Electric Power Systems Research 221 (2023) 109387

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Weather sensitive short term load forecasting using dynamic mode

decomposition with control
Amir Mansouri a , Amir H. Abolmasoumi a,b ,∗, Ali A. Ghadimi a,b
a Department of Electrical Engineering, Arak University, Arak, 38156-8-8349, Iran
b
Research Institute of Renewable Energy, Arak University, Arak, 38156-8-8349, Iran

ARTICLE INFO ABSTRACT

Keywords: The problem of short term load forecasting (STLF) for power grids using the dynamic mode decomposition
Short term load forecasting with control (DMDc) is considered. A forecasting model is discovered from time-series data based on the
Dynamic mode decomposition with control dynamic mode decomposition algorithm in which the effect of climatic factors on electric power consumption
Hierarchical clustering
is considered. An input selection method is also proposed to provide more informative dataset that efficiently
Data-driven modeling
reflects the load pattern changes. The meteorological data are processed through a hierarchical clustering
Weather sensitive load forecasting
method and is used by the DMDc algorithm as the inputs. The forecasting results with three datasets from
Electric Reliability Council of Texas, ISO New England, and Australian Energy Market Operator show the
effective performance of the proposed method compared to several other well-known forecasting methods
within the literature of STLF such as ARIMAX, SVR, and DMD. Specifically, the average daily load forecasting
errors are 4.78%, 7.6%, and 3.94% for the load datasets of three companies which indicates an improvement
of 21.64%, 15.55% and 10.45%, respectively, compared to the DMD method without considering the effect of
the climatic factors.

1. Introduction STLF which may fall into three main categories: classical methods, arti-
ficial intelligence(AI-)-based methods and hybrid methods. In classical
The ability to predict the precise behavior of loads in power grids methods, the load’s future value is predicted using a linear combina-
contributes to solve some essential problems regarding the operation tion of previous load values and the previous and current values of
and maintenance such as real-time control, price management, and exogenous factors such as meteorological variables [4]. Among the
grid protection. Forecasting the electric power is a requirement for classical methods, time-series methods are of the most popular ones.
providing the required operation reliability and meeting the econom- For example, in [5] the short-term load forecasting using different time-
ical consumer demands. In fact, load forecasting modeling is a key to series implementations for the load data from ten European countries
power systems planning that, if not done efficiently, can damage the has been compared, and finally it has been concluded that the Holt–
entire planning process. For example, a conservative study showed that Winters exponential smoothing-AR model has the best performance.
a %1 reduction in forecasting error for a 10,000 MW power company In [6] a seasonal autoregressive integrated moving average (SARIMAX)
could save up to annual $1.6 million [1]. From time-frame viewpoint,
model provides a two-step process for forecasting load demand over
the load forecasting can be put into three categories; i.e. short-term,
a period of one to nine days. These models use simple mathematical
medium-term and long-term. Among which the short-term load fore-
formulas and are effective in linear prediction problems, but are not
casting (STLF), i.e. forecasting over a period of one hour to one week,
efficient for processing nonlinear and complex load time series [7].
plays an important role in various operational functions of the power
Artificial neural networks create nonlinear mapping between input
systems, such as unit commitment, economic dispatch, energy transfer
and demand variables by learning patterns from historical data and
scheduling and real-time control [2].
exogenous variables. Classical neural networks [8,9], support vector
The measured dataset in the form of time-series is commonly com-
plex and non-stationary. Moreover, it depends on multiple external fac- machines [10,11] and, recently, the deep neural networks [12,13] are
tors such as the meteorological variables and/or the mental consumer the most popular methods from the second category. Authors in [8]
behavior, etc. Therefore, load forecasting is usually a complicated propose a short-term electric load forecasting model based on an arti-
task [3]. In recent decades, various methods have been suggested for ficial neural network in which statistical methods are used to select

∗ Corresponding author.
E-mail address: [email protected] (A.H. Abolmasoumi).

https://doi.org/10.1016/j.epsr.2023.109387
Received 5 November 2022; Received in revised form 1 April 2023; Accepted 3 April 2023
Available online 12 April 2023
0378-7796/© 2023 Elsevier B.V. All rights reserved.
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

variables and models in order to obtain the optimal generalization addressed by this paper. To address this issue, those variants of DMD
of neural networks. Also in [9], in order to increase the prediction which take into account the exogenous inputs should be considered.
accuracy, a boosted neural network which consists of a set of iteratively Such DMD variants have been referred to as DMD with control (DMDc)
trained ANNs is proposed. In each iteration, in order to reduce the and are presented in [38,39], where the input-affine linear state–
forecasting error, a new ANN model is trained using previous iteration space model is calculated from the data snapshots of the features and
models. In [14], a machine-learning-based load forecasting method inputs. The DMDc makes it possible to extract the role of the weather-
is presented using smart meters and weather sensor data from 114 related factors in the load variation rule. It should be noted that the
residential consumers in New England. A novel feed-forward ANN application of DMDc-based method for STLF is not straightforward and
algorithm implemented by a backtracking adjustment of the learning the selection of inputs such as climatic temperature and humidity is a
rate is utilized and the results show a good prediction accuracy for critical part that needs accurate analysis and adjustment. In this paper,
short-term forecasting of household loads. In [10], a method for short- firstly, the temperature and humidity datasets are pre-processed to
term load forecasting based on the support vector regression together provide the average real-feel temperature which is utilized as the raw
with Grasshopper optimization algorithm is presented and the results input of the DMDc-STLF method. Secondly, a hierarchical clustering
are compared with other methods of parameter optimization for the method is used to select suitable inputs. The input selection step serves
support vector machines. The load forecasting using feature selection to enhance the quality of the STLF by removing the restrictions of
through the auto-correlation function and the least squares support previously proposed methods.
vector regression is presented in [11] where the parameters have been As mentioned, in this paper, an accurate, fast and non-parametric
optimized using the Gray-Wolf optimization algorithm. In [12], a long- DMDc-based method for short-term load forecasting is presented which
term short-term memory (LSTM) neural network-based method, as one also considers the weather factors and improves the accuracy of the
of the most popular deep learning techniques, is applied to the STLF load prediction by using the hierarchical clustering for the input selec-
problem. The proposed method has been tested on a set of residential tion.
smart meter data and has been concluded to perform more efficient The contributions of this paper to the body of STLF literature
than the classical neural network methods for the residential house- knowledge can be highlighted and summarized as follows:
holds. In [13], the wavelet transform is used to remove the fluctuation
• This paper fills the gap regarding explicit incorporation of the
in the load and price time series. Then a feature selection method based
climatic factor in highly efficient DMD-based methods. Since the
on entropy and mutual information is used to rank candidate inputs
consumption of electricity in summer is affected by the tempera-
and to remove the redundant inputs according to their information
ture and humidity, and in order to involve these factors, after the
values. Finally, load and price are predicted using the LSTM algo-
processing of the input meteorological data, the DMDc method is
rithm. All these AI-based methods exhibit acceptable load prediction
adopted to address the load forecasting in presence of available
accuracy. However, they also call for time-consuming computations,
weather-related data. These will improve the results of [29].
are dependent on the proper choice of parameters, may suffer from
• While [29] and many other papers utilize four previous days
over-fitting problem and are not usually capable of generalization.
(i.e., two immediate previous days, same day in the previous week
The hybrid methods aim to combine two or more STLF methods in
and previous day in the previous week) for historical load data,
order to eliminate the disadvantages of the individual methods [15].
in this study, more richness is given to the input data by using the
Although hybrid models have higher accuracy, they are usually highly
clustering technique by adopting more appropriate clustering to
complex and their performance depends on parameters that are difficult
provide more appropriate data for extracting the underlying load
to adjust [7].
patterns and discovering the consumer behavior.
Dynamic mode decomposition (DMD) [16,17] has recently received
• It is observed that the consumption of electric energy on week-
considerable attention as a method to discover the underlying dynamics
ends and Mondays is different from the rest of the week due to
governing time-series data. The DMD has shown to be efficient in
the change in the consumption pattern. Such weekend prediction
the reconstruction and prediction of the spatio-temporal data [18].
is also addressed in this study.
Also DMD has a strong connections to the Koopman operator the-
ory [19], i.e. it discovers the linear dynamics as an approximation of The remainder of this paper is arranged as follows: Section 2 describes
the Koopman operator [20]. Several applications have been reported the preliminaries and reviews the technical requirements. Section 3
for DMD such as in fluid mechanics [19], power system analysis [21], focuses on the implementation of the forecasting method. In Section 4,
neuroscience [22], finance [23], climate science [24], and transporta- the assessment of the proposed model on three datasets is given.
tion [25]. Several modifications on the original DMD has been also Finally, the conclusions and the future perspectives are discussed in
proposed to address the data quality issues such as [26–28]. The DMD Section 5.
has been applied to the problem of STLF in [29] using a Hankelized
variant of the original DMD together with an averaging method. The 2. Preliminaries
results from comparison of the DMD with other existing STLF methods
show that the DMD-STLF has low computational complexity, is robust 2.1. Dynamic mode decomposition with control
and accurate, and easy to implement. In another study [30], the use
of dynamic mode decomposition method in forecasting the load of the Dynamic mode decomposition is a data-driven method that can
Irish grid has been investigated, the results of which indicate the high produce a linear reduced order model of a complex nonlinear dynamics
accuracy and performance of DMD-STLF. such that the temporal and spatial modes of the system are obtained.
The electric loads are usually affected by the climatic and season-to- This method was first introduced by Schmid [40] in the field of fluid
season fluctuations. Meteorological factors such as temperature and hu- dynamics. The increasing success of DMD stems from the fact that it is
midity directly affect the profile of the energy consumption. an equation-free data-driven. The DMD can provide accurate analysis of
Temperature-to-load dependency has been investigated in a number of a data flows into coherent spatio-temporal structures that can be used
articles such as [31–37]. However, the effect of climatic factors has for short-term forecasting and control [38]. As an important modifica-
not been addressed by the DMD-STLF. The DMD-STLF method does tion to DMD, authors in [39] propose the dynamic mode decomposition
not explicitly consider the weather factors such as the temperature with control (DMDc) which extends DMD to take into account the effect
and humidity of the region where the load is being predicted. Due to of exogenous force or the control input on system dynamics. In other
the demonstrated high effectiveness of the DMD-based load forecast- words, DMDc is able to discover the approximate dynamics from data
ing [29] the lack of considering weather factors is a big gap which is collected in the presence of external excitation. The main contribution

2
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

of this paper is to employ the DMDc for load forecasting where the humidity. The body loses some of its heat as a result of the process
climatic variables such as the temperature and humidity are considered of transpiration and evaporation, and as a result, it cools down. When
as the exogenous inputs. The previously presented DMD-STLF method, the relative humidity is high, the rate of evaporation decreases and the
however, relies only on load power time-series which may reflect the body loses less heat, which results in more body heat. Therefore, the
implicit effect of the weather variable. Instead, we suggest to discover perceptible temperature increases with an increase in the temperature
the explicit influence of these variables through DMDc and to use it and the relative humidity. Heat index is the result of extensive biolog-
for STLF. Since weather variables such as temperature and humidity ical meteorology research described by Steadman in 1979 [41]. There
directly stimulate the energy demand the proposed DMDc-STLF algo- are many formulas for approximating the original tables developed
rithm is a way to account for such variables in the load forecasting. by Steadman. Among them the result obtained by multiple regression
In the original DMDc, the main goal is to determine the relationship analysis performed by Rothfusz is well known and is [42] described as
between the three vector of measured or known variables, the current
state vector 𝐱𝑘 , the future state vector 𝐱𝑘+1 , and the current control 𝐻𝐼 = 𝑐1 + 𝑐2 𝑇 + 𝑐3 𝐻 + 𝑐4 𝑇 𝐻 + 𝑐5 𝑇 2 + 𝑐6 𝐻 2 + 𝑐7 𝑇 2 𝐻
vector 𝐮𝑘 . This relationship can be represented as + 𝑐8 𝑇 𝐻 2 + 𝑐9 𝑇 2 𝐻 2 , (11)
𝐱𝑘+1 ≈ 𝐀𝐱𝑘 + 𝐁𝐮𝑘 , where the constants are given as
𝐱𝑘 ∈ R𝑛 , 𝐮𝑘 ∈ R𝑙 , 𝐀 ∈ R𝑛×𝑛 , 𝐁 ∈ R𝑛×𝑙 (1) 𝑐1 = −42.379, 𝑐2 = 2.04901523, 𝑐3 = 10.14333127,
The spatio-temporal data is assumed to be composed of 𝑚 snapshots of 𝑐4 = −0.22475541, 𝑐5 = −6.83783 × 10−3 ,
the state vector organized in two matrices as 𝑐6 = −5.481717 × 10−2 𝑐7 = 1.22874 × 10−3 ,
𝐗1 = [𝐱1 𝐱2 𝐱3 ...𝐱𝑚−1 ] ∈ R 𝑛×(𝑚−1)
𝑐8 = 8.5282 × 10−4 , 𝑐9 = −1.99 × 10−6 .
𝐗2 = [𝐱2 𝐱3 𝐱4 ...𝐱𝑚 ] ∈ R𝑛×(𝑚−1) (2) In (11) 𝐻𝐼 represents the heat index in Fahrenheit, 𝑇 is the dry bulb
temperature in Fahrenheit and 𝐻 is the relative humidity that is a
The matrix of the collected snapshots from the inputs, or controls, is
percentage value between 0 and 100. Here, regardless of factors such as
also defined as
sunlight, wind speed, precipitation and other factors which affect our
Ψ = [𝐮1 𝐮2 𝐮3 … 𝐮𝑚−1 ] ∈ R𝑙×(𝑚−1) (3) feelings, the approximated real feel temperature is assumed to be equal
to the calculated heat index by using (11).
The system dynamics are to be approximated through DMDc as the
following matrix form:
2.3. Clustering
𝐗2 = 𝐀𝐗1 +𝐁Ψ (4)
In the literature, clustering methods have been used extensively
which can be rewritten as
( ) to improve the forecasting accuracy [43–53]. The load consumption
( ) 𝐗1 features varies from one location to another and from time to time.
𝐗2 = 𝐆Ω = 𝐀 𝐁 , (5)
Ψ Therefore, it is a good idea to examine the load profiles in order
̂ ∈ R𝑛×(𝑛+𝑙) is obtained by minimizing the to build some classified zones based on similar load behaviors [43].
The least squares solution 𝐆
Clustering analysis aims to group observations with similar charac-
Frobenius norm ‖ ‖𝐗 2 − 𝐆𝜴 ‖ . If the matrix 𝜴 ∈ R(𝑛+𝑙)×(𝑚−1) has the
‖𝐹 teristics within the same cluster. The similarities between any pairs
singular value decomposition (SVD) as
of observations are normally evaluated using distance based metrics,
Ω = 𝐔Σ𝐕∗ (6) such as the Manhattan, Euclidean and auto-correlation metrics. The
clustering process aims to maximize observation similarities within
̂ matrix is obtained as
then the least square estimation 𝐆
the same cluster and minimize the similarities between different clus-
̂ = 𝐗2 Ω† = 𝐗2 𝐕Σ−1 𝐔∗ ,
𝐆 (7) ters [46]. Therefore, selecting similar days and classifying them into
the same category as the training samples may be effective on the
̂ and 𝐁
Matrices 𝐀 ̂ are obtained as sub-matrices of 𝐆:
̂
accuracy of load forecasting [51]. We show that the clustering is a
( ) ( −1 ∗ −1 ∗
) more powerful tool than the auto-correlation function in discovering
̂ = 𝐀
𝐆 ̂ 𝐁 ̂ ≈ 𝐗2 𝐕Σ 𝐔 𝐗2 𝐕Σ 𝐔 , (8)
1 2
underlying load patterns. Also using the clustering method enables us
where to employ a wider range of data as to be included in the training
( ∗ ) dataset for future load predictions. One of the most well-known clus-
𝐔1 𝐔∗2 = 𝐔∗ , 𝐔1 ∈ R(𝑛+𝑙)×𝑛 , 𝐔2 ∈ R(𝑛+𝑙)×𝑙 , (9)
tering models is the hierarchical clustering. Hierarchical clustering is a
By obtaining matrices 𝐀 ̂ and 𝐁,
̂ the linear dynamic evolution rule of flexible and deterministic algorithm which outperforms other methods
the state vector can be obtained as such as K-centers [50,54]. The hierarchical algorithm produces a tree
or dendrogram by either agglomerative (bottom-up) or divisive (top-
𝐗(𝑡 + 1) = 𝐀𝐗(𝑡) + 𝐁Ψ(𝑡), 𝑡 = 1, 2, … , 𝐹 . (10)
down) methods. The agglomerative method makes use of a distance
The DMDc discovered rule (10) can be used for both the reconstruc- matrix in which the distance between each pair of observations (load
tion and future prediction of the state vector where F represents the curves) is calculated based on a distance criterion. Depending on the
run-time time horizon. elements of such distance matrix, the observations or clusters that have
the least distance, or the most similarity merge to form a new cluster.
2.2. Real-feel temperature In the next step, the distance between new observations or clusters is
calculated to form an updated distance matrix. It then continues until
Among meteorological variables, temperature is the most important only one cluster remains. Hierarchical clustering is different from other
independent factor influencing the energy demand. This is especially clustering techniques as it measures distances between clusters. In this
true in areas with hot climates during the summer due to the high way, the two clusters that have the most similarity are merged by the
use of cooling devices. High levels of humidity make the temperature linkage criteria to form a new cluster. For further reading, see [55].
more effective on the energy demand. The heat index shows the human The selection of distance measures is very determining on the cluster-
feeling towards the heat of the environment, the amount of which is ing results. The Euclidean distance is commonly used as the distance
determined by considering two factors of air temperature and relative metric, however, it is blind to capture vector or feature correlation it

3
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

series data based on their similarities. Therefore, we suggest to employ

the hierarchical clustering with correlation criterion before the input
selection step. Here, the number of training days is selected to be 28
days and the Calinski Harabasz criterion is utilized to calculate the
optimal number of clusters (See (12)). It should be noted that prior
to analysis, the data was processed to fill the missing data using linear
interpolation. There are no general rules to follow in the selection of
input variables. It depends largely on experience, professional judgment
and preliminary experimentation [62]. Although the load profile of the
forecasting day is unknown it is known that, according to the principle
of load consumption, the daily consumption has a similar trend to that
of the next day. Also there is a weekly periodicity. Therefore, we select
training data from the days belonging to the clusters in which the
previous day and the same day in previous are located. However, it is
important to note that to forecast Saturdays and Mondays, due to their
nature, the previous day is not a good pattern. Therefore, to forecast
Fig. 1. A typical dendrogram of hierarchical clustering method. these days, only the days belonging to the cluster in which the same
day in previous week is located are used. In short, by relying on the
clustering technique, a wider range of training data is obtained, which
is a more realistic representation of the nonlinear and non-stationary
treats the time series as a vector, and as a result, ignores the temporal
nature of the load.
correlations in the data [56]. For time-series data, analyzing similarities
based on Pearson’s correlation is more suitable [57]. There are also
3.2. Temperature processing
various criteria for finding the optimal number of clusters. In this
study, we use the method proposed by Calinski and Harabaz [58]. The
To consider the simultaneous effect of temperature and humidity
effectiveness and robustness of this criterion has been studied through
on energy consumption, the real feel temperature is obtained through
several researches such as [59–61]. The criterion is defined as the ratio
(12). It should be noted that in addition to the temperature of the
of the sum of between-cluster and the inter-cluster dispersions for all
current hour the temperatures of the previous hours also affect the
clusters as follows:
∑𝑘 load. A common method for incorporating the temperature value of
𝑛𝑖 ||𝑚𝑖 − 𝑚|| 𝑁 − 𝑘
2
the previous hours is to utilize the average temperature of the previous
𝐶𝐻𝑘 = ∑ 𝑖=1 ∑ (12)
𝑘 | |2 𝑘 − 1 hours [63–65]. The average temperature can be calculated as
𝑖=1 𝑥∈𝑐𝑖 |𝑥 − 𝑚𝑖 |
1∑
𝑎
where 𝑘 is the number of clusters, 𝑁 is the number of observations, 𝑛𝑖 𝑇𝑅𝑎 (𝑘) = 𝑇 (𝑘 − 𝑖 + 1) (13)
is the number of observations in cluster 𝑖, 𝑚𝑖 is the centroid of cluster 𝑖, 𝑎 𝑖=1 𝑅
𝑚 is the overall mean of the sample data. By maximizing with respect
where 𝑇𝑅 (𝑘) is the current real feel temperature and 𝑇𝑅𝑎 (𝑘), is the
to 𝑘, the optimal value of clusters is obtained. A summary of applied
average real feel temperature over 𝑎 hours. To find the proper number
hierarchical time series clustering method is depicted in Fig. 1. The
of needed hours to achieve an appropriate average temperature, the
load series on the left side of the figure are the training data of the
Pearson correlation coefficient (PCC) between the demand and the real
forecasting model, which belongs to the last 9 days. It should be noted
feel temperature, i.e. as a well-known method for measuring the linear
that in this paper, 28 days are considered as training data. However, for
dependence between two quantities, is used. The most appropriate
simplicity, only 9 days are displayed in Fig. 1. As seen, the employed
averaging hour is then obtained from the highest value of PCC.
clustering algorithm that makes use of the distance criterion and the
optimal number of clusters (here 4 clusters) puts the days 2, 5, 6, 8,
3.3. Data pre-processing
and 9 in the first cluster, days 4, and 7 in the second cluster, day 3 in
third cluster and finally puts day 1 in fourth cluster.
After selecting the appropriate days whose temperature sequences
is to be used as the training data, the corresponding average real
3. DMDc-based short term load forecasting
feel temperatures is selected. Afterwards, it is necessary to normalize
the load and averaged real feel temperature series. For simplicity, the
This section describes the proposed DMDc-based STLF method. The
averaged real feel temperature is represented as 𝑇 in the rest of the
presented method is composed of five steps: (1) Input selection step,
study. Assume that the time series of the load is represented as 𝑥𝑖 ∈
(2) Temperature processing step, (3) Data pre-processing, (4) DMDc
𝑅, 𝑖 = 1, 2, … , 𝑚 and the real feel temperature time series is represented
calculation, and (5) Forecasting step. In what follows each step is
as 𝑇𝑖 ∈ 𝑅, 𝑖 = 1, 2, … , 𝑚, then
described in details.
𝑥𝑖 − min(𝐗)
𝑥𝑖 (𝑛𝑜𝑟𝑚) =
3.1. Similar day-based load input selection using hierarchical clustering max(𝐗) − min(𝐗)
𝑇𝑖 − min(𝐓)
𝑇𝑖 (𝑛𝑜𝑟𝑚) = (14)
As mentioned, DMDc is a method to extract the underlying dynamic max(𝐓) − min(𝐓)
patterns of the data by taking into account the external forces or where min(𝐗) and max(𝐗) represent the minimum and maximum values
inputs. This lets us to include also the environmental factors such of the load data sequence, and min(𝐓), max(𝐓) represent the minimum
as the temperature and humidity in the STLF process. In order to and maximum values of the average real feel temperature sequence,
utilize the DMDc for forecasting, it is firstly required to provide it respectively. As mentioned in Section 2, we need data matrices includ-
with the load information regarding past days. In this regard, it is very ing the temporal load data to approximate the underlying dynamics
crucial to provide the load data sequence from days with maximum using the DMDc algorithm. Time delay embedding is a way to convert
similarity to the forecasting day. This process we refer to as input a time series to a matrix of time-dependent data. In fact, it can be
selection stage. As discussed in the previous section, the hierarchical thought as a mapping that transmits a one-dimensional time series to
time series clustering is an appropriate method to cluster the time a multidimensional matrix. The normalized load values and real feel

4
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

temperature series data are converted to multidimensional augmented The conceptual illustration of the proposed DMDc-based STLF is
data matrix as depicted in Fig. 2. As can be seen, the proposed method consists of six
⎛ 𝑥1 𝑥2 ⋯ 𝑥𝐿 ⎞ modules. Firstly, the data of the historical load is collected along with
⎜𝑥 𝑥3 … 𝑥𝐿+1 ⎟ temperature and humidity. Then, by clustering the load time series,
⎜ 2 ⎟ suitable days are selected for forecasting, also the temperature and
𝐗=⎜⋮ ⋮ ⋱ ⋮ ⎟
⎜⋮ ⋮ ⋱ ⋮ ⎟ humidity data are processed in parallel to obtain the average real feel
⎜ ⎟ temperature. At the next step, the load data of the selected days and
⎝𝑥𝑆 𝑥𝑆+1 … 𝑥𝑆+𝐿−1 ⎠
the corresponding processed temperatures are normalized. Afterwards,
⎛ 𝑇1 𝑇2 ⋯ 𝑇𝐿 ⎞
⎜𝑇 the DMDc algorithm is implemented by performing the time delay
𝑇3 … 𝑇𝐿+1 ⎟
⎜ 2 ⎟ embedding producer on load time-series data and finally at the last
Ψ=⎜⋮ ⋮ ⋱ ⋮ ⎟ (15)
⎜⋮ step the load forecasting is carried out based on the discovered DMDc
⋮ ⋱ ⋮ ⎟
⎜ ⎟ characteristics.
⎝𝑇𝑆 𝑇𝑆+1 … 𝑇𝑆+𝐿−1 ⎠
where 𝐿 is called window length and is the only parameter that can 4. Experiment and analysis
be determined at this step. Since the time series have a daily cycle
the value of this parameter is chosen equal to the size of observations This section aims to verify the performance of the proposed STLF
during one day. If the number of the time series load measurement method for real load data. Three datasets from different energy mar-
is equal to 𝑚 then the parameter 𝑆 in (15) is determined to be 𝑆 = ket operators are utilized, i.e. Electric Reliability Council of Texas
𝑚 − 𝐿 + 1. (ERCOT) [66], ISO new England [67], and Australian Energy Market
Operator(AEMO) [68]. Also the meteorological data for the first two
3.4. DMDc algorithm cases are taken from [69] and for the third case from [70]. Simulations
are performed within Matlab software environment version 2021a
After time series data of the measured loads and average real feel
by Windows7 operating system. The hardware used has a 2.67 GHz
temperature are arranged as the matrix form in (15) the operator 𝐀
processor and 8 GB of internal memory.
and 𝐁 are estimated through DMDc from (6)–(10). Then (4) models the
evolution of the data time series which is in fact the inherent dynamics
4.1. Performance measures
of consumers behavior embedded in the load series. Moreover, the esti-
mated value of 𝐁 describes the effect of changes in real feel temperature
The assessment of the proposed forecasting method is based on three
on amount of electrical load consumption.
standard evaluating measures including mean absolute percentage er-
3.5. Forecasting ror (MAPE), mean absolute error (MAE), and mean square error (MSE)
which are defined as follows:

1 ∑ || 𝑥𝑎 (𝑖) − 𝑥𝑓 (𝑖) ||
After identifying the linear approximation of the underlying load/ 𝑁

temperature dynamics, the future load values can be predicted. It 𝑀𝐴𝑃 𝐸 = | | (18)
𝑁 𝑖=1 || 𝑥𝑎 (𝑖) |
|
should be noted that in the traditional usage of DMD/DMDc algorithm
1 ∑|
𝑁
the arrangement of the data snapshots of the state vectors are rep- |
𝑀𝐴𝐸 = |𝑥 (𝑖) − 𝑥𝑓 (𝑖)| (19)
resented as (2) while in time-series load forecasting the Hankelized 𝑁 𝑖=1 | 𝑎 |
representation of the data series in (15) is utilized. Therefore, the √
√
discovered dynamics, although obtained in the form of (10), represents √1 ∑ 𝑁
𝑅𝑀𝑆𝐸 = √ (𝑥 (𝑖) − 𝑥𝑓 (𝑖))2 (20)
the evolution rule of the system corresponding to (15). Using the 𝑁 𝑖=1 𝑎
discovered matrix gains 𝐴, 𝐵 in (10), the future value of the load
data matrix 𝐗 can be predicted. Afterwards, the predicted load values where 𝑁 denotes the number of predicted points 𝑥𝑎 (𝑖) indicates the
are de-normalized to the original data range. Assume the matrix of actual value in 𝑖th time and 𝑥𝑓 (𝑖) is the corresponding predicted value.
de-normalized predicted values to be MAE depicts the average error due to magnitude difference between
the observed and estimated values. MAPE reflects the accuracy of the
⎛ 𝑥̂ 1 𝑥̂ 2 ⋯ 𝑥̂ 1𝐹 ⎞ forecasting method and MSE measures the general deviation of the
⎜ 𝑥̂ 2′ 𝑥̂ 3′ … ⋮ ⎟
⎜ ⎟ estimated values from true measurements. The best forecasting method
⋮ ⋮ ⋯ ⋮
𝐗̂ = ⎜ ⎟ (16) is the one which provides the smallest value on each of these metrics.
⎜ ⋮ ⋮ ⋯ ⋮ ⎟
⎜ ⋮ ⋮ ⋯ ⋮ ⎟
⎜ ⎟ 4.2. Experiment 1: Effect of appropriate input selection on DMD-based
⎝𝑥̂ 𝑆 ′ ...′ 𝑥̂ 𝑆 ′ ...′ +1 … 𝑥̂ 𝑆 ′ ...′ +𝐹 −1 ⎠𝑆×𝐹
forecasting
⌢
Since matrix 𝑋 is not necessarily of Hankel form (for example 𝑥2′ ≅ 𝑥2
DMD algorithm is a data-driven, model-free method that recon-
and not 𝑥2′ = 𝑥2 ), in order to obtain an accurate forecast, it is necessary
to perform a sub-diagonal averaging operation on the predicted matrix structs the underlying dynamics from snapshot measurements only.
elements. The goal of diagonal averaging is to convert the given matrix In the case of STLF, due to the non-stationary characteristics of the
into a Hankel matrix form which can be subsequently turned into a time load and its dependence on various factors such as time of the day
series. Set 𝑆 ∗ = min(𝑆, 𝐹 ), 𝐹 ∗ = max(𝑆, 𝐹 ), 𝑁 = 𝑆 + 𝐹 − 1, 𝑞 = 𝑖 + 𝑗, and day of the week (working days, holidays), it is necessary to select
𝑥̂ ∗𝑖𝑗 = 𝑥̂ 𝑖𝑗 if 𝑆 < 𝐹 and 𝑥̂ ∗𝑖𝑗 = 𝑥̂ 𝑗𝑖 otherwise. Then the element 𝑥̃ 𝑖𝑗 of the the appropriate input. In this section, it is shown that the hierarchical
Hankel matrix is calculated as clustering provides higher performance than methods based on the
auto-correlation functions (ACF) such as [29]. Mohan et al. [29] use
⎧ 1 ∑ ∗
𝑞−1
⎪ 𝑥 2 ≤ 𝑞 ≤ 𝑆 ∗ − 1, four days as the input for the day of prediction: two immediate previous
⎪ 𝑞 − 1 𝑙=1 𝑙,𝑞−1 days, same day in the previous week and previous day in the previous
⎪
𝑆∗ week. In this section, to obtain a large number of statistical samples,
⎪
𝑥̃ 𝑖𝑗 = ⎨ 1 ∑ ∗ (17)
𝑥 𝑆∗ ≤ 𝑞 ≤ 𝐹 ∗, we predict the load of day-ahead and calculate the results for all days
⎪ 𝐹 ∗ 𝑙=1 𝑙,𝑞−1
⎪ in the 2018 for South Zone of ERCOT, year 2019 for ME Zone of New
𝑆∗
∑
⎪ 1
𝑥∗𝑙,𝑞−1 𝐹 ∗ + 2 ≤ 𝑞 ≤ 𝑁 + 1.
England and year 2016 for NSW Zone of AEMO. The average results for
⎪𝑁 − 𝑞 + 2 the whole years are shown in Table 1. The best performance values are
⎩ 𝑙=𝑞−𝐹 ∗

5
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

Fig. 2. Block diagram of DMDc with clustering method for STLF.

represented in bold. It is observed that the proposed combination of Table 1

the DMD with the hierarchical clustering method offers more accurate The averaged MAPE, MAE, and RMSE of DMD with clustering and DMD with ACF
prediction results for all datasets. Such improvement comes from higher methods for three datasets.
quality training of the forecasting by clustering the historical load series Forecasting method Evaluation criteria South ME NSW
and extracting the most appropriate data. Year Year Year
2019 2018 2016
4.3. Experiment 2: Weather sensitive STLF using DMDc MAPE 7.35 5.45 6.19
DMD with ACF MAE 258.5 72.22 510.4
RSME 302 84.8 611
The DMDc inherits the advantages of DMD. Besides, it provides
accurate input–output models for complex systems which include actu- MAPE 6.36 4.64 5.25
DMD with clustering MAE 218.3 61.1 430.2
ation in their dynamics [39]. Meteorological factors are the of critical RSME 257.2 71.9 515.7
importance in stimulating demand especially during warm days of
the year. This can be clearly seen from Fig. 3 where the electricity
consumption in three different areas are depicted, i.e. the COAST
Zone of ERCOT and the WCMA Zone of New England, both in 2020,
demand is significantly increased due to activity of cooling equipment.
and Queensland in 2017. As seen, for COAST and WCMA the energy
consumption in summer (months June, July and August) is higher than As Queensland is in the southern hemisphere it sees the same effect in
other seasons. It is because during summer season, the electrical load January, February and March. The purpose of this section is to study

6
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

Fig. 3. Annual electrical load consumption for (a) ERCOT In 2020, (b) WCMA In 2020, (c) Queensland In 2017.

the STLF problem warm months where the temperature and humidity have been used for prediction. To show the effectiveness and accuracy
have a significant impact on the load consumption profile. of the proposed method in forecasting the weekend days, as well as
In order to assess the performance of the DMDC-based STLF method, on Monday, the data of the warm months of the South-Central zone of
we compare it with three well-known forecasting methods, i.e. Support ERCOT in 2018 will be selected.
Vector Regression (SVR), ARIMAX and DMD with auto-correlation As an instance of DMDc-based STLF performance in comparison
function as its input classification step. All methods are tested using with other four well-known methods, the predicted and real weekend
the same database. MATLAB hyperparameters optimization capability load curves from the last week of June 2018 are depicted in Fig. 5.
has been utilized to find the optimal parameters in SVR. Also Bayesian The average MAPE for weekends and Mondays throughout the summer
information criteria has been utilized to find optimal parameters for is also listed in Table 3. The results show the satisfactory accuracy of
ARIMAX. The DMD with auto-correlation function is implemented the proposed method in predicting the weekend days and Mondays. To
according to paper [29]. In order to obtain a large number of statistical accurately measure the computational time required to compare the
samples, we do one-day-ahead load value prediction for all days of proposed method; First, the total time required for training and fore-
the warm month. The resulting average MAPE, MAE, RSME of load casting all days of summer 2018, which includes 92 days, is calculated,
forecasting models are represented in Table 2. The best performance and then the mean for one day is calculated and listed in Table 4.
values appear in bold. The results show that the accuracy of the
proposed method for all datasets are satisfactory Higher performance
5. Research and practical implications
of the proposed DMDc-based STLF is more tangible in July and August
for the COAST Zone, in August for the WCMA Zone and in March for
the Queensland Zone. Also, the comparison of the DMD with clustering The proposed method of this paper is a combination of the DMDc
and DMDc with clustering confirms the effect of considering temper- and hierarchical clustering method for STLF. As a step forward in
ature and humidity in increasing the forecasting accuracy, so that the research literature, the DMDc-STLF is advantageous over the previously
accuracy of the proposed method is higher in all evaluation months. proposed DMD-STLF as it is capable of incorporating the weather vari-
Fig. 4 depicts the bar graph and line graph of mean and standard ables. The hierarchical clustering method is also included in the pro-
deviation of MAPE for each STLF method for all summer days. Also posed method to enhance the input selection leading to improvement
the boxplot charts representing the distribution MAPEs are depicted in in prediction accuracy. Generally, in time-series forecasting methods,
the right column of the figure. In boxplot charts, the horizontal center- it is assumed that there is not going to be vast change in the signal to
line represents the median, the edge locations correspond to the upper be predicted at different times. Therefore, based on the past behavior
and lower quartiles, and the ‘‘+’’ symbol represents the outliers. The of the data sequence, its future can be estimated. The change in
statistical illustration shows that not only the proposed method has a weather conditions can challenge the time-series-based STLF methods.
lowest error mean but also has a more uniform distribution with less On the other hand, including the effect of the exogenous factors in the
dispersion and less error outliers. forecasting process can increase the forecasting accuracy. This is what
DMDc-STLF can add to the DMD-STLF. It was shown that this method
4.4. Experiment 3: Weekend forecast is simple to implement and does not impose almost any assumptions
on the system. At the same time it has a high accuracy that makes it
As mentioned before, the load patterns on Saturdays and Mondays a good choice for real-time predictions for purposes such as electricity
have no similarity to the load patterns of their previous days, therefore, consumption, energy prices, wind power, solar energy.
considering the previous day in their prediction causes deviation in The method presented in this paper can provide a tool for energy
the results. Therefore, for the prediction of these two days, only the companies to forecast the future electrical load especially in summer
weekly similarity has been taken into account. In other words, only days or in regions with tropical climates without getting involved with
days belonging to the cluster of the similar day in the previous week different settings and parameters.

7
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

Table 2
The averaged MAPE, MAE, and RMSE of proposed and benchmark methods for three datasets.
Forecasting method Evaluation criteria COAST WCMA Queensland
Year2020 Year2020 Year2017
June July Aug. June July Aug. Jan. Feb. Mar.
MAPE 5.21 5.01 4.98 7.1 6.59 9.01 5.49 4.22 3.68
SVR with clustering MAE 782.2 756.1 770.7 131.8 145.6 180.2 397.9 306.6 249.8
RSME 965.5 920.6 918.6 158.6 174.8 210.9 480.8 383.8 302
MAPE 4.72 4.92 4.97 11 6.66 8.79 4.83 3.25 3.64
ARIMAX MAE 681.1 759.2 767.9 198.2 148.6 176.9 351.1 238.6 250.4
RSME 840.8 938.1 940.1 228.5 180.2 207.8 428.3 294.6 319.9
MAPE 6.09 6.3 5.9 9.22 8.67 9.13 4.99 3.82 4.33
DMD with ACF MAE 880.1 971.7 898.2 168.3 191.7 188.3 358.9 280.2 295.6
RSME 1039 1152 1049 196.3 226.9 219.7 433.1 335.1 361.3
MAPE 5.34 5.13 4.98 9.02 7.28 8.38 5.15 3.56 3.45
MAE 772.4 769.7 750.5 159.3 163.1 169.9 370.1 260.5 226.9
DMD with clustering
RSME 940.4 933.9 897.5 184.2 194.6 197.7 435.7 307.8 267.8
MAPE 4.92 4.88 4.53 8.2 6.84 7.74 4.94 3.39 3.43
DMDc with clustering MAE 716.2 733.9 697.2 147.9 152.9 156.4 354.3 246 225.6
RSME 875.3 895.7 834.1 172.3 183.2 183.1 414 292 265.8

Fig. 4. Statistical results of the MAPE values obtained for day forecasting of summer using five forecasting methods (a), (b) COAST zone, (c), (d) WCMA zone, (e), (f) Queensland
zone.

Comparison of the proposed DMDc-STLF with other STLF method, consumers behavior. Specifically, compared to DMD-STLF, the average
such as ARIMAX, SVR, and DMD implies its prediction accuracy espe-
cially in the areas where the climatic factors are more determining on daily load forecasting errors shows improvement of 21.64%, 15.55%

8
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

Fig. 5. Forecasting results of three days using proposed and benchmark methods for South-C datasets.

Table 3 models. The meaningful findings obtained demonstrate the high quality
The averaged MAPE of proposed and benchmark methods for South-C datasets.
performance of the proposed method and its low computational burden.
Forecasting method South-Central Nonlinearities in the underlying system, however, can challenge our
Weekend Monday assumptions on simply using the measurements directly in DMD [38].
SVR with clustering 4.09 5.75 Therefore, the idea of using kernel trick within the DMD architecture
ARIMAX 4.3 6.99 can be a suggestion for future research.
DMD with ACF 4.11 7.56
DMDc with clustering 3.69 6.96
CRediT authorship contribution statement

Table 4 Amir Mansouri: Conceptualization, Methodology, Software,

The averaged runtime of the proposed and benchmark methods for one-day ahead
Writing – original draft. Amir H. Abolmasoumi: Conceptualization,
forecasting.
Methodology, Validation, Writing – review & editing, Supervision.
Forecasting method South-Central
Running time (s)
Ali A. Ghadimi: Conceptualization, Methodology, Investigation,
Writing – review & editing.
SVR with clustering 15.5
ARIMAX 16.1
DMD with ACF 0.04 Declaration of competing interest
DMDc with clustering 0.1
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to
influence the work reported in this paper.
and 10.45%, for three datasets Electric Reliability Council of Texas, ISO
New England, and Australian Energy Market Operator respectively.
Data availability
As a result, using DMDc-STLF, the benefits of an accurate and
straightforward forecast are available to energy companies. It is impor-
Data will be made available on request.
tant to mention that in common forecasting methods such as neural
networks, a large amount of historical data is generally needed for
training, while the proposed method can work well with historical data References
with a small range. This can be very fruitful especially when access
[1] B.F. Hobbs, S. Jitprapaikulsarn, S. Konda, V. Chankong, K.A. Loparo, D.J.
to historical load or weather data is limited. Consequently, the DMDc- Maratukulam, Analysis of the value for unit commitment of improved load
STLF can be packed as a low-computation easy-to-implement software forecasts, IEEE Trans. Power Syst. 14 (4) (1999) 1342–1348.
with the capability of load prediction in different geographical area. [2] A. Abdelhay, P. Om, Load forecasting, in electric distribution systems, IEEE
(2019) 41–71.
6. Conclusions [3] E. Paparoditis, T. Sapatinas, Short-term load forecasting: The similar shape
functional time-series predictor, IEEE Trans. Power Syst. 28 (4) (2013)
3818–3825.
This paper proposed a data-driven method to improve the accuracy [4] R. Weron, Modeling and Forecasting Electricity Loads and Prices: A Statistical
of short-term electrical load forecasting based on the DMDc method Approach, John Wiley & Sons, 2007.
which is capable of including the impact of weather variables. By [5] J.W. Taylor, P.E. McSharry, Short-term load forecasting methods: An evaluation
utilizing the clustering theory, more appropriate and extensive train based on european data, IEEE Trans. Power Syst. 22 (4) (2007) 2213–2219.
[6] A. Tarsitano, I.L. Amerise, Short-term load forecasting using a two-stage sarimax
data were extracted for forecast. Then, using the DMDc, which is a model, Energy 133 (2017) 108–114.
method of identifying system dynamics in the presence of an external [7] A. Rafati, M. Joorabian, E. Mashhour, An efficient hour-ahead electrical load
force, temperature and humidity factors entered the load forecasting forecasting method based on innovative features, Energy 201 (2020) 117511.
problem using the real temperature. The results of the indicate an [8] N. Ding, C. Benoit, G. Foggia, Y. Besanger, F. Wurtz, Neural network-based model
design for short-term load forecast in distribution systems, IEEE Trans. Power
improvement in forecasting accuracy compared to the DMD method,
Syst. 31 (1) (2015) 72–81.
which has already been presented for short-term load forecasting. [9] A. Khwaja, X. Zhang, A. Anpalagan, B. Venkatesh, Boosted neural networks for
Also the proposed approach was compared with other widely known improved short-term electric load forecasting, Electr. Power Syst. Res. 143 (2017)
forecasting methods, such as the ARIMAX and the optimized SVM 431–437.

9
A. Mansouri et al. Electric Power Systems Research 221 (2023) 109387

[10] M. Barman, N.D. Choudhury, S. Sutradhar, A regional hybrid GOA-SVM model [40] P.J. Schmid, Dynamic mode decomposition of numerical and experimental data,
based on similar day approach for short-term load forecasting in Assam, India, J. Fluid Mech. 656 (2010) 5–28.
Energy 145 (2018) 710–720. [41] R.G. Steadman, The assessment of sultriness. Part I: A temperature-humidity
[11] A. Yang, W. Li, X. Yang, Short-term electricity load forecasting based on feature index based on human physiology and clothing science, J. Appl. Meteorol.
selection and Least Squares Support Vector Machines, Knowl.-Based Syst. 163 Climatol. 18 (7) (1979) 861–873.
(2019) 159–173. [42] L.P. Rothfusz, N.S.R. Headquarters, The Heat Index Equation (Or, More Than
[12] W. Kong, Z.Y. Dong, Y. Jia, D.J. Hill, Y. Xu, Y. Zhang, Short-term residential You Ever Wanted to Know About Heat Index), Vol. 9023, National Oceanic and
load forecasting based on LSTM recurrent neural network, IEEE Trans. Smart Atmospheric Administration, National Weather Service, Office of Meteorology,
Grid 10 (1) (2017) 841–851. Fort Worth, Texas, 1990.
[13] G. Memarzadeh, F. Keynia, Short-term electricity load and price forecasting by a [43] H.H. Aly, A proposed intelligent short-term load forecasting hybrid models of
new optimal LSTM-NN based prediction algorithm, Electr. Power Syst. Res. 192 ANN, WNN and KF based on clustering techniques for smart grid, Electr. Power
(2021) 106995. Syst. Res. 182 (2020) 106191.
[14] S.-V. Oprea, A. Bâra, Machine learning algorithms for short-term load forecast [44] I. Benítez, A. Quijano, J.-L. Díez, I. Delgado, Dynamic clustering segmentation
in residential buildings using smart meters, sensors and big data solutions, IEEE applied to load profiles of energy consumption from Spanish customers, Int. J.
Access 7 (2019) 177874–177889. Electr. Power Energy Syst. 55 (2014) 437–448.
[15] T. Hong, P. Pinson, Y. Wang, R. Weron, D. Yang, H. Zareipour, Energy [45] Z. Chen, Y. Chen, T. Xiao, H. Wang, P. Hou, A novel short-term load forecasting
forecasting: A review and outlook, IEEE Open Access J. Power Energy 7 (2020) framework based on time-series clustering and early classification algorithm,
376–388. Energy Build. 251 (2021) 111375.
[16] P. Schmid, J. Sesterhenn, Dynamic mode decomposition of numerical and [46] C. Fan, F. Xiao, S. Wang, Development of prediction models for next-day building
experimental data, in: Sixty-First Annual Meeting of the APS Division of Fluid energy consumption and peak power demand using data mining techniques,
Dynamics, San Antonio, Texas, USA, 2008. Appl. Energy 127 (2014) 1–10.
[17] P.J. Schmid, Dynamic mode decomposition of numerical and experimental data, [47] H. Hino, H. Shen, N. Murata, S. Wakao, Y. Hayashi, A versatile clustering method
J. Fluid Mech. 656 (2010) 5–28. for electricity consumption pattern analysis in households, IEEE Trans. Smart
[18] J.N. Kutz, S.L. Brunton, B.W. Brunton, J.L. Proctor, Dynamic Mode Grid 4 (2) (2013) 1048–1057.
Decomposition: Data-Driven Modeling of Complex Systems, SIAM, 2016. [48] S. Lu, G. Lin, H. Liu, C. Ye, H. Que, Y. Ding, A weekly load data mining approach
[19] C.W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, D.S. Henningson, Spectral analysis based on hidden Markov model, IEEE Access 7 (2019) 34609–34619.
of nonlinear flows, J. Fluid Mech. 641 (2009) 115–127. [49] H. Mori, A. Yuihara, Deterministic annealing clustering for ANN-based short-term
[20] J.N. Kutz, S.L. Brunton, B.W. Brunton, J.L. Proctor, Dynamic Mode load forecasting, IEEE Trans. Power Syst. 16 (3) (2001) 545–551.
Decomposition: Data-Driven Modeling of Complex Systems, SIAM, 2016. [50] A. Rajabi, M. Eskandari, M.J. Ghadi, L. Li, J. Zhang, P. Siano, A comparative
[21] Y. Susuki, I. Mezić, Nonlinear Koopman modes and coherency identification of study of clustering techniques for electrical load pattern segmentation, Renew.
coupled swing dynamics, IEEE Trans. Power Syst. 26 (4) (2011) 1894–1904, Sustain. Energy Rev. 120 (2020) 109628.
http://dx.doi.org/10.1109/TPWRS.2010.2103369. [51] W. Wu, M. Peng, A data mining approach combining 𝐾-Means clustering with
[22] B.W. Brunton, L.A. Johnson, J.G. Ojemann, J.N. Kutz, Extracting spatial– bagging neural network for short-term wind power forecasting, IEEE Internet
temporal coherent patterns in large-scale neural recordings using dynamic mode Things J. 4 (4) (2017) 979–986.
decomposition, J. Neurosci. Methods 258 (2016) 1–15. [52] B. Yildiz, J.I. Bilbao, J. Dore, A.B. Sproul, Recent advances in the analysis of
[23] J. Mann, J.N. Kutz, Dynamic mode decomposition for financial trading strategies, residential electricity consumption and applications of smart meter data, Appl.
Quant. Finance 16 (11) (2016) 1643–1655. Energy 208 (2017) 402–427.
[24] J.N. Kutz, X. Fu, S.L. Brunton, Multiresolution dynamic mode decomposition, [53] S. Yilmaz, J. Chambers, M.K. Patel, Comparison of clustering approaches for
SIAM J. Appl. Dyn. Syst. 15 (2) (2016) 713–735. domestic electricity load profile characterisation-Implications for demand side
[25] A.M. Avila, I. Mezić, Data-driven analysis and forecasting of highway traffic management, Energy 180 (2019) 665–677.
dynamics, Nature Commun. 11 (1) (2020) 2090. [54] C. Si, S. Xu, C. Wan, D. Chen, W. Cui, J. Zhao, Electric load clustering in smart
[26] T. Askham, P. Zheng, A. Aravkin, J.N. Kutz, Robust and scalable methods for grid: Methodologies, applications, and future trends, J. Mod. Power Syst. Clean
the dynamic mode decomposition, 2017, Preprint, arXiv:1712.01883v1. Energy 9 (2) (2021) 237–252.
[27] M.S. Hemati, C.W. Rowley, E.A. Deem, L.N. Cattafesta, De-biasing the dynamic [55] G. Gan, C. Ma, J. Wu, Data Clustering: Theory, Algorithms, and Applications,
mode decomposition for applied Koopman spectral analysis of noisy datasets, SIAM, 2020.
Theor. Comput. Fluid Dyn. 31 (4) (2017) 349–368. [56] O. Lauwers, B. De Moor, A time series distance measure for efficient clustering
[28] A.H. Abolmasoumi, M. Netto, L. Mili, Robust dynamic mode decomposition, IEEE of input/output signals by their underlying dynamics, IEEE Control Syst. Lett. 1
Access 10 (2022) 65473–65484. (2) (2017) 286–291.
[29] N. Mohan, K. Soman, S.S. Kumar, A data-driven strategy for short-term electric [57] F. Iglesias, W. Kastner, Analysis of similarity measures in times series clustering
load forecasting using dynamic mode decomposition model, Appl. Energy 232 for the discovery of building energy patterns, Energies 6 (2) (2013) 579–597.
(2018) 229–244. [58] T. Caliński, J. Harabasz, A dendrite method for cluster analysis, Comm. Statist.
[30] T. Namas, S. Dzakmic, I. Dzafic, Testing the feasibility of dynamic mode Theory Methods 3 (1) (1974) 1–27.
decomposition for short term load forecast, in: 2021 7th International Conference [59] U. Maulik, S. Bandyopadhyay, Performance evaluation of some clustering al-
on Electrical Energy Systems (ICEES), IEEE, 2021, pp. 229–233. gorithms and validity indices, IEEE Trans. Pattern Anal. Mach. Intell. 24 (12)
[31] J.R. Cancelo, A. Espasa, R. Grafe, Forecasting the electricity load from one day to (2002) 1650–1654.
one week ahead for the spanish system operator, Int. J. Forecast. 24 (4) (2008) [60] G.W. Milligan, M.C. Cooper, An examination of procedures for determining the
588–602. number of clusters in a data set, Psychometrika 50 (2) (1985) 159–179.
[32] C.-M. Huang, C.-J. Huang, M.-L. Wang, A particle swarm optimization to [61] Y. Zhang, W. Chen, R. Xu, J. Black, A cluster-based method for calculating
identifying the ARMAX model for short-term load forecasting, IEEE Trans. Power baselines for residential loads, IEEE Trans. Smart Grid 7 (5) (2015) 2368–2377.
Syst. 20 (2) (2005) 1126–1133. [62] A. Jain, B. Satish, Short term load forecasting by clustering technique based on
[33] C.-L. Hor, S.J. Watson, S. Majithia, Analyzing the impact of weather variables on daily average and peak loads, in: 2009 IEEE Power & Energy Society General
monthly electricity demand, IEEE Trans. Power Syst. 20 (4) (2005) 2078–2085. Meeting, IEEE, 2009, pp. 1–7.
[34] M. De Felice, A. Alessandri, P.M. Ruti, Electricity demand forecasting over Italy: [63] J. Foster, X. Liu, S. McLoone, Short-term load forecasting with high levels of
Potential benefits using numerical weather prediction models, Electr. Power Syst. distributed renewable generation, in: 2017 52nd International Universities Power
Res. 104 (2013) 71–79. Engineering Conference (UPEC), IEEE, 2017, pp. 1–6.
[35] K.-B. Song, S.-K. Ha, J.-W. Park, D.-J. Kweon, K.-H. Kim, Hybrid load forecasting [64] T. Hong, Short Term Electric Load Forecasting, North Carolina State University,
method with analysis of temperature sensitivities, IEEE Trans. Power Syst. 21 (2) 2010.
(2006) 869–876. [65] T. Hong, M. Shahidehpour, Load Forecasting Case Study, EISPC, US Department
[36] S. Ruzic, A. Vuckovic, N. Nikolic, Weather sensitive method for short term load of Energy, 2015.
forecasting in electric power utility of Serbia, IEEE Trans. Power Syst. 18 (4) [66] Electric reliability council of texas, 2022, Accessed 2022-04-30 https://www.
(2003) 1581–1586. ercot.com/gridinfo/load/load_hist/.
[37] A. Selakov, D. Cvijetinović, L. Milović, S. Mellon, D. Bekut, Hybrid PSO– [67] Recently published demand data, ISO New England, https://www.iso-ne.com/
SVM method for short-term load forecasting during periods with significant isoexpress/web/reports/load-and-demand/-/tree/zone-info.
temperature variations in city of Burbank, Appl. Soft Comput. 16 (2014) 80–88. [68] Aggregated price and demand (AEMO), https://aemo.com.au/energy-systems/
[38] J.N. Kutz, S.L. Brunton, B.W. Brunton, J.L. Proctor, Dynamic Mode electricity/national-electricity-market-nem/data-nem/aggregated-data.
Decomposition: Data-Driven Modeling of Complex Systems, SIAM, 2016. [69] Local climatological data, national centers for environmental information, https:
[39] J.L. Proctor, S.L. Brunton, J.N. Kutz, Dynamic mode decomposition with control, //www.ncdc.noaa.gov/cdo-web/datatools/lcd.
SIAM J. Appl. Dyn. Syst. 15 (1) (2016) 142–161. [70] Weather underground, https://www.wunderground.com.

10