For quick and accurate retrieval in massive grid events, this study proposes the parallel similarity calculation of grid events based on keyword generation. The method generates grid event keywords through the pointer-generator network based on the encoder of a bidirectional LSTM network and the decoder of a unidirectional LSTM network. It uses a memory network to store sequence information and applies the attention mechanism to the input sequence, enabling more important information into the decoder. It also introduces the overwriting mechanism to avoid duplicate texts. After the keywords are generated, the overall similarity is calculated based on the structural similarity and situational similarity of events. GPU is utilized to accelerate the LSTM network and similarity calculation. Experimental results show that the method has better performance in similarity calculation than methods based on machine learning, with a speedup ratio reaching 4.04 times.