Time series forecasting of NOx concentration based on Informer for MSWI

Accurate time series forecasting of NOx concentration in flue gas is an important part of MSWI (municipal solid waste incineration). Time series prediction methods based on sliding window prediction, such as CNN and LSTM, have the problem of error accumulation, that is, the longer the predicted time series, the greater the error. Aiming at the above problems, we innovatively introduce Informer for time series prediction of NOx concentration. In this paper, we propose a method in which Informer fused with learnable positional encoding is trained with the dataset generated with variables selected by TLCC (time-lagged cross-correlation) between NOx concentration time series and other parameter time series to achieve time series forecasting of NOx concentration. Firstly, the TLCC was used to select the appropriate variables to generate the data set. Secondly, the global timestamp encoding of Informer is discarded, and the experiment proves that the global timestamp does not help the time series forecasting of NOx concentration. Thirdly, the learnable positional encoding is integrated into Informer to enhance the model's flexibility to obtain the location information of data at different time steps to compensate for the lack of the global timestamp. Finally, the prediction results of the proposed method are compared with those of other models and methods using several evaluation indicators. The experimental results show that this method has better prediction ability.