SVBTformer: A Decomposition-Enhanced Hybrid Transformer for Long-term Time Series Forecasting

Time series forecasting is a fundamental task in many domains, such as finance, energy, and intelligent systems. It is increasingly important in modern computing environments, including cloud computing and distributed resource management. However, real-world time series often exhibit complex temporal dependencies, high volatility, and multi-scale nonlinear patterns, making accurate forecasting challenging. To address these issues, this work proposes SVBTformer, a novel and effective forecasting model that enhances the Transformer-based Informer architecture with structured temporal learning modules. Specifically, SVBTformer integrates Savitzky-Golay (SG) filtering for noise reduction and signal smoothing, followed by Variational Mode Decomposition (VMD) to extract multi-resolution temporal components. Then, an improved Informer network called BTformer is employed to enhance the modeling capability for time series and strengthen the extraction of temporal dependencies. This work extensively experiments on publicly available benchmarks spanning multiple domains, including the ETT dataset for electric power demand, foreign exchange rates, and meteorological measurements. The results demonstrate that SVBTformer consistently outperforms state-of-the-art models, such as Informer and Autoformer, across most evaluation metrics, delivering superior accuracy and robustness. These gains underscore SVBTformer's strong generalization capability and suitability for deployment in various real-world time series applications.