Hybrid ML-RL Approach for Smart Grid Stability Prediction and Optimized Control Strategy

Anandi Dutta; Kazi Sifatul Islam; Shivani Mruthyunjaya

REVIEW 4 major objections 5 minor 21 references

Hybrid ML-RL Approach for Smart Grid Stability Prediction and Optimized Control Strategy

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A stacking classifier plus DQN restores smart-grid stability with 100 percent success in 44 episodes.

desk verdict The ML baseline is a routine benchmark, but the RL control claim is unsupported because the environment's transition dynamics are never specified. read the letter →

arxiv 2508.19541 v1 pith:DLTM3F7M submitted 2025-08-27 eess.SY cs.SY

Kazi Sifatul Islam , Anandi Dutta , Shivani Mruthyunjaya This is my paper · ORCID

classification eess.SYcs.SY

keywords smartgridstabilitystackingensemblereinforcementlearningDQNPPOA2CpredictionhybridML-RL

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that splitting smart-grid stability work into two stages—machine learning for fast prediction and reinforcement learning for control—is faster and more reliable than either approach alone. It trains several ML classifiers on an augmented simulated grid dataset, selects a stacking ensemble as the best predictor, and then applies PPO, A2C, and DQN agents to unstable states. The reported result is a hybrid Stacking+DQN pipeline that reaches 100% success in restoring stability, converges in 44 episodes, and requires the least training time of the three RL agents. If true, this would offer a practical template for real-time smart-grid control that combines rapid instability detection with adaptive corrective action.

What carries the argument

The key mechanism is the two-stage hybrid pipeline. Stage one is a stacking ensemble classifier that predicts stable versus unstable grid states from a 13-dimensional state vector (reaction times, power values, gamma function values, and a stability indicator). Stage two is a reinforcement-learning agent that acts on predicted unstable states by choosing among three discrete actions (decrease, maintain, or increase power) and receives a reward of +20, -20, or 0 depending on whether stability improves. DQN's fast convergence is attributed to experience replay and a fixed target network, which let it learn an effective stabilization policy with fewer environment interactions.

What would settle it

Run the Stacking+DQN pipeline on a simulator with an explicitly defined state-transition function, or on logged real grid disturbances, and check whether the 100% success rate and 44-episode convergence reproduce; if not, the reported control results are environment-specific.

Watch

Extended reading notes

Core claim

The paper claims that a hybrid pipeline—a stacking ensemble classifier for stability prediction followed by a reinforcement-learning controller for stabilization—outperforms both ML-only and RL-only approaches on the augmented UCI smart-grid dataset. The stacking classifier, built from Random Forest, XGBoost, LightGBM, and ANN base learners with logistic regression as meta-learner, achieves the most balanced performance with F1-scores of 0.98 for both stable and unstable classes. In the control stage, DQN stabilizes the grid in all 100 test episodes, converges in 44 episodes, and has the lowest training time, making the Stacking+DQN configuration the paper's recommended hybrid model.

Load-bearing premise

The results stand on the assumption that the custom simulation environment accurately represents smart-grid dynamics, but the paper never specifies the rule that moves the grid from one state to the next.

Editorial extensions

If this is right

If the Stacking+DQN pipeline holds up, a fast classifier can gate a reinforcement-learning controller, reducing the state space the RL agent must explore and making real-time stabilization more practical.
The reported results suggest DQN is more suitable than PPO or A2C for deployment scenarios where training time and episode budget are constrained.
The augmented 60,000-instance dataset appears to raise achievable accuracy compared with prior 10,000-instance studies, indicating that dataset augmentation itself is a significant contributor to the new benchmark numbers.
The hybrid design implies that operators could run the lightweight classifier continuously and invoke the RL controller only on predicted instability events, lowering overall computational load.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper's claims, the 100% success rate and 44-episode convergence are likely sensitive to the specific reward structure and three-action discretization; a more granular action space or a noisy reward signal would probably increase episodes-to-convergence and reduce the success rate.
Beyond the paper's claims, the comparison with prior work is not strictly apples-to-apples because the augmented dataset changes the training distribution; a reader should not attribute all accuracy gains to the model architecture.
Beyond the paper's claims, a natural testable extension would be to replace the simulated environment's implied transition rule with a physics-based grid simulator to see whether the same hybrid pipeline retains its speed and reliability.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, here's my read on arXiv:2508.19541.

The ML half is a straightforward benchmark on the UCI grid stability dataset, and it's fine as far as it goes. The RL half is not salvageable as written: the paper never defines the environment's transition dynamics, so the 100% success rate and 44-episode convergence are meaningless numbers. This matches the stress-test note.

What's actually new: nothing much. The stacking classifier hits 97.88% accuracy, which is consistent with published results on the same data. The augmentation from 10k to 60k rows is standard practice. The paper is readable, the figures are clear, and the ML comparison includes several standard baselines. That's about it.

The soft spots are serious. In Section IV-B.1, the state vector includes stab, the action set is given as {0,1,3} then described as 0/1/2, and the reward compares stab_new to stab_old. There is no P(S'|S,a) anywhere. Without it, the Bellman equation in Eq. (4) has no operational meaning. The statement that RL agents are "iteratively tested on the unseen data" confirms the authors treat RL as a classifier, not a control loop. Eq. (4) is also wrong: it uses Q(s',a') instead of Q(s,a) as the old value. The DQN description refers to "Equation (1)", which is majority voting. These are not typos; they indicate the RL section was never actually implemented as described. The ML comparison in Table II is also biased: prior work is on 10k instances, this paper uses 60k, and there are no error bars. So the claim of a new benchmark is not established.

Who should read this? Someone looking for a quick survey of ML baselines on the UCI stability dataset might skim Part 1. But the control results should not be cited or trusted.

My recommendation: this would not survive serious peer review as is. It needs either a real environment specification with reproducible code, or the RL claims should be removed entirely. I would not send it to a good journal; it's a desk reject for the load-bearing flaw. If the authors fix the environment question, it might become a modest conference paper, but that's speculative.

Referee Report

4 major / 5 minor

Summary. This paper proposes a two-stage hybrid ML-RL framework for smart grid stability. Stage 1 uses a stacking ensemble (Random Forest, XGBoost, LightGBM, ANN with a logistic regression meta-learner) to classify grid states as stable/unstable on the UCI Electrical Grid Stability dataset, augmented from 10,000 to 60,000 instances. Stage 2 applies RL algorithms (PPO, A2C, DQN) to select power control actions to restore stability, with a reward based on the change in the stability indicator. The authors report the stacking classifier achieves 97.88% accuracy and F1 scores around 0.98, and the Stacking+DQN hybrid achieves 100% success rate with convergence in 44 episodes and the lowest training time. The paper claims this hybrid approach effectively stabilizes the grid in real-time while reducing training time compared to RL alone.

Significance. The problem is relevant: combining rapid ML-based instability detection with RL-based corrective control is a plausible design for smart-grid operation. The ML classification stage is based on a reproducible public dataset and uses standard metrics, which is a positive aspect. However, the RL part lacks a well-defined environment: no transition dynamics are specified, and the agent is described as being tested on 'unseen data,' suggesting a static dataset rather than an interactive closed-loop system. The central control claim—100% success and 44-episode convergence for Stacking+DQN—is therefore unsupported. No code, data, or reproducibility artifacts are provided, and the comparison with prior work is confounded by dataset augmentation. The paper does not meet the standard for a systems-control publication in its current form.

major comments (4)

[Section IV-B.1 (RL Environment Setup)] The transition function P(S_{t+1}|S_t,a_t) is never specified; the paper defines only the state space, action space, and reward. The text says agents are 'iteratively tested on the unseen data,' implying a static labeled dataset rather than an interactive environment. Without a rule mapping actions to next states (especially how stab_new is computed), Eq. (4) has no operational meaning, and the 100% success rates and 44-episode convergence in Table III cannot be verified or interpreted as closed-loop control results. This is the central load-bearing flaw.
[Section IV-B.1, Eq. (4)] The Q-learning update is misstated. Standard Q-learning subtracts the current Q(s,a), but Eq. (4) subtracts Q(s',a'). Also, the text says DQN follows the Bellman equation 'indicated in Equation (1)', but Eq. (1) is the random forest majority-vote rule. These errors indicate the RL formalism is not carefully handled and undermine confidence in the reported RL results.
[Section IV-B.1 (Action Space)] The action set is written At = {a1=0, a2=1, a3=3}, but the description says a1=0, a2=1, and a3=2 for Decrease/Maintain/Increase Power. This internal inconsistency makes the action semantics ambiguous; since actions are what the RL agent outputs, the ambiguity must be resolved before the control claim can be evaluated.
[Section V-A, Table II] The comparison with prior work is confounded because previous studies use 10,000 instances while this study uses 60,000 augmented instances. The accuracy gain may reflect added data rather than model superiority, so the claim of setting a new benchmark is not supported. In addition, Fig. 6 reports no error bars or repeated-run variance for any ML model, so differences such as 97.88% vs 98.21% cannot be assessed for statistical significance.

minor comments (5)

[Section IV-B] The section numbering repeats '1)' for both 'RL Environment Setup' and 'Algorithms Used'; use distinct sub-numbering.
[Section IV-A.5] Stacking is described as 'bagging all the base learners'; stacking is not bagging. The wording should be corrected.
[Section IV-B.1.c] The reward condition 'stab_new ≈ stab_old' is not defined quantitatively; specify a tolerance or use exact equality.
[Section V-B] Fig. 7 caption says 'including Hybrid ML-RL', but the text compares only RL agents; clarify whether success rate is for the full stacking+RL pipeline or the RL stage alone.
[References] Reference [20] is incomplete ('p. to be added'). Also, no data or code availability statement is provided.

Circularity Check

2 steps flagged · score 6.0 of 10

RL control claims reduce to dataset-label reward because no environment transition is specified.

self definitional [Section IV-B.1c (Reward Function) and Eq. (4)]
"R(a_t, S_t) = { +20, if stab_new > stab_old; −20, if stab_new < stab_old; 0, if stab_new ≈ stab_old } The reward function encourages the agents to learn the environment according to the reward and penalty values."

The reward and the claimed success are both expressed as changes in the single variable stab. The paper never defines how an action a_t transforms S_t to S_{t+1}; no transition function or simulator is given. Because the agents are "iteratively tested on the unseen data" (a static table), the only computable stab_new/stab_old values come from the dataset's label column. Thus the policy is rewarded for matching or selecting already-labeled stable rows; "stabilizing the grid" reduces to the label-dependent reward, making the 100% success rate an artifact of the training objective rather than a closed-loop control result.
fitted input called prediction [Section V-B (Performance Analysis of RL-based Control Stage)]
"The success rate in this stage reflects how effectively each RL model restores grid stability after instability is predicted by the stacking classifier in the ML stage. ... PPO achieved a 98% success rate, meaning it was able to successfully stabilize the grid in 98 out of 100 test episodes."

The success rate is presented as a control or prediction outcome, but it is computed on the same UCI dataset whose stab labels define both the reward and the ML classifier's target. Without an environment transition, an episode is just a row (or sequence of rows) of the static data; "successfully stabilize the grid" therefore names a label-based match, not an independent dynamical outcome. This is a fitted value learned from the stab-dependent reward, reported as a predicted control result.

full rationale

The ML stability-prediction leg is self-contained: it trains standard classifiers on the public UCI electrical-grid dataset and compares accuracy with prior work on the same data; that is conventional supervised learning and not circular. The circularity is confined to the RL control leg. The reward function in Section IV-B.1c is defined entirely on stab_new versus stab_old, and the state vector already contains the same stab variable, but no transition function or grid simulator is specified that would let an action determine stab_new. The paper instead says the agents are "iteratively tested on the unseen data," i.e., a static table. Consequently, the Q-update in Eq. (4) has no operational environment; the only way to evaluate the reward is to read stab labels from the dataset. A policy trained to maximize that reward is therefore fitted to the dataset's labels, and the reported 100% success rate and 44-episode convergence are properties of that label-fitting procedure, not of a validated closed-loop control system. This makes the central "effective stabilization" claim reduce, by construction, to the label-based reward defined on the training data. The ML benchmark claims remain unaffected, so the overall circularity is substantial but not total.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper's core contribution depends on a reward function with hand-picked constants and an undefined simulation environment. The dataset is a standard public resource, but the RL setup introduces a black-box controller that is not independently verifiable.

free parameters (2)

Reward function constants = +20 for improvement, -20 for worsening, 0 for no change
Chosen by hand to shape learning; no sensitivity analysis is provided.
Augmentation factor = 6 (3! permutations)
The dataset is expanded from 10,000 to 60,000 instances; this choice directly affects the reported accuracy and the comparison with prior work.

assumptions (3)

domain assumption UCI Electrical Grid Stability dataset represents real smart grid behavior
The paper uses this simulated dataset as ground truth for both ML and RL without validating against physical grid data.
ad hoc to paper The unseen RL environment transition function is well-defined and controllable
The paper never specifies how an action a_t changes the state S_t to S_{t+1}; this is the load-bearing premise of the control claim.
domain assumption Reward based on stab_new > stab_old guides the agent to stable operation
The reward is a heuristic proxy; no proof or simulation detail shows it corresponds to real grid stabilization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hybrid ML-RL Approach for Smart Grid Stability Prediction and Optimized Control Strategy." pith.science (2026). https://pith.science/paper/DLTM3F7M

@misc{pith2026250819541,
  author       = {Pith},
  title        = {Pith review of: Hybrid ML-RL Approach for Smart Grid Stability Prediction and Optimized Control Strategy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLTM3F7M}},
  note         = {Machine review of arXiv:2508.19541}
}

read the original abstract

Electrical grids are now much more complex due to the rapid integration of distributed generation and alternative energy sources, which makes forecasting grid stability with optimized control a crucial task for operators. Traditional statistical, physics-based, and ML models can learn the pattern of the grid features, but have limitations in optimal strategy control with instability prediction. This work proposes a hybrid ML-RL framework that leverages ML for rapid stability prediction and RL for dynamic control and optimization. The first stage of this study created a baseline that explored the potential of various ML models for stability prediction. Out of them, the stacking classifiers of several fundamental models show a significant performance in classifying the instability, leading to the second stage, where reinforcement learning algorithms (PPO, A2C, and DQN) optimize power control actions. Experimental results demonstrate that the hybrid ML-RL model effectively stabilizes the grid, achieves rapid convergence, and significantly reduces training time. The integration of ML-based stability classification with RL-based dynamic control enhances decision-making efficiency while lowering computational complexity, making it well-suited for real-time smart grid applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

[1]

Smoliński, A., Smoliński, A., & Bąk, A. (2022). Clean Coal Technologies as an Effecti ve Way in Global Carbon Dioxide Mitigation. Energies, 15(16), 5868.J

work page 2022
[2]

A distributed control approach for enhancing smart grid transient stability and resilience,

M. Ayar, S. Obuz, R. D. Trevizan, A. S. Bretas, and H. A. Latchman, “A distributed control approach for enhancing smart grid transient stability and resilience,” IEEE Transactions on Smart Grid, vol. 8, no. 6, pp. 3035–3044, 2017

work page 2017
[3]

Improving the stability of an interconnected power system using genetic eigenvalue technique,

U. K.I, “Improving the stability of an interconnected power system using genetic eigenvalue technique,” Journal of Energy Technologies and Policy, vol. 11, no. 1, pp. 22–41, 2020

work page 2020
[4]

Research on stability of the power system,

D. Xie, D. Zang, and P. Gao, “Research on stability of the power system,” in Proc. 2nd Int. Conf. on Materials Science, Energy Technology and Environmental Engineering (MSETEE), Zhuhai, China, Apr. 2017, IOP Conf. Ser.: Earth Environ. Sci., vol. 81, p. 012163. doi: 10.1088/1755-1315/81/1/012163

work page doi:10.1088/1755-1315/81/1/012163 2017
[5]

Performance monitoring and redesign of power system stabilizers based on system identification techniques,

X. Bombois and L. Vanfretti, “Performance monitoring and redesign of power system stabilizers based on system identification techniques,” Sustainable Energy, Grids and Networks, 2023

work page 2023
[6]

8. optimized lstm for accurate smart grid stability prediction using a novel optimization algorithm,

F. K. Karim, D. S. Khafaga, E.-S. M. El-kenawy, M. M. Eid, A. Ibrahim, L. Abualigah, N. Khodadadi, and A. A. Abdelhamid, “8. optimized lstm for accurate smart grid stability prediction using a novel optimization algorithm,” Frontiers in Energy Research, 2024

work page 2024
[7]

Machine - Learning Based Prediction of Stability of Smart Grid,

G. P. Gunjal, M. M. Teke, A. M. Deulkar and D. B. Pardeshi, "Machine - Learning Based Prediction of Stability of Smart Grid," 2024 International Conference on Expert Clouds and Applications (ICOECA), Bengaluru, India, 2024, pp. 930 -937, doi: 10.1109/ICOECA62351.2024.00164

work page arXiv 2024
[8]

Application of Deep Learning and Neural Networks for Smart Grid Stability Predictions,

E. Marevac, E. Kadusic, C. Ruland and N. Zivic, "Application of Deep Learning and Neural Networks for Smart Grid Stability Predictions," 2024 IEEE 12th International Conference on Smart Energy Grid Engineering (SEGE), Oshawa, ON, Canada, 2024, pp. 84 -89, doi: 10.1109/SEGE62220.2024.10739561

work page arXiv 2024

Show all 21 references

[9]

Proactive Semi -Supervised Machine Learning Method for Stability Estimation in Smart Grids,

B. Abegaz and E. Hernandez, "Proactive Semi -Supervised Machine Learning Method for Stability Estimation in Smart Grids," 2024 19th Annual System of Systems Engineering Conference (SoSE), T acoma, WA, USA, 2024, pp. 215 -219, doi: 10.1109/SOSE62659.2024.10620968

2024
[10]

Mewada, H., Syam Sundar, L., Patel, B., Desai, M., & Khan, A. (2024). A Lightweight Sequential Convolutional Neural Network for Smart Grid Stability Analysis. Transdisciplinary Journal of Engineering & Science, 15. https://doi.org/10.22545/2024/00259

2024 doi
[11]

Robust Defense Against Extreme Grid Events Using Dual-Policy Reinforcement Learning Agents,

B. Peter and M. Korkali, "Robust Defense Against Extreme Grid Events Using Dual-Policy Reinforcement Learning Agents," arXiv preprint,
[12]

Optimizing load scheduling in power grids using reinforcement learning and Markov decision processes,

D. Luo, “Optimizing load scheduling in power grids using reinforcement learning and Markov decision processes,” arXiv preprint arXiv:2410.17696, 2024. [Online]. Available: https://arxiv.org/abs/2410.17696

2024 arXiv
[13]

Study on a General Reinforcement Learning Simulation Platform for Regional Power Grid Control Requirements,

L. Zhou, Z. Yi, Q. Chen, W. Shen, L. Li, and T. Wang, "Study on a General Reinforcement Learning Simulation Platform for Regional Power Grid Control Requirements," in Proc. IEEE International Conference on Energy Power Grid (ICEPG), 2024, pp. 691–696

2024
[14]

GridLearn: Multiagent reinforcement learning for grid -aware building energy management,

A. Pigott, C. Crozier, K. Baker, and Z. Nagy, "GridLearn: Multiagent reinforcement learning for grid -aware building energy management," Electric Power Systems Research, " vol. 213, p. 108521, 2022, doi: 10.1016/j.epsr.2022.108521

2022
[15]

Electrical Grid Stability Simulated Data ,

V. Arzamasov, “Electrical Grid Stability Simulated Data ,” UCI Machine Learning Repository, 2018, DOI:https://doi.org/10.24432/C5PG66

2018 doi
[16]

A machine learning-based model for stability prediction of decentralized power grid linked with renewable energy resources,

M. Ibrar et al. “A machine learning-based model for stability prediction of decentralized power grid linked with renewable energy resources,” Wirel. Commun. Mob. Comput., vol. 2022, pp. 1–15, Aug. 2022

2022
[17]

Protection of a smart grid with the detection of cyber-malware attacks using efficient and novel machine learning models,

S. Aziz, M. Irshad, S. A. Haider, J. Wu, D. N. Deng, and S. Ahmad, “Protection of a smart grid with the detection of cyber-malware attacks using efficient and novel machine learning models,” Front. Energy Res., vol. 10, Aug. 2022

2022
[18]

Assessment and classification of grid stability with cost -sensitive stacked ensemble classifier,

K. Ramasamy, A. Sundaramurthy, and D. Velusamy, “Assessment and classification of grid stability with cost -sensitive stacked ensemble classifier,” Automatika, vol. 64, no. 4, pp. 783–797, 2023

2023
[19]

A novel approach to predicting the stability of the smart grid u tilizing mlp-elm technique,

A. Alsirhani, M. M. Alshahrani, A. Abukwaik, A. I. Taloba, R. M. A. El-Aziz, and M. Salem, “A novel approach to predicting the stability of the smart grid u tilizing mlp-elm technique,” Alexandria Engineering Journal, vol. 74, pp. 495–508, 2023

2023
[20]

Per- performance analysis of machine learning -based traditional and ensemble techniques for smart grid stability prediction,

M. S. Hassan, M. A. Siddique, N. Mohammad, and A. I. Ikram, “Per- performance analysis of machine learning -based traditional and ensemble techniques for smart grid stability prediction,” in Proceedings of the 6 th International Conference on Electrical Engineering and Informa...

2024
[2024]

Available: https://doi.org/10.48550/arxiv.2411.11180

[Online]. Available: https://doi.org/10.48550/arxiv.2411.11180

doi arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Smoliński, A., Smoliński, A., & Bąk, A. (2022). Clean Coal Technologies as an Effecti ve Way in Global Carbon Dioxide Mitigation. Energies, 15(16), 5868.J

work page 2022

[2] [2]

A distributed control approach for enhancing smart grid transient stability and resilience,

M. Ayar, S. Obuz, R. D. Trevizan, A. S. Bretas, and H. A. Latchman, “A distributed control approach for enhancing smart grid transient stability and resilience,” IEEE Transactions on Smart Grid, vol. 8, no. 6, pp. 3035–3044, 2017

work page 2017

[3] [3]

Improving the stability of an interconnected power system using genetic eigenvalue technique,

U. K.I, “Improving the stability of an interconnected power system using genetic eigenvalue technique,” Journal of Energy Technologies and Policy, vol. 11, no. 1, pp. 22–41, 2020

work page 2020

[4] [4]

Research on stability of the power system,

D. Xie, D. Zang, and P. Gao, “Research on stability of the power system,” in Proc. 2nd Int. Conf. on Materials Science, Energy Technology and Environmental Engineering (MSETEE), Zhuhai, China, Apr. 2017, IOP Conf. Ser.: Earth Environ. Sci., vol. 81, p. 012163. doi: 10.1088/1755-1315/81/1/012163

work page doi:10.1088/1755-1315/81/1/012163 2017

[5] [5]

Performance monitoring and redesign of power system stabilizers based on system identification techniques,

X. Bombois and L. Vanfretti, “Performance monitoring and redesign of power system stabilizers based on system identification techniques,” Sustainable Energy, Grids and Networks, 2023

work page 2023

[6] [6]

8. optimized lstm for accurate smart grid stability prediction using a novel optimization algorithm,

F. K. Karim, D. S. Khafaga, E.-S. M. El-kenawy, M. M. Eid, A. Ibrahim, L. Abualigah, N. Khodadadi, and A. A. Abdelhamid, “8. optimized lstm for accurate smart grid stability prediction using a novel optimization algorithm,” Frontiers in Energy Research, 2024

work page 2024

[7] [7]

Machine - Learning Based Prediction of Stability of Smart Grid,

G. P. Gunjal, M. M. Teke, A. M. Deulkar and D. B. Pardeshi, "Machine - Learning Based Prediction of Stability of Smart Grid," 2024 International Conference on Expert Clouds and Applications (ICOECA), Bengaluru, India, 2024, pp. 930 -937, doi: 10.1109/ICOECA62351.2024.00164

work page arXiv 2024

[8] [8]

Application of Deep Learning and Neural Networks for Smart Grid Stability Predictions,

E. Marevac, E. Kadusic, C. Ruland and N. Zivic, "Application of Deep Learning and Neural Networks for Smart Grid Stability Predictions," 2024 IEEE 12th International Conference on Smart Energy Grid Engineering (SEGE), Oshawa, ON, Canada, 2024, pp. 84 -89, doi: 10.1109/SEGE62220.2024.10739561

work page arXiv 2024