TDformer: a domain knowledge-constrained transformer model for photovoltaic power forecasting

This is the authors' abstract. We don't add key points for this paper.

Photovoltaic (PV) power forecasting is essential for the reliable integration of renewable energy into power systems. However, it remains a challenging due to the complex temporal patterns of PV generation and the need for physically consistent predictions. This paper addresses the challenges of insufficient collaboration between long-term and short-term features and the lack of physical consistency in PV power forecasting. A model based on the Transformer architecture, termed TDformer, is proposed by integrating time frequency feature modeling with domain knowledge constraints. Specifically, the Discrete Cosine Transform (DCT) is employed to decompose the original sequence into frequency components, and a Temporal Convolutional Network (TCN) is introduced to construct a dual-branch time frequency collaborative modeling framework, enabling adaptive extraction of multi-scale periodic patterns and short-term fluctuations. Meanwhile, a lightweight physics informed constraint mechanism is incorporated at both input and output stages, including sunrise and sunset based data filtering and non-negative output constraints, to ensure physically consistent predictions. Experimental results on the Sanyo and Solibro datasets demonstrate that the proposed model consistently outperforms representative baseline methods across different forecasting horizons, with improved accuracy and stability.