oa Machine Learning-Based Building Energy Consumption Prediction

Energy is the lifeblood of modern societies. In the past decades, the world's energy consumption and associated CO2 emissions have increased rapidly due to the increases in population and comfort demand of people. In this decade, the increase in energy consumption and associated CO2 emissions are expected to continue due to the demand coming from developing countries such as China, India, and some Middle East countries. Negative environmental impacts, such as air pollution and global warming, are being triggered by the generation and use of non-renewable energy, including oil and natural gas. Buildings are a significant source of the world's energy consumption. The building sector is responsible for 39% and 40% of the energy consumption and 38% and 36% of the CO2 emissions in the U.S. (Becerik-Gerber et al. 2014) and Europe (Ahmad et al. 2014), respectively. Buildings, therefore, offer a great potential for reducing the world's energy consumption and limiting the negative impacts caused by the use of non-renewable sources. Improving building energy efficiency is one of the best strategies for reducing the energy consumption of buildings, while maintaining the comfort and well-being of the building occupants. Building occupants care about their comfort and well-being, as well as about the energy cost and the environment. A recent study showed that thermal comfort, visual comfort, indoor air quality, health, personal productivity, energy cost saving, and environmental protection are moderately important or higher to residential and office building occupants (Amasyali and El-Gohary 2016).

In this regard, building energy efficiency drew a lot of research attention; a relatively large number of research studies have been undertaken in the field of building energy efficiency. These efforts can be classified into five categories: (1) efforts to improve the efficiency of building appliances and materials; (2) efforts toward increasing the use of renewable energy sources; (3) new policies, incentives, and regulations to reduce energy consumption; (4) efforts toward improving occupant behavior, and (5) efforts to automate building control. Studies in all these categories require accurate building energy consumption prediction for improved energy decision making. Building energy software tools (physical models), such as EnergyPlus and eQuest, are being widely used for energy consumption prediction. These tools are, however, very elaborate and therefore requires a significant number of input parameters that are not always available to users. In order to predict energy consumption of buildings without many input parameters, data-driven models were developed. Data-driven approaches utilize historical input data (e.g., outdoor weather conditions, electricity consumption) for developing a prediction model. In any data-driven approach, developing a model consists of four steps: data collection, data preprocessing, model training, and model testing. In the area of building energy consumption prediction, the types of data collected could be data from sensors that are utilized in empirical building energy studies, data generated by a building energy simulation software, or data from publicly available generic datasets (e.g., datasets provided for energy consumption prediction competitions). Data preprocessing includes data cleaning, data transformation, data normalization, and data interpolation. Model training is the training of the prediction model using the historical data. Support Vector Machine (SVM), Artificial Neural Networks (ANN), decision tree, and statistical techniques are the most commonly used training algorithms and techniques. Model testing is the evaluation of the prediction model using some standard evaluation measures.

This paper presents a data-driven building energy consumption model. The authors used publicly available generic data and sensor data: the ASHRAE's Great Building Energy Predictor Shootout dataset (ASHRAE dataset) and a dataset gathered from an office building in Philadelphia, PA which was instrumented and monitored for this study (PA dataset). The lengths of the ASHRAE and PA datasets are six and two months, respectively. These datasets were cleaned, transformed to the format required by the learning algorithm, and normalized. SVM was used as the training algorithm. SVM is a kernel-based machine learning (ML) algorithm that can be used for both regression and classification (Wu et al. 2008). The goal of this algorithm is to find a function f(x) that has at most epsilon (ε) deviation from the actually obtained target yi for all the training data and at the same time is as flat as possible (Vapnik, 1995). The algorithm can solve non-linear problems even with a small amount of training data (Zhao and Magoules 2012). SVM is one of the most robust and accurate algorithms and has been listed in the top ten most influential data mining algorithms in the research community by the IEEE International Conference on Data Mining (Wu et al. 2008). It was found to outperform other ML algorithms in numerous applications. For model testing, the coefficient of variation (CV) and the mean bias error (MBE) were used to evaluate the performance of the models in predicting energy consumption. CV and MBE are performance criteria, provided by ASHRAE, for evaluating energy consumption level prediction algorithms. CV determines how much the overall prediction error varies with respect to the target's mean and MBE determines how likely a particular model is to over-estimate or under-estimate the actual data (Edwards et al. 2012).

The LIBSVM software package was used to implement the SVM algorithm. Eight input parameters were used for the prediction model: outdoor dry-bulb temperature of the current hour, outdoor dry bulb temperature of the previous hour, solar radiation intensity of the current hour, solar radiation intensity of the previous hour, wind speed of the current hour, relative humidity of current hour, energy consumption of the previous hour, and energy consumption of the two hours ago. The following model parameters were used: nu-SVR (type of SVM), radial basis function (kernel type), and 1500000 (cost). As shown in Fig. 1, the predicted results of the model on the ASHRAE dataset, showed a good fitness with the actual energy consumption. In the end, the models showed that it has many promising features that could make it more reliable for effective energy decision making. The model achieved 3.71% CV and 0.30% MBE.

The authors are currently working on improving the accuracy of the model based on the ASHRAE dataset, as well as extending the model to an occupant-behavior-sensitive energy consumption prediction model based on the PA dataset. The model will be able to predict overall building energy consumption on a daily, hourly, and sub-hourly basis. It will predict energy consumption based on (1) indoor environmental condition data (e.g., indoor temperature and relative humidity), (2) occupant energy use behavior data (e.g., thermostat setpoints), and (3) outdoor weather condition data (e.g., ambient temperature and ambient relative humidity). The authors will also focus on: (1) assessing the effectiveness of utilizing indoor environmental condition data and occupant energy use behavior data for energy consumption prediction, using sensitivity analysis that are going to be conducted individually for the daily, hourly, and sub-hourly prediction models, and

(2) comparing the performances of the daily, hourly and sub-hourly prediction models. References: Ahmad, A. S., Hassan, M. Y., Abdullah, M. P., Rahman, H. A., Hussin, F., Abdullah, H., & Saidur, R. (2014). A review on applications of ANN and SVM for building electrical energy consumption forecasting. Renewable and Sustainable Energy Reviews, 33, 102–109.

Amasyali, K., & El-Gohary, N. M. (2016). Energy-related values and satisfaction levels of residential and office building occupants. Building and Environment, 95, 251–263.

Becerik-Gerber, B., Siddiqui, M. K., Brilakis, I., El-Anwar, O., El-Gohary, N., Mahfouz, T., … & Kandil, A. A. (2013). Civil engineering grand challenges: Opportunities for data sensing, information analysis, and knowledge discovery. Journal of Computing in Civil Engineering, 28(4), 04014013.

Edwards, R. E., New, J., & Parker, L. E. (2012). Predicting future hourly residential electrical consumption: A machine learning case study. Energy and Buildings, 49, 591–603.

Vapnik, V (1995). The Nature of Statistical Learning Theory. Springer-Verlag New York, Inc., New York, NY, USA.

Wu, X., Kumar, V., Quinlan, J. R., Ghosh, J., Yang, Q., Motoda, H., … & Steinberg, D. (2008). Top 10 algorithms in data mining. Knowledge and Information Systems, 14(1), 1–37.

Zhao, H. X., & Magoulès, F. (2012). A review on the prediction of building energy consumption. Renewable and Sustainable Energy Reviews, 16(6), 3586–3592.