To address the limitations of current photo voltaic power forecasting methods，which rely heavily on meteorological monitoring and classification techniques and are unable to achieve accurate all-weather forecasting under large-scale complex data，a Multiformer-TSA method based on a two-stage attention （TSA） mechanism is proposed for photovoltaic power forecasting.First，a cross-scale embedding layer is constructed to generate staged sampling tokens and collect photovoltaic sequences at different scales，thus extracting cross-scale features.Then，point-based segments of different dimensions in multivariate time series are embedded to form new feature vectors，capturing cross-dimensional dependencies.Finally，cross-scale and cross-dimensional dependency information is fused through the TSA mechanism to achieve accurate all- weather photovoltaic power forecasting.Multi-scale forecasting comparison experiments and ablation experiments are conducted on a publicly available photovoltaic power dataset from Alice Springs，Australia.The experimental results demonstrate that the proposed method accurately captures cross-scale and cross-dimensional features of multivariate time series and improves the multi-scale forecasting accuracy of photovoltaic power generation.Compared with existing methods，the proposed method achieves the best performance in terms of root mean square error （RMSE） and mean absolute error （MAE）.