Commercial Building Load Forecasts with Artificial Neural Network
Commercial Building Load Forecasts with Artificial Neural Network
复制标题
DOI：
10.1109/isgt.2019.8791654
复制
发表时间：
2019-02
期刊：
影响因子：
--
通讯作者：
Zejia Jing;Mengmeng Cai;M. Pipattanasomporn;S. Rahman;Raghavan Kothandaraman;A. Malekpour;E. Paaso;S. Bahramirad
中科院分区：
文献类型：
--
作者：
Zejia Jing;Mengmeng Cai;M. Pipattanasomporn;S. Rahman;Raghavan Kothandaraman;A. Malekpour;E. Paaso;S. Bahramirad
本文提出了一种基于人工神经网络（ANN）的建筑级小时电力负荷预测方法，该方法除考虑建筑历史负荷和室外天气数据外，还将暖通空调设定值作为输入参数之一。本文所提供的数据仅涉及冷负荷。利用人工神经网络对数据集进行训练和测试，基于人工神经网络的负荷预测模型提供一天中每小时的预测负荷。研究了Levenberg-Marquardt、缩放共轭梯度反向传播和贝叶斯正则化（BR）三种训练算法。研究结果表明，基于br的神经网络在预测精度方面表现最好。此外，以伊利诺斯州芝加哥市的一座商业建筑为例，对所开发的基于人工神经网络的模型的性能进行了比较。1小时前负荷预测的预测误差在5%左右，12小时前负荷预测的预测误差在8%左右。
This paper presents an Artificial Neural Network (ANN)-based building-level hourly electrical load forecasting method that takes into account HVAC set points as one of the input parameters, in addition to the historical building load and outdoor weather data. The data presented in this paper deal with cooling load only. ANN is used to train and test the dataset, and the ANN-based load forecasting model provides the predicted load for each hour of the day. Three training algorithms are explored, including Levenberg-Marquardt, Scaled Conjugate gradient back-propagation and Bayesian Regularization (BR). Findings indicate that the BR-based neural network offers the best performance in terms of forecasting accuracy. In addition, a case study using a commercial building in Chicago, Illinois is presented where performances of the developed ANN-based models are compared. The forecasting error is around 5% or less for hour-ahead load forecasting, and around 8% or less for 12-hour ahead load forecasting.