







































S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

47 

 

 

 

Review 

Deep-learning-based multi-timescale load 

forecasting in buildings: opportunities and 

challenges from research to deployment 
Sakshi Mishra*, Stephen M. Frank, Anya Petersen, Robert Buechler, Michelle Slovensky 

Intelligent Campus Program, National Renewable Energy Laboratory, Golden, Colorado, United States 

A R T I C L E   I N F O 
 

Article history: 
Received 12 July 2024  
Received in revised form 
18 August 2024 
Accepted 26 August 2024 
 
Keywords: 
Deep learning, Recurrent neural networks, Long 
short-term memory, Building load forecasting, 
Grid-interactive efficient buildings, Smart grid 
 
*Corresponding author 
Email address:  
Sakshi.M@outlook.com  

 
DOI: 10.55670/fpll.fuen.3.4.3 

A B S T R A C T 
 

Electricity load forecasting for buildings and campuses is becoming 
increasingly important as the penetration of distributed energy resources 
(DERs) grows. Efficient operation and dispatch of DERs require reasonably 
accurate predictions of future energy consumption in order to conduct near-
real-time optimized dispatch of on-site generation and storage assets. Electric 
utilities have traditionally performed load forecasting for load pockets 
spanning large geographic areas, and therefore, forecasting has not been a 
common practice by buildings and campus operators. Given the growing trends 
of research and prototyping in the grid-interactive efficient buildings domain, 
characteristics beyond simple algorithm forecast accuracy are important in 
determining the algorithm’s true utility for smart buildings. Other 
characteristics include the overall design of the deployed architecture and the 
operational efficiency of the forecasting system. In this work, we present a 
deep-learning-based load forecasting system that predicts the building load at 
1-hour intervals for 18 hours in the future. We also discuss challenges 
associated with the real-time deployment of such systems as well as the 
research opportunities presented by a fully functional forecasting system that 
has been developed within the National Renewable Energy Laboratory’s 
Intelligent Campus program. 
 

 
1. Introduction  

The commercial buildings sector in the United States 

consumed 1,671.61 trillion Btu of energy in January 2020 

alone (immediately prior to the onset of the COVID-19 

pandemic) and a total of 18,177.95 trillion Btu in 2019 [1]. 

Overall, buildings account for nearly 40% of the total energy 

consumption in the United States [2], and building energy 

consumption is projected to expand by an annual 1.5% 

globally [3]. Even with many commercial facilities operating 

at reduced occupancy during the COVID-19 pandemic, 

preliminary industry reports indicate that commercial 

building energy is again increasing as operators increase 

ventilation rates [4]. Because buildings are among the largest 

consumers of energy globally, research to increase the design 

and operational energy efficiencies of the commercial sector 

(i.e., commercial buildings) will play a significant role in 

meeting energy and greenhouse gas emissions reduction 

targets. The penetration of behind-the-meter distributed 

energy resources (DERs) is also increasing, owing to the 

decreasing cost of renewable energy technologies [5]. Smart 

building operations can simultaneously reduce energy 

consumption (and, by extension, greenhouse gas emissions) 

and optimize behind-the-meter DER dis-patch to save money 

for building owners/operators via additional value streams 

such as demand management (peak shaving), energy 

arbitrage, user-initiated demand-response, and optimal 

electric vehicle charging. Properly deployed, these 

capabilities also benefit the utility grid by enhancing grid 

reliability and resilience, deferring or reducing capital 

expenditures required to upgrade the distribution grid, and 

helping balance the supply of renewable energy as its 

penetration increases. The U.S. Department of Energy’s Grid-

interactive Efficient Buildings initiative [6] describes the 

multi-faceted benefits of these smart building capabilities. 

Smart building controls are also an integral component of 

autonomous energy grids [7], where they provide 

opportunities for granular device-level controls. 

Proportional–integral–derivative (PID) controllers are still 

widely used in building controls for heating, ventilating, and 

air-conditioning (HVAC) systems, largely because of their 

simplicity and fast solution times, as they employ numerical 

methods to determine the controlling parameters [8]. 

 

 

Future Energy 

Open Access Journal 

https://doi.org/10.55670/fpll.fuen.3.4.3 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

November 2024| Volume 03 | Issue 04 | Pages 47-55 

Journal homepage: https://fupubco.com/fuen 

 
ISSN 2832-0328 

mailto:Sakshi.M@outlook.com
https://doi.org/10.55670/fpll.fuen.3.4.3
https://fupubco.com/fuen


S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

48 

 

However, the time-varying system dynamics characteristics 

of HVAC control systems result in the inconsistent 

performance of traditional controllers, often leading to less-

economic performance. Moreover, the objectives of building 

control systems have shifted from being solely occupant-

comfort and economic operations to being grid-interactive 

prosumers. As prosumers, building energy management 

systems have goals such as (i) maximizing the use of on-site 

distributed energy resources, (ii) managing consumption 

profiles based on time-of-use energy rates, and (iii) using on-

site energy storage to reduce the peak demand for cost 

savings. Such a complex set of objectives requires the use of 

advanced control techniques like model-predictive control 

[9]. Implementation of intelligent dispatch and model-

predictive control algorithms that help achieve the above-

stated goals to reduce overall energy consumption and 

operational costs of the building nearly always requires a 

forecast of the energy consumption of the building. These 

algorithms use these building load predictions to schedule the 

dispatch of flexible loads, on-site clean energy resources, 

and/or energy storage systems to optimize a desired metric, 

such as utility cost. Forecast accuracy is a key driver of model-

predictive control algorithm effectiveness [8-10]. Prediction 

of building energy consumption patterns is also important for 

detecting faults or operational anomalies in energy systems. 

Therefore, building load forecasting systems are an essential 

component of smart buildings. 

Building load forecasting is not a new research question; 

much research work in the literature has addressed this 

problem space [11-13]. Commercial load forecasting tools for 

building energy predictions are also available [14,15]. The 

journey of successful research outcome to industry 

deployment, however, inevitably passes through the 

research-prototyping phase, and advanced building load 

forecasting systems are no exception. This is especially 

relevant for data-driven modeling projects where the type, 

quality, and pre-processing of data largely dictate the 

performance of the forecasting system. The work presented 

in this article sheds light on the end-to-end pipeline of 

building load forecasting systems- in a campus-wide 

prototyping setup- using advanced data-driven methods 

(recurrent neural networks). The following paragraphs 

delineate the types of modeling algorithms used in the 

literature for the building energy forecasting problem and 

specifically discuss works that have utilized data-driven 

modeling methods. 

As building systems become more advanced through 

next-generation sensors, controls, connectivity, and 

communications, they produce a large volume of empirical 

data available to building operators for decision-making. This 

data, along with meteorological parameters, can be harnessed 

to predict building and campus energy consumption. There 

are two main approaches available for this prediction task: (i) 

physics-based (or “white box”) modeling and (ii) data-driven 

(or “black box”) modeling. (Physics-based models are termed 

“white box” because the inner workings of the model are 

typically open to the modeler, whereas data-driven models 

are termed “black box” because their prediction logic is often 

opaque.) A third category, “grey box” models [16], represents 

a combination of the physics-based and data-driven 

approaches. They combine a partial theoretical structure with 

data, offering relatively simpler model architectures. Grey box 

models are generally trained or fit like data-driven models. 

Physics-based building models such as EnergyPlus® [17,18] 

model the physical relationships between the building 

characteristics (construction details, operation schedules, 

shading information) and environmental parameters (sky 

conditions) to calculate building energy consumption [19]. 

Data-driven models, on the other hand, make predictions by 

learning the pattern empirically from historical data. Data-

driven models have two major subcategories: (i) statistical 

models and (ii) machine learning models. Deep learning 

models are a subset of machine learning models that are 

capable of learning complex nonlinear relationships between 

the inputs and the predicted variable(s). Recurrent neural 

networks (RNNs) are an advanced variant of deep neural 

networks that are capable of incorporating temporal 

dependencies between the input and output variables. Long 

short-term memory networks (LSTM) are a variant of RNNs 

that are effective at capturing longer-term temporal 

dependencies in the data sets.  

There are numerous advanced applications of predictive 

analytics in the renewable energy field [20]- spanning the 

energy system from generation (solar and wind forecasting) 

to consumption (smart buildings energy forecasting to fault 

predictions). Because the focus of this article is smart 

buildings’ predictive analytics application, herein, we discuss 

literary works focused on data-driven building load 

forecasting. Recent research has established the utility of 

neural networks for forecasting the energy consumption of 

both individual buildings and groups of buildings. For 

example, a machine-learning-based forecasting model is 

presented in [21], whereas [22] covers a review of the 

intelligent system for power load forecasting. A Polish power 

system study with a deep learning approach is presented in 

[23], and [24] takes a deeper look at a hybrid system based on 

LSTM for short-term power load forecasting; similarly, 

reference [25] presents an RNN-based robust short-term load 

forecasting framework. An online adaptive RNN-based load 

forecasting algorithm with smart meter data is presented in 

[26]. Short-term load forecasting for urban loads using an 

artificial neural network is tackled in reference [27]. An 

LSTM-based method is improved for short-term load 

forecasting [28]. On the research end of the building load 

forecasting spectrum, the existing literature has extensively 

focused on assessing the effectiveness of machine learning 

(specifically deep learning) models for the building load 

prediction problem. The commercial end of building load 

forecasting spectrum, on the other hand, focuses on using 

proven and tested traditional methods in prepackaged 

software products. Both these extreme ends of the building 

load forecasting problem space do not address the “research 

prototyping” step for incrementally adopting cutting-edge 

prediction methodologies. That is to say, none of these works 

have covered a holistic picture of the deep-learning-based 

prediction system deployment process—in a research set-

up—which starts with the data collection and curation 

process and continues to critically assess the challenges and 

opportunities with its end-to-end deployment and continued 

use. The work presented in this paper contributes a 

perspective on how to move data-driven algorithms from 

research to practical deployment by presenting a case study 



S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

49 

 

of designing, building, and deploying a research prototype of 

a deep-learning-based building load forecasting system. The 

current section lays out the introduction and background 

behind the presented work, and the rest of the paper is 

organized as follows: Section II provides a brief introduction 

of the real-time testbed (Intelligent Campus) utilized for 

deploying the deep learning models. Section III discusses the 

algorithm employed in modeling, methodology, and 

benchmarking. Section IV presents the challenges faced, 

lessons learned, and opportunities that a fully functional and 

deployed load forecasting system presents for modern 

buildings. Section V concludes the paper with a discussion of 

outlook. This paper is an extended version of our preprint 

[29]. 

2. Intelligent campus 

The Intelligent Campus platform at the National 

Renewable Energy Laboratory (NREL) was established to 

collect historical and real-time building performance data to 

support analytics that enhance operational awareness and 

decision-making with respect to energy use. The platform was 

developed using open standards and protocols (Figure 1) 

with the intent to provide an architecture that is readily 

transferable to other campuses [30]. Since its inception, the 

Intelligent Campus platform has evolved into a program 

providing an ecosystem for the use of the NREL campus as a 

living laboratory. The Intelligent Campus program offers 

nascent energy analytics technologies and algorithms a 

pathway to mature toward commercialization. The Intelligent 

Campus program includes an interdisciplinary portfolio of 

projects aimed at prototyping the future of smart, sustainable, 

resilient, and self-healing buildings. By providing an 

environment for technology demonstration in real facilities, 

the Intelligent Campus program is bridging the gap between 

theoretical research into building efficiency technologies 

(which are often published but rarely deployed) and practical 

application. Intelligent Campus pilot projects include detailed 

monitoring and analysis in order to identify and improve 

technology shortcomings. 

 

Figure 1. End-to-end data collection architecture (from [29]) 

 

2.1 Data: measurement, collection, and curation 

Advanced analytics and model-predictive control 

algorithms for smart buildings require reliable, internally 

consistent building performance data. NREL collects a variety 

of performance data at its South Table Mountain campus. The 

campus has a single common utility electric meter for all 

facilities, including the campus’s central heating and chilled 

water plant. Eighteen campus buildings have whole-building 

electrical meters; several other small buildings do not have 

dedicated electrical meters. Several newer buildings have 

submetering by end-use, per the requirements of ASHRAE 

90.1 [31]. Facilities that consume heating and chilled water 

from the central plant have thermal meters for heating and 

chilled water consumption. NREL captures the electric and 

thermal meter interval data using a central energy 

management information system (EMIS). The EMIS also 

collects interval data from the campus building automation 

system and syncs data from several cloud data sources via 

web application programming interfaces, including utility 

data from the campus main electricity meter and NREL’s on-

site research weather station [32]. To maximize accessibility 

for operations and research, Intelligent Campus organizes 

collected interval data per the Project Haystack standard 

[30,33]. These interval data are available to serve as inputs 

for predictive load models. 

2.2 Predictive analytics 

Intelligent Campus’s predictive analytics project applies 

recent machine learning and deep learning advances to 

develop and continuously improve load forecasting capability 

for NREL’s South Table Mountain campus. Intelligent Campus 

executes research, prototyping, and enhancement in a cycle 

so that practical deployment lessons can be folded back into 

early-stage research, increasing the ultimate effectiveness 

and eventual impact of the work. Intelligent Campus’s 

building load forecasting work builds from prior NREL 

research that predicts expected building performance based 

on exogenous inputs [34].  

 

 



S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

50 

 

As an academic discipline, load forecasting approaches 

can be studied using static postprocessed data sets. 

Transferal of these results to a real-world setting requires, at 

a minimum, some underlying sensor infrastructure, a 

mapping of sensor data streams to model inputs, and quality 

control on the measured data being fed to the model. 

Therefore, the focus of the current effort is twofold: (i) to 

automate, replicate, and scale the predictive algorithms and 

(ii) to improve accuracy by leveraging state-of-the-art neural 

networks. 

3. Load forecasting using deep learning 

3.1 Deep learning: long short-term memory network 

Artificial neural networks are universal function 

approximators. They are capable of representing complex 

nonlinear relationships in high-dimensional data sets such as 

the one being employed for building load forecasting in this 

work. Deep learning algorithms, which employ multiple 

hidden layers, have powerful generalizing capabilities. That 

is, they are able to make reasonably accurate predictions for 

previously unseen scenarios because of their ability to learn 

the intricate structures in large data sets. Feedforward neural 

networks, sometimes referred to as “vanilla” deep neural 

networks, have the fundamental drawback of independence 

among the time-series samples/data points, which makes 

them an ineffective choice of algorithm for time-series-based 

prediction problems. This is because the entire state of the 

feedforward neural network is cleared after processing each 

time-series sample, which means that the network starts 

mapping the inputs and output for the next time step from 

scratch, thereby failing to account for the impact of the 

previous time step’s input variables. Recurrent neural 

networks (RNN) are advanced types of deep neural networks 

that overcome feed-forward neural networks’ limitations. 

They are different from feedforward neural networks 

because of the presence of additional directed edges that 

introduce temporal memory components, enabling them to 

capture complex nonlinear relationships between the 

temporally related inputs and outputs across multiple time 

steps. This makes them effective at modeling time-series 

forecasting problems. A long short-term memory (LSTM) 

network [35], shown in Figure 2, is an RNN variant that is able 

to learn long-term dependencies between the input features 

and the predicted variable [36]. 

 

Figure 2. LSTM memory cell diagram (from [29,36]) 

 

Figure 2 displays a typical memory cell for an LSTM 

network. In the figure, 𝐶 denotes the cell states at different 

points in time 𝑡; 𝑓𝑡 , 𝑖𝑡, and 𝑜𝑡 denote the forward, input, and 

output gates, respectively; ℎ(𝑡)  is the hidden state at time 𝑡; 

and 𝑊𝑓 , 𝑊𝑖 , 𝑊𝑐 , 𝑊𝑜 are the weights. Further explanation can 

be found in [36]. 

3.2 Performance measure 

The evaluation metric used to measure the performance 

of the forecasting algorithm is the mean squared error (MSE). 

The difference between the ground truth (i.e., the actual 

future time-step value of the measurement being predicted) 

and forecasting values (i.e., output of the deep learning 

model) is calculated using MSE: 

MSE =  
1

𝑛
∑ (AV𝑗

(𝑡)
− OV𝑗

(𝑡)̂
)

2
𝑛
𝑗 = 1            (1) 

where 𝐴𝑉𝑗
(𝑡)

 is a vector of actual (ground truth) values and 

𝑂𝑉𝑗
(𝑡)̂

 is a vector of forecasted (output) values. 

3.3 Case study: building load forecasting on a research 
campus 
    In this case study, the energy consumption at the main 

electricity meter for the Café building on the NREL South 

Table Mountain campus is predicted at an hourly resolution. 

There are six input features used for training the model: 

relative humidity, barometric pressure, dry bulb 

temperature, global horizontal irradiance, total cloud cover, 

and wind speed. These data are from NREL’s Solar Radiation 

Research Laboratory data set [32]; however, they also 

represent measurements commonly available from high-

quality weather stations worldwide. The model is run for 200 

epochs with the data spanning a year’s interval: 10- and 2-

months train/test split, respectively. We use a single-layered 

network with 35 neurons. The following MSE plot (Figure 3) 

shows the train and test losses over several thousand 

iterations. Despite the difference between final train and test 

losses, no overfitting was observed in the experiment. 

 

 

Figure 3. Mean squared error (MSE) error plot for train and test data 

(from [29]) 

 

 



S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

51 

 

The model can produce load forecasts at multiple future 

time steps, as shown in Figure 4 for the Café building's main 

power. The models use the most recent 12 hours of weather 

data to capture the transient effect of weather on energy 

consumption and use the time of day and day of the week to 

capture time-based energy consumption patterns, such as 

those related to building occupancy. Because neural networks 

can learn non-linear relationships, the trained models are 

able to infer complex behaviors, such as the pre-occupancy 

energy spike in the Café due to early morning food 

preparation. Each model outputs a prediction of building load 

from the current time step to 18 hours in the future in 1-hour 

increments. This input-output structure applies to all 

buildings that were studied on the NREL campus. Model 

performance was not seen to change significantly as the 

length of the forecast window increased/decreased, but the 

training time was significantly impacted by the length of the 

training window (number of hours of historical data used as 

input to the model). 

4. Operational load forecasting system deployment 

A primary research objective of the case study was to 

establish deployment and continuous improvement 

pathways for the developed algorithms. That is, developed 

algorithms should be able to be rapidly deployed for testing 

within the Intelligent Campus platform and iteratively 

improved with minimal retooling. The Intelligent Campus 

team intends to expand the algorithm to forecast other 

quantities (such as building electricity meters, thermal 

meters, and photovoltaic system generation) and to support 

additional exogenous variables. Therefore, from a software 

development point of view, the system’s input-output 

architecture must be flexible, replicable, and scalable. Due to 

the nature of the deep learning algorithms used, ongoing 

access to data is an integral part of the continued operation of 

the system. Periodic retraining of the models as more data 

gets collected has the potential to increase model accuracy.  

 

Figure 4. Multi-time horizon forecast (from [29]) 

The following two subsections describe the challenges 

and opportunities that open up for smart buildings’ research 

and prototyping once those challenges are addressed. 

Understanding the deployment aspect of a deep-learning-

based forecasting system, therefore, completes the re-search-

prototyping loop. 

4.1 Challenges 

4.1.1 Data availability and quality control 

“Garbage in, garbage out”––a common phrase used in the 

machine learning community––captures the importance of 

this step of data collection and curation. Assembling a 

machine learning training data set begins with proper 

commissioning and maintenance of meters, sensors, and 

other data inputs. Data should have proper units and scaling. 

(Although in some circumstances, a machine learning model 

could be successfully trained from improperly scaled or 

otherwise corrupted input data, if those data stream errors 

were corrected in the future, the model would need to be 

retrained).  

Automated accumulation of new training data (without 

extensive human review) presents data quality challenges. 

NREL’s Intelligent Campus team has implemented quality 

check mechanisms for the measurements coming from the 

sensors in real time before they are fed to the model. One 

example is removing the outliers identified by the values lying 

outside of three standard deviations. Another is operating 

basic fault detection rules that check for stuck sensors 

(constant value for long periods of time), logically out-of-

range data (such as significant solar radiation recorded at 

night), and similar anomalies. These checks not only improve 

machine learning model quality but also provide an 

opportunity to detect and correct errors in the measurement 

systems and equipment. 

 

 

 

 



S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

52 

 

4.1.2 Integration with EMIS and scalability 

The practical implementation of a forecasting system 

requires reliable interfaces for transferring data to the model 

and for transferring model outputs to an analytics system that 

consumes them. The forecasting system developed for this 

case study communicates with the NREL EMIS via an 

application programming interface defined by the Project 

Haystack standard [33]. The EMIS provides forecast inputs, 

and the forecast system, in turn, writes forecasts back to the 

EMIS. Adherence to a metadata standard and an open 

application programming interface allows the forecasting 

system to be rapidly retrained and redeployed for new sites. 

Any database that adheres to the Project Haystack standard 

and correctly tags the required input points can be connected 

to a new instance of the deep learning algorithm, and a new 

model can be trained with minimal effort. When the 

forecasting system runs in integration with EMIS, it is also 

important to be able to redeploy the model (with either 

architectural or training updates) without service 

interruption. Because forecasts cover a span of time but are 

frequently updated, managing caching is a key challenge. In 

addition, the forecast read/write system must be flexible 

enough to handle arbitrary forecast inputs and outputs. 

Additionally, given the computationally intensive nature of 

deep learning models, scalability in terms of available 

computing power for retraining the models is another 

interesting challenge. 

4.1.3 Online vs. offline training 

Training is an important consideration for deep-

learning-based forecasting systems because deep-learning 

models’ accuracies have been shown to keep rising with more 

training. Offline training is a relatively simpler way of 

implementation in which the machine learning models are 

retrained by a manual process where an engineer copies the 

old model to a local machine, trains it further with the new 

data, and transfers it back to the servers where the systems 

are deployed. Offline training, therefore, requires more 

human hours over the course of the system’s operation 

because human intervention is needed to train the models 

further at regular intervals. The online training approach, on 

the other hand, is an automated process. At its simplest, an 

online training system trains models on a schedule (for 

example, every 1 to 2 months) without any manual 

intervention [37]. Though setting up the automated training 

pipeline initially requires additional planning and coding, this 

online training eliminates the maintenance dependency on 

human intervention to move the models offline, retrain them, 

and upload them back into the forecasting system pipeline 

regularly. Thus, online training can increase operational 

efficiency and reduce opportunities for human error to 

impact the forecasting system. Moreover, because a robust 

metering platform has already been established as the first 

step in the process of our prototype of Intelligent Campus, 

high-quality data are constantly accumulating and can be 

used to increase the accuracy rates of our models. For our 

final implementation, we have chosen the online training 

option.  

4.1.4 Continuous improvement 

As described above, deep learning models are capable 

of increasing accuracy as more data is made available for 

training. Also, in the initial stages, when the accuracy of the 

model is not satisfactory, given the shortage of data points to 

train it, large changes in predictive behavior can be off-

putting to end users. Therefore, along with iterative training, 

we conduct model enhancement with hyperparameter tuning 

as the operational circumstances change over time. This is an 

infrequent yet desirable step. 

4.1.5 Generalizability 

Another typical challenge in deploying deep learning 

models for multiple prediction points (i.e., building-wise 

energy consumption forecasts) is generalizability. Because 

each prediction point has its own model, the deployment 

architecture must address the need for automated model 

generation. To address the challenge of generalizability, the 

forecasting system’s data preprocessing and postprocessing 

modules are designed with the capability to take in a generic 

variable named “point-id” that is used as a reference to fetch 

the respective machine learning model and input data 

streams.  

4.2 Opportunities 

An automated end-to-end pipeline for load forecasting 

deployed on-site, with an online training mechanism in place, 

enables many opportunities to harness the predicted data for 

furthering the smart buildings’ research. In the following 

subsections, we describe three potential applications. 

4.2.1 Demand management 

Motivations for electricity demand management (peak 

shaving), including utility demand charge reductions and 

reduced carbon footprint. (When the peak of multiple 

buildings and campuses coincides with the utility peak, 

lower-efficiency “peaking” power plants fired by gas or diesel 

are used to support the grid, which translates to 

proportionally higher greenhouse gas emissions.) Reducing 

peak demand requires the ability to predict the timing and 

magnitude of peak demand, then preemptively shift load or 

dispatch energy storage assets to avoid the peak. 

4.2.2 Energy arbitrage 

Energy arbitrage is another cost-saving opportunity 

available for grid-interactive buildings in cases where on-site 

energy storage is available. Energy arbitrage is performed by 

participating in the energy markets by charging energy 

storage (for example, batteries) when electricity prices are 

low and discharging storage to sell power back to the grid 

when prices spike. Even in utility markets where bidirectional 

electricity exchange is not possible, a form of energy arbitrage 

is possible by buying energy when prices are low, storing the 

energy locally, and dispatching the energy locally to meet load 

when energy prices are high. Energy arbitrage can be 

accomplished by employing near-real-time or real-time 

optimization algorithms. However, the real-time optimized 

storage dispatch requires reasonable foresight into future on-

site energy consumption, thereby making building load 

forecasts an important component of the system. 

4.2.3 Load flexibility 

A third opportunity to save energy costs for building 

operations is shifting the load based on the time-of-use 

energy rates. Building load can be classified as critical, which 

consists of loads that cannot be deferred, and flexible, which 



S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

53 

 

consists of loads whose consumption can be scheduled within 

a given time window. Examples of critical loads include lights, 

fans, microwaves, computing systems, etc.; these loads must 

be available for use when needed. Examples of flexible loads 

include dishwashers, laundry, and even heating and cooling 

(buildings can be precooled/preheated). With adequate 

foresight, flexible loads can be shifted to minimize energy 

costs. This also requires predicting the building’s future 

energy consumption. 

5. Conclusion 

The paradigm of grid-interactive efficient buildings, i.e., 

dynamically operating buildings that constantly exchange 

information and energy with the utility grid, is increasingly 

gaining traction in both the research community and in 

industry. Such smart buildings will operate in harmony with 

the grid to make electricity more affordable and integrate a 

larger share of DERs while meeting the comfort and 

productivity needs of the buildings’ occupants. NREL’s 

Intelligent Campus is a living laboratory that is dedicated to 

accelerating the research and deployment of various pieces of 

the smart buildings puzzle, serving as a vital testbed for the 

innovative solutions that push the frontier of smart building 

operations. In this article, we describe and critically discuss a 

deep learning building load forecasting system that is an 

important module of smart buildings’ overall controls. 

Through the critical discussion, we highlight that the 

characteristics beyond simple algorithm forecast accuracy 

are important in determining the algorithm’s true utility for 

smart buildings. We demonstrate that metering, collection, 

and curation of data is a crucial part of the prototyping 

process of such an advanced forecasting system. Using LSTMs, 

the deep learning model can predict the building and campus 

load for an 18-hour’ time horizon. The architecture is flexible 

with the time resolution; sub-hourly predictions can be 

enabled with minimal modifications to the code. Using 

Intelligent Campus as the platform for testing the research 

findings in real-time settings, we shed light on the 

deployment aspect of such advanced forecasting systems. We 

conclude that a well-architected design of the forecasting 

system is key to its effectiveness. To take a forecasting 

algorithm from the research phase to deployment for 

practical use, building engineers will need to focus on the end-

to-end pipeline, which requires considerations for multiple 

submodules such as access to data, data handling, seamless 

integration with the existing building data platform, model 

training, and update mechanism, periodic performance 

checks, and model enhancement with hyperparameter tuning 

as operational circumstances change. The end-to-end 

forecasting system presented in this work forms a robust 

platform for furthering research in the building energy 

predictive analytics problem space. We plan to expand this 

work in the following directions: 

5.1 Probabilistic forecasting 

Rather than generating a point forecast for every time 

step, a range can be outputted with the estimated minimum 

and maximum values serving as the upper and lower bounds 

for a prediction band within which the actual consumption in 

the future time step will fall. Having a range instead of point 

forecasting is especially helpful in the fault detection 

application where consistent outliers (outside of the band 

points) indicate a potential issue that can be further 

investigated by the building’s operations team. Moreover, 

probabilistic forecasting using deep neural networks is an 

active inquiry front on the research end of the forecasting 

technology’s spectrum. 

5.2 Predictive maintenance 

Current practice in the building industry is to perform 

“schedule-based” maintenance for fault prevention, which is 

not effective in terms of flagging the equipment vulnerable to 

a fault beforehand. The next generation of the fault-

prevention mechanism is “condition-based” or “preventive” 

maintenance. This involves using statistical analysis to assess 

the health of the patients and estimate the probability of 

failure in order to inform the operators. The presented 

forecasting system can be further developed for predictive-

maintenance applications where the occurrence of a specific 

fault on the system, which may occur due to unusual wear-

and-tear on equipment, can be predicted well in advance, 

thereby preventing equipment failure from triggering faults. 

Author contributions  

Conceptualization, Sakshi Mishra, Stephen M. Frank, and Anya 
Petersen; method-ology, Sakshi Mishra and Stephen M. Frank; 
software, Anya Petersen and Sakshi Mishra; validation, 
Stephen M. Frank and Robert Buechler; formal analysis, 
Sakshi Mishra; resources, Stephen M. Frank and Michelle 
Slovensky; data curation, Stephen M. Frank; writing—original 
draft preparation, Sakshi Mishra, Stephen M. Frank, and Anya 
Petersen; writing—review and editing, Robert Buechler and 
Michelle Slovensky; visualization, Sakshi Mishra, Robert 
Buechler, and Marj Schott; supervision, Michelle Slovensky; 
project administration, Anya Petersen; funding acquisition, 
Michelle Slovensky. All authors have read and agreed to the 
published version of the manuscript. 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically concerning authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with policies on research ethics. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere in any language. 

Funding 

This research was funded by the U.S. Department of Energy 
Office of Energy Efficiency and Renewable Energy Building 
Technologies Office. 

Acknowledgments 

 This work was authored by NREL, operated by Alliance for 
Sustainable Energy, LLC, for the U.S. Department of Energy 
(DOE) under Contract No. DE-AC36-08GO28308. Funding 
provided by the U.S. Department of Energy Office of Energy 
Efficiency and Renewable Energy Building Technologies 
Office. The views expressed in the article do not necessarily 
represent the views of the DOE or the U.S. Government. The 
authors wish to acknowledge the National Renewable Energy 
Laboratory’s continued support in enhancing the research 
and prototyping capabilities of the Intelligent Campus 
program. The authors would like to thank Marj Schott (NREL) 
for contributing Figure 1 to the article. 

 

 



S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

54 

 

Data availability statement 
Data sharing does not apply to this article as no datasets were 

generated or analyzed during the current study.  

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1] Total Energy - Monthly Energy Review, U.S. Energy 

Information Administration,, 02 05 2020. [Online]. 

Available: 

https://www.eia.gov/totalenergy/data/monthly/inde

x.php. [Accessed 02 05 2020]. 

[2] U. DOE, Buildings energy data book. 

Avalable: https://data.openei.org/submissions/144 

[3] P. G. Jordan, “Global Markets” in Solar Energy Markets, 

2014, pp. 127-133. 

[4] Energy-Spectral, During COVID-19 Lockdown, Office 

Buildings are Consuming Even More HVAC Energy 

Than When Fully Occupied, Spectral Energy, 2020. 

[Online]. Available: 

https://spectral.energy/news/covid19-office-

building-energy-consumption-increasing/. [Accessed 

25 07 2020]. 

[5]  P. Donohoo-Vallett, P. Gilman, D. Feldman, J. Brodrick, 

D. Gohlke, R. Gravel, A. Jiron, C. Schutte, S. Satyapal, T. 

Nguyen, P. Scheihing, B. Marshall and S. Harman, 

"Revolution Now The Future Arrives for Five Clean 

Energy Technologies," U.S. DOE, 2016. 

[6]  M. Neukomm, V. Nubbe and R. Fares, "Grid-interactive 

Efficient Buildings Technical Report Series," DoE Office 

of Energy Efficiency & Renewable Energy, 2019. 

[7]  B. Kroposki, E. Dall’Anese, A. Bernstein, Y. Zhang and 

B.-M. Hodge, "Autonomous Energy Grids," 2017.  

[8]  A. Afram and F. Janabi-Sharifi, "Theory and 

applications of HVAC control systems - A review of 

model predictive contol (MPC)," Building and 

Environment, vol. 72, pp. 343-355, 2014.  

[9]  E. Khanmirza, A. Esmaeilzadeh and A. H. D. Markazi, 

"Predictive control of a building hybrid heating system 

for energy cost reduction," Applied Soft Computing, 

vol. 46, pp. 407-23, 2016.  

[10]  T.-T. Nguyen, H.-J. Yoo and H.-M. Kim, "Analyzing the 

Impacts of System Parameters on MPC-Based 

Frequency Control for a Stand-Alone Microgrid," 

Energies, vol. 10, no. 4, 2017.  

[11]  Y. Iino, M. Murai, D. Murayama, I. Motoyama, S. 

Kuzusaka and K. Ueta, "Hybrid modeling with physical 

and JIT model for building thermal load prediction and 

optimal energy saving control," in ICCAS-SICE, 

Fukuoka, Japan, 2009.  

[12]  F. Zhang and X. Zhou, "Gray-Regression Variable 

Weight Combination Model for Load Forecasting," in 

2008 International Conference on Risk Management & 

Engineering Management, Beijing, China, 2008.  

[13]  Y. Iwafune, Y. Yagita, T. Ikegami and K. Ogimoto, 

"Short-term forecasting of residential building load for 

distributed energy management," in 2014 IEEE 

International Energy Conference (ENERGYCON), 

Cavtat, Croatia, 2014.  

[14]  Itron, "Energy Forecasting". https://na.itron.com/ 

[15]  Enverus, "Electric Gas and Load Forecast". 

https://www.enverus.com/ 

[16]  B. Sohlberg and Jacobsen, "Grey Box Modelling - 

Branches and Experiences," IFAC Proceedings 

Volumes, vol. 41, no. 2, pp. 11415-11420, 2008.  

[17]  D. B. Crawley, L. K. Lawrie, F. C. Winkelmann, W. F. 

Buhl, Y. H. C. O. Pedersen, R. K. Strand, R. J. Liesen, D. E. 

Fisher, M. J.Witte and J. Glazer, "EnergyPlus: creating a 

new-generation building energy simulation program," 

Energy and Buildings, vol. 33, no. 4, pp. 319-331, 2001.  

[18]  EERE, "EnergyPlus," energy.gov, 2020. [Online]. 

Available: 

https://www.energy.gov/eere/buildings/downloads/

energyplus-0. 

[19]  S. S. Kwok and E. W. M. Lee, "A study of the importance 

of occupancy to building cooling load in prediction by 

intelligent approach," Energy Conversion and 

Management, vol. 52, no. 7, pp. 2555-2564, 2011.  

[20]  S. Mishra, A. Glaws and P. Palanisamy, "Predictive 

Analytics in Future Power Systems: A Panorama and 

State-Of-The-Art of Deep Learning Applications," in 

Optimization, Learning, and Control for 

Interdependent Complex Networks, SpringerLink, 

2020, pp. 147-182. 

[21]  J. Hwang, D. Suh and M.-O. Otto, "Forecasting 

Electricity Consumption in Commercial Buildings 

Using a Machine Learning Approach," Energies, vol. 13, 

no. 22, 2020.  

[22]  I. S. Jahan, V. Snasel and S. Misak, "Intelligent Systems 

for Power Load Forecasting: A Study Review," 

Energies, vol. 13, no. 22, 2020.  

[23]  T. Ciechulski and S. Osowski, "Deep Learning Approach 

to Power Demand Forecasting in Polish Power 

System," Energies, vol. 13, no. 22.  

[24]  Y. Jin, H. Guo, J. Wang and A. Song, "A Hybrid System 

Based on LSTM for Short-Term Power Load 

Forecasting," Energies, vol. 13, no. 23, 2020.  

[25]  G. Chitalia, M. Pipattanasomporn, V. Garg and S. 

Rahman, "Robust short-term electrical load forecasting 

framework for commercial buildings using deep 

recurrent neural networks," Applied Energy, vol. 278, 

2020.  

[26]  M. N. Fekri, H. Patel, K. Grolinger and V. Sharma, "Deep 

learning for load forecasting with smart meter data: 

Online Adaptive Recurrent Neural Network," Applied 

Energy, vol. 282, 2021.  

[27]  N. Gautam, A. S. Mayal, V. S. Ram and A. Priya, "Short 

Term Load Forecasting Of Urban Loads Based On 

Artificial Neural Network," in 2nd International 

Conference on Power and Embedded Drive Control 

(ICPEDC), 2019, 2019.  

[28]  J. Cui, Q. Gao and D. Li, "Improved Long Short-Term 

Memory Network Based Short Term Load 

Forecasting," Hangzhou, China, 2020.  

[29]  S. Mishra, S. M. Frank, A. Petersen, R. Buechler and M. 

Slovensky, "Deep-Learning-Based, Multi-Timescale 

Load Forecasting in Buildings: Opportunities and 

Challenges from Research to Deployment (preprint)," 

vol. 1, no. 1, pp. 1 - 13, 2020.  

[30]  D. Cutler, S. Frank, M. Slovensky, M. Sheppy and A. 

Petersen, "Creating an Energy Intelligent Campus: Data 

https://data.openei.org/submissions/144


S. Mishra et al. /Future Energy                                                                                               November 2024| Volume 03 | Issue 04| Pages 47-55 

55 

 

Integration Challenges and Solutions at a Large 

Research Campus," in ACEEE Summer Study on Energy 

Efficiency in Buildings, Pacific Grove, CA, 2016.  

[31]  ASHRAE, "Energy Standard for Buildings Except Low-

Rise Residential Buildings, Standard 90.1-2019," 2019. 

[32]  T. Stoffel and A. Andreas, "NREL Solar Radiation 

Research Laboratory (SRRL): Baseline Measurement 

System (BMS)," National Renewable Energy 

Laboratory, Golden, Colorado (Data), 1981. 

[33]  "Project Haystack," 2020. [Online]. Available: 

https://project-haystack.org/. 

[34]  G. P. Henze, S. Pless, A. Petersen, N. Long and A. T. 

Scambos, "Control limits for building energy end use 

based on frequency analysis and quantile regression," 

Energy Efficiency, vol. 8, p. 1077–1092, 2015.  

[35]  S. Hochreiter and J. Schmihuber, "Long Short-Term 

Memory," Neural Computation, vol. 9, no. 8, pp. 1735-

1780, 1997.  

[36]  S. Mishra and P. Palanisamy, "An Integrated Multi-

Time-Scale Modeling for Solar Irradiance Forecasting 

Using Deep Learning," arxiv.org, 2019.  

[37]  Cloud-Architecture-Center, Google, "MLOps: 

Continuous delivery and automation pipelines in 

machine learning," 2021. 

 

 
This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

 

 

https://creativecommons.org/licenses/by/4.0/

