Modeling and Optimization Method for Building Energy

Performance in the Design Stage

ed
Cong Lia,b, Youming Chena,b,*
a College of Civil Engineering, Hunan University, Changsha, Hunan 410082, China
b Key Laboratory of Building Safety and Energy Efficiency of the Ministry of Education, Hunan University,
Changsha, Hunan 410082, China

iew
Abstract
Building energy consumption is responsible for 30% to 40% of the total society's energy
consumption. There is great energy-saving potential in the building design stage. Good building
designs can substantially reduce energy consumption in the long term. This study introduces a data-
driven, modeling and optimization method for building energy performance in the design stage. The

v
method has four steps: building model establishment, dataset generation, surrogate modeling, and
optimization. Six potential regression methods and five potential optimization methods are applied

re
and compared in two cases; where the regression models are used to train surrogate models, and the
optimization methods are used to find the optimal designs. One case uses a constantly operating AC
system and the other case uses an intermittently operating AC system. The results show the same
trends in both cases. RFR has the best accuracy in predicting the annual building energy
consumption, with an R-squared value consistently above 0.99 for varying subcases. AdaBoost has
er
the poorest performance among the six regression methods. The results indicate that 2000 to 3500
subcases are enough for the optimization of one case. In the case with a constantly operating AC
system, the combination of XGBoost and DE found the best design with minimal annual building
pe
energy consumption of 1060 GJ; In the case with an intermittently operating AC system, the
combination of GB and PSO found the best design with minimal annual building energy
consumption of 565 GJ. The combination of RFR and PSO failed to find the optimal design in a
limited time for both cases.
Abbreviation
ot

AC, Air Conditioning; SVR, Support Vector Regression; RFR, Random Forest Regression; GB,
Gradient Boosting; XGBoost, Extreme Gradient Boosting; LightGBM, Light Gradient Boosting;
AdaBoost, Adaptive Boosting; PSO, Particle Swarm Optimization; DE, Differential Evolution; NM,
tn

Nelder-Mead; CG, Conjugate Gradient; CG-TR, Conjugate Gradient with Trust Region; BFGS,
Broyden-Fletcher-Goldfarb-Shanno; L-BFGS-B, Limited-memory Broyden-Fletcher-Goldfarb-
Shanno with Bounds; HSCW, Hot Summer and Cold Winter
Keywords
rin

Building energy performance; Building design; Surrogate model; Optimization; Regression; data-
driven
1. Introduction
Since the 20th century, the world has faced a global energy crisis. Efforts have been made in
ep

various areas to reduce energy consumption. The building energy consumption is responsible for 30%
to 40% of the total energy consumption [1]. In many countries, air conditioning plays an important
role in building energy consumption [2, 3]. Energy conservation plans in the building design stage
can provide the most valuable strategies with fewer constraints. Extensive research shows that good
Pr

building designs can substantially reduce energy consumption in the long run [4, 5]. A good building
design can significantly reduce energy consumption to maintain indoor thermal comfort [6].
The building energy performance is influenced by multiple factors. To find the optimal building
1

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 design in the design stage, a dataset of simulated data is needed to establish a surrogate model. A
building energy simulation tool that considers the impacts of multiple building design parameters,

ed
such as envelope heat transfer, solar heat gain, heat and mass transfer from ventilation, air-
conditioning loads, et al. There are many popular building energy simulation tools, and EnergyPlus
is one of them. EnergyPlus is a whole-building energy simulation program developed by the USA
Department of Energy [7]. Its energy simulation includes solar gains, HVAC systems, equipment

iew
loads, lighting, and occupant thermal comfort. The simulation is based on solving the heat balance
and mass balance equations in each thermal zone. There is an increasing amount of research using
EnergyPlus for building energy simulation [8, 9]. It is also widely used for building energy
simulation in optimizations. Guo et al. combined EnergyPlus and CFD to optimize the passive
ventilation mode of a medium-sized gymnasium in subtropical regions [10]. Kamal et al. used

v
EnergyPlus to optimize the strategic control and cost of thermal energy storage in buildings [11].
DesignBuilder is a building simulation software that integrates the EnergyPlus simulation engine at

re
its core while providing a visual modeling environment [12]. Compared with EnergyPlus,
DesignBuilder is more user-friendly and commonly used for commercial applications. Goenaga-
Pérez et al. used DesignBuilder to build a virtual building and evaluate multiple designs in five
climate zones [13]. Liu et al. combined DesignBuilder and a random forest model to predict the
building energy consumption with different envelope designs [14]. TRNSYS is a powerful software
er
package for dynamic energy simulations [15]. TRNSYS, which stands for Transient systems
simulations, is widely used for modeling the solar system of buildings [16]. Dezhdar et al. used
TRNSYS to optimize a renewable system with two potential wind and solar energies for electricity
pe
production, cooling, and heating [17]. ESP -r is a time-based building energy simulation software
first developed by the University of Strathclyde [18]. Strachan et al. introduced the history and
development of validation of ESP-r [19]. There are other building energy simulation tools, such as
DeST, Modelica, IESVE, and TRACE 700, for research or commercial use [20, 21].
To optimize the building energy simulation after the dataset has been generated, a surrogate
ot

model must be trained to find out the relationship between the variables and the building energy
consumption. Since the building energy consumption is determined by multiple variables, a
powerful machine learning method is required. Machine learning methods are usually used for two
tn

kinds of problems: classification and regression. Training a surrogate model to predict building
energy consumption needs a regression method. There is a lot of research in this area. Ali et al. used
Extreme Gradient Boosting to train a surrogate model for the prediction of energy consumption
under 16 setting variables [22]. The result shows that the surrogate model has a very high accuracy
rin

that the R2 value reaches 0.99. Shen et al. proposed a BIM-supported energy performance analysis
for green building design [23]. In their study, Bayesian optimization-LightGBM is used to predict
three objectives on behalf of building energy consumption, and the results show very high
accuracies. Hosamo et al. developed a system to investigate the impact of building factors on energy
ep

usage and find the optimal design [24]. In this research, the Group Least Square Support Vector
Machine is selected as the best prediction method for building energy consumption after 11 machine
learning algorithms were compared. Ye et al. developed a framework to model the relationship
between city-block-level building-oriented features and building energy consumption [25]. The
Pr

framework is based on a Random Forest model; a dataset of 881 city blocks was employed. The
framework identified 7 important features among 59 potential features in the case study. Seyedzadeh
et al. proposed a machine-learning-supported model for energy performance prediction of non-

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 domestic buildings [26]. The model was evaluated by a case study of three thousand retrofit
variations, the model achieved an error of just 1.7%.

ed
After the surrogate model is built by using regression machine learning methods, the final step is
to find the optimal design that has minimal annual building energy consumption. There is a lot of
research using optimization methods for building energy. Cong et al. applied Particle Swarm
Optimization into a thermal comfort based optimization method for building energy performance

iew
with natural ventilation [6]. The optimal design found by this method in the case study has much
better energy performance compared with commonly used building designs in hot summer and cold
winter climate zone. Chegari et al. developed a multi-objective optimization method to improve the
indoor thermal comfort and energy performance of residential buildings [27]. This method is based on
the integration of Artificial Neural Networks and coupled with Multi-Objective Particle Swarm

v
Optimization. The results show that the energy consumption for thermal needs can be reduced to 74.52%
of the total for a Moroccan ground floor + first floor house. Imran et al. proposed a loT (Internet of

re
Things) task management mechanism for energy consumption minimization in smart residential
buildings [28]. The task management mechanism is based on a predictive optimization module, and
the Particle Swarm Optimization and Genetic Algorithm are used in this module. Afroz et al.
proposed a modeling and optimization approach to minimize the energy consumption of the HVAC
systems without concession to indoor environmental quality [29]. The optimization part of this
er
approach uses the PSO algorithm. The results indicate that 7.8% of total energy can be reduced by
this approach under certain indoor environmental quality requirements.
In the former research, there are various machine learning methods used to predict building
pe
energy consumption. However, there is not much research about comparing and selecting regression
methods for surrogate modeling of building energy consumption in the design stage, and there is no
suggested simulation number of different variable combinations for reference. Moreover, there is
not much research on different combinations of regression and optimization methods to minimize
the annual building energy consumption in the design stage. To fill this gap, this study introduces a
ot

data-driven, surrogate modeling and optimization method for building energy performance in the
design stage. The method has four main steps: building model establishment, dataset generation,
surrogate modeling, and optimization.
tn

Different building energy simulation tools have different advantages: some provide many details,
some are more suitable for single-zone simulations, and some are outstanding for one part of
building energy simulation like solar heat transfer. This study needs a tool that is capable of handling
whole building simulation, which combines high accuracy and fast computation speed. EnergyPlus
rin

is the best choice for generating the dataset. In this study, six potential regression methods and five
potential optimization methods are compared and evaluated in two cases. The two cases use two
typical AC system operation strategies respectively: constantly and intermittently operating.
2. Methodology
ep

In this study, the modeling and optimization method for annual building energy consumption in
the design stage has four steps, as shown in Fig. 1.
Step 1, building model establishment. The building geometric model is built in SketchUp [30]
with Euclid plug-in [31], and the specific parameters are set in EnergyPlus [32, 33].
Pr

Step 2, dataset generation. Important building variables in the design stage are selected. A number
of subcases with different building variable combinations are simulated in EnergyPlus. The results
are collected into a dataset.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 Step 3, surrogate modeling. A regression method is selected to train a surrogate model using the
dataset from Step 2. The surrogate model is also validated.

ed
Step 4, optimization. An optimization method is selected. The surrogate model is introduced as
the objective function in the optimization method to find the optimal design.
Step 2 to Step 4 are accomplished in Python [34].
Fig. 1. Research framework

iew
Step1: Building model establishment

v
re
Step2: Dataset generation

variable 𝑥1
variable 𝑥2
variable 𝑥3
subcase
random()
er …
s
Repeat

variable 𝑥𝑚―1

variable 𝑥𝑚
pe
Dataset

Step3: Surrogate modeling

Extreme Adaptive Random

ot

Gradient Boost Forest

Light
Boost
Support Gradient
Gradient Vector Boosting
tn

Boost Regression Machine

Step4: Optimization
rin

Nelder-Mead Conjugate Gradient Differential

Method With Trust Region Evolution
Method

Limited-memory
ep

Broyden-Fletcher-Goldfarb-Shanno Method
with Bounds

Optimal design
Pr

As shown in Fig. 1, six potential regression methods and five potential optimization methods are
selected in this study. They are introduced as follows.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 2.1. Regression Methods
In this study, six potential regression methods are selected. They all have both fast speed and

ed
good accuracy in training surrogate models for annual building energy consumption.
2.1.1. Support Vector Regression (SVR)
Support Vector Regression is a typical machine learning algorithm used for regression analysis
[35, 36]. SVR identifies a hyperplane in a higher dimensional space, maximizing the margin

iew
between the hyperplane and the data points. The basic idea is to find a hyperplane f(x) with at most
ε deviation from the actual objective value y for all training data while f(x) is as flat as possible. The
function f(x) can predict the value of a dependent variable y based on a set of independent variables
x, subject to some constraints. For any data point, there is a loss function as Eq. (1).

|y - f(x)| = | 0 if |y - f(x)| ≤ 𝜀
{
y ― f(x)| ― 𝜀 otherwise 1
#( )

v
Taking the linear form

re
f(x) = w ∙ ϕ(x) + b#(2)
This problem can be seen as a convex optimization problem.
1
minimize ‖w‖2
2
subject to yi - w ∙ ϕ(x) ― b ≤ ε (3)#
w ∙ ϕ(x) + b ― yi ≤ ε
er
A soft margin is adopted by introducing slack variables to avoid the effect of abnormal data [37].
n
1
min ‖w‖2 + C ξi + ξ*i
2
pe
i=1
subject to yi ― w ∙ ϕ(x) ― b ≤ ε + ξi (4)#
w ∙ ϕ(x) + b ― yi ≤ ε + ξ*i
ξi,ξ*i ≥ 0
Where C is a penalty parameter to try to get a big ε and a small number of outliers simultaneously
and ξi, ξ*i are slack variables.
ot

This problem can be solved by constructing a Lagrange function:

n n
1
L(w, b, ξ, ξ , α, α , β, β ) = ‖w‖2 + C
* * *
ξi + ξ*i ― * *
(αξi + α ξi )
tn

2 i=1 i=1

n n

― β(yi ― w ∙ ϕ(x) ― b ― ε ― ξi) ― β*(w ∙ ϕ(x) ― b ― yi ― ε ― ξ*i) (5)

i=1 i=1
rin

Where L is the Lagrangian function, and α, α , β, β* are Lagrange multipliers. The original
*

problem becomes:
max min L(w, b, ξ, ξ*, α, α*, β, β*)
α, α , β, β w, b, ξ, ξ*
* *
ep

This problem can be solved by using Karush-Kuhn-Tucker (KKT) conditions.

A kernel function maps the data to a higher dimension space for nonlinear regression to transform
the problem into a linear regression [37]. There are different kernels, such as linear kernel,
polynomial kernel, sigmoid function kernel, and Gaussian radial basis function kernel. The
Pr

following are the kernel functions:

Linear kernel: k(xi, x) = xi ∙ x #(6)
Polynomial kernel: k(xi, x) = (γ(xi, x) + coefθ)degree#(7)
5

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 Sigmoid kernel: k(xi, x) = tanh(γ(xi, x) + coefθ)#(8)

ed
Radial basis function (RBF) kernel: k(xi, x) = exp ( - γ‖xi - x‖2) #(9)

Where 𝛾 is the kernel parameter.

2.1.2 Random Forest Regression (RFR)
The Random Forest algorithm is a popular machine-learning technique that has proven highly

iew
effective in classification and regression. It is widely used to predict building energy consumption
[38, 39]. The algorithm builds multiple decision trees and aggregates their predictions to make more
accurate predictions [40]. One of the main benefits of Random Forest is its use of randomness in
the training process. Each tree in the Random Forest is built on a random subset of the data, which
helps to reduce overfitting and increase the diversity of the trees. At each tree split, a random subset

v
of the features is considered, further increasing the diversity and robustness of the model. To make
predictions, the RFR aggregates the predictions of each tree by taking the average vote for regression,

re
which means that each tree's prediction is weighted equally in the final prediction, resulting in a
more robust and accurate overall prediction. However, to achieve the best possible performance
from RFR, it is crucial to identify the optimal values for two key parameters: the number of variables
chosen in each tree node and the total number of trees in the model. These parameters can
significantly impact the model's performance and must be carefully tuned to ensure the best results.
There are two basic steps in RFR:
er
(1) Bagging: Bagging is a technique for randomly selecting training samples with replacements
from an original sample D that contains p feature variables. The resulting training sample sets, n,
pe
are used to build the model, while the out-of-bag (OOB) dataset comprises elements not included
in n and is utilized as the test sample set.
(2) Random Feature selection: In the RFR model, k feature variables (k < p) are randomly chosen
from p variables as branching nodes of the classification tree in each training sample set. The
impurity of the variables is measured using the Gini coefficient, and an unpruned regression tree is
ot

constructed for each training sample. The resulting n regression trees make up the RFR model, and
the predicted response value for the OOB data is averaged over all trees (NTree).
2.1.3 Gradient Boosting (GB)
tn

Gradient Boosting is a widely used machine learning algorithm that can tackle both regression
and classification tasks [41, 42]. It is an ensemble technique that combines multiple weak learners
to produce a more accurate prediction [43]. The fundamental concept behind GB is to train a series
of weak learners on the residuals (errors) of the previous model iteratively, aiming to minimize the
rin

final model's overall loss function, as shown in Fig. 2. In simpler terms, each new model in the
sequence is designed to rectify the errors of the preceding models, with the expectation that the
combined effect will yield a more accurate prediction model. During training, the model's gradient
(the partial derivative of the loss function) concerning the training data is calculated by GB. The
ep

gradient is employed to update the model to reduce the error using techniques such as gradient
descent. By continuously refining the model, GB can create an exact prediction model that can
handle complicated datasets and nonlinear relationships between features.
There are different models that GB can be taken as based learners, such as Decision Tree, Linear
Pr

Regression, and Neural Network. The decision tree is the most typical learner in GB for regression
problems. The Decision Tree learner can capture nonlinear relationships between the features and
the target variable. For large datasets, the Decision Tree learner has good efficiency. The Decision

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 tree learner can handle a mixture of categorical and numerical features. Due to these advantages,
decision-tree-based GB is a good choice for optimizing building energy consumption.

ed
Fig. 2. Schematic of decision-tree-based gradient boosting
2.1.4 Extreme Gradient Boosting (XGBoost)
y h
1 h
n-1

iew
X X
X

Tree 1 Tree 2
Tree n
…

v
y h h
n-

re
1
X X 1 X

h = y – f (X) h = h – f (X)
h =h - f (X)
1 1 2 1
er 2
n

Extreme Gradient Boosting is an improved algorithm based on gradient boosting [44]. XGBoost
n-1 n

uses L1 and L2 regularization to prevent overfitting problems. XGBoost simplifies the decision
pe
trees to improve the efficiency. XGBoost is much faster due to the convenience of parallel
computation. It is also very flexible compared with gradient boosting, such as having more
acceptable input data forms and supporting user-defined objective functions. XGBoost performs
better in dealing with large-scale data than traditional gradient boosting [45].
2.1.5 Light Gradient Boosting Machine (LightGBM)
ot

Light Gradient Boosting Machine is another improved algorithm based on gradient boosting [46].
LightGBM has an outstanding training speed by using histogram-based algorithms. LightGBM uses
discrete bins instead of continuous values, which saves much memory usage. It uses one-side
tn

sampling to select a small subset of the training data that is highly informative for splitting, which
reduces the cost of calculating the gain for each split. Same as XGBoost, LightGBM also supports
parallel computation. LightGBM has excellent accuracy with large-scale data. Guo et al. used
LightGBM to predict heating and cooling loads [47]. The coefficient of determination (R2) of results
rin

reaches 0.9981.
2.1.6 Adaptive Boosting (AdaBoost)
Adaptive Boosting is an algorithm that combines multiple weak learners into a strong learner
[48]. It works by assigning weights to each training example and then iteratively training a weak
ep

learner on the weighted examples. In each iteration, the weights of the poorly predicted data points
are increased so that the next weak learner focuses more on these points. The final learner is a
weighted sum of all the weak learners. As an iterative algorithm, AdaBoost works well when the
data is noisy or contains outliers, and it can handle a wide range of models as base learners [49].
2.2. Optimization Methods
Pr

In this study, five potential optimization methods are selected. They all have fast speed in
optimizing annual building energy consumption.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 point B point B
worst point A point B

ed
worst point A worst point A Point B’
reflected point D contacted point F reflected point D
point A’
best point C best point C best point C
expanded point E reflected point D

iew
(1) Expansion (2) Contraction (3) Shrinkage

2.2.1 Particle Swarm Optimization (PSO)

Particle Swarm Optimization is an evolutionary computation algorithm [50]. The main idea is
that every particle of the swarm shares the information with others at each timestep to receive the
pbest and gbest values, where pbest is the best-ever objective value of this particle, gbest is the best-

v
ever objective value of all particles. Each individual particle adjusts its speed and direction based
on the previous timestep. It's necessary to introduce random numbers to prevent the swarm from

re
converging to local optimal solutions. For particle i,
vi,t+1 = ωvi,t + c1r1(pbesti,t - zi,t) + c2r2(gbestt - zi,t)#(10)
zi,t+1 = zi,t + vi,t+1#(11)
Where vi,t+1, zi,t+1, vi,t, zi,t are the speed and location of particle i at timestep t+1 and timestep t, ω
er
is momentum, it shows how much the particle’s speed is influenced by the speed of last timestep,
pbesti,t is the best ever objective value of particle i at timestep t, and gbestt is the best-ever objective
value of all particles in the swarm at timestep t, c1 and c2 are the learning factors that show how
pe
much the particle’s speed is influenced by pbest and gbest of last timestep, and r1, r2 are random
numbers from 0 to 1.zi,t+1, the location of particle i at timestep t+1, can be calculated by using Eq.
(10) and Eq. (11). The iteration calculation stops when the difference between gbestt and gbestt+1 is
small enough, or it reaches the maximum of iterations. PSO is a simple and efficient optimization
method with a fast convergence speed [51].
ot

2.2.2 Differential Evolution (DE)

Differential Evolution is a popular metaheuristic optimization algorithm used to solve
optimization problems. It was proposed by Storn and Price in 1997 and has been applied to various
problems in various fields [52]. The basic idea of DE is to maintain a population of candidate
tn

solutions and iteratively improve the population by applying three simple operators: mutation,
crossover, and selection. The mutation operator generates new candidate solutions by perturbing
existing solutions in the population. The crossover operator combines two solutions to generate a
new solution, and the selection operator determines which solutions should be retained for the next
rin

iteration. DE is a simple and powerful algorithm for optimization problems. DE can effectively
search through the search space and converge towards the optimal solution.
2.2.3 Nelder-Mead (NM) method
The Nelder-Mead method is a popular heuristic optimization algorithm used to find the minimum
ep

of a nonlinear function without requiring gradient information. It was proposed by J. A. Nelder and
R. Mead in 1965 and has been widely used [53]. The basic idea of the NM method is to maintain a
simplex, a geometric figure consisting of n+1 points in an n-dimensional space. The simplex is
iteratively modified by four operations: reflection, expansion, contraction, and shrinkage. When the
Pr

NM method is applied to searching an optimization in two-dimensional space, the simplex is a

triangle, and this can be easily demonstrated by graphs, as shown in Fig. 3.
Fig. 3. Three operations of Nelder-Mead method
8

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 In each iteration, the simplex has three points: A, B, and C. Point A has the worst performance
among the three points; point C has the best performance. To optimize the simplex, first is to find a

ed
reflection point of point A, which is point D in Fig. 3. Point D is obtained by connecting point A
and the midpoint of BC and extending along this direction by certain times (usually one time, as
shown in Fig. 3).
(1) Expansion: If point D performs better than point C, the optimizing direction is correct, and it is

iew
possible to get a better performance point with further distance along this direction. In this case,
point E can be obtained after expansion. After comparing the performance of point D and point
E, the point with better performance will replace point A and form a new simplex with point B
and point C.
(2) Contraction: If point D performs worse than point B, it means that point D went too far from

v
point A. In this situation, point F is found on the line segment from point A to the midpoint of
BC (normally the midpoint). If point F performs better than point A, point F will form a new

re
simplex with points B and C. Otherwise, the Shrinkage step will be taken.
(3) Shrinkage: In this case, the entire simplex will be scaled down. The best performance point,
point C, will be kept. Usually, point A' is the midpoint of AC, and point B' is the midpoint of
BC.
The best optimization is found after the maximum iteration number is reached or the simplex is
er
small enough. The NM method is a robust optimization algorithm that can handle many functions
without requiring gradient information. The NM method can effectively search the parameter space
and converge toward the optimal solution by iteratively modifying the simplex based on the function
pe
values at the vertices.
2.2.5 Conjugate Gradient with Trust Region (CG-TR) method
The Conjugate Gradient method is a kind of conjugate direction method. It is an iterative
algorithm that can solve unconstrained optimization problems. The CG method requires derivatives.
When the function is quadratic, the CG method is more efficient than the conjugate direction method.
ot

It was first proposed by Magnus Hestenes and Eduard Stiefel in 1952 and has become a popular
method for solving large-scale optimization problems in many fields [54].
The basic idea of the CG method is to search for the minimum of a quadratic function by
tn

iteratively finding a sequence of conjugate directions. A direction is considered conjugate if it is

orthogonal to all the previous directions in the sequence. The algorithm starts with an initial point
and a search direction and updates the point and the direction at each iteration. The updating formula
of each iteration can be generalized as follows:
rin

rTtrt
αt = #(12)
dTtAdt
xt+1 = xt + αtdt#(13)
rt+1 = rt - αtAdt#(14)
ep

rTt+1rt+1
βt = #(15)
rTtrt
dt+1 = rt+1 + βtdt#(16)
Where t is the step (iteration) number, 𝑟𝑡 is residual at t step, 𝛼t is α coefficient at t step, 𝛽𝑡
Pr

is 𝛽 coefficient at t step, 𝑑𝑡 is the direction vector at t step. The algorithm can converge towards
the optimal solution efficiently by ensuring that the search directions are conjugate.
To search for the solution within boundaries, we combine Conjugate Gradient with Trust Region.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 The Trust Region method is a valuable tool for solving constrained optimization problems [55]. The
Trust Region will check every step to make sure the boundaries are not violated. If the step would

ed
cause a boundary violation, the Trust Region method adjusts the step size or direction to keep the
solution within the boundaries.
2.2.6 Limited-memory Broyden-Fletcher-Goldfarb-Shanno with Bounds (L-BFGS-B) algorithm
The Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm is a popular iterative algorithm for

iew
solving unconstrained optimization problems [56-59]. It is a quasi-Newton method. The Newton
method operates through the iterative determination of intersections between the tangent line,
representing the function's derivative at a certain point (𝑥𝑘, f(𝑥𝑘)), and the x-axis. The intersection's
x-coordinate value will be updated to 𝑥𝑘. Subsequently, these updated 𝑥𝑘 values are refined to
approximate the solution of the function. In mathematics, the method becomes solving the following

v
equation.
f' (xk) + f'' (xk)(x - xk) = 0, k = 0, 1, 2 …#(17)

re
For a multi-factor optimization, 𝒙𝑘 is a multi-dimension vector, which makes the derivatives
more complex. The second-order derivative of the function is a matrix called the Hessian matrix.
Obtaining the Hessian matrix 𝑯𝑘 is difficult when x is relevant to many dimensions. BFGS uses a
positive definite matrix 𝐵𝑘 to replace the Hessian Matrix at any point while:
Bk ≈ Hk = ∇2f (xk)#(18)
er
If the exact line search or Wolfe condition search is used in BFGS, 𝐵𝑘 is a symmetric positive
definite matrix.
pe
Bksk(sk)TBk yk(yk)T
Bk+1 = Bk - + #(19)
(sk)TBksk (yk)Tsk

Where sk is the displacement, sk = xk+1 ― xk, 𝑦𝑘 is the gradient difference, yk = ∇f(x + 1) - ∇f

(x).
If the Armijo condition search is used in BFGS to guarantee that 𝐵𝑘 is a symmetric positive definite
ot

matrix:

{
Bk, (yk)Tsk ≤ 0
Bk+1 = Bksk(sk)TBk yk(yk)T (20)
tn

Bk - + , (yk)Tsk > 0 ##
(sk)TBksk (yk)Tsk

BFGS is an optimization algorithm designed for primarily designed for unconstrained

optimization problems. Limited-memory Broyden-Fletcher-Goldfarb-Shanno with Bounds (L-
rin

BFGS-B) algorithm is an extension of BFGS, which is suitable for optimization problems with
boundaries [60]. L-BFGS-B adjusts step sizes and modifies the search direction by the constraints
to ensure that the solution remains within the bounds. The L-BFGS-B algorithm ensures that the
approximation remains positively definite at each iteration, ensuring that the algorithm converges
ep

to a local minimum of the objective function. The L-BFGS-B algorithm is fast and efficient for
solving unconstrained optimization problems.
3. Case study
The case study includes two cases; two cases use the same building but with different AC
Pr

operation strategies. The case building is located in Changsha, China. Changsha has a typical hot
summer and cold winter climate. In summer, temperatures vary between 27℃ and 39℃, with
humidity ranging from 80% to 99%, while in winter, temperatures can drop to -5℃ or reach 10℃,

10

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 with humidity levels ranging from 50% to 90%. In China, many big cities are located in the HSCW
climate zone, with enormous energy consumption. This study can provide a design solution that

ed
helps significantly reduce building energy consumption. Changsha's Chinese standard weather data
was used as the input weather data file in EnergyPlus; file name and download link are given in
[61]. The building energy consumption is simulated in EnergyPlus, while other calculations in this
method are all programmed and run in Python.

v iew
re
er
Fig. 4. The case building
3.1. Building modeling
The geometric model of the case building was built in SketchUp with the Euclid plug-in, as shown
pe
in Fig. 4. The case building has seven floors and 2016 m2 total area. Each floor contains six rooms
and one corridor. Each room has one exterior window, and the corridor has two exterior windows.
All exterior windows have overhangs right above them. The height of each floor is 3.5 m. The size
of all six rooms is 8 m×5 m×3.5 m, and the width of the room windows is 4m. The corridor is 24
m×2 m×3.5 m, and the two corridor windows both have 1.2 m width and 1.8m height. Except for
ot

the corridor windows, the height of all other windows is changeable to get different window wall
ratios, while the width of all windows remains constant. In this building, except for the first and top
floors, the other five floors in the middle have repetitive structures and similar thermal boundaries.
tn

The zonegroup object in EnergyPlus was used to reduce the amount of input parameters. Every
envelope has the boundary conditions the same as shown in Fig. 4. Because of the use of zonegroup,
the outside boundary condition of the top story floor is the ceiling of the middle story, and vice
versa; the outside boundary condition of the entry-level story ceiling is the floor of the middle story,
rin

and vice versa. This building is designed according to the General Code of Energy Efficiency and
Renewable Energy Application in Buildings, GB 55015-2021 [62]. Envelope designs are listed in
Table 1. The AC system is a VAV system. The heating energy source is a boiler with 80% efficiency
using natural gas. The cooling system has an electric reciprocating Chiller whose nominal
ep

coefficient of performance (COP) is 3.2. In the case with a constantly operating AC system, the case
building is assumed to be a senior care facility. The lights are designed as 9 W/m2 according to
GB50034-2019 [63]. According to JGJ450-2018, the occupant density is 12 m2 per person, and the
electric equipment is set to 16 W/m2 [64]. In the case with an intermittently operating AC system,
Pr

the case building is assumed to be an office building. The lights are designed as 10 W/m2 according
to GB50034-2019 [63]. According to JGJ/T67-2019, the occupant density is 9 m2 per person, and
the electric equipment is set to 14.5 W/m2 [65].

11

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 Table 1
Envelope materials and parameters

ed
Layers Thickness Conductivity Density Heat capacity
Construction
(from exterior to interior) (m) (W/(m ∙ K)) (kg/m3) (kJ/(kg ∙ K))

Cement mortar 0.02 0.93 1800 0.837

iew
Brick wall 0.24 0.81 1800 0.879
Shale hollow brick 0.24 0.58 1400 1.05
Exterior wall
Aerated concrete 0.1 0.18 700 1.05
EPS insulation — 0.039 20 1.38
Lime mortar 0.02 0.814 1600 0.837
Gravelly sand 0.005 2.04 2400 0.92

v
Heavyweight concrete 0.2 1.95 2240 0.9
Roof
Cement expanded perlite 0.025 0.16 400 1.17

re
Lime mortar 0.02 0.814 1600 0.837
Gypsum board 0.019 0.33 1050 1.05
Interior
Heavyweight concrete 0.1 1.95 2240 0.9
ceiling
Gypsum board er0.019 0.33 1050 1.05
Clear glass 0.006 0.9 2500 0.84
Exterior Glass with lowE coating 0.006 0.9 2500 —
window Air 0.012 0.0267 1.29 1.005
pe
Argon gas 0.012 0.0173 1.784 0.52
Blinds 0.001 0.1 — —
3.2. Variables selection
Both the two cases have six relevant variables, as shown in Table 2. Exterior wall masonry and
exterior window structure are explained in Table 3. x1 refers to the angles that the building has
ot

rotated clockwise from the position that the building entrance is facing due south. When x1 is 0°, the
building entrance is facing due south. Since this building is nearly symmetrical, the meaningful
range is from 0° to 90°. x3 refers to the window wall ratio of the exterior building wall that has the
tn

building entrance and its opposite building wall. All the windows on these two walls have the same
width and height, the width remains the same, x3 changes with the window height.
Table 2
Variables’ ranges and intervals
rin

Variable Range Interval

Building orientation, x1 (°) 0-90 15
Exterior wall masonry, x2 #M1, #M2, #M3 —
17%-31%
ep

Window wall ratio, x3 Window height 0.1 m

(window height 1.2 m-2.2 m)
Shading overhang depth, x4 (m) 0.6 -1.5 0.1
Interior blinds slat angle, x5 (°) 15-90 15
Exterior window structures, x6 #W1,#W2, #W3, #W4 —
Pr

Table 3
Construction of exterior wall and window
Exterior wall Exterior window U-value (W/(m2 ∙
12

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 K))
M1: Cement mortar + Brick wall

ed
W1: Clear glass + Air + Clear glass 3.38
+ EPS insulation + Lime mortar
M2: Cement mortar + Shale hollow brick W2: Clear glass + Argon gas + Clear
3.19
+ EPS insulation + Lime mortar glass
M3: Cement mortar + Aerated concrete W3: Glass with lowE coating + Air +

iew
3.24
+ EPS insulation + Lime mortar Clear glass
W4: Glass with lowE coating + Argon
— 3.0
gas + Clear glass
Subcases are generated by using the random function to get the variable values.
3.3 Constantly operating AC system

v
Some buildings, such as hospitals, healthcare facilities, data centers, laboratories, and research
facilities, need constantly operating AC systems. In this case, the case building is assumed to be a

re
senior care facility. Five thousand subcases were simulated in this case study and the results are
collected into a dataset. Eighty percent of the data is used to train surrogate models and the rest is
used to validate the surrogate models. Fig. 5 shows the regression results by different regression
methods with different subcase numbers. The variables are shown in Table 2; the objective value is
er
annual building energy consumption. The results are shown by R-squared value (poorest value: 0,
best value: 1).
pe
ot
tn
rin

Fig. 5. The R-squared values of different regression methods with different subcase numbers
AdaBoost has a relatively low accuracy compared with the other five regression methods. The R-
squared value does not keep an upward trend with the subcase number increase. The other five
regression methods always have R-squared values bigger than 0.98. The R-squared value keeps
ep

increasing as the subcase number grows. When the subcase number is 2500, all five regression
methods have an R-squared value bigger than 0.99. Among all regression methods, RFR performs
best for all subcase numbers. When the subcase number is 5000, the R-squared value of RFR reaches
0.99847. RFR can predict the annual building energy consumption with a constantly operating AC
Pr

system with very high accuracy.

According to the above results, it is suggested to simulate 2000 to 3500 subcases for one case.
Here, we choose the subcase number of 2500. With this subcase number, most cases' time cost and

13

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 regression fitness are acceptable. The R-squared values using different regression methods with
2500 subcases are shown in Table 4. All regression methods have good performance, with an R-

ed
square value higher than 0.89. The RFR has the best performance; the R-squared value is as high as
0.997. Among all the regression methods in Table 4, AdaBoost has the poorest performance, with
an R-squared value of 0.897.
Table 4

iew
The R-squared values using different regression methods (2500 subcases)
Regression Methods SVR RFR GB XGBoost LightGBM AdaBoost
R-squared Value 0.994 0.997 0.991 0.992 0.993 0.897
After applying the trained surrogate models (2500 subcases) to different optimization methods,
the results differ with different combinations, as shown in Table 5. The optimization is required to

v
find the optimal design in a limited time; here, the limitation is 5 minutes. When the computing time
is more than 5 minutes, it is considered that the optimization method has failed to converge. The

re
variable values of minimum annual building energy consumption found by XGBoost & DE are
shown in Table 6.
Table 5
The optimization results using different methods (2500 subcases)
er
Minimum annual building energy consumption (GJ)
Optimization
Regression methods
Methods
SVR RFR GB XGBoost LightGBM AdaBoost
pe
PSO 1063 - 1061 1061 1063 1095
DE 1063 1061 1061 1060 1070 1095
NM 1067 1131 1135 1143 1129 1139
CG-TR 1064 1096 1093 1103 1087 1104
L-BFGS-B 1063 1139 1136 1143 1129 1140
ot

*"-" means the optimal design is not found in 5 minutes.

Table 6
The minimum annual building energy consumption (y) and the corresponding variables (x1 to x6)
tn

Method y x1 x2 x3 x4 x5 x6
XGBoost & DE 1060 GJ 4° M2 18% 1.0 m 17° W4
The results in Table 5 show that PSO method works well with the surrogate models trained by
SVR, GB, XGBoost, and LightGBM. With the surrogate models trained by RFR, PSO failed to find
rin

an optimal design in 5 minutes. DE method works well with the surrogate models trained by SVR,
RFR, GB, and XGBoost. The NM method works well only with the surrogate model trained by SVR,
and so do CC-TR and L- BFGS-B. All optimization methods work well with the surrogate model
trained by SVR. With the surrogate model trained by AdaBoost, the optimal designs found by all
ep

five optimization methods have relatively big annual building energy consumption. Of all
combinations, the combination of XGBoost and DE found the best design with the minimal annual
building energy consumption of 1060 GJ. All optimization methods work well with the surrogate
model trained by SVR, and all the combinations found the designs with very close annual building
Pr

energy consumptions. There are many other combinations found the designs with very close annual
building energy consumptions, such as RFR & DE, GB & PSO, and GB & DE.

14

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 3.4. Intermittently operating AC system
In many buildings, the AC system is not always required to operate. For example, the office

ed
building, the museum, or the shopping center are mainly used during the daytime. At night, the
building has no occupants, and the AC system is off. In this case, the case building is assumed to be
an office building. A common AC operation strategy is used: the AC system is on from 8 a.m. to 6
p.m. from Monday to Friday; the other time the AC system is off. All the windows stay closed at

iew
night. Five thousand subcases are simulated and the results are collected into a dataset. Eighty
percent of the data is used to train surrogate models and the rest is used to validate the surrogate
models. The variables are shown in Table 2; the objective value is annual building energy
consumption. Fig. 6 shows how the R-squared value changes with the subcase number while the
windows are closed at night.

v
re
er
pe

Fig. 6. The R-squared values of different regression methods with different subcase numbers
ot

AdaBoost has a lower accuracy compared with the other five regression methods. The R-squared
value decreases with the subcase number increase when the subcase number is not bigger than 2500.
AdaBoost has the biggest R-squared value when the subcase number is 1000. The R-squared value
of AdaBoost has never been bigger than 0.9 when the subcase number is not bigger than 5000. The
tn

other five regression methods have an R-squared value bigger than 0.94 all the time. When the
subcase number increases from 1000 to 1500, the R-squared values of all five regression methods
grow. However, the R-squared value doesn't keep increasing with the subcase number when the
subcase number is bigger than 1500. The R-squared values of SVR, RFR, GB, and LightGBM do
rin

not change much when the subcase number exceeds 1500. The R-squared value of XGBoost does
not change much when the subcase number is bigger than 2000. Among all regression methods,
RFR has the best performance with all different subcase numbers; its R-squared value is always
bigger than 0.99. RFR can predict annual building energy consumption with intermittently operating
ep

AC systems with very high accuracy.

According to the above results, it is suggested to simulate 2000 to 3500 subcases with SVR, RFR,
GB, XGBoost, or LightGBM for one case. Most cases' time cost and regression fitness are
acceptable when the subcase number is 2500. It is not suggested to use AdaBoost for annual building
Pr

energy consumption with intermittently operating AC systems. Table 7 shows the R-squared values
using different regression methods with 2500 subcases (Windows are closed at night).
Table 7
15

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 The R-squared values using different regression methods (2500 subcases)
Regression Methods SVR RFR GB XGBoost LightGBM AdaBoost

ed
R-squared Value 0.975 0.994 0.967 0.959 0.967 0.864
Among all the regression methods, the AdaBoost method has the poorest performance, with an
R-squared value of 0.864. All the other five methods have very good performance, with an R-square
value higher than 0.95. The RFR has the best performance; the R-squared value is as high as 0.994.

iew
After applying the trained surrogate models to different optimization methods, the results differ with
different combinations, as shown in Table 8. The optimization is required to find the optimal design
in a limited time; here, the time limitation is 5 minutes. It is considered that the optimization method
has failed to converge when the computing time is more than 5 minutes. The variable values of
minimum annual building energy consumption found by GB & PSO are shown in Table 9.

v
Table 8
The optimization results using different methods (2500 subcases).

re
Minimum annual building energy consumption (GJ)
Optimization
Regression methods
Methods
SVR RFR GB XGBoost LightGBM AdaBoost
PSO 567 -
er 565 566 567 579
DE 567 567 566 566 570 579
NM 567 592 592 595 592 593
CG-TR 567 577 578 581 575 584
pe
L-BFGS-B 567 592 593 595 592 593
*"-" means the optimal design is not found in 5 minutes.
Table 9
The minimum annual building energy consumption (y) and the corresponding variables (x1 to x6)
Method y x1 x2 x3 x4 x5 x6
ot

GB & PSO 565 GJ 0° M2 17% 1.3 m 15° W4

Table 8 shows that most combinations of regression and optimization methods can find the
optimal design in a limited time. PSO failed to find the optimal design with surrogate models trained
tn

by RFR in a limited time. PSO found the best design with the surrogate model trained by GB, which
has the minimal annual building energy consumption of 565GJ. There are many other combinations
found the designs with very close annual building energy consumptions, such as RFR & DE, GB &
DE, and XGBoost & PSO. All optimization methods work well with the surrogate model trained by
rin

SVR, and all the combinations found the optimal designs with annual building energy consumptions
of 567 GJ. PSO method works well with the surrogate models trained by SVR, GB, XGBoost, and
LightGBM. DE method doesn't work well with the surrogate models trained by AdaBoost, but
works well with the other five regression methods. The NM, CC-TR, and L- BFGS-B methods work
ep

well only with the surrogate model trained by SVR.

4. Discussion
The results from the two cases show the very same trends. Both cases show that RFR is an
outstanding regression method for the prediction of annual building energy consumption, while
Pr

AdaBoost has a relatively low accuracy. It is suggested to simulate 2000 to 3500 subcases with
SVR, RFR, GB, XGBoost, or LightGBM for one case. All optimization methods work well with
the surrogate model trained by SVR in both cases. In the case with a constantly operating AC system,
16

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 the combination of XGBoost and DE found the best design with the minimal annual building energy
consumption of 1060 GJ; In the case with an intermittently operating AC system, the combination

ed
of GB and PSO found the best design with the minimal annual building energy consumption of 565
GJ. Many other combinations of regression and optimization methods are able to find optimal
designs with very close annual building energy consumption. AdaBoost has a poor performance
compared with other regression methods. It is not suggested to use AdaBoost to train the surrogate

iew
model of annual building energy consumption. The combination of RFR and PSO failed to find the
optimal design in a limited time in both cases. It is not suggested to use this combination to optimize
building designs.
There are several possible reasons why different combinations of regression methods and
optimization methods can show very different results: (a) Algorithmic differences: Methods differ

v
in their ability to handle different data types, landscapes, noise, and uncertainties. (b) Model
assumptions: Regression training methods make different data and variable relationship

re
assumptions, resulting in different optimization effectiveness. (c) Sensitivity to noise and outliers:
different methods may have different sensitivities to data noise and outliers. (d) Exploration-
exploitation trade-off: Optimization methods vary in their abilities to balance search space
exploration and exploitation of promising regions. Some methods have a higher risk of getting stuck
in local optima. To find the optimal building design with the method proposed by this study, it's
er
suggested to follow the steps in Fig. 1: building model establishment, dataset generation, surrogate
modeling, and optimization. For the steps of surrogate modeling and optimization, it’s suggested to
use the combinations of regression and optimization methods that perform well in Table 5 and Table
pe
8.
5. Conclusion
This study introduces a data-driven, modeling and optimization method for building energy
performance in the design stage. This method has four steps: building model establishment, dataset
generation, surrogate modeling, and optimization. This study selects six potential regression
ot

methods (Support Vector Regression, Random Forest Regression, Gradient Boosting, Extreme
Gradient Boosting, Light Gradient Boosting, and Adaptive Boosting) to train surrogate models; and
five potential optimization methods (Particle Swarm Optimization, Differential Evolution, Nelder-
tn

Mead method, Conjugate Gradient with Trust Region method, and Limited-memory Broyden-
Fletcher-Goldfarb-Shanno with Bounds algorithm) to find the optimal designs with the minimal
annual building energy consumption. These methods are applied to two cases with different AC
systems operation strategies. The two cases use the same seven-floor building under the same
rin

weather conditions in the HSCW climate zone. One case has a constantly operating AC system, and
the other case has an intermittently operating AC system.
The results from the two cases have the very same trends. For both cases, the results show that
RFR has the best performance among the six regression methods, while AdaBoost has the poorest
ep

performance. Except for AdaBoost, the other five regression methods can achieve very good
accuracies when the subcase number is above 2000. 2500 is the suggested subcase number with
good computation speed and accuracy. Different combinations of regression and optimization
methods get different results. In the case with a constantly operating AC system, the combination
Pr

of XGBoost and DE found the best design with the minimal annual building energy consumption
of 1060 GJ; In the case with an intermittently operating AC system, the combination of GB and
PSO found the best design with the minimal annual building energy consumption of 565 GJ. There

17

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 are many other combinations of regression and optimization methods that are able to find optimal
designs with very close annual building energy consumption. All optimization methods work well

ed
with the surrogate model trained by SVR in both cases. The combinations of RFR and PSO failed
to find the optimal design in a limited time in both cases. For a specific case, it is suggested to try
different combinations of regression and optimization methods that perform well in this study.
The main findings of this study could be a valuable reference for energy-saving building design.

iew
However, there are still some limitations in this study that can be investigated in future works. For
example, this method can be applied to buildings located in different climate zones. This method
can also be applied to the optimization of a specific part of the building energy consumption, such
as the energy consumption for indoor thermal comfort.
Acknowledgement

v
The authors would like to express their gratitude to the National Natural Science Foundation of
China (No.52130802).

re
References
[1] European Communities, EU energy and transport in figures, Luxembourg (2009) 228,
https://op.europa.eu/en/publication-detail/-/publication/6d938db3-7fe0-4a09-89f9-37377c0500d0.
[2] EIA, 2020 Residential Energy Consumption Survey (RECS) Survey Data,
https://www.eia.gov/consumption/residential/data/2020/.
er
[3] Eurostat, Energy consumption in households (2021), https://ec.europa.eu/eurostat/statistics-
explained/index.php?title=Energy_consumption_in_households.
[4] S. C. Zhang, Y. J. Fu, X. Y. Yang, W. Xu, Assessment of mid-to-long term energy saving
pe
impacts of nearly zero energy building incentive policies in cold region of China, Energy Build. 241
(2021) 110938, https://doi.org/10.1016/j.enbuild.2021.110938.
[5] R. F. De Masi, V. Festa, A. Gigante, S. Ruggiero, G. P. Vanoli, The incidence of smart windows
in building energy saving and future climate projections, Energy Rep. 8 (16) (2022) 283-289,
https://doi.org/10.1016/j.egyr.2022.10.222.
ot

[6] C. Li, Y. M. Chen, A multi-factor optimization method based on thermal comfort for building
energy performance with natural ventilation, Energy Build. 285 (2023) 112893,
https://doi.org/10.1016/j.enbuild.2023.112893
tn

[7] Department of Energy, EnergyPlus, https://energyplus.net/.

[8] K. A. Barber, M. Krarti, A review of optimization based tools for design and control of building
energy systems, Renew. Sustain. Energy Rev. 160 (2022) 112359,
https://doi.org/10.1016/j.rser.2022.112359.
rin

[9] T. L. Garwood, B. R. Hughes, M. R. Oates, D. O’Connor, R. Hughes, A review of energy

simulation tools for the manufacturing sector, Renew. Sustain. Energy Rev. 81(1) (2018) 895-911,
https://doi.org/10.1016/j.rser.2017.08.063.
[10] W. H. Guo, S. Liang, Y. W. He, W. W. Li, B. Xiong, H. Y. Wen, Combining EnergyPlus and
ep

CFD to predict and optimize the passive ventilation mode of medium-sized gymnasium in
subtropical regions, Build Environ. 207 (A) (2022) 108420,
https://doi.org/10.1016/j.buildenv.2021.108420.
[11] R. Kamal, F. Moloney, C. Wickramaratne, A. Narasimhan, D. Y. Goswami, Strategic control
Pr

and cost optimization of thermal energy storage in buildings using EnergyPlus, Appl. Energy 246
(2019) 77-90, https://doi.org/10.1016/j.apenergy.2019.04.017.
[12] DesignBuilder Software Ltd. DesignBuilder, https://www.designbuilder.co.uk/.

18

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 [13] A. Goenaga-Pérez, M. Álvarez-Sanz, J. Terés-Zubiaga, A. Campos-Celador, Cost-
effectiveness and minimum requirements of nZEB for residential buildings under the new Spanish

ed
Technical Building Code, Energy Build. 287 (2023) 112986,
https://doi.org/10.1016/j.enbuild.2023.112986.
[14] Y. Liu, H. Y. Chen, L. M. Zhang, Z. B. Feng, Enhancing building energy efficiency using a
random forest model: A hybrid prediction approach, Energy Rep. 7 (2021) 5003-5012,

iew
https://doi.org/10.1016/j.egyr.2021.07.135.
[15] University of Wisconsin-Madison, TRNSYS, http://www.trnsys.com/.
[16] R. L. Shrivastava, V. Kumar, S. P. Untawale, Modeling and simulation of solar water heater:
A TRNSYS perspective, Renew. Sustain. Energy Rev. 67 (2017) 126-143,
https://doi.org/10.1016/j.rser.2016.09.005.

v
[17] A. Dezhdar, E. Assareh, N. Agarwal, A. Bedakhanian, S. Keykhah, G. Y. Fard, N. Zadsar, M.
Aghajari, M. Lee, Transient optimization of a new solar-wind multi-generation system for hydrogen

re
production, desalination, clean electricity, heating, cooling, and energy storage using TRNSYS,
Renewable Energy 208 (2023) 512-537, https://doi.org/10.1016/j.renene.2023.03.019.
[18] W. A. Beckman, L. Broman, A. Fiksel, S. A. Klein, E. Lindberg, M. Schuler, J. Thornton,
TRNSYS the most complete solar energy system modeling and simulation software, Renewable
Energy 5(1-4) (1994) 486–488. https://doi.org/10.1016/0960-1481(94)90420-0.
er
[19] P. A. Strachan, G. Kokogiannakis, I. A. Macdonald, History and development of validation
with the ESP-r simulation program, Build Environ. 43 (4) (2008) 601-609,
https://doi.org/10.1016/j.buildenv.2006.06.025.
pe
[20] Y. X. Chen, C. H. Yang, X. Pan, D. Yan, Design and operation optimization of multi-chiller
plants based on energy performance simulation, Energy Build. 222 (2020) 110100,
https://doi.org/10.1016/j.enbuild.2020.110100.
[21] K. A. Barber, M. Krarti, A review of optimization based tools for design and control of building
energy systems, Renew. Sustain. Energy Rev. 160 (2022) 112359,
ot

https://doi.org/10.1016/j.rser.2022.112359.
[22] A. Ali, R. Jayaraman, A. Mayyas, B. Alaifan, E. Azar, Machine learning as a surrogate to
building performance simulation: Predicting energy consumption under different operational
tn

settings, Energy Build. 286 (2023) 112940, https://doi.org/10.1016/j.enbuild.2023.112940.

[23] Y. X. Shen, Y. Pan, BIM-supported automatic energy performance analysis for green building
design using explainable machine learning and multi-objective optimization, Appl. Energy 333
(2023) 120575, https://doi.org/10.1016/j.apenergy.2022.120575.
rin

[24] H. H. Hosamo, M. S. Tingstveit, H. K. Nielsen, P. R. Svennevig, K. Svidt, Multiobjective

optimization of building energy consumption and thermal comfort based on integrated BIM
framework with machine learning-NSGA II, Energy Build. 277 (2022) 112479,
https://doi.org/10.1016/j.enbuild.2022.112479.
ep

[25] Z. N. Ye, K. Cheng, S. C. Hsu, H. H. Wei, C. M. Cheung, Identifying critical building-oriented

features in city-block-level building energy consumption: A data-driven machine learning approach,
Appl. Energy 301 (2021) 117453, https://doi.org/10.1016/j.apenergy.2021.117453.
[26] S. Seyedzadeh, F. P. Rahimian, S. Oliver, S. Rodriguez, I. Glesk, Machine learning modelling
Pr

for predicting non-domestic buildings energy performance: A model to support deep energy retrofit
decision-making, Appl. Energy 279 (2020) 115908, https://doi.org/10.1016/j.apenergy.2020.115908.
[27] B. Chegari, M. Tabaa, E. Simeu, F. Moutaouakkil, H. Medromi, Multi-objective optimization of

19

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 building energy performance and indoor thermal comfort by combining artificial neural networks and
metaheuristic algorithms, Energy Build. 239 (2021) 110839,

ed
https://doi.org/10.1016/j.enbuild.2021.110839.
[28] Imran, N. Iqbal, D. H. Kim, IoT Task Management Mechanism Based on Predictive Optimization
for Efficient Energy Consumption in Smart Residential Buildings, 257 (2022) 111762,
https://doi.org/10.1016/j.enbuild.2021.111762.

iew
[29] Z. Afroz, G. M. Shafiullah, T. Urmee, M. A. Shoeb, G. Higgins, Predictive modelling and
optimization of HVAC systems using neural network and particle swarm optimization algorithm, Build
Environ. 209 (2022) 108681, https://doi.org/10.1016/j.buildenv.2021.108681.
[30] Trimble Inc., SketchUp, https://www.sketchup.com/.
[31] Bigladder Software, Euclid, https://bigladdersoftware.com/projects/euclid/.

v
[32] D. B. Crawley, C. O. Pedersen, L. K. Lawrie, F. C. Winkelmann, EnergyPlus: energy
simulation program, ASHRAE J. 42 (2000) 49-56, https://www.researchgate.net/profile/Drury-

re
Crawley/publication/230606369_EnergyPlus_Energy_Simulation_Program/links/546a40d80cf2f5eb18
077919/EnergyPlus-Energy-Simulation-Program.
[33] D. B. Crawley, L. K. Lawrie, F. C. Winkelmann, W. F. Buhl, Y. J. Huang, C. O. Pedersen, R.
K. Strand, R. J. Liesen, D. E. Fisher, M. J. Witte, J. Glazer, EnergyPlus: creating a new-generation
building energy simulation program, Energy Build. 33 (4) (2001) 319-331,
er
https://doi.org/10.1016/S0378-7788(00)00114-6.
[34] M. F. Sanner, Python: a programming language for software integration and development, J.
Mol. Graph. Model. 17 (1) (1999) 57-61,
pe
https://www.academia.edu/download/25505223/10.1.1.35.6459.pdf.
[35] H. Drucker, C. J. C. Burges, L. Kaufman, A. Smola, V. Vapnik, Support Vector Regression
Machines, Advances in Neural Information Processing Systems 9, MIT Press (1996) 155-161,
https://proceedings.neurips.cc/paper_files/paper/1996/file/d38901788c533e8286cb6400b40b386d-
Paper.pdf.
ot

[36] A. J. Smola, B. Schölkopf, A tutorial on support vector regression, Stat Comput 14 (2004) 199-
222, https://alex.smola.org/papers/2004/SmoSch04.pdf.
[37] C. Cortes, V. Vapnik, Support vector networks, Mach Learn 20 (1995) 273–297,
tn

http://dx.doi.org/10.1007/BF00994018.
[38] T. R. Tooke, N. C. Coops, J. Webster, Predicting building ages from LiDAR data with random
forests for building energy modeling, Energy Build. 68 (A) (2014) 603-610,
https://doi.org/10.1016/j.enbuild.2013.10.004.
rin

[39] Z. Y. Wang, Y. R. Wang, R. C. Zeng, R. S. Srinivasan, S. Ahrentzen, Random Forest based

hourly building energy prediction, Energy Build. 171 (2018) 11-25,
https://doi.org/10.1016/j.enbuild.2018.04.008.
[40] L. Breiman, Random Forests Machine Learning, Kluwer Academic Publishers (2001),
ep

https://doi.org/10.1023/a:1010933404324.
[41] S. Touzani, J. Granderson, S. Fernandes, Gradient boosting machine for modeling the energy
consumption of commercial buildings, Energy Build. 158 (2018) 1533-1543,
https://doi.org/10.1016/j.enbuild.2017.11.039.
Pr

[42] P. Nie, M. Roccotelli, M. P. Fanti, Z. F. Ming, Z. W. Li, Prediction of home energy

consumption based on gradient boosting regression tree, Energy Rep. 7 (2021) 1246-1255,
https://doi.org/10.1016/j.egyr.2021.02.006.

20

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 [43] L. Breiman, Arcing The Edge, Technical Report 486. Statistics Department, University of
California, Berkeley CA.94720 (1997), https://statistics.berkeley.edu/tech-reports/486.

ed
[44] T. Q. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, Proc. 22nd acm sigkdd Int.
Conf. Knowl. Discov, data Min (2016) 785-794, https://doi.org/10.1145/2939672.2939785.
[45] H. F. Lu, F. F. Cheng, X. Ma, G. Hu, Short-term prediction of building energy consumption
employing an improved extreme gradient boosting model: A case study of an intake tower, Energy

iew
203 (2020) 117756, https://doi.org/10.1016/j.energy.2020.117756.
[46] G. L. Ke, Q. Meng, T. Finley, T. F. Wang, W. Chen, W. D. Ma, Q. W. Ye, T. Y. Liu, LightGBM:
A Highly Efficient Gradient Boosting Decision Tree, 31st Conference of Neural Information
Processing Systems (NeurIPS) (2017),
https://proceedings.neurips.cc/paper_files/paper/2017/file/6449f44a102fde848669bdd9eb6b76fa-

v
Paper.pdf.
[47] J. X. Guo, S. N. Yun, Y. Meng, N. He, D. F. Ye, Z. N. Zhao, L. Y. Jia, L. Yang, Prediction of

re
heating and cooling loads based on light gradient boosting machine algorithms, Build Environ. 236
(2023) 110252, https://doi.org/10.1016/j.buildenv.2023.110252.
[48] Y. Freund, R. E. Schapire, A desicion-theoretic generalization of on-line learning and an
application to boosting, EuroCOLT 1995: Computational Learning Theory (1995) 23–37,
https://doi.org/10.1007/3-540-59119-2_166.
er
[49] Z. J. Qu, H. X. Liu, Z. X. Wang, J. Xu, P. Zhang, H. Zeng, A combined genetic optimization
with AdaBoost ensemble model for anomaly detection in buildings electricity consumption, Energy
Build. 248 (2021) 111193, https://doi.org/10.1016/j.enbuild.2021.111193.
pe
[50] J. Kennedy, R. Eberhart, Particle Swarm Optimization, Proceedings of ICNN'95 - International
Conference on Neural Networks, IEEE, Perth, Austrilia (1995) 1942–1948,
https://doi.org/10.1109/ICNN.1995.488968.
[51] H. Molavi, M. M. Ardehali, Utility demand response operation considering day-of-use tariff
and optimal operation of thermal energy storage system for an industrial building based on particle
ot

swarm optimization algorithm, Energy Build. 127 (2016) 920-929,

https://doi.org/10.1016/j.enbuild.2016.06.056.
[52] R. Storn, K. Price, Differential evolution - a simple and efficient heuristic for global
tn

optimization over continuous spaces, J Glo Optim 11 (1997) 341–359,

https://doi.org/10.1023/A:1008202821328.
[53] J. A. Nelder, R. Mead, A simplex method for function minimization, Computer Journal, 7 (4)
(1965): 308–313, http://doi.org/10.1093/comjnl/7.4.308.
rin

[54] M. R. Hestenes, E. Stiefel, Methods of Conjugate Gradients for Solving Linear Systems, J Res
Natl Bur Stand 49 (6) (1952) 409-436, http://doi.org/10.6028/jres.049.044.
[55] D. C. Sorensen, Newton's Method with a Model Trust Region Modification, SIAM J. Numer.
Anal. 19 (2) (1982) 409–426, http://doi.org/10.1137/0719026.
ep

[56] C. G. Broyden, The convergence of a class of double-rank minimization algorithms, J. Inst.

Math. Appl. 6 (1970) 76–90, http://doi.org/10.1093/imamat/6.1.76.
[57] R. Fletcher, A New Approach to Variable Metric Algorithms, Comput J 13 (1970) (3) 317–
322, http://doi.org/10.1093/comjnl/13.3.317.
Pr

[58] D. Goldfarb, A Family of Variable Metric Updates Derived by Variational Means, Math.
Comput. 24 (109) (1970) 23–26, http://doi.org/10.1090/S0025-5718-1970-0258249-6.
[59] D. F. Shanno, Conditioning of quasi-Newton methods for function minimization, Math.

21

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602
 Comput 24 (111) (1970) 647–656, http://doi.org/10.1090/S0025-5718-1970-0274029-X.
[60] R. H. Byrd, P. Lu, J. Nocedal, C. Zhu, A Limited Memory Algorithm for Bound Constrained

ed
Optimization, SIAM J. Sci. Comput 16 (5) (1995) 1190–1208, http://doi.org/10.1137/0916069.
[61] CHN_Hunan.Changsha.576870_CSWD.epw, https://energyplus.net/weather-
location/asia_wmo_region_2/CHN/CHN_Hunan.Changsha.576870_CSWD.
[62] Ministry of Housing and Urban-Rural Construction of the People’s Republic of China, General

iew
Code of Energy Efficiency and Renewable Energy Application in Buildings, GB55015-2021, China
Building Industry Press, Beijing 2021.
[63] Ministry of Housing and Urban-Rural Construction of the People’s Republic of China,
Standard for Lighting Design of Buildings, GB50034-2019, China Building Industry Press, Beijing
2019.

v
[64] Ministry of Housing and Urban-Rural Construction of the People’s Republic of China,
Standard for design of care facilities for the aged, JGJ450-2018, China Building Industry Press,

re
Beijing 2018.
[65] Ministry of Housing and Urban-Rural Construction of the People’s Republic of China, Design
Code for Office Building, JGJ/T67-2019, China Building Industry Press, Beijing 2019.

er
pe
ot
tn
rin
ep
Pr

22

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4612602