Time Series Prediction Based on Random Convolution Kernel

Abstract

In the wake of fast transformation and growth of industrial manufacturing digitization, time series data recording important data of industrial machinery performance and functions has become a key fulcrum for processing and understanding industrial Internet of Things data. The prediction of time series data is an urgent problem in the industry. For the prediction of time series data in industry, the prediction accuracy of traditional machine learning methods is not high, and the training time of complex neural network models is too long to meet industrial needs. To solve this problem encountered in practical applications, this paper uses the Rocket method, which uses a simple linear regressor with random convolution kernels to achieve high accuracy with less computational cost. This method is improved in combination with the characteristics of the industrial dataset used, so that it has an acceptable high model prediction accuracy on the shield dataset. In this paper, we also selected four other time series datasets for evaluation, and our results show that Rocket and improvements to it exhibit higher overall accuracy and adaptability in regression than the state-of-the-art machine learning algorithm XGBoost and the recurrent neural network algorithm LSTM.