Article

Quantitative Analysis on the Proportion of Renewable Energy

Generation Based on Broadband Feature Extraction
Li Chen 1,2, Jian Shen 1, Bin Zhou 1, Qingsong Wang 3,* and Giuseppe Buja 4

1 State Key Laboratory of Smart Grid Protection and Control, Nari Group Corporation,
Nanjing 211106, China
2 School of Instrument Science and Engineering, Southeast University, Nanjing 210096, China

3 School of Electrical Engineering, Southeast University, Nanjing 210096, China

4 Department of Industrial Engineering, University of Padova, 35131 Padova, Italy

* Correspondence: [email protected]

Abstract: With the massive access of distributed renewable energy sources, many uncertain renew-
able energy power components have been added to the low-voltage lines in substations in addition
to the loads of definite classification. From the perspective of economy and cleanliness, it is neces-
sary to quantitatively analyze the renewable energy share among them and improve the power
quality level of users. For the power quality information at low-voltage feeders, this paper proposes
a quantitative analysis algorithm based on improved wavelet energy entropy and LSTM neural net-
work. The method is based on wavelet transform, based on sym8 wavelet basis function; it divides
the long time sequence into equal-length small time sequences, calculates each feature component
obtained from wavelet transform decomposition separately, then borrows the concept of infor-
mation entropy to find its energy entropy. After obtaining the energy entropy sequence of each
Citation: Chen, L.; Shen, J.; Zhou, B.; feature component, it then borrows the concept of kurtosis to weighted differentiation of each en-
Wang, Q.; Buja, G. Quantitative ergy entropy sequence to highlight the feature information and finally, uses the LSTM neural net-
Analysis on the Proportion of work to classify the power quality information of different renewable energy sources to determine
Renewable Energy Generation Based to which interval segment they belong. By building a simulation model to simulate the actual data
on Broadband Feature Extraction. in the field, the percentage of renewable energy can be quantitatively analyzed efficiently and ac-
Appl. Sci. 2022, 12, 11159. curately.
https://doi.org/
10.3390/app122111159
Keywords: renewable energy source; quantitative analysis; improved wavelet energy entropy;
Academic Editor: Andrea Frazzica LSTM neural network; kurtosis

Received: 12 October 2022

Accepted: 2 November 2022
Published: 3 November 2022
1. Introduction
Publisher’s Note: MDPI stays neu-
The natural environment and energy tension are the main issues facing the global
tral with regard to jurisdictional
claims in published maps and institu-
economic development at this stage. On the basis of the basic requirement of ecological
tional affiliations.
environmental protection as well as the goal of carbon neutrality, the cleanliness of
China’s electricity continues to accelerate, and renewable energy, mainly wind power and
photovoltaic power, will see accelerated development. Considering the process of eco-
nomic globalization, some large enterprises focus on the cleanliness of electricity when
Copyright: © 2022 by the authors. Li- purchasing electricity and are committed to using electricity with low pollution levels.
censee MDPI, Basel, Switzerland. Therefore, this puts forward higher data mining requirements for power system dispatch-
This article is an open access article ing [1,2].
distributed under the terms and con- In order to improve the macro economy and enhance the development of these in-
ditions of the Creative Commons At- dustries, it is necessary to be able to grasp the power consumption attributes, power con-
tribution (CC BY) license (https://cre- sumption scale, and energy consumption trend more accurately at the dispatching side.
ativecommons.org/licenses/by/4.0/). Therefore, there is an urgent need to study the macroscopic load identification method at the
station side to realize the control of energy-using groups and their proportion by dispatch [3].

Appl. Sci. 2022, 12, 11159. https://doi.org/10.3390/app122111159 www.mdpi.com/journal/applsci

Appl. Sci. 2022, 12, 11159 2 of 12

To quantitatively analyze the share of renewable energy generation from the load
side, the collected data information needs to be analyzed and processed firstly to extract
its features from the broadband spectrum. Fourier transform is one of the most basic fre-
quency domain analysis methods, whose function is to reflect the overall information of
the signal at a certain time by finding the spectrum of the original signal [4]. However, it
cannot reflect the local characteristics of the signal and is not suitable for analyzing the
power quality problems with fast changes and short durations, such as transient pulses
and transient oscillations due to renewable energy access. In the literature [5], the short-
time Fourier transform is used to obtain multiple short-time signals by sliding and adding
windows to the signals, after which Fourier transform is performed separately to obtain
the time-frequency characteristics of the signals; this improves, to some extent, the draw-
back that the Fourier transform cannot reflect the local characteristics, but it is not adap-
tive because the length of the selected time-frequency window is fixed. The wavelet trans-
form is a multi-scale time–frequency transform that can reflect both the full process and
local information of the signal, so it is widely used to analyze abrupt signals and unsteady
signals [6], but there are problems, such as the difficulty of selecting the basis function.
The literature [7] uses Hilbert–Huang transform, combining two algorithms of empirical
mode decomposition and Hilbert transform, which has better adaptivity, but has disad-
vantages, such as modal confusion [8] and endpoint effects. Wavelet energy entropy [9,10]
is an analysis algorithm that performs a secondary calculation of the feature vector, and
the energy entropy of each component is obtained as a new feature vector with more ob-
vious spectral characteristics by further analysis and calculation of the feature vectors ob-
tained from the wavelet transform.
At present, algorithms involving quantitative analysis are less studied, using mostly
machine learning methods. The more-frequently applied methods based on feature selec-
tion are population-based genetic algorithms [11], particle swarm optimization [12] and
ant colony optimization [13] algorithms, and other population intelligence algorithms,
which are all effective for the input data drive algorithms of classifiers. The literature [14]
based on improved GA to filter and combine wavelet extracted features and classify them
by multiple models effectively improves the training speed and accuracy of the classifier,
but the feature screening algorithm is not integrated with the subsequent classifier, and
the GA tuning reference is more complicated, and the scene fit is lower. The literature [15]
improves the accuracy by merit screening the wavelet-extracted feature vectors through
PSO, but the overall algorithm complexity is high. The ACO is computationally intensive,
convergence is relatively slow, and it is not applicable to the feature selection problem of
PQ perturbation. The traditional BP neural network [16] has stronger adaptivity, but the con-
vergence results are slower and not suitable for dealing with long time series problems.
On the basis of the above considerations, this paper proposes a quantitative analysis
algorithm based on improved wavelet energy entropy method and LSTM neural network
for the problem of renewable energy power share on low-voltage feeders. The method in
this paper firstly pre-processes the sampled voltage and current signals by wavelet trans-
form; secondly, it finds the energy entropy of each signal characteristic component, ana-
lyzes the internal variability of the energy entropy sequence separately, introduces the
weighting coefficients, and obtains the improved energy entropy sequence after mathe-
matical processing; finally, the improved energy entropy sequence is used as the training
sample of LSTM neural network to test the judgment accuracy of the algorithm under
different new energy occupation ratio. By building the distribution network model of re-
newable energy access and obtaining the simulation data, the effectiveness and accuracy of
the quantitative analysis of the algorithm in this paper are verified by comparing this method
with the traditional method.
Appl. Sci. 2022, 12, 11159 3 of 12

2. Principles of Improved Wavelet Energy Entropy and LSTM Neural Network

2.1. Wavelet Transform
Due to the strong random fluctuation of new energy, a high proportion of new en-
ergy connected to the grid will lead to a significant increase in power generation fluctua-
tion and a change in the harmonic components of voltage and current in the grid. There-
fore, in order to study the external characteristics of current, power, and frequency of new
energy sources (such as distributed photovoltaic and wind power after grid connection),
extract their typical features, and construct filtering algorithms and frequency analysis
algorithms suitable for end-user load current analysis, suitable feature extraction algo-
rithms should be selected.
Wavelet transform—which is a local transformation of time and frequency and can
simultaneously analyze the data jointly in the time domain and frequency at multiple
scales with multi-scale refinement analysis—is based on various specific functions to ex-
pand the data signal into a series of levels. Therefore, the signal can be analyzed effectively
at different decomposition levels and with different wavelet basis functions. In this paper,
the wavelet transform approach is chosen as a major tool for spectral analysis.
The method of signal processing requires a basis function as a tool for the transform.
The basis function used in wavelet transform, on the other hand, is different from the
stable and periodic sinusoidal function in Fourier transform, except for a small segment
of the signal, thus reflecting the local characteristics.
Suppose   t   L  R  , if it satisfies [17].
2

   t  dt  0 (1)

Then,   t  can be used as the base wavelet (or mother wavelet). If the basis wave-
let is scaled and translated, a series of wavelet families can be obtained as  a ,  t  :

 a ,  t  | a |
1/ 2  t  
  (2)
 a 
where a and τ are the scale parameter and the translation parameter, respectively, and t is
the time. The scale parameter determines the degree of scaling of the fundamental wave-
let. However, in general practical calculations, the data are stored in discrete form; there-
fore, the discrete wavelet transform is generally chosen to change the parameters using
the dichotomy method.
j j
Take a  2 ,   2 k , the discrete wavelet family is given by

 j ,k  t   2  j /2  2  j t  k  (3)

For a certain signal f  t   L  R  the discrete wavelet transform is calculated as fol-

2

lows:

  R
 
W f 2 j , 2 j k  2  j /2  f  t  * 2  j t  k dt  f  t  , j ,k  t   d j  k  (4)

where dj(k) denotes the wavelet coefficients, and its expression is given by:

d j  k    f  t   t  dt
*
j ,k
(5)
R

Multi-resolution analysis, also known as multi-scale analysis, is able to decompose

the signal well in both the low- and high-frequency bands, showing the details and the
full picture of the signal. Its reconstruction equation is:
Appl. Sci. 2022, 12, 11159 4 of 12

f t  
1   1
W f  a ,  
 t   dad
C
 
 
a
2 
 a


(6)

Using the decomposition and reconstruction of the signal, multi-resolution decom-

position can be achieved.
Let Wj denote the high-frequency part of the multi-resolution decomposition and Vj
denote the low-frequency part. Then, Wj is Vj in Vj+1, the orthogonal complement in
V j  W j  V j 1 , j  Z (7)

Therefore, for an arbitrary signal, the multi-resolution decomposition can be divided

into a high-frequency detail part and a low-frequency approximation part, which is im-
plemented as follows:
Suppose that the frequency distribution of a signal f(t) has a frequency distribution
in (0,π) above; then, this frequency band is defined as the space V0. After the first level of
decomposition, the space is divided into two subspaces: the low-frequency V1 (frequency
band (0,π/2)) and the high-frequency W1 (frequency band (π/2,π)). In the second level of
decomposition, the object is the low-frequency space of the previous level V1, and it is
decomposed into the low-frequency part V2 (frequency band (0,π/4)) and the high-fre-
quency part W2 (frequency band (π/4,π/2)), and so on.
Let the high- and low-frequency components, after decomposition and reconstruc-
tion, be D j  n  and Aj  n  ; the specific equation is as follows:
m
f  n   D1  n   A1  n   D1  n   D2  n   A2  n      D j  n   Am  n  (8)
j 1

It is not necessary to let Am  n   Dm1  n  , then the above equation is:

m 1
f n   Dj n (9)
j 1

where D j  n  denotes the f  n  components at different frequency bands.

2.2. Wavelet Energy Entropy

The wavelet energy entropy here is actually borrowed from the concept of infor-
mation entropy. By information entropy, we mean that for some discrete random variable
n
X, its probability distribution is pi  P  X  xi , i  1, 2,3, , and  pi  1 , then the ex-
i 1

pression of information entropy is:

n
H  x    pi ln  pi  (10)
n 1

Before deriving the wavelet energy entropy, we ask for the wavelet energy at each
frequency band. According to the assumptions in the previous subsection, the signal se-
quence f  n  wavelet coefficients on each frequency band are D j  n  , then the wavelet
energy can be expressed by Equation (11).
n
Ej   i 1
Dj
2
i  (11)

Since the probability p used for the information entropy satisfies and is 1, the relative
energy is found:

Ej
Pj  n (12)
 i 1
Ej
Appl. Sci. 2022, 12, 11159 5 of 12

Therefore, the wavelet energy entropy of a segment of the sequence is:

   jPj ln( Pj ) (13)

It should be noted that the wavelet energy entropy represents the feature information
contained in a certain segment of the sequence, so when conducting feature extraction on
the data, it can be considered to segment it and split it into multi-group sequences of equal
length for processing.

2.3. Improving Wavelet Energy Entropy

Since the new energy source is connected to the grid—although it will lead to changes
in the characteristic quantities, such as grid frequency as well as harmonic characteris-
tics—the fluctuation is still within a certain range. Therefore, if the wavelet energy entropy
of each time period is directly used as the characteristic parameter, it will lead to little
difference in the characteristic quantities under different new energy percentages, and
further processing is needed.
A weighted wavelet energy entropy theory is proposed: let there be k subsequences
within a certain period of time sequence, and the energy entropy size of each sequence is
Entropy(k). Calculate the energy entropy under the entire time period and obtain the re-
sulting entropy mean. Let the weighting factor be A, then the processed energy entropy
sequence is:

Entropynew  k   A   Entropy  k   Entropy _ mean  (14)

where the weighting factor A is calculated by the following formula:

4
A  B /  Entropy   
4
(15)

The letter B in the equation is taken to be approximately 103 and represents the scale
factor. μ represents the mean value of the energy entropy series, and σ represents the
standard deviation of the energy entropy series. The value of the weighting coefficient is
taken to be inversely proportional to the kurtosis of the energy entropy series. Since the
kurtosis reflects the steepness of the data, and the value of kurtosis is small for energy
entropy with inconspicuous feature information, the inverse of kurtosis is used to high-
light the internal variability in order to highlight the internal variability. At the same time,
it makes the level of internal variability of each type of energy entropy maintain on the same
baseline, thus improving the validity of the overall information and the reliability as training
samples.

2.4. LSTM Neural Network

LSTM neural network is a special type of recurrent neural network, which is a type
of neural network with memory and suitable for processing sequences with time scales
[18]. Especially for the problem of time series with long span, it is able to memorize longer
historical information. The structure of LSTM neural network is shown in the following
Figure 1 [19].

ht ht 1
− ht+1

tanh tanh tanh

σ σ tanh σ σ σ tanh σ σ σ tanh σ

xt 1
− xt xt+1
Appl. Sci. 2022, 12, 11159 6 of 12

Figure 1. LSTM neural network structure diagram.

The key to LSTM is the cell state, which is analogous to a transmission band, a
straight line located at the top of the structural map. The state of the cell moves along this
line with only a small number of linear operations. In this way, cellular information is not
overly modified after passing through a cell. The superiority of the LSTM is its ability to
selectively modify the state of the cell, and this is achieved by a gate which selectively
allows information to pass through. The gate consists of a sigmoid neural network layer
and a point multiplication operation.
The operation of a single LSTM unit can be divided into four steps as follows:
(1) Step 1
First, the LSTM must decide to discard some of the information in the cell state, i.e.,
the “forgetting layer”. The sigmoid function is used to determine the amount of infor-
mation to be retained by using the output, ht-1, of the previous moment and the input, xt,
of the current moment.
(2) Step 2
The amount of information stored in the cell state is determined. First, the sigmoid
layer, which is the “input gate layer”, also outputs a value between 0 and 1 to determine
the amount of information to be remembered; second, the information passes through the
tanh layer to create a candidate vector, Ct’. This vector is then fed as additional infor-
mation to the cell’s state.
(3) Step 3
Update the value of Ct. First, multiply the forgetting factor, ft, obtained in the first step
to obtain the information to be forgotten. Add the candidate vector to get the new cell state,
Ct.
(4) Step 4
Require the new cell state to pass through the tanh layer to normalize the value be-
tween −1 and 1; pass the two external inputs through the sigmoid layer to determine the
amount of cell information output. The product of the two quantities is used as the final
output quantity.

3. Simulation and Discussions

3.1. Grid Model Simulation with Renewable Energy
In order to study the proportion and spectral characteristics of distributed new en-
ergy sources on the low-voltage feeder, a simulation model with multiple renewable en-
ergy sources connected to the grid is built in Simulink, and the model structure is shown
in Figure 2. The voltage level of the main network is 750 kV, and at the 35 kV bus, wind
farms and photovoltaic farms are connected. The wind turbine is a double-fed asynchro-
nous wind turbine with a rated output power of 1.5 MW, and the PV cell is SPR-415E-
WHT-D with a rated output power of 10 MW. A three-phase RLC load is connected to the
10-kV low-voltage feeder with a load power of 200 MW. The detailed parameters of each
part of the simulation are shown in the Table 1.

Table 1. Parameters of the simulation.

Parameter Value Parameter Value (pu)

Main grid voltage 750 kV Resistance of Transformer 1 (main grid) 0.08
DFIG output voltage 575 V Inductance of Transformer 1 (main grid) 0.08/30
PV output voltage 10 kV Resistance of Transformer 1 (new energy) 0.025
Load power 200 MW Inductance of Transformer 1 (new energy) 0.025/30
Appl. Sci. 2022, 12, 11159 7 of 12

750kV
Transformers
35kV
Wind Power Transformers PCC Grid

DFIG Rectifier Inverter

35kV
PCC

PV power Transformers

PV
MPPT Inverter
arrays

10kV
Load

Figure 2. Simulation model.

In this paper, the fixed load power is 200 MW, and the change of the new energy
share at the low-voltage feeder is achieved by changing the number of input units of the
wind turbine and the number of components of the PV array. On the basis of the consid-
eration of the actual situation, a total of 500 sets of simulations are performed, and the
new energy share increases from 0.1% to 50% by 0.1% increments.
Through the power acquisition module, the PV output power (P_PV), the wind tur-
bine output power (P_wind), and the load-side power (P_load) are obtained, and the new
energy share ratio is calculated for each simulation as follows:
ratio  P _ PV  P _ wind / P _ load (16)
At the same time, there is more noise in the electrical signal due to the actual meas-
urement in the field, for example, white noise interference, among others. To simulate this
situation, adding Gaussian white noise (i.e., superimposing a certain percentage of nor-
mally distributed noise signal in the output current) simulates white noise interference.
As an example, the time and frequency domain diagrams of the single-phase current
at the low-voltage feeder are shown in the Figure 3 (taking the A-phase current signal for
analysis), taking the 10% and 40% sets of data. With the increase of the proportion of new
energy, the distortion of current waveform has been improved, to a certain extent. It can
be seen from Figure 3 that there are obvious harmonics at the wave crest and wave trough.

10%rate

2000
Amplitude

−2000

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

Time(s)
40%rate

2000
Amplitude

−2000

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

Time(s)

Figure 3. Current signal of low-voltage feeder under different proportions.

However, it can be seen that the degree of signal distortion at the low-voltage feeder
does not vary much when the new energy ratio varies, so it is not possible to quantitatively
analyze it visually.
Appl. Sci. 2022, 12, 11159 8 of 12

3.2. Specific Process to Improve Wavelet Energy Entropy

Since this simulation is a discrete simulation, the set step size is 5 × 10−6, thus 4000
data points in one work frequency cycle are taken for analysis. The 4000 points are divided
into 40 groups of 50 points each, and the time series composed of them are wavelet-trans-
formed, and the energy entropy is obtained. In this way, a set of current data is turned
into a 40-point energy entropy sequence. Since six sets of voltage and current data are
obtained from each simulation group, a 6 × 40 feature matrix is composed, representing
40 six-dimensional column feature vectors, as the training object of the neural network.
The specific flow is shown in the following Figure 4.

improve energy entropy

Calculate the energy at each

frequency band based on the
wavelet coefficients obtained from
the decomposition
Splitting 1000 data Use sym8-based Using energy
points into 40 time wavelet transform entropy as the new
blocks for each data set time series
Calculate the energy entropy based
on the energy value, and then use
its cliff characteristics to obtain the
improved energy entropy

Figure 4. Flow chart of wavelet energy entropy method.

The 500 sets of training data were divided into five categories of levels with new
energy percentages of 0–10%, 10–20%, 20–30%, 30–40%, and 40–50%, respectively, and
were recorded as R1, R2, R3, R4, and R5. Each set of data from each category was analyzed
for its energy entropy data, as shown in the following Figure 5. (R1-R5 is conducted in the
same simulation model.)

600
R1 R2 R3 R4 R5

400

200
Entropy

−200

−400

−600
0 5 10 15 20 25 30 35 40
Time block

Figure 5. Improved energy entropy of different categories.

From the figure, it can be found that the peak of the improved energy entropy is
concentrated in the starting phase as well as the midpoint phase, and the other phases are
relatively smooth, indicating that the new energy fluctuation amplitude is more obvious
in these three phases. Therefore, if we want to improve the operation speed of the algo-
rithm and reduce the storage burden, we can segment the waveform and intercept the key
segments for feature extraction, which can effectively improve the subsequent training speed.
Appl. Sci. 2022, 12, 11159 9 of 12

At the same time, observing the waveform information at different levels under the
same stage, it can be found that in the start–stop stage, the improved energy entropy
shows an increasing trend as the share of new energy increases, with the most drastic
change in the end stage. In the intermediate stage, the energy entropy shows a small range
of sinusoidal characteristics and changes in a similar way as the start–stop stage. There-
fore, it indicates that the improved energy entropy taps the information of signal charac-
teristics under various types of levels.

3.3. Classification Algorithm and Result Analysis

In this paper, traditional backpropagation neural network (BPNN) [20] and LSTM
neural network are chosen for the classification algorithm, and the performance of each
method is compared by changing the input side data. Four sets of control experiments are
taken: (1) traditional wavelet energy entropy + BPNN; (2) traditional wavelet energy en-
tropy + LSTM neural network; (3) improved wavelet energy entropy + BPNN; and (4)
improved wavelet energy entropy + LSTM neural network.
The generated 500 sets of data are divided into 300 training samples and 200 test
samples, with 60 training samples and 40 test samples for each type of data. The implicit
layer of BPNN is configured as 200, the maximum number of rounds is 200, and the learn-
ing rate is set to 0.001; the implicit layer of LSTM neural network is configured as 100, the
maximum number of rounds is set to 100, and the learning rate is set to 0.001. Specific
neural network parameters are shown in the Table 2, and the training results of each group
are listed in the following Table 3. (Traditional refers to traditional wavelet energy en-
tropy, while improved refers to improved wavelet energy entropy.)

Table 2. Specific neural network parameters.

LSTM Setting BPNN Setting

Num Classes 5 Num Classes 5
Max Epochs 100 Max Epochs 200
Mini Batch Size 50 Mini Batch Size 50
Initial Learn Rate 0.005 Initial Learn Rate 0.001
Execution Environment Auto Execution Environment Auto
Shuffle Every Epoch Shuffle Every Epoch

Table 3. Comparison of training effects of different algorithms.

Algorithm Hidden Units Epochs Time (s) Accuracy

Traditional + BPNN 200 2400 50 89.7%
Traditional + LSTM 100 1600 38 91.3%
Improved + BPNN 200 1700 42 92.5%
Improved + LSTM 100 1000 33 98.6%

It can be seen from Table 1 that the training effect of improved wavelet energy en-
tropy + LSTM performs the best, which has 1000 cycles and 33 s training time, both of
which are the lowest; in addition, its accuracy rate reaches 98.6%, which is the highest
among the four methods. At the same time, it is not difficult to find that the training ac-
curacy of the traditional wavelet energy entropy method is much less than that of the
improved wavelet energy entropy method. In the case where the selected neural network
remains unchanged, the accuracy rate is improved by 3–7% after using the improved
wavelet energy entropy. In addition, the use of LSTM neural network also has a significant
impact on the total training time. BPNN is not suitable for dealing with long time se-
quences due to its inherent characteristics, so it is inferior to LSTM neural network in
terms of efficiency and accuracy. For a given training data set, the LSTM training time is
reduced by 12 s and 9 s for the traditional and improved algorithms, respectively,
Appl. Sci. 2022, 12, 11159 10 of 12

compared with their corresponding BPNN times, which proves LSTM’s superior perfor-
mance.
The above analysis shows that compared with BPNN neural network and traditional
wavelet energy entropy method, the improved wavelet energy entropy + LSTM neural
network algorithm proposed in this paper has great superiority. The algorithm requires
low number of layers for the implicit layer configuration, few cycles, and short training
time, but the accuracy rate still remains high enough to prove the effectiveness of the im-
proved wavelet energy entropy method.
The test samples were analyzed using this method; the test results are shown in the
Figure 6, and the test accuracies for each type of occupancy are shown in the Table 4. From
the table, it can be found that the overall accuracy rate is 97.5%, which shows the validity
and reliability of the method. However, the variability of the prediction accuracy of each
category is large, due, in part, to the small number of test samples. At the same time, it
can be found in the figure that the points of error in judgment are all concentrated in the
stratification of each type of group. Since this classification is based on the new energy
occupancy ratio, the feature information at the intersection of R1 and R2 is continuous,
and the generation of misjudgment is within the acceptable range. If the percentage range
is subdivided, for example, into 10 categories with 5% intervals, it can be expected that
there will be a small decrease in the model training accuracy but an improvement in the
practical significance of quantitative analysis.

Table 4. Test accuracies for each type of occupancy.

Category R1 R2 R3 R4 R5
Accuracy 100% 92.5% 97.5% 100% 97.5
Total Accuracy 97.5%

Pred
4.5 Real

3.5
type

2.5

1.5

1
0 20 40 60 80 100 120 140 160 180
group

Figure 6. Comparison between predicted results and accurate values.

3.4. Discussions
From the above simulation and training results, it can be seen that the proposed
method in this paper has a certain improvement in accuracy and a shorter training time
compared with the traditional algorithm, indicating a certain reliability. However, limited
by the capacity of the simulation data and the performance of the machine, the span of the
percentage band in this paper is large, reaching 10%. As a result, it leads to an uneven
distribution of correct predictions in the same class of data sets, and there are some mis-
judgments at the occupancy threshold. If conditions allow, after acquiring more data, and
then using the feature variability of improved energy entropy concentrated in some data
Appl. Sci. 2022, 12, 11159 11 of 12

segments to reduce the length of feature vectors in the training samples and reduce re-
dundant information, this algorithm should be able to further improve the prediction ac-
curacy and reduce the training time. Meanwhile, considering that the regression model
can more accurately describe the proportion of renewable energy, the further study is
considered.
Therefore, the following advantages and disadvantages should be pointed out. Re-
garding the advantages:
(1) There is no need to specifically analyze the topology and power flow of the power
grid, and the required data types are easy to obtain;
(2) The intelligent algorithm has higher fault tolerance and can adapt to various network
structures and operating conditions;
(3) Compared with the traditional algorithm, the algorithm in this paper can highlight
the power quality characteristic level with renewable energy more.
However, this paper also has the following shortcomings:
(1) The training of neural network requires a large amount of data and has certain re-
quirements for sampling devices and storage capacity;
(2) At present, the proportion cannot be given very accurately, only the general scope
can be specified.
Based on the above analysis and simulation verification, this method has certain
practical application value. In the context of global promotion of low carbon emissions,
this method can judge the proportion of renewable energy components from the user side,
so as to provide a more accurate basis for carbon emissions for power users. Electricity
enterprises can use it as a basis to choose cleaner energy.

4. Conclusions
In this paper, we propose a quantitative analysis method based on improved wavelet
energy entropy and LSTM neural network for quantitative analysis of renewable energy
power share at low-voltage feeders in the context of high proportion of renewable energy
access. By building a simulation model to simulate field data and conducting a large num-
ber of training tests, we obtain the following main conclusions:
(1) The feature extraction method based on improved wavelet energy entropy can effec-
tively solve the problem of difficult spectrum analysis caused by complex perturba-
tions in power quality due to the massive access of new energy sources, especially
the improved energy entropy algorithm, which improves the traditional energy en-
tropy method not applicable to the problem of low-kurtosis data processing. At the
same time, this is the innovation of this paper.
(2) The quantitative analysis of renewable energy occupancy ratio by a traditional ma-
chine learning algorithm has problems such as long training time and low accuracy
rate. Therefore, in this paper, we use the data of Conclusion (1) to build an LSTM
neural network and simulate it in MATLAB/Simulink; the simulation shows that the
LSTM neural network can successfully identify the voltage and current data under
different occupancy ratios with low complexity and training time, and the average
accuracy rate is improved.
(3) The next step is to improve the accuracy of quantitative analysis and effectively solve the
problem of low success rate of identification at the renewable energy occupation score.

Author Contributions: L.C. and J.S. conceived the idea of this manuscript, performed the simula-
tions, and wrote the manuscript; B.Z., Q.W., and G.B. provided guidance, revised the manuscript,
and added their thinking to validate the idea. All authors have contributed equally to the analysis
and discussions. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported in part by the State Key Laboratory of Smart Grid Protection
and Control, Nari Group Corporation under project SGNR0000KJJS2200302.
Appl. Sci. 2022, 12, 11159 12 of 12

Conflicts of Interest: The authors declare no conflict of interest.

References
1. Zhuo, Z.; Zhang, N.; Xie, X. Key Technologies and Developing Challenges of Power System with High Proportion of Renewable
Energy. Autom. Electr. Power Syst. 2021, 45, 171–191.
2. Wang, Q.; Chen, P.; Deng, F.; Cheng, M.; Buja, G. The State of the Art of the Control Strategies for Single-Phase Electric Springs.
Appl. Sci. 2018, 8, 2019.
3. Kuo, W.-C.; Chen, C.-H.; Hua, S.-H.; Wang, C.-C. Assessment of Different Deep Learning Methods of Power Generation Fore-
casting for Solar PV System. Appl. Sci. 2022, 12, 7529.
4. Tong, X.; Ye, Z.; Xu, Y.; Gao, S.; Xie, H.; Du, Q.; Liu, S.; Xu, X.; Luan, K. Image Registration with Fourier-Based Image Correlation:
A Comprehensive Review of Developments and Applications. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2019, 12, 4062–4081.
5. Chen, Z.; Wang, L.; Yi, Y. Computation of radio interference excitation current of DC Corona based on short-time Fourier trans-
form. High Volt. Eng. 2019, 45, 1866–1872.
6. Jia, R.; Zhao, J.; Wu, H. Application of correlated probabilistic wavelet transform in partial discharge detection. High Volt. Eng.
2017, 43, 2896–2902.
7. Hong, X.; Liu, Y.; Lin, X.; Luo, Z.; He, Z. Nonlinear Ultrasonic Detection Method for Delamination Damage of Lined Anti-
Corrosion Pipes Using PZT Transducers. Appl. Sci. 2018, 8, 2240.
8. Koganezawa, S.; Tsuda, S.; Tani, H.; Lu, R.; Tagawa, N. Frequency Analysis of Disturbance Torque Exerted on a Carriage Arm
in Hard Disk Drives Using Hilbert–Huang Transform. IEEE Trans Magn. 2018, 54, 1–6.
9. Deng, Y.; Lin, S.; Fu, L. New Criterion of Converter Transformer Differential Protection Based on Wavelet Energy Entropy. IEEE
Trans Power Deliv. 2019, 34, 980–990.
10. Li, Y.; Ning, F.; Jiang, X.; Yi, Y. Feature Extraction of Ship Radiation Signals Based on Wavelet Packet Decomposition and Energy
Entropy. Math. Probl. Eng. 2022, 2022, 8092706. https://doi.org/10.1155/2022/8092706.
11. Qu, H.; Liu, H.; Li Xi. Feature combination optimization for multi-disturbance classification of power quality. Electr. Power
Autom. Equip. 2017, 37, 146–152.
12. Ahila, R.; Sadasivam, V.; Manimala, K. An integrated PSO for parameter determination and feature selection of ELM and its
application in classification of power system disturbances. Appl. Soft Comput. 2015, 32, 23–37.
13. Biswal, B.; Dash, P.; Mishra, S. A hybrid ant colony optimization technique for power signal pattern classification. Expert Syst.
Appl. 2011, 38, 6368–6375.
14. Wang, Y.; Zhang, K.; Zheng, C.; Chen, H. An Optimal Energy Management Method for the Multi-Energy System with Various
Multi-Energy Applications. Appl. Sci. 2018, 8, 2273.
15. Przednowek, K.; Krzeszowski, T.; Przednowek, K.H.; Lenik, P. A System for Analysing the Basketball Free Throw Trajectory
Based on Particle Swarm Optimization. Appl. Sci. 2018, 8, 2090.
16. Chen, J.; Gao, Z.; Luo, L. Error correction of UHF partial discharge location method based on BP network. High Volt. Eng. 2018,
44, 2040–2047.
17. Ingrid, D. Ten Lectures on Wavelets; Society for Industrial and Applied Mathematics (SIAM): Philadelphia, PA, USA, 1992; pp.
14–22.
18. Ali, A.; Devinder, K. Immunocomputing-Based Approach for Optimizing the Topologies of LSTM Networks. IEEE Access 2021,
9, 78993–79004.
19. Sepp, H.; Jürgen, S. Long Short-Term Memory. Neural Comput. 1997, 9, 1735–1780.
20. Paul, J.W. The Roots of Backpropagation: From Ordered Derivatives to Neural Networks and Political Forecasting; John Wiley & Sons:
Hoboken, NJ, USA, 1994; pp. 86–132.