{"id":"726ad6c7-302e-4711-add3-5a622acfdb79","arxiv_id":"1908.08077","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An adapted hysteretic control policy for on-off loads achieves convergence, no chattering, and near-optimal power allocation in primary frequency regulation.","lead":"This paper designs a control scheme for on-off electrical loads that helps stabilize grid frequency without causing rapid switching (chattering) or endless oscillation (limit cycles). It guarantees the final power allocation is close to the cheapest possible, with a proven margin of error.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's Lyapunov proof fails in the marginal case pc=p^c_j: an admissible solution can keep σ=1 and make V increase, so convergence is not established as written.","rationale":"The paper's central claim is valuable and the NPCC simulations give plausible qualitative support. However, the analytical path to Theorem 4 has a load-bearing gap in the marginal-load case where pc equals a threshold p^c_j and two equilibria coexist. The proof fixes a single equilibrium x* and asserts via (36a) that V is nonincreasing. In the one-load case, starting at the σ=0 equilibrium with σ=1 is admissible, and the branch that keeps σ=1 moves away from x* while V increases; no single V anchored at one equilibrium can be nonincreasing for branches converging to the other equilibrium. This is not cosmetic: the marginal interval is exactly where Theorem 5's epsilon bound accounts for the discrete gap. The separate omission of α_j in (32) makes the appendix algebra unreliable, though it is likely repairable by taking V_M=Σ(τ_j/α_j)(pM_j-pM*_j)^2. I do not claim the main theorem is false; the simulations and structure suggest a corrected argument may exist. As written, the convergence guarantee is not established, which strengthens the reader's CONDITIONAL verdict rather than supporting ACCEPT. This differs from the reader's identified caveat about exact ℓ and α_j=c_j^{-1}, so my agreement assessment is disagree.","tokens_in":26411,"tokens_out":17447,"duration_ms":179211,"concrete_test":"Run a one-bus, one-load instance of (27) with Design Condition 2 and pc=p^c_1, initial continuous state equal to the σ=0 equilibrium and discrete state σ=1. Follow the admissible branch that keeps σ=1 and compute V from (33) relative to the σ=0 equilibrium. If V increases on any interval, the proof's claimed monotonicity (36a) and its invocation of [27, Cor. 8.7(b)] fail; Theorem 4 then needs a revised Lyapunov argument that handles the two coexisting equilibria.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Under Design Condition 2 with a single bus and one load, set pc=p^c_1. Then (27) has two equilibria: (ω^0_1, pM=-αω^0_1, σ=0) and (ω^0_1-\\bar d_1/D, pM=-α(ω^0_1-\\bar d_1/D), σ=1), both allowed by (24). Start at the first equilibrium's continuous state with σ=1. This is an admissible initial condition because (24) places σ=1 in {0,1} at ω=ω^0_1, pc=p^c_1. The branch that keeps σ=1 has \\dot ω=-\\bar d_1/M<0 at t=0, while pM is at its σ=0 equilibrium value. The Lyapunov function V of (33), taken relative to the σ=0 equilibrium, is 0 at t=0 and becomes positive for small t, so V is not nonincreasing. This contradicts the asserted bound in (36a): here (ω-ω*)(d^c-d^{c,*})<0 because σ=1>σ*=0 while ω<ω^0_1=ω*. Hence [27, Cor. 8.7(b)] cannot be invoked in the marginal interval where Theorem 5's epsilon bound is needed. A separate algebraic issue in eq. (32) also drops α_j; with V_M=1/2 τ(pM-pM*)^2 the derivative contains the indefinite term (1-α)(ω-ω*)(pM-pM*), so the stated Lyapunov inequality is not generally true for α_j≠1, the case required by Theorem 5.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies primary frequency regulation in power networks with on-off controllable loads. It first analyzes loads that switch at fixed frequency thresholds, proving convergence of Filippov solutions (Theorem 1) while noting that such schemes can chatter. It then introduces hysteretic switching (Section IV), proves existence of equilibria under a hysteresis-width condition (Theorem 2), and states absence of chattering (Proposition 1), but observes that limit cycles can still occur. The main proposal is an adapted hysteretic scheme (Section V and Section VI) that uses an aggregate-demand power command to prevent further switching for a subset of loads. Under Design Condition 1 the authors claim convergence and no chattering (Theorem 3, Proposition 3); under Design Condition 2 they claim convergence (Theorem 4) and, when the droop coefficients satisfy alpha_j = c_j^{-1}, an epsilon-optimal steady-state allocation for a mixed-integer optimization problem (Theorem 5), with epsilon = (1/(2D)) max_j (bar_d_j)^2. The analytical results are illustrated by simulations on the NPCC 140-bus system.","tokens_in":26779,"tokens_out":12628,"duration_ms":129625,"significance":"If the convergence and optimality results are correct, this is a valuable contribution to demand-side primary frequency control. The paper clearly identifies why naive on-off control chatters, shows how hysteresis removes chattering, and proposes a constructive design whose equilibria are provably close to optimal for an NP-hard mixed-integer allocation problem, with an explicit and plausibly small error bound. The numerical validation on a realistic 140-bus model is a definite strength. However, the current proof of the central Lyapunov estimates contains a parameter-dependent algebraic error, the convergence proof has a gap at the marginal power-command value, and several auxiliary proofs are deferred to a companion paper. These issues must be repaired before the main claims can be regarded as established.","major_comments":[{"comment":"The Lyapunov derivative for V_M is computed as if alpha_j = 1. From (2a), tau_j \\dot{p}^M_j = -(p^M_j - p^{M,*}_j) - alpha_j(omega_j - omega^*_j), so the derivative of V_M = (1/2) sum tau_j (p^M_j - p^{M,*}_j)^2 contains the cross term -alpha_j (p^M_j - p^{M,*}_j)(omega_j - omega^*_j), not -(p^M_j - p^{M,*}_j)(omega_j - omega^*_j) as written in (32). Consequently the inequalities in (34) and (36a) are not valid for general alpha_j, and Theorem 1, Theorem 3, and Theorem 4 are stated without the restriction alpha_j = 1. The argument can be repaired by taking V_M = (1/2) sum (tau_j/alpha_j)(p^M_j - p^{M,*}_j)^2, which cancels the cross term, but the proof as written is incorrect for the parameter range claimed.","section":"Appendix, proof of Theorem 1, Eq. (32); proof of Theorem 4, Eq. (36a)"},{"comment":"The sign assertion (omega_j - omega^*_j)(d^c_j - d^{c,*}_j) >= 0 fails in the marginal case p^c = p^c_j. For a single-bus, single-load network with Design Condition 2 and p^c_1 = D omega^0_1, take the continuous state x equal to the sigma = 0 equilibrium and set sigma = 1. This point is admissible under (24), but the dynamics give dot{omega} = -bar{d}_1/M < 0 at t = 0 while p^M is initially constant. The function V defined in (33) relative to the sigma = 0 equilibrium therefore increases for small t: (omega - omega^*)(d^c - d^{c,*}) is negative because omega < omega^* while d^c - d^{c,*} = bar{d}_1. Part (a) of the proof only fixes sigma for buses in N2, so the argument does not cover this case, which is exactly the case where the epsilon bound in Theorem 5 is active.","section":"Appendix, proof of Theorem 4, part (b), inequality (36a)"},{"comment":"The proofs of Propositions 1, 3, and 5 are not provided; they are deferred by analogy to Lemma 4 and Proposition 1 of [16]. Since [16] treats secondary frequency control with a different hybrid setup, and Proposition 5 supplies the completeness, finite-dwell-time, and no-chattering properties that are used in Theorem 4 and in the paper's main claims, the analogy is not sufficient. The authors should include self-contained proofs, or at least a precise transfer argument showing that the different flow and jump sets of (10), (18), and (27) do not affect the conclusion.","section":"Appendix, 'Proofs of Propositions 1, 3 and 5'"},{"comment":"The introduction lists as a contribution a distributed mechanism for obtaining the required demand measurements, and Remark 4 refers to '[30, Appendix B]' for this mechanism. However, [30] is the present manuscript and the current version contains no Appendix B and no other description of the distributed scheme. Either the distributed scheme and the proof that stability and optimality are preserved must be included, or the corresponding contribution claim should be removed.","section":"Remark 4 and Section I contribution list"}],"minor_comments":[{"comment":"The term 'non-conservative' for epsilon is not formally justified: the proof bounds C* - C_opt by hat{q}^2/(2D) and then by epsilon, but it does not demonstrate that some equilibrium attains the bound. Please either provide a tightness argument or soften the wording to 'explicit and small'.","section":"Theorem 5 and its proof"},{"comment":"In the paragraph after equation (41), the interval notation '[p^c_j, p^c_j]' appears to be a typo; it should presumably be '[p^c_j, p^c_j + bar{d}_j]' to match the surrounding argument and equation (22d).","section":"Appendix, proof of Theorem 5"},{"comment":"The optimality guarantee requires exact aggregate demand ell in (12) and the matching condition alpha_j = c_j^{-1}. Remark 9 correctly notes that replacing exact values by bounds preserves only stability and not optimality, but this limitation should also be stated explicitly in the abstract or theorem statement to avoid overclaiming.","section":"Theorem 5 and Remark 9"},{"comment":"The notation for p^c_j, bar{p}^c_j, and p^c is visually easy to confuse; consistent typesetting of the subscripts and bars would improve readability.","section":"Equations (11), (15), (24)"}],"recommendation":"major_revision","confidential_remarks":"The paper is part of the authors' own line of work, and the reliance on [16] for the proofs of Propositions 1, 3, and 5, together with the self-reference to [30, Appendix B] for the distributed scheme, needs editorial attention. The central idea is promising and the flaws identified in the main report appear fixable, but the proof gaps are substantial enough that a major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: this is a genuinely useful idea in power-systems control, but the proof of the main convergence theorem is not fully sound as written. I agree with the conditional verdict, and I'd weigh the proof problems a bit more heavily than the reader's summary does.\n\nWhat is new and good: the adapted hysteretic policy in (21), the design conditions, and the epsilon-optimality statement for a mixed-integer allocation problem are real extensions beyond the authors' secondary-control work. The paper does a service by separating chattering, limit cycles, and suboptimality, and by giving explicit design rules. The NPCC simulations are a reasonable sanity check, not a substitute for the proofs, but they support the qualitative claims.\n\nThe soft spots, in order of importance:\n\n1. Equation (32) is wrong as written. For V_M = (τ/2)(p_M−p_M*)^2, differentiating (2a) gives an α_j in front of the cross term: −α_j(p_M−p_M*)(ω−ω*), not −(p_M−p_M*)(ω−ω*). Without that α_j, the cross term with V_F does not cancel, and the claimed negative semidefiniteness of V-dot is not true for general α_j. This affects the Lyapunov arguments in Theorems 1, 3, and 4. It looks fixable by rescaling V_M with 1/α_j, but the fix has to be written and checked.\n\n2. The marginal case pc = p^c_j in Theorem 4 is a genuine hole. At that threshold, a solution can start with σ=1 at the σ=0 equilibrium point and initially move away from that equilibrium, so V is increasing rather than nonincreasing. This is exactly the boundary case where the epsilon-optimality argument needs to work. A patient reviewer can probably patch it, but the present text invokes [27, Cor. 8.7(b)] without the needed monotonicity.\n\n3. The current version points to a distributed scheme in “Appendix B” of the arXiv preprint [30], but that appendix is not here. Either include the scheme or clearly mark it as future work. Minor but needs fixing.\n\nThe optimality guarantee also depends on exact aggregate demand and on α_j = c_j^{-1}; Remark 9 honestly concedes that bounds on ℓ and D preserve stability but not the epsilon bound. That limitation should be prominent in any revision.\n\nWho is this for? Control theorists and power-system engineers working on demand response and primary frequency control. It deserves a serious referee, but I would not accept it in its current form. I'd ask for a repaired Lyapunov proof and a clean handling of the boundary cases, then it could be a solid contribution.","headline":"Useful adapted-hysteresis scheme with a real proof gap: the Lyapunov derivative drops the droop coefficient α_j, and the marginal threshold case breaks monotonicity; likely fixable, but not ready as is.","tokens_in":27264,"tokens_out":6041,"would_cite":false,"duration_ms":64999,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C30","90C11","93B52","34D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that an adapted hysteretic on-off load policy, driven by local frequency and total demand, converges without chattering or limit cycles and reaches power allocations within epsilon of the global optimum.","keywords":["frequency control","on-off loads","hysteresis","hybrid systems","chattering","limit cycles","optimal power allocation","mixed-integer optimization"],"falsifier":"Run the adapted hysteresis scheme on a test system with a deliberately wrong total-demand signal, say 10% lower than the true value, and check whether the steady-state generation-and-load cost still lies within $\\epsilon = \\max_j \\bar{d}_j^2/(2D)$ of the H-OSLC optimum; Theorem 5 depends on $p^c = -\\ell$ exactly, so this directly probes the bound. A second check is to violate $\\alpha_j = c_j^{-1}$ and observe whether the equilibrium cost gap exceeds the stated epsilon.","tokens_in":26218,"feed_emoji":"⚡","tokens_out":5562,"duration_ms":53373,"temperature":0.7,"pith_summary":"The paper asks whether loads that can only be on or off can safely help with primary frequency regulation, the fast seconds-scale balancing that follows a disturbance in a power grid. It shows that naive frequency-threshold switching makes loads chatter, and plain hysteresis can create limit cycles. Its main claim is that an adapted hysteretic scheme, where loads switch based on local frequency plus a signal of aggregate demand, eliminates both problems: all solutions converge to an equilibrium, and the steady-state power allocation is within epsilon of the global optimum of a mixed-integer supply-and-load cost problem. The paper also gives an explicit, non-conservative value for epsilon and verifies the behavior on a realistic 140-bus system.","feed_headline":"On-off loads can aid grids without chattering or limit cycles","feed_subtitle":"Hysteretic load switching provably converges and keeps allocation cost near the global optimum.","key_machinery":"The central object is the adapted hysteretic load policy (21) with power-command signal $p^c = -\\ell$ and thresholds specified by Design Condition 2. It gives each load three modes, stay on under high demand, standard hysteresis under low demand, and a one-way switch when demand is intermediate, which prevents chattering and provides a Lyapunov function for the hybrid system. The same thresholds encode the KKT conditions of the continuous relaxation of H-OSLC, so equilibrium cost is forced close to the global optimum.","core_discovery":"Under Design Condition 2, the hybrid power-network system with adapted hysteretic on-off loads has bounded solutions that converge to equilibria, and every such equilibrium is epsilon-optimal for the mixed-integer H-OSLC problem with $\\epsilon = \\frac{1}{2D}\\max_j (\\bar{d}_j)^2$. The design ranks loads by their cost per unit of demand and aligns that ranking with the KKT conditions of the continuous relaxation of the mixed-integer problem; when the aggregate demand parameter falls between design thresholds, the equilibrium allocation is exactly optimal, and the only suboptimality comes from the single load whose threshold interval contains that parameter.","pith_inferences":["Because epsilon does not depend on the number of loads or the network size, the scheme should remain near-optimal as many small on-off devices join, with the largest single device setting the size of the cost gap.","An operator that cannot obtain exact total demand could use a conservative bound for stability and periodically refine the estimate; the cost penalty should be driven mainly by the mismatch between the true and the used demand signal.","The ranking by cost per unit demand suggests an immediate extension to heterogeneous loads of different sizes: reducing the largest load magnitude $\\bar{d}_j$ would shrink epsilon more effectively than retuning any other single parameter."],"forward_implications":["Under Design Condition 2, every maximal solution of the hybrid system (27) is bounded and converges to an equilibrium, so limit cycles and chattering are ruled out.","At every equilibrium, the total cost of generation, uncontrollable demand, and load switching is within $\\epsilon = \\frac{1}{2D}\\max_j (\\bar{d}_j)^2$ of the global minimum of the mixed-integer H-OSLC problem.","When the aggregate demand parameter $p^c$ falls outside all intervals $[p^c_j,\\bar{p}^c_j]$, the equilibrium allocation is exactly optimal, not merely epsilon-optimal.","A simpler design, Design Condition 1, needs only lower bounds on $D$ and retains convergence, but gives no optimality guarantee.","The convergence properties survive when only an upper bound on $|\\ell|$ and a lower bound on $D$ are known, so the scheme is robust to some measurement uncertainty."],"supporting_citations":[{"why":"Supplies the hysteretic on-off load framework that this paper extends from secondary to primary frequency control.","marker":"[16]"},{"why":"Provides the Filippov solution framework used to analyze the non-hysteretic switching loads in Theorem 1.","marker":"[26]"},{"why":"Defines hybrid time domains, maximal solutions, and the Lyapunov corollary used in the convergence proofs of Theorems 3 and 4.","marker":"[27]"},{"why":"Gives the nonsmooth Lyapunov convergence result used to show convergence to equilibria in the discontinuous switching case.","marker":"[34]"},{"why":"Establishes NP-hardness of mixed-integer optimization, motivating the epsilon-optimality approach for the H-OSLC problem.","marker":"[31]"},{"why":"Describes the distributed scheme for obtaining aggregate demand without altering stability and optimality properties.","marker":"[30]"},{"why":"Provides the Power System Toolbox model used for the NPCC 140-bus simulations that verify the analytical results.","marker":"[32]"}],"fun_headline_variants":["Hysteretic on-off loads cure chattering and limit cycles in grids","On-off loads with hysteresis: stable frequency, near-optimal allocation","Grid frequency control without chattering via hysteretic on-off loads","Chattering-free on-off loads achieve epsilon-optimal power allocation","Load hysteresis eliminates limit cycles and ensures convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimality guarantee depends on the control scheme knowing the exact total uncontrollable demand $\\ell$ and on setting every generator's droop coefficient to the reciprocal of its generation cost; with only bounds on $\\ell$ and $D$, the paper says stability remains but the epsilon cost bound does not.","fun_headline_variants_meta":{"raw":{"variants":["Hysteretic on-off loads cure chattering and limit cycles in grids","On-off loads with hysteresis: stable frequency, near-optimal allocation","Grid frequency control without chattering via hysteretic on-off loads","Chattering-free on-off loads achieve epsilon-optimal power allocation","Load hysteresis eliminates limit cycles and ensures convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3339,"prompt_tokens":936,"completion_tokens":2403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2318}},"tokens_in":552,"tokens_out":2403,"duration_ms":15037,"temperature":1.0,"reasoning_tokens":2318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:50:31.157793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the adapted hysteresis scheme on a test system with a deliberately wrong total-demand signal, say 10% lower than the true value, and check whether the steady-state generation-and-load cost still lies within $\\epsilon = \\max_j \\bar{d}_j^2/(2D)$ of the H-OSLC optimum; Theorem 5 depends on $p^c = -\\ell$ exactly, so this directly probes the bound. A second check is to violate $\\alpha_j = c_j^{-1}$ and observe whether the equilibrium cost gap exceeds the stated epsilon.","supporting_citations":[{"cited_title":"Secondary frequency control with on-off load side participation in power networks","cited_arxiv_id":"1708.09351","evidence_quote":"Supplies the hysteretic on-off load framework that this paper extends from secondary to primary frequency control."},{"cited_title":"Discontinuous dynamical systems,","cited_arxiv_id":null,"evidence_quote":"Provides the Filippov solution framework used to analyze the non-hysteretic switching loads in Theorem 1."},{"cited_title":"Stability and stabiliz ation of discon- tinuous systems and nonsmooth lyapunov functions,","cited_arxiv_id":null,"evidence_quote":"Gives the nonsmooth Lyapunov convergence result used to show convergence to equilibria in the discontinuous switching case."},{"cited_title":"Reducibility among combinatorial problem s,","cited_arxiv_id":null,"evidence_quote":"Establishes NP-hardness of mixed-integer optimization, motivating the epsilon-optimality approach for the H-OSLC problem."},{"cited_title":"Primary frequency regulation in power grids with on-off loads: chattering, limit cycles and convergence to optimality","cited_arxiv_id":"1908.08077","evidence_quote":"Describes the distributed scheme for obtaining aggregate demand without altering stability and optimality properties."},{"cited_title":"Power system toolbox , v 3.0,","cited_arxiv_id":null,"evidence_quote":"Provides the Power System Toolbox model used for the NPCC 140-bus simulations that verify the analytical results."}],"review_version":1}