Search

• Deep reinforcement learning (DRL) based models are proposed for HVAC system energy consumption prediction. • Autoencoder (AE) algorithm is used to extract the high-level features of state space of DDPG. • The DDPG based models are applied for energy consumption forecasting over short-term time horizon. • Improvement in prediction performance is observed when combine AE with DDPG. • The DDPG based prediction models, outperform common supervised models like BP neural network and support vector machine. Short-term energy consumption prediction has fundamental importance in many HVAC system management tasks, such as demand-side management, short-term maintenance, etc. Currently, the prevailing data-driven techniques, especially supervised machine learning methods, are widely applied for short-term energy consumption prediction. Deep reinforcement learning (DRL), as the state-of-the-art machine learning techniques, have been applied for HVAC system control, but rarely for energy consumption prediction. In this paper, a DRL algorithm, namely Deep Deterministic Policy Gradient (DDPG), is firstly introduced for short-term HVAC system energy consumption prediction. Moreover, Autoencoder (AE), which is powerful in processing data in their raw form, is incorporated into DDPG method to extract the high-level features of state space and optimize the prediction model. The operation data of the ground source heat pump (GSHP) system of an office building in Henan province, China is used to train and assess the proposed models. The results demonstrate that the proposed DDPG based models can achieve better prediction performance than common supervised models like BP Neural Network and Support Vector Machine. This study is an enlightening work which may inspire other researchers to tap the potential of DRL algorithms in this field. [ABSTRACT FROM AUTHOR]

• The potentials of deep reinforcement learning (DRL) techniques are investigated. • Three commonly-used DRL techniques, i.e. A3C, DDPG and RDPG are utilized for building energy consumption forecasting. • The prediction performances of six popular predictive algorithms are studied. • DDPG and RDPG can evidently enhance prediction performance while accounting for more computation time. • A3C doesn't show advantage in forecasting building energy consumption. Reliable and accurate building energy consumption prediction is becoming increasingly pivotal in building energy management. Currently, data-driven approach has shown promising performances and gained lots of research attention due to its efficiency and flexibility. As a combination of reinforcement learning and deep learning, deep reinforcement learning (DRL) techniques are expected to solve nonlinear and complex issues. However, very little is known about DRL techniques in forecasting building energy consumption. Therefore, this paper presents a case study of an office building using three commonly-used DRL techniques to forecast building energy consumption, namely Asynchronous Advantage Actor-Critic (A3C), Deep Deterministic Policy Gradient (DDPG) and Recurrent Deterministic Policy Gradient (RDPG). The objective is to investigate the potential of DRL techniques in building energy consumption prediction field. A comprehensive comparison between DRL models and common supervised models is also provided. The results demonstrate that the proposed DDPG and RDPG models have obvious advantages in forecasting building energy consumption compared to common supervised models, while accounting for more computation time for model training. Their prediction performances measured by mean absolute error (MAE) can be improved by 16%-24% for single-step ahead prediction, and 19%-32% for multi-step ahead prediction. The results also indicate that A3C performs poor prediction accuracy and shows much slower convergence speed than DDPG and RDPG. However, A3C is still the most efficient technique among these three DRL methods. The findings are enlightening and the proposed DRL methodologies can be positively extended to other prediction problems, e.g., wind speed prediction and electricity load prediction. [ABSTRACT FROM AUTHOR]