基于深度学习考虑人员行为的空调负荷预测
Air conditioning load prediction based on deep learning considering human behavior
摘要：
本文以提高空调负荷预测精度为目的,首先通过k均值聚类,得到人员进出房间行为模式,以确定室内人数;其次基于马尔可夫移动模型描述室内各房间人员移动,以确定室内房间有无人员;然后通过LSTM预测室内参数,并采用PMV-PPD方法对室内热环境进行评价,以确定房间内人员开关空调行为;最后以室外参数、室内参数及空调开关状态参数为输入特征使用LSTM、GRU及BiLSTM预测空调负荷。案例验证表明:预测中引入空调开关状态参数可以更好地挖掘空调开关的特征与规律,引入室内参数可以抑制由空调负荷变化剧烈引起的预测精度差的问题,同时引入2种参数能够显著降低预测误差;相比于LSTM和GRU,BiLSTM具有更高的预测精度。
Abstract：
In this paper, with the purpose of improving the accuracy of air conditioning load prediction, firstly, the entry and exit behavior patterns of personnel are obtained through k-means clustering to determine the number of people in the room. Secondly, the movement of people in each indoor room is described based on the Markov movement model to determine the presence or absence of people in indoor rooms. Then, the indoor parameters are predicted by LSTM and the indoor thermal environment is evaluated by the PMV-PPD method to determine the behavior of turning on and off the air conditioner by the personnel in the rooms. Finally, the outdoor parameters, indoor parameters and air conditioning switch status parameters are used as input features to predict the air conditioning load using LSTM, GRU and BiLSTM. The case verification shows that the introduction of air conditioning switch status parameters in the prediction can better explore the characteristics and rules of air conditioning switch, the introduction of indoor parameters can inhibit the poor prediction accuracy caused by the drastic changes of air conditioning loads, and the introduction of the two types of parameters at the same time can significantly reduce the prediction error. Compared with LSTM and GRU, BiLSTM has higher prediction accuracy.
Keywords:air conditioning load; prediction; human behavior; clustering; deep learning; environmental assessment