A Semi-supervised Modeling Method for Dioxin Emission Prediction Based on Random Forest

Abstract

Dioxin (DXN) is a kind of pollutants with cumulative effects in the municipal solid waste incineration (MSWI) process. Its emission concentration is difficult to detect online and in real-time, which restricts the operational optimization of the MSWI process. At the same time, it is difficult to meet actual needs through traditional supervised modeling methods because of the high time and economic cost of directly measuring DXN. Therefore, a DXN emission prediction model based on semi-supervised random forest (SSRF) is established to make full use of the unlabeled data obtained in the actual industrial process. First, the training subsets are acquired through randomly sampling the labeled data. Second, the training subsets are utilized to build multiple random forest (RF) models and pseudo-label the unlabeled data. Finally, the mixed samples composed of pseudo-labeled data and labeled data are used to train an RF model for predicting the DXN emission concentration. The proposed method is verified by the actual DXN dataset.