Multi-Task Learning Based on Deep Architecture for Various Types of Load Forecasting in Regional Energy System Integration

Abstract

To alleviate environmental pollution and improve energy efficiency, energy system integration (ESI) becomes an important paradigm in energy structure evolution. Power, gas and heat systems become tightly interlinked with each other in ESI. Accurate energy loads forecasting has significant impact on ESI dispatching and optimal operation. This paper proposes a method of short-term hourly load forecasting for various energy sources based on deep multi-task learning. Firstly, the algorithm architecture consists of a deep belief network (DBN) at the bottom and a multi-task regression layer at the top. The DBN can extract abstract and effective characteristics in an unsupervised fashion, and the multi-task regression layer above the DBN is used for supervised prediction. Then, a two-stage load forecasting system based on off-line training and on-line prediction is deployed subject to the conditions of practical demand and model integrity. Finally, validity of the algorithm and accuracy of the load forecasts for ESI system are verified with simulations using actual operating data from load system. Results demonstrate that deep learning and multi-task learning are promising approaches in energy demand forecast research.