Research On Complex Prediction Of Industrial Time Series Data Based On Collaborative And Efficient Neural Networks

Since entering the 21st century,global industrial production has been developing at a high speed,and a new generation of intelligent information technology and industrial production have been integrated with each other.The popularity of data collection systems in the industrial production process has collected a large amount of high-dimensional multivariate industrial time series data.These data contain a wealth of information on the adjustment of working conditions,operation rules,and abnormal states in the production process.Meanwhile,operators want to know the future trend of certain key indicators in advance,to achieve performance prediction,saving energy,reducing emissions,and improving production efficiency and other functions.Therefore,the task of forecasting industrial time series data has become one of the main research objects.However,with the increasing scale and complexity of current industrial production,the collected industrial data contain complex information and are highly correlated,strongly nonlinear,and multimodal,making it more difficult to analyze and predict the data.Therefore,this paper conducts a study on industrial time series data forecasting tasks with the above characteristics:（1）In complex industrial systems,time series data mostly exist in multivariate form.The variables in the data interact with each other and contain complex information.To adequately extract this implied information,we propose a multi-stream neural network with point attention feature fusion（MS-PANet）model for industrial time-series data short-time prediction tasks.Inspired by the idea of the division and collaboration,we designed a two-stage network model.In the division of labor phase,multiple branches divide the work among themselves and work independently to improve the specialization ability of their respective branches.In the collaboration phase,a point attention mechanism approach is proposed that can effectively fuse the multiple feature information extracted from the division of labor phase to improve the model prediction performance.And it is verified in experiments on numerous domain datasets that MSPANet has better prediction performance.（2）Compared with short time series forecasting,long time series forecasting is more meaningful in real industrial production.To effectively capture the long-term dependencies in very long input data,while efficiently processing the input sequences and improving the efficiency of model operations.We propose an efficient recurrent neural network model based on the divide-and-conquer algorithm,called DC-RNN.The standard gated recurrent unit（GRU）model is improved using the idea of global library,and the global GRU model is proposed to retain more information about long-range sequences.In addition,we use the divide-and-conquer algorithm,which enables DC-RNN to efficiently complete the feature information extraction of long-range input sequences.Compared with other models,the DC-RNN model has high operational efficiency and low time complexity,and achieves parallelization.Better prediction performance is also obtained.（3）The large amounts of industrial time-series data obtained from actual industrial production processes are usually high-dimensional and redundant.Direct use of these highdimensional data for prediction can create problems such as high model complexity and noise.Therefore,a dynamic feature selection model（DFS-GRU）is proposed by combining the attention mechanism and the GRU model.The method first fuses the input variable data of alltime steps to focus on more distant moments and improve the memory capability of the model.Subsequently,all feature variables are scored using an attention mechanism to achieve a ranking of variable importance and remove redundant feature variables,while helping us to reveal the principles behind process industry production.Finally,the prediction of key indicators is accomplished by using GRU as a predictor.The experimental results show that the proposed DFS-GRU has better prediction performance compared with other baseline models.