Elsevier

Applied Energy

Volume 263, 1 April 2020, 114683
Applied Energy

Building thermal load prediction through shallow machine learning and deep learning

https://doi.org/10.1016/j.apenergy.2020.114683Get rights and content

Highlights

  • Building load prediction informs chiller plant and thermal storage optimization.

  • We used and compared 9 machine learning algorithms and 3 heuristic prediction methods.

  • XGBoost and LSTM are found to be the best shallow and deep learning algorithm.

  • LSTM is better for short term prediction, while XGBoost for long term prediction.

  • It is better to train the model with uncertain rather than accurate weather data.

Abstract

Building thermal load prediction informs the optimization of cooling plant and thermal energy storage. Physics-based prediction models of building thermal load are constrained by the model and input complexity. In this study, we developed 12 data-driven models (7 shallow learning, 2 deep learning, and 3 heuristic methods) to predict building thermal load and compared shallow machine learning and deep learning. The 12 prediction models were compared with the measured cooling demand. It was found XGBoost (Extreme Gradient Boost) and LSTM (Long Short Term Memory) provided the most accurate load prediction in the shallow and deep learning category, and both outperformed the best baseline model, which uses the previous day’s data for prediction. Then, we discussed how the prediction horizon and input uncertainty would influence the load prediction accuracy. Major conclusions are twofold: first, LSTM performs well in short-term prediction (1 h ahead) but not in long term prediction (24 h ahead), because the sequential information becomes less relevant and accordingly not so useful when the prediction horizon is long. Second, the presence of weather forecast uncertainty deteriorates XGBoost’s accuracy and favors LSTM, because the sequential information makes the model more robust to input uncertainty. Training the model with the uncertain rather than accurate weather data could enhance the model’s robustness. Our findings have two implications for practice. First, LSTM is recommended for short-term load prediction given that weather forecast uncertainty is unavoidable. Second, XGBoost is recommended for long term prediction, and the model should be trained with the presence of input uncertainty.

Introduction

The building sector is a major energy consumer and carbon emitter in modern society [1]. To reduce building energy usage and its associated carbon emissions, building thermal load prediction could play an important role. It has wide applications in HVAC control optimization [2], thermal energy storage operation [3], energy distribution system planning [4], and smart grid management [5] among others.

Because of its wide application, much research has been conducted to predict building thermal load, and those studies date back to the 1980s [6]. The approaches to forecasting building thermal load could be generally classified into three categories: white-box physics-based models, gray-box reduced-order models, and black-box data-driven models,1 as shown in Fig. 1.

White-box models predict building loads with detailed heat and mass transfer equations. Some mature software tools (such as EnergyPlus, Dest, and TRNSYS) are commercially available to set up white-box models [7]. To develop a detailed physics-based model, many detailed inputs are needed, and it is time-consuming to collect that information. More important, the uncertain and inaccurate inputs lead to a marked gap between the model results and reality [10].

Gray-box models simplify the building thermal dynamics to reduced order Resistance and Capacity (RC) models. Typically, the parameter values (Rs and Cs) are inferred from measured data (a.k.a. parameter identification) by minimizing the prediction errors, rather than from the specification of building physical parameters, as in the white-box model [11]. With new data available, the inferred parameters might change to reflect different building operations (window closing/opening) or material deterioration. In this regard, the gray-box model can be self-adaptive [12]. Models with different orders (how many Rs and Cs to represent the thermal zone or building envelope) have been proposed and studied, ranging from 1R1C [13], 2R2C [14], 3R2C [14], 3R3C [15], and 5R4C [15].2 Harb et al. compared RC models with different orders, and recommended the 4R2C model configuration [16]. A key shortcoming of a gray-box model is it only considers external heat gains, and overlooks internal heat gains. The gray-box models lack ways to identify and reflect the schedule and intensity change of occupant, lighting, and plug loads. As indicated in [17], the internal heat gains account for an increasingly higher proportion in modern buildings with a high-efficiency level of envelope.

Black-box models are purely data-driven; they predict building thermal load using historical data. As increasing amounts of data are monitored and collected, it becomes possible to learn the load patterns from historical data and use these learned patterns to make predictions. Early load forecast models were statistics-based models. For instance, in the 1980s, linear regression [18] (1984) and autoregressive integrated moving-average (ARIMA) [6] (1989) were applied to forecast building load. In subsequent years more complicated methods have been used. Some popular choices include support vector machine (SVM) [19]; artificial neural networks (ANN) [20]; extreme learning machine [21]; regression tree [22]; random forest [23]; and Hierarchical Mixture of Experts [24].

With the rapid development of deep learning in recent years, deep neural networks have been introduced as tools for load prediction. The major distinction between “shallow” and “deep” machine learning models lies in the number of linear or non-linear transformations the input data experiences before reaching an output. Deep models typically transform the inputs multiple times before delivering the outputs, while shallow models usually transform the inputs only one or two times. As a result, deep models can learn more complicated patterns, allowing end-to-end learning without manual feature engineering, and they perform well in tasks such as computer vision and sequential data analytics.

The most straightforward deep model is multilayer perceptron (MLP). MLP adds multiple hidden layers to an ordinary neural network to enhance its capability to learn more complicated patterns. Massana et al. applied MLP to predict the load for non-residential buildings [25]. However, MLP is not specially designed for time series data analytics, as it fails to capture and retain sequential information. Contrarily, a recurrent neural network (RNN), as a special form of the deep neural network, is specially designed to deal with time-series data. RNN is suitable for building load forecasts, as building load is essentially a time series. The long short term memory (LSTM) is a special form of RNN, which is designed to handle long sequential data. LSTM has been used successfully to forecast internal heat gains [17] or building energy usage [26]. However, to the best of the authors’ knowledge, LSTM has not been used for building thermal load prediction.

Because many different algorithms are used to develop black-box load prediction models, a natural question is: “Is there a particular algorithm that is superior to others?” Li et al. compared SVM and ANN and found SVM performed better [27]. Guo et al. compared multivariable linear regression, SVM, ANN, and extreme learning machine and found extreme learning machine outperformed others [21]. Fan et al. compared seven machine learning algorithms (multiple linear regression, elastic net, random forests, gradient boosting machines, SVM, extreme gradient boosting, and deep neural network) and found extreme gradient boosting combined with deep auto-coding performed best [28]. Wang et al. compared LSTM and ARIMA and found LSTM performs better than ARIMA in plug load prediction [29]. Rather than selecting the best predictor, ensembling multiple different predictors into one model could provide better generalization performance [23].

In addition to the prediction algorithms, determining what features should be used plays an important role in determining black-box model prediction accuracy. Time-related variables (hour of day, day type) are usually used, as they could reflect occupancy pattern, internal heat gains, and building usage schedule (like temperature set point) [28]. Outdoor weather variables (temperature and relative humidity) are also widely selected as features because weather conditions markedly influence the fresh air thermal load [30] and heat transfer through the building envelope. Some studies [21] used the return chilled water temperature as a predicting variable, as the return chilled water temperature could indicate real-time cooling demand and thermal mass of the current time step. However, the return chilled water temperature can only be used for short-term load prediction (on the scale of hours); it is not suitable for long-term forecasting (on the scale of days).

As a rapidly developing area, new machine learning algorithms are being developed consistently. Though various literature has compared different algorithms, it is always worthwhile to investigate and reflect which algorithm performs best under the context of building load prediction. Additionally, to the best of the authors’ knowledge, some cutting-edge machine learning techniques (such as Extreme Gradient Boosting [XGBoost] and LSTM) have not been applied to forecast the building thermal load. To address this research gap, we aimed to predict building cooling load with two cutting-edge machine learning techniques—XGBoost and LSTM—which represent shallow machine learning and deep learning, respectively. XGBoost and LSTM were implemented with the XGBoost library [31] and Keras [32] in Python. These two tools are powerful, as most Kaggle competition3 winners used either the XGBoost library (for shallow machine learning) or Keras (for deep learning) [33]. The first objective of this study was to compare the performance of shallow and deep learning under the context of building load prediction.

The second research gap in past studies is how the input uncertainty would influence the performance and selection of machine learning algorithms. Under the context of building load prediction, weather data is one of the inputs that might be associated with uncertainty. For instance, to predict building load, we need to input weather forecast, which unavoidably has forecast uncertainty. The influence of weather forecast uncertainty is overlooked in previous studies. Given the existence of input uncertainty, for instance weather forecast uncertainty, which machine learning algorithm performs better and how to make the prediction model more robust is the second research question to be answered in this study.

Section snippets

Methodology

This study’s research roadmap is illustrated in Fig. 2. The feature and the algorithm are two pillars of any data-driven models, and these will be discussed in 2.1 Algorithms, 2.2 Features. Section 2.3 details how we established the baseline models and introduced the evaluation index for model comparison. Prior to presenting the results, we introduce the case study building and the associated data collection process in Section 2.4.

Performance of baseline models

We compared the three baseline models and selected the one that performed the best for later comparison with the two machine learning models. As shown in Table 3 and Fig. 6, the most naïve approach—using the load of the previous day as the forecast of the coming day—outperforms the other two baseline models. Therefore, the persistence_daily approach was selected as the baseline model of this study.

Considering the daily periodical behavior of the occupant, lighting, and plug-load schedules, the

Algorithm comparison

In this paper, we compared 12 approaches for load prediction: three heuristic methods, seven shallow machine learning algorithms, and two deep learning algorithms. XGBoost delivers the most accurate prediction in the shallow machine learning category, and LSTM performs the best in the deep learning category. Support Vector Machine, Random Forest, and vanilla Deep Neural Network provide similar prediction accuracy. The above five algorithms outperform the best baseline model developed from

Conclusions

Building load prediction has wide applications in the fields of HVAC control, thermal storage operation, smart grid management, and others. There are three approaches for building load prediction: white-box physics-based models, gray-box reduced-order models, and black-box data-driven models. However, a white-box model requires a huge amount of input parameters, which could degrade with time and might be challenging to measure. A gray-box model overlooks the variation of internal heat gains. In

CRediT authorship contribution statement

Zhe Wang: Conceptualization, Methodology, Data curation, Investigation, Writing - original draft. Tianzhen Hong: Supervision, Conceptualization, Methodology, Writing - review & editing. Mary Ann Piette: Conceptualization, Writing - review & editing, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This research was supported by the Assistant Secretary for Energy Efficiency and Renewable Energy, Office of Building Technologies of the United States Department of Energy, under Contract No. DE-AC02-05CH11231.

References (42)

  • S. Wang et al.

    Simplified building model for transient thermal performance estimation using GA-based parameter identification

    Int J Therm Sci

    (Apr. 2006)
  • D.H. Blum et al.

    Practical factors of envelope model setup and their effects on the performance of model predictive control for building heating, ventilating, and air conditioning systems

    Appl Energy

    (Feb. 2019)
  • H. Harb et al.

    Development and validation of grey-box models for forecasting the thermal response of occupied buildings

    Energy Build

    (Apr. 2016)
  • Z. Wang et al.

    Data fusion in predicting internal heat gains for office buildings through a deep learning approach

    Appl Energy

    (Apr. 2019)
  • Y. Wei

    Prediction of occupancy level and energy consumption in office building using blind system identification and neural networks

    Appl Energy

    (Apr. 2019)
  • Y. Guo

    Machine learning-based thermal response time ahead energy demand prediction for building heating systems

    Appl Energy

    (Jul. 2018)
  • J.-S. Chou et al.

    Modeling heating and cooling loads by artificial intelligence for energy-efficient building design

    Energy Build

    (Oct. 2014)
  • C. Fan et al.

    Development of prediction models for next-day building energy consumption and peak power demand using data mining techniques

    Appl Energy

    (Aug. 2014)
  • R.E. Edwards et al.

    Predicting future hourly residential electrical consumption: A machine learning case study

    Energy Build

    (Jun. 2012)
  • J. Massana et al.

    Short-term load forecasting in a non-residential building contrasting models and attributes

    Energy Build

    (Apr. 2015)
  • W. Wang et al.

    Forecasting district-scale energy dynamics through integrating building network and long short-term memory learning algorithm

    Appl Energy

    (Aug. 2019)
  • Cited by (239)

    View all citing articles on Scopus
    View full text