Robust and efficient multihorizon photovoltaic power forecasting with a dilated multi-scale Transformer

Accurate and reliable multi-horizon photovoltaic (PV) power forecasting is essential for seamless grid integration, economic dispatch, and power system stability. However, the inherent variability of solar irradiance and complex nonlinear dynamics pose significant challenges. To address these issues, this study proposes WinFlashDilated-former, a model specifically designed for efficient LSTF in the PV domain.Built upon an encoder-decoder architecture, the model introduces three key innovations to overcome the limitations of existing methods:(1) a precise multi-scale window self-attention mechanism deeply optimized with FlashAttention-3, which captures multi-scale dependencies, from short-term fluctuations to long-term trends, without loss of information while achieving excellent computational efficiency;(2) a dilated convolutional feed-forward network in the encoder, which expands the receptive field to enhance modeling of long-range contextual information; and (3) a time-enhanced windowed autoregressive generation strategy in the decoder, which strengthens the perception of periodic patterns by leveraging future temporal attributes.Extensive experiments were conducted on real-world datasets provided by DKASC, covering multiple PV technologies and multi-year operational records. The results demonstrate that WinFlashDilated-former consistently outperforms both traditional time-series models and several SOTA LSTF approaches across all forecasting horizons from 5-min to 12-h. In the 5-min horizon, the model achieved MAE(0.0009), RMSE(0.0146), and R²(0.9972). In the 12-h horizon, compared to the baseline method, MAE and RMSE were reduced by 13.2% and 12.03%, respectively, while maintaining strong robustness.Ablation studies further confirm the critical contributions of the multi-scale attention mechanism and dilated convolutional feed-forward network. Moreover, WinFlashDilated-former achieves an excellent balance between accuracy and efficiency, demonstrating competitive inference speed and GPU memory utilization. The code and data are publicly available at: https://github.com/YingyingLiu2002/winflashdilated-former.