Xufeng Wu, Min Chen, Nan Dong, Yuwen Wu, Zhanzhi Liu, Buyun Su
View Full Paper
Power load forecasting; Adaptive graph convolutional network;Multi-head attention mechanism; Multi-scale features; Spatio-temporal modeling
With the high proportion of renewable energy connected to the grid, the non-stationarity and multi-scale characteristics of power load are becoming increasingly prominent, which puts higher demands on prediction accuracy. Therefore, a multi-scale power load forecast- ing model combining an adaptive graph convolutional network and a multi-head attention mechanism was proposed. The core innova- tion lies in the fact that adaptive graph convolutional networks dy- namically evolve the spatial dependency relationships between nodes through learnable adjacency matrices, breaking through the limita- tions of traditional graph convolutional networks that rely on fixed topologies. The multi-head attention mechanism extracts multi-scale temporal features in parallel from the load sequence. The two work together to achieve deep integration of spatiotemporal features. Ex- periments on the ISO-NE and BuildingsBench datasets show that the model maintains the highest accuracy in both short-term and medium to long-term predictions, with root mean square error, mean absolute error, and mean absolute percentage error of 46.88MW, 34.29MW, and 5.13%, respectively. Its anti-interference ability and inference speed are also superior to mainstream comparison mod- els. The results indicate that the MS-AGCN-MHA model can ef- fectively improve the robustness and accuracy of load forecasting in complex power grid environments, providing reliable technical sup- port for real-time scheduling of smart grids.
Important Links:
Go Back
