Online performance prediction using the fusion model of LightGBM and TabNet for large laser facilities

Performance prediction is a crucial aspect of ensuring optimal facility operation and maintenance across diverse fields. Performance predictions for large-scale optical integrated systems in inertial confinement fusion (ICF) can enhance output performance and facilitate predictive maintenance of optical components. Yet, there is no existing implementation for the known inertial confinement fusion (ICF) facilities, which rely entirely on expert experience to estimate each optical path in the past manually. Although deep learning algorithms have found numerous applications in various industries, their applicability in tabular data mining that integrates multiple types of parameters has remained somewhat limited. To address this gap, this research proposes introducing an integrated learning model based on practical business scenarios. This model includes data cleaning and well-founded feature engineering methods. The study specifically utilizes the LightGBM and TabNet fusion models to address inefficiencies associated with traditional ensemble learning techniques, such as XGBoost and GBDT. Furthermore, a recursive feature elimination approach is implemented to reduce redundancy during model construction, resulting in improved online performance prediction. In simulations conducted on large-scale laser facilities, the predictive bias MAE of this model was reduced by 32.66% compared to traditional artificial prediction methods through the utilization of time series cross-validation. These findings underscore the significance of the integrated learning model in enhancing predictive accuracy, which is a crucial component of effective facility management.