Abstract: Combined with the feature extraction ability of self-coding and the time series analysis ability of

long short-term memory（LSTM），a short-term load forecasting model based on LSTM self-coding was proposed.

First of all，based on the basic principle and flow of the depth model，the functional characteristics of the automatic encoder were given. Secondly，the outlier judgment method based on Goblas criterion and the missing value completion method based on Lagrangian spline interpolation were proposed. Finally，using Tensorflow architecture to call Keras library to build an experimental platform to carry out forecasting experiments on light industrial load，heavy industrial load and photovoltaic industrial user load respectively，which verifies the conclusion that LSTM self-coding model is more suitable for short-term industrial load forecasting.

    Key words: load forecasting；industrial load；deep learning；long short-term memory（LSTM）；automatic

encoder