Modeling temporal dual variations for return air temperature prediction of mK-level temperature-controlled clean chamber

Abstract

Clean chamber with mK-level temperature stability is required for precision manufacturing and optical equipment to achieve nanoscale manufacturing tolerances and measurement accuracy, however, the equipped precision air conditioning system cannot respond immediately to temperature fluctuation in the chamber due to large time delay. Predicting the future return air temperature gives the control module a prescient ability to act in advance, thus aiding in temperature regulation. To this end, this paper deals with the mK-level temperature prediction problem under a high-precision temperature control situation, and a novel prediction network model dubbed TDVNet is designed based on convolutional neural network (CNN) and gated recurrent unit (GRU), where Temporal Dual Variations of time series data, namely variations-between-phases and variations-across-periods, are modeled to achieve a higher accuracy of return air temperature prediction. Utilizing chamber temperature data collected on-site for model training and evaluation, the experimental results show that capturing data change patterns from dual perspectives has reduced the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the TDVNet model by 21.4% and 24.3%, respectively, compared to the model only capturing the variations-between-phases. Under a signal sampling interval of 10 s and a temperature fluctuation of (66.9 ± 25.8) mK@10min, the TDVNet model can accurately predict the return air temperature change in the next 5 min, and the proportion of prediction errors within 15 mK in the test set can reach over 90.1%.