Elsevier

Energy and Buildings

Volume 174, 1 September 2018, Pages 323-334
Energy and Buildings

A hybrid teaching-learning artificial neural network for building electrical energy consumption prediction

https://doi.org/10.1016/j.enbuild.2018.06.017Get rights and content

Abstract

Numerous data-driven models have been successfully adopted for electrical energy consumption forecasting at building and larger scales. When the data set for forecasting is multi-sourced, heterogeneous or inadequate, single data-driven model may lead to convergence problem or poor model accuracy. The combination of advanced evolutionary algorithms (EAs) and data-driven models is proved effective in terms of prediction accuracy and robustness improvements. However, some of them are very time consuming to converge. In this paper, a novel EA, i.e. teaching learning based optimization (TLBO), is proposed for short-term building energy usage prediction. To enhance its convergence speed and optimization accuracy, the basic TLBO algorithm is further modified in three aspects. The improved algorithm is combined with artificial neural networks (ANNs) and applied to hourly electrical energy prediction of two educational buildings located in USA and China respectively. Performance comparisons show that the proposed model has superior performances than previously reported GA-ANN and PSO-ANN methods in terms of convergence speed and predictive accuracy, and is suitable for online energy prediction in the future.

Introduction

Rapid economic growth, accompanied by structural changes, strongly affects the global trend of energy consumption [1]. For building sector, energy use prediction contributes to effective building energy management, energy systems commissioning through detecting system faults, building energy operation and control, etc. Many computer softwares have been developed for energy efficiency design of new buildings, e.g. EnergyPlus, eQUEST, BLAST, DeST, etc. But for existing buildings, it is difficult to estimate future energy usage because a number of multi-sourced factors influence the building energy behavior, e.g. weather conditions, building materials’ thermal properties, occupancy schedule, not to speak of the complex interactions of HVAC and lighting systems.

In the past two decades, numerous data-driven methods were introduced to the area of building energy prediction. These techniques modeled building energy usage patterns based on previously recorded time series data, such as past energy usage data, weather conditions, occupancy schedule, etc. Recent review studies had offered detailed classification of the existing predictive models and their characteristics. Zhao and Magouls [2] provided a review on major prediction methods of building energy consumption and classified them into five categories, i.e. engineering approaches, statistical models, artificial neural networks (ANNs), support vector machines (SVMs), and grey models (GMs). Wang and Srinivasan [3] reviewed artificial intelligence (AI) based building energy forecasting. Two categories of AI based prediction methods, i.e. the single prediction methods and the ensemble prediction methods, were discussed and compared. Deb et al. [4] provided a review on time series forecasting techniques for building energy consumption. Nine major time series techniques (ARIMA, ANN, SVM, Hybrid, etc.) with respect to building energy usage were compared and analyzed. Daut et al. [5] reviewed conventional and AI methods for building electrical energy consumption prediction. For the purpose of accuracy improvement, the swarm intelligence methods were surveyed to be hybridized with single data-driven models.

Above reviews provided very beneficial information on numerous forecasting models at different space and time scales. Hybrid AI models for building energy forecasting are popularly used, and recent case studies from 2010 to 2017 are briefly reviewed in the next section. It is noted that most research works focused on prediction accuracy improvement, but the modeling time issue, which is important for online application, is seldom addressed. This study emphasizes on the development of a new evolutionary algorithm (EA) based predictive models for short-term building electrical energy prediction. The novel EA, i.e. teaching learning based optimization (TLBO) is proposed to improve the modeling performances of regular artificial neural networks (ANNs). To enhance its capabilities of convergence speed and precision, the basic TLBO is further modified using three different measures. By integrating the characteristics of improved TLBO (iTLBO) and ANN, the iTLBO-ANN hybrid predictive models are applied to two educational buildings for hourly electrical energy consumption forecasting. Performance comparisons with previous AI models are investigated. Results validate the effectiveness and efficiency of the proposed method.

The study is organized as the following: Section 2 provides a short review of current research trends of energy use prediction. Section 3 presents the theories, structure and characteristics of the iTLBO-ANN model. Section 4 gives the data description of two applications for model validation. Section 5 provides the predictive applications’ results and discusses the capabilities of the hybrid model compared with other EAs based methods. Finally, conclusions are provided in Section 6.

Section snippets

Short review of building energy usage prediction from 2010 to 2017

During the past decade, there were numerous data-driven models for building energy usage prediction, which included regression model, artificial neural networks, support vector regression, fuzzy model, grey model, etc. For example, Yun et al. [6] used an indexed fourth order auto regressive model for one hour ahead building heat load forecasting. Korolija et al. [7] developed bivariate and multivariate regression models for predicting long term building’s heating, cooling and auxiliary energy

Data-driven model using artificial neural network

ANNs, analogous to the biological neurons of the human brain, are composed of a number of simple and highly interconnected processors [25]. Weighted links connect each processor and pass signals from one unit to another. For instance, a three-layer network’s mapping function is formulated below: Y=f(b0+j=1kh(ψj+i=1mpiwij)bj)where Y is the output of the network, denoted by nonlinear transfer function f( · ); b0 is the output bias; bj represents the weights of links from hidden layer to output

Data sets: Input variables and data pre-processing

For data-driven models, the selection of input variables is very important for precise building energy forecasting. In general, input variable types usually fall into five categories, i.e., the meteorological data, calendar data, occupancy data, historical energy data, and all others. In this section, different data sets from two buildings are collected to verify the forecasting ability of the proposed iTLBO-ANN model. Both data sets are derived from actual building energy usage data, which are

Data set A

From PCA results, it is shown that the outdoor air dry bulb temperature and solar radiation intensity have a great influence on electricity energy usage in this case study. While the relative humidity and wind speed are less correlated with the energy consumption. Considering the importance of building occupancy for energy usage, the weekends (or holidays) and normal working days are labeled with ‘0’ and ‘1’ respectively. The daily hour of the sine and cosine values are calculated to make it in

Conclusions and future work

To improve the prediction accuracy of the regular ANN model for short-term building energy forecasting, a new hybrid method called TLBO-ANN was utilized to optimize ANN’s parameters in the global scope. To accelerate the convergence speed and enhance the optimization accuracy, the basic TLBO was further improved using three measures, i.e., adding a feedback stage, adding an accuracy factor and eliminating the worst solution. A plenty of benchmark functions were carried out for performance

Acknowledgments

This work was supported by National Natural Science Foundation of China (grant nos. 61304075 and 51705206), China Postdoctoral Science Foundation (grant nos. 2018T110457 and 2016M601741), Jiangsu Provincial Natural Science Foundation of China (grant no. BK20150525) and Project Foundation for Priority Academic Program Development of Jiangsu Higher Education Institutions.

References (37)

Cited by (0)

View full text