1.1. Rationale
Buildings represent a large portion of a country’s energy consumption and associated greenhouse gas emissions. For example, Canadian residential and commercial/institutional sectors consumed approximately 27.3% of the country’s total secondary energy usage in 2013. Amongst both sectors, the energy needed for space heating, space cooling and hot water needs accounted for 21% of the overall total secondary energy usage [
1]. Consequently, the energy needed in order to maintain internal conditions within buildings accounts for a significant portion of the overall energy usage and greenhouse emissions. Therefore, increasing energy efficiency and utilization in buildings is of great importance to our overall sustainability.
Over the past few decades, researchers have dedicated themselves to improving building energy efficiency and usage through various techniques and strategies. The forecasting of energy use in an existing building is essential for a variety of applications such as demand response, fault detection and diagnosis, model predictive control, optimization, and energy management.
Energy estimation models are a growing area of research, this is especially true with new advancements in artificial intelligence and machine learning. Such models have widely been applied to both energy systems, buildings, and HVAC (heating, ventilation, and air conditioning) systems alike as they can help with a variety of tasks. ASHRAE breaks energy estimation models into two main categories: physics-based/forward models, and data-driven/inverse models [
2].
Physics-based models also called white-box models/forward models, are based on physical laws. Such models require a large number of inputs about the building and HVAC system from existing buildings, which are unknown at the current and future times, e.g., (t + 1, t + 2…), where (t) is the current time, for the purpose of forecasting. These models are used in building energy simulation software such as EnergyPlus, eQuest, and TRNSYS. Such models are more useful at the design stage of a building rather than real-time forecasting for existing buildings. This is a result of the time and budgetary constraints needed in the development and calibration of physics models when trying to model existing buildings. For such cases, often too many parameters are needed, access to them all is not feasible, and the overall construction of such models may require tedious amounts of work in calibration.
Data-driven models, in contrast, are based on a strictly mathematical model and measurements. Consequently, they do not require such detailed knowledge of the building or equipment. Their forecasts are mostly based on historical data which are more readily available from control systems implemented within the building (e.g., building automation systems or BAS, building energy management systems or BEMS). The accuracy of these models, when applied to forecasting, depends on the quality of the selected forecasting model, the quality, and quantity of data available. Such models are easily adaptable to changing conditions, can model nonlinear phenomena, and are relatively easy to train and use. In most cases, the relationship between the forecasted variable and its driving physical functions is not explicitly derived.
Data-driven models are classified by the American Society of Heating, Refrigeration and Air-conditioning Engineers (ASHRAE) [
2] into two main categories: (i) black-box models, which apply a strictly mathematical approach calibrated with measurements; and (ii) grey-box models, which couple a physical model of the HVAC system or building, with a black-box model applied at key parameters within the physical model. Due to their ease of development, accuracy, and applications, data-driven models have gained significant popularity within the past two decades.
Data-driven models typically have two main approaches for the mathematical-based model, a statistical approach or a machine learning algorithm. The statistical approach typically applies a pre-set mathematical function and has shown good performance for medium to long term energy forecasting [
3]. In addition, such models have shown acceptable performance forecasting short term whole building loads [
4]. Furthermore, it was found that they can outperform machine learning methods at a whole city level, however, machine learning methods outperformed them at a more building level [
5]. Machine learning, a subfield of artificial intelligence (AI), in contrast typically applies an algorithmic approach (which may non-linearly transform the data), in order to provide a forecast [
6]. Many such algorithms have shown to be effective for forecasting and include decision trees [
7], random forest [
8,
9], gradient boosting machines [
10], k-nearest neighbors [
11], case-based reasoning [
12], support vector machines [
13], etc.
Exploring the effectiveness of different data-driven models is beyond the scope of this paper and has been previously reviewed over the past decade. However, a brief summary of the findings within previous literature reviews related to building energy forecasting and prediction is provided.
Zhao et al. published a review in 2012 focusing on the main approaches for energy prediction and forecasting in buildings [
14]. Specifically, the authors compared machine learning, statistical, and physics-based models. They noted that the machine learning-based models obtained the highest accuracy and flexibility, especially when compared to statistical models. Support vector machines were elaborated as having superior performance to artificial neural network (ANN) models, however, the following paper was published prior to breakthroughs in deep learning. One of the recommended areas of future investigation was to focus on the application with regards to optimizing parameters for the data-driven models.
Similarly, Daut et al. [
15] published a review in 2017 with a comparison of conventional methods (e.g., times series, regression) and machine learning-based models for forecasting building electrical consumption. They too noted the improved performance with the application of machine learning-based models. Specifically, the authors noted that the conventional methods lack flexibility in dealing with nonlinear patterns which emerge. Such nonlinear patterns occurring in weather, indoor conditions, and occupancy data can greatly influence the overall forecasts of the conventional models resulting in low performance.
Wang and Srinivasan [
16] in 2017 explored the usage of AI models and ensemble models for prediction and forecasting of building energy use. Firstly, the authors provided a breakdown of how AI as a whole have been applied to building energy prediction. The majority of AI-based papers were found to be applied to a whole building load with hourly data. Secondly, the authors explored how ensemble methods have currently been applied within building energy prediction. The authors noted that such ensembles have been widely applied to fields outside building energy, and their results show improved performance compared with single prediction models. However, they noted a lack of papers which have applied ensemble models for prediction of short-term building energy use.
Similarly, Amasyali et al. [
17], in 2018 explored the usage of AI models for building energy forecasting and prediction. The authors found that the majority of papers developed the AI models with hourly data, to a target variable of an overall building energy load, and used measurement data in their case studies. The ANN models were found to be deployed approximately at a two to one ratio when compared to support vector machine learning algorithms. The authors concluded with postulating future research directions.
Wei et al. [
18] in 2018 provided a review for data-driven approaches of both prediction and classification in buildings. The authors noted the wide range of practical applications of ANN models to date including forecasting and prediction of energy loads, ascertaining the current energy performance of a building, and predicting potential for energy saving through retrofit strategies. Their analysis concluded that ANN prediction models have been applied to a commercial building, targeting a total energy load, and with a short-term horizon. However, approximately 15 papers were used in order to provide that conclusion of both prediction and forecasting. Future work was suggested to modify the framework of data-driven approaches responding to the unique requirements of the building prediction models. In addition, exploring the models at different buildings, with various climate conditions and incorporating multiple indices (e.g., thermal comfort) should be considered in future models.