Integration of continuous restricted Boltzmann machine and SVR in NOx emissions prediction of a tangential firing boiler

Abstract

The NOx emissions prediction modeling of a coal-fired boiler is a complicated problem because of its highly nonlinear and strongly correlated multi-variables. To address this problem, this paper presents a novel deep structure using continuous restricted Boltzmann machine (CRBM) with support vector regression (SVR). A new steady-state identification method is first established by applying kernel principal component analysis (KPCA). Subsequently the stationary samples are subjected to the stacked CRBM network for extracting implied features. The combination of the optimal features and target NOx values are further utilized for training SVR to establish the regression pzmart of the CRBM-SVR prediction model. Additionally, particle swarm optimization (PSO) is applied for optimizing hyper-parameters of SVR, and weights of CRBMs are fine-tuned with gradient descent on account of prediction errors. Compared with the existing state of the art, the proposed structure achieves a good performance of 4.83 mg/m3, 3.55 mg/m3 and 0.963 of RMSE, MAE and R2.