The extension of in-home time has boosted people's increase demand for comfortable indoor temperatures. This paper aims to propose a smart home-based temperature control framework based on deep learning and reinforcement learning to automatically control indoor temperature. The heat pump power and the ventilation system volume could be adjusted by the proposed framework to minimize costs while maintaining indoor comfort. The uncertainty of future information and the requirement for continuous control make it challenging to build an intelligent temperature control system. Simultaneously, the human body is sensitive to different temperatures. Therefore, the selection of appropriate evaluation criteria for people’s thermal comfort is essential. To address this problem, the Predicted Mean Vote (PMV) criterion is applied to scientifically evaluate human thermal comfort. The indoor temperature control problem is transformed into a Markov decision process. A deep learning framework (Proximal Policy Optimization PPO) based on continuous control actions is proposed to solve this problem. Combined with the deep learning method Long Short-Term Memory (LSTM), future state information is applied to improve the control stability. The proposed framework performs well under different conditions in summer and winter. The proposed framework performs well under different conditions in summer and winter. It is compared by the discrete action control method Deep Q Network and the PPO algorithm, which does not contain future state information. As a result, the proposed framework achieves 24.29% and 23.63% cost savings in winter and summer, respectively.

居家时间的延长，推动了人们对舒适室内温度的需求增加。本文旨在提出一种基于深度学习和强化学习的智能家居温度控制框架来自动控制室内温度。热泵功率和通风系统体积可以通过拟议的框架进行调整，以最大限度地降低成本，同时保持室内舒适度。未来信息的不确定性和对连续控制的要求使得构建智能温度控制系统具有挑战性。同时，人体对不同的温度也很敏感。因此，为人们的热舒适度选择合适的评价标准至关重要。为了解决这个问题，预测平均投票 (PMV) 标准用于科学评估人体热舒适度。将室内温度控制问题转化为马尔可夫决策过程。提出了一种基于连续控制动作的深度学习框架（Proximal Policy Optimization PPO）来解决这个问题。结合深度学习方法长短期记忆（LSTM），应用未来状态信息来提高控制稳定性。所提出的框架在夏季和冬季的不同条件下表现良好。所提出的框架在夏季和冬季的不同条件下表现良好。通过离散动作控制方法Deep Q Network和不包含未来状态信息的PPO算法进行比较。结果，提议的框架达到了 24.29% 和 23。