The Research On Improved Elman Neural Network And Parameter Optimization Algorithm

Elman neural network(Elman Neural Network,ENN)is a kind of common feedback neural network.It has a strong adaptive ability of time-varying characteristic.Therefore,ENN is suitable for the prediction of time series data.The former researchers have shown that the performance of neural network mainly depends on the design of network structure,including the topology structure and the number of nodes in the hidden layer and so on.In order to solve different problems,we need to design different network structures,especially to determine the suitable number of hidden layer nodes.Traditional Elman neural networks adopt gradient descent learning as learning algorithm,which causes the problems that the convergence speed is slow and the convergence process is unstable and network is easy to fall into local optimal solution.In order to solve above problems of traditional Elman neural network,this paper improved the topology structure of traditional Elman neural network firstly,and proposed a novel Elman neural network structure with double hidden layers and output-hidden feedback(DHOHF-Elman).Then,combined with the improved adaptive genetic algorithm,an optimization algorithm for network parameter based on AGA was proposed.Finally,the optimized DHOHF-Elman neural network was applied to the prediction of air quality time series.The work and innovations of this paper mainly includes the following three aspects:(1)Improved Elman neural network topology structure.Based on the traditional Elman neural network,and combining the advantages of OIF-Elman and OHF-Elman network proposed by Shi et al,we put forward a kind of new Elman neural network with double hidden layers which includes internal feedback and external feedback by increasing the special output feedback.In order to verify the effectiveness of theimproved DHOHF-Elman network structure,the DHOHF-Elman network was compared with the traditional Elman neural network and the OIF-Elman proposed by Shi et al.The experimental results showed that the DHOHF-Elman neural network had higher prediction accuracy and faster convergence speed.(2)Proposed an optimization algorithm for network parameters based on genetic algorithm.An improved adaptive genetic algorithm(AGA)was used to determine the numbers of hidden layer nodes and the initial weights and thresholds.The numbers of hidden layer nodes and the initial weights and thresholds respectively used different encoding methods and the corresponding operations for selection,crossover and mutation of genetic evolution.Then decoding the chromosome with the best fitness value can get our desired solution.Two groups of contrast experiments were designed to verify the effectiveness of AGA determining the numbers of hidden layer nodes and the initial weights and thresholds.The experimental results showed that:1 determining the number of hidden layer nodes in the network with AGA was less time-consuming than using the enumeration method to find the best hidden layer node number;2 determining the initial weights and thresholds with AGA before training that spent a lot of time,but in the training process,it had faster convergence speed,better accuracy and more stable convergence process than using the random method to determine initial weights and threshold.(3)Predicted the air quality time series data.Based on the DHOHF-Elman neural network structure and network parameter optimization algorithm proposed by this paper,we used the air quality time series data to predict the ozone concentration in the next period.In order to verify the validity of the experiment,the DHOHF-Elman neural network was compared with other Elman neural networks,such as Elman neural network and OIF-Elman neural network.At the same time,we compared with other time series forecasting methods,such as grey prediction and NARX neural network.The experimental results showed that the DHOHF-Elman network and the proposed network parameter optimization algorithm can improve the prediction accuracy in the aspect of time series prediction,and can avoid falling into local optimal solution,reduce the concussion in the process of convergence,and accelerate the convergence speed to some extent.