Abstract: Under the goal of carbon peak and neutrality,for the problem of harmonic influence caused by temperature and illumination changes on the grid-connected photovoltaic system,we propose a harmonic prediction method based on ip-iq harmonic extraction method and improved bi-directional long and short-term memory network（BI-LSTM）to provide a new solution for harmonic suppression. Firstly,the grid-connected photovoltaic system is established by MATLAB/SIMULINK tool,and the actual harmonic variation data are obtained by ip-iq harmonic extraction method based on instantaneous reactive power theory, and the data are simplified by differential interpolation. Secondly,the grid search optimization BI-LSTM neural network algorithm is used to predict harmonic data,and compared with BP,LSTM,GRU and BI-LSTM deep learning methods,and the LOSS function and prediction results of MSE,MAE and MAPE are obtained. Finally,a practical example in Longdong area is used to simulate the grid connection of PV,and the results show that the method can accurately predict the harmonic.