Abstract:The development of smart grid makes the data obtained from the grid gradually increasing. In order to obtain useful information from multidimensional big data and accurately predict the shortterm power load, this paper proposes a shortterm power load forecasting method based on dimension reduction with improved deep sparse autoencoder (IDSAE) and extreme learning machine (ELM) optimized with fruit fly optimization algorithm (FOA). Adding L1 regularization to the deep sparse autoencoder （DSAE） can induce better sparsity, and the improved deep sparse autoencoder is used to reduce the dimensionality of highdimensional data that affects the accuracy of power load prediction, which eliminates the multicollinearity among the indexes and realizes compression coding from highdimensional data to lowdimensional space. The fruit fly optimization algorithm (FOA) is used to optimize the weights and thresholds of the extreme learning machine (ELM), and the optimal weights and thresholds are obtained, which can overcome the shortcomings of low prediction accuracy caused by the extreme learning machine randomly selecting the weights and thresholds. In this paper, the meteorological factors are first dimension reduced by the IDSAE to obtain the sparse comprehensive meteorological factor characteristic indexes, and the coordinated power load data are used as the input vector of the FOA optimized ELM prediction model to perform power load prediction. The comparison experiments with DSAEFOAELM, DSAEELM, IDSAEELM and other models prove that the proposed prediction model can effectively improve the prediction accuracy, and the improved accuracy is about 8% after calculation.