View of AI-Driven Framework to Optimize Smart Grid Operations, Enhance Energy Efficiency, and Facilitate Seamless Integration Using Hybrid (LSTM-CNN) Models.pdf

View of AI-Driven Framework to Optimize Smart Grid Operations, Enhance Energy Efficiency, and Facilitate Seamless Integration Using Hybrid (LSTM-CNN) Models

1. Journal of Information Systems Engineering and

Management
2025, 1085)

ago

peso mao Research Article

Al-Driven Framework to Optimize Smart Grid
Operations, Enhance Energy Efficiency, and Facilitate
Seamless Integration Using Hybrid (LSTM-CNN)
Models

Dr. Amir Jalaly Bidgoly* , Mohammed Al Yousef®,
‘Unversity of Qom, Department of Information Technology Engineering, ran
ll @gomacir

‘mohammedalyousef@uobabylonedig

ED Mode power ce emp allg BNE Be
ot Fea Pen of termi eb energy sores, tuning demand pair,
152025 su he deterioration of existing foutre hs study introduces an innovate

Accepted: 39 Febaoas Adrien framework tha integrates Long Short Term Memor? (LST) network
‘sith Conautonal Neural Network (CNX) o optimo the operon fac of
mar grid, improve ene le, and ese the seuss incorporation of
enable sure. The byl rta eee migo the shrcengs of
tao! model y concurrent analing temporal and sata etre: LSTA.
les mange nears data (eg lod demand, meteors variable) where
den pat pater ro ri topology maps and sensor networks. Ao,
layer upped ih tention mechani adapte weigh he cottons of
‘otk oda tating contest rare deci making
‘The famevork ets enhanced performance by empikal validated using rel
‘wold datt. ining high ecto Smart meer data rom he Pan See
Project, meteorlgkal monde om NOAA, and ibe ri opos ro
MANPOWER It reales an 10% enhanoament nad forecasting aca (AE =
(087) compared to standalone LSTMS and achieves a 94% accua te aime
Ha detection, thereby iminishing gi dote by 30% na silted scenario
featuring 0% sa nergy penton and ud ind aah he framework
ais vage sal within 248 of nominal val, repens corel
tol by 22% In preicion er reduction. Furthermore the system focos
predictive maintenance reulig la 35% reduction In operational expendiures
ern 6 month til conducto wt European ity gd. Ptae vetas
‘sl dee it erat earning for privacy presi ple and quant
Inspired optimization for yperarameteraihg.
Keywords Smart grid, LSTM, CNN, renevableintegmtion, spatiotemporal
nas anomaly deletion

INTRODUCTION

1: Background and Motivation
Largescale integration of renewable energy sources (RES) distributed energy resources (DER), and
bidirectional power lows deves moder power networks paradigm shit. Since renewable energy
sources (RES) ike solar and wind are intrinsically intermittent and weather-depndent, traditional gid
systems built for centralized generation using fossil fuels have dificlty managing their
173
Cort 16h An cesa SEM Ti nas re sald Ct Cos Aa
ets pm en udn mery mein pe gd oks po ed

2. Journal of Information Systems Engineering and
Management

2025, 1085)

ago

peso mao Research Article

unpredictability, For example, under cloud cover, solar production can fall by 70% in minutes,
upsetting grid frequeney and voltage [1]. Likewise, the spread of electric cars (EVs) brings
unantcipate load surges during peak charging ines. These dificuies cal for predictive ability and
real-time adaptation that traditional rule-based systems lack.

Adrien solutions and intense learning provide disruptive potential by fclitating dta-informed
decision making, For instance, Google's DecpMind achieved a 40% reduction in energy usage in data
centers by applying neural networks, ustrating As ability tenance complicated systems. However,

balance time-vaving demand pattems with spatial grid structure, smart(2] grids need more than
jst temporal foreasting—they also need spatiotemporal analysis This research adresses this gap by
fering hybrid LSTM-CNN system for brillant grid dynamic

Figures: Challenges in modern smart rds including (a renewable intermittency (>) EV-indued
Toad spikes, nd (transmission ine fas

Figure 1: Challenges in Smart Grids

1. Research Objectives
Creates hybrid framework combining STA and CNN to study spatial (eg grid topology, sensor
date) and temporal (eq, od, weather) properties,

Energy dispatch strategies might be optimized by forecasting RES generation and demand within a
limit fleas than 5 percent mean abate mistake (MAB).

Use real time anomaly detection to enhance grid rdliece, cutting the time needed to adress a
nu to eas than thre seconds,

Toscaletesting, use datasets rom many geographies (eg, Pecan Street, EU Grid) and grid sizes (10~
100 buses).

174
Cort 16h An cesa EN Tin nas re under Ct Cos Ata
ets pm en udn mery mein pe gd oks po ed

3. Journal of Information Systems Engineering and
Management

2025, 1085)

ago

peso mao Research Article

“These aims aes significant shortcomings in existing artificial intligence solutions, which ll oo
ten stress isolated spatial or temporal analyse, For instance, LSTMs are highly effective at ously
Toad forecasting [a), but they fail o consider grd topology, which results in suboptimal dispatch
decisions during ine flares.

1.3 Contributions
A unique hybrid architecture result from combining CNN for grd topology analysis and LSTM for
time-series forecasting. By combining the two result, the blend ayer uses attention mechanisms to
priori important aspects

A realtime anomaly detection module surpasses SVML-based methods by 12%, reaching 94 percent
“accuracy in spoting errors, including line outages and transformer breakdowns.

‘On mult-seae grid, empirical eating show 18% better forecast acuracy and 30% faster ul recovery
‘an top models.

“The open-source implementation of the platform supports regional grid imitations, aiding adaptability
and repestabiiy.

a. LITERATURE REVIEW
24 Alin Smart Grids
Although mos studies concentrate on minor uses, dep learning has revolutionize grid optimization

“Though they do not adjust to sudden weather changes, LSTM predic day-shead load with 90%
aura 5]

se erosin grid topology maps but have no temporal contest fo proactive maintenance (6.

Although Reinforcement Learning (RL) dynamically optimizes energy pric it demands unrealistic
Fini times surpasing one week for ast ri [7

‘Wile hybrid systems are understudied, they are sil advancing. For wind forecast, a 2022 study
combined GRU and CNN but left out grd topology data (8), restricting its usefulness for dispatch
optimization. Transformer-based models also achieve improved accuracy but demand signifiant
processing power, s they are unrealistic or real-time uses [9]

15
Cort ©0200 y Arn cesa SEM Ti nas re under Ct Cos Atco
ets pm en udn mery mein pe gd oks po ed

4. Journal of Information Systems Engineering and
Management

2025, 1085)

ago

peso mao Research Article

Figure 2: Taxonomy ofA applications in smart grids, highlighting the dominance ofisolated
temporal (LSTM) or spatial (CNN) models

Figure 2: Taxonomy of AI Applications

2.2 Research Gaps and Opportunities
‘Three major gaps persist inthe literature

1. Current theories analyze temporal and geographic data individually, excluding their interrelated
aspects. although current approaches lack thorough investigation, temporal surge in EV charging
could tax a substation serving a speiiloctin

2. Model created using data from a particular place (eg, California) often underperform in other areas
(08, Scandinavia) because to climate and infrastructural differences iting generalization,

3. Complex models such as Vision Transformers (VITS ofer better accuracy but requir tenfold the
training time a compared to LSTMs [o], therefore ipeding realtime use

"Tis sudy adresses these drawbacks using mult regional validation anda moe ficient hybrid
dein

Table: Comparative Analysis of Hybrid LSTM-CNN Models

ENT Key Pater Fertomane eier Gr

Arie

GLI | NaS ope

Shean [vin can | Dir nl cs, FR
ES

Kassa | tee arena | Raa amie

ES SS tal
[A

176
Cort 16h An Lay SEM Ts nac ne ral nd Ct Cos Aa
ets pm en udn mery mein pe gd oks po ed

5. Journal of Information Systems Engineering and

Management
2028.10)
LS és ere
Dar ar pa Research Article
GREEN | Conbines GRU and TEN Tor Oups
‘tenon Jonge dependency prediction cc and Pr
ei opens Spion iene
TESSA | Combine EST wi Fe Hp ara rc
A | asus nme m)
nengrdemandfoecating | er management
CIS | Enhances LST ih ‘Redes MAPE in solar ner
I D a (5
ONIS | Optimise ppp NE
‘tng Coat opimintion | (62% in wind poner loan
Satan a ua

3. METHODOLOGY
3 Framework Architecture
‘Theoutlned framework (Fig, ) consists of five distinct modules:

1. Data Preprocessing: Standardises diverse datasets (load, weather, rd topology) and addresses
missing vales by applying kenearest neighbors (KN.

2. LSTM Submodel This submodel analyzes time-series data through three LSTA layers cach
containing 6 units, incorporating a dropout rat of oto mitigate overfiting,

3. CNN Submodel: This submodel examine rd topology through 2D heatmaps, tiling convolutional
layers (9x3 kernels, 32 filters) and max pooling techniques,

4 Fusion Layer This layer ntegrates the output of LSTM and CNN through an atenton-based
concatenation method. The attention mechanism lloats significance to essential features, suchas
solar generation a peak hous

3. Decision Layer: This layer uses a filly connected network to suggest actions (eg. activate backup
storage and reroute power).

Figure: Framework architecture wit data low from ra inputs to decisions,

m
Cort 2 Arm cesa EN Ts nac re ua nd Cri Cos Atco
Es pm doin udn ag mein poe gd oks pa

6. Journal of Information Systems Engineering and
Management

2025, 1085)

ago

peso mao Research Article

3.2 Mathematical Formulation

LSTM Gates:
lo oleo] +b) Anput gat)
Fe mo leazd+b) Forget gate) o
& oo (Pax) +2.) Output gate)

CNN Feature Maps:
Die RLU ta on +) o

Fusion with Attention:
Sofimax( (one hoe). Mini = sn 0) ho 0)

Figure 4: Attention mechanism in the fusion Laye, showing weighted contributions of LSTM and
ES

Figure 4: LSTM-CNN Fusion Mechanism

3:3 Datasets and Preprocessing

Data Sources:

Load and Distributed Energy Resources (DER) Dat: Pecan Steet Dataset

1000 houses) 7

Weather Data: NOAA (temperature, humidity sola iradianc),

rid Topolog: Synthti 0-bus and 18-bas systems derived from MATPOWER [18], annotated with

line capacity and alae records.

Steps for Preprocessing

“Temporal Alignment: Resampleall data to15minate intervals

‘Normalisation: Implement Min-Max scaling fr lad data (ranging rom

‘normalization for weather data.

‘Topology Encoding; Transform grid configurations into aD matrices, cach pixel denotinga bus rin.
178

Cort 16h An Lay SEM Tin nas re sad Ct Cos Aion
ets pm en udn mery mein poe gd oks po ed

7. Journal of Information Systems Engineering and

Management
2025, 1085)

ago

peso mao Research Article

“Tables Summary of datasets and preprocessing techniques.

‘Table: Datasets and Preprocessing
Dataset Jeune [Rouen | Preprocessing
Toad Data | Penso | rnin | RENN Imputaton, Normal
WesiherData [NOR Hou |Zscore Sealing
Grid Topology | MATPOWER [N/A | aD Mates Encoding

A RESULTS

4.1 Load Forecasting
“The ybrid model decreased the Mean Absolute Error (MAE) by 18% in comparison tothe slo Long
Short-Term Memory (LSTM model Table 2) During a heatwave in Tess, Ihehybrid model forested
120% lod increase four ours in advance allowing pronctve ri tabiizaion. The LSTM-only model
failed to detet the surge owing to inadequate spatial contest, such as regional aie conditioning
consumption tends

Figure 5: Forscastvs actual loud during heatwave event.

qu...

Figure 3: Load Forecasting Results

42 Renewable Integration Case Study
Imovercst weather, grid with 40% solar penetration vs simulated The hybrid model achieved a 22%
decrease in sola forecating error compared to ARIMA while maintaining voltage stability within 25%
of nominal (Fg. ©). Without the model, voltage fluctuations exceeded 12%, tigerng safety relay

179
Cort 16h An Lary SEM Ts nas re under Ct Cos Atco,
ets pm en udn mery mein pe gd oks po ed

8. Journal of Information Systems Engineering and
Management

2025, 1085)

ago

peso mao Research Article

Figure 6: Voltage profiles (a) with and () without the hybrid mode during loud cover.

ra
>
F
5
de
Le a

|

Figure 6: Voltage Stability During Cloud Cover
43 Anomaly Detection

“The framework detected 49/52 line issus inthe 18-bus system with 94% accuracy and a mean
monitoring time of 23 seconds. In contrast, SVM-based approaches attained an accuracy of 82% but
required, seconds whieh presents ra eascade failures. The CNN submodel identified failure in
specifi grid ares, while the LSTM detected corresponding temporal regularities, including sudden
Heed reductions.

“Table 2: Performance comparison across models (MAE, RMSE, Accuracy, P-Score).
‘Table a: Model Performance Comparison

po MA | RMSE | 66) Accuracy | F-Score
snr us [158 [82 075
EN 195 [168 178 o7

5. DISCUSSION
4 Practical Ramifiations

(Cost Reduction: Peditve maintenance reduced operating expenses by 35% during a 6-month
experiment with a European ut.

RES Integration: The system allowed 55% integration ofrenewables in simulations, according to EU
2030 objectives.

Scalability: The model was rained on,000-bus ridin around 4 hours using single GPU,
demonstrating industria ab.

180
Cort 16h An Lay SEM Ti nas re ral nd Ct Cos Aa
ets pm en udn mery mein pe gd oks po ed

9. m

Journal of Information Systems Engineering and
Management
2025, 1085)

pesos salen Research Article

52 Limitations
Data Dependeney: Performance decreased by 8% when asesed on grids with limited sensor coverge
Edge Deployment Realtime execution on edge devices (eg, Raspberry PD required model
quantization, resulting ina 9% decreas in accuac:

Figure 7: Trade-off between model accuracy and computational load.

Training Time (Hous)

5.9 Future Work
Federated Leaning: Tn the model using decentralied data while preserving privacy.

Quantum: Inspired Optimisation: Espeditehyperparsmeter tweaking by quantum annealing.
Hardware Sofware Co-despn: Create ASIC special designed for LSTM-CNN integration.

The hybrid LSTML-CNN architecture signifies notable advance in Al-enhanced Innovative gid
management. Using the advantages ofboth CNN and LSTMs, the mode tains elevated accuracy in
energy forecasting, facitates the incorporation of renewnble energy sources, and enhances grid
‘operations. Its uses include household, commercial, and grd-sal energy management, making ita
‘exible instrament for contemporary smart grids As the energy landscape evolves, more research nto
optimization approaches, salty, and iterpretabity il maintain the frameworks relevance and
‘tty.

REFERENCES

Bostan, A, Kulshrshtha, K, Agatha, A À, Kartik, K, Saath, K, Pawar, À M, & Ashreth, B.
(2024). Adaptive Energy Management Sytem for Electric Vehicle Charging Stations: Leveraging Alor

1

10. 15]
sl

ta
6
16)
a

0]

ol
Lo]
tu)

ha

bal
bal
bs

us

in

bs]

Journal of Information Systems Engineering and
Management

2025, 1085)

ago

peso mao Research Article

‘Real-Time Grid Stabilization and Efficiency In EgS Web of Conferences (Vol. 591, . 04002). EDP
Scenes

sae, EAE, Ko, VV, e Grigoe A. A (2029). Data centr efficiency model: a new approach
“and the role of atic eigene. Mareusmavecxan Guonorta u Gnounopnanu, 180), 215-227.
1, L Wo, Zhang, J, Zhang, L Tan, 8 Tan, Z. (202). VMD and LSTM based hybrid mode of
load forecasting for power grid security. IEEE Transactions on Industrial Informatis, 189), 6474-
6483,

Alene, M, Ana, F,Packinather M, & Shouran, M. (2024) Enhancing transformer protection: A
‘machine learning framework for cry ful detection Sustainability, 1629), 10759.

Qing, X. & Ni, Y. (2018. Hourly day-ahead solr irradiance prediction using weather forecasts by
STM, Enero, 148, 461-468.

Almasoudi, F. M. (2035). Enhancing power grid rence through real-time fault detection and
‘remediation using advanced hybrid machine learning models. Sustainability, 1510), 8248.

Kim, B.G., Zhang, Y, Van Der Schaar, M, & Lee, J. W. (2015). Dimamie pricing and energy
consumption sebedulin with reinforcement learning. IEEE Transactions on smart ri, 75) 2187
2198.

Dhungans, H. (2035). A machine learning approach for wind turbine power foreasing for
‘maintenance planning, Energy Informati, 8), 2-

Hemmat, A Bazar, Rahman A.M. & Moosad, H. (2025) A Systemati Review on Optimization
Approches for Transformer and Large Language Model. Authorea Preprints.

‘Alomar, K, Aysel LL, Cu, X. (2024) RNNS, CNN and Transformers in Human Acton Recognition:
A Survey and A Hybrid Model arXiv preprinariv2407.06162.

BARTOULI, M, HELALI,A, & HASSEN, F.(2024, Ape). Applying Bayesian Optimized CNN-BILSTM
to Real-Time Load Forecasting Model for Smart Grids. In 2024 IEEE Intemational Conference on
Advance Systems and Emergent Technologies (IC_ASET) (pp 1-6). IEEE.

‘Ashore, MF, Habib, Des D. À, Alattah, W. Ilm, M, & Alabl,. (2032). Imposing the
eficieney ef must "erde electricity load forcasting via REN with ML
ESTA. Sensors, 22018), 619

Wen, X. Liao J, Nin, Q, Shen, N, & Bao, Y. (2021) Dep Iearning-driven hybrid model fr short-term
load forecasting and smart grid information management Scene report, 10, 19720

Baviiet, D. P. AiEnabled Metaheurstie Optimization for Predictive Management of Renewable
Energy Produetion in Smart Grids.

Zafar, A, Che, ¥,Sehnan, M, Al, U Algar, A.D, & Elmannal, H. 2029). Optimizing solar power
generation forecasting in smart rds: À Hybrid convoluional neural network autoencoder long shor
term memory approach. Physica Seripis, 000), 095249.

Abu Hour, M, Bari, 5. M.S, Zar, M.H. Mansoor, M, & Chen, W. (2023). COA-CNN-ASTM:
Coat optimization algorithm-based hybrid deep learning model for PV/wind power forecasting in
mat grid applications Applied Energy, 349, 121638.

Hossan, 5, Zulkefl, A, Suanarıyan, $, & Nair A. (2024) Riscinformed Hierarchical Control of
[Behind-the Meter DERs with AMI Data Integration (Final Technica Report) (No. DOE Eston-09023)
Eaton Corporation, Golden, CO United States)

hanes, Turin, K, & Chertkon, M. (201, Ju). Saiten lasifiaton of easeading fire |
power grid. In 201 IEEE Power and Energy Society General Meeting (pp. 1-3) IEEE.

182
Cort 16h An Lay SEM Ts nas re ua nd Ct Cos Ati,
ets pm en udn mery mein pe gd oks po ed