Abstract: As an important index of stable operation of power system, power load is difficult to achieve accurate prediction due to its nonlinearity and timing. As the ordinary nonlinear autoregressive exogenous（NARX） model is difficult to properly capture the long-term time dependence and select the relevant driving series for prediction, a recursive LSTM neural network based on double attention mechanism is established in this paper. In the encoder stage, the LSTM structure is used as the nonlinear function of the input data mapped to the hidden state to capture the long-term dependence of the hidden state. At the same time, the input attention mechanism is introduced to adaptively extract the input features, that is, to calculate the attention weight of the relevant driving series to the driving sequence when predicting the target sequence. In the decoder stage, before decoding with LSTM structure, this paper adds a time attention mechanism to adaptively select the importance weight of the hidden state of the relevant encoder to time when predicting the target sequence. Using this two-stage attention mechanism RNN network, the model in this paper can not only effectively predict power load, but also be easy to explain and have good robustness.