空调控制参数用户设定偏好的预测方法
Prediction method of air conditioner control parameters by user’s preferences
摘要：
空调器智能化提升的方向,是使平时需要用户设定的控制参数值自动地调整为用户的偏好值。这就需要空调器能够准确预测出用户偏好的设定参数,在实现上则有赖于对用户历史操作数据的学习,但这又超出了空调器单片机的计算能力。本文提出云端学习及本地计算相结合的空调用户偏好的预测方法,将学习用户历史操作数据等复杂的计算任务放到云端服务器,使得对于控制参数的预测不会超过空调器单片机的计算能力。该方法中,云端服务器完成对用户历史数据的预处理和递归式特征消除后,使用梯度提升框架训练数据并得到学习模型;空调器单片机下载由云端生成的多维矩阵,建立对预测值的插值查询规则后即得到本地预测方法。本文在对上述建立的学习算法和预测方法进行验证时,将上海、重庆和广州3个不同地域的城市的用户夏季使用数据(不含8月)作为训练集,8月的数据作为测试集。验证结果表明:用户实际设定温度与预测值的误差在±0.5 ℃内的占比平均为84%,最高为88%;用户实际设定风速与预测值的误差在±10%内的占比平均为92%,最高为94%。验证结果表明本文提出的云端学习及本地计算相结合的用户偏好的预测方法能够准确地预测用户偏好。
Abstract：
The direction of intelligent optimization of the air conditioner is to automatically adjust the control parameter that usually needs to be set by the users to their preferred value. This requires the air conditioner to be able to accurately predict the setting parameters based on users’ preferences. This implementation depends on the learning of the users’ historical operation data, which is beyond the computing power of the microcomputer in the air conditioner. This paper proposes a prediction method for users’ preferences by combining cloud learning and local computing. This method puts the complex computing tasks such as learning users’ historical data on the cloud server, so that the prediction of control parameters will not exceed the computing power of the microcomputer in the air conditioner. In this method, after the cloud server completes the preprocessing of historical data and the recursive feature elimination, the gradient boosting framework will be used to train the data and obtain the learning model. The microcomputer downloads the multi-dimensional matrix generated by the cloud, establishes the interpolation query rule, and obtains the local prediction method. When validating the learning algorithm and prediction method established above, summer data (excluding August) in three different regions including Shanghai, Chongqing, and Guangzhou is chosen as the training set, and August data is chosen as the test set. The validation result shows that the error between the actual set temperature and the predicted value within ±0.5 ℃ accounts for an average of 84% and a maximum of 88%. The error between the user’s actual set wind speed and the predicted value within ±10% accounts for an average of 92% and a maximum of 94%. Therefore, the prediction method combined with cloud learning and local computing proposed in this paper can accurately predict users’ preferences.
Keywords:room air conditioner; control parameter; user preference; data mining; machine learning; preference prediction