ANTDCC: A Lightweight and Generalizable Framework for Long-Term Photovoltaic Power Forecasting via Adaptive Normalization and Dynamic Channel Clustering

The intermittency and uncertainty of photovoltaic (PV) power generation have posed significant challenges to grid stability and efficient integration of renewable energy, especially in long-term forecasting tasks where accurately modeling multi-channel temporal dependencies and adapting to dynamic distribution shifts have become critical issues. To address the limitations of existing methods in channel modeling strategies and normalization mechanisms, this paper has proposed an efficient PV power forecasting framework—ANTDCC—that integrates adaptive normalization, dynamic channel clustering, and TSMixer. First, an Attention-alike Structural Re-parameterization Dynamic tanh (ASRDyT) has been introduced to enhance the model adaptability to temporal distribution changes. Second, TSMixer has been employed as the backbone network, leveraging its linear computational complexity and efficient time-feature mixing capability to effectively capture global temporal dependencies while balancing accuracy and efficiency. Third, a Dynamic Channel Clustering Mechanism (DCCM) has been designed to excavate inter-channel synergies, improving model interpretability and parameter utilization. Finally, systematic evaluations on PV datasets from Australia and China have demonstrated that the proposed method has significantly outperformed baseline models across four forecasting horizons, exhibiting superior generalization ability, thereby providing effective support for intelligent PV plant scheduling and optimized power system operation.