As an important means of engine development and optimization, modelbuilding plays an increasingly important role in reducing carbon dioxide emissions of the internal combustion engines (ICEs). However, due to the non-linearity and high dimension of the engine system, a large amount of data is required to obtain high model accuracy. Therefore, a modelling approach combining the experimental data and prior knowledge was proposed in this study. With this method, an artificial neural network (ANN) model simulating the engine brake specific fuel consumption (BSFC) was established. With mean square error (MSE) and Kullback-Leibler divergence (KLD) serving as the fitness functions, the 86 experimental samples and constructed physical models were used to optimize the ANN weights through genetic algorithms. To improve the performance of the model, model-based feature selection method constructed by generalized regression neural network (GRNN) is introduced reducing the input dimension from 8 to 4. Subsequently, different fitness functions and features were applied to construct the models. Through the comparison of the models, the ANN model trained with MSE + KLD and selected features (ANNM+L,S) obtained the best comprehensive performance. ANNM+L,S can well reproduce the operating characteristics of the engine, such as the impact of air-fuel ration and ignition timing on the fuel consumption, while maintaining high accuracy. With the constraints of KLD, ANNM+L,S provides a higher predictive ability which is proved by the better performance in the test data set. With the embedding of prior knowledge, precise engine models with higher stability and predictive abilities can be established even under the conditions of small sample size.