You are on page 1of 6

Research and Application of AI Services Based on

5G MEC in Smart Grid


Li Zhang1, Jiakai Hao1, Guanghuai Zhao1*, Mingshi Wen1, Tianxiang Hai1, Kun Cao1

1 State Grid Beijing Electric Power Company, Beijing, China


*sg_zhaoguanghuai@163.com

Abstract—5G networks will provide services for a wealth of analysis process, which could improve the interaction and
vertical industry applications, bringing more demand for edge response time of the business analysis process. By providing
services. Multi-access edge computing (MEC) is one of the an information technology service environment and cloud
important technologies of 5G. By providing an information computing capabilities close to the mobile users, 5G MEC
technology service environment and cloud computing can better support low-latency and high-bandwidth services
capabilities close to the mobile users, it can better support low- in 5G networks.
latency and high-bandwidth services in 5G networks. This
paper constructs an AI capability center based on 5G MEC, Users can directly use the core algorithm capabilities to
which can provide rich AI services for smart grids. The center build applications for different application scenarios, greatly
supports the dynamic construction and operation of the improving the utilization rate of research and development
intelligent data analysis application, and realizes the rapid results. The main innovations of this article include:
development and flexible deployment of the business analysis
process, which could improve the interaction and response  Artificial intelligence analysis center: 5G uses edge
time of the business analysis process. Based on AI algorithm computing technology to deploy applications to the
capabilities, we focus on the typical cases of artificial customer's data side instead of sending all data to a
intelligence service applications in smart grid e.g., monitoring centralized data center, which improves the real-time
of electrical equipment based on similar audio recognition. We performance of applications. Smart grid is a very
describe the applications of artificial intelligence technology in commercial scale and typical edge computing
various business scenarios from the perspective of application demand scenario, which can provide users with safe
scenario, application scheme, algorithm principle and practice and reliable edge computing services. Figure 1 shows
effect. The applications have been put in use and obtained good the MEC architecture defined by ETSI [7]. The
application effect. mobile edge platform provides users with open
capabilities and users can deploy applications in the
ME app. We can deploy AI center in the ME app to
Keywords—Artificial Intelligence, Audio Recognition, Smart
provide AI analysis capabilities for the smart grid.
Grid, 5G, MEC, Service Application  Optimize execution mechanism: In order to
optimize execution efficiency, a chain-optimized
I. INTRODUCTION
execution model is proposed. At the same time, to
SG (Smart grid) is an inevitable trend in the development meet the needs of different types of data analysis, a
of the energy and power industries, and all countries are fusion method of batch computing and streaming
actively promoting the research and application of SG computing modes are proposed. The mechanism
technology. The U.S. department of energy defines SG as the effectively combines the two computing modes,
use of digital technology to improve the reliability, safety, which can more flexibly execute distributed data
and operating efficiency of the power system (including analysis algorithms, and improves the robustness and
power generation, transmission and distribution networks, flexibility of the system.
power consumption, including distributed generation and
distributed energy storage). The European Union defines SG  Smart grid application: Based on AI algorithm
as the ability to intelligently integrate the behaviors and capabilities, we focus on the typical cases of artificial
actions of all users connected to it to ensure the sustainability, intelligence service applications in smart grid e.g.,
economy and safety of power supply [1-2]. SG will converge monitoring of electrical equipment based on similar
and integrate new measurement, communication, control and audio recognition. We describe the applications of
decision-making technologies to achieve technological artificial intelligence technology in various business
changes in the power industry. The core of SG is scenarios from the perspective of application scenario,
"intelligence" [3-5]. application scheme, algorithm principle and practice
effect. The applications have been put in use and
With the development of AI and 5G, the intelligence of obtained good application effect.
the power grid can be greatly improved [6], and the safe and
efficient operation of the power grid can be effectively Through practical application, it is proved that the
guaranteed. This paper constructs an AI capability center performance of artificial intelligence analysis center can
based on 5G MEC (Multi-access edge computing), which meet the needs of research and development application
can provide rich AI services for smart grids. The center incubation. The application of monitoring of electrical
supports the dynamic construction and operation of the equipment based on similar audio recognition was carried
intelligent data analysis application, and realizes the rapid out, and the practical application showed that this method
development and flexible deployment of the business can fill the gap of the existing scheme, and strengthen the
information security protection.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Authorized licensed use limited to: Mukesh Patel School of Technology & Engineering. Downloaded on February 25,2024 at 10:11:44 UTC from IEEE Xplore. Restrictions apply.
CFS Mx1

Mobile edge
system level
Operations Support System
portal Mm8
User
UE app Mm1
app Mx2 LCM Mm9
proxy Mobile edge
orchestrator

Other
Mp3

Mobile edge hostlevel


Mobile ME Mm2 Mm3
Edge service
Mm4
platform
Service registry
ME ME ME
Mp1 Mp1 Traffic DNS
Mm5 platform platform platform
rules handing
element element element
mgmt mgmt mgmt
Service control
ME ME ME Mobile edge platform Moblie edge platform manager
app app app Mm
Mp2
Other 6
Data plane Virtualisation
mobile
infrastructure manager
edge host Virtualisation infrastructure Mobile edge host Mm7

Fig. 1. ETSI MEC Reference Architecture (ETSI GS MEC 010-2-2019).


The reference architecture is composed of two levels, System level and
Host level. Through the cooperation of two levels of functional components,
MEC provides deployment, management and operation and maintenance
capabilities for edge APPS. Fig. 2. Artificial intelligence center. The AI center includes four parts:
infrastructure, data collection and storage, data calculation and analysis,
II. ARTIFICIAL INTELLIGENCE ANALYSIS CENTER and core algorithm.

A. Center function a) Common Work Flow (Common WF):


The artificial intelligence analysis center mainly The daemon in Hadoop generates non-linear running
includes three parts according to the data flow and dependencies between jobs by analyzing the DAG graph
processing replacement (see Figure 2): network and cloud of the workflow. This kind of dependency can be realized
infrastructure, data acquisition and storage, and by relying on combined MapReduce. Hadoop provides
calculation and analysis, respectively. Job and JobControl classes. The Job class maintains sub-
 Network and cloud infrastructure: This part is task configuration information and sub-task dependency.
the foundation of the artificial intelligence analysis The JobControl class controls the entire work flow, adds
platform, mainly providing network, computing, all sub-task jobs to the JobControl, and executes the
storage and environmental resources. The artificial JobControl's run () method to ensure the sequential
intelligence analysis center is deployed on a cluster. execution of the flow. In this case, the transactions in the
workflow are executed independently and sequentially.
 Data acquisition and storage: This module is the Common mode can be defined as follows: [MAP |
core of the platform and mainly includes open REDUCE?] +. REDUCE is the reducer, after the mapper
source components such as distributed file system named MAP. [·] represents a Job, [MAP | REDUCE? ]
(HDFS), distributed database system (HBase), data represents the sequence which can be repeated one or more
collection (Flume, Kafka) and so on. times to complete a task.

 Calculation and analysis: It mainly includes batch b) Chain Work Flow (Chain WF)
computing (MapReduce), streaming computing In common mode, the output data of each component
(Storm), and the fusion of the two computing mode. must be written to the hard disk. For component-intensive
The platform also provides big data interaction workflows, frequent data landing will reduce the
components (Hive, Pig) and machine learning performance of the workflow. The common mode is
algorithm components (such as Mahout, KALDI, inefficient because the intermediate results of each step in
HTK, Tensorflow, Caffe). the process require I/O and storage resources. For
multiple simple components, we can adopt a modular and
B. Computing mode optimization combined mode to build the execution process. If there is
In order to meet the needs of different types of data only one reduce in multiple jobs, it can be combined into
analysis, on the basis of existing work, a fusion method of one job (see Figure 2). The advantage of this mechanism
batch processing mode and stream processing mode is is that all pre-processing and post-processing are run in a
proposed. This method organically combines batch data single job, and no intermediate files are generated, which
analysis mode and streaming data analysis mode, which greatly reduces I/O operations. Based on the above
can execute distributed data analysis algorithms flexibly analysis, we propose a chain workflow mode. In the cloud
and improve the robustness and flexibility of the system. platform, Daemon dynamically merges and combines the
jobs in the workflow. Chain workflow can be defined as
follows: MAP+ | REDUCE| MAP*.
In practice, the chain mode optimizes and combines
the transactions in the execution process, which can
greatly improve the performance of the workflow.

Authorized licensed use limited to: Mukesh Patel School of Technology & Engineering. Downloaded on February 25,2024 at 10:11:44 UTC from IEEE Xplore. Restrictions apply.
c) Fusion Mode
In response to highly dynamic real-time information
analysis, such as real-time trend analysis, real-time search,
etc., we introduce Storm into a distributed data analysis
platform. Storm is a memory-level calculation framework.
The data is directly imported into the memory through the
network, and the disk access delay is about 75,000 times the
memory access delay. The pseudo-regular expression
symbolically expressing the link of the streaming execution
mode is [Spout | Bolt+] +. In order to meet the needs of
different types of data analysis, we have integrated offline
batch processing (MapReduce) and online streaming
Fig. 3. Fusion of batch computing and streaming computing.
processing (Storm) to enable them to run in the same
cluster. According to research [8], Storm-yarn is used to
Similar audio recognition is mainly used in abnormal
integrate two different processing modes (see Figure 3).
sound detection and identification. The method is to
Zookeeper in the cluster is used to perform unified
compare the audios with the equipment malfunction audio
management and scheduling of resources and share data
library to determine whether they are malfunction sound.
through storage media (such as HDFS or HBase). This
Figure 4 shows the framework for audio recognition which
method not only completes the fusion of the two processing
includes audio data acquisition, audio pre-processing, audio
modes, but also relatively independently retains their
feature extraction and match. Firstly, the audio features are
respective computing frameworks. The lightweight fusion of
extracted, and the hash technology is used to establish the
the two computing models improves the efficiency of data
audio index. Secondly, the feature of the audio library is
analysis.
used to search for candidate audio, and then audio matching
C. Algorithm capability is performed. Finally, the search results are obtained. In this
The platform can achieve resource intensive, overall process, the robustness and distinguish ability of feature
planning, and refined management, and can achieve retrieval directly determine the final retrieval effect.
resource sharing, reuse, and dynamic collaboration and
aggregation on demand.
TABLE I. LIST OF CORE ALGORITHMS.
 Core algorithm capability: The platform
integrates many algorithms including audio Category Algorithm
analysis, video analysis, image analysis, text Text clustering algorithm
analysis, and general machine learning. Table I Text classification algorithm
Text
shows all the core algorithms.
Violation text identification
 Application capability building: Based on the Similar audio fuzzy recognition
storage and computing capabilities provided by the
Similar audio clustering algorithm
artificial intelligence platform, users can conduct Audio
application research and development according to Keyword-based speech retrieval
their needs, and can test on the platform. Speech conversion text recognition
Image and video OCR recognition
III. MONITORING OF ELECTRICAL EQUIPMENT BASED ON
AUDIO RECOGNITION Image and video fuzzy matching algorithm

The sound of the normal operation of power equipment Image&video Face detection and recognition algorithm
has a certain regularity, but when the equipment fails, the Obscene and pornographic detection and recognition
sound will also change due to changes in the operating state Image and video logo detection and recognition
or operating structure. For example, when a mechanical Cyber
Network security situation awareness analysis
failure occurs, its vibration characteristics or vibration security
energy in some frequency bands will change, and it will be
accompanied by harsh or sharp noise. Therefore, the sound Audio retrieval is usually based on this assumption that
signal of the power equipment contains a lot of operating successfully matched audio clips are derived from
state information, and it can be judged whether the homologous audio. The homologous audio means that the
equipment is in an abnormal operating state through the audio has the same source but is obtained through different
learning and training of the abnormal sound of the field interference, such as compression, quantization,
equipment [9]. environmental noise, etc.

Authorized licensed use limited to: Mukesh Patel School of Technology & Engineering. Downloaded on February 25,2024 at 10:11:44 UTC from IEEE Xplore. Restrictions apply.
S u (hR , hT (n2 ))  S (hR , hT (n1 ))  (n2  n1 ) .
In this way, the window shift of the observation
window can be calculated according to R T 1 S (h , h (n ))
without moving only one frame at each time.
The following are the implementation steps of
histogram search:

Fig. 4. Framework for audio recognition. (1) Set the number of template audio frames as N ,
extract features, and count the histograms
a) Feature extraction
hR  (hR1 , hR2 , , hRB ) , where B is the number of
Feature extraction process uses sub-band energy to
bars in the histogram.
extract features, and this method requires less computation.
FFT (Fast Fourier Transform) is used to find the frequency (2) Use the moving window to take the previous frame N

spectrum
X (k ) of audio and calculate the sub-band of test audio, extract features, and count the histogram

energy:
hT  (hT1 , hT2 , , hTB ) .
1 Un (3) Calculate the intersection area of the template
e( n )   | X (k ) |2
U n  Ln  1 k  Ln
histogram and the test histogram:
B

L U S (hR , hT )   min(hRi , hTi ) 


where n and n are the lower limit and upper limit of the i 1 .
frequency label, respectively.
S (h , h )  T T
Then take the logarithm of the sub-energy band and use the (4) If R T h and h is the threshold, it is

inverse DCT (Discrete Cosine Transform) transform to considered that template speech has been detected.
transform it into the cestrum domain. This can reduce the (5) Calculate the window shift:
dynamic range of the feature, so that the effect of VQ T  S (hR , hT )  1, if S (hR , hT )  Th
(Vector Quantization) is significantly improved. Lifter w h
processing can also be used to remove the effect of 1, otherwise
.
amplitude scaling. h
b) Hash search and fingerprints match (6) As the window moves, update the test histogram T
for the newly moved frame. Go to the step (3) and
After the hash search, the histogram method is used for
repeat this process until all the test audios are searched.
refined search. The idea is to extract the features of the
template audio frame and count its histogram, and then IV. EXPERIMENTS
perform the same operation on the equal length of the audio
to be detected (it can be considered that an observation A. Center performance evaluation
window is added to the test speech). Then the intersection Streaming computing supports continuous calculation
area of the two histograms is calculated. If the area is greater and logistical analysis in real time, which can respond to
than a certain threshold, the match is successful. In this way, events and data changes dynamically. Meanwhile, it
the timing information of the audio can be lost. The two supports fine-grained logistical analysis which can better
audio histograms in the same time period may not be the provide personalized services and operations. The targets of
same. The advantage of this operation is that when the streaming computing are high throughput and low
observation window moves backward, it is only necessary to processing delay. Figure 5 shows the test workflow which
consider the effect of the newly moved in and out audio on includes data input, statistical analysis, aggregation
the histogram, which can greatly reduce the amount of operation and tables join. The test data is gradually input
calculation. into the system until increased to 1,000,000.
In addition, as each frame of the observation window
Bolt
moves, the histogram will be updated. Suppose that the
newly moved frame contributes to the intersection of the Kafka Spout Bolt Result
histogram, and the moved frame does not contribute to the
intersection of the histogram, which will increase the Bolt
intersection area by one.
If the observation window has the intersection area Fig. 5. Streaming computing test work-flow.
S (hR , hT (n1 )) n
at frame 1 , then the maximum intersection The throughput of system changes smoothly over time
area (ie, the upper bound of the intersection area) when and the average is 34,702 data/min, which can meet the
n2 needs of real-time analysis (see Figure 6). The processing
moving to frame is
delay is in milliseconds mostly and there are a few delays in
hundred milliseconds (see Figure 7). The reason is that due

Authorized licensed use limited to: Mukesh Patel School of Technology & Engineering. Downloaded on February 25,2024 at 10:11:44 UTC from IEEE Xplore. Restrictions apply.
to task scheduling, the data middleware has more resources
at a certain moment, so more data is sent to Spout. The
processing component can process these data at any time, so TABLE II. TRAINING DATA.
that the Spout module will not accumulate too much data. Category Subclass Amount
Machine wear 1,265

Abnormal Machine fouling 1,606

Noise 13396

Normal Normal 92,716

TABLE III. TEST DATA.

Subclass Week1 Week2 Week3 Week4 Week5 Week6 Week7


Machine
4,680 790 917 277 611 266 472
wear
Machine
1,615 1,731 2,668 411 1,868 787 1,227
fouling
Fig. 6. Throughput result. The test data is gradually input into the system
until increased to 1,000,000. The X-axis is time, and the Y-axis is data size. Noise 13,349 11,331 10,008 2,216 8,310 127 2
Normal 112,755 77,472 74,890 26,850 52,671 35,260 62,790

The accuracy and recall of abnormal audio are 98.85%


and 75.42% respectively (see Figure 8). The accuracy and
recall of Normal audio are 98.64% and 90.86% (see Figure
9). The overall accuracy and recall are 98.75% and 83.14%,
respectively.
The length of a single audio is about 30 seconds, and
the average processing time is about 200 to 300
milliseconds. Since iterative training is not performed on the
template library, the recall rate will gradually decrease as
Fig. 7. Processing delay analysis. The X-axis is the data size, and the Y- the abnormal audios increasing. In practical applications, it
axis is the response time.
is necessary to dynamically update the template library and
B. Experiment on monitoring of electrical equipment add newly appeared abnormal audios to the template library.
The overall accuracy of the algorithm is high (98.75%), and
Take the failure analysis of a power station boiler it has preliminary practical application conditions.
induced draft fan as an example, and use audio recognition According to the recall (83.14%), it is expected to reduce
technology to analyze the cause of the failure. The induced the amount of manual review by about 80%.
draft fan is one of the important auxiliary machines of the
power station boiler. The main reasons for the failure
include machine wear, machine fouling and noise. The
similar audio recognition algorithm is employed into the
monitoring of electrical equipment can automatically
determine and handle the similar audio to the historical
audios, and improve the recognition efficiency. Eight weeks
of data is used to test the algorithm, the information is as
follows:
 Training data: One week of audio data (about
110,000 pieces) is selected to train the algorithm and
adjust the parameters. The audio data can be divided
into two major categories: Abnormal and Normal, Fig. 8. Abnormal audio test results. The X-axis refers to the 7 weeks,
where the abnormal category includes the machine while the Y-axis refers to the accuracy rate and recall rate. Since iterative
wear, Machine fouling and Noise. The training data is training is not performed on the template library, the recall rate will
shown in Table II. gradually decrease as the abnormal audios increasing. In practical
applications, it is necessary to dynamically update the template library and
 Test data: Seven weeks of audio data (about 500,000 add newly appeared abnormal audios to the template library.
pieces) are selected to test the algorithm, which
contains the same category as the training data. The
testing data is shown in Table III.
 Test index: Accuracy represents that the algorithm can
correctly identify the proportion of similar audio.
Recall represents that the algorithm can find out the
proportion of similar audio from all similar audio.

Authorized licensed use limited to: Mukesh Patel School of Technology & Engineering. Downloaded on February 25,2024 at 10:11:44 UTC from IEEE Xplore. Restrictions apply.
artificial intelligence technology in various business
scenarios is introduced in detail, and good results have been
achieved in actual use.
In future, we will further promote the application of
artificial intelligence technologies in the field of smart grid.
Meanwhile, we will provide security protection to ensure
the safe operation of artificial intelligence applications.
ACKNOWLEDGMENT
This work is supported by 2020 State Grid Beijing
Municipality Science and Technology project "Research
Fig. 9. Normal audio test results. The X-axis refers to the 7 weeks, while and application of ubiquitous business agile access
the Y-axis refers to the accuracy rate and recall rate. technology in 5G integrated smart grid ".
V. CONCLUSION REFERENCES
This paper constructs an AI capability center based on [1] EC JRC, U.S DOE. Assessing Smart Grid Benefits and Impacts: EU
5G MEC, which can provide rich AI services for smart grids. and U.S. Initiatives [R/OL].
The center supports the dynamic construction and operation [2] D Von Dollen. Report to NIST on the smart grid interoperability
standards roadmap[J]. 2009.
of the intelligent data analysis application, and realizes the
[3] ZHOU Xiaoxin, LU Zongxiang, LIU Yingmei, et al. Development
rapid development and flexible deployment of the business models and key technologies of future grid in china [J]. Proceedings
analysis process, which could improve the interaction and of the CSEE, 2014, 34(29): 4999-5008.
response time of the business analysis process. In order to [4] LIU Zhenya. Smart grid technology [J]. 2010.
meet the needs of different types of data analysis, a fusion [5] ZHANG Dongxia, YAO Liangzhong, MA Wenyuan. Development
method of batch computing mode and streaming computing strategies of smart grid in China and abroad [J]. Proceedings of the
mode is also proposed, which can execute distributed data CSEE, 2013, 33(31):1-15.
analysis algorithms more flexibly and improve the [6] XIAO Zeqing, HUA Haochen, CAO Junwei. Overview of the
Application of Artificial Intelligence in Energy Internet[J]. Electric
robustness and flexibility of the system. Experiments results Power Construction, 2019, 040(005):63-70.
show that the platform can effectively handle batch analysis [7] ETSI GS MEC 003-2016, Multi-access Edge Computing (MEC);
and real-time analysis. The platform has unified deployment Framework and Reference Architecture.
and integration of self-developed algorithms to form the [8] O'Neill, Brian. Storm blueprints: patterns for distributed real-time
core algorithm capabilities. This article focuses on the computation [M]. Packt Publishing, 2014.
application of artificial intelligence security such as [9] ZHAI Yongjie, YANG Xu, PENG Yani. Research overview on state
monitoring of electrical equipment based on computer hearing
monitoring of electrical equipment based on similar audio technology[J]. Guangdong electric power, 2019, 032(009):24-32.
recognition, in terms of application scenarios and schemes,
algorithm principles and practical effects. The application of

Authorized licensed use limited to: Mukesh Patel School of Technology & Engineering. Downloaded on February 25,2024 at 10:11:44 UTC from IEEE Xplore. Restrictions apply.

You might also like