摘要
准确的建筑能耗预测对建筑能源系统的智能高效运行起着至关重要的作用。为了更好挖掘海量数据中蕴含的有效信息,提高短期能耗预测精度,针对能耗数据时序性和非线性的特点,本研究基于湖南省某办公楼4年的能耗数据,提出了一种基于时间序列的集成模型用于电力负荷预测。该方法将季节性自回归综合移动平均(SARIMA)模型与长短期记忆网络(LSTM)模型相结合,综合考虑了能耗数据的时序性和非线性特点,利用最小二乘法动态计算每个模型的权重因子。此外,本研究将该方法与其他预测模型在预测精度和计算效率两方面进行对比,结果表明,该集成模型的均方根误差(RMSE)为12.065,拟合优度(R^(2))为97.45%,相比最优的单一模型(LSTM)预测误差降低了20.3%,显著提高了能耗预测的准确率。同时,本研究分别统计了模型在供暖季和供冷季的RMSE分布并进行预测精度对比,结果表明采用该混合模型在供暖季具有更优的预测性能。另外,在特征数据构建过程中,本研究采用混合比例为0.5的改进Boruta算法进行特征选择,不仅降低了阴影特征样本的复杂度,同时更高效地提取出与建筑能耗相关的特征集合。
An accurate prediction of building energy consumption plays an important role in the intelligent and efficient operation of building energy management system.In order to better mine useful information from massive data and improve the accuracy of short⁃term electrical load prediction,in response to the timing and non⁃linear features of energy consumption data,an integrated model for electrical load forecasting via time series method was proposed based on the four⁃year measured energy consumption data of an office building in Hunan Province.The method proposed combines the Seasonal Autoregressive Integrated Moving Average(SARIMA)model and the Long Short⁃Term Memory(LSTM)network,which considers the timing and nonlinear characteristics of electrical load data comprehensively.The least square method was used to calculate the weight factor of each model dynamically.In addition,the model was compared with other frequently⁃used machine learning algorithms in terms of accuracy and computational efficiency.The result shows that the root mean square error(RMSE)of the integrated model is 12.065 and the goodness of fit(R^(2))is 97.45%.Through contrastive analysis,the model proposed significantly improves the accuracy of electrical load prediction,the error of which is decreased by 20.3%compared with the optimal single model.This paper also calculated the RMSE distribution of the model in heating and cooling season and compared the prediction accuracy,the results show that the integrated model has better performance in heating season.In the process of feature data construction,this study used an improved Boruta algorithm with a mixing ratio of 0.5 for feature selection,which not only reduces the complexity of shadow feature samples,but also extracts the feature set related to building energy consumption more efficiently.
作者
李毅
彭晋卿
廖维
邹斌
曹静宇
LI Yi;PENG Jinqing;LIAO Wei;ZOU Bin;CAO Jingyu(Hunan University,Changsha 410082,China)
出处
《建筑科学》
CSCD
北大核心
2022年第10期190-197,共8页
Building Science
基金
湖南省科技创新计划资助(2020RC5003)
中国博士后科学基金面上资助(2020M682559)
住房和城乡建设部2020年科学技术项目计划(2020⁃K⁃165)。
关键词
集成算法
数据采集
特征提取
最小二乘法
integrated model
data acquisition
feature extraction
least square method
