The Study Of Prediction And Optimization Of Ejector Performance Based On Artificial Neural Networks

The energy crisis has become one of the greatest concerns today. With the development of the world, especially the popularity of air-conditioning, more and more energy is used to support buildings. In China, refrigeration and air-conditioning energy consumption has accounted for about one-fifth of the total energy consumption. The power grid is facing with the impact of peak in summer. If the energy consumption of refrigeration and air conditioning can be reduced, energy crisis will be calmed in our country, which will play an important role in energy saving and emission reduction. Steam-jet refrigeration system is a promise option.There are no moving parts in steam-jet refrigeration system and it does not require electricity. Steam-jet refrigeration system is energy-saving and environment-friendly when steam, as refrigerant, is obtained by solar, waste heat or geothermic. However, the COP of traditional jet-refrigeration system is lower than other refrigeration system, which led to the restriction on the development of steam jet-refrigeration system. Ejector is a core component of steam-jet refrigeration system. If the performance of ejector can be enhanced, the COP of jet-refrigeration system may be improved. Therefore, it is particularly important to focus on ejector.For many years, Researchers have proposed various theories to design ejector and predict the performance of ejector. However, experiment is the only way to examine the correctness of these theories and the cost of experiment is high. If forecasting model of ejector which based on part experimental data can predict similar or the same series ejectors, it will be helpful to develop, design and operate ejectors. The great practical significance of ANN is saving time and costs for experiments, expanding the scope of application parameters of ejectors. The performance of ejector has a highly non-linear relationship with its own structural parameters and operating conditions. Artificial Neural Networks(ANN) is suitable to simulate non-linear systems. The use of ANN for research on the performance of ejector is a totally new method.Activation functions and learning methods of ANN have significant impact on learning speed and prediction accuracy. In this paper, three kinds of ANN based different activation functions are used to predict the performance. These activation functions are Morlet, Mexi-hat and Gauss1 mother-wavelet. Numerical experiments show that learning speed and predictive ability of the three kinds of wavelet neural network are higher than traditional ANN. The number of neurons in hidden-layer plays an important role in improving the performance of ANN. The optimal number of hidden-layer neurons was decided through repeated attempts. The learning process of ANN is essentially an optimization process. The traditional learning method is slow and easy to fall into local minimum. Ant colony optimization(ACO) is a new algorithm for global optimization. In this paper, two kinds of continuous ACO are introduced to train ANN, which improve the learning speed and prediction accuracy of ANN. Regularization can effectively improve the generalization of ANN. Numerical simulation results show that the combination of ACO and regularization can improve the prediction accuracy of ANN.In this paper, models of ANN used to predict ejector performance is obtained according to a large number of ejector experimental data. The prediction accuracy of ANN is better than theories. Optimized structure and operation parameters of the ejector are obtained to improvement entrainment ratio. It is worth noting that ANN can be easy to solve two-parameter and multi-parameter optimization problems, whereas it is unrealistic to do multi-parameter optimization for traditional experimental method. The results of multi-parameter optimization obtained by ANN are more accurate than that from traditional single optimization. The combining of ANN and experimental data simplifies the process of analysis. This method provides an economical, rapid and sufficiently accurate way for design and operation optimization of ejector.