Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Sustainable Energy, Grids and Networks 36 (2023) 101191

Contents lists available at ScienceDirect

Sustainable Energy, Grids and Networks


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

Assessment of hybrid transfer learning method for forecasting EV profile


and system voltage using limited EV charging data
Paul Banda, Muhammed A. Bhuiyan, Kazi N. Hasan *, Kevin Zhang
School of Engineering, RMIT University, Melbourne, VIC 3000, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: The number of electric vehicles (EV) is increasing exponentially, significantly affecting the planning and oper­
Electric vehicle ation of future electricity grids, albeit the availability of EV data is very limited to perform power system studies.
Forecasting To overcome the poor forecasting accuracy associated with limited available data, this research proposes a time-
Transfer learning
series-based hybrid transfer learning forecasting approach, namely CNN-BiLSTM (Convolutional Neural Network
Time series
Voltage profile
– Bidirectional Long Short-Term Memory) to forecast EV charging profile. Additionally, the proposed algorithm
has been applied to forecast network voltage from the EV data without performing power flow. EV charging
demand datasets collected over a year for residential, slow commercial, and fast commercial charging stations
and their corresponding voltage profiles have been used to test the effectiveness of the proposed hybrid transfer
learning framework. The results confirm the improved accuracy of the proposed hybrid CNN-BiLSTM model
compared to the conventional CNN model, newly created models in predicting the EV charging demand and
voltage profiles.

1. Introduction separable-stationary data if the data is adequately prepared and the


model is appropriately tuned [4].
The number of electric vehicles (EVs) adoption is expected to in­ On the other hand, AI-based methods involve using computers to do
crease exponentially in the next decade due to favourable government tasks that typically require human intelligence, such as learning and
policies, innovation and incentives [1]. Many countries in the world are pattern recognition. The AI-based techniques include genetic program­
still in the race for early adoption of EVs. As the number of EVs increases, ming (GP), machine learning, reinforcement learning [5] and deep
the EV charging demand will increase, and hence, it would be chal­ learning-based methods, such as LR (Linear Regression), SVR (Support
lenging to manage the electricity grid with a high penetration of EVs [2]. Vector Regression), CNN (Convolutional Neural Network), and LSTM
Moreover, secure operation of the power grid would be challenging with (Long Short-Term Memory) methods [6]. The EV charging demand
the deployment of large capacity fast EV charging stations [3]. There­ forecasting has been conducted by using fuzzy clustering and back­
fore, forecasting the EV charging demand and estimating their likely propagation neural network models [7], evolutionary optimisation
impact on the grid is critical for the secure operation of the electricity techniques [8], and various data mining techniques [9]. Early detection
grids. of cyber attack on fast EV charging stations has been demonstrated by
using machine-learning approach [10]. Furthermore, decision trees,
multi-layer perceptron, and support vector machines have been imple­
1.1. Overview of the forecasting techniques mented for EV charging demand forecasting [9].

Data-driven forecasting techniques can be broadly classified into two


categories: (1) traditional time series (TS) models and (2) artificial in­ 1.2. Applications of deep and hybrid deep learning models
telligence (AI)-based techniques. The TS-based methods include ARMA
(Autoregressive Moving Average), ARIMA (Autoregressive Integrated The deep learning methods are empowered by unprecedented
Moving Average), and SARIMA (Seasonal Autoregressive Integrated learning ability from extensive data analysis for solving challenging
Moving Average) models. Classical TS models perform best for linearly forecasting tasks [11]. These methods [12,13] have shown outstanding

* Correspondence to: School of Engineering, RMIT University, 402 Swanston Street, Melbourne, VIC 3000, Australia.
E-mail address: kazi.hasan@rmit.edu.au (K.N. Hasan).

https://doi.org/10.1016/j.segan.2023.101191
Received 19 May 2023; Received in revised form 29 August 2023; Accepted 15 October 2023
Available online 24 October 2023
2352-4677/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-
nc/4.0/).
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

performance in pattern recognition [14], image classification, and networks will be significantly affected. The low-voltage network is
medical sciences [15–17]. spread over a wide geographical area that spans the street level – home-
There are applications of a hybrid lion algorithm, which is a com­ to-home. It is impossible to perform power flow studies for such a large
bination of CNN and LSTM, that has been used for EV charging load network to maintain the operational limits (as defined by the grid
forecasting [11,17]. Estimation of the EV import/charging and codes). In such a case, the transfer learning approach could be per­
export/discharging capability to the grid has been identified by using formed to analyze the EV profiles for predicting the network voltage
deep learning and attention mechanism [18], and deep learning and profiles. This approach will be used for screening purposes to identify
NANN (nonlinear autoregressive neural network) [19]. The bidirec­ the likely voltage problems in the network. Once the vulnerable part of
tional LSTM has shown superior performance over the unidirectional the network is identified, a power flow simulation could be performed
LSTM in applications such as phoneme classification [20], speech for the selected part of the network where the voltage problem is
recognition [21], stock market forecasting [22], and text classification identified by the transfer learning approach. This sequential approach
[23]. Studies implementing different architectures of the hybrid will reduce the computational requirement and time for the whole
CNN-BiLSTM have shown effective results in the speech recognition network simulation.
domain [24,25]. EV charging demand data belongs to the time-series Therefore, this research presents the following significant
domain, which has a different data structure (1-dimensional) to contributions:
speech recognition (2-dimensional); thus, it requires a different data
handling protocol. • Exploring feature extraction strategies using the hybrid CNN-BiLSTM
transfer learning model’s architectural variations to apply for (resi­
1.3. Applications of transfer learning models dential to commercial and vice versa) EV charging demand predic­
tion with limited available data.
The advancement in the machine learning approaches leads to • Demonstration of network voltage profile prediction by using the
transfer learning "that focuses on storing knowledge gained while hybrid CNN-BiLSTM transfer learning models. This would enable
solving one problem and applying it to a different but related problem". quick prediction of the system voltage without performing a power
Such as the knowledge gained while learning to ’forecast residential EV flow study.
charging profile’ could be applied to ’forecast commercial EV charging
profile’. The advantage of the transfer learning approach is that it can The rest of the paper is organised as follows: Section 2 discusses the
predict the system voltage profile by analysing the EV charging profiles proposed transfer learning modelling approach, Section 3 provides a
(without running a power flow solution), which may provide quick in­ highlight of the dataset, Section 4 outlines the results and discussion,
puts to system operational decisions. and Section 5 presents the conclusions and future works.
Transfer learning has attracted attention in image processing [26],
natural language processing tasks [27],[28], time-series classification 2. Proposed transfer learning modeling approach
[29], and building energy forecasting [30]. These studies have mostly
applied the CNN model. Two encoder-decoder hybrid models were 2.1. Transfer learning
proposed for parameter-sharing neural network transfer learning to
predict hourly residential EV energy using limited data [31]. In [31], the Transfer learning is the process of using information gathered during
authors proposed two hybrid transfer learning frameworks that out­ a task (TS ) from a source domain (DS ) to improve a task (TT ) in the target
performed the conventional CNN model for EV charging demand pre­ domain (DT ). Transfer learning differs from conventional machine
diction using limited data. An extension to [31], in this study we learning approaches that assume similar data distributions in the
explore; (i) the effect of different feature extraction strategies to deter­ training and test datasets. The goal of transfer learning is to reduce the
mine whether freezing only the CNN module or freezing both CNN and amount of data and computation costs associated with the creation of
BiLSTM modules and retraining the last dense layer can lead to different customised data-driven models.
results in EV charging demand prediction, and (ii) the CNN-BiLSTM The transfer learning formal expression using a single source and
transfer learning is tested for voltage forecasting using limited data target domain is given as: a domain comprises feature space X and label
and its performance is compared against its new and retrained space Y, thus the source and target domains are expressed as in (1) and
CNN-BiLSTM models. (2), respectively.
Ds = {(xs1 , ys1 ), (xs2 , ys2 ), ⋅⋅⋅⋅⋅(xsi , ysi ) (1)
1.4. Contributions of the study
DT = {(xT1 , yT1 ), (xT2 , yT2 ), ⋅⋅⋅⋅⋅(xTi , yTi ) (2)
The limited availability of the EV charging demand dataset is a
barrier to accurate EV charging profile prediction, which adversely af­ where, xsi and ysi represent the input and corresponding output data
fects power systems. This study proposes a solution to this latter problem instances in the source domain, respectively. Similarly, xTi and yTi are
by using a hybrid transfer learning strategy that improves EV charging input and output instances from the target domain.
demand and voltage profiles’ prediction with a limited dataset using the Four major approaches can be used to transfer knowledge learnt
knowledge learned from a data-rich EV charging/voltage dataset (such from the source domain, namely feature-based, instance-based,
as ’forecasting residential EV charging profile’). Dependencies in the parameter-based, and relation-based strategies [32,33]. The
charging demand or voltage profiles can be leveraged to forecast a feature-based approach common with heterogeneous transfer learning
related dataset (such as ’forecasting commercial EV charging profile’ or aims to create coincident variables through feature transformation
’forecasting system voltage profile caused by the residential EV charging processes [33]. The instance-based method utilises weighting tech­
profile’). niques to weight source data. It assigns greater weights to data sources
To the best of our knowledge, no study has applied the CNN-BiLSTM more similar to the target domain. The parameter-based transfer
hybrid transfer learning model to predict EV loads and subsequent learning reprocesses the model parameters learnt from the source
network voltage profiles. This prediction study would benefit future grid domain in the target domain. Deep learning models are easily adaptable
planning and operational decision-making while the scarcity of EV data to these techniques and have thus seen a more widespread application. A
exists. limited amount of data in the target domains is utilised for fine-tuning
With the rapid integration of EVs into the grid, the total energy de­ the model. This study has adopted this fine-tuning-based
mand will increase, and subsequent voltage profiles of the low-voltage

2
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

parameter-sharing approach for all transfer learning experiments. gate, output gate, and forget gate. The three gates control the flow of
Relational-based transfer learning primarily handles relational data, information from and into the memory cell [35]. The input gate man­
where logical relationships and rules are transferred across domains. ages input activations into the memory cells, while the output gate
Fig. 1 shows a schematic representation of transfer learning, showing controls output activations. The forget gate works to keep useful infor­
how knowledge transfer differentiates it from the customised traditional mation from the previous timestep while filtering out irrelevant
machine learning workflow. In machine learning, EV demand prediction information.
is per individual datasets (1 and 2). However, in transfer learning, The hidden vector (ht ) in LSTM, which differentiates it from the RNN
knowledge learnt while processing EV dataset 1 is transferred in pre­ (recurrent neural network), can be obtained using the (5− 9) below [36].
dicting EV dataset 2, as seen in Fig. 1. This knowledge can be in the form
it = σ (Wxi xt + Whi ht− 1 + Wci ct− 1 + bI ) (5)
of model parameters, features, data instances etc. The procedure of the
transfer learning is shown in Fig. 2.
ot = σ ( Wxo xt + Who ht− 1 + Wco ct + b0 ) (6)
( )
ft = σ Wxf xt + Whf ht− 1 + Wcf ct− 1 + bf (7)
2.2. Deep learning-based models
ct = ft ct− 1 + it tanh( Wxc xt + Whc ht− 1 + bC ) (8)
The deep learning models employ hierarchical processing using
many layers of architecture and extract features directly from the data ht = ot tanh(ct ) (9)
without the need for manual feature extraction [12]. Two deep
learning-based models implemented in this study are discussed below. where it, ot, and ft represent the input gate, output gate, and forget gate
values, respectively. Ct is the memory cell, bi, bo, bf, and bc are the cor­
2.2.1. Convolutional neural network (CNN) responding bias values, and σ is the sigmoid function. Wx(i,o,f) , Wh(i,o,f) )
In performing the EV demand/voltage forecasting, a predetermined and Wc(i,o,f) represents the weights between input nodes and hidden
number of EV demand and system voltage data are fed into the con­ nodes, weights between hidden nodes and memory cells, and weights
volutional layers. In the first step, a one-dimensional convolution of this between the memory cell and output nodes, respectively. The network
data is performed. Then in the next step, the final processing would be output yt is then obtained as expressed in (10) below [36].
performed by passing the output data from the convolution operation to
the fully connected layers. A CNN operation is expressed, as shown in (3) yt = Why ht + by (10)
[34]. in which,
s = (x ∗ w) (3) ht = F(Wxh xt + Whh ht− 1 + bh ) (11)
where, s is the output (feature map) of the convolution operation, x
represents the input function, and w denotes the weighting function where xt and yt represent the input and output at time step t, ht and ht− 1
(kernel). The vector representing the predicted next hour can be are the hidden states at the time step t and t-1, respectively, Wxh , Whh and
expressed as (4). Why are the weight matrices between input and hidden vectors,
different time steps of the hidden vectors and hidden vectors to the
̅→
y = yt + yt+1 (4) output vectors, bh and by represents the corresponding bias vectors of
̅→ weight matrices Whh and Why , respectively. F represents the nonlinear
Where y t denotes the predicted EV charging demand at the given activation function for hidden nodes, for which the rectified linear unit
timestep t. function has been implemented [37].

2.2.2. Long short-term memory (LSTM)


The basic unit of the LSTM block comprises the memory block, input

EV dataset 1 Machine Learning


Model I
EV demand
Knowledge prediction I
Transfer

EV dataset 2 Machine Learning


Model II
EV demand
prediction II

Fig. 1. Schematic layout of the operations of transfer learning.

3
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

Compare
predictions
Target domain
EV demand/Voltage Predict using same EV dataset
dataset
Target EV
prediction Target Ev
label prediction

Collect and
Preprocess Source
Train on andCNNBiLSTM
target domainRegression
temporal Hybrid
Analyse predictor:

Fine tuning
feature
data
data dataset and calendar
extractor inputs
dense layer model fine
tuning
Source EV
prediction Training
label pretrained
hybrid model
Source domain
EV demand/ voltage
dataset Target EV
test dataset

Fig. 2. Transfer learning workflow using the proposed CNN-BiLSTM model.

2.2.3. Bidirectional LSTM (BiLSTM) CNN and BiLSTM models discussed above with a fully connected layer
The operation of the general LSTM works in such a way that remains (artificial neural network), as shown in Fig. 3. It is a hybrid model that
unidirectional. In case of the bidirectional LSTM (BiLSTM), it has two accumulates the powerful feature extraction capabilities of the 1D- CNN
groups of EV sequence – one in the positive time direction and another in module and excellent temporal sequence modelling traits of the LSTM,
the negative time direction. However, the hidden layers of these two which are processed bidirectionally. The 1D-CNN with a BiLSTM model
groups are independent of each other. In such a process, the outputs are aims to solve the vanishing gradient problem. In this hybrid system, a
concatenated and linked to the same output layer. convolutional operation is done by the encoder 1D-CNN module
In the BiLSTM architecture, the underlying mathematical represen­ responsible for interpreting the input sequence from the EV charging
tation is the same as that of the unidirectional LSTM, however the demand/voltage time series. The CNN encoder model outputs a
BiLSTM creates two hidden states at timestep t. The hidden steps are the sequence which is then passed on for interpretation to the BiLSTM model
f
forward and backward hidden states, which are presented by, Ht and Hbt , (decoder) and further temporal sequencing in forward and backward
f directions. The LSTM usually remember the short-term and long-term
respectively. The hidden states Ht and Hbt constitutes the final hidden
temporal sequences; thus, by combining the CNN and BiLSTM, which
state, as follows [38]:
automatically learns features from sequence data, the hybrid CNN-
Ht = H ft + H bt (12) BiLSTM model offers improved accuracy.
The CNN-BiLSTM architecture for voltage prediction is discussed as
the one for EV charging demand is similar to the one implemented in
2.3. Proposed hybrid deep CNN-BiLSTM inspired transfer learning [31] as this work extends this previous work. Fig. 3 illustrates the hybrid
CNN-BiLSTM used for EV voltage-transfer learning. The hybrid
The proposed hybrid CNN-BiLSTM model is developed by fusing CNN-BiLSTM expect the input EV demand data to have a 3D form

EV database

Fixed Feature extraction Fixed Sequence learning

Max
pooling Bi-LSTM Bi-LSTM

Bi-LSTM Bi-LSTM

Data Bi-LSTM Bi-LSTM


Pre-
Process Conv. Conv.
…….

ing Conv. (32 filters) (16 filters) ……...


(64 filters) Bi-LSTM Bi-LSTM

Flattening
Dense
Convolution & pooling operation BiLSTM operation layer

Fig. 3. Implemented CNN-BiLSTM model for EV voltage prediction and transfer learning.

4
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

(sample, input, and timestep). The simulated system voltage with and ∑
without EV penetration are input into the CNN encoder architecture for (y − y′)2
R2 = 1 − ∑ − (15)
reading. The first convolutional layer reads across this input sequence (y − y)2
using a filter of size three timesteps (1×3) and then projects its output
onto 32 feature maps. The second convolutional layer performs a similar 3. Simulation procedure and data collection
operation on the previous layer’s output in an attempt to expose more
salient features. 3.1. Simulation procedure
The max-pooling layer simplifies the feature map sizes while pre­
serving the most significant signal values. The final extracted features The conceptual framework of the transfer learning simulation pro­
maps from the max-pooling layer are flattened into a vector. The vector cedure has been presented in Fig. 4. The EV charging stations dataset
is passed to the BiLSTM decoding module, which processes the cell state from many residential, slow commercial, and fast commercial EV
forward and backwards before passing the output to the fully connected charging stations have been deposited into a data repository. These
layer for forecasting. The decoding BiLSTM model was defined as a datasets are then used for simulation and validation.
single layer BiLSTM model with 100 neurons. The fully connected layer A power flow simulation has been performed in a parallel work­
with 30 neurons interpreted the features learnt by the BiLSTM layer, and stream using an IEEE test network. The error (RMSE) of the forecasting
lastly, the output layer predicts the next hour EV voltage value. has been calculated for all scenarios.
This study highlights the limitations of forecasting with a limited
number of datasets. Hence, it proposes a new hybrid transfer learning
2.4. Input processing
model for EV charging and voltage prediction to overcome that limita­
tion. Therefore, the trajectory of the error with respect to the granularity
The CNN-BiLSTM model takes the EV time series data and performed
of the datasets has been presented in the results.
supervised learning. The sliding window method is used to reframe the
This section outlines transfer learning implementation using the
EV sequence into input-output variable pairs [39]. The inputs of the
discussed hybrid model in the preceding section. The first procedure
algorithm is the current EV observation at time t and previous EV ob­
involves forecasting residential EV/voltage with a full data complement
servations at times t1, t2, ..., tn, which predict the output EV observation
in the source domain and saving the weights of this model. To do this,
at time t + 1. The length of the sliding window is defined based on [39].
data is pre-processed, which includes data cleaning, identifying outliers
and filling in missing data and normalisation to allow efficient training.
The saved model is then re-loaded and initialised using saved weights
2.5. Performance evaluation criteria
during training (fine-tuning) with limited data from the target domain
(slow EV/voltage and fast EV/voltage). Test data from the target domain
The assessment of the accuracy of the models has been evaluated
is used to validate the pre-trained models, and performance is measured
using the root means square error (RMSE), mean absolute error (MAE),
using the evaluation metrics discussed above. The test data sizes range
and R2 evaluation metrics, as presented in (13− 15), where y is the
from almost 0% to full data complement. The selected data size ranges
observed EV charging load, y′ is the predicted EV charging load, and N is
allow for tracking model performance with increasing data size. The
the number of data pairs considered.
data size is limited to less than 50%, which presents challenges for
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅

√1 ∑ N training the new model (limited data).
RMSE = √ [y − y′]2 (13)
N i=1
3.2. Electric vehicle (EV) dataset
1 ∑N
MAE = |y′ − y| (14) The EV charging demand dataset is collected over an entire year from
N i=1
01/09/2017–31/08/2018, equating to 8760 data points, representing
hourly EV charging power (in kilowatt, kW) of a year [40]. The three

Power system approach

Residential EV System
voltage
profile
EV data Power Flow
repository
Commercial slow EV Data analytics approach

EV Voltage
forecasting forecasting
Transfer Learning
Commercial fast EV
Fig. 4. Conceptual framework of the transfer learning approach for network voltage prediction from EV charging dataset.

5
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

types of EV charging station data were collected from (i) aggregated commercial EVs and vice versa.
residential [41], (ii) slow commercial [40], and (iii) fast commercial The initial training results of the three models on commercial EV
[40] charging stations. The database records 8760 hourly data points forecasting and voltage forecasting before transfer learning are shown in
throughout the entire annual cycle. Table 1 describes the EV charger Table 2. Both the conventional CNN and hybrid CNN-BiLSTM models are
types and statistical parameters of the dataset. As a single residential EV evaluated during EV forecasting. For the subsequent voltage-transfer
profile is highly random, collective residential EV profiles were created learning experiments, only the proposed CNN-BiLSTM model is
by grouping data from 100 single residential profiles. The slow com­ considered and compared against a newly developed and pre-trained
mercial profile is collected from a shopping centre. Fast commercial EV hybrid CNN-BiLSTM. The expectation is for the transfer-based model
charging station profiles were created from roadside charging station. to outperform both the new and pre-trained models. Outperforming the
The weekday-hourly EV charging demand and the corresponding new models will validate the need for transfer learning, and out­
voltage profiles are given in Fig. 5 and Fig. 7. The residential EV performing the pre-trained model demonstrates that the data distribu­
charging demand is the highest during weekdays between 5 p.m. and tions of the source and target datasets are different, which is a
8 p.m., which coincides with the time most people come home from prerequisite for transfer learning.
work. The lowest EV demand is between 4 a.m. and 8 a.m., when most Algorithms have been implemented in the Python programming
people will have stopped charging their EVs. Slow commercial EV language. The deep learning models have been built using the Keras
charging demand is highest midweek around 1 p.m., as most people use (https://keras.io/) library with Tensorflow (https://www.tensorflow.
their lunch break for a quick charge. The lowest EV demand is on org/) backend. All experimental tasks have been conducted on a macOS
weekends for slow commercial EV stations. Fast commercial EV demand Mojave (2.90 GHz Intel Core i9 16 GB 2400 MHz DDR4) machine.
shows an irregular demand pattern, with peak demand experienced on
weekdays and weekends. The irregular patterns reflect the random na­ 4.1. Model definitions
ture in which people charge their vehicles since fast EV charging stations
provide fast charging. The irregular charging patterns cause challenges The baseline transfer learning-based model (CNN-T) is compared
during both modelling and transfer learning. In contrast, voltage profiles against the proposed hybrid CNN-BiLSTM-T transfer learning-based
in Fig. 7 show the lowest and highest voltage values at the exact times model to determine which model is superior for the transfer learning
when the charging demand was highest and lowest, respectively. More task. The CNN-BiLSTM-Dt model is a pre-trained model created in the
details on this dataset are given in [31]. source domain (Table 2) that directly predicts the target dataset and has
no access to the target dataset. The performance of the CNN-BiLSTM-Dt
3.3. System voltage profiles is expected to be unstable as it does not utilise any data from the target
domain and thus does not account for the differences in data distribution
Power flow studies have been performed using the developed EV between the source and target datasets. The newly created CNN-N and
profiles in the IEEE 33 bus test network (Fig. 6) to obtain the voltage CNN-BiLSTM-N models are non-transfer learning newly created models
profiles. To illustrate the impact of the consumption of EV charging compared against the earlier defined transfer learning models. If a new
stations on the system voltage profile, power flow simulation has been model without transfer learning can outperform the transfer learning-
performed in MATPOWER [42], followed by data processing in MAT­ based models, we can conclude that transfer learning may not be
LAB. Fig. 7 presents the voltage profiles of the power network based on necessary; thus, we may need to fine-tune the transfer learning model to
the 30% EV penetration from residential, slow, and fast EV, respectively, perform better or abandon it. Tested models’ accuracies are evaluated
which is the average of one-year data. The details of the voltage profile and compared using the RMSE, MAE, and R2 metrics.
analysis have been presented in [43].
The challenges of voltage forecasting from the EV data are well un­ 4.2. EV charging forecasting using hybrid transfer learning
derstood. The voltage is a local phenomenon of a power network
depending on the network topology, feeder characteristics, system The next-hour residential EV charging demand forecast has been
loading and several other parameters. This study applies a transfer implemented as discussed in [31] using the previous 47 h of historical
learning approach to forecast EV charging demand and voltage profiles. residential EV charging demand and calendar inputs, namely, the hours
Hence, this voltage forecasting will capture the general trend of EV of the day, the days of the week, the weeks of the year, the months of the
loading on the network and its link with system voltage. year, the quarters of the year, weekend or not weekend, the days of the
month, and the days of the year.
4. Results and discussion The data has been split into 70% training, 10% validation, and 20%
testing while maintaining the temporal order of the residential EV de­
Transfer learning is tested in two phases; firstly, from residential EV mand time series. The developed models have been implemented to
(source domain) to fast commercial (target domain) and then to slow perform forecasts for the next hour. Input data is standardised using the
commercial (target domain). The aim is to enhance forecasting in the min-max scaler function [44].
target datasets with limited labelled data. Secondly, transfer learning is The tested models demonstrated similar performance in predicting
tested for voltage forecasting (that is influenced by high EV charging) the next-hour charging load for the residential EV (source domain), and
using transformed EV datasets. In this experiment, transfer learning is the results are shown in Table 2. The CNN and CNN-BiLSTM models
tested on the voltage profiles caused by residential EVs to fast have comparable RMSE scores of 27.87 kWh and 27.17 kWh.

Table 1
Different Types of EV Charging Stations and Their Information [40,41].
EV charging EV charger type Nonzero values Maximum Mean Median Mode
stations (%) (kW) (kW) (kW) (kW)

Residential In-home plug-in electric vehicle recharging profiles associated with 100 8739 (99%) 357 130 46
households. 131
Slow Shopping Centre. Free to park and charge. 8402 (95%) 17 0.55 00
Commercial 1.33
Fast Roadside charging. Park and ride. 1455 (16%) 107 00 00
Commercial 4.77

6
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

Fig. 5. Weekly hourly EV charging demand patterns for residential, slow commercial, and fast commercial charging stations.

Fig. 6. Single line diagram of the IEEE 33-bus radial distribution system.

Fig. 7. Weekly hourly voltage profiles for residential, slow commercial, and fast commercial charging stations.

Table 2 dominates the baseline CNN-T model in both limited data and adequate
Testing phase results for best-selected models for the residential EV and voltage data availability situations at both tested transfer learning scenarios.
forecasting. The results in Table 3 have also been published in the previous paper
Model RMSE MAE R2 [31]. However, to smoothen the reader’s understanding and for refer­
EV forecasting
ence to the following discussed sections, we briefly summarise the
CNN 27.87 22.13 0.86 results.
CNN-BiLSTM 27.17 21.25 0.87
Voltage forecasting 4.2.2. Comparing the different hybrid model architectures on transfer
CNN-BiLSTM-Residential 0.006 0.005 0.79
learning
CNN-BiLSTM-Fast commercial 0.006 0.004 0.33
CNN-BiLSTM-Slow commercial 0.008 0.006 0.25 In this study segment, an investigation was set up to determine which
feature extraction strategy is more effective for EV charging prediction
using transfer learning. Precisely, two feature extraction strategies were
4.2.1. Performances for EV forecasting using transfer learning tested, freezing only the encoder module (CNN part of the hybrid model)
In this section, a comparison is made between two transfer learning while retraining the dense layer. The second strategy involves freezing
methods to show which transfer learning method performs better. the weights of both the encoder and decoder modules and retraining
Table 3 shows the findings from this trial. only the fully connected layers (dense layer) weights using target data.
When comparing the transfer learning-based models, it is observed The investigation is tested on the residential EV to slow commercial EV
that the hybrid-based transfer learning models outperform the baseline and residential EV to fast commercial EV transfer learning scenarios, and
CNN-T model at most tested data sizes (17 out of 20 instances). The the results are shown in Table 4.
latter points to the effectiveness of the hybrid model’s architecture in According to Table 4, for the residential to slow commercial transfer
enhancing transfer learning. The hybrid CNN-BiLSTM-T model learning scenario, freezing both the CNN and BiLSTM modules resulted

7
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

Table 3
Comparison of the best performing transfer learning methods.

Residential EV to slow EV Residential EV to fast EV


Data
CNN-T CNN-BiLSTM -T CNN-T CNN-BiLSTM -T
size
RMSE MAE RMSE MAE RMSE MAE RMSE MAE
12 44.87 39.23 0.08 0.07 22.64 18.54 42.9 36.33
24 0.26 0.25 0.39 0.39 57.77 51.74 32.29 29.58
36 1.05 1.04 0.42 0.41 44.24 36.35 19.7 18.7
48 2.56 2 0.82 0.04 45.77 39.07 42.14 36.02
60 0.27 0.26 0.24 0.19 27.13 19.05 12.82 10.21
72 1.69 1.5 1.43 1.11 33.3 26.81 22.51 17.92
96 2.17 1.21 2.32 1.32 27.67 20.95 18.73 15.24
120 2.37 1.58 2.27 1.62 25.18 20.06 22.95 18.01
1200 1.25 0.65 1.4 0.62 28.03 21.38 25.9 20.28
6961 1.99 1.17 1.38 0.98 26.2 20.13 24.58 19.61

Table 4
Comparing the transfer learning strategies.

Residential to slow commercial EV Residential to fast commercial EV


Data Freeze CNN Freeze CNN + Freeze CNN Freeze CNN +
size only BiLSTM only BiLSTM
RMSE MAE RMSE MAE RMSE MAE RMSE MAE
12 0.15 0.13 0.08 0.07 29.06 29.06 42.9 36.33
24 0.66 0.66 0.39 0.39 47.57 47.57 32.29 29.58
36 0.06 0.05 0.42 0.41 48.94 48.94 19.7 18.7
48 0.8 0.45 0.82 0.04 101.59 101.59 42.14 36.02
60 0.1 0.09 0.24 0.19 40.25 40.25 12.82 10.21
72 1.56 1.25 1.43 1.11 21.88 21.88 22.51 17.92
96 2.87 1.69 2.32 1.32 22.15 22.15 18.73 15.24
120 3.77 2.62 2.27 1.62 25.11 25.11 22.95 18.01
1200 1.51 0.8 1.4 0.62 26.6 26.6 25.9 20.28
6961 1.47 1.03 1.38 0.98 26.74 26.74 24.58 19.61

in lower error scores except for data size instances 36, 48 and 60 (3 out This approach is adopted for transfer learning in voltage forecasting
of 10). In the residential to fast commercial transfer learning scenario, 8 tasks discussed below.
out of 10 instances had lower error scores if the CNN and BiLSTM
modules were frozen and only the dense layer was retrained. The results 4.3. Voltage forecasting using hybrid transfer learning
generally show better performance if both the CNN and the BiLSTM
model’s weights were frozen while retraining the dense layers weights The next-hour voltage forecasting was the most challenging for the
with the target dataset. These results further support the importance of a tested models. This can be attributed to many electrical appliances
hybrid model in enhancing transfer learning compared to using the in­ having different electrical usage signatures (at different times). As such,
dividual CNN model. The combined encoder and decoder modules act as they present significant challenges during modelling. As seen in Table 2,
feature extractors in this architecture and have better information predicting residential voltage profiles (source domain) was relatively
abstraction capabilities than a single CNN feature extractor. The com­ easier for the CNN and CNN-BiLSTM models, while the fast and com­
bined frozen CNN and BiLSTM feature extraction strategy is adopted for mercial EV voltage profiles were particularly challenging. The R2 scores
voltage forecasting and subsequent transfer learning experiments. The for predicting the next hour of voltage profile for residential, fast, and
dark colour in Table 4 represents the frozen layers of the CNN-BiLSTM slow voltage profiles are 0.79, 0.35, and 0.23, respectively. The poor
model. performance during fast and slow voltage prediction is expected to
In the residential EV to fast commercial EV transfer learning negatively impact transfer learning results since the models were not
schemes, the most influential architecture involves freezing the weights adequately trained.
of both the encoder and decoder models and retraining the dense layer. The initial transfer learning test for voltage forecasting involved
The results confirm this strategy’s effectiveness as most data size in­ comparing the transfer learning-based CNN-BiLSTM-T model’s perfor­
stances (8 out of 10) outperformed the other approach (freezing the CNN mance against the newly created CNN-BiLSTM-N model. The rationale is
weights only) except at data instances 12 and 72. In the residential EV to that a transfer learning-based model should better predict the target
slow commercial EV scheme, 7 out of 10 findings confirm the superiority dataset than a newly developed model since it uses ’knowledge’ learned
of freezing both the encoder and decoder weights during the training of from the source dataset. The results from this experiment are given in
hybrid models. While no clear trend could be observed, freezing both the Table 5 and illustrated in Fig. 8.
encoder and decoder modules and retraining the dense layers is
preferred when using hybrid architectures in transfer learning settings.

8
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

Table 5
RMSE Comparison of a new and transfer learning-based models for voltage forecasting.

Scenario 12 24 36 48 60 72 84 96 108 120 132 200 250 300 450 500 550 600 800 1200 4200 6200
CNN-
Residential EV commercial EV
Fast commercial Residential EV

0.334 0.548 0.363 0.047 0.041 0.032 0.077 0.014 0.017 0.018 0.017 0.023 0.019 0.038 0.008 0.005 0.005 0.005 0.007 0.014 0.018 0.014
BiLSTM-T
CNN-
to fast

0.749 0.822 0.064 0.139 0.129 0.232 0.312 0.257 0.196 0.125 0.165 0.068 0.074 0.057 0.010 0.013 0.005 0.022 0.018 0.018 0.008 0.006
BiLSTM-N
CNN-
0.031 0.018 0.024 0.035 0.031 0.035 0.030 0.033 0.033 0.034 0.033 0.034 0.034 0.034 0.034 0.033 0.033 0.033 0.032 0.033 0.048 0.055
BiLSTM-Dt
CNN-
0.499 0.594 0.465 0.024 0.053 0.020 0.083 0.028 0.035 0.025 0.043 0.015 0.015 0.023 0.017 0.013 0.016 0.016 0.018 0.012 0.006 0.011
BiLSTM-T
CNN-
EV to

0.337 0.436 0.095 0.010 0.085 0.221 0.080 0.043 0.033 0.040 0.055 0.042 0.027 0.046 0.026 0.027 0.011 0.022 0.036 0.008 0.006 0.008
BiLSTM-N
CNN-
0.077 0.059 0.058 0.060 0.059 0.067 0.062 0.058 0.059 0.057 0.055 0.058 0.059 0.055 0.061 0.060 0.059 0.060 0.054 0.058 0.055 0.053
BiLSTM-Dt

Residential EV to Fast Commercial EV Forecasting Fast Commercial EV to Residential EV Forecasting


1 1

0.1 0.1
RMSE

RMSE
0.01 0.01

0.001 0.001
1200
4200
6200
12
24
36
48
60
72
84
96
108
120
132
200
250
300
450
500
550
600
800

1200
4200
6200
48
12
24
36

60
72
84
96

200
108
120
132

250
300
450
500
550
600
800
CNN-BiLSTM-T CNN-BiLSTM-N CNN-BiLSTM-Dt CNN-BiLSTM-T CNN-BiLSTM-N CNN-BiLSTM-Dt

Fig. 8. RMSE comparison between the fast commercial EV to slow commercial EV and slow commercial EV to fast commercial EV.

4.3.1. Transfer learning for residential EV voltage to the fast-commercial predict the next-hour residential EV voltage profile are shown in Table 5
EV voltage and Fig. 8. In comparison to the new model, the CNN-BiLSTM T model
According to the RMSE scores presented in Table 5, the transfer starts with higher RMSE scores but recovers after 48 h. The transfer-
learning-based CNN-BiLSTM-T outperforms the newly created model based model dominates the new model until the 1200th data size
CNN-BiLSTM-N. Interestingly it outperformed (for a limited time) the instance and an isolated high RMSE point at the 550th sample size.
CNN-BiLSTM-Dt model trained on a residential dataset but is not Beyond 1200 h the new model dominates again as expected as it will
exposed to the target dataset labels. The target data samples before the have amassed adequate training examples. A similar trend with earlier
4th day did not provide valuable information to enhance transfer transfer learning schemes is also observed, where the CNN-BiLSTM-Dt
learning. However, at the 96th-hour mark (four days) and beyond, the dominates the CNN-BiLSTM-T model on the first data instances up to
CNN-BiLSTM-T model supersedes the CNN-BiLSTM-Dt model for all the 36th hour and at the 84th hour after which the transfer-based model
tested samples. The transfer learning-based model outperformed the takes over superiority. While the pre-trained model CNN-BiLSTM-Dt
new model at all data instances except the 36th-hour data size instance maintained very low RMSE scores at tested data instances, it is
and at 4200 and 6200 training samples (almost 50%). After 4200 h, the observed that it also has the least standard deviation on its RMSE scores
new model (CNN-BiLSTM-N) outperforms the CNN-BiLSTM-T model. on both transfer learning schemes. This low standard deviation on the
The latter is expected as the new model will have seen adequate data RMSE score shows no improvement in model performance as data size
samples (50%) as in the traditional model prediction scenarios. Since in increases (or decreases), which is attributed to non-exposure to the
transfer learning, the interest is in instances where there is limited data target data (as per experiment design).
size (below 4200), this finding suggests that the effect of weight initi­
alisation outweighed the differences in data distribution, thus validating 4.4. Main findings and underlying justification
the use of transfer learning.
Fig. 8 shows a visual representation of how the RMSE values for the The experiments conducted led to three main findings, which are
three models vary as the training data size increases. A sharp decline in summarised below:
the RMSE score is observed after 24 h for the CNN-BiLSTM-T model, and
it continues to decline. The new model remains with the highest RMSE 1) The proposed hybrid transfer learning (CNN-BiLSTM) model per­
until the last two data instances, where it records the least RMSE scores, forms better than the baseline CNN-N transfer learning model, new
demonstrating the importance of adequate data in learning a good and pre-trained CNN-BiLSTM models in predicting the EV charging
model. The pre-trained model remains constant throughout. and network voltage profiles.
2) A transfer learning feature extraction strategy involving freezing
4.3.2. Transfer learning for fast commercial EV voltage to residential EV both the encoder and decoder models and retraining only the dense
voltage layer weights yielded better results as opposed to just freezing the
The results of a reverse transfer learning scheme in which knowledge encoder CNN module of the CNN-BiLSTM model. Understanding
learned from the fast commercial EV voltage prediction is transferred to

9
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

transfer learning training strategy is useful for improving accuracy, Declaration of Competing Interest
model design and interpretability.
3) The proposed approach could be used for a limited number of The authors declare that they have no known competing financial
datasets. However, a very low number of data size instances (below interests or personal relationships that could have appeared to influence
36 h for EV charging demand and below 48 for voltage prediction) the work reported in this paper.
could not provide enough helpful information to enhance prediction
in the target domain (transfer learning) in the studied datasets. Data Availability

5. Conclusions and future works Data will be made available on request.

This paper presents a novel hybrid transfer learning approach to References


predict EV charging demand and network voltage profiles using EV
charging datasets. The CNN model is the standard model for transfer [1] M. Grubb, C. Vrolijk, D. Brack, Routledge revivals: kyoto protocol. A Guide and
Assessment, Routledge, 1999–2018.
learning in various domains due to its enabling properties. Experiments [2] D. Hilson, Managing the impacts of renewably powered electric vehicles on
using EV charging demand datasets and voltage profiles were set up to distribution networks, Australian Renewable Energy Agency (ARENA) (2019).
determine the performance of the hybrid CNN-BiLSTM in transfer [3] C. Zhang, Z. Yang, K. Li, Modeling of electric vehicle batteries using RBF neural
networks, in 2014 International Conference on Computing, in: Management and
learning settings. The transfer-based hybrid model was then compared Telecommunications (ComManTel), IEEE, 2014, pp. 116–121.
to the commonly used CNN transfer-based model, a newly created [4] R. J. Hyndman and G. Athanasopoulos, Forecasting: principles and practice, 2013,
hybrid model and a pre-trained hybrid model in the designed transfer URL: https://www. otexts. org/fpp [accessed 2018–02-15][WebCite Cache ID
6xFJlXCQI], 2017.
learning experiments. Improved accuracy has been achieved by the [5] M. Ahrarinouri, M. Rastegar, K. Karami, A.R. Seifi, Distributed reinforcement
proposed hybrid model compared to newly created, pre-trained hybrid learning energy management approach in multiple residential energy hubs, ISSN
models and the standard CNN model by employing the proposed method 2352-4677, Sustain. Energy, Grids Netw. Volume 32 (2022), 100795, https://doi.
org/10.1016/j.segan.2022.100795.
on datasets with as low as 48 h of charging demand for EV forecasting
[6] M.Q. Raza, A. Khosravi, A review on artificial intelligence based load demand
and 60 h for voltage forecasting. forecasting techniques for smart grid and buildings, Renew. Sust. Energy Rev. vol.
This research has demonstrated the following contributions: 50 (2015) 1352–1372.
[7] W.G. Zhang, F.X. Xie, M. Huang, J. Li, Y.F. Li, Research on short-term load
forecasting methods of electric buses charging station, Power Syst. Prot. Control
• Demonstration of the feature extraction strategies using the hybrid vol. 41 (4) (2013) 61–66.
CNN-BiLSTM transfer learning model’s architectural variations to [8] D. Niu, T. Ma, H. Wang, H. Liu, Y. Huang, Short-term load forecasting of electric
apply for (residential to commercial and vice versa) EV charging vehicle charging station based on KPCA and CNN parameters optimised by NSGA
II, Electr. Power Constr. no. 3 (2017) 12.
demand prediction with limited available data. [9] S. Xydas, C. Marmaras, L.M. Cipcigan, A. Hassan, and N. Jenkins, Electric vehicle
• Demonstration of network voltage profile prediction by using the load forecasting using data mining methods, IET Hybrid and Electric Vehicles
hybrid CNN-BiLSTM transfer learning models. This would enable Conference, 06-07 November, London, UK, 2013.
[10] Z.S. Warraich, W.G. Morsi, Early detection of cyber–physical attacks on fast
quick prediction of the system voltage without performing a power charging stations using machine learning considering vehicle-to-grid operation in
flow study and subsequently reduce the computational burden for microgrids, ISSN 2352-4677, Sustain. Energy, Grids Netw. Volume 34 (2023),
power grid operation. 101027, https://doi.org/10.1016/j.segan.2023.101027.
[11] J. Zhu, et al., Electric vehicle charging load forecasting: a comparative study of
deep learning approaches, Energies vol. 12 (14) (2019) 2692.
With high integration of EVs into the grid, the total system load will [12] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, nature, 521 (7553) (2015)
increase significantly. Subsequently, the voltage profile of the low- 436–444.
[13] S.G. Kim, J.Y. Jung, M.K. Sim, A two-step approach to solar power generation
voltage networks will be affected. As the low-voltage network is
prediction based on weather data using machine learning, Sustainability 11 (5)
spread over a wide geographical area and the size of the distribution grid (2019) 1501.
is significantly large – it would not be possible to perform power flow [14] J. Long, E. Shelhamer, and T. Darrell, Fully convolutional networks for semantic
simulation for the whole network. The transfer learning approach segmentation, in Proceedings of the IEEE conference on computer vision and
pattern recognition, 2015, pp. 3431–3440.
described in this paper will be performed to train models with the EV [15] D. Cireşan, U. Meier, J. Schmidhuber, Multi-column deep neural networks for
profiles to predict the voltage profiles. This approach will be used for image classification, Prepr. arXiv 1202 (2012) 2745.
screening purposes to identify the likely voltage problems in the [16] Z. Liang, G. Zhang, J.X. Huang, Q.V. Hu, Deep learning for healthcare decision
making with EMRs, in 2014, in: IEEE Int. Conf. on Bioinformatics and Biomedicine
network. Then, a power flow simulation will be performed for the (BIBM), IEEE, 2014, pp. 556–559.
selected part of the network where the voltage problem is identified by [17] Y. Li, Y. Huang, M. Zhang, Short-term load forecasting for electric vehicle charging
the transfer learning approach. Consequently, the voltage regulating station based on niche immunity lion algorithm and convolutional neural network,
Energies vol. 11 (5) (2018) 1253.
equipment could be installed into the network where required as veri­ [18] M. Longo, S.M. Miraftabzadeh, High-resolution PV power prediction model based
fied by the power flow simulation. This sequential approach will reduce on the deep learning and attention mechanism, ISSN 2352-4677, Sustain. Energy,
the computational requirement and time for the whole network Grids Netw. 34 (2023), 101025, https://doi.org/10.1016/j.segan.2023.101025.
[19] H.S. Nogay, Estimating the aggregated available capacity for vehicle to grid
simulation. services using deep learning and Nonlinear Autoregressive Neural Network, ISSN
This study pioneers transfer learning in EV charging demand and 2352-4677, Sustain. Energy, Grids Netw. 29 (2022), 100590, https://doi.org/
subsequent network voltage forecasting (without using power system 10.1016/j.segan.2021.100590.
[20] A. Graves, J. Schmidhuber, Framewise phoneme classification with bidirectional
simulation). Future work involves using additional variables that affect
LSTM and other neural network architectures, Neural Netw. vol. 18 (5–6) (2005)
the EV dataset such as meteorological variables, and socio-economic 602–610.
factors data to EV forecasting which consequently improves transfer [21] A. Graves, N. Jaitly, A.-rA. R. Mohamed, Hybrid speech recognition with deep
learning. Defining a similarity index between the source and target bidirectional LSTM, 2013 IEEE Workshop Autom. Speech Recognit. Underst.
(2013) 273–278 (IEEE).
datasets will help identify which datasets are similar as transfer learning [22] K. A. Althelaya, E. S. M. El-Alfy, and S. Mohammed, Evaluation of bidirectional
is enhanced when datasets are similar. Introducing more transfer lstm for short-and long-term stock market prediction, in 2018 9th international
learning models to provide a comprehensive assessment of transfer conference on information and communication systems (ICICS), 2018, pp.
151–156: IEEE.
learning-based models helps in model selection and designing models [23] G. Liu, J. Guo, Bidirectional LSTM with attention mechanism and convolutional
for improved prediction under limited data. layer for text classification, Neurocomputing vol. 337 (2019) 325–338.
[24] V. Passricha, R.K. Aggarwal, A hybrid of deep CNN and bidirectional LSTM for
automatic speech recognition, ( Engl. ), J. Intell. Syst. 29 (1) (2020) 1261, 01 Jan.
2020.

10
P. Banda et al. Sustainable Energy, Grids and Networks 36 (2023) 101191

[25] Q. Zhang, Y. Li, Y. Hu, X. Zhao, An encrypted speech retrieval method based on [36] Y. LeCun, Y. Bengio, Convolutional networks for images, speech, and time series,
deep perceptual hashing and CNN-BiLSTM, IEEE Access, 8 (2020) 148556–148569. Handb. brain Theory Neural Netw. vol. 3361 (10) (1995) 1995.
[26] C. Szegedy et al., Going deeper with convolutions, in Proceedings of the IEEE [37] F.A. Gers, J. Schmidhuber, Recurrent nets that time and count, in: Proceedings of
conference on computer vision and pattern recognition, 2015, pp. 1–9. the IEEE-INNS-ENNS International Joint Conference on Neural Networks. IJCNN
[27] D. Bahdanau, K. Cho, and Y. Bengio, Neural machine translation by jointly learning 2000. Neural Computing: New Challenges and Perspectives for the New
to align and translate, [1409.0473] (arxiv.org), 2014. Millennium 3, IEEE, 2000, pp. 189–194.
[28] I. Sutskever O. Vinyals Q. V. Le Sequence to sequence learning with neural [38] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. vol. 9 (8)
networks (neurips.cc) 2014. (1997) 1735–1780.
[29] H.I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, P.A. Muller, Transfer learning for [39] V. Nair, G.E. Hinton. Rectified linear units improve restricted boltzmann machines,
time series classification, 2018 IEEE Int. Conf. Big Data (Big Data) (2018) ICML, 2010.
1367–1376. [40] Dandee_City_Council. (2019). Electric Vehicle Charging Data. Available: https://
[30] C. Fan, et al., Statistical investigations of transfer learning-based methodology for data.dundeecity.gov.uk/ dataset/ev-charging-data.
short-term building energy predictions, Applied Energy 262 (2020), 114499. [41] M. Muratori, M.J. Moran, E. Serra, G. Rizzoni, Highly-resolved modeling of
[31] P. Banda, M.A. Bhuiyan, K.N. Hasan, K. Zhang, A. Song. Timeseries based deep personal transportation energy consumption in the United States,, Energy vol. 58
hybrid transfer learning frameworks: A case study of electric vehicle energy (2013) 168–177.
prediction, International Conference on Computational Science, Krakow, Poland, [42] R.D. Zimmerman, C.E. Murillo-Sanchez, R.J. Thomas, MATPOWER: steady-state
2021, pp. 259–272, 16–18 June. operations, planning and analysis tools for power systems research and education
[32] S. J. Pan, Q. Yang, A survey on transfer learning, IEEE Transactions on Knowledge (Feb), IEEE Trans. Power Syst. vol. 26 (1) (2011) 12–19 (Feb).
and Data Engineering, 22(10), 1345 - 1359, 2010. [43] K. N. Hasan, K. M. Muttaqi, P. Borboa, J. Scira, Z. Zhang, M. Leishman,
[33] O. Day, T.M. Khoshgoftaar, A survey on heterogeneous transfer learning, Journal of Measurement-based electric vehicle load profile and its impact on power system
Big Data, 4 (1) (2017) 1–42. operation, 9th International Conference on Power and Energy Systems (ICPES), p.
[34] T.Y. Kim, S.B. Cho, Predicting residential energy consumption using CNN-LSTM 1–6, Perth, Australia, Dec 10–12, 2019.
neural networks, Energy vol. 182 (2019) 72–81. [44] F. Pedregosa, et al., Scikit-learn: machine learning in python, J. Mach. Learn. Res.
[35] F.U.M. Ullah, A. Ullah, I.U. Haq, S. Rho, S.W. Baik, Short-term prediction of vol. 12 (Oct) (2011) 2825–2830.
residential power energy consumption via CNN and multi-layer bi-directional
LSTM networks, IEEE Access, vol. 8 (2019) 123369–123380.

11

You might also like