A NOx Emission Forecasting Model Using Multiresolution LSSVM Based on Data-driven and SF-KPCA

Abstract

A NOx emission forecasting model using multiresolution least squares support vector machine (LSSVM) based on data-driven and SF-KPCA is proposed. Firstly, the data-driven model based on real operation data contains a lot of redundant information. Therefore, a Kernel Principal Component Analysis based on Similarity Function (SF-KPCA) is proposed to reduce the redundant information and optimize the modeling complexity. Then, inspired by multiresolution analysis, a multiresolution LSSVM with online adaptive update ability is conducted, which employs the total forecast error as the threshold to update the forecasting model parameters in real time. Finally, the proposed method is applied to predict the NOx emission of a coal-fired boiler. Results reveal that the online multiresolution LSSVM model with SF-KPCA maintains the prediction accuracy at a high level even if the process characteristics vary. Meanwhile, the model only needs 0.15 seconds for each forecast, which is suitable for real time monitoring in field.