Enhancing Long-Term Time Series Forecasting via Hybrid DLinear-PatchTST Ensemble Framework

Long-term time series forecasting is critical for strategic decision-making in energy management and supply chain optimization domains. While deep learning models like DLinear and PatchTST have advanced forecasting accuracy, existing approaches often fail to systematically integrate linear and nonlinear modeling for extended horizons, limiting their ability to address complex temporal dynamics. This study proposes a novel ensemble framework combining DLinears linear trend modeling with PatchTSTs transformer-based pattern recognition to improve long-term forecasting robustness. Experiments on the ETTh1 temperature dataset demonstrate that the hybrid model achieves a 2.7% reduction in mean square error (MSE) over standalone PatchTST and 1.1% improvement compared to DLinear for 720-step forecasts. The ensemble outperforms baseline models (Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing) across all tested horizons (96720 steps), with mean absolute error (MAE) values consistently below 0.45. Results validate that integrating linear and nonlinear approaches captures both seasonality and complex disruptions more effectively than single-model strategies. This work advances ensemble methodologies for time series analysis, offering a scalable solution for applications requiring reliable long-term predictions.