LiFT-MPC: Language-in-the-Loop Feedback Tuning of Cost Previews for MPC

Ioannis Lestas; Xinyi Yi

REVIEW 4 major objections 6 minor 16 references

Language context can correct MPC cost previews online from control performance, with a proven gap to the best fixed correction and large gains on battery arbitrage.

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 · grok-4.5

2026-07-30 10:58 UTC pith:FF4J6EHD

load-bearing objection Clean residual language-correction for cost-preview MPC with a real ESS study; the delay-term proof needs a standard fix but the claim is probably fine. the 4 major comments →

arxiv 2607.23832 v1 pith:FF4J6EHD submitted 2026-07-26 math.OC cs.SYeess.SYstat.AP

LiFT-MPC: Language-in-the-Loop Feedback Tuning of Cost Previews for MPC

Xinyi Yi , Ioannis Lestas This is my paper

classification math.OC cs.SYeess.SYstat.AP MSC 93B4549N1090C90

keywords model predictive controlpreview refinementlanguage-in-the-looponline gradient descentenergy storageelectricity pricesresidual correctiontime-varying objectives

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.

When model predictive control optimizes a cost that depends on future signals such as electricity prices, those signals are hard to forecast from numbers alone because they also depend on events described in text. This paper proposes LiFT-MPC: keep a standard numerical preview, add a language-informed residual correction built from a library of historical residual shapes, and retune that correction inside the control loop using a delayed loss that measures how much preview error hurt the actual control cost. The authors prove that the closed-loop cost stays within a bound of the best fixed correction in their class, separating ordinary online-learning terms, delay, and the gap between the true and proxy losses. On a realistic battery energy-storage task with real New South Wales prices and day-ahead news, the method multiplies arbitrage reward relative to the numerical baseline, both with strong and weak forecasters. A sympathetic reader cares because it turns messy contextual text into a control-relevant, online-tunable correction with guarantees rather than a one-shot forecast fix.

Core claim

LiFT-MPC shows that residual corrections of cost-signal previews, formed as scenario-weighted historical trajectories and updated by delayed gradient steps on a control-performance proxy loss, yield a closed-loop cost within an explicit bound of the best fixed refinement in that structure, and that this improves economic performance of preview-based MPC on real price-and-news energy storage operation.

What carries the argument

The LiFT residual correction plus delayed performance update: the preview is baseline plus a mixture of fixed historical residual trajectories reweighted by a parameter θ updated with delayed gradient descent on the proxy per-round loss L_i; Theorem 1 bounds J_MPC(θ online) − J_LiFT-OPT by standard online-gradient, delay, and surrogate-mismatch terms.

Load-bearing premise

Day context is fixed before the horizon starts, and every useful correction must be a reweighting of a finite library of past residual trajectory shapes—if a new day is outside that library, the method’s benchmark and gains can fail.

What would settle it

On held-out news days whose price residuals are poorly spanned by the 2018–2022 scenario library, check whether online LiFT still beats Base-MPC on realized arbitrage reward and whether the measured gap tracks Theorem 1’s bound; collapse of both would falsify the central claim in practice.

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

If this is right

Preview mismatch in economic MPC can be treated as an online residual-tuning problem driven by realized control loss, not only by forecast error.
Day-ahead text can be wired into MPC as additive cost-preview corrections with an explicit closed-loop performance certificate relative to the best fixed mixture.
Stronger numerical baselines leave residuals that language correction can exploit more effectively than weak baselines alone.
Practical tracking of the offline optimal trajectory remains a tube whose radius scales with feedforward preview mismatch.
Within-day adaptation carries most of the economic gain; carrying parameters across days is secondary in the reported setting.

Where Pith is reading between the lines

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

The same residual-plus-delayed-control-loss pattern could apply to other exogenous cost drivers carried in text (fuel, congestion, weather alerts) without changing the MPC plant model.
If the scenario library were refreshed online when residuals persistently miss the span of prototypes, the method might extend beyond fixed day-ahead context.
Clipping of power and SoC in deployment is outside the affine-law theory; quantifying that gap would be the natural next certificate.
Zero-shot mixture weights already help, so cheap offline language indexing of residual shapes may be valuable even where online gradients are restricted.

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

Clean residual language-correction for cost-preview MPC with a real ESS study; the delay-term proof needs a standard fix but the claim is probably fine.

read the letter

Punchline: this is a solid CDC methods paper that puts text into economic MPC the right way—residual correction on cost previews, not a black-box LLM controller—and backs it with a multi-year NSW price+news ESS study. The distinctive pieces are real: keep a numerical baseline, correct with a finite scenario mixture from news, and update θ online on a delayed control-gap proxy with an explicit regret-style bound relative to best fixed θ in that class.

What works. Lemma 1’s gap identity and the proxy construction from realized-interval backward recursion are clean LQ dynamic-programming algebra. Treating language as an additive residual on π̂ rather than regenerating disturbances is the right architectural choice for cost-entering prices, and it makes the role of context explicit. The experiments are the other half of the value: Chronos and AR(48) baselines, zero-shot vs online, reset vs carry, and large reward lifts on raw-news days (Table II, Fig. 1). That is more than a toy demo.

Soft spots, in proportion. The stress-test on Appendix C is fair: Eq. (38) writes a cross-round gradient difference as if it were pure parameter drift, but L_t and L_i are different losses. A standard delayed-feedback pairing (gradient at i applied at i+k) recovers the same O(√T)+delay order, so I treat this as a derivation gap at the paper’s central term, not a false theorem. Still, referees will want that rewrite. Minor but material: theory is unclipped affine LQ while the ESS clips power/SoC; Assumption 3 plus a fixed historical residual library means novel residual shapes are outside the guarantee and the empirical story; no code/data release for a methods paper whose headline includes economic multiples. Free parameters (k, |S|, temperature, η) are normal for this venue if sensitivity is shown.

Who it’s for: people doing preview/economic MPC and BESS operation who want a control-in-the-loop path for text. Citation pattern is appropriate (InstructMPC, price forecasting, residual correction). Math structure is serious; empirics directionally strong but not uncertainty-quantified.

I’d send it to peer review. Engage if you work on economic MPC or language-informed control; skim the ESS tables and Theorem 1 if you only need the idea.

Referee Report

4 major / 6 minor

Summary. The paper proposes LiFT-MPC, a receding-horizon MPC scheme for linear systems with time-varying economic stage costs (ℓ_t = x⊤Qx + u⊤Ru + 2απ π_t⊤u) in which a baseline numerical preview of the exogenous signal is additively corrected by a language-informed residual term. The correction is a weighted mixture of fixed historical residual-trajectory prototypes, with weights refined online by a delayed gradient update on a proxy control-performance loss. The main results are: Lemma 1, an exact identity expressing J_MPC − J_Off-OPT as a sum of per-round quadratic losses ψ_t⊤Hψ_t; Lemma 2, an O(ρ^{k−1}) bound on the gradient discrepancy between the exact per-round loss and its online-computable delayed proxy; Theorem 1, a bound on J_MPC(θ_{0:T−1}) − J_LiFT-OPT decomposing into a standard OGD term, a delay term, and a surrogate-mismatch term, with the first two terms O(√T) under η_t = D/(G√(t+1)); and Theorem 2, an ISS-type practical tracking bound. Experiments on a battery-storage arbitrage problem with NSW (Australia) prices and day-ahead news (2015–2024) report >4× reward improvement over Base-MPC with a Chronos-T5 baseline and >74% improvement with an AR(48) baseline.

Significance. If the claims hold, the paper contributes (i) a clean reduction of preview-induced suboptimality to a sum of per-round quadratic losses (Lemma 1), enabling control-performance-driven online tuning of a predictor inside the MPC loop; (ii) an explicit decomposition of the online guarantee into standard OGD, delay, and surrogate-mismatch terms, which is falsifiable and interpretable; and (iii) a realistic evaluation on 2015–2024 NSW price and news data with ablations over baseline preview quality (Chronos-T5 vs AR(48)), showing large and consistent economic gains (>4× arbitrage reward vs Base-MPC on Chronos; >74% on AR(48)). The residual-correction structure (keeping the numerical preview nominal and adding a context-driven correction) is a practically appealing design choice that makes the role of language context explicit and auditable. The scenario-library construction and semantic weighting are described in enough detail (Appendices E–F, named encoder, clustering and temperature-softmax pipeline) to permit reproduction. These are genuine strengths; the delayed-feedback analysis is the paper's load-bearing theoretical component and must be watertight.

major comments (4)

[Appendix C, Eq. (38)] The bound on Σ_t (r_t − r_i)ᵀ(θ_t − θ*) is not justified as written, and the invoked analogy to Eq. (12) of [11] does not apply. Here r_t := ∇L_t(θ_t) and r_i := ∇L_i(θ_i) with i = t−k are gradients of *different* losses: L_t is the proxy built from the interval realized at round t, while L_i is built from the interval realized k rounds earlier. The realized signals π and s^real over these two intervals differ generically, so ‖r_t − r_i‖ is O(1) per round under Assumption 5 alone, and the sum is only bounded by ~2GDT = O(T). With the stated stepsize η_t = D/(G√(t+1)) this would render the published bound vacuous rather than O(√T). The (k−1)GD constant in (38) appears to account only for the first k rounds (where r_i ≡ 0), suggesting a drift-only argument was applied to a term that is not drift. The result itself is very likely correct at the stated order, via the standard delayed-feedbac
[Theorem 1, Eq. (29)] The bound (29) is presented with emphasis on the O(√T) behavior of the first two terms under η_t = D/(G√(t+1)), but the surrogate-mismatch term D·C_LD Σ_t ρ^{T(t)−t} is Θ(T·ρ^{k−1}) for interior rounds: it grows linearly in T. The per-round gap therefore does not vanish; it converges to a floor of order D·C_LD·ρ^{k−1}. This is inherent to the proxy construction (Lemma 2) and not an error, but it should be stated explicitly, and the trade-off in k deserves discussion: increasing k shrinks the mismatch floor geometrically while inflating the delay term linearly. A brief remark giving the per-round asymptotic bound and any guidance on selecting k would make the guarantee's content transparent.
[§IV-A, footnote 6; relation between §III theory and Table II] The theory (Lemma 1, Theorems 1–2) is developed for the unclipped affine law (12), while the ESS experiments enforce hard limits u_t ∈ [−Pmax, Pmax] and SoC ∈ [0, Emax] by clipping (footnote 6). Clipping destroys the affine structure on which the optimality-gap identity (15), the affinity of ψ_t in θ (Assumption 4 usage), and the convexity of F_t all rely. The authors acknowledge this and defer analysis to future work, which is acceptable, but the connection between Theorem 1 and the Table II claims then rests on clipping being a mild perturbation. Since §IV-B.2 reports that power and SoC limits are in fact reached in closed loop, please quantify this: e.g., report the fraction of clipped steps per method and, if feasible, a small-scale comparison of the clipped closed loop against the unclipped affine prediction of J_MPC, so the reader can gauge how far the experiment sits from the anal
[§III-A, Eq. (22); §IV, scenario library construction] The residual correction (22) is confined to the span of a finite library of fixed historical prototypes {h^s} built from 2018–2022 data. The theory is internally consistent (the benchmark LiFT-OPT is the best fixed parameter within the same class), but the empirical gains depend on 2023–2024 residual patterns being well represented by this library; a novel outage or weather regime outside the span would degrade both zero-shot and online variants. Please add a diagnostic supporting robustness on the evaluation period: e.g., the distribution of the projection error of realized daily residuals onto span{h^s} over 2023–2024, or per-day performance dispersion identifying the worst-case days and whether they coincide with poorly spanned residuals. This would substantiate the 'robust benefit' claim beyond the aggregate numbers in Table II and Fig. 1.

minor comments (6)

[§III-A.2, Eq. (24)] The update (24) is unconstrained gradient descent, yet Assumption 5 requires θ_t ∈ Θ for all t with Θ compact. Please add the Euclidean projection onto Θ in (24) (the telescoping bound (37) is unaffected, since projection is nonexpansive), and state the initialization θ_0 explicitly.
[Appendix F, Eq. (40)] Symbol clash: T denotes the horizon length throughout, but T > 0 is reused as the softmax temperature in Eq. (40). Please rename the temperature (e.g., τ or β).
[Table II] Table II reports only aggregate rewards over 2023–2024. Please add day-level dispersion (standard deviation or quantiles across the 420 news days / all days) so the '>4.6×' improvement can be assessed for consistency rather than being driven by a few high-price days.
[Various] Typographical/grammatical: 'prediciton' (end of §I); abstract 'predicted signals need to be often incorporated' → 'often need to be incorporated'; §III-A the sentence ending '...extreme weather conditions in power systems)' is missing a period before 'This is quantified in equation (22)'; §II-B 'needs to be often incorporated' phrasing also appears in the abstract.
[Figs. 2–3 and §IV-B.4] In Figs. 2–3, the caption states that the reported mean/median are computed from day-level differences while the displayed MAE/RMSE are averages of day-level values; this dual reporting is confusing on first read. Please either align the statistics or add one clarifying sentence in the main text where the figures are discussed.
[§II-D.2 / Remark 2] Remark 2 and the surrounding text would benefit from one sentence making the information structure explicit: the loss L_i incurred at round i only becomes computable at round i+k, which is precisely the source of the delay in (24). This is currently implicit and would help readers parse Appendix C.

Circularity Check

0 steps flagged

No significant circularity: regret bound is standard OGD-vs-best-fixed-comparator; scenario library and evaluation use a proper temporal split.

full rationale

The load-bearing theoretical claim (Theorem 1) bounds realized closed-loop cost of the online parameter sequence against LiFT-OPT, defined as the best fixed refinement parameter in the same parametric class (Eq. 28). That is ordinary online-gradient regret, not a self-definitional loop: the comparator is openly restricted to fixed θ in Θ, and the bound separates standard OGD, delay, and surrogate-mismatch terms derived from Lemmas 1–2 and Assumptions 1–5. Lemma 1’s exact gap and Definition 2’s delayed proxy are obtained from dynamic programming / completing-the-square identities, not by fitting the target cost into the update. Scenario prototypes h^s and semantic features are built on 2018–2022 residuals and news and evaluated on 2023–2024 (Appendix E–F, §IV), a proper hold-out. Citation [11] for the delayed-gradient estimate is by different authors. Empirical gains (Table II, Fig. 1) are out-of-sample closed-loop rewards under real prices, not refits of the evaluation metric. No step reduces a claimed prediction or first-principles result to its own inputs by construction. (A separate correctness concern about the delay-term proof does not constitute circularity.)

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 2 invented entities

The central guarantee rests on classical LQ/MPC structure (Riccati, stable Φ), bounded prices, a fixed finite scenario library with affine reweighting in θ, compact Θ with bounded loss gradients, and delayed access to a proxy gap that approximates the true DP gap when the preview tail is long. Empirical claims further depend on hand-built residual clusters, embedding similarities, and several controller/forecast hyperparameters. No new physical entities; the invented objects are methodological (LiFT residual mixture, proxy loss L_i).

free parameters (6)

Preview length k = 8
Set to k=8 (4 hours) in experiments; enters delay term (k−1) and surrogate decay ρ^{k−1} in Theorem 1.
OGD stepsizes η_t and gradient bound G / diameter D
Theory uses η_t=D/(G√(t+1)); practical η, and realized G,D for the ESS loss are not fully specified in the main text.
Scenario-library size |S| and K-means residual prototypes h^s
Historical residuals 2018–2022 are clustered and rescaled; number of clusters and median-norm rescaling are design choices that define the span of all corrections.
Softmax temperature T in initial weights w(c)
Controls how peaky news-to-scenario weights are before online refinement (Appendix F).
ESS cost weights Q, R, α_π and hardware limits = Q=1e-4, R=1e-3, α_π=2
Q=1e-4, R=1e-3, α_π=2, Emax=200 MWh, Pmax=100 MW shape both the economic objective and closed-loop behavior.
Refinement map parameterization θ (matrix multiplying w) = θ_0 = I
Experiments take w_θ=θw with θ initialized at I; dimension and any projection onto Θ are implementation choices under Assumption 4–5.

axioms (7)

standard math Assumption 1: (A,B) controllable and (Q^{1/2},A) detectable so a stabilizing Riccati P and Φ=A−BK exist.
Standard LQR prerequisite for the affine law and exponential decay of Φ used throughout §§II–III.
domain assumption Assumption 2: exogenous signal π_t is componentwise bounded for all t.
Used to bound s-differences and surrogate mismatch; reasonable for prices with caps but idealized for spikes.
domain assumption Assumption 3: contextual information is collected before the horizon and held fixed over [T].
Matches day-ahead news for intra-day ESS, but rules out intra-day breaking news; load-bearing for ‘best fixed θ’ benchmark.
ad hoc to paper Assumption 4: θ ↦ f(θ,w) is differentiable and affine in θ, so ψ_t(θ) is affine and F_t convex.
Enables convex OGD analysis; experiments specialize to w_θ=θw.
standard math Assumption 5: Θ convex compact, diameter ≤ D, and ||∇L_t||≤G uniformly.
Standard online convex optimization regularity for the regret bound.
ad hoc to paper Residual correction lies in the span of fixed historical trajectories {h^s} (Eq. 22).
Defines the hypothesis class of LiFT-OPT; not implied by physics, only by the chosen library.
domain assumption Linear plant and quadratic stage cost with linear price–input coupling (Eqs. 1–2).
Makes exact DP affine structure available; ESS model is a shifted integrator approximating SoC.

invented entities (2)

LiFT residual scenario-mixture correction Δπ(θ) no independent evidence
purpose: Map fixed news context plus tunable θ into an additive preview correction without altering the baseline forecaster.
Methodological construct: weighted sum of historical residual prototypes with language-initialized, performance-tuned weights.
Proxy per-round loss L_i from realized-interval backward recursion s^{real} independent evidence
purpose: Online-computable surrogate of the true preview-induced optimality gap F_i that needs full-horizon s⋆.
Defined in Def. 2 / Eq. 23; Lemma 2 bounds gradient mismatch by O(ρ^{k−1}).

pith-pipeline@v1.2.0-grok45-kimik3 · 19814 in / 4178 out tokens · 88925 ms · 2026-07-30T10:58:39.407807+00:00 · methodology

0 comments

read the original abstract

In model predictive control (MPC) with time-varying objectives, predicted signals need to be often incorporated in the cost function, such as prices in energy system operation. These are, however, often difficult to predict from the historical trajectory of these signals alone, as they may depend on other contextual events. We propose LiFT-MPC, an MPC framework that integrates a LiFT (Language-in-the-Loop Feedback Tuning) correction scheme to refine such predictions within the MPC loop. The prediction mechanism is updated online via a control-performance loss function, and we establish a performance guarantee for the resulting closed loop system. Numerical experiments using a realistic example of energy-storage management with real prices and news context to improve predictions, demonstrate an improved economic performance

Figures

Figures reproduced from arXiv: 2607.23832 by Ioannis Lestas, Xinyi Yi.

**Figure 2.** Figure 2: Histograms of day-level forecasting error differences between Zero [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Histograms of day-level forecasting error differences between Zero [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 2 linked inside Pith

[1]

An overview of systems-theoretic guaran- tees in data-driven model predictive control,

J. Berberich and F. Allg ¨ower, “An overview of systems-theoretic guaran- tees in data-driven model predictive control,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 8, no. 1, pp. 77–100, 2025

2025
[2]

Economic model predictive control for time-varying system: Performance and stability results,

L. Gr ¨une and S. Pirkelmann, “Economic model predictive control for time-varying system: Performance and stability results,”Optimal Control Applications and Methods, vol. 41, no. 1, pp. 42–64, 2020

2020
[3]

Forecasting day- ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,

J. Lago, G. Marcjasz, B. De Schutter, and R. Weron, “Forecasting day- ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,”Applied Energy, vol. 293, p. 116983, 2021

2021
[4]

Chronos: Learning the language of time series,

A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. P. Arango, S. Kapoor, J. Zschiegner, D. C. Maddix, H. Wang, M. W. Mahoney, K. Torkkola, A. G. Wilson, M. Bohlke-Schneider, and Y . Wang, “Chronos: Learning the language of time series,”arXiv preprint, 2024

2024
[5]

Lag-llama: Towards foundation models for probabilistic time series forecasting,

K. Rasul, A. Ashok, A. R. Williams, H. Ghonia, R. Bhagwatkar, A. Khorasani, M. J. D. Bayazi, G. Adamopoulos, R. Riachi, N. Hassen et al., “Lag-llama: Towards foundation models for probabilistic time series forecasting,”arXiv preprint arXiv:2310.08278, 2023

Pith/arXiv arXiv 2023
[6]

Context matters: Leveraging contextual features for time series forecasting,

S. Chattopadhyay, P. Paliwal, S. S. Narasimhan, S. Agarwal, and S. P. Chinchali, “Context matters: Leveraging contextual features for time series forecasting,”arXiv preprint arXiv:2410.12672, 2024

Pith/arXiv arXiv 2024
[7]

A hybrid system based on ensemble learning to model residuals for time series forecasting,

D. S. d. O. S. J ´unior, P. S. de Mattos Neto, J. F. de Oliveira, and G. D. Cavalcanti, “A hybrid system based on ensemble learning to model residuals for time series forecasting,”Information Sciences, vol. 649, p. 119614, 2023

2023
[8]

A machine-learning-based event- triggered model predictive control for building energy management,

S. Yang, W. Chen, and M. P. Wan, “A machine-learning-based event- triggered model predictive control for building energy management,” Building and Environment, vol. 233, p. 110101, 2023

2023
[9]

Efficient context-aware model predictive control for human-aware navigation,

E. Stefanini, L. Palmieri, A. Rudenko, T. Hielscher, T. Linder, and L. Pallottino, “Efficient context-aware model predictive control for human-aware navigation,”IEEE Robotics and Automation Letters, vol. 9, no. 11, pp. 9494–9501, 2024

2024
[10]

Large language model-augmented model predictive control for marine navigation,

T. Zenget al., “Large language model-augmented model predictive control for marine navigation,”Ocean Engineering, 2026

2026
[11]

Instructmpc: a human-llm-in-the-loop frame- work for context-aware control,

R. Wu, J. Ai, and T. Li, “Instructmpc: a human-llm-in-the-loop frame- work for context-aware control,” in2025 IEEE 64th Conference on Decision and Control (CDC). IEEE, 2025, pp. 172–179

2025
[12]

Scenario-based stochastic optimization for energy and flexibility dispatch of a microgrid,

K. Antoniadou-Plytaria, D. Steen, O. Carlson, B. Mohandes, M. A. F. Ghazviniet al., “Scenario-based stochastic optimization for energy and flexibility dispatch of a microgrid,”IEEE Transactions on Smart Grid, vol. 13, no. 5, pp. 3328–3341, 2022

2022
[13]

Bidding strategy for microgrid in day-ahead market based on hybrid stochastic/robust optimization,

G. Liu, Y . Xu, and K. Tomsovic, “Bidding strategy for microgrid in day-ahead market based on hybrid stochastic/robust optimization,”IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 227–237, 2015

2015
[14]

Nsw-epnews: A news- augmented benchmark for electricity price forecasting with llms,

Z. Bi, L. Huang, H. Jin, Q. Zeng, and H. Chen, “Nsw-epnews: A news- augmented benchmark for electricity price forecasting with llms,”arXiv preprint, 2025

2025
[15]

Sentence-bert: Sentence embeddings using siamese bert-networks,

N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP- IJCNLP), 2019, pp. 3982–3992

2019
[16]

Some methods of classification and analysis of mul- tivariate observations,

J. B. McQueen, “Some methods of classification and analysis of mul- tivariate observations,” inProc. of 5th Berkeley Symposium on Math. Stat. and Prob., 1967, pp. 281–297

1967

[1] [1]

An overview of systems-theoretic guaran- tees in data-driven model predictive control,

J. Berberich and F. Allg ¨ower, “An overview of systems-theoretic guaran- tees in data-driven model predictive control,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 8, no. 1, pp. 77–100, 2025

2025

[2] [2]

Economic model predictive control for time-varying system: Performance and stability results,

L. Gr ¨une and S. Pirkelmann, “Economic model predictive control for time-varying system: Performance and stability results,”Optimal Control Applications and Methods, vol. 41, no. 1, pp. 42–64, 2020

2020

[3] [3]

Forecasting day- ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,

J. Lago, G. Marcjasz, B. De Schutter, and R. Weron, “Forecasting day- ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,”Applied Energy, vol. 293, p. 116983, 2021

2021

[4] [4]

Chronos: Learning the language of time series,

A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. P. Arango, S. Kapoor, J. Zschiegner, D. C. Maddix, H. Wang, M. W. Mahoney, K. Torkkola, A. G. Wilson, M. Bohlke-Schneider, and Y . Wang, “Chronos: Learning the language of time series,”arXiv preprint, 2024

2024

[5] [5]

Lag-llama: Towards foundation models for probabilistic time series forecasting,

K. Rasul, A. Ashok, A. R. Williams, H. Ghonia, R. Bhagwatkar, A. Khorasani, M. J. D. Bayazi, G. Adamopoulos, R. Riachi, N. Hassen et al., “Lag-llama: Towards foundation models for probabilistic time series forecasting,”arXiv preprint arXiv:2310.08278, 2023

Pith/arXiv arXiv 2023

[6] [6]

Context matters: Leveraging contextual features for time series forecasting,

S. Chattopadhyay, P. Paliwal, S. S. Narasimhan, S. Agarwal, and S. P. Chinchali, “Context matters: Leveraging contextual features for time series forecasting,”arXiv preprint arXiv:2410.12672, 2024

Pith/arXiv arXiv 2024

[7] [7]

A hybrid system based on ensemble learning to model residuals for time series forecasting,

D. S. d. O. S. J ´unior, P. S. de Mattos Neto, J. F. de Oliveira, and G. D. Cavalcanti, “A hybrid system based on ensemble learning to model residuals for time series forecasting,”Information Sciences, vol. 649, p. 119614, 2023

2023

[8] [8]

A machine-learning-based event- triggered model predictive control for building energy management,

S. Yang, W. Chen, and M. P. Wan, “A machine-learning-based event- triggered model predictive control for building energy management,” Building and Environment, vol. 233, p. 110101, 2023

2023

[9] [9]

Efficient context-aware model predictive control for human-aware navigation,

E. Stefanini, L. Palmieri, A. Rudenko, T. Hielscher, T. Linder, and L. Pallottino, “Efficient context-aware model predictive control for human-aware navigation,”IEEE Robotics and Automation Letters, vol. 9, no. 11, pp. 9494–9501, 2024

2024

[10] [10]

Large language model-augmented model predictive control for marine navigation,

T. Zenget al., “Large language model-augmented model predictive control for marine navigation,”Ocean Engineering, 2026

2026

[11] [11]

Instructmpc: a human-llm-in-the-loop frame- work for context-aware control,

R. Wu, J. Ai, and T. Li, “Instructmpc: a human-llm-in-the-loop frame- work for context-aware control,” in2025 IEEE 64th Conference on Decision and Control (CDC). IEEE, 2025, pp. 172–179

2025

[12] [12]

Scenario-based stochastic optimization for energy and flexibility dispatch of a microgrid,

K. Antoniadou-Plytaria, D. Steen, O. Carlson, B. Mohandes, M. A. F. Ghazviniet al., “Scenario-based stochastic optimization for energy and flexibility dispatch of a microgrid,”IEEE Transactions on Smart Grid, vol. 13, no. 5, pp. 3328–3341, 2022

2022

[13] [13]

Bidding strategy for microgrid in day-ahead market based on hybrid stochastic/robust optimization,

G. Liu, Y . Xu, and K. Tomsovic, “Bidding strategy for microgrid in day-ahead market based on hybrid stochastic/robust optimization,”IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 227–237, 2015

2015

[14] [14]

Nsw-epnews: A news- augmented benchmark for electricity price forecasting with llms,

Z. Bi, L. Huang, H. Jin, Q. Zeng, and H. Chen, “Nsw-epnews: A news- augmented benchmark for electricity price forecasting with llms,”arXiv preprint, 2025

2025

[15] [15]

Sentence-bert: Sentence embeddings using siamese bert-networks,

N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP- IJCNLP), 2019, pp. 3982–3992

2019

[16] [16]

Some methods of classification and analysis of mul- tivariate observations,

J. B. McQueen, “Some methods of classification and analysis of mul- tivariate observations,” inProc. of 5th Berkeley Symposium on Math. Stat. and Prob., 1967, pp. 281–297

1967