Long-term time series forecasting is essential for predicting future values based on historical patterns over extended periods. However, it poses significant challenges in effectively capturing both global trends and local variations. This research introduces FusionNet-FR, a novel DualNet Transformer architecture based on ensemble deep learning, which integrates fast-learning with slow-learning components to improve forecast accuracy. The fast learners are designed to extract high-frequency components and short-term variations, enabling the model to adapt to rapid changes in data patterns. In contrast, the slow learner is responsible for modeling long-term dependencies by leveraging global temporal structures through an inverted self-attention mechanism. In addition, incorporating frequency domain representation and interseries dependency modeling improves the robustness of the model against noisy data. To further strengthen learning, a controlled reconstruction loss strategy is integrated into the slow learner, facilitating self-supervised learning mechanisms. We evaluate FusionNet-FR on real-world multivariate and univariate datasets—Exchange, ETT, Weather, and ILI—using mean squared error (MSE) and mean absolute error (MAE) as performance metrics. FusionNet-FR achieves consistent improvements across both settings, with average percentage gains of 38.08% in MSE and 29.06% in MAE over baseline models on multivariate datasets, and 51.58% in MSE and 36.53% in MAE on univariate datasets. These results demonstrate the model’s effectiveness in addressing the inherent challenges of long-term time series forecasting, delivering superior accuracy compared to existing approaches.
Shiddiqi et al. (Tue,) studied this question.