Abstract:Accurate photovoltaic power prediction plays an important role in optimizing the operation and management of photovoltaic power stations and improving the efficiency of photovoltaic power generation. In this paper, a short-term photovoltaic power prediction method based on clustering algorithm and conversion network is proposed. In this method, short-term photovoltaic power data is preprocessed based on unsupervised clustering algorithm of autoencoder to reduce the influence of instability of photovoltaic output data on power prediction. Then, the method uses a conversion network with self-attention mechanism and multi-attention mechanism to predict short-term photovoltaic power. The conversion network consists of encoder and decoder. Compared with traditional recurrent neural networks (RNN), transformational networks are better at mining the relationship between time sequences. The attention mechanism makes the switching network have the capability of parallel computation, which can accelerate the speed of network training. Finally, the proposed method is verified on the photovoltaic data set of the Australian Photovoltaic Power and Meteorological Data Center (DKASC). Experimental results show that the proposed method has satisfactory prediction accuracy.