Local data augmentation of biomass gasification performance prediction based on improved generative adversarial network

Data-driven modeling is widely used in biomass gasification research. However, due to the complex operating conditions and wide range of gasification parameters in biomass gasification, there is often an imbalanced data distribution problem, seriously affecting the accuracy of data-driven models. An improved generative adversarial network (GAN) is proposed for local data augmentation in the prediction performance of biomass gasification. The traditional GAN is improved by the gradient penalty and the new training strategy. Based on the improved GAN method, the small-sample local gasification condition is expanded to construct an augmented dataset. The results showed that by local data augmentation based on the improved GAN, the global RMSE of gas yield, LHV, and cold gas efficiency decreased by 8.70 %, 14.32 %, and 11.83 %, respectively. The global R2 of the prediction models for gas yield, LHV, and CGE reached 0.94, 0.93, and 0.97, respectively. The global and local RMSE of LHV after local data augmentation decreased by 37.18 % and 13.01 %, respectively, compared to traditional GAN, demonstrating that the improved GAN can improve the data-augmentation effect. In addition, the case further found that as the imbalance level increases, the global and local accuracy of LHV decreases, and the effect of the local data augmentation is also weakened. Beyond biomass gasification, the proposed method has the potential for broader applications in other fields where data imbalance and the need for accurate process prediction are common challenges. The improved GAN model could provide a new solution for enhancing model performance in these areas.