The Long- and Short-Term Time Series Forecasting Based on L-Former

Time series forecasting is a fundamental task in fields such as finance, meteorology, and energy, as it facilitates the analysis of historical trends and the prediction of future changes. Traditional methods such as ARIMA and XGBoost perform well for short-term forecasting; however, they face limitations when addressing complex periodicity and long-term dependencies. In recent years, Transformer-based deep learning models have demonstrated exceptional performance in long-sequence forecasting, yet their high computational complexity and input length constraints impede the effective capture of multi-period signals. Conversely, while simple linear models can efficiently capture long-term trends and seasonal variations, they struggle with intricate multi-period data. To overcome these challenges, this paper introduces L-Former, a hybrid network model that integrates a Transformer with a linear model. Specifically, the proposed approach first decomposes the raw time series data and employs Fourier transform along with sequence segmentation to enhance the Transformer encoder's ability to capture complex, multi-period seasonal patterns in long sequences. Simultaneously, the linear model efficiently extracts simple trend components. By synergistically combining the strengths of each sub-network module, the model achieves improved overall performance, enabling effective handling of both short-term and long-term forecasting tasks. Experimental results on six datasets demonstrate that L-Former significantly enhances forecasting accuracy for both short-term and long-term time series prediction tasks.