller scales additional design issues arise which raise the specific price $/kW even futher, thus limiting the market potential of this technology.
Small scale ORCs represent a viable method to retrofit Diesel generators in stationary applications. The ORC power plant receives as an input the thermal energy of internal combustion engine exhaust gas and converts it into electricity by means of a thermodynamic cycle. The inherently low-temperature source leads inevitably to low recovery efficiencies, and therefore it is important for designer to optimize both the cycle parameters and the working fluid, the two issues being of course intimately connected.
Machine learning techniques (MLT) are receiving interest in the optimization of ORC plants. The reason is twofold: firstly, the thermodynamic problem is highly non-linear, which makes it impervious the design of efficient optimization algorithms. Secondly, MLTs allow the user to perform feature selection and eliminate from the optimization problem all these variables which do not affect significantly the objective function. Among MLTs, artificial neural networks (ANN) methods have proven to perform well in energy optimization problems.
This paper presents a case study of a 20 kW ORC system which converts the sensible heat of the exhaust gas of a Diesel engine into electrical energy. A thermodynamic model of the ORC system and a detailed 1-D model of the expander have been developed using the software MATLAB. Hence, an in-house code has been written which implements Neural Networks, using various classes of activation functions, to evaluate and compare their performance.
First, the procedure used for the optimization process is described. Subsequently, both the cost and size of the system have been minimized to meet reasonable specific cost for the plant. This study underlines the need for an integrated approach in which thermodynamic, technical and economic criteria are considered simultaneously in order to design an efficient and cost-effective system.