A Multi-Scaler Hybrid Autoformer for Enhanced Time Series Forecasting in Energy Consumption
Wenhao Wang, Xiuqin Li, Pengjia Yan
Abstract
The Autoformer model exhibits limitations in capturing local information in time series forecasting, which is crucial for improving prediction accuracy. To address this, this paper proposes the Multi-Scaler Hybrid Autoformer (MSHA) model, which combines multi-scale series decomposition modules and a hybrid attention mechanism to enhance long-term dependency modeling and multi-scale feature extraction. The study utilizes the classic time series forecasting dataset, Appliances Energy Consumption, and conducts experiments with three different forecasting horizons (36, 72, and 144 steps). Results demonstrate that the MSHA model consistently outperforms others across all horizons, achieving an average MSE reduction of 83.33%(from 0.0012 to 0.0002). The MSHA model effectively addresses the local feature extraction shortcomings of existing models, not only enhancing time series forecasting accuracy but also offering valuable insights for future applications in other time series tasks. This model holds significant practical value in energy consumption forecasting, providing a robust foundation for enhancing accuracy across various time series prediction tasks.