{"id":"5e7fa66f-cbc3-4021-83fe-db3b1ffc8cd3","arxiv_id":"2608.08790","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":17,"one_line_summary":"A risk-aware stochastic pump scheduling framework with probabilistic forecasts reduces imbalance volumes and improves reliability, with modest cost savings versus a price-invariant benchmark.","lead":"This paper builds a stochastic optimization framework for scheduling water pumps when water demand and electricity prices are uncertain, and tests it on data from a German water supplier. It finds the main benefit is operational robustness rather than lower expected costs, and that water demand forecasts matter more than electricity price forecasts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RES forecast leakage in day-ahead price models violates non-anticipativity and may inflate the 9% cost-saving claim and the conclusion that price-forecast improvements have limited value.","rationale":"The reader identified the RES non-anticipativity violation as the weakest assumption; I agree. This concern is load-bearing because it sits inside the forecasting models that drive the policy optimization (Eqs. 29 and 30) and is explicitly admitted in Section 1.2. The paper's own statement “we accept this” does not neutralize the effect. The central claim—9% cost savings and limited value of further price-forecast improvements—depends on the quality of the day-ahead price forecasts; leakage of post-cutoff information can only improve them, potentially overstating both the savings and the “limited value” conclusion. A clean re-run without the unavailable regressors is a concrete, feasible check. Other concerns, such as the conflation of stochasticity with the sequential overlapping horizon in SH vs DET FH comparisons, are also present and would require an isolated design (e.g., a stochastic full-horizon variant), but the RES leakage is the most direct threat to the headline number. The paper's otherwise transparent and detailed methodology, with reproducible components, supports a conditional rather than reject verdict.","tokens_in":25746,"tokens_out":7639,"duration_ms":76855,"concrete_test":"Re-run the numerical study with day-ahead price forecasting models (Eqs. 29 and 30) re-estimated without the dRES regressors (and any other data published after the 12:00 day-ahead gate closure). Regenerate the joint forecast ensembles, re-apply the same scenario reduction and policy optimization, and recompute the cost savings relative to FB DET FH and the price-forecast sensitivity analysis. If the headline 9% savings drops by more than a few percentage points or the marginal value of price-forecast improvements becomes non-negligible, the non-anticipativity violation materially overstates the results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the acknowledged non-anticipativity violation in Section 1.2: “RES forecasts are published only after DA gate closure, so their inclusion violates non-anticipativity. Nevertheless, we accept this.” The day-ahead price models in Eqs. (29) and (30) include dRES (forecast renewable generation) as a regressor. Because scheduling decisions for day d+1 are fixed by 12:00 on day d, while RES forecasts appear only later, the price forecasts used in the stochastic optimization are informed by information unavailable to a real operator. This leakage directly affects the quantitative headline: the 9% cost savings versus the price-invariant fall-back and the conclusion that further price-forecast improvements have limited value. If the realistic (non-leaked) price forecasts are worse, the optimal policy will be less able to exploit price variations, reducing the savings; conversely, the “limited value of price improvements” may be an artifact of already-leaked information. The paper's acceptance of the violation is not a fix. To test materiality, remove all RES-based regressors from DAPF models, re-estimate, regenerate forecasts, and re-run the full policy evaluation. If the 9% savings or the price-value conclusion shifts materially, the central claims need qualification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a sequential decision-making framework for pump scheduling in a drinking water supply system, integrating probabilistic forecasts of water demand, day-ahead electricity prices, and imbalance prices into a risk-aware stochastic optimization problem. The objective function is lexicographic, combining expected operational costs, Conditional Value-at-Risk of economic costs, and reliability penalties based on exceedance risk measures. The framework explicitly models day-ahead procurement with subsequent imbalance settlement and uses scenario reduction and a MILP reformulation for tractability. In a numerical study with data from a German water supplier, the authors compare forecast-driven policies (Naive, LEAR, GAMLSS) against a price-invariant fall-back policy and perfect-foresight baselines, report cost savings of up to 9% versus the fall-back, and conclude that the primary benefit of stochastic optimization lies in improved operational robustness rather than lower expected costs.","tokens_in":26069,"tokens_out":5997,"duration_ms":59531,"significance":"If the quantitative findings are robust, the paper is a valuable methodological contribution at the interface of probabilistic forecasting and stochastic optimal control for water systems. It offers a detailed integration of forecast ensembles with a risk-aware, imbalance-aware scheduling model, a careful evaluation of scenario reduction methods, and a clean separation of forecast and policy evaluation. The explicit treatment of the imbalance market and the lexicographic risk objective are strengths. However, the headline quantitative claims currently rest on an acknowledged non-anticipativity violation in the price forecast models and on policy performance figures reported without any measures of uncertainty, which limits the confidence that can be placed in the reported rankings and the 9% savings figure.","major_comments":[{"comment":"The day-ahead price forecasting models include dRES (forecast renewable generation) as a regressor, while the paper itself states that RES forecasts are published only after day-ahead gate closure and that their inclusion violates non-anticipativity. Since day-ahead procurement decisions for delivery day d+1 are fixed at 12:00 on day d, price forecasts used in the optimization are informed by information that a real operator would not possess at decision time. This leakage can inflate the reported 9% cost savings relative to the price-invariant benchmark and can also distort the conclusion that further price-forecast improvements have limited value, because the leaked information may already capture part of the exploitable price signal. The paper's acknowledgment of the violation does not address its materiality. I request a robustness analysis in which the dRES terms are removed from the DAPF models, the forecasts are regenerated, and the full policy evaluation is rerun; if the cost-savings figure or the price-value conclusion changes materially, the abstract and conclusions must be qualified.","section":"Section 1.2, Eqs. (29) and (30)"},{"comment":"All policy performance indicators are reported as single averages over 100 runs and 3 repetitions, with no standard errors, confidence intervals, or statistical tests. Several policy comparisons that carry the paper's message involve very small cost differences, for example GAMLSS SH with λrisk=1 (6100.44 EUR), LEAR SH (6097.34 EUR), and GAMLSS SH (6102.17 EUR). Without a measure of dispersion, the ranking of policies and the extrapolated annual savings of 236,355.75 EUR (about 9%) cannot be distinguished from sampling noise. I request that the authors report run-level variability (e.g., standard errors or quantile ranges) for the Table 4 indicators and, where appropriate, apply tests to the cost differences between the main policies.","section":"Table 4"},{"comment":"The sensitivity-analysis conclusions, in particular that the marginal value of further improvements in electricity price forecasts appears comparatively limited and that most exploitable price information is already captured, are based on comparisons such as GAMLSS PF EPF SH versus GAMLSS SH that are reported without any uncertainty measures. As these claims are central takeaways of the paper, they should be either supported by statistical evidence (error bars, confidence intervals, or formal tests) or restated more cautiously as qualitative observations that are not yet statistically grounded.","section":"Section 3.3"}],"minor_comments":[{"comment":"The assumption numbering skips Assumption 3, and Assumption 5 refers to Assumption 1.1 for the retroactive day-ahead participation; please renumber the assumptions consistently.","section":"Section 1.1, Assumptions"},{"comment":"In the paragraph on electricity procurement, 'Wile preliminary IBP estimates' should read 'While preliminary IBP estimates'.","section":"Section 1.1"},{"comment":"At the start of Section 2.3, 'In the proceeding we first derive' should read 'In the following we first derive'.","section":"Section 2.3"},{"comment":"The sentence 'The resulting model are denoted as DAPF GAMLSS tdist' should read 'The resulting models are denoted as DAPF GAMLSS tdist'.","section":"Section 2.3.3"},{"comment":"The notation HDI(w_WD_t-L, t) is introduced without defining w_WD_t-L, while the surrounding text refers to D_t-L; please align the notation.","section":"Equation (31)"},{"comment":"The footnote markers asterisk and double asterisk are explained in the table region, but the explanation is not integrated into the table caption; please move the notes into a proper table note so that the meaning of the starred rows is unambiguous.","section":"Table 4"},{"comment":"The extrapolation from the normalized costs in Table 4 to the annual savings of 236,355.75 EUR is not described; please provide the calculation formula so that the 9% headline figure can be reproduced.","section":"Section 3.3, extrapolation"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is single-authored and the author explicitly acknowledges the non-anticipativity violation that affects the main quantitative claims. The lack of uncertainty quantification in Table 4 is also a concern. In my view the paper's framework is defensible and the issues are fixable, but the robustness analysis and statistical reporting are prerequisites for publication. The editor may also wish to verify that the data and code are available or will be made available upon request."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper's qualitative finding—stochastic pump scheduling improves reliability more than it reduces expected cost, and water-demand forecasts matter more than price forecasts—is plausible and likely to survive scrutiny. The paper's quantitative headline, up to 9% savings over a price-invariant benchmark, is on shakier ground, because the day-ahead price forecasts use renewable-generation forecasts published only after gate closure. The author openly tags this as a non-anticipativity violation but then proceeds anyway. That means the 9% and the \"limited value of better price forecasts\" conclusion are both contaminated by information a real operator would not have. The fix is straightforward: re-estimate the price models without the RES regressors, regenerate forecasts, rerun the policy evaluation, and report whether the savings and the value-of-information conclusion move. This must happen before the numbers can be trusted.\n\nWhat is genuinely new here is the integration: probabilistic demand and price forecasts feeding a risk-aware lexicographic objective (expected cost, CVaR on cost, exceedance penalties on storage and head violations) with explicit imbalance-market settlement, tested on real utility data. The empirical observation that scenario reduction preserves the overall distribution but not necessarily the tails that drive CVaR is useful and non-obvious. The forecasting work is competent: proper scoring rules, Diebold–Mariano tests, copula-based joint trajectories, and a sensitivity analysis around scenario count and risk weights. The author is also transparent about the study design's shortcuts, including the finite-horizon retroactive day-ahead participation and the fallback benchmark's perfect demand foresight.\n\nThe soft spots beyond the leakage are minor. Policy costs in Table 4 are point estimates without error bars, so the few-euro differences among top policies are not statistically supported. Reducing the ensemble to five scenarios is aggressive for a problem where tail risk is the central concern. And the fallback benchmark, which is partially price-invariant but demand-perfect, makes the 9% hard to interpret as a pure value-of-price-forecasting claim. None of these sink the qualitative conclusions; they do argue for treating all quantitative comparisons as directional.\n\nThis is a paper worth refereeing seriously. A good referee would push for the non-leaked re-run, confidence intervals on policy costs, and a more tail-aware scenario reduction. The framework and the qualitative findings merit that effort.","headline":"The qualitative findings are credible, but the headline 9% savings rest on price forecasts that leak future information; re-run without RES regressors before trusting the numbers.","tokens_in":26599,"tokens_out":3523,"would_cite":false,"duration_ms":37732,"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":"Probabilistic pump schedules cut water-supply power costs up to 9 percent.","keywords":["pump scheduling","water supply","probabilistic forecasting","stochastic optimization","conditional value-at-risk","imbalance market","day-ahead electricity price","scenario reduction"],"falsifier":"Re-run the numerical study with day-ahead price models estimated strictly from information available before gate closure (dropping RES forecast covariates), keeping everything else fixed. If the best stochastic policy's saving over the price-invariant benchmark falls well below 9 percent, the claimed economic benefit depends on the non-anticipativity violation; if the saving is unchanged, the violation is immaterial.","tokens_in":25475,"feed_emoji":"💧","tokens_out":7561,"duration_ms":71497,"temperature":0.7,"pith_summary":"This paper argues that a drinking-water utility can cut operating cost by up to 9 percent relative to a price-insensitive benchmark by scheduling pumps with probabilistic forecasts of water demand and electricity prices, while explicitly accounting for day-ahead procurement and imbalance-market settlement. The central finding is that the main gain from stochastic optimization is operational robustness: stochastic policies reduce imbalance energy, delay constraint violations, and improve reliability, while expected costs are only slightly lower than deterministic alternatives. The paper also reports that water demand uncertainty is the dominant driver of policy quality, and that further improvements in electricity-price forecasts add little value because the economically exploitable price signal is already captured. If correct, the results imply that utilities should invest in demand forecasting and risk-aware control rather than chasing ever-more-accurate price forecasts.","feed_headline":"Probabilistic pump schedules cut water-supply power costs up to 9%","feed_subtitle":"Stochastic scheduling buys reliability: less imbalance energy and later constraint violations, not just cheaper power.","key_machinery":"The load-bearing mechanism is a stochastic lookahead policy solved as a mixed-integer linear program over a multistage scenario tree. Probabilistic forecasts of water demand, day-ahead prices, and imbalance prices are generated by naive, LASSO-autoregressive (LEAR), and GAMLSS distributional-regression models, merged into a 1,000-member joint ensemble, reduced to five scenarios by energy-distance forward selection, and branched only on the water-demand process. The objective is lexicographic: expected economic cost plus Conditional Value-at-Risk (the average loss in the worst-performing cost scenarios), pump-smoothing penalties, and two-regime exceedance-risk penalties for storage and head violations. This structure lets the optimizer trade a controlled increase in imbalance exposure against cheaper day-ahead procurement while keeping reliability constraints binding.","core_discovery":"On the paper's own terms, the discovery is that a lexicographic, risk-aware formulation of the multi-stage pump-scheduling problem—combining expected cost, Conditional Value-at-Risk of cost, and exceedance-risk penalties for storage and pump-head violations—translates forecasting skill into operational value. In a numerical study on real-world data from a German water supplier (2021–2025, with 2025 held out), the best stochastic policy achieves normalized costs of about 6,097 EUR per optimization run versus 6,595 EUR for the price-invariant fall-back benchmark, an extrapolated annual saving of roughly 236,356 EUR, i.e., about 9 percent. The paper emphasizes that the benefit is primarily risk reduction: stochastic policies cut imbalance energy and delay first constraint violations by several hours relative to deterministic forecasts, even though expected costs are similar.","pith_inferences":["If the renewable-forecast non-anticipativity violation were corrected (training price models only on information available before gate closure), the 9 percent saving might shrink; a natural test is to re-estimate the day-ahead price models without the leaked RES covariates and re-run the policy comparison.","A decision-aware scenario-reduction criterion that targets the cost tail rather than the full distribution could recover part of the gap between reduced-set and full-ensemble performance and is a direct, testable extension of the energy-distance approach used here.","The finding that price-forecast improvements have limited marginal value is conditional on the studied market (German day-ahead plus imbalance settlement); in markets with different imbalance-price dynamics or stronger price spikes, price-forecast skill may matter more.","Because the paper assumes zero-mean imbalance deviations (no deliberate arbitrage), the 9 percent figure is a conservative estimate of what a fully optimizing trader could achieve if allowed to take intentional imbalance positions; extending the market-neutrality constraint is a natural next step."],"forward_implications":["Utilities that adopt stochastic, risk-aware scheduling can expect cost savings on the order of 9 percent relative to conventional price-invariant operation, with most of the gain coming from reduced imbalance exposure and delayed reliability violations rather than lower expected procurement cost.","Improvements in water-demand forecasts matter more for policy quality than equally-sized improvements in electricity-price forecasts, so forecast-development effort should be weighted toward demand.","Because deterministic policies incur substantially more imbalance energy and earlier violations, using deterministic point forecasts in place of probabilistic ones sacrifices robustness even when average costs look similar.","Scenario reduction that preserves the overall distribution does not automatically preserve the tail events that drive CVaR- and exceedance-risk-based decisions; set sizes and reduction criteria should be chosen with the downstream risk objective in mind.","The framework is transferable: because the hydraulic model is deliberately simplified to a storage mass balance plus affine pump curves, the same forecasting-and-optimization pipeline can be extended to site-specific networks without changing the core approach."],"supporting_citations":[{"why":"Supplies the unified sequential-decision framework (state, decision, exogenous information, transition, policy search) used to formulate the base model.","marker":"Powell (2019)"},{"why":"Provides the CVaR formulation used both in the economic risk objective and in deriving risk-aware storage lower bounds.","marker":"Rockafellar and Uryasev (2000)"},{"why":"Provides the energy-distance forward-selection scenario reduction with reweighting that compresses the 1,000-member ensemble to five scenarios.","marker":"Ziel (2021)"},{"why":"Provides the multistage scenario-tree construction used to represent branching in the stochastic lookahead policy.","marker":"Heitsch and Römisch (2009)"},{"why":"Defines the naive day-ahead-price benchmark model that is converted into probabilistic forecasts and serves as the baseline forecast method.","marker":"Marcjasz, Uniejewski, and Weron (2020)"},{"why":"Motivates the treatment of imbalance prices as a hard-to-predict process, framing the imbalance-forecast modeling and its expected limited skill.","marker":"Browell and Gilbert (2022)"}],"fun_headline_variants":["Risk-aware pump scheduling trims water-supply energy bills up to 9%","Probabilistic demand forecasts make pump schedules robust, not just cheap","Stochastic pump control balances cost and risk in water supply","Forecast skill pays off in water pump scheduling via risk hedging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline saving assumes the day-ahead price forecasts can use renewable-generation forecasts that are only published after the day-ahead market closes; if that leaked information materially improves the price forecasts, the simulated schedules and the 9 percent saving are better than what a real operator could achieve in real time.","fun_headline_variants_meta":{"raw":{"variants":["Risk-aware pump scheduling trims water-supply energy bills up to 9%","Probabilistic demand forecasts make pump schedules robust, not just cheap","Stochastic pump control balances cost and risk in water supply","Forecast skill pays off in water pump scheduling via risk hedging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":1978,"prompt_tokens":933,"completion_tokens":1045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":970}},"tokens_in":549,"tokens_out":1045,"duration_ms":8409,"temperature":1.0,"reasoning_tokens":970,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:24:07.884751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the numerical study with day-ahead price models estimated strictly from information available before gate closure (dropping RES forecast covariates), keeping everything else fixed. If the best stochastic policy's saving over the price-invariant benchmark falls well below 9 percent, the claimed economic benefit depends on the non-anticipativity violation; if the saving is unchanged, the violation is immaterial.","supporting_citations":[{"cited_title":"Tyrrell and Uryasev, Stanislav , title =","cited_arxiv_id":null,"evidence_quote":"Provides the CVaR formulation used both in the economic risk objective and in deriving risk-aware storage lower bounds."},{"cited_title":"Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume =","cited_arxiv_id":null,"evidence_quote":"Provides the energy-distance forward-selection scenario reduction with reweighting that compresses the 1,000-member ensemble to five scenarios."},{"cited_title":"Scenario tree modeling for multistage stochastic programs , journal =","cited_arxiv_id":null,"evidence_quote":"Provides the multistage scenario-tree construction used to represent branching in the stochastic lookahead policy."},{"cited_title":"International Journal of Forecasting , volume =","cited_arxiv_id":null,"evidence_quote":"Defines the naive day-ahead-price benchmark model that is converted into probabilistic forecasts and serves as the baseline forecast method."},{"cited_title":"Energies , VOLUME =","cited_arxiv_id":null,"evidence_quote":"Motivates the treatment of imbalance prices as a hard-to-predict process, framing the imbalance-forecast modeling and its expected limited skill."}],"review_version":1}