Incremental fast relevance vector regression model based multi-pollutant emission prediction of biomass cogeneration systems

Abstract

Exact and trusty prediction of pollutant emissions is pivotal for optimal combustion control in biomass cogeneration systems, which possess multiple variables, high-volume data streams, and dynamic characteristics. Aiming at the multivariate dynamic systems, this paper extends a classical fast relevance vector regression (FRVR) algorithm into a multivariate form to accomplish synchronous multi-pollutant prediction. Meanwhile, a flexible and effective online training strategy is proposed to solve the problems of low accuracy of multi-step prediction and lack of dynamic updating capability. First, the given dataset is divided utilizing the k-means clustering method to enhance the clustering of similar features and expedite the prediction process. Then, the classical FRVR algorithm is extended into a multiple-output form, enabling the simultaneous prediction of multiple pollutant emissions. Moreover, the incremental learning method is introduced into the proposed multivariate FRVR model to improve its dynamic performance and online learning ability. Finally, the proposed method’s effectiveness is verified through a biomass cogeneration systems case. Experimental findings fully illustrate that the proposed method provides the lower RMSE and MAE while runtime decreases by 50% and $R^2$ reaches 96%. The proposed method significantly outperforms others, showing excellent potential in the pollutant prediction field.