The rapid development of electric vehicles will affect the charging pile load on the power grid. Therefore, a charging pile load prediction method is proposed by using deep learning quantile regression. In this method, based on historical data, the Adam stochastic gradient descent method is used firstly to train the LSTM neural network parameter estimation in different quantile conditions. Then the results in each quantile condition of the next 96 hours are predicted. After that, the probability density function of results at the same time is built by using the kernel density estimation. Finally, the load probability density prediction is obtained. According to the actual charging pile load results, the proposed probability density prediction method can predict the real value more accurately. Furthermore, it has higher accuracy and reference value than the BP neural network quantile regression method. This work is supported by Education Innovation Platform Open Fund of Hunan Provincial Department (No. 17K001).