A multi-scale patch embedding and dual-view feature learning network for multi-step forecasting in high-reliability low-carbon energy systems: Case studies in nuclear and wind

As the global energy system undergoes a low-carbon transition, modern power systems are increasingly reliant on highly variable renewable energy and stable zero-carbon baseload nuclear power. Accurate multivariate multi-step forecasting is critical to ensuring their reliable operation. However, existing methods face four challenges: insufficient modeling of multivariate physical couplings, limited ability to capture multi-scale temporal dynamics, prominent trade-off between accuracy and efficiency, and weak cross-scenario generalizability. To address these limitations, this paper proposes a multivariate multi-step forecasting framework based on Multi-Scale Patch Embedding and Dual-View Feature Learning Network (MPDFNet). The model integrates multi-scale patch embedding, seasonal-trend decomposition, and dual-view feature learning to model multi-scale temporal dynamics and cross-variable dependencies. A parallel decoder is employed to mitigate error accumulation in multi-step forecasting. The framework is rigorously validated on four real-world datasets and compared against nine state-of-the-art models. Experimental results demonstrate that MPDFNet outperforms all baseline models in both main experiments: achieving an average R2 > 0.99 for 20-step nuclear accident forecasting, and an average MAPE of 4.81% for 12–48-step wind speed forecasting. The generalization capability of MPDFNet is validated in two supplementary cross-scenario experiments. Furthermore, MPDFNet exhibits exceptional computational efficiency, completing 200-s nuclear forecast in just 0.21-s and training for a 4-h wind forecast in approximately 10-min. Interpretability analysis reveals the information-driven behavior of the attention mechanism, while ablation studies confirm the contributions of each core component. This study proposes an accurate, lightweight, generalizable, and interpretable forecasting framework to support the safe and reliable operation of low-carbon energy systems.