基于机器学习的建筑能耗模型适用性研究

摘要

This paper focuses on a comparison of predicting accuracy of six different machine learning approaches for estimating energy use in office buildings,including linear regression,GP(Gaussian process),MARS(multivariate adaptive regression splines),bagging MARS,RF(random forest) and SVM(support vector machine).The results indicate that three methods(Bagging MARS,MARS,and RF)have better accuracy in predicting heating energy, whereas the bagging MARS performs best in estimating cooling energy.It is also found out that the prediction of cool-ing energy is more difficult than that of heating energy in office buildings.These conclusions can be used to provide some reference for machine learning method choosing in building energy assessment.Moreover,the models obtained from this research can also be used to create a building stock model at urban scales.%为进一步分析不同机器学习方法用于建筑能耗模型的适用性，重点比较了6种常用机器学习方法用于预测办公建筑能耗时的准确性，包括线性回归、高斯过程、多元自适应回归样条法、自助多元自适应回归样条法、随机森林和支持向量机。结果表明：多元自适应回归样条法、自助多元自适应回归样条法和随机森林法适用于取暖能耗的模型建立；对于制冷能耗预测，自助多元自适应回归样条法的计算精度最高。同时发现制冷能耗与取暖能耗相比，由于存在更加复杂的非线性关系，其预测难度更大。研究结果不仅可用于在建筑节能分析中确定最佳机器学习方法，而且所得机器学习方法可用于城市建筑能耗模型的建立。