A novel building energy consumption prediction method using deep reinforcement learning with consideration of fluctuation points. (1st January 2023)

• Record Type: Journal Article
• Authors: Jin, Wei; Fu, Qiming; Chen, Jianping; Wang, Yunzhe; Liu, Lanhui; Lu, You; Wu, Hongjie
• Is Part Of: Journal of building engineering. Volume 63(2023)Part A
• Publication Date: 2023-01-01
• Subjects: Energy consumption prediction -- Deep reinforcement learning -- Deep forest -- Deep deterministic policy gradient -- Fluctuation points; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2022.105458
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: Accurate building energy consumption prediction plays an irreplaceable role in building energy-saving fields. Deep Reinforcement Learning (DRL), as an innovative artificial intelligence method, has been increasingly applied to prediction problems. The existing prediction methods suffer from the fluctuation points of the energy consumption, and the prediction accuracy of the fluctuation points is often difficult to be guaranteed because the fluctuation points are often sparse and irregular. This paper proposes a novel building energy consumption prediction method using DRL with consideration of fluctuation points. Firstly, since the energy consumption is time-dependent, the timestamp features of the energy consumption data are extracted, which can potentially improve the final prediction accuracy and help solve the subsequent fluctuation point problem. Then, according to the fluctuation point types defined by the fluctuation degree, the fluctuation point features are obtained by the Deep Forest method. Finally, the prediction problem is modeled as a Markov Decision Process (MDP), where the prediction result can be considered an action selection in DRL. Additionally, by using timestamp and fluctuation point features, the state space is enriched, and meanwhile the action space is reduced, which can help the DRL agent to make the correct decision when it encounters fluctuation points. Experimental results show that the proposed method achieves higher prediction accuracy and more stable convergence than the other eight comparable methods. Moreover, compared to the representative DRL method—Deep Deterministic Policy Gradient (DDPG) method, MAE, MAPE, and RMSE are decreased by 7.15%, 12.71%, and 18.33%, respectively, and R 2 is increased by 1.3%. Highlights: The potential of Reinforcement Learning for energy consumption prediction is investigated, and the fluctuation point problem is found. The correlation analysis on available meteorological data features is conducted, and the timestamp features are extracted, which can potentially improve the final prediction accuracy. A modeling method for fluctuation points classification based on the Deep Forest is proposed by defining the types of fluctuation points, providing a high-precision classification results. An overall framework of the DF-DDPG method for single-step ahead energy consumption prediction with consideration of fluctuation points is designed. Real data sampled from an office building in Shanghai, China, is used to illustrate the prediction process and validate the proposed method.