Multi objective optimization of an evolutionary feedforward neural network for the automotive air conditioning system performance prediction

In this paper, a novel multi-objective evolutionary artificial neural network approach is proposed to predict the performance of an automotive air conditioning (AAC) system. A Feedforward Neural Network (FNN) was used to simulate the cooling capacity and compressor power under different combination of input compressor speeds, evaporator inlet air speeds, air temperature upstream of the condenser and evaporator. Differential Evolution (DE) algorithm was employed to automatically optimize the FNN's parameters, involving the number of hidden layers and the number of neurons in each hidden layer. The training of connection weights and biases is carried out using the basic backpropagation algorithm with Levenberg Marquardt nonlinear optimization method. For the purpose of multi-objective optimization, the DE algorithm is incorporated with two key elements of the NSGA-II (Non-dominated Sorting Genetic Algorithm II), namely the non-dominated sorting method and the crowding distance metric. A parametric study was performed on the proposed algorithm and the best DE base variant was determined. The experimental results show that the proposed algorithm with DE based variant 'DE/Best/1' exhibited its superiority in term of prediction performance. The best neural network obtained is FNN with 4×18×2 network configuration and its network complexity is equivalent to 108 connection weights. It yields an average relative error of 0.60% for the prediction of cooling power and one of 3.0% for the prediction of compressor power