RampMixer: An attribution-aware Tiny Time Mixer for photovoltaic power forecasting under coupled environmental dynamics

With the growing penetration of photovoltaic (PV) generation in power systems, reliable PV power forecasting has become important for generation scheduling and secure dispatch. However, power variations during ramp events are often governed by the coupled dynamics of multiple environmental drivers, which obscure the dominant causes of abrupt power changes and undermine forecast reliability. This paper therefore presents RampMixer, an attribution-aware framework that leverages time-varying environmental driver contributions for PV power forecasting under ramp conditions. The framework evaluates the time-varying contribution of each environmental driver to ramp-related power changes and encodes these contributions in a coupling attribution matrix. This matrix guides the decomposition of coupled ramp power signals into contribution-aligned modes, separating components governed by different drivers and reducing interference from multi-driver interactions. A ramp-oriented tiny time mixer then learns temporal dependencies within each decomposed mode through a temporal-mixing structure that captures ramp dynamics regulated by environmental drivers while maintaining the mode-specific organization. Analysis of the decomposed modes indicates that the proposed contribution-aligned decomposition effectively reduces interference caused by environmental driver coupling. Experiments on two PV datasets demonstrate that RampMixer outperforms state-of-the-art models across different ramp intensities, achieving root mean square error reductions of at least 5.73% on UNISOLAR and 7.63% on HIT05667 relative to the evaluated baselines under both low-ramp and high-ramp conditions. These results demonstrate the effectiveness of the proposed framework for reliable PV power forecasting under coupled ramp dynamics.