Abstract: In the context of smart grids, accurate estimation and prediction of power load has become an important prerequisite for grid power planning, and is of great significance to grid operating safely and economically. Aiming at the cyclical fluctuation and non-periodic influence of power load data, a short-term power load forecasting method based on dual graph regularized non-negative low-rank decomposition is proposed. The method constructs the power load space-time matrix using historical data, and performs robust non-negative low-rank matrix decomposition on the matrix to simultaneously acquire the periodic mode and non-periodic influence of the power load. On this basis, the spatial and temporal correlation of the power load is integrated to further improve the matrix decomposition results, and finally the short-term prediction of the power load is obtained through matrix recovery. This method takes into account the missing complement and trend analysis of power load and derives an effective learning algorithm. The experimental analysis shows that compared with the related methods, the proposed method achieves better accuracy and robustness under multiple evaluation criteria of short-term forecasting of power load.

Key words: forecasting of power load, matrix completion, low-rank representation, non-negative matrix decomposition