Research on Dual GRU model driving by time series alignment for multi-step prediction of ash content of dense medium clean coal

Abstract

ABSTRACT Dense medium cyclone separation is a widely used coal processing method. However, duetween the ash content measurement and the other sensors. However, existing studies either do not consider the impact of time lag or only conduct single-step ash prediction, making it difficult to consider the systematic changes and uncertainties of multiple time steps in the future entirely. To solve these problems, first, this study uses time lag correlation to quantify and eliminate the lag effect between each variable and the ash content. Secondly, this study designs a Dual GRU model combining time series and regression prediction. Finally, this approach is adapted to some of the time series forecasting models proposed in recent years to compare with the model proposed in this study. Experiments conducted on production data from a coal preparation plant showed that input and prediction lengths were determined to be 7 based on the time lag correlation. Then, the mean absolute error of the 7-step ash prediction was 0.1495. The MAE value was reduced by 4.66% compared to the model without time series alignment. Compared to the model without aligned process variables, the MAE value was reduced by 6.77%. Compared to Informer, the best performer among the Transformer-based models, Dual GRU accuracy is improved by about 7.94%. Time series alignment and Dual GRU provide technical support for realizing density control based on ash content multi-step prediction.