An Adaptive Multi-Scale Framework for Accurate Forecasting of Performance Metrics in Complex

Abstract

This study proposes an adaptive multi-scale representation learning method for system metric prediction to address the dynamic characteristics and non-stationary distributions of multi-dimensional metric sequences in complex systems. The method extracts temporal features at different time granularities through a multi-scale convolutional feature decomposition module, capturing both short-term fluctuations and long-term trends in system state changes. An adaptive feature fusion mechanism is introduced to dynamically weight multi-scale features and enforce consistency across scales, thereby enhancing the model's capability to represent complex time-varying patterns. Structurally, the model integrates hierarchical normalization and gated update units to improve the stability of feature flow and the continuity of temporal dependencies, avoiding prediction degradation under high-frequency disturbances and distribution shifts. In addition, a residual propagation-based dynamic feature transformation layer is constructed to jointly model local information and global semantics, further improving robustness and generalization in multi-dimensional signal interactions. Experimental results show that the proposed method achieves lower MSE, MAE, MAPE, and RMSE values than mainstream models such as Autoformer, EDFormer, and TimesNet on benchmark system metric datasets, confirming its superiority in multi-scale feature reconstruction and dynamic temporal modeling. This research provides an efficient and scalable modeling framework for system performance prediction, intelligent operations, and multi-dimensional time-series analysis, enabling accurate forecasting and structured representation of non-stationary sequences in complex system environments.