Computing Power Scheduling Method Based on Long-Term Workload Prediction for Computing Power Platform

Users' requests for the computing power often exhibit suddenness and unpredictability, which creates challenges for workload scheduling of computing power platforms. Traditional scheduling methods struggle to adapt to such situations as they typically schedule based on real-time load conditions, ignoring long-term trends in load variations. To address these limitations, we propose a computing power scheduling method based on long-term workload prediction. First, we leverage a tree structure to aggregate temporal-spatial information from multi-variate operating time-series data, effectively describing the characteristics of workload on computing power platform. Building upon Informer, we employ residual mechanisms and dilated causal convolutions to extract deep features. Additionally, we integrate Wasserstein Generative Adversarial Network (WGAN) with Informer to accurately forecast future changes in workload for computing platforms, and introduce a novel training algorithm to enhance efficiency. We validate our method through experiments using real-world workload datasets, demonstrating the significant reduction in Mean Absolute Error (MAE) compared to LSTM. Furthermore, our model exhibits sensitivity to the length of encoder input sequences, indicating improved performance with longer series inputs.