Abstract
Backend system load prediction is a fundamental issue for ensuring stable service operation, optimizing resource scheduling efficiency, and enhancing system autonomy. Addressing the common challenges in backend system load sequences, such as multi-timescale coupling, complex local fluctuations, volatile global states, and difficulties in fully characterizing temporal dependencies, this paper proposes a multi-scale dynamic evolution time series modeling method for backend system load change prediction. This method first organizes the original load sequence into multiple scales to separate change information across different time ranges. Then, it extracts key representations at each scale through a multi-branch time series coding structure and utilizes an adaptive fusion mechanism to enhance the synergistic effect between features at different scales. Furthermore, a dynamic evolutionary state update unit is introduced to jointly model historical operating states and current load information, thereby enhancing the model's ability to express complex temporal structures and key change features. Research on real open- source backend operating data shows that the proposed method can more effectively mine deep correlations in load sequences, demonstrating good comprehensive capabilities in error control, prediction accuracy, and overall stability. This research provides a targeted modeling approach for backend system load prediction and offers a methodological foundation for tasks such as intelligent resource orchestration, service capacity configuration, and system operating status awareness.
