Open AccessA multi-task learning framework for multi-location short-term load prediction
Author(s) -
Zhaorui Meng,
Jun Sun
Publication year - 2021
Publication title -
iop conference series. earth and environmental science
Language(s) - English
Resource type - Journals
eISSN - 1755-1307
pISSN - 1755-1315
DOI - 10.1088/1755-1315/651/2/022083
Subject(s) - computer science , regularization (linguistics) , task (project management) , term (time) , graph , machine learning , power consumption , multi task learning , artificial intelligence , predictive power , similarity (geometry) , data mining , power (physics) , engineering , theoretical computer science , philosophy , physics , systems engineering , epistemology , quantum mechanics , image (mathematics)
Short-term power load prediction is crucial to management of power system. The traditional load prediction methods are based on learning data from single location. However, load consumption is related among different locations. In this paper, we propose a novel approach to train load predictors for multi-locations in a collaborative way based on multi-task learning. Specifically, the load predictor in each location is splitted into two parts: a general one and a location-specific one. The general predictor is to capture information shared by various locations. And the location-specific predictor is to obtain location-specific information. In addition, a location similarity graph is built and incorporated into the model as regularization. Extensive experiment result shows that our approach outperformed single task learning method and another two multi-task learning methods in at least 80% of locations.