Abstract
Univariate time series forecasting is pivotal in domains such as climate modeling, finance, and healthcare, where both short-term precision and long-term reliability are essential. This study introduces AUNET (Attention-based Unified Network)—a novel extension of the N-BEATS model—integrating multi-head self-attention to capture both short-term fluctuations and long-term dependencies while reducing architectural redundancy and enhancing interpretability. Empirical evaluations on the (baseline dataset for this research) CIMIS weather dataset show that AUNET achieves a 0.3580 MAE, 0.4632 RMSE, 0.68% MAPE, and an R2 of 0.9988, outperforming baseline models including N-BEATS and attention-augmented variants by margins exceeding 87% in MAE and RMSE. AUNET’s generalization capabilities are further validated on diverse domains: in finance it attains an R2 of 0.9990, and in healthcare, it outperforms baselines in RMSE and R2 despite inherent volatility. Statistical assessments confirm the significance and robustness of these improvements. Moreover, AUNET demonstrates reduced training time, lower FLOPs, and higher convergence stability. Its attention-guided modular architecture not only enhances predictive accuracy but also promotes transparency, making AUNET a compelling candidate for deployment in high-stakes, real-world forecasting tasks across varied domains.