Research On Reservoir Computing Algorithm Andapplication Of Time Series Prediction

As a challenging task,time series prediction has attracted researchers in many fields.At present,many new timing prediction models have been proposed and the existing models have been modified.The purpose of these studies is to achieve high prediction accuracy.With the progress and development of machine learning and deep learning,deep learning gradually replaces traditional models and becomes the mainstream method of timing prediction.With the advent of this method,the accuracy of timing sequence prediction is higher.At the same time,this method is still being improved to solve more complex timing analysis problems.Based on the theory of RNN and ESN,combined with the knowledge of complex network,a new reservoir calculation model is proposed to predict nonlinear timing system.In this thesis,the reservoir model is applied to the time series prediction of onedimension and multi-dimension respectively.The model consists of three parts: input layer,reservoir computing network and output layer.The input layer is a fully connected network,and its weight keeps the initial value in the training process and will not be updated.The output layer is also a fully connected network.Different from the input layer,the weight of the network is mainly trained in the training process.The reservoir computing network located in the middle of two fully connected networks can be composed of three types of networks: random network,scale-free network and smallworld network.In addition to the weight matrix of the reservoir network itself,the model also introduces the state vector of network nodes,which is constantly updated in combination with the structural characteristics of the weight matrix,and then input to the output layer.The final output result is determined by the input data and the state value of network nodes.This thesis also analyzes which network structure is more suitable for one-dimensional time series or multidimensional time series.In the one-dimensional time series prediction,seven major data sets were collected and collated for training and testing,with each data set containing an average of 2272 time points.The experimental results show that the proposed model is generally superior to the LSTM deep learning method in most cases.To further assess the predictive power of this model,this thesis compares this model with two other deep learning methods in recent studies.The results show that the model is more competitive in the prediction of one-dimensional time series than the existing deep learning methods.In multidimensional time series prediction,based on project requirements,the proposed model is used to predict specific multidimensional time series.The prediction model proposed in this thesis completed the training in the training set consisting of the first 80% of the data,and then completed the evaluation in the remaining 20% of the data set.The experimental results show that the model performance meets the project requirements.The software package development based on reservoir calculation model is also completed.In order to verify the feasibility of this algorithm in hardware environment,this thesis finally realized the optical circuit simulation of this model,and proved that this reservoir computing model can be realized in the optical physical environment.