Smart Grid Stability Prediction with Machine Learning

GIL-VERA VICTOR DANIEL

Faculty of Engineering, Luis Amigó Catholic University,

Trans. 51 A N° 67 B-90,

COLOMBIA

Abstract: - Smart grids refer to a grid system for electricity transmission, which allows the efficient use of

electricity without affecting the environment. The stability estimation of this type of network is very important

of power supply.

Key-Words: analysis; artificial intelligence; control, machine learning; smart grid; stability.

Received: June 28, 2021. Revised: July 15, 2022. Accepted: September 16, 2022. Published: October 6, 2022.

1 Introduction

Smart grids are networks that control power

delivery and provide several advantages, including

the development and effective management of

distribution network, reducing costs, and saving

energy [3].

In essence, a smart grid has advanced technology

and incorporates information and communication

technologies (ICT), utilizing technology for

metering, communications, and control in the

facilities' generating, transmission lines, substations,

feeders (circuits), and meters [4]. The objectives of

smart grids are; to generate faster performance for

the benefit of the end consumer (services, tariffs,

of transmission networks to attacks or failures and

facilitate their rapid location in urban and rural areas

[5].

According to [6], modern electric power systems'

technical and commercial disturbances are often

referred to as "smart grid", encompassing everything

integrated into them, what uses the grid services and

what interacts with them. On the other hand, [7]

defines them as a complex system of technological,

electricity trading, and service subsystems

have close ties to other energy sources and domains

due to the coupling of sectors and electrification of

energy domains like building heating and cooling,

transportation, and industrial processes [7]. The key

to making the best use of abundant energy resources

is smart grid engineering, which enables the

efficient dispatching of power generated by hybrid

renewable energy sources (RES) over long distances

via DC transmission lines using high voltage DC

(HVDC) transmission technology [8].

business culture, technology, process maturity, and

the current market, as well as the socioeconomic and

environmental situation of their concession region,

are what drive the deployment of smart grids [10].

These generate benefits for utilities, better grid

management, increased customer choice, greater

understanding of energy use, reduced electricity

cost, increased communication with customers and

themselves to a more stable and effective system,

and their primary functions include real-time

monitoring and reaction, allowing the system to

constantly change to an ideal condition. This is one

of its key qualities [12]. Self-healing enables them

to identify anomalous signals, carry out adaptive

reconfigurations, and isolate disturbances, reducing

or eliminating electrical disturbances during storms

and disasters. They can also reduce power outages

and shorten their duration when they do occur [13].

While grid operators manage the system's

balance, provide supply stability and security,

physically connect producers and consumers and

facilitate energy transactions, smart grids also

provide services that enable an electricity system's

efficient and secure running [15]. Smart grids aim to

improve the functioning of energy markets, use

existing transmission infrastructures more

effectively, increase the capacity of renewable

energy sources, electric vehicles, heat pumps, and

information to determine their usage because the

entire process is time-dependent, it is crucial to

understand and plan for disturbances and

fluctuations in energy consumption and production

introduced by system participants dynamically,

considering not only technical considerations but

also how participants react to changes in energy

prices [17].

dynamic security assessment and prediction are

in electric power distribution and generation

systems, the frequency of the electric signal

increases in times of excess generation, therefore,

measuring the frequency of the grid at each

customer's location is sufficient to give the manager

the necessary information on the present grid energy

balance, so that it can price its energy supply and

alert consumers, while it reduces in times of

underproduction [19].

improve the predictability of the grid stability

prediction model, which obtained high performance.

In the research conducted by [22], they predict the

stability of smart grids using multidirectional short-

term memory (LSTM). Meanwhile, [23] employed

a symmetric non-negative latent factor model based

on matrix factorization. In the research developed

by [24], they concluded that neural networks can

achieve high performance in predicting network

stability; however, they claim that most existing

different models to predict the stability of smart

grids using machine learning techniques (Random

Forests, Support Vector Machine (SVM), Logistic

Regression, K-Nearest Neighbors (KNN), Decision

Trees, ANN-MLP, Naïve Bayes), compare the

performance of each technique and identify the

optimal one to predict the stability of this type of

grids. The utility of this study in practical

applicability is the identification of the optimal

technique in terms of accuracy that can help smart

unknown. Machine learning and data analytics are

interdependent and related fields of study that

primarily focus on acquiring decisive knowledge

[26]. Models are developed using training data and

evaluated with test data. Machine learning is

currently widely employed in many fields of

knowledge to generate predictions and facilitate

decision-making. The objective of Machine

Learning is to let computers learn how to carry out

tasks without being explicitly taught to do so [27].

be used for classification (to predict the membership

of a class or label) and regression (to predict a

numerical value) tasks. Threesome several

specialized tools or programs allow the use of

machine learning; some of them are Keras,

TensorFlow, KNIME, Shogun, IBM Watson,

Apache Mahout, R, Apache Spark MLlib, Weka,

Oryx 2, RapidMiner, H20.ai, and Pytorch.

Support Vector Machine (SVM), Logistic

entries as the class label for the new data point [29].

the feature on which it will be divided, and (T, X)

the entropy calculated after dividing the data on the

feature X. Random Forests is a technique based on

decision trees, which are assembled by bags and

trained independently [32]; this technique forecasts

an output based on features using a collection of

decision trees. The prediction is the outcome of

consecutive binary decisions that are divided

orthogonally in the multivariate space of variables;

in essence, it is a meta-learning of numerous

neurons, which work simultaneously to process

information. The activation function (softmax, tanh,

relu) is in charge of returning output from an input

value, often the set of output values in a certain

range such as (0,1) or (-1,1) [33]. As universal

approximators, multilayer perceptrons are neural

network models that can approximate any

continuous function. They are made up of

perception, which is neurons. A perceptron takes n

characteristics as input (x = x1, x2, ..., xn), and each

Smart grids are the future of energy supply. Their

instability can cause problems in the supply of

energy to consumption nodes, for this reason, is

important to predict their stability. In this type of

network, generation must match demand at all

times, a reserve must be maintained for immediate

outages, and sufficient capacity must be provided

study aimed to compare various machine learning

approaches to identify the best technique for

predicting a smart grid's stability. The database used

contains the results of stability simulations of a star

network (three consumption nodes and one

generation node) as presented in Fig.3.

4 Problem Solution and Discussion

A free database accessible from the following link

was used to build the models: https://onx.la/46d79,

the dataset contains 60,000 observations, twelve

primary predictive characteristics, and one

dependent variable. The database's structure is

shown in Table 1.

Table 1. Database structure

variations in price data, and the reaction time refers

to the response time of network participants to

adjust consumption and/or production in response to

price changes, and the power balance refers to the

nominal power produced or consumed at each

network node. The models were trained in a ratio of

75/25 (75% for training and 25% for testing), thanks

to this division it is possible to identify the accuracy

of the models, which were developed in Python

using Google Colab. This tool provides free virtual

Services. These Google virtual machines are

restarted every 12 hours, allow running and

programming in Python in a web browser, do not

require configuration, allow free access to Graphics

Processing Units (GPUs), and allow sharing content.

This tool can be used by students, data scientists, or

artificial intelligence researchers.

blending of executable code and rich text in a single

document, as well as graphics, HTML, and LaTeX.

importing own data from a Google Drive account

and GitHub, it also allows importing image datasets,

training image classifiers, and evaluating

classification and regression models. It should be

noted that these notebooks run code on Google's

cloud servers, which allows taking advantage of the

power of Google hardware regardless of the

computer power on which it is used. Table 2

presents the libraries and optimal parameters for

each of them.

representation of the prediction’s outcomes made,

was used to assess the accuracy of the constructed

models (Table 3).

were also positive in the real values.

were not negative in the real values.

were not positive in the real values.

metrics presented in equations (10), (11), (12), (13),

(14), and (15) were calculated.

Accuracy: percentage of correct predictions.

Sensitivity, Exhaustiveness, or Recall:

Receiver operating characteristics curve (ROC):

differentiate between classes, and AUC = 0 means

score to the model with the lowest score.

(Accuracy=0.935, F1-Score=0.948). Other models

grid zones depending on energy usage (high/low),

each zone was subdivided into several subzones and

assigned to Random Forest branches. In this

research, the authors confirm the effectiveness of

this technique compared with others (SVM, K-NN,

and Naïve Bayes) and conclude that it can identify

the exact location of energy availability in minimum

time, which allows providing quick responses to

grid users.

In the research developed by [35] on the

from clinical datasets, where they highlight that the

Random Forests technique is one of the most

accurate.

It should be noted that the objective of

employing this type of technique in predictive

modeling is that they discover by themselves

patterns that generalize well the data that were not

analyzed instead of memorizing data that they

learned during training; all accuracy metrics should

be evaluated to decide which is the best and not only

technique obtained the best results in the metrics

that were studied (Accuracy, Precision, Recall,

Specificity, and F1 Score). When one class is less

frequent than others, this technique can

automatically balance data sets; it is less

computationally expensive and does not require a

graphics processing unit (GPU). This technique is

commonly used in classification exercises since,

unlike artificial neural networks, it doesn't need a lot

of data to be effective. However, it is not correct to

optimal parameters, and inadequate processing can

considerably affect its performance, is very

important to normalize the data, fill in missing data

with null values and eliminate inconsistencies

before training the classification models.

constructing predictive models using combined

Machine Learning techniques (Bagging, Boosting,

Random Subspaces, and others) and compares

presented in this work. Finally, Google Colab

[5] Panda, D. K., & Das, S., “Smart grid architecture

[6] Dileep, G., “A survey on smart grid technologies

[7] Selvam, M. M., Gnanadass, R., & Padhy, N. P.,

[9] Mollah, M. B., Zhao, J., Niyato, D., Lam, K. Y.,

[10] Liu, D., Zhang, Q., Chen, H., & Zou, Y., “Dynamic

[12] Fan, D., Ren, Y., Feng, Q., Liu, Y., Wang, Z., &

[13] Ashrafi, R., Amirahmadi, M., Tolou-Askari, M., &

[15] Emmanuel, M., Rayudu, R., & Welch, I.,

[17] Babar, M., Tariq, M. U., & Jan, M. A., “Secure and

[18] Mukherjee, R., & De, A., “Development of an

[20] Breviglieri, P., Erdem, T., & Eken, S., “Predicting

[21] Massaoudi, M., Abu-Rub, H., Refaat, S. S., Chihi,

[24] Massaoudi, M., Chihi, I., Sidhom, L., Trabelsi, M.,

[26] Ibrahim, M. S., Dong, W., & Yang, Q., “Machine

[27] Lei, Y., Yang, B., Jiang, X., Jia, F., Li, N., & Nandi,

[28] Kotsiopoulos, T., Sarigiannidis, P., Ioannidis, D., &

[31] Naganandhini, S., & Shanmugavadivu, P.,

[32] Golden, C. E., Rothrock Jr, M. J., & Mishra, A.,

[33] Heidari, A. A., Faris, H., Aljarah, I., & Mirjalili, S.,

[34] Durairaj, D., Wróblewski, Ł., Sheela, A.,

[35] Vafeiadis, T., Diamantaras, K. I., Sarigiannidis, G.,

Contribution of Individual Authors to the

Creation of a Scientific Article (Ghostwriting

Victor Daniel Gil Vera has performed the

normalization of the database, trained the predictive

models in Python, and performed the statistical

analysis.

This research was funded by the Universidad

Católica Luis Amigó and was one of the results of

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en

_US