Response surface methods based in artificial intelligence for superstructure thermoeconomic optimization of waste heat recovery systems in a large internal combustion engine

Abstract

Listen

Translate article

In a large internal combustion diesel engine, less than half of the fuel energy is converted into useful energy, while the remaining energy is lost, mainly by the exhaust gases and cooling water. Thus, the implementation of waste heat recovery (WHR) systems has been one of the main research areas for increasing power, and reducing specific fuel consumption and pollutant emissions, which favors the improvement of these engines. Currently, there are two important challenges regarding the use of WHR systems in internal combustion diesel engines: (i) identifying the best technology, or combination of technologies, for the WHR in an internal combustion diesel engine from an economic point of view; and (ii) determining the ideal configuration and the best design parameters for the most suitable system. This paper proposes an optimization procedure for automatic selection of the most economically suitable WHR technology for a large internal combustion diesel engine, as well as the definition of its optimal configuration and ideal design parameters. For this, superstructures composed by the Organic Rankine Cycle (ORC), Kalina Cycle (KC), and Conventional Rankine Cycle (CRC) are modeled and optimized using a genetic algorithm. For the formulation of the thermoeconomic optimization problem—Mixed Integer Nonlinear Programming (MINLP)—the economic model of each component of the superstructures is evaluated and the objective function is proposed based on the specific cost. However, due to the complexity of the superstructure optimization problem, the use of the response surface method to solve this problem can become more attractive than the original modelling optimization. The superstructure optimization problem, as well as its thermodynamic modeling, economic modeling and response surfaces based on radial base function and artificial neural network, were solved using the programs Engineering Equation Solver (EES) and Octave. The results show that for a medium-sized superstructure (384 variables and 17 decision variables), despite the high computational effort, the use of the response surface method in the optimization problem proved to be more efficient than the original modelling optimization, and for a large superstructure (944 variables and 45 decision variables) it is only possible to solve the optimization problem using the response surface method. Furthermore, after the superstructure’s optimizations, the additional net power is produced by ORC through of the recovery waste heat from the exhaust gases mass flow and the cooling water mass flow provided by the diesel engine. The additional net power represents around 7% of the nominal net power.