LAUSR

Abstract Deep learning methods, especially the RNN based approaches, have been applied to the industrial time-series modeling these years successfully. However, the RNN based methods cannot overcome the enormous amount of iterative calculation. Traditional CNN based methods can calculate much faster; however, the CNN based methods lack in the explanation of the variables. In this paper, we proposed a novel adapted receptive field temporal convolution networks integrating regularly updated multi-region operations based on principal component analysis (PIMRO-ARFTCN). First, the principal component analysis (PCA) method is used to select the most concerned variables. Then, the Levenshtein distance based hierarchical clustering method is used to extract the multi-region operating features in the variables, and the related variables are used as additional input samples. The adapted receptive field decides the primary input samples. Last, the temporal convolution network structure is used to obtain the final results, and the multi-region operations can be regularly updated with various working situations. The proposed regularly updated PIMRO-ARFTCN method can not only take the most useful human experience into consideration but also combine the advantages of both RNN and CNN. The causal relationship between variables and the calculation speed are both considered. The prediction error of the proposed method is 0.1979, and the remaining errors of the other techniques, the MRO-ARFTCN method, the LSTM method, and the traditional RNN method, are 0.2464, 0.2829, and 0.4168, respectively. Compared with other means, the proposed method shows better prediction performance in time-series modeling.