Analysis And Improvement Of Genetic Algorithm And Artificial Neural Network

This paper focuses on analyzing and improving genetic algorithm and artificial neural network, and cooperating genetic algorithm with artificial neural network to learn together. Finally, this cooperative technique is applied to estimate battery state of charge. The main works include:(1) Analyze the functions and search scopes of genetic operations, and also the effect of population diversity, by probability. Then, adjust the parameters of genetic algorithm during evolution according to population diversity in order to restrain premature convergence. The efficiency of the improved algorithm is validated by simulation results of six different validation functions.(2) Take the different data set, learning algorithm and network topology into account to improve genetic neural network performance by adding learning algorithms code to chromosome. The adaptability of this algorithm is proved by simulation results of Proben1 test set.(3) Analyze the weights difference of neural network between pre-training and post-training by statistics, improve the connection pruning algorithm to construct non-full connection neural network. The simulation results of Proben1 test set show that the improved algorithm is capable of enhancing the neural network efficiency with meeting the requirement of output error. (4) Analyze the weights distribution status of well-performed neural networks by statistics to generalize distributive principle. Then based on this principle, we initialize weights before training neural network. The simulation results of Proben1 test set illuminate that the training efficiency could be enhanced to some extent by initiating weights according to normal distribution.(5) Considering the important roles of activation functions of neural node, we combine different functions to the same neural network, and analyze the drawbacks of traditional code methods, then construct hybrid code method based genetic neural network. The output error can be generally decreased using the neural network with combined activation functions achieved by evolution method. (6) We construct adaptive estimation model of battery state of charge by genetic neural network. During the process of modeling, neural network topology is optimized by genetic algorithm after it learned the relationship between state of charge and terminal voltage, discharge current. Thus the neural network model for battery of certain type is achieved. The adaptability and efficiency of the cooperative technique with two methods are validated by the simulation results of three batteries of different type.