Operational optimisation of domestic refrigerators based on user behaviour prediction and deep reinforcement learning.

The refrigerator holds a significant place in domestic energy usage as a 24-hour operating appliance. Thus, one of the current research hotspots is the development of advanced energy-saving control systems. This study proposes a control technique based on user behaviour prediction and deep reinforcement learning. It addresses the issue
that existing control methods struggle to identify the optimal settings among components and fail to account for the influence of user behaviour on energy consumption. First, Transformer coding in conjunction with Kolmogorov-Arnold Networks (KAN) is used to build a prediction system for user door-opening and closing behaviour. This method outperforms the baseline KAN by more than 85 %. Based on this, an enhanced Pre-LSTMSAC (PreLSAC) method is proposed to combine the Long Short-Term Memory (LSTM) networks and prediction information, thereby enhancing the agent’s control capacity during refrigerator operation. According to experimental results, PreLSAC has a considerable advantage over conventional control methods in terms of temperature control and energy consumption reduction. It also shows good generalisation and robustness under various thermal loads and random door-opening and closing events, indicating its potential for use in real-world scenarios.