Short‐term prediction of behind‐the‐meter PV power based on attention‐LSTM and transfer learning

Abstract

Distributed photovoltaic (PV) systems often lack adequate measurements due to cost considerations, which makes it very difficult to predict them accurately. Here, an approach is proposed for behind‐the‐meter (BTM) PV power prediction using attention‐LSTM neural network and transfer learning. First, the weather is classified into four types based on the deviation ratio β. Second, the correlation analysis algorithm identifies the weather factors that contribute the most to PV power generation as GHI, DNI, humidity, and temperature. Then, attention‐LSTM pre‐trained model is constructed, in which the LSTM network fully extracts the temporal characteristics of PV power generation, while the attention mechanism enhances the attention to the important information in the input. Finally, a novel BTM PV short‐term power prediction method using transfer learning that the upper layer parameters of the pre‐trained model were frozen, and a small amount of DP‐NB data was used to fine‐tune the model. The proposed methodology has higher prediction accuracy compared with other benchmark methods under four weather types, and the time cost is saved by freezing the parameters of the first layer of the network by 37.2%.