IT Operations KPI Time Serie Forecasting Method Based on Time-Frequency Hybrid Attention Mechanism

As the scale of internet services continues to expand, traditional maintenance methods have become inadequate to meet the growing demands, leading to the emergence of AIOps (Artificial Intelligence for IT Operations). AIOps combines artificial intelligence, big data analytics, and machine learning to enhance operational efficiency and reduce manual intervention. In the AIOps framework, trend forecasting based on KPI time series data is one of the core technologies that effectively ensures the stable operation of systems. However, existing methods face challenges in handling complex periodic features and long-term dependencies, resulting in suboptimal prediction accuracy and high false-negative rates in anomaly detection. To address these issues, this paper proposes a time-frequency hybrid attention mechanism (TFHA) for prediction. The proposed method leverages a frequency-domain attention mechanism based on Fourier Transform to extract multi-scale periodic features, integrates a time attention mechanism to enhance the model's perception of temporal dependencies, and introduces a gating mechanism to dynamically fuse frequency and time features, thereby improving the model's ability to capture both short- and long-term dependencies. Experimental results demonstrate the effectiveness of the proposed method on the AWS and GAIA datasets.