Comparison of linear regression and artificial neural networks models to predict heating and cooling energy demand, energy consumption and CO₂ emissions

Highlights

• Comparison of linear regression and ANN models.

• Predictive methods for energy demand, consumption and CO₂ emissions.

• Multilayer perceptron accuracy with an R2 coefficient above 99%.

• Multiple variable iterations on the first stages of building design.

Abstract

An attempt has been made to develop linear regression models and Artificial Neural Networks (ANN) to predict the heating and cooling energy demands, energy consumptions and CO₂ emissions of office buildings in Chile. The calculation of dependent variables to calibrate and evaluate the models has been determined starting from the ISO 13790:2008 standard, assigning constructive characteristics to each of the geometries studied based on the Chilean standards, studying 77,000 cases. A total of 8 fundamental variables have been considered to cover the design parameters. In energy consumption and CO₂ emissions cases, the linear regression models that offer a better performance are those where the predictive variables have been transformed. Whereas, the multilayer perceptron adjusted over the variables without being transformed, provides greater accuracy in the determination of the demand, consumption and CO₂ emissions both for heating and cooling, offering ECM values closer to 0, with an R² coefficient above 99%. It is foreseen that the models developed can be used to estimate the energy saving between different design outlines during the project phases when the construction standards, systems and internal loads are defined.

Introduction

Construction is a complex and dynamic sector. The construction industry's total budget in 2013 was 8.2 billion dollars [1] and it is foreseen that it will grow to 15 trillion dollars in 2025; some 13% of the world's GDP. This is expected to grow to 15% by 2020 [2]. Apart from its economic impact at a global level, its environmental impact is also representative. According to diverse sources [3], [4], around 30–40% of the total energy consumption for human activities is attributed to the construction sector. In 2010, its global consumption was 23.7 PWh and the International Energy Agency indicates that it may reach 38.4 PWh in 2040 [5], being responsible for 38% of greenhouse gas emissions, becoming a crucial action area against climate change [6]. Due to this, energy efficiency and energy demand reduction strategies are one of the main challenges for the construction sector [7].

The need to predict energy consumption and at least contain the increase of CO₂ emissions in the construction sector has compelled official organizations, like the European Union since 2002 [8], to order that the energy efficiency of buildings had to be quantified. One of the calculation methods recommended by the Commission Delegated Regulation (EU) No. 244/2012 [9] is the one contemplated in the ISO 13790:2008 [10]. This method has been widely used in the scientific community [11] in the first stages of design for both simple or complex envelopes [12]; it has even been optimized for specific climates through the use of the factor method [13], as it is a validated tool and is relatively easy to make iterations, unlike the dynamic simulation methods [14]. When it comes to simplifying the calculation methods, there have been several investigations that relate the heating and cooling demand in regards to their energy consumption [15], [16] and CO₂ emissions [17] with regression models.

In addition, the use of artificial neural networks has spilled over to the construction sector [18] because they give higher feasibility and reliability than other traditional regression techniques. Most of the recent research is focused on energy used in buildings [19], with various applications ranging from predicting the consumption of the building per-se, studies that predict energy demand [20], energy consumption [21], both cooling and heating [22], [23], and the performance of different air-conditioning systems [24], [25].

Moreover, the ANN approach has allowed researchers to combine several parameters in the decision-making analyses [26] and the use of online forecasting tools [27], [28]. These approaches are also applied to electricity load forecasting [29], hourly energy consumption [30] and bioclimatic buildings [31]. Other studies consider the analysis of climate parameters in the built environment as well as indoor air temperature [32] and relative humidity [33], and there are also investigations related to thermal [34] and visual comfort [35] levels gathered with energy consumption.

A revision of the relevant literature points out that those studies related to ANN mainly rely on the study of a single case study or a group of case studies. L. Mba et al. tested the performance of ANN in predicting comfort parameters in a small room (6.5 m²) located in a building in Cameroon [33]. R. Kumar et al. proposes the application of ANN to study heating and cooling demands for a group of 250 buildings, with areas ranging from very small spaces (1–2 m²) to medium sized ones (100 m²) [18]. S. Karatasou et al. compared the feasibility of an ANN model against two case studies, one from a prediction tool and another from a real case study of an office building in Athens [19]. A single case-study of an administrative building located in Sao Paulo [21] set results from a simulation software and ANN off against each over. Another single case-study was used as a test model for ANN in Southern Spain, in this case for a bioclimatic building with a peculiar energy demand [31]; while, a secondary school located in Portugal was used as a test model because of its suitability for predictive neural networks [32]. A similar case-study was used as a test model for ANN for a medium sized tertiary building in Italy [23] The energy demand of three institutional buildings were assessed over 2 years and then compared with the results of an ANN model [22]. A similar study proved the suitability of using Machine learning methodologies to accurately predict the energy demand of three educational buildings in Spain [29].

With regard to tertiary buildings, X. Li et al. provided evidence of the adaptability of these regression techniques to predict energy demands, testing the model against two commercial premises, one small and one medium-sized property [27]. Energy demand for an entire rooftop air conditioning system of a single large-scale commercial building could be also accurately predicted using ANN [24], and similar results were provided by M. Kljajic et al. [25] for a larger sample size: 65 boilers located randomly across 50 buildings in Serbia. Another case study, with regard to sample size, tested the feasibility of meta-learning based systems against 48 test buildings and 1 real building [26]. Studies have been made not only for energy demand, but also for daylighting prediction using ANN, such as S.I. Wong et al. [35], where a 35 × 35 m simulation of a 40 storey building was carried out as a single case study using Energy Plus software to then compare the results against ones given by the ANN model.

Other authors focus on predicting the energy demand on a large scale (nationwide study for the USA) [20] focusing on the evolution of socioeconomic parameters or using existing data obtained from past records to predict the short-term energy consumption [30]. Additionally, other earlier studies provide the necessary basis on how to adapt the architecture of an ANN to the requirements of a study of this kind [28] and how to improve the algorithms to better predict comfort parameters and energy consumption [34].

Other research projects are more similar to the approach of this one, such as F. Khayatian et al. [36], who uses data generated from a simulation software for energy certification in Italy as the training set in order to model an ANN that is capable of predicting the expected outcomes. Dall'O et al. use a broad database of around 175.000 elements to provide inputs for predicting outcomes for the energy certification of buildings in a designated area [37]. Alternatively an ample database can be used to establish benchmarking methods for the energy performance assessment of buildings [38].

Those existing studies provide a reliable scientific corpus that widely demonstrates that ANN methodology is applicable for predicting several variables in relation to building's energy demand and energy consumption. In this way, the authors have made some contributions on this matter in the development of prediction models related to energy demand in buildings [17] and found out that there is room for further development in this field.

This research intends to propose an advance in the use of ANN to predict energy demand, energy consumption and CO2 emissions, taking tertiary buildings located in Chile as a target, considering that this approach is compelling for the following reasons. At first, Chile is a country that is experimenting a profound transformation in the construction industry, as a consequence of being the first South American country to join the OECD [39]. The legislative framework for energy efficiency has strengthened remarkably over the last few years, changing from an old 2007 standard [40] to some legislative texts that, despite not yet being mandatory, provide useful guidelines to reduce CO2 emissions and contain energy demand both for public [41] and residential buildings [42]. Additionally, the Chilean government has promoted a multilateral agreement with the objective of implementing the BIM (Building Information Modelling) standard in all public projects in order to improve efficiency and productivity [43]. This transformation is being done in cooperation with countries such as Spain or the UK, who provide technology and know-how about how to implement energy efficiency policies, in particular with regard to the European ISO calculation standard and energy-rating systems. Given this context, the application of regression techniques, such as ANN, are useful in helping this transformation into an energy-efficient and reduced-CO₂ policy.

In terms of the scale of the study, it has focused on tertiary office buildings, because of this sector's impact on construction activities in Chile [44], comprising facilities of all kinds and types in relation to the built area, location, shape or conditioning systems, amongst other variables. According to data from the Chilean Bureau of Statistics, office buildings are a growing sector, totalling 8.9 million square meters, approximately 12.75% of the total built area since the 2012–2015 period [45]. Considering just the public policies and the Chilean Integrated Repository of Projects [46], more than 50% of the projects currently developed by the Chilean Government are offices, and their average area is approximately 1500 m², with most of them located in the country's capital, Santiago. That is why, in order to assist this transition, this study attempts to make a relevant contribution in forecasting the three variables that are considered crucial for energy policies in the building industry: Energy demand, energy consumption and CO₂ emissions. As pointed out before, there is not yet a standardized quasi-dynamic calculation procedure in Chile, as in European countries, that guides the aforementioned calculations, and for this reason it has been considered a novel approach to implement the ISO 13790:2008 standard in the Chilean context.

Given this context, this research intends on clarifying how energy demand, energy consumption and CO2 emissions of office buildings can be predicted in the event of the enforcement of ISO 13790:2008 standard. Two points have to be remarked, as they pose a novel approach in this field. First, the variegated climatic context of Chile, ranging from dry desert to cold polar climates [47]. Second, the scale of the study, giving as a result 154.000 possible cases with an estimated built surface of 1.1 billion m², which would equal approximately 130 times the built surface from 2012 to 2015, comfortably covering the size of the study. In this way, the objective is to demonstrate the performance and reliability of ANN in predicting large scale data not only for a single parameter, but for three of them (energy consumption, energy demand and CO2 emissions) in relation to a large-scale sample of buildings, with all the issues associated to them, such as the nonlinearity of problems related to building design and performance.

It is also necessary to indicate that the building's CO₂ emissions do not just depend on the building's intrinsic parameters, as its emissions are also associated to the type of energy that this uses to cover its different demands. Thus, following the Standard International Energy Classification (SIEC) in the International Recommendation on Energy Statistics (IRES) [48], we distinguish between primary and secondary energy; primary energy includes those materials that are directly burned in the building thus producing CO₂, such as petrol, coal and natural gas. Secondary energy includes those sources of energy, in this case electric power, that are generated elsewhere and consumed in the building. A building can consume both types of energy, so with the idea of unifying all of these into a common unit, each type of energy has a CO₂eq emission factor associated to quantify the emissions that their combustion or production generate. In the case of primary energy, this information is obtained directly from the type of fuel that is used (Table 1).

Chile is supplied by four major electrical systems, which are all independent from each other. In the case of Santiago, this city is supplied by the Central Interconnected System (SIC by its Spanish acronym), which provides 71.03% of the total capacity nationwide. Each system is comprised by a different type of generation, which is called the energy mix, combining petrol and diesel power plants, hydroelectric plants, run-of-river plants, gas combined cycle plants, etc. In turn, each generation has an equivalent CO₂ emission factor (Table 1) associated to it. For the case of the SIC system, an equivalent in CO₂ emission factor for each year can be established by combined all of these (Table 2). Logically, this figure varies every year depending on the different percentages that the energy mix has.

The following points will be cleared up as this investigation advances:

• Generating a procedure to determine the data set based the ISO 13790:2008 method to generate multiple iterations in office buildings under the TDRe Standard.

• Clarifying the parameters to be considered in the first design stages for the construction of office buildings in Santiago (Predictor variables).

• Generating, through diverse statistical models, simplified methods for the prediction of heating and cooling demands, energy consumptions and CO₂ emissions introducing variables associated to the design.

It is expected that the results of the study will be of help in two main areas. First, assisting designers in the early stages of design, being able to grasp the demand of resources of their design with a few variables. Second, the reference values expected from the ANN results will be of help in assisting future policy makers in establishing realistic goals in order to reduce, or at least contain, the energy demand, energy consumption and CO2 emissions for these buildings.