The invention discloses a short-term wind-power power forecasting method based on a coding / decoding long-short term memory network. The method comprises steps of firstly, using an E-D based LSTM network to carry out AE processing on power, and extracting a trained network middle state as abstract expression of a time sequence relation in WP data; then, combining the network middle state extractedin the first step with weather data in a prediction period, and inputting the result into a new LSTM network, thereby finishing prediction of wind-power power. Compared with a multi-layer LSTM networkmethod where AE preprocessing is not used, according to the invention, by use of the WP time frequency relation information extracted in the AE process, the model misspecification risk of a model isreduced; the generalization ability is improved; and by combining the time sequence characteristic and the weather prediction information, the prediction precision is further improved.