数据分布特性对空调系统能耗预测的影响

空调系统能耗预测是实现智能调控、能源需求管理、系统节能的重要手段和前提之一,当前的空调系统能耗预测主要是基于机器学习算法。诸多机器学习算法的重要理论前提是数据的分布应尽量满足正态分布,然而空调系统的实际运行数据很少能满足正态分布特性,目前的研究鲜有涉及数据分布特性对空调系统能耗预测的影响。首先基于实际项目的空调系统能耗数据,从偏度和峰度两个指标分析了实际能耗数据分布与正态分布呈现出的偏离;然后通过对数变换对能耗数据进行数据变换,使能耗数据更接近于正态分布;接着以常见的4种能耗预测机器学习算法(广义线性回归算法、支持向量回归算法、人工神经网络算法、随机森林算法)对原始数据和经过数据变换后的数据分别进行空调系统能耗预测工作,分析负荷预测结果的RMSE和R~2统计量。结果对比发现,数据的分布特性对能耗预测有着重要的影响,合适的数据变换可以有效地提高空调系统能耗预测机器学习算法模型的预测效果。

Study on the Impact of Data Distribution Characteristics on Energy Consumption Forecasting of Air Conditioning System

Energy consumption prediction of air conditioning system is one of the important means and premises to realize intelligent control, energy demand management and system energy saving. The current energy consumption prediction of air conditioning system is mainly based on machine learning algorithm. The important theoretical premise of many machine learning algorithms is that the data distribution should meet the normal distribution as much as possible, but the actual operation data of air conditioning system rarely meet the normal distribution characteristics, and the current research rarely involves the influence of data distribution characteristics on the energy consumption prediction of air conditioning system. Based on the energy consumption data of the air conditioning system of the actual project, the deviation between the actual energy consumption data distribution and the normal distribution is analyzed from two indicators of skewness and kurtosis. Then, the energy consumption data are transformed by logarithmic transformation, which makes the energy consumption data closer to the normal distribution. Finally four common energy consumption prediction machine learning algorithms (generalized linear regression algorithm, support vector regression algorithm, artificial neural network algorithm, random forest algorithm) are used to predict the energy consumption for the original data and the data after data transformation respectively, and RMSE and R2 are analyzed. The results show that the distribution characteristics of data have an important impact on energy consumption prediction, and appropriate data transformation can effectively improve the prediction effect of machine learning algorithm model for energy consumption prediction air conditioning system.