Abstract: Running data from the photovoltaic power system is time indexed,which may be incomplete caused by low quality communication, and always contain the amount of abnormal data from the inverter. This paper studies the algorithm to preprocess photovoltaic data before being used to evaluate the whole system performance. The preprocessing includes labeling abnormal data and cleaning noise data. One optimized GRU neural network is used to do that, which is trained on our lab-built dataset. The GRU network is optimized to process photovoltaic data more efficiently with the activation function, loss function, and hidden layer. The best accuracy is as good as 99.84% from the test dataset consisting of actual photovoltaic data which is not used in training.