{"id":"9511527c-30f2-41d5-916e-7884d5e37538","arxiv_id":"2502.09280","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An adaptive Bayesian multi-objective optimizer with automatic reference point and noise settings is applied to joint capacity sizing of four heat-source technologies in cold-region CHP systems.","lead":"This paper proposes a planning model that sizes heat pumps, thermal storage, electric boilers, and storage heaters together to reduce renewable curtailment in cold-region power systems. It also introduces an adaptive version of a known multi-objective Bayesian optimization algorithm and claims it produces better trade-off solutions with fewer simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation is in-sample: typical scenarios and the SAA 'true' benchmark in Eqs. (3) and (41) come from the same historical season, so the claimed noise-modeling benefit may not transfer to unseen years.","rationale":"The reader's weakest assumption identified the same issue, and I agree it is the load-bearing one. The paper's three contributions are: (i) AMBO algorithm, (ii) noise-model based handling of simulation deviation, (iii) coupled time-series scenario generation. Contributions (ii) and (iii) are validated exclusively through Fig. 5 and Fig. 9, both of which use the full-season SAA of Eq. (41) as 'actual' performance. Since the typical days are selected from and adjusted to the same historical season used for the SAA, the experiments demonstrate only that the typical-day compression plus GP filtering reconstructs that season's average behavior. They do not test whether either the scenario set or the learned noise level sigma_n remains valid for a future, statistically different heating season. This matters because the motivating problem is precisely the unpredictability of future RES output (Section II-B). A validation split by year would settle the question: if the posterior-mean advantage persists on held-out years, the claim is robust; if not, the conclusion should be weakened to 'the method reduces in-sample simulation compression error'. The AMBO algorithm's algorithmic novelty over qNEHVI (adaptive reference point and marginal-likelihood noise estimation) is plausible, and the large-scale case in Section V-D is a useful stress test, but it is also in-sample and single-run. There is a secondary notational concern in Eq. (38): for a minimization problem the reference point should be worse than the observed worst case, while the formula as written moves it in the opposite direction; this may be a typographical error, but it should be corrected or clarified. Overall, the provided evidence supports a conditional acceptance pending out-of-sample validation and multi-seed comparisons, which matches the reader's conditional verdict.","tokens_in":16756,"tokens_out":7952,"duration_ms":82059,"concrete_test":"Hold out one full heating season (e.g., the final year) from the province dataset. Generate typical scenarios using K-medoids clustering and Eq. (3) on the remaining training years only. Re-run AMBO, NMBO, PBO, and NSGA-II using these training-derived typical scenarios, with at least 10 random seeds. Then compute the SAA benchmark of Eq. (41) on the held-out year's daily scenarios and compare e_B (GP posterior-mean error) vs e_typ (raw typical-scenario error) on the held-out year, and compare the hypervolume of each algorithm's returned Pareto front evaluated on the held-out year. If e_B no longer substantially beats e_typ, or AMBO's hypervolume advantage over the baselines disappears, the in-sample validation is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AMBO's posterior-mean filtering and adaptive noise modeling reduce evaluation error is supported only by Fig. 9, which compares against the SAA benchmark in Eq. (41) computed by averaging the full historical heating season. But the typical scenarios Styp are generated by K-medoids clustering from that same heating-season data (Section V-A2) and then adjusted in Eq. (3) to match its monthly means and variances. The errors e_B and e_typ in Eqs. (39)-(40) therefore measure in-sample reconstruction of the very year from which the typical days were selected, not predictive accuracy on unseen years. The noise term epsilon(x|Styp) in Eq. (1) is motivated by the unpredictability of future RES output, yet no experiment exposes the method to a different year: the same data both builds the typical-day compressor and defines 'ground truth'. If a validation year has different heat-load or RES statistics, the K-medoids prototypes and the GP noise variance sigma_n would have to be re-estimated, and the advantage in Fig. 9 could shrink or disappear. Thus the paper's strongest practical claim—that the proposed planning results perform well under real operational conditions in cold regions—is not established by the provided evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a capacity-planning model for hybrid heat sources (heat pump, thermal energy storage, electric boiler, and combined storage heater) in CHP-dominated cold-region power systems. The model has two objectives, annual cost and consumed renewable energy, and is evaluated through a time-series operation simulation. The authors introduce AMBO, an adaptive multi-objective Bayesian optimization algorithm that extends qNEHVI by updating the hypervolume reference point and the Gaussian-process noise standard deviation during optimization, and they combine it with a K-medoids-based typical-day scenario generation method with monthly mean/variance adjustment. Case studies on one heating season in northeastern China and on a larger provincial system compare AMBO against the original qNEHVI-style method (NMBO), PBO, and NSGA-II, reporting a more diverse and evenly distributed Pareto front, higher sample efficiency, and lower evaluation error relative to a full-season sample-average approximation benchmark.","tokens_in":16938,"tokens_out":8155,"duration_ms":86909,"significance":"If the reported performance survives the correctness and validation issues noted below, the paper would make a useful contribution. It transfers modern multi-objective Bayesian optimization to a relevant power-system planning problem, proposes a concrete mechanism for adapting the reference point and noise level, and adapts scenario generation to the electric-heat coupling that is central in cold regions. The formulation is documented in enough detail to be reproduced from the equations and tables, and the scalability study is a valuable addition. However, the central empirical claims currently rest on in-sample comparisons, single runs, and a weak NSGA-II baseline, so the strengths are mostly in the problem framing and algorithmic extension rather than in the demonstrated out-of-sample benefit.","major_comments":[{"comment":"The error-reduction claim is in-sample. The sample-average approximation benchmark in Eq. (41) is computed by averaging over the same full heating-season data from which the typical scenarios S_typ are selected by K-medoids (Section V-A2) and then adjusted in Eq. (3). Consequently, e_typ and e_B in Eqs. (39)-(40) measure how well typical-day evaluation reconstructs the same season from which the typical days were drawn, not predictive accuracy on an unseen year. Since the noise term in Eq. (1) is motivated by future RES variability, the paper needs a temporal out-of-sample test, for example by generating typical days from one heating season and evaluating the planning schemes on a different historical year. Without such a test, the conclusion that the proposed planning results 'perform well under real operational conditions in cold regions' is not established by the evidence presented.","section":"Section V-B, Eqs. (39)-(41)"},{"comment":"The upper bound in the CHP feasible region is self-referential: the inequality reads P^CHP_{i,t} ≤ P^CHP_{i,t} - c_cAB_i H^CHP_{i,t}, which contains the variable to be bounded on both sides and is therefore ill-posed. The intended bound is presumably P^CHP_{i,t} ≤ P^CHP_{max,i} - c_cAB_i H^CHP_{i,t} or the corresponding upper envelope. Because this constraint defines the operation simulation model used to generate every Pareto front in the paper, the equation must be corrected and the simulations confirmed to have used the intended feasible region.","section":"Eq. (17)"},{"comment":"The Gaussian-process posterior covariance is misprinted. The standard posterior covariance is k(x,x') - K(x,X)[K(X,X)+sigma_n^2 I]^{-1} K(X,x'), whereas the displayed formula has a plus sign before the final K(X_j,x') and appears to omit the multiplication by the inverse. As written, the expression is dimensionally inconsistent (the first term is a scalar and the additive term is a vector) and is not the covariance that Algorithm 3 would use to build the posterior in Eq. (36). Since the posterior distribution is the basis of the expected hypervolume improvement acquisition function, this needs correction.","section":"Eq. (37)"},{"comment":"The adaptive reference point formula needs clarification. For the stated minimization problem, a hypervolume reference point should be worse than all observed points, i.e., larger than ymax, but Eq. (38) sets r = ymax - ymin*10%, which is smaller than the worst observed value. If the hypervolume is instead computed on transformed maximization objectives such as -C_ann and P_RES, that transformation should be stated explicitly and the reference point defined consistently on that transformed space. As written, the formula appears to place the reference point on the dominated side of the Pareto front, which would bias the hypervolume computation and affect the core algorithmic claim.","section":"Section IV-C1, Eq. (38)"},{"comment":"The comparison does not provide statistical support for the sample-efficiency and diversity claims. NSGA-II is run with a population size of 12, which is far below common practice and is likely to handicap it severely. The hypervolume curves in Figs. 7-8 and the error comparisons in Figs. 5 and 9 are reported without multiple independent runs, confidence intervals, or significance tests. To support the central claim that AMBO is superior in sample efficiency and Pareto diversity, please repeat each algorithm with several random seeds, use a more standard NSGA-II configuration, and report the mean and spread of the performance metrics.","section":"Section V-A3 and Figs. 6-9"}],"minor_comments":[{"comment":"The algorithm names are inconsistent: Section V-A3 lists PBO, while Figs. 6-7 use NMBO and PMBO without defining these labels. Please unify the nomenclature and state which abbreviation corresponds to which baseline.","section":"Section V-A3, Figs. 6-8"},{"comment":"The statement that NSGA-II 'evaluates over four times the number of capacity allocation schemes in each iteration' is difficult to reconcile with a population size of 12. Please clarify the evaluation budget per iteration for each algorithm, since this directly affects the interpretation of the sample-efficiency comparison.","section":"Section V-B"},{"comment":"For the comparison methods from [32] and the random method, four distinct days are drawn per month for the electric load, heat load, wind power, and solar power curves. Table II lists, for example, '20, 22, 13, 29' but does not indicate which day corresponds to which variable. Please clarify how a coherent daily scenario is assembled for these methods.","section":"Section V-A2, Table II"},{"comment":"The caption and text do not explain what the labels 'AMBO 1 2 3 4' refer to, whether these are four random initial noise standard deviations or four repeated runs, and no error bars are shown. Please describe the experimental design and add a measure of variability.","section":"Fig. 9"},{"comment":"The nomenclature contains typos such as 'Invement' for 'investment' and 'equimpment' for 'equipment'. In addition, references [18] and [19] appear to cite the same Applied Energy article under different author lists; please merge them or clarify the distinction.","section":"Nomenclature and references"}],"recommendation":"major_revision","confidential_remarks":"The novelty is incremental relative to qNEHVI (Daulton et al.), but the application to hybrid heat-source planning is appropriate for a transactions journal if the validation is strengthened. The main risk for the editor is that the empirical benefits are demonstrated only in-sample and without statistical replicates; I would ask for out-of-sample validation and corrected equations before considering acceptance. I also did not find any data or code availability statement, which is worth requesting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, incremental engineering paper, but the headline contribution—that AMBO's noise-aware filtering beats raw typical-scenario evaluation—is only shown on the same heating season that produced the typical days. The stress-test note is right. That doesn't sink the whole paper, but it does mean the claim that the planning results 'perform well under real operational conditions' isn't supported by the evidence.\n\nWhat's genuinely new: the AMBO variant (adaptive reference point from Eq. 38, sigma_n by marginal likelihood in Algorithm 2) is a modest but sensible extension of qNEHVI [29]. The time-series scenario construction in Section II-C, with K-medoids over four coupled curves then affine adjustment in Eq. (3), is a reasonable practical choice for cold-region studies. The operation simulation model is detailed, and the equations are mostly clear. Posterior-mean filtering in Algorithm 3 is a straightforward idea and a legitimate way to damp noise in the final Pareto front. Credit where due: the paper is readable, the engineering context is real, and the authors compare against three baselines.\n\nSoft spots, in order of severity. First, the validation is in-sample. The typical days are clustered from the same heating season that the SAA benchmark in Eq. (41) averages over, so Fig. 9 measures how well the typical days reconstruct that particular season, not how the method would cope with a different year. The error bars in Fig. 9 show the adaptive noise model helps, but there is no unseen-year test. Second, the algorithm comparisons in Figs. 6–8 and 11 are single runs; NSGA-II is given a population of 12, which is suspiciously weak for a population-based method. The sample-efficiency conclusion would be more convincing with multiple seeds and a properly tuned baseline. Third, Eq. (17) has a self-referential upper bound (P on both sides); it's likely a typo but should be fixed. Minor: the 'eliminates predefined parameters' claim overstates things—the 10% reference-point scalar and the kernel hyperparameters are still set by the user, and the sigma_n initialization in Algorithm 2 is random.\n\nOverall the central engineering model holds up and the algorithm is coherent; the weak point is the evidence for the noise-modeling benefit. Who gets value: people working on MOBO for energy-system planning, and anyone who wants a concrete application of qNEHVI-style acquisition. It deserves a serious referee, but the revision needs multi-seed runs, a fairer NSGA-II setting, and ideally a leave-one-year-out validation. If I were handling it, I'd ask for those before accepting.","headline":"A plausible engineering extension of qNEHVI for cold-region heat planning, but the noise-modeling benefit is only demonstrated in-sample.","tokens_in":17561,"tokens_out":2817,"would_cite":false,"duration_ms":27452,"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":"This paper claims that making two Bayesian-optimization parameters adaptive—the hypervolume reference point and the noise standard deviation—lets capacity planners for cold-region electric-heat systems produce more diverse, more evenly…","keywords":["Hybrid heat sources","Electric-heat coupling systems","Capacity planning","Multi-objective optimization","Time series scenarios","Bayesian optimization","Hypervolume indicator","Renewable energy curtailment"],"falsifier":"Run the proposed scenario generation and AMBO on a held-out future heating season: select typical days from a training year, evaluate the final planned capacities against that year's full-season sample-average cost and renewable consumption, and compare the posterior-mean errors ($e_{ann}^B$, $e_{RES}^B$) with the raw typical-scenario errors ($e_{ann}^{typ}$, $e_{RES}^{typ}$). If the posterior-mean filtering no longer reduces the renewable-consumption error below the raw typical-scenario error, the paper's central noise-modeling claim would be contradicted. A second check: apply AMBO to a synthetic multi-objective benchmark with a known true Pareto front and known heteroscedastic noise; if its achieved hypervolume does not beat NMBO with optimally tuned parameters across many seeds, the adaptive-reference-point advantage is not general.","tokens_in":16485,"feed_emoji":"🔥","tokens_out":8608,"duration_ms":80778,"temperature":0.7,"pith_summary":"This paper argues that capacity planning for hybrid heat sources in cold-region electric-heat coupling systems should be treated as a noisy multi-objective optimization problem, and that an adaptive Bayesian optimizer—AMBO—can solve it better than standard baselines. AMBO removes two user-set parameters that other Bayesian multi-objective methods require: the reference point used in the hypervolume indicator and the noise standard deviation of the Gaussian-process surrogate. Both are updated automatically from the data, and the final Pareto front is taken from the posterior mean rather than from noisy simulation outputs. The paper also proposes a time-series scenario generation method that clusters on the coupled statistics of heat load and net load, then rescales the selected days to match the original variability. If the claims hold, planners in cold regions get a more diverse and evenly distributed set of cost-versus-renewable-consumption trade-offs with far fewer expensive simulation runs than evolutionary search.","feed_headline":"Adaptive optimizer finds better heat-source trade-offs in fewer runs","feed_subtitle":"Cold-region planners get a diverse, evenly spread set of cost-versus-renewable options without hand-tuning optimizer parameters.","key_machinery":"The load-bearing machinery is AMBO, a Bayesian multi-objective optimizer built on the noisy expected hypervolume improvement acquisition function—the expected gain, under Gaussian-process posterior samples, in the hypervolume (volume of dominated objective space) of the current Pareto front. AMBO makes two components adaptive: the reference point $r = \\hat{y}_{\\max} - \\hat{y}_{\\min} \\times 10\\%$, recomputed from the worst observed objectives, and the noise standard deviation $\\sigma_n$, re-estimated each iteration by gradient ascent on the Gaussian-process marginal log likelihood. In the final step the Pareto front is taken from the posterior mean $\\mu_{D_J}$ rather than from the noise-corrupted observations $Y_J$. A second supporting mechanism is the time-series scenario generator: one day per heating-season month is chosen by K-medoids clustering on the mean and variance of heat load and net load, and the selected days are linearly rescaled via $S_{typ} = (S'_{typ})a + b$ to match the original curves' statistics.","core_discovery":"The discovery is that the planner-chosen parameters in Bayesian multi-objective optimization—the reference point for hypervolume and the noise standard deviation—are not innocuous defaults: fixing them narrows and clusters the Pareto front, while adapting them to the observed data spreads it out. AMBO sets the reference point as the observed worst case minus 10% of the observed objective range, and re-estimates the noise standard deviation each iteration by gradient ascent on the marginal log likelihood of the Gaussian process. The final non-dominated set is computed from the posterior mean $\\mu_{D_J}(x)$ rather than the raw objective values, which the authors show cuts the evaluation error of the renewable-consumption objective relative to the sample-average-approximation benchmark. The accompanying scenario-generation method, which selects typical days by K-medoids clustering on the mean and variance of heat load and net load and then applies the linear adjustment $S_{typ} = (S'_{typ})a + b$, reduces simulation error against the full-season benchmark compared with independent typical-day selection.","pith_inferences":["The same adaptive-noise mechanism could transfer to other expensive multi-objective design problems where the objective noise level is unknown and non-stationary, not just power-system planning; a natural test is applying AMBO to standard noisy MOO benchmarks with known ground-truth Pareto fronts.","Since the SAA benchmark uses the same historical heating season from which the typical days are drawn, the claimed accuracy gains are in-sample; a held-out-year test could either confirm the gains or show that the scenario-adjustment method overfits the clustering year.","The linear rescaling in Eq. (3) matches means and variances only, leaving higher-order temporal dependencies unadjusted; checking whether the Pareto front changes when autocorrelation or ramping statistics are also matched would clarify how much of the benefit comes from the clustering versus the scaling.","The claimed 'no predefined parameters' still leaves choices such as the 10% reference-point margin, the kernel hyperparameters, and the number of clusters; a sensitivity analysis over these residuals would show how much of the robustness is truly automatic."],"forward_implications":["AMBO's adaptive reference point and noise standard deviation remove the planner-tuned parameters that can bias capacity planning, so the same optimizer can be applied to different regions or system sizes without manual recalibration.","Because AMBO needs far fewer operation-simulation evaluations than NSGA-II, the planning model can afford more detailed or longer-horizon simulations at equal compute.","Filtering the Pareto front with the Gaussian-process posterior mean instead of raw evaluations reduces the error in the renewable-consumption objective from over 10% to a more usable level, making typical-scenario planning results more trustworthy.","The cold-region scenario generator that clusters on coupled heat-load and net-load statistics yields lower average evaluation error than independent typical-day selection, so it better represents the real operating environment for planning.","Integrating all four heat sources—electric boiler, heat pump, TES, and combined storage heaters—produces a markedly wider and more uniformly distributed Pareto front than planning either generation-side or demand-side equipment alone."],"supporting_citations":[{"why":"Supplies the noisy expected hypervolume improvement acquisition function and Gaussian-process framework that AMBO extends.","marker":"[29]"},{"why":"Provides the PBO baseline, a Bayesian multi-objective method with adaptive coefficients, compared in the case study.","marker":"[21]"},{"why":"Supplies the NSGA-II evolutionary baseline whose sample efficiency is compared against AMBO.","marker":"[18]"},{"why":"Provides the alternative typical-scenario selection method used as a comparison for the proposed scenario generation.","marker":"[32]"},{"why":"Establishes the sample average approximation benchmark that defines the true objective values in the noise-modeling evaluation.","marker":"[33]"},{"why":"Defines the heating-season operation simulation model that the planning objectives are evaluated with.","marker":"[28]"},{"why":"Gives the expected hypervolume improvement formulation referenced for the acquisition function calculations.","marker":"[30]"},{"why":"Provides the software library used to implement the Gaussian-process optimization in the case study.","marker":"[31]"}],"fun_headline_variants":["Adaptive optimizer widens Pareto front for heat-source planning","Fewer runs, broader trade-offs in cold-region heat planning","No hand-tuning: adaptive optimizer improves heat planning","Sample-efficient optimizer diversifies cold-region heat options"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that the typical days extracted from one historical heating season, and the full-season sample-average benchmark built from that same season, faithfully represent how the planned system will actually perform; the validation never tests on a truly unseen year, so the noise-modeling and scenario-generation benefits could shrink or vanish for future data.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive optimizer widens Pareto front for heat-source planning","Fewer runs, broader trade-offs in cold-region heat planning","No hand-tuning: adaptive optimizer improves heat planning","Sample-efficient optimizer diversifies cold-region heat options"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1528,"prompt_tokens":979,"completion_tokens":549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":483}},"tokens_in":595,"tokens_out":549,"duration_ms":5749,"temperature":1.0,"reasoning_tokens":483,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:02:52.947674+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed scenario generation and AMBO on a held-out future heating season: select typical days from a training year, evaluate the final planned capacities against that year's full-season sample-average cost and renewable consumption, and compare the posterior-mean errors ($e_{ann}^B$, $e_{RES}^B$) with the raw typical-scenario errors ($e_{ann}^{typ}$, $e_{RES}^{typ}$). If the posterior-mean filtering no longer reduces the renewable-consumption error below the raw typical-scenario error, the paper's central noise-modeling claim would be contradicted. A second check: apply AMBO to a synthetic multi-objective benchmark with a known true Pareto front and known heteroscedastic noise; if its achieved hypervolume does not beat NMBO with optimally tuned parameters across many seeds, the adaptive-reference-point advantage is not general.","supporting_citations":[{"cited_title":"Parallel bayesian optimization of multiple noisy objectives with expected hypervolume improvement,","cited_arxiv_id":null,"evidence_quote":"Supplies the noisy expected hypervolume improvement acquisition function and Gaussian-process framework that AMBO extends."},{"cited_title":"Parego: a hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems,","cited_arxiv_id":null,"evidence_quote":"Provides the PBO baseline, a Bayesian multi-objective method with adaptive coefficients, compared in the case study."},{"cited_title":"Multi-objective stochastic expansion plan- ning based on multi-dimensional correlation scenario generation method for regional integrated energy system integrated renewable energy,","cited_arxiv_id":null,"evidence_quote":"Supplies the NSGA-II evolutionary baseline whose sample efficiency is compared against AMBO."},{"cited_title":"Resilience oriented planning of urban multi- energy systems with generalized energy storage sources,","cited_arxiv_id":null,"evidence_quote":"Provides the alternative typical-scenario selection method used as a comparison for the proposed scenario generation."},{"cited_title":"Adap- tive characteristic modeling of long-period uncertainties: A multi-stage robust energy storage planning approach based on the finite covering theorem,","cited_arxiv_id":null,"evidence_quote":"Establishes the sample average approximation benchmark that defines the true objective values in the noise-modeling evaluation."},{"cited_title":"Evaluation method of renewable energy consumption capacity of integrated heat and power energy system in heating season based on time series production simulation,","cited_arxiv_id":null,"evidence_quote":"Defines the heating-season operation simulation model that the planning objectives are evaluated with."},{"cited_title":"Differentiable expected hypervolume improve- ment for parallel multi-objective bayesian optimization,","cited_arxiv_id":null,"evidence_quote":"Gives the expected hypervolume improvement formulation referenced for the acquisition function calculations."},{"cited_title":"Botorch: A framework for efficient monte-carlo bayesian optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the software library used to implement the Gaussian-process optimization in the case study."}],"review_version":1}