Research on Short-term Load Forecast Using Multi-task With Low-rank Representation

For load forecasting in power system, latent relationships among different locations are not considered by traditional single task learning method. Neglecting the associated information transfer between multiple locations may lead to poor learning. In order to use latent relationships fully, an approach is proposed to forecast load for multi-locations simultaneously by using multi-task learning with low-rank representation. By extracting shared low dimension representation among load forecasting models at different locations, the approach can mine relationships between multiple tasks, and distinguish differences between different tasks. It is verified by simulation result that the multi-task load forecasting performs on average better than single-task learning algorithms such as decision trees and random forests, and the load forecasting accuracy has been improved.