An inductive transfer regression framework for small sample modeling in power plants

Abstract

Small sample size presents a significant challenge in process modeling, making machine learning (ML) models prone to overfitting and reduced accuracy. To address this issue, this study develops a novel inductive transfer regression framework called double-weight least squares support vector regression (DWLSSVR). First, sample weights are incorporated to minimize the multi-kernel maximum mean discrepancy (MK-MMD) between domains, thereby promoting joint distribution adaptation and decreasing domain discrepancy. Second, the impact of unrelated source domain samples is further mitigated by iterative weights derived from fitting errors. In addition, a two-step strategy is developed to optimize the hyperparameters in DWLSSVR, which introduces a new criterion based on Wasserstein distance (WD). A numerical simulation demonstrates the effectiveness of the developed framework. Then, the proposed method is applied to the small sample modeling of a complex chemical process. The results of predicting NO_x emissions from a coal-fired boiler demonstrate that the DWLSSVR model achieves superior prediction accuracy, with a coefficient of determination (R²) of 0.942 under the new operating condition. In contrast, the best LSSVR model achieves an R² of 0.844 under the same condition.