Time Series Forecasting with Multi-Scale Feature Extraction and Explicit Periodic Modeling

Time series forecasting is critical in many applications such as electrical power scheduling and weather prediction. However, existing models often insufficiently extract local details and inefficiency use computational resources when handling complex time series data. Although attention-based models improve the global dependency modeling via sparse attention designs, their capability for explicit modeling of short-term fluctuations and periodic patterns remains limited. To address these issues, this paper proposes an Informer-based time series forecasting framework that enhances the local feature extraction and dynamic periodic modeling to improve prediction accuracy and efficiency. Specifically, a local feature extraction mechanism is designed to extract multi-scale local features and capture short-term fluctuations; A explicit periodic feature modeling module is established to identify dominant periods and enhances periodic patterns via dynamic weight fusion. Experimental results demonstrate that the proposed model reduces prediction error (MSE) by 49.34 % compared to the original Informer model across four benchmark datasets, covering three practical domains: electricity, weather, and energy. Ablation studies show that the local feature extraction and periodic modeling modules contribute 26.73 % and 36.69 % error reductions individually.