A novel dynamic integration approach for multiple load forecasts based on Q‐learning algorithm. (29th July 2019)

• Record Type: Journal Article
• Authors: Ma, Minhua; Jin, Bingjie; Luo, Shuxin; Guo, Shaoqing; Huang, Hongwei
• Other Names: Wang Yi guestEditor.
• Is Part Of: International transactions on electrical energy systems. Volume 30:Number 7(2020)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2019-07-29
• Subjects: action space -- environmental status -- multiple load forecasts integration -- Q‐learning algorithm -- reinforcement learning -- return function; Electric power -- Periodicals; Electric power systems -- Periodicals; Electrical engineering -- Periodicals; 621.3
• DOI: 10.1002/2050-7038.12146
• Languages: English
• ISSNs: 2050-7038

Abstract

Summary: A higher accurate load forecasts can significantly reduce the potential cost for power system operation and power market decision. There have been extensive load forecasting methods proposed by now. Ensemble learning is an effective way to integrate various forecasts and further improve the forecasting accuracy. Traditional ensemble learning algorithms static. However, how to select a suitable forecasting method from a large number of load forecasting methods under the complex and varying environment conditions is still challenging. In this context, this paper proposes a dynamic integration approach for multiple load forecasts based on reinforcement learning. Compared with the traditional integration methods, this method adopts Q‐learning algorithm to realize the adaptive selection of the forecasting methods and dynamic weight calculation for different and varying environment. The basic principles and implementation processes of Q‐learning algorithm is firstly introduced. Then, the implementation methodology of dynamic integration approach of multiple load forecasts is proposed. On this basis, the environment state, action space and return function are designed according to the requirements of Q‐learning algorithm. Finally, a case study based on the real‐world data in Guangzhou shows that the average forecasting accuracy could be improved by almost 3% by adopting the dynamic integration approach with the traditional ones.