Study And Application Of LSTM Structure Based On Improved Particle Swarm Optimization

The swarm intelligence algorithm is a widely used optimization algorithm,but its population diversity will decrease rapidly because the population falls into the local optimal area,resulting in poor algorithm convergence accuracy.A novel reverse learning strategy is proposed to improve the algorithm performance.Long-term and short-term memory network is a hot field of current research,and its model has many hyperparameters and needs to be manually set and debugged.In this paper,the swarm intelligence algorithm is used to optimize the combination of long-term and short-term memory network hyperparameters,build an automatic parameter tuning model,and verify it through experiments The validity of the model,the specific research content is as follows:(1)A reverse learning strategy based on orthogonal complement space is proposed,which calculates the orthogonal complement space of the optimal individual of the population and generates reverse solutions accordingly.The reverse solution can jump out of the local extreme value area,increase the diversity of the population,and fully explore the solution space.The strategy is applied to the standard particle swarm and genetic algorithm,and 16 benchmark functions are optimized.The experimental results verify the effectiveness and universality of the algorithm.(2)Combining the orthogonal complementary space reverse learning strategy with the center of gravity reverse learning and random topology structure,the improved particle swarm algorithm has sufficient space exploration ability in the early stage,excellent development ability in the later stage of iteration,fast convergence speed and high precision.Using CEC2013 as the test function set,compared with four classic or excellent reverse learning particle swarm optimization algorithms,the experimental results show that the improved algorithm has the best performance.(3)The improved particle swarm algorithm is used to optimize the hyperparameters of the long and short-term memory network model,and the learning rate,the number of hidden layer units,the batch sample size and the parameters of the Adam algorithm are selected as the optimization goals.Taking the prediction of stock prices and temperature as an example,the prediction accuracy of the optimized model is verified and compared with the standard particle swarm optimization and genetic algorithm model.The experimental results show that the improved particle swarm optimization model has higher accuracy.