{"id":"2a03af52-2292-4359-9a31-132312dbed6d","arxiv_id":"2412.08370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A noise-aware Bayesian optimization method, with the simulation error modeled as Gaussian noise, is applied to distributed energy resource capacity planning and outperforms BO and PSO in test networks.","lead":"This paper treats the difference between simulated and real-world performance of renewable-rich distribution grids as random noise, and uses a noise-aware computer search method to size wind, solar, and battery installations. Test results on two benchmark grids suggest the method finds lower annual cost and converges faster than two common alternatives.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's adaptive noise update uses the sample standard deviation of all noisy objective values as σn, conflating objective variation with observation noise and miscalibrating the GP posterior that NEI and final selection rely on.","rationale":"The paper's noiseless validation is a genuine strength: NBO closely matches the QP benchmark and converges faster than BO and PSO, so the optimization machinery is plausible. The risk is concentrated in the noisy case, which is the paper's advertised contribution. The reader's weakest assumption targets the zero-mean, homoscedastic Gaussian noise model in Eq. (16). My concern is adjacent but distinct: even if the noise were exactly Gaussian with fixed variance, Algorithm 1 does not estimate that variance correctly. The update in lines 7 and 13 uses the sample standard deviation of all noisy objective values, which conflates the underlying function's variation with observation noise. This is an internal inconsistency rather than a disagreement with external consensus. The subsequent test-benchmark tuning of the initial σn means the reported noisy-case accuracy is not independent evidence that the posterior is well calibrated. These issues are addressable: report ζ and the initialization size, replace or justify the σn update with a proper estimator such as maximum likelihood, and validate hyperparameters on a held-out set. Because the central idea remains plausible and the noiseless results are credible, I do not move the reader's CONDITIONAL verdict; I would keep it unchanged but require these revisions.","tokens_in":14640,"tokens_out":7447,"duration_ms":82612,"concrete_test":"Re-run Algorithm 1 on a synthetic test problem with a known nonconstant objective and known additive noise, e.g., a quadratic with the same cost scale and true noise standard deviation σ_true = 4.5e5. Record the sequence σn produced by lines 7 and 13 and compare it to σ_true and to the marginal standard deviation of the observed f_i values. If σn tracks the marginal standard deviation rather than σ_true, the GP posterior is miscalibrated, and the noisy-case claims require re-evaluation. Separately, select σn and ζ on a validation set instead of the Section 5.3 test benchmark and report the held-out error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central noisy-case claim to hold, the GP posterior used in the NEI acquisition and in the final 'simple reward' selection must be a valid posterior under the stated model Eq. (16). Algorithm 1 (lines 7 and 13) sets σn to a weighted average of the previous value and sqrt(Σ(fi − fbar)^2/(N+j)), i.e., the sample standard deviation of all observed objective values. That statistic estimates the marginal spread of the noisy annual costs across the sampled capacity plans, not the variance of the noise term ε in Eq. (16); it includes the variation of the underlying cost function over the input domain. Unless CT(x) is constant over the sampled plans, this systematically overestimates σn. An overestimated noise level shrinks the GP posterior mean toward the constant prior, attenuating the differences between candidate plans that the simple-reward selection (Algorithm 1, lines 17–18) uses to choose x*. The paper provides no derivation for this update, no value for the adaptive rate ζ, and no sensitivity analysis. Compounding this, the initial σn is tuned against the Section 5.3 test benchmark, so the reported 0.2044% error does not independently validate posterior calibration. The issue is not that the heuristic cannot work empirically; it is that, as written, the algorithm is not the noise-aware Bayesian procedure described by Eqs. (16)–(21), undermining the claimed connection between the probabilistic surrogate model and the noisy optimization results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a noise-aware Bayesian optimization (NBO) algorithm for the collaborative capacity planning of distributed energy resources (DERs) in an active distribution network (ADN). The planning model is formulated as an annual-cost minimization over wind, PV, and storage capacities, with a lower-level quadratic programming dispatch simulation subject to security constraints. The deviation between ADN simulation and real operation is modeled as zero-mean Gaussian noise added to the objective (Eqs. 15-16). The NBO algorithm uses a Gaussian-process surrogate, a noisy expected improvement acquisition function, and a 'simple reward' selection based on the GP posterior mean. Validation is performed against BO, PSO, and a QP benchmark in noiseless cases (Table 2), against a 200-scenario test value in the noisy case (Section 5.3), and on a 118-bus network for scalability (Section 5.4). The paper claims that NBO overcomes environmental noise, achieves lower annual cost, and is more sample-efficient than the baselines.","tokens_in":14940,"tokens_out":5456,"duration_ms":54271,"significance":"If the claims hold, the work would offer a practical way to account for simulation-to-reality deviation in distribution network planning without massive scenario enumeration. The noiseless validation (Table 2) is a genuine strength: NBO closely matches or improves upon the QP benchmark across all five models, and the convergence curves in Fig. 5 indicate faster convergence than PSO. The scalability experiment on the 118-bus network, comparing centralized and distributed configurations, is also valuable. However, the central noise-aware claim rests on two points that are not yet fully supported: the adaptive noise-variance update in Algorithm 1 and the validation of the noise level in Section 5.3. The paper provides reproducible equations and a clear QP benchmark, but it does not ship code or machine-checked proofs. The reported 0.2044% error in the noisy case is a fitted quantity because σn is tuned on the test benchmark, so the quantitative noise-handling result is not yet independent. With revision addressing these issues, the paper would be a useful contribution.","major_comments":[{"comment":"The adaptive noise variance update sets σn to a weighted average of its previous value and the sample standard deviation of all observed objective values. That statistic estimates the marginal spread of the noisy annual costs across the sampled capacity plans, not the variance of the noise term ε in Eq. (16), because it conflates the input-dependent variation of CT(x) with the observation noise. An overestimated σn shrinks the GP posterior mean toward the prior constant, attenuating the differences between candidate plans on which the final 'simple reward' selection (lines 17–18) relies. The paper provides no derivation for this update, no value for the adaptive rate ζ, and no sensitivity analysis. As written, the GP posterior used for NEI and final selection is not a valid posterior under the stated model, which undermines the claimed connection between the probabilistic surrogate and the noisy optimization results.","section":"Algorithm 1, lines 7 and 13; §3.3 Eq. (16)"},{"comment":"The error e used to assess the GP inference is computed against the test value C_test, and the optimal initial noise standard deviation σn = 4.5e5 is selected by minimizing this same error on the test benchmark. Consequently, the reported 0.2044% error for NBO is a fitted result, not an independent out-of-sample validation of the noise-handling claim. This circularity also affects the paper's assertion in §3.3 that the Gaussian approximation 'has enough accuracy', because that assertion is justified only by the same Section 5.3 experiment. To support the central claim, the authors should estimate σn from data independent of the test benchmark (e.g., from historical simulation-vs-observation mismatch pairs) or report performance over a range of σn without selecting the best on the test set.","section":"§5.3, Eq. (30) and Fig. 7a"},{"comment":"The comparison against 'basic BO' is not fully specified. The paper does not state what acquisition function the baseline uses, whether it is a noiseless or noisy EI variant, how the GP hyperparameters are initialized or re-estimated, or whether multiple restarts are performed. Since the central claim is that NBO outperforms BO, the baseline configuration must be described in enough detail to be reproducible and to ensure a fair comparison. Please provide the exact BO setup and cite a specific standard implementation, or otherwise document the configuration used.","section":"§5.2, Table 2 and Fig. 5"}],"minor_comments":[{"comment":"There is a typo: 'Inter CORE' should be 'Intel Core'.","section":"§5.1"},{"comment":"The phrase 'quasi-Monte Carol' should be 'quasi-Monte Carlo'.","section":"§4, after Eq. (24)"},{"comment":"The notation for the noisy objective is inconsistent: Eq. (15) uses CobT(x), Algorithm 1 uses y and f(x), and Section 5.3 uses µD and e. Please unify the symbols for the observed noisy cost and the GP mean.","section":"§3.3 and Algorithm 1"},{"comment":"The caption describes a 'stacked bar chart', but the plotted curves for BO and NBO appear to be lines or single bars; please clarify the plot type and whether the error bars represent standard deviations across repeated optimization runs.","section":"Fig. 7a"},{"comment":"The legend labels 'BO7/NBO7' and 'BO3/NBO3' are not defined in the caption; add a note that 7 and 3 refer to the number of installable nodes in the distributed and centralized models, respectively.","section":"Fig. 9"},{"comment":"The selection of Nres = 200 is stated to be 'by experiment', but the sensitivity of the test value to Nres is not shown; please report the experiment or provide a brief discussion.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of Sustainable Energy Grids & Networks, and the noiseless validation is a solid piece of evidence. However, the noisy-case evaluation is partially circular because σn is tuned on the test benchmark, and the adaptive noise update in Algorithm 1 appears statistically unjustified. I recommend major revision. The authors should also consider benchmarking against a standard BO implementation with explicit noise handling (e.g., GPyOpt or scikit-optimize) and report the baseline hyperparameters. The paper's novelty relative to standard noisy BO is modest, so the revision should clearly articulate what is new beyond the Letham et al. framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is the packaging, not the optimizer: the authors take Letham et al.'s noisy expected improvement, wrap it in a Gaussian-noise model for simulation-to-reality deviation, and apply it to DER capacity planning in an active distribution network. That application is real and is reasonably well executed on the noiseless side: the QP benchmark comparisons in Table 2 show NBO matching Gurobi closely, and the convergence plots suggest it gets there faster than PSO and basic BO. The 118-bus scalability study is a nice addition.\n\nThe soft spots are where the noisy claims live. First, the initial noise standard deviation sigma_n is selected by minimizing the Section 5.3 error on the test benchmark itself (Fig. 7a). So the headline 0.2044% error is a fitted quantity, not an independent prediction. Second, the stress-test note is right: Algorithm 1 updates sigma_n with the sample standard deviation of all observed objective values, which mixes variation of the underlying cost function across the input space into what is supposed to be observation noise. That systematically overestimates sigma_n, shrinks the GP posterior toward the constant prior, and undermines the posterior used by NEI and the simple-reward final selection. This is not necessarily fatal to the empirical approach, but as written the algorithm is not the noise-aware Bayesian procedure described by Eqs. (16)-(21), and there is no derivation or sensitivity analysis for the update. Third, the BO baseline is underspecified (kernel, hyperparameters, acquisition details), and no code or data are provided, which makes the comparisons hard to audit.\n\nOn the math: the Gaussian deviation model in Eq. (16) is acknowledged as an approximation, and that would be acceptable if the validation were independent. The noise term is also treated as independent of the plan x, which is a strong assumption for simulation-to-reality gaps. The QP linearization and constraints look standard.\n\nVerdict: this paper deserves a serious referee, but only conditional acceptance. The application of NEI to ADN capacity planning is worth publishing if the evaluation is cleaned up. For a revision, I would want validation-set-based tuning of sigma_n, a corrected or justified noise update, full baseline and hyperparameter details, and code/data release. I would send it to peer review.","headline":"A credible noiseless application of noisy expected improvement to DER capacity planning, but the noisy-case validation is partially circular and the adaptive noise update is miscalibrated as written.","tokens_in":15464,"tokens_out":2010,"would_cite":false,"duration_ms":21700,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Noise-aware Bayesian optimization is proposed for capacity planning of distributed energy resources in active distribution networks, treating simulation deviation as Gaussian noise and reaching near-benchmark annual costs.","keywords":["active distribution network","distributed energy resources","capacity planning","Bayesian optimization","noisy expected improvement","Gaussian process","renewable energy uncertainty","demand response"],"falsifier":"Feed the optimizer a synthetic problem whose true cost is known and whose simulator noise is deliberately non-Gaussian (for example, skewed or with variance that grows with the plan's cost). If the posterior-mean selection rule then shows a systematic bias away from zero when compared with a large test set, the zero-mean fixed-variance Gaussian assumption that carries the method is falsified.","tokens_in":14437,"feed_emoji":"⚡","tokens_out":4309,"duration_ms":43270,"temperature":0.7,"pith_summary":"Capacity planning for active distribution networks must fix the sizes of wind, solar, and battery assets, but the yearly operating cost of any plan can only be estimated through simulations that never fully capture future renewable variability. This paper claims that the gap between simulation and reality can be treated as zero-mean Gaussian noise added to the annual cost, and that a noise-aware Bayesian optimization algorithm can then find near-optimal plans with far fewer simulation calls than particle swarm optimization or standard Bayesian optimization. On a 33-bus test network with no noise, the proposed optimizer reaches annual costs within roughly 0.05% of the quadratic-programming benchmark. When random renewable scenarios inject noise, the selected plan's cost error against a 200-scenario test value is about 0.2%, and the optimizer stays stable even when the assumed noise level is misspecified. The payoff is a planning method that does not need thousands of scenarios to be reliable, and that can be scaled to larger networks with distributed rather than centralized resource siting.","feed_headline":"Noise-aware Bayesian optimizer beats PSO on grid capacity planning","feed_subtitle":"Treating simulation error as Gaussian noise finds cheaper wind-solar-storage plans without rerunning 200 scenarios.","key_machinery":"The load-bearing machinery is a Gaussian-process surrogate model with a Matérn kernel, updated by a noisy expected improvement acquisition function, together with an adaptive estimate of the noise standard deviation. The Gaussian process receives noisy annual-cost outcomes from the simulation, treats the noise variance as a learned parameter, and produces a posterior mean and covariance for the true cost across the space of capacity plans. The noisy expected improvement function, approximated by quasi-Monte Carlo sampling from the posterior, chooses the next capacity plan to simulate by balancing predicted improvement against uncertainty. At the end, a simple reward strategy selects the plan with the lowest posterior mean, which is what keeps the final answer robust to the noise that would mislead a heuristic search that trusts each raw simulation.","core_discovery":"The paper's central discovery is that the unavoidable deviation between an active distribution network's operation simulation and its real behavior can be converted from a nuisance into a handled quantity: instead of trying to eliminate the deviation with typical scenarios or very long dispatch cycles, the authors model it as a noise term appended to the annual cost objective. On top of this noisy objective they build a Bayesian optimizer whose probabilistic surrogate distinguishes signal from noise, and they show that the resulting plans are both economically better and computationally cheaper than those found by particle swarm optimization or standard Bayesian optimization. In the noiseless setting the optimizer essentially matches the quadratic-programming benchmark, and under noisy evaluation it attains roughly 0.2% error relative to a costly 200-scenario test value while heuristic search degrades sharply. The same machinery solves a larger 118-bus case, where distributing resources across more nodes lowers annual cost and increases renewable accommodation.","pith_inferences":["If the zero-mean, fixed-variance Gaussian noise assumption is replaced by a heteroscedastic model (noise growing with plan cost or with renewable penetration), the same framework should be extendable by letting the GP noise parameter vary with the input; the paper does not test this.","The noise-aware treatment of simulation error is not specific to distribution networks: any expensive black-box simulator with stochastic inputs could adopt the same Gaussian-noise-plus-NEI recipe, so the method may transfer to other infrastructure planning problems.","A testable extension would be to run NBO on a synthetic benchmark with known true objective and skewed or input-dependent noise, checking whether the 0.2% error level is preserved or whether the simple reward selection becomes biased.","The paper tunes the initial noise standard deviation against the test benchmark; an automatic way to infer it from the data (for instance by marginal likelihood) would make the method fully parameter-free and is an implied next step."],"forward_implications":["Capacity planners can use the noise-aware optimizer in place of PSO when simulation outcomes are contaminated by renewable-scenario variability, since the surrogate separates signal from noise.","The method removes the need for exhaustive scenario enumeration: a handful of random renewable scenarios per iteration suffices, with accuracy checked by a 200-scenario test value.","Collaborative sizing of wind, solar, and storage beats single-resource planning: on the 33-bus case, combining all three gives the lowest annual cost, and adding storage raises renewable accommodation.","Distributed placement of resources across more nodes in the 118-bus network yields lower annual cost and more renewable absorption than centralized placement, at the price of a more complex optimization.","The reported convergence curves indicate that, even in the noiseless setting, the noise-aware optimizer reaches its near-benchmark solution in less computation time than standard BO and PSO."],"supporting_citations":[{"why":"Supplies the constrained Bayesian optimization with noisy experiments framework, including the noisy expected improvement approximation and the methodology the NBO algorithm builds on.","marker":"[28]"},{"why":"Provides the linearized power flow model used to formulate the active distribution network operation constraints in the simulation.","marker":"[27]"},{"why":"Supplies the 118-node distribution network model used for the scalability experiments with distributed and centralized configurations.","marker":"[29]"},{"why":"Provides an optimal siting and sizing procedure for energy storage systems that the paper contrasts with its scenario-free planning approach.","marker":"[16]"},{"why":"Supplies a heuristic renewable-energy system optimization baseline that the paper compares against NBO and PSO.","marker":"[22]"},{"why":"Provides a reconfiguration model for active distribution network restoration that represents a scenario-based planning comparison.","marker":"[15]"},{"why":"Supplies another heuristic distribution network planning method used as a baseline for the inefficiency of heuristic search under noise.","marker":"[26]"}],"fun_headline_variants":["Bayesian optimizer turns simulation noise into cheaper grid plans","Noise-aware BO finds cost-effective DER capacity for active grids","Modeling simulation error as noise boosts grid planning efficiency","Bayesian approach beats PSO on noisy distribution network planning","Smart capacity planning: Bayesian optimization handles real-world noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the premise that the error between the simulation and the real network is zero-mean Gaussian noise with a fixed variance that does not depend on the capacity plan being tested.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian optimizer turns simulation noise into cheaper grid plans","Noise-aware BO finds cost-effective DER capacity for active grids","Modeling simulation error as noise boosts grid planning efficiency","Bayesian approach beats PSO on noisy distribution network planning","Smart capacity planning: Bayesian optimization handles real-world noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1411,"prompt_tokens":876,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":492,"tokens_out":535,"duration_ms":6669,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:52:20.202011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the optimizer a synthetic problem whose true cost is known and whose simulator noise is deliberately non-Gaussian (for example, skewed or with variance that grows with the plan's cost). If the posterior-mean selection rule then shows a systematic bias away from zero when compared with a large test set, the zero-mean fixed-variance Gaussian assumption that carries the method is falsified.","supporting_citations":[{"cited_title":"Constrained bayesian optimization with noisy experiments,","cited_arxiv_id":null,"evidence_quote":"Supplies the constrained Bayesian optimization with noisy experiments framework, including the noisy expected improvement approximation and the methodology the NBO algorithm builds on."},{"cited_title":"Novel linearized power flow and linearized opf models for active distribution networks with application in distribution lmp,","cited_arxiv_id":null,"evidence_quote":"Provides the linearized power flow model used to formulate the active distribution network operation constraints in the simulation."},{"cited_title":"An improved ts algorithm for loss- minimum reconfiguration in large-scale distribution systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the 118-node distribution network model used for the scalability experiments with distributed and centralized configurations."},{"cited_title":"Optimal planning of dis- tributed energy storage systems in active distribution networks embed- ding grid reconfiguration,","cited_arxiv_id":null,"evidence_quote":"Provides an optimal siting and sizing procedure for energy storage systems that the paper contrasts with its scenario-free planning approach."},{"cited_title":"Economic analysis and optimization of a renewable energy based power supply system with different energy storages for a remote island,","cited_arxiv_id":null,"evidence_quote":"Supplies a heuristic renewable-energy system optimization baseline that the paper compares against NBO and PSO."},{"cited_title":"A multi-step reconfiguration model for active distribution network restoration integrating dg start-up sequences,","cited_arxiv_id":null,"evidence_quote":"Provides a reconfiguration model for active distribution network restoration that represents a scenario-based planning comparison."},{"cited_title":"Distribution network planning enhancement via network reconfiguration and dg integration using dataset approach and water cycle algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies another heuristic distribution network planning method used as a baseline for the inefficiency of heuristic search under noise."}],"review_version":1}