Online improving NOX concentration prediction at SCR system outlet through continual learning

Abstract

Background

Accurately predicting NO_X concentration at SCR system outlet is crucial for optimizing process parameters and reducing NO_X emissions. Existing prediction models often struggle to maintain accurate long-term online predictions when operating conditions change and new data arrive. Therefore, frequent model updates are required in practical applications. However, new models often exhibit catastrophic forgetting of learned patterns, leading to a deterioration in NO_X concentration prediction accuracy.

Methods

A novel continual learning algorithm, termed TDIR, is proposed. This algorithm dynamically identifies historical temporal samples that are most susceptible to interference from new data and prioritizes their replay. The TDIR algorithm is integrated with the iTransformer architecture to establish TDIRformer, an online self-updating model for NO_X emission prediction. The model utilizes variate tokens to capture cross-feature correlations, effectively addressing prediction inaccuracies caused by multivariate coupling. It also employs the TDIR algorithm for online updates, which mitigates catastrophic forgetting and improves the prediction accuracy of NO_X concentrations.

Significant findings

Experimental results show that TDIRformer significantly outperforms LSTM, Informer, PatchTST, and iTransformer, while TDIR also surpasses the continual learning methods EWC and MAS. Additionally, the TDIR algorithm demonstrates strong generalization capability, achieving RMSE reductions of 10.7 %, 9.3 %, and 11.9 % when applied to LSTM, Informer, and PatchTST models, respectively.