Potential energy of an alloy is an essential indicator for evaluating the stability of the structure in predicting new materials. Therefore, how to calculate the potential energy in material design has become an inevitable problem. While first-principles calculations can provide chemical accuracy for arbitrary atomic arrangements, they are prohibitive in terms of computational effort and time. To enable atomistic-level simulations of both the processing and performance of Aluminum alloys, neural network potential was proposed to predict the binding energy of vacancy-containing aluminum alloys in a highly accurate state. This method combined first-principles calculations and machine learning techniques to explore the intrinsic link between solid solution structure and binding energies. In this study, four binary alloys (aluminum-silicon, aluminum- zirconium, aluminum-magnesium and aluminum-titanium alloys) were investigated. The mean squared errors were used to quantify the quality of the neural network potential models and it was found that the trained model is more stable and exhibits high accuracy for energy prediction. The Monte Carlo simulation results show that using this neural network potential successfully simulated aging process of aluminum alloys, and the neural network potential can be much faster than first-principles calculations, even with high accuracy.