Accurate Photovoltaic (PV) generation forecasts can reduce power redeploy from the grid, thus increasing the supplier’s profit in the day-ahead electricity market. However, the PV process is affected differently by various factors under different weather conditions, resulting in significantly different energy output curves. In this context, this paper proposes a day-ahead PV power forecasting method with weather conditioned attention mechanism. We propose a Multi-Stream Attention Fusion Network (MSAFN) which utilizes an algorithm to derive the optimal decomposition algorithm for different weather conditions. The proposed Conditional Decomposition (CD) algorithm searches for the decomposition algorithms and corresponding hyperparameters of the prediction model, aiming to achieve the optimal prediction performance. The MSAFN incorporates multiple attention modules to learn the energy output patterns under various weather conditions. Notably, the attention modules adeptly learn patterns under diverse conditions, while simultaneously, the sharing of weights among the remaining components of the model effectively enhances prediction accuracy and facilitates a reduction in training time. We compare the state-of-the-art decomposition algorithms (VMD, EEMD, MSTL, etc.) and prediction models (BPN, LSTM, XGBoost, transformer, etc.) commonly used in PV prediction. The results show that the MSAFN model is more accurate than the models above, which has a noticeable improvement compared to other recent day-ahead PV predictions on Desert Knowledge Australia Solar Centre (DKASC) dataset.