基于青岛某办公建筑2015年全年逐时总用电能耗及空调用电能耗数据，利用kmeans聚类算法对其进行聚类，将全年能耗水平分为四大类。利用求平均值法得到每一类典型设备使用率曲线。将典型曲线的数据、日前两周数据以及气象数据一同作为BP神经网络的输入，预测未来24小时的建筑总用电和空调用电，该方法比单用日前两周数据及气象数据进行负荷预测能获得更低的相对误差、均方根误差、平均绝对百分误差。 BP负荷预测相对误差在5%以内，而kmeans-BP负荷预测算法控制在±2.5%以内；BP预测得到的均方根误差和平均绝对百分误差范围分别在4.6~9.0之间、2.3%~4.4%之间，kmeans-BP将该误差缩小到3.1、2.0%以内，对于负荷预测精度要求上是阶跃性的突破。

Based on the data of hourly total energy consumption and energy consumption for air conditioning of an office building in Qingdao in 2015, kmeans clustering analysis was used to cluster all of the data. As a result, they were devided into four categories according to annual energy consumption level.The average value method was utilized to obtain each kind of typical usage curves.Typical curve data, two weeks’data before the forecasting day and meteorological data were used as input of BP neural network, and then the power load of next 24 hours was predicted. This method was better than that without considering typical curve data, as shown by the lower relative error, root mean square error and mean absolute percent error. The relative error of the BP load forecasting is less than 5%, while the kmeansBP load forecasting algorithm controls it within ±2.5%. The root mean square error and absolute absolute error range predicted by BP are between 4.6-9.0 and 2.3%-4.4%, respectively. KmeansBP reduces the errors to 3.1 and 2%, respectively, which is a breakthrough for the accuracy requirement of load forecasting.

TK01+8

国家重点研发计划项目“可再生能源绿色建筑领域应用效果研究”（2016YFC0700104）