Real-time monitoring and time-series collecting system often faces to the issue of disruption due to common issues such as hardware/software failure or network disconnection, leading to the problem of missing data or time-delayed data. Online learning-based time-series prediction model trained by those missing data could be affected negative impact and degrade its performance. Recently, a method (namely, ERL) leveraging time-delayed complete data to enhance the time-series representation learning is introduced and shown the efficacy to address this time-delayed issue. In this paper, we present technical extension of ERL to reduce the resource consumption during training process.