Enabling Effective Computing-Power Collaboration through VAE-based Time Series Generation

Effective computing-power collaboration is a critical challenge in IoT-driven power systems, where noisy, sparse, and imbalanced time-series data from IoT devices hinders intelligent decision-making. Traditional augmentation methods increase the amount of data but often fail to capture temporal dependencies and statistical properties essential for robust downstream performance. We propose a Variational Autoencoder (VAE) based Time Series Generation (TSG) method for augmentation, improving model performance and robustness. By using a VAEbased model to learn latent representations of time series data, it can effectively generate synthetic sequences that preserve both statistical structure and temporal coherence. Our results demonstrate enhanced accuracy and generalizability across time series classification, anomaly detection, and demand forecasting tasks, highlighting the potential of VAE-based TSG to optimize resource allocation and improve system reliability, particularly in environments with limited and noisy data.