The Value of Perfect Endpoint Forecasts for Offshore-Wind Thermal Firming

Arash Khojaste; Geoffrey Pritchard; Golbon Zakeri

REVIEW 2 major objections 5 minor 15 references

A one-hour perfect endpoint forecast cuts offshore-wind firming cost by 8.07%, while a six-hour endpoint cuts only 2.28%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 04:36 UTC pith:ZXCA5U5D

load-bearing objection The non-nested endpoint forecast value idea is the real contribution; the specific VPI numbers need a stronger kernel-consistency check before being taken at face value. the 2 major comments →

arxiv 2607.13627 v1 pith:ZXCA5U5D submitted 2026-07-15 math.OC

The Value of Perfect Endpoint Forecasts for Offshore-Wind Thermal Firming

Arash Khojaste , Geoffrey Pritchard , Golbon Zakeri This is my paper

classification math.OC MSC 90C4090C05

keywords offshore wind integrationthermal firmingendpoint forecastsvalue of perfect informationMarkov decision processesquantile regressioncyclostationary

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks how much a perfect forecast of net demand at a single future hour is worth to a thermal firming operator, independently of forecast accuracy. It embeds a one-target 'endpoint' signal into a cyclostationary Markov decision process calibrated to ISO New England load and offshore-wind data, solving hour-by-hour linear programs for horizons of one to six hours. The central finding is that endpoint value declines steeply with horizon: the one-hour target reduces annual cost by 8.07%, while the six-hour target reduces it by only 2.28%. The reason is structural: the controller's action today positions the firming level for one hour ahead, so the most actionable future state is the next one. Because endpoint information sets are not nested across horizons, this decline does not contradict the usual principle that more information is weakly better.

Core claim

For a one-step ramping decision, perfect knowledge of the net-demand state one hour ahead is worth substantially more than perfect knowledge of a more distant state. In the ISO New England case study, the value of perfect endpoint information falls from 1,049 cost units (8.07% of the no-forecast baseline) at h=1 to 297 units (2.28%) at h=6. The paper's conceptual contribution is to separate this diagnostic endpoint question from rolling path forecasts: with endpoints, the information sets are non-nested, so there is no monotonicity theorem; the observed decline is explained by the one-hour actuation lag of the firming resource.

What carries the argument

The forecast-augmented cyclostationary MDP with state (ℓ_t, z_t, e_t^(h)), where e_t^(h)=z_{t+h} is the single perfect endpoint. The joint transition kernel for (z_t, z_{t+h}) → (z_{t+1}, z_{t+1+h}) is estimated by maximum likelihood under a periodic Fourier form (Eq. 11), with row-wise stochasticity and a marginal-consistency condition against the baseline one-step chain. The model is solved as a state-action frequency linear program; the difference between the baseline and forecast-augmented optimal costs defines VPI_end(h). The non-nested information sets are the key structural feature that allows value to decrease with horizon.

Load-bearing premise

The quantitative profile (8.07% to 2.28%) rests on the estimated joint endpoint transition kernel faithfully representing real net-demand dynamics; if the kernel is misspecified—for example, if net demand has non-periodic residual structure or longer memory than the M=4 states capture—the reported VPI values could shift.

What would settle it

Solve the same endpoint-augmented LP with a known ground-truth transition model (e.g., a calibrated ARMA or a higher-order Markov chain) instead of the estimated Fourier kernel; if the VPI_end(h) ordering changes or the h=1 advantage disappears, the decreasing profile is an artifact of the fitted kernel. Alternatively, run the same experiment with a two-level-per-hour ramp limit: if VPI_end(2) exceeds VPI_end(1), the one-hour actuation-delay explanation would be contradicted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Forecast value should be evaluated with the same information structure and timing as the decision model that consumes the forecast; endpoint-only and rolling path forecasts answer different economic questions.
For thermal firming with a one-hour ramp delay, the one-hour-ahead target is the most actionable endpoint; resources with other actuation delays could exhibit different profiles.
The reported VPI numbers are upper bounds for imperfect forecast products, since the forecasts are assumed perfect.
The method generalizes to other storage-like or fast-ramping technologies by changing the state and transition dynamics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The declining endpoint-value profile likely reflects a general principle for any resource with a one-step actuation delay: information about the immediately affected state dominates information about more distant states, even when the distant state is correlated with it. This could be tested by varying the ramp limit.
The 2.28% value at h=6 is not a bound on the value of a six-hour rolling path forecast; a path forecast would be nested and would be at least as valuable as h=1 under the same cost structure. Combining an endpoint signal with intermediate path information would likely restore monotonicity.
A natural extension is to compute the value of the intermediate states z_{t+1},...,z_{t+h-1} themselves, e.g., by adding them one at a time, to decompose how much of the rolling-forecast value comes from each horizon.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

The non-nested endpoint forecast value idea is the real contribution; the specific VPI numbers need a stronger kernel-consistency check before being taken at face value.

read the letter

Here's the short version: the paper makes a genuinely new conceptual point about forecast value—endpoint-only information sets are not nested across horizons, so a longer perfect endpoint forecast need not be more valuable—and backs it with a clean LP formulation and a plausible case study. The punchline empirical claim, VPI_end decreasing from 8.07% at h=1 to 2.28% at h=6, is exactly the kind of result that gets cited. But I wouldn't put much weight on the specific numbers until the transition-kernel issue is resolved.

What the paper does well: Proposition 1 is correct, and the argument that a forecast coordinate can't hurt when the kernel satisfies the marginal-consistency condition (13) is clean. The authors are also honest about the diagnostic nature of the endpoint structure and about normalized cost weights. The distinction between rolling path forecasts (nested, monotone VPI) and endpoint forecasts (non-nested, no monotonicity) is new to this literature and worth taking seriously.

The soft spot is real. The augmented transition kernel (Eq. 11) is estimated separately for each horizon h from paired state sequences. Under the authors' own baseline QFR chain (Eq. 3), the joint transition of (z_t,z_{t+h}) is determined by the baseline one-step matrices and their composition. The separately fitted kernels are not checked against this implied kernel; the reported consistency check only compares the current-state marginal (max discrepancy 0.0199). That is not enough to ensure the six LPs describe the same stochastic process. So the VPI differences could be partly an artifact of kernel estimation, not a pure measure of endpoint information value. This is addressable—derive the augmented kernels from the baseline model, or bootstrap the kernel estimates and report uncertainty—but as it stands, the headline numbers are not well-grounded. This is a moderate flaw, not a fatal one: the monotone ordering is smooth and consistent with the paper's interpretation, so the qualitative finding probably survives.

Readers in forecast evaluation and renewable integration will get value from the non-nested framing even if the case-study numbers are treated tentatively. I'd send this to a competent referee with a request for revision on the kernel consistency and some sensitivity analysis. It doesn't deserve desk rejection, but it does need the authors to make the comparison across horizons apples-to-apples.

Referee Report

2 major / 5 minor

Summary. The paper studies the value of a perfect endpoint forecast in a cyclostationary Markov decision process for offshore-wind thermal firming. For each horizon h=1,...,6, the controller observes the current net-demand state and one perfect future state z_{t+h}, but no intermediate path or past endpoint messages. The augmented MDP is solved by a state-action-frequency LP, and the value of perfect endpoint information is defined as the cost reduction relative to the no-forecast baseline. The central empirical claim is that this value decreases strictly with h: a one-hour endpoint saves 8.07% of annual firming cost while a six-hour endpoint saves 2.28%. The paper argues that endpoint information sets are not nested, so no monotonicity theorem applies, and that the profile reflects the one-hour ramping actuation delay of the resource.

Significance. If the empirical profile is robust, the paper makes a useful diagnostic contribution: it separates forecast value from forecast target horizon and cautions that forecast evaluation must match the information structure of the operating model. The LP formulation is clean, the non-nested information-structure argument is conceptually valuable, and the authors are explicit about the diagnostic rather than operational nature of the endpoint signal. The use of public data and solver-certified optima is a strength. The main caveat is that the central numerical profile depends on a set of separately estimated transition kernels whose consistency with the baseline model is only partially verified.

major comments (2)

[§III.B, §V, Table II] The central decreasing VPI profile is computed from endpoint kernels (Eq. 11) that are fitted by maximum likelihood separately for each h. Proposition 1 requires the marginal-consistency condition (13), but the reported check in §V verifies only one-step current-state marginals, not the full pair transition law; the maximum absolute discrepancy is 0.0199 and the mean is 0.0026. Under the paper's own baseline QFR chain (Eq. 3), the joint endpoint kernel is uniquely determined by the h-step composition of the one-step transition matrices and would satisfy (13) exactly. Because each C_h is instead the optimum of a slightly different fitted process, the reported VPI differences may reflect estimation discrepancies across h rather than the genuine value of endpoint information. I request either (a) recomputing the profile with kernels derived from the baseline chain, (b) a sensitivity analysi
[§VI.B, Table II] The paper's main claim is the strict ordering VPI_end(1) > ... > VPI_end(6). The VPI numbers are point estimates obtained from transition kernels estimated on a finite sample, yet no uncertainty quantification is provided. Even if the kernel inconsistency in the first major comment is resolved, the reader cannot assess whether the monotone profile is statistically meaningful or within estimation noise. A bootstrap or a sensitivity analysis over the QFR/MLE estimation noise should be reported for at least the VPI profile, not just for a single marginal discrepancy.

minor comments (5)

[§I] The phrase 'would require the net-demand component (z_t, z_{t+1}, ..., z_{t+h}), which has M^{h+1} possible values' is correct but could be made clearer by explicitly counting the h+1 coordinates.
[§II.B] In Eq. (5), D_t(z) is described as the representative net demand in state z; it would help to state explicitly how D_t(z) is recovered from the QFR quantiles (e.g., midpoint or conditional mean), since the cost coefficients are normalized and the absolute numbers depend on this choice.
[§III.A] The information sets in Eq. (9) use a slight abuse of sigma-algebra notation; this is understandable but should be flagged in a sentence.
[§V] The variable count 'N L_R M^2 |A| ≈ 5.9e6' is consistent with the stated parameters, but it assumes |A|=3 for all states; boundary states have fewer actions. A short note about this would avoid confusion.
[§VI.C] Figure 2 is informative but the caption is long and the 'white cells are state-hour pairs outside that support' wording is slightly confusing: it would be clearer to say 'outside the support of the optimal policy on this day' rather than 'outside that support'.

Circularity Check

0 steps flagged

No significant circularity: the endpoint VPI profile is an emergent LP optimum, not an input or a fitted parameter.

full rationale

The value of perfect endpoint information is computed by solving the forecast-augmented LP (14) with fixed cost parameters, ramp constraints, and transition kernels (11) estimated by maximum likelihood from paired state sequences. The reported costs in Table II and the ordering (20) are optimal objective values produced by the solver; no parameter is calibrated to reproduce that ordering, and no equation defines VPI_end(h) in terms of the kernel coefficients in a way that would force the decreasing profile. The h=1 deterministic support condition j=i' is part of the definition of a one-hour perfect endpoint, not a fitted shortcut. Proposition 1 is a genuine feasible-policy argument requiring the marginal-consistency condition (13), which the paper states and checks. The self-citations [8]–[10],[15] supply the QFR baseline and related applications, but the endpoint augmentation, the LPs, and the numerical comparison are performed in this paper. Any concern that separately fitted endpoint kernels describe slightly different stochastic processes is a modeling-validity question, not a circularity: it does not make the output equal to the input by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper is a case study built on a prior QFR-MDP model; the main free parameters are the fitted transition kernels and hand-chosen cost/state-space parameters. No new physical entities are introduced.

free parameters (6)

cost coefficients c_r and c_d = 0.1 and 3
Normalized objective weights chosen by hand (Table I); the percentage savings may depend on their ratio.
number of net-demand states M = 4 (quantile levels 0.25, 0.5, 0.75)
Discretization choice from [8]; affects the state space granularity.
number of firming levels L_R = 14
Thermal stack discretization (Table I).
QFR harmonics r = 2 (d=5 basis terms)
Following [8]; trade-off between fit and parameter count.
offshore wind capacity scaling = 21,687 MW
Scenario capacity chosen for the case study (Section V).
joint endpoint transition kernel coefficients (alpha, beta^c, beta^s) for each h = estimated by ML
Fitted to the ISO-NE paired state sequence (Sec. III-B); the central VPI numbers are computed from these fitted kernels.

axioms (4)

domain assumption The QFR-discretized net-demand process is a cyclostationary Markov chain of order 1
Baseline model from [8]; used throughout, e.g., Eq. (3). If the real process has memory beyond the QFR state, the baseline and forecast-augmented LPs are misspecified.
domain assumption The joint endpoint process (z_t, z_{t+h}) is Markovian with transition kernel (11)
By construction the fitted kernel defines a Markov chain, but the true dynamic may not be first-order in this pair; the Fourier form is a parametric approximation.
domain assumption The marginal-consistency condition (13) approximately holds for the fitted kernels
Proposition 1 requires exact equality; the authors verify max |discrepancy| = 0.0199, so the theoretical nonnegativity is approximate in the empirical study.
standard math Standard optimization duality and LP strong duality hold
Used to solve (6) and (14) with Gurobi.

pith-pipeline@v1.3.0-alltime-deepseek · 7869 in / 11831 out tokens · 99756 ms · 2026-08-02T04:36:26.977500+00:00 · methodology

0 comments

read the original abstract

Forecast value depends not only on accuracy but also on the information structure available to the operating model. We study a diagnostic current-endpoint-only forecast for offshore-wind thermal firming: at hour t, the controller observes the current net-demand state z_t and one perfect future target z_{t+h}, but not the intermediate path or earlier endpoint messages. Unlike rolling path forecasts, these endpoint information sets are not nested in h. We embed the signal in a cyclostationary MDP using quantile Fourier regression states estimated from ISO New England load and offshore wind data, and solve annual state-action-frequency LPs for h = 1, . . . , 6. A one-hour endpoint forecast reduces annual firming cost by 8.07%, while a six-hour endpoint reduces it by 2.28%. The decreasing profile shows that a single farther endpoint is less actionable for a one-step ramping decision, without implying that longer rolling forecasts are less valuable.

Figures

Figures reproduced from arXiv: 2607.13627 by Arash Khojaste, Geoffrey Pritchard, Golbon Zakeri.

**Figure 2.** Figure 2: Optimal ramping policies on July 15. The top panel shows the no-forecast baseline, which conditions only on the firming level and current net-demand [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

15 extracted references

[1]

Economic valuation of reserves in power systems with high penetration of wind power,

J. M. Morales, A. J. Conejo, and J. P ´erez-Ruiz, “Economic valuation of reserves in power systems with high penetration of wind power,”IEEE Transactions on Power Systems, vol. 24, no. 2, pp. 900–910, 2009

2009
[2]

Dynamic scheduling of operating reserves in co-optimized electricity markets with wind power,

Z. Zhou and A. Botterud, “Dynamic scheduling of operating reserves in co-optimized electricity markets with wind power,”IEEE Transactions on Power Systems, vol. 29, no. 1, pp. 160–171, 2014

2014
[3]

Commitment and dispatch with un- certain wind generation by dynamic programming,

J. J. Hargreaves and B. F. Hobbs, “Commitment and dispatch with un- certain wind generation by dynamic programming,”IEEE Transactions on Sustainable Energy, vol. 3, no. 4, pp. 724–734, 2012

2012
[4]

J. R. Birge and F. Louveaux,Introduction to Stochastic Programming, 2nd ed. New York, NY , USA: Springer, 2011

2011
[5]

A. J. Conejo, M. Carri ´on, and J. M. Morales,Decision Making Under Uncertainty in Electricity Markets. Springer, 2010

2010
[6]

M. L. Puterman,Markov Decision Processes: Discrete Stochastic Dy- namic Programming. Wiley, 1994

1994
[7]

A unified framework for stochastic optimization,

W. B. Powell, “A unified framework for stochastic optimization,”Eu- ropean Journal of Operational Research, vol. 275, no. 3, pp. 795–821, 2019

2019
[8]

Quantile fourier regressions for decision making under uncertainty,

A. Khojaste, G. Pritchard, and G. Zakeri, “Quantile fourier regressions for decision making under uncertainty,”Ann. Oper. Res., vol. 355, no. 3, pp. 2843–2858, Dec. 2025

2025
[9]

Electricity price-aware scheduling of data center cooling,

A. Khojaste, J. Pearce, G. Zakeri, and Y . Sang, “Electricity price-aware scheduling of data center cooling,” in2025 57th North American Power Symposium (NAPS). IEEE, Oct. 2025, pp. 1–6

2025
[10]

Baseline hy- dropower generation offer curves,

J. Pearce, A. Khojaste, G. Zakeri, and G. Pritchard, “Baseline hy- dropower generation offer curves,” in2025 57th North American Power Symposium (NAPS). IEEE, Oct. 2025, pp. 1–6

2025
[11]

Koenker,Quantile Regression

R. Koenker,Quantile Regression. Cambridge University Press, 2005

2005
[12]

Form No. 714: Annual Elec- tric Balancing Authority Area and Planning Area Report,

Federal Energy Regulatory Commission, “Form No. 714: Annual Elec- tric Balancing Authority Area and Planning Area Report,” ISO New England hourly demand, 2006–2020, 2021

2006
[13]

National Data Buoy Center: Station 44025 (New York Bight),

National Oceanic and Atmospheric Administration, “National Data Buoy Center: Station 44025 (New York Bight),” NOAA NDBC, 2021

2021
[14]

Definition of the IEA 15-megawatt offshore reference wind turbine,

E. Gaertner, J. Rinkeret al., “Definition of the IEA 15-megawatt offshore reference wind turbine,” National Renewable Energy Laboratory, Tech. Rep. NREL/TP-5000-75698, 2020

2020
[15]

Risk-averse markov decision processes: Applications to electricity grid and reservoir management,

A. Khojaste, J. Pearce, D. P. de Farias, G. Pritchard, and G. Zakeri, “Risk-averse markov decision processes: Applications to electricity grid and reservoir management,” arXiv preprint arXiv:2601.02207, Jan. 2026, preprint

arXiv 2026

[1] [1]

Economic valuation of reserves in power systems with high penetration of wind power,

J. M. Morales, A. J. Conejo, and J. P ´erez-Ruiz, “Economic valuation of reserves in power systems with high penetration of wind power,”IEEE Transactions on Power Systems, vol. 24, no. 2, pp. 900–910, 2009

2009

[2] [2]

Dynamic scheduling of operating reserves in co-optimized electricity markets with wind power,

Z. Zhou and A. Botterud, “Dynamic scheduling of operating reserves in co-optimized electricity markets with wind power,”IEEE Transactions on Power Systems, vol. 29, no. 1, pp. 160–171, 2014

2014

[3] [3]

Commitment and dispatch with un- certain wind generation by dynamic programming,

J. J. Hargreaves and B. F. Hobbs, “Commitment and dispatch with un- certain wind generation by dynamic programming,”IEEE Transactions on Sustainable Energy, vol. 3, no. 4, pp. 724–734, 2012

2012

[4] [4]

J. R. Birge and F. Louveaux,Introduction to Stochastic Programming, 2nd ed. New York, NY , USA: Springer, 2011

2011

[5] [5]

A. J. Conejo, M. Carri ´on, and J. M. Morales,Decision Making Under Uncertainty in Electricity Markets. Springer, 2010

2010

[6] [6]

M. L. Puterman,Markov Decision Processes: Discrete Stochastic Dy- namic Programming. Wiley, 1994

1994

[7] [7]

A unified framework for stochastic optimization,

W. B. Powell, “A unified framework for stochastic optimization,”Eu- ropean Journal of Operational Research, vol. 275, no. 3, pp. 795–821, 2019

2019

[8] [8]

Quantile fourier regressions for decision making under uncertainty,

A. Khojaste, G. Pritchard, and G. Zakeri, “Quantile fourier regressions for decision making under uncertainty,”Ann. Oper. Res., vol. 355, no. 3, pp. 2843–2858, Dec. 2025

2025

[9] [9]

Electricity price-aware scheduling of data center cooling,

A. Khojaste, J. Pearce, G. Zakeri, and Y . Sang, “Electricity price-aware scheduling of data center cooling,” in2025 57th North American Power Symposium (NAPS). IEEE, Oct. 2025, pp. 1–6

2025

[10] [10]

Baseline hy- dropower generation offer curves,

J. Pearce, A. Khojaste, G. Zakeri, and G. Pritchard, “Baseline hy- dropower generation offer curves,” in2025 57th North American Power Symposium (NAPS). IEEE, Oct. 2025, pp. 1–6

2025

[11] [11]

Koenker,Quantile Regression

R. Koenker,Quantile Regression. Cambridge University Press, 2005

2005

[12] [12]

Form No. 714: Annual Elec- tric Balancing Authority Area and Planning Area Report,

Federal Energy Regulatory Commission, “Form No. 714: Annual Elec- tric Balancing Authority Area and Planning Area Report,” ISO New England hourly demand, 2006–2020, 2021

2006

[13] [13]

National Data Buoy Center: Station 44025 (New York Bight),

National Oceanic and Atmospheric Administration, “National Data Buoy Center: Station 44025 (New York Bight),” NOAA NDBC, 2021

2021

[14] [14]

Definition of the IEA 15-megawatt offshore reference wind turbine,

E. Gaertner, J. Rinkeret al., “Definition of the IEA 15-megawatt offshore reference wind turbine,” National Renewable Energy Laboratory, Tech. Rep. NREL/TP-5000-75698, 2020

2020

[15] [15]

Risk-averse markov decision processes: Applications to electricity grid and reservoir management,

A. Khojaste, J. Pearce, D. P. de Farias, G. Pritchard, and G. Zakeri, “Risk-averse markov decision processes: Applications to electricity grid and reservoir management,” arXiv preprint arXiv:2601.02207, Jan. 2026, preprint

arXiv 2026