A Differentially Private Quantum Neural Network for Probabilistic Optimal Power Flow

Yan Xu; Yue Chen; Yuji Cao

REVIEW 3 major objections 6 minor 17 references

A Differentially Private Quantum Neural Network for Probabilistic Optimal Power Flow

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A quantum neural network can approximate probabilistic optimal power flow while provably protecting individual load data, using clipped and Gaussian-noised gradient updates that yield differential privacy and beat a private classical…

desk verdict The privacy proof has a factor-of-2 sensitivity error that invalidates the central guarantee, but the empirical QNN-vs-MLP noise tolerance is interesting and the application is new. read the letter →

arxiv 2411.16117 v2 pith:IALSMYLW submitted 2024-11-25 eess.SY cs.SY

Yuji Cao , Yue Chen , Yan Xu This is my paper · ORCID

classification eess.SYcs.SY

keywords probabilisticoptimalpowerflowquantumneuralnetworkvariationalcircuitdifferentialprivacyGaussiannoiseload-privacyrenewableuncertaintyIEEE33-bussystem

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

This paper tries to show that a quantum neural network (QNN) can approximate probabilistic optimal power flow (OPF) solutions while provably protecting the privacy of individual load data. The proposed method clips each sample's gradient to a fixed norm, adds Gaussian noise of calibrated scale, and updates the variational quantum circuit parameters with the noisy gradient, which is claimed to make the whole training process $(\varepsilon, \delta)$-differentially private. If true, a grid operator could release a fast OPF surrogate without revealing customer load patterns, and the aggregate statistics of the OPF solution would still be accurate. The paper further claims that this private QNN is far more noise-resilient and parameter-efficient than a classically trained private neural network, using 165 parameters versus 1,281 at comparable or better accuracy.

What carries the argument

The load-bearing mechanism is Algorithm 1: for each OPF sample, compute the VQC gradient, clip it to $\ell^2$ norm $C$, add Gaussian noise $\mathcal{N}(0, \sigma^2 C^2 I)$ to the batch sum, average, and update. The privacy analysis combines the Gaussian mechanism with subsampling amplification and advanced composition over $T$ epochs. The expressiveness side is the strongly entangled variational layer, built from angle encoding, parameterized rotations, and adjacent CNOT gates, which the paper uses to fit the nonlinear OPF map with only 165 parameters.

What would settle it

Enumerate all adjacent pairs in a small training set (or randomly sample many pairs), compute the batch-averaged clipped gradient for each, and measure the maximum $\ell^2$ difference; if that measured sensitivity exceeds $C$, then Theorem 1's noise scale $\sigma$ is not sufficient for the claimed $(\varepsilon, \delta)$-differential privacy.

Watch

Extended reading notes

Core claim

The central claim is that injecting Gaussian noise into the gradient updates of a variational quantum circuit yields an $(\varepsilon, \delta)$-differentially private learning algorithm for probabilistic OPF, with privacy amplification by random subsampling and advanced composition providing the final guarantee. The authors show numerically that a private QNN trained on the IEEE 33-bus system produces voltage statistics that stay close to the Monte Carlo benchmark while visually masking the customer's load pattern, and that the QNN maintains positive $R^2$ values at noise scales where a classical MLP collapses to large negative $R^2$. The discovery is intended to establish that quantum machine learning can carry differential privacy without sacrificing the utility needed for OPF approximation, and that the strong entanglement in the ansatz supports the required nonlinear mapping.

Load-bearing premise

The privacy guarantee rests on the claim that clipping each per-sample gradient to length $C$ limits the change in the batch-averaged gradient to $C$ when one sample changes; if the true sensitivity is larger, the claimed $(\varepsilon, \delta)$ level does not follow at the stated noise scale.

Editorial extensions

If this is right

A distribution operator can train and release a QNN-based OPF surrogate that hides individual customer load signatures while preserving mean and standard deviation of voltage within 0.002% of the Monte Carlo reference.
At noise scales where a private classical MLP becomes unusable ($R^2$ negative), the private QNN still tracks the OPF solution, suggesting DP noise does not break quantum-circuit training the same way.
The parameter reduction from 1,281 to 165 gives a concrete path to running private OPF surrogates on near-term quantum hardware with limited qubits.
Choosing a moderate noise scale ($\sigma = 1$) provides a practical privacy-utility tradeoff: strong visual masking of the load pattern with minimal statistical distortion.

Reading between the lines

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

The stated sensitivity bound in Theorem 1 may undercount the true sensitivity: clipping each per-sample gradient to norm $C$ bounds each gradient, but replacing one sample can change the batch-averaged clipped gradient by up to $2C$, so the effective $\varepsilon$ could be larger than advertised unless the analysis is tightened or the noise scale increased.
The paper defines adjacency on a single load value in a vector, while the training algorithm consumes a dataset of samples; a fully convincing DP claim would define adjacency on the dataset (add or remove one sample) and account for the sampling process consistently.
An empirical membership-inference or sensitivity-audit test on a real load dataset would settle whether the realized privacy level matches the theoretical $(\varepsilon, \delta)$, and would also quantify how much privacy is lost by the $2C$ gap.
The noise-resilience advantage over the MLP may depend on the small circuit and single-bus output; testing on larger grids and deeper circuits would show whether the pattern holds at scale.

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

The thing you should know about this paper: the privacy guarantee in Theorem 1 is wrong, and the mistake is load-bearing. The paper clips per-sample gradients to norm C and then adds Gaussian noise scaled by σC. The proof claims the sensitivity of the average batch gradient is C, but it's actually 2C: two adjacent batches differ in one sample, and the worst-case difference after clipping is 2C (one gradient C along e1, the other −C along e1). So the noise scale σ ≥ √(2 log(1.25)/δ)/ε is a factor of 2 too small. The subsampled and composed bounds in Eq. (17) inherit the error.

There's also an adjacency mismatch: Definition 1 protects load vectors D,D′ differing in one element, but Algorithm 1 trains on a dataset of N samples (x_i,y_i). The per-step analysis is about replacing one training sample, so the protected object and the analyzed adjacency don't line up. Fixing this will change the strength of the guarantee and possibly the utility of the method.

What's genuinely new here: this is the first application of a differentially private QNN to probabilistic OPF that I know of. The empirical comparison against a classical MLP is interesting—the QNN maintains positive R² at σ=5 while the MLP collapses to -22,000. That noise-resilience observation is worth taking seriously, though the paper doesn't explain why QNNs behave this way.

The rest of the paper is competent. The OPF formulation, the variational circuit (strongly entangled layers from Schuld et al.), and the DP-SGD setup are standard. The case study is limited: single node (Node 30), single run, no code or data, and the quantum time estimates are speculative. But for a first look, those are secondary.

My overall take: the empirical results are suggestive, but the central theoretical claim—the thing that makes this a "differentially private" method—doesn't hold as stated. This is fixable by doubling the noise (which weakens the privacy guarantee) and rewriting the adjacency definition to match the training data. As is, the paper should not be accepted.

Who is this for? Researchers interested in quantum ML for power systems, or in DP for non-standard models. It's worth a serious referee because the application is novel and the empirical observation about QNN robustness is nontrivial, but the referee should check the sensitivity argument carefully.

My recommendation: send it to review, but flag the privacy proof as needing a major revision.

Referee Report

3 major / 6 minor

Summary. The paper proposes a differentially private quantum neural network (QNN) for probabilistic optimal power flow (OPF). The method uses a variational quantum circuit with angle encoding and strongly entangled layers to approximate the mapping from uncertain inputs (wind, solar, load) to OPF solutions, and it modifies DP-SGD by clipping per-sample gradients and adding Gaussian noise. The central theoretical claim is Theorem 1, which states that Algorithm 1 achieves (ε′, TBδ/N + δ′)-differential privacy under a specific noise multiplier condition. Case studies on an IEEE 33-bus system compare the private QNN against a classical MLP, reporting accuracy, parameter counts, and noise resilience.

Significance. If the privacy guarantee were correct, the paper would make a useful contribution by extending differential privacy to quantum neural networks for a power-system application. The empirical part shows a potentially interesting phenomenon of noise resilience in QNNs compared to classical MLPs, and the parameter reduction is noteworthy. However, the theoretical privacy proof—the main advertised contribution—contains load-bearing errors. The sensitivity calculation is off by a factor of two, the adjacency definition does not match the training data, and the composition is undercounted. Because differential privacy is a worst-case formal guarantee, the numerical experiments cannot remedy these flaws. The contribution is therefore substantially weakened, and the central claim is not established.

major comments (3)

[Section 4.1 Definition 1 vs. Algorithm 1] The proof claims that clipping each per-sample gradient to norm C bounds the ℓ2 sensitivity of the batch-averaged gradient by C. This is incorrect. For two batches of size B that differ in exactly one sample, the clipped gradients of that sample can be C·e1 and -C·e1, so the ℓ2 norm of the difference of the sums of clipped gradients is 2C; after the 1/B averaging in Algorithm 1 line 10, the sensitivity is 2C/B. The added noise in line 10 has standard deviation σC/B. The Gaussian mechanism therefore requires σC/B ≥ (2C/B)·√(2 log(1.25/δ))/ε, i.e., σ ≥ 2√(2 log(1.25/δ))/ε, not the factor-of-one condition stated in Theorem 1. Consequently, the per-step (ε,δ) guarantee, the subsampling-amplified bound, and the final ε′ in Eq. (17) are all unsupported.
[Section 4.3, composition] Definition 1 and the adjacency relation in Eq. (15) are stated for two load vectors D,D′ ∈ R^n that differ in one element (one load value at one node). Algorithm 1, however, trains on a dataset of N samples {(x_i, y_i)}, where each x_i contains wind, PV, and load values, and the sensitivity analysis treats adjacency as replacing one sample in a batch. These are different objects: the former protects a single load value in a vector, while the latter protects a sample in a training set. The theorem therefore does not establish privacy for the data representation used in the algorithm, and the claim that the method protects individual load data is not formally supported.
[Section 4.3, composition] The proof applies advanced composition over T epochs, but Algorithm 1 has an inner loop over batches. In each epoch there are N/B gradient updates, each adding fresh Gaussian noise and each constituting a separate mechanism invocation. The composition must be over T·N/B steps, not T steps. This undercounts the privacy loss in Eq. (17) even if the per-step guarantee were correct. Additionally, the subsampling amplification step uses a fixed-size batch of size B, whereas the cited result of Balle et al. applies to Poisson subsampling (each sample included independently with probability B/N). The application of that amplification bound to fixed-size random batches requires additional justification and is not supplied.

minor comments (6)

[Abstract and Section 5.3] The abstract and Section 5.3 claim that the QNN reduces the number of parameters by 90%, but Table 3 lists 165 parameters for the QNN versus 1,281 for the MLP, which is an 87.1% reduction. Please report the exact percentage or say 'about 90%'.
[Section 5.2, Table 2] The notation ε^{V30}_{mean} in Table 2 is not defined before the table; please state that this denotes the relative error of the mean voltage magnitude at node 30 and clarify the entry '≤ 10−5' for σ = 0.
[Figure 3] The caption states that the solid line represents the mean predicted value from 10 measurements, each of 100 shots, but it is unclear whether this averaging also applies to the 'Non-private QNN' and 'Non-private OPF' curves; please clarify.
[Algorithm 1, line 10] The noise is added to the sum of clipped gradients before division by B; equivalently, Gaussian noise with scale σC/B is added to the averaged gradient. Stating this equivalence explicitly would help readers connect the algorithm to the sensitivity calculation.
[Theorem 1] The condition on σ is typeset as 'σ ≥√2 log(1.25)/δ ϵ', which appears to be a typographical error for σ ≥ √(2 log(1.25/δ))/ε. Please correct the typesetting and also clarify that this condition is the per-step Gaussian-mechanism condition.
[Section 5.1] The periodic load pattern formula p(t) in Section 5.1 is not cited; if this test pattern is taken from reference [7] (Dvorkin et al.), please add the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the privacy proof uses external differential-privacy mechanisms and the QNN results are benchmarked against Monte Carlo and an MLP, not derived from the paper's own claims.

full rationale

The derivation chain is not circular. The POPF approximation claim is checked against Monte Carlo reference statistics on the IEEE 33-bus system (Section 5.2), and the QNN-vs-MLP comparison trains both models with the same Algorithm 1 (Section 5.3), so the accuracy and noise-resilience results are benchmarked rather than built into the definitions. The privacy guarantee of Theorem 1 is composed from the standard Gaussian mechanism, subsampling amplification [14], and advanced composition [13]; these are external results and the noise scale is not fitted to the observed data. The strongly entangled ansatz is adopted from Schuld et al. [12] by citation, not presented as a derived consequence of the OPF problem. The only self-citation, [17], supplies hardware timing constants (Tp+TM=1 µs, TG=10 ns) for the secondary computation-time column in Table 3; it is not load-bearing for the central privacy or accuracy claims. The possible 2C sensitivity error in the proof of Theorem 1 and the adjacency/dataset mismatch are formal correctness concerns, not instances of a claim reducing to its inputs by construction, and therefore are not scored as circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard differential privacy theorems, the choice of clipping and noise parameters, and the domain assumptions about the OPF model and input distributions. No new physical entities are introduced. The false sensitivity assumption in the DP proof is the most important load-bearing premise.

free parameters (5)

clipping norm C = 1
Chosen by hand in Table 1; controls the sensitivity bound in the DP proof and the noise scale.
noise multiplier sigma = varied 0,1,2,5,10
Chosen by hand to trade privacy and utility; the DP theorem gives a condition on sigma.
batch size B = 32
Appears in the subsampling amplification ratio q=B/N in Theorem 1.
number of variational layers = 10
Architecture choice setting the number of trainable parameters and expressiveness.
learning rate eta = 0.05
Optimizer step size; affects convergence and the reported loss curves.

assumptions (6)

standard math Gaussian mechanism provides (epsilon, delta)-DP when noise scale is at least Delta * sqrt(2 log(1.25)/delta) / epsilon
Invoked in Section 4.3 to relate clipping sensitivity, noise, and privacy parameters.
standard math Advanced composition theorem for differential privacy
Used in Theorem 1 to accumulate privacy loss over T epochs.
standard math Privacy amplification by uniform subsampling
Used to scale per-step privacy by q=B/N in Theorem 1.
domain assumption Clipping per-sample gradients to norm C bounds the sensitivity of the batch gradient by C
The proof assumes this, but the correct sensitivity is 2C, making the assumption false.
domain assumption The branch flow model and its convex relaxation accurately represent the distribution grid
OPF labels are generated from this model; if the relaxation is inexact, the training data are not true OPF solutions.
domain assumption The Weibull, Beta, and Normal input distributions represent real wind, solar, and load
Used to generate the 1,000 samples; unrepresentative distributions would limit the reported accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Differentially Private Quantum Neural Network for Probabilistic Optimal Power Flow." pith.science (2026). https://pith.science/paper/IALSMYLW

@misc{pith2026241116117,
  author       = {Pith},
  title        = {Pith review of: A Differentially Private Quantum Neural Network for Probabilistic Optimal Power Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IALSMYLW}},
  note         = {Machine review of arXiv:2411.16117}
}

abstract

The stochastic nature of renewable energy and load demand requires efficient and accurate solutions for probabilistic optimal power flow (OPF). Quantum neural networks (QNNs), which combine quantum computing and machine learning, offer computational advantages in approximating OPF by effectively handling high-dimensional data. However, adversaries with access to non-private OPF solutions can potentially infer sensitive load demand patterns, raising significant privacy concerns. To address this issue, we propose a privacy-preserving QNN model for probabilistic OPF approximation. By incorporating Gaussian noise into the training process, the learning algorithm achieves ($\varepsilon, \delta$)-differential privacy with theoretical guarantees. Moreover, we develop a strongly entangled quantum state to enhance the nonlinearity expressiveness of the QNN. Experimental results demonstrate that the proposed method successfully prevents privacy leakage without compromising the statistical properties of probabilistic OPF. Moreover, compared to classical private neural networks, the QNN reduces the number of parameters by 90% while achieving significantly higher accuracy and greater stability.

Figures

Figures reproduced from arXiv: 2411.16117 by the authors.

**Figure 3.** The left figure presents the load pattern of the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 2.** Topology of modified IEEE 33-bus power distri [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Voltage magnitude comparisons between differentially private QNN and non-private OPF under load changes. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]

**Figure 4.** Figure 4: Training loss of QNN and MLP when σ = 1. The classical MLP suffers from training stability while QNN is DP noise-resilient. To investigate the underlying reasons for the distinct performance between the QNN and MLP, we analyze the loss values during the training proce…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

[1]

Analysis of probabilis- tic optimal power flow taking account of the variation of load power,

X. Li, Y. Li, and S. Zhang, “Analysis of probabilis- tic optimal power flow taking account of the variation of load power,” IEEE Transactions on Power Systems , vol. 23, no. 3, pp. 992–999, 2008

work page 2008
[2]

Deepopf- v: Solving ac-opf problems efficiently,

W. Huang, X. Pan, M. Chen, and S. H. Low, “Deepopf- v: Solving ac-opf problems efficiently,” IEEE Transac- tions on Power Systems , vol. 37, no. 1, pp. 800–803, 2021

work page 2021
[3]

Quantum com- puting based hybrid solution strategies for large-scale discrete-continuous optimization problems,

A. Ajagekar, T. Humble, and F. You, “Quantum com- puting based hybrid solution strategies for large-scale discrete-continuous optimization problems,” Comput- ers & Chemical Engineering , vol. 132, p. 106630, 2020

work page 2020
[4]

Towards provably efficient quantum algorithms for large-scale machine-learning models,

J. Liu, M. Liu, J.-P. Liu, Z. Ye, Y. Wang, Y. Alex- eev, J. Eisert, and L. Jiang, “Towards provably efficient quantum algorithms for large-scale machine-learning models,” Nature Communications, vol. 15, no. 1, p. 434, 2024

work page 2024
[5]

Privacy pre- serving in non-intrusive load monitoring: A differen- tial privacy perspective,

H. Wang, J. Zhang, C. Lu, and C. Wu, “Privacy pre- serving in non-intrusive load monitoring: A differen- tial privacy perspective,” IEEE Transactions on Smart Grid, vol. 12, no. 3, pp. 2529–2543, 2020

work page 2020
[6]

Differential privacy,

C. Dwork, “Differential privacy,” in International col- loquium on automata, languages, and programming . Springer, 2006, pp. 1–12

2006
[7]

Differentially private optimal power flow for distribution grids,

V. Dvorkin, F. Fioretto, P. Van Hentenryck, P. Pin- son, and J. Kazempour, “Differentially private optimal power flow for distribution grids,” IEEE Transactions on Power Systems , vol. 36, no. 3, pp. 2186–2196, 2020

work page 2020
[8]

Decentralized optimal power flow for multi-agent active distribution networks: A differen- tially private consensus admm algorithm,

C. Lei, S. Bu, Q. Chen, Q. Wang, Q. Wang, and D. Srinivasan, “Decentralized optimal power flow for multi-agent active distribution networks: A differen- tially private consensus admm algorithm,”IEEE Trans- actions on Smart Grid , 2024

work page 2024

Show all 17 references

[9]

Quantum machine learning with differential privacy,

W. M. Watkins, S. Y.-C. Chen, and S. Yoo, “Quantum machine learning with differential privacy,” Scientific Reports, vol. 13, no. 1, p. 2453, 2023

2023
[10]

Branch flow model: Relax- ations and convexification—part i,

M. Farivar and S. H. Low, “Branch flow model: Relax- ations and convexification—part i,” IEEE Transactions on Power Systems , vol. 28, no. 3, pp. 2554–2564, 2013

2013
[11]

C. Z. Mooney, Monte carlo simulation. Sage, 1997, no. 116

1997
[12]

Circuit-centric quantum classifiers,

M. Schuld, A. Bocharov, K. M. Svore, and N. Wiebe, “Circuit-centric quantum classifiers,” Physical Review A, vol. 101, no. 3, p. 032308, 2020

2020
[13]

The algorithmic foundations of differential privacy,

C. Dwork, A. Roth et al., “The algorithmic foundations of differential privacy,” Foundations and Trends ® in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211– 407, 2014

2014
[14]

Privacy ampli- fication by subsampling: Tight analyses via couplings and divergences,

B. Balle, G. Barthe, and M. Gaboardi, “Privacy ampli- fication by subsampling: Tight analyses via couplings and divergences,” Advances in neural information pro- cessing systems, vol. 31, 2018

2018
[15]

Quantum computing in the nisq era and beyond,

J. Preskill, “Quantum computing in the nisq era and beyond,” Quantum, vol. 2, p. 79, 2018

2018
[16]

Qaoa for max- cut requires hundreds of qubits for quantum speed-up,

G. G. Guerreschi and A. Y. Matsuura, “Qaoa for max- cut requires hundreds of qubits for quantum speed-up,” Scientific reports, vol. 9, no. 1, p. 6903, 2019

2019
[17]

Carbon market risk estimation using quan- tum conditional generative adversarial network and amplitude estimation,

X. Zhou, H. Zhao, Y. Cao, X. Fei, G. Liang, and J. Zhao, “Carbon market risk estimation using quan- tum conditional generative adversarial network and amplitude estimation,” Energy Conversion and Eco- nomics, 2024. 7

2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Analysis of probabilis- tic optimal power flow taking account of the variation of load power,

X. Li, Y. Li, and S. Zhang, “Analysis of probabilis- tic optimal power flow taking account of the variation of load power,” IEEE Transactions on Power Systems , vol. 23, no. 3, pp. 992–999, 2008

work page 2008

[2] [2]

Deepopf- v: Solving ac-opf problems efficiently,

W. Huang, X. Pan, M. Chen, and S. H. Low, “Deepopf- v: Solving ac-opf problems efficiently,” IEEE Transac- tions on Power Systems , vol. 37, no. 1, pp. 800–803, 2021

work page 2021

[3] [3]

Quantum com- puting based hybrid solution strategies for large-scale discrete-continuous optimization problems,

A. Ajagekar, T. Humble, and F. You, “Quantum com- puting based hybrid solution strategies for large-scale discrete-continuous optimization problems,” Comput- ers & Chemical Engineering , vol. 132, p. 106630, 2020

work page 2020

[4] [4]

Towards provably efficient quantum algorithms for large-scale machine-learning models,

J. Liu, M. Liu, J.-P. Liu, Z. Ye, Y. Wang, Y. Alex- eev, J. Eisert, and L. Jiang, “Towards provably efficient quantum algorithms for large-scale machine-learning models,” Nature Communications, vol. 15, no. 1, p. 434, 2024

work page 2024

[5] [5]

Privacy pre- serving in non-intrusive load monitoring: A differen- tial privacy perspective,

H. Wang, J. Zhang, C. Lu, and C. Wu, “Privacy pre- serving in non-intrusive load monitoring: A differen- tial privacy perspective,” IEEE Transactions on Smart Grid, vol. 12, no. 3, pp. 2529–2543, 2020

work page 2020

[6] [6]

Differential privacy,

C. Dwork, “Differential privacy,” in International col- loquium on automata, languages, and programming . Springer, 2006, pp. 1–12

2006

[7] [7]

Differentially private optimal power flow for distribution grids,

V. Dvorkin, F. Fioretto, P. Van Hentenryck, P. Pin- son, and J. Kazempour, “Differentially private optimal power flow for distribution grids,” IEEE Transactions on Power Systems , vol. 36, no. 3, pp. 2186–2196, 2020

work page 2020

[8] [8]

Decentralized optimal power flow for multi-agent active distribution networks: A differen- tially private consensus admm algorithm,

C. Lei, S. Bu, Q. Chen, Q. Wang, Q. Wang, and D. Srinivasan, “Decentralized optimal power flow for multi-agent active distribution networks: A differen- tially private consensus admm algorithm,”IEEE Trans- actions on Smart Grid , 2024

work page 2024