Towards Intelligent Backend Operations via Unified Temporal Representation and Prediction Modeling

Abstract

This study proposes a unified modeling approach for long- and short-term temporal dependencies to address the challenges of complex metric structures, dynamic distribution changes, and intertwined multidimensional dependencies in backend systems, aiming to achieve high-accuracy and high-robustness system state prediction. The method first performs structured preprocessing and normalization on multi-source monitoring data, extracts multidimensional temporal features through a feature encoding module, and constructs a dual-channel architecture for short-term dynamic modeling and long-term context capture to fully characterize temporal properties of local fluctuations and global trends. On this basis, a multi-scale fusion mechanism is introduced to enable adaptive interaction and integration of features across different temporal granularities, enhancing the model's ability to represent non-stationary signals, sudden load changes, and cross-dimensional dependencies. In addition, residual calibration and dynamic aggregation strategies are designed to mitigate feature shifts in high-noise environments, ensuring stable prediction performance and strong generalization under complex operating conditions. Validation across various dynamic scenarios shows that the proposed method outperforms existing baseline models on key metrics such as MSE, MAE, MAPE, and RMSE, while exhibiting strong adaptability to data disturbances, environmental changes, and sampling granularity variations. The results demonstrate that the proposed approach effectively captures multi-level dependencies in complex time series, providing strong technical support for state modeling, performance prediction, and anomaly detection in backend systems, and laying a methodological foundation for building high-precision intelligent operations and maintenance systems.