Research On Time Series Forecasting Based On Evolutionary Echo State Network

The task of time series forecasting is to predict the future trends of data based on collected historical data,providing theoretical and data support for human judgement and decisionmaking.From scientific research to industrial production applications,these statistics reflect the state and degree of change of the target system under study over time,from a wide variety of sources,and are of strategic research value.Echo state network,as a simplified artificial recurrent neural network,only need to train the connection weights of the output layer,and their simple structure and fast training speed have been widely used in various research and application areas of time series analysis.However,there are some problems that need to be solved: for example,the echo state network is heavily influenced by its reservoir parameters,which are difficult,time-consuming and inconvenient to design manually,and finding the optimal network parameters that are more suitable for a specific task is a problem that needs attention;the reservoir is the core of the echo state network,and its original internal topological connections are randomly generated and only limited by the control of sparsity,which often performs poorly on more complex tasks and affects the overall predictive performance of the network.To address the above issues,this paper focuses on two main areas of research as follows.(1)To address the problem of designing the parameters of the reservoir of the echo state network,this paper proposes a particle swarm optimization algorithm based on an adaptive strategy for time series prediction.The particle swarm optimization algorithm is used to encode the parameters of the echo state network,and an adaptive framework is constructed for the parameter optimization of the echo state network.During the training process,an automatic strategy selection mechanism is used for each particle to select strategies from the pool of candidate strategies to evolve to the next generation,and the selection probability of strategies is continuously updated according to the performance of the strategies in the strategy pool.Leaky integrator neurons are also introduced to replace the simulated neurons inside the traditional reservoir,adding their leaking rate parameters as the optimization objects.The experimental results show that compared with the traditional echo state network and the echo state network with optimized parameters based on the evolutionary computation method of a single strategy,the echo state network with the particle swarm optimization algorithm based on the adaptive strategy selection pool can better adapt to the target task and obtain higher prediction accuracy.(2)To address the problem of designing the internal topological connections of the reservoir of the echo state network,a topological structure optimization model based on the particle swarm optimization algorithm encoding for time series prediction is proposed in this paper.For the topological connection of the reservoir,the population-based optimization method is introduced into the topological structure optimization of the echo state network.The topology of the reservoir is encoded and evolved using particles in the population,which is then decoded and used to generate a matrix of connection weights within the reservoir,and finally the echo state network is trained to evaluate its fitness values.As the population continues to evolve,the reservoir topology is continuously updated and searched for the best fit for the task at hand until a pre-determined stopping condition is met.The proposed model is evaluated on the Mackey-Glass benchmark time series and EEG signal prediction.Experimental results show that the proposed scheme is better adapted to complex data application scenarios than traditional hand-designed topologies,and has higher prediction accuracy and stability especially in the case of large data fluctuations and more complex motion trends.