Industrial Process Time Series Predictions Using A Novel Esn Concerning Dynamic Time Delay Matching Of Multivariate Data

With strong dynamic performance learning capabilities,the Echo State Network(ESN)has been applied to predictions of industrial process multivariate time series,contributing to solving key problems associated with process supervision,control and optimization.The traditional ESN uses a randomly generated sparse reservoir and a single neuron model in the network structure,which results in strong node coupling and limited computing power,rarely applying to complex industrial processes.Additionally,static time delays are used to select multivariate data in a fixed time dimension,which ignores the dynamic time delay characteristics of practical processes,possibly causing misalignments of network input/output data or losing key data information.In response,this thesis conducts an in-depth investigation on the above issues,achieving the following research results.1.Considering structural defects of the traditional ESN model,a hybrid cycle reservoir with jumps(HCRJ)network is proposed.It combines a low-complexity topology with a hybrid wavelet neuron model,constructing an optimal network model by means of combinations of the mixed neuron model in the cyclic structure and jump structure of the reservoir.Predictions of key variables in an oil drilling process demonstrate the superiority of the proposed network.2.Based on the dynamic time delay theory,an approach to matching multivariate time series for the network is proposed.According to the dynamic transmission time between variables,the time dimension of the selected multivariate data is specified for each predicting point,so that the network can accurately and quickly capture the largest contribution of input data.Applied to the prediction of distillation column temperature,satisfactory results were achieved obtained,verifying the effectiveness of the contributions.