As the installed capacity of photovoltaic power stations continues to grow, and with the advancement of power system reforms and new energy systems, the accuracy requirements for photovoltaic output forecasting have become increasingly stringent. A photovoltaic power station output model integrating MIC (maximum information coefficient), NRBO (natural randomized optimization), FMD (finite modal decomposition), and BiGRU (bidirectional gated recurrent unit) with an attention mechanism was proposed. The MIC method was employed to identify key meteorological factors affecting photovoltaic power output. Furthermore, NRBO and FMD techniques were innovatively applied to decompose photovoltaic output data into modal components, which helped eliminate high-frequency noise that could affect the training results. A bidirectional attention mechanism was then introduced into GRU model to enhance its ability to learn from complex time-series data, ultimately leading to more accurate forecasting of photovoltaic power station output. Experimental results demonstrate that the proposed model significantly outperforms other models in terms of forecasting accuracy.