Research on air-conditioner energy saving control method based on Q-learning and adaptive constraint network

In order to solve the problem of air-conditioner’s high energy consumption during operation under traditional PID control, this paper proposes an air-conditioner energy-saving control method based on Q-learning and adaptive constraint network. Firstly, standard reward matrix is constructed according to the “expert system” and “reward function”, and the air-conditioner operating parameters corresponding to matrix elements are divided into A and B datasets. Secondly, initialising radial basis function (RBF) neural network model and training constrain network with A datasets. Besides, using B datasets doesn’t verify model performance until accuracy rate is over 90% and this model is regarded as Standard Comfort Energy-saving Constraint Network (SCECN). Finally, air conditioner operation strategy is formulated by combining Q-learning algorithm of reinforcement learning with actual reward matrix that is constructed according to SCECN. If this strategy dissatisfies needs of user, SCECN will be updated by actual feedback data and defined as Self-Adaptive Comfort and Energy Constraint Network (SACECN). Experiments show: comparing with traditional control logic of air-conditioner, the algorithm saves up to 18.3% electric energy when air-conditioner operates.