Abstract
<title>Abstract</title> <p>Driven by the global energy structure transition and the dual-carbon goals, multivariate time series forecasting (MTSF) for power load is of great practical significance. In the current field of MTSF, however, the pursuit of a universal architecture may be a misconception. In fact, existing universal models struggle to balance prediction accuracy with the limited computing resources of terminal power grid devices, which necessitates the design of a tailored and efficient MTSF architecture specifically for the power industry.State Space Models serve as a critical solution to this challenge. However, existing Mamba-based forecasting models suffer from inherent limitations such as redundant architectural stacking and insufficient mining of multivariate correlation relationships. To address these issues, this paper proposes a novel PatchMamS2S method. The patch representation mechanism is adopted to capture the periodic patterns of power load, and a physical prior-guided weight fusion strategy is designed to reconstruct seasonal and trend features. Furthermore, a lightweight Seq2Seq architecture named MamS2S is constructed based on Mamba-2 to capture temporal dependencies. This design avoids the blind deep stacking of Mamba modules, enables efficient dependency modeling at the linear complexity level, lowers hardware deployment barriers, and breaks the application bottleneck of high-complexity models on terminal power grid devices. Extensive experiments on five power-domain datasets and three datasets from other domains demonstrate that the proposed model achieves superior overall performance and strong generalization ability. Code at https://github.com/IHAN-1212/PatchMamS2S.</p>