Adaptive Multi-Scale Representation Learning for System Indicator Prediction in Complex Operational Environments

Abstract

This paper proposes an adaptive multi-scale representation learning method for system indicator prediction to address the challenges of non-stationarity, dynamic dependencies, and scale heterogeneity in multidimensional indicator sequences of complex systems. The proposed method employs a multi-scale convolutional decomposition module to extract features at different temporal granularities, capturing both short-term fluctuations and long-term trends in system state variations. An adaptive feature fusion mechanism is further introduced to dynamically weight and constrain cross-scale features, enabling joint modeling and balanced representation across multiple temporal levels. Structurally, the model integrates hierarchical normalization and gated update units to enhance the stability of feature flow and the continuity of temporal dependencies, effectively mitigating prediction degradation under high-frequency disturbances and distribution shifts. In addition, a residual propagation-based dynamic feature transformation layer is designed to achieve collaborative modeling between local information and global semantics, further improving the model's representational capacity and generalization for complex time-varying signals. Validation on representative system indicator datasets demonstrates that the proposed approach outperforms mainstream baseline models in key metrics such as MSE, MAE, MAPE, and RMSE, maintaining high accuracy and stability under complex dynamic conditions and providing an effective solution for intelligent prediction and robust modeling of multidimensional system indicators.