FreqMLNet: Non-transformer network with frequency domain reconstruction and multi-scale representation for time series forecasting

Time series forecasting is essential in finance, meteorology, healthcare, and industrial process control. Traditional time-domain forecasting methods struggle to capture complex patterns and structures. Frequency-domain analysis offers an alternative perspective for identifying periodic and oscillatory behaviour. However, current frequency-domain methods often lose vital information by excluding high-frequency components, which contain crucial short-term variations and abrupt changes that impact forecasting accuracy. Additionally, existing multilevel analyses inadequately integrate local and global features, making it difficult for forecasting algorithms to capture both detailed local information and broader global trends. To address these issues, this study introduces FreqMLNet, a novel non-transformer architecture that combines frequency-domain reconstruction and multilevel feature representation to enhance time series forecasting. FreqMLNet extracts periodic patterns through a frequency-domain reconstruction module and integrates information from multiple scales using multilevel feature representation, resulting in a more comprehensive time series feature representation. Experimental results demonstrate that the proposed model achieves an average improvement of 14.39% in mean squared error across seven long-term datasets and 11.03% in symmetric mean absolute percentage error across four short-term datasets compared to state-of-the-art models. Moreover, an in-depth analysis reveals that FreqMLNet exhibits greater robustness and prediction accuracy, particularly in complex time series forecasting tasks.