Deep attention convolutional neural network-based adaptive multi-source information fusion for accurate short-term photovoltaic power forecast

Solar energy is one of the most important renewable energy sources. Photovoltaic (PV) power generation is currently one of the most important ways to utilize solar energy. PV power has strong uncertainties and is adverse to the stability of the electricity grid. PV power forecast is an important method to tackle this issue. Conventional methods only use historical PV power data. This paper proposes a new method that uses deep Convolutional Neural Networks (CNN)-based multivariable information fusions for PV forecast. Multiple variables including global horizontal irradiation (GHI), relative humidity, air temperature, cloud thickness, wind speed, etc. are used for PV forecast. Meanwhile, clear-sky GHI estimated by McClear clear-sky models are used as physical prior knowledge. Through the attention mechanism, these variables are adaptively fused with the historical PV power data to realize an accurate PV power forecast. Forecast experiments in three-year (2017-2019) actual data of Brussels PV power stations verify the significant superiority of the proposed method over conventional forecast methods.