Bidirectional Generative Pre-training for Improving Time Series Representation Learning