Review Reports
- Xiangjin Zeng 1 and
- Cong Xi 2,*
Reviewer 1: Abraham Vadillo Morillas Reviewer 2: Anonymous
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsSummary
The manuscript proposes an intelligent multivariable control strategy (F-DDPG) for phosphogypsum pretreatment, combining a Mamdani-type fuzzy controller (baseline control actions) with a DDPG agent that adaptively adjusts per-channel gain coefficients for pH, moisture content, and flow rate. The overall architecture is clear at block-diagram level and the simulated comparisons against PID, fuzzy, and fuzzy-PID show improved transient and steady-state indicators in several scenarios.
However, the paper currently lacks critical methodological detail and reproducibility elements required to assess scientific validity and generality. In particular, the process/plant model is not specified, the baseline controller tuning procedures are not documented, the fuzzy-controller numerical choices are incomplete (notably defuzzification and signal scaling), and the RL setup is missing key hyperparameters and statistical validation. I therefore recommend Major Revision.
Additionally, there are internal-consistency issues (e.g., Table 2 lists Experiment A and Experiment C identically, and the text refers to “Section 2.2.3” which does not exist) (p. 13, lines 409-418).
Major Comments
1) The process/plant model is not specified (reproducibility and generality)
The manuscript does not provide a mathematical/engineering description of the controlled plant (transfer functions, differential equations, nonlinear blocks, time constants, delays, coupling structure, operating ranges). Without this, readers cannot reproduce results or assess whether the method generalizes beyond a specific Simulink implementation.
Paper-specific detail: Section 2.1 provides a process flow diagram (Figure 1) and qualitative statements, including that the authors “reconstruct the dynamic control behaviors … within the MATLAB/Simulink environment” but no explicit plant dynamics/parameters are reported (p. 4, lines 109-137).
Required actions: Provide the full plant model (equations or explicit Simulink block description), parameter values, coupling relations, and how these were obtained/validated.
At minimum, report: (i) sample time and any transport delays; (ii) saturation/limits and rate constraints on actuators; (iii) nonlinearities (e.g., neutralization kinetics, mixing/residence-time dynamics, sensor filtering); and (iv) how cross-coupling was implemented.
2) Fuzzy controller numerical specification is incomplete (defuzzification, scaling, composite signals)
The combination of error and error-rate into composite signals cpH,cm,cfc_{pH}, c_m, c_f is stated but not defined (weights, normalization, derivative filtering, sampling time). The defuzzification method is not specified, and output scaling/saturation limits are not documented.
Paper-specific detail: the manuscript explicitly states that errors and error-rates are “combined to form three composite signals, denoted as c_pH, c_m, and c_f” but does not give the formulas or signal conditioning (p. 6, lines 195-202). It also lists “(4) Select the defuzzification algorithm” but does not name the chosen method (p. 7, lines 223-230).
Required actions: Explicitly define signal processing and defuzzification, and document scaling/saturations/rate limits.
Please specify (i) exact formulas for each composite signal and how derivatives are computed/filtered; (ii) inference operators (AND/OR, implication, aggregation); (iii) defuzzification method; and (iv) mapping from crisp outputs to valve openings (units, saturations, and rate limits).
3) Action bounds allow negative “gain” values; safety/stability constraints are missing
Actions {kpH,km,kf} ∈ [−1,1] are applied multiplicatively to fuzzy outputs. Negative gains can invert control directions and potentially destabilize loops; no safety mechanism (positive mapping, clipping, constraints, safe exploration) is described.
Paper-specific detail: the actor-network output is explicitly bounded to [-1,1] via tanh (Figure 7), which permits negative gains unless additional constraints are enforced (p. 10, lines 312-316). The integration mechanism also indicates a multiplicative use of gains on fuzzy outputs (p. 5, lines 159-166).
Required actions: Map actions to strictly positive gains, add saturations/rate limits, and discuss safety/constraints (pH limits, actuator constraints).
A practical remedy is to re-parameterize gains to be strictly positive (e.g., k = k_min + (k_max-k_min)*sigmoid(a)) and enforce channel-wise constraints (pH safety band, valve limits, maximum rate of change). If negative gains arebati used intentionally, provide a clear stability/safety argument and evidence.
4) Baseline controller tuning is not documented (fairness of comparisons)
PID/fuzzy/fuzzy-PID tuning procedures and parameters (including anti-windup, constraints, and decoupling) are not provided, limiting fairness and reproducibility.
Paper-specific detail: the comparative study specifies common target setpoints (pH=5.5, moisture=12%, flowrate=100 L/min) and compares PID, fuzzy, fuzzy-PID, and F-DDPG, but does not disclose baseline tuning parameters (p. 12, lines 368-376).
Required actions: Provide tuning procedures and final parameters for all baselines.
Provide a table of baseline parameters (PID Kp/Ki/Kd per loop; fuzzy-PID coupling structure; filters; saturations; anti-windup) and a short description of the tuning protocol (manual, rule-based, optimization, etc.).
5) RL training details and statistical validation are insufficient
Key hyperparameters are missing (discount factor, replay buffer size, batch size, target update parameters, exploration noise type/magnitude, normalization). Results are not reported across multiple seeds; training curves and variability are absent.
Paper-specific detail: the manuscript reports learning rates (actor 1×10^-4, critic 5×10^-4), 400 max steps/episode, and 300 training episodes (p. 12, lines 371-373), but does not report key DDPG configuration parameters required for replication (gamma, replay buffer length, batch size, target update, exploration noise).
Required actions: Report full RL setup and provide mean±std (or CI) across multiple runs/seeds, plus training curves.
Given typical seed sensitivity in DDPG, please report results over multiple random seeds and include learning curves (episodic reward and key tracking metrics vs episode). Clarify whether the reported traces use the best checkpoint or the final checkpoint.
6) Metrics definitions and internal consistency need strengthening
Metrics such as “Error (%)”, settling time, overshoot, and peak counts require formal definitions (tolerance bands, normalization, counting rules). The manuscript also contains internal referencing/section inconsistencies that must be corrected.
Paper-specific detail: Table 2 lists Experiment A and Experiment C identically, yet the text indicates Experiment C is discussed elsewhere (“Section 2.2.3”) which does not exist (p. 13, lines 409-418). Table 3 defines N as the number of overshoot peaks and reports “Error (%)” values (e.g., 0.01-0.18) without clarifying whether these are true percentages or normalized errors (p. 17, lines 515-520).
Required actions: Define all metrics precisely; fix section references and ensure text aligns with figures/tables.
Define settling-time tolerance bands (e.g., ±2% or ±5% of setpoint) and provide the peak-counting rule for N. For Table 3, clarify whether “Error (%)” is percent or a normalized/absolute error measure and revise units/labels accordingly.
7) Claims about “energy/low-carbon” benefits are not substantiated by measured outputs
Energy/low-carbon claims are made without an explicit energy model or proxy (e.g., valve movement cost, pump power proxy).
Paper-specific detail: although the reward includes a control-effort term r_a = -0.1(k_pH^2 + k_m^2 + k_f^2) (Equation (3)), this is not an energy model; yet the text frames the reward as balancing accuracy, stability, and “energy consumption” (p. 10, lines 291-298).
Required actions: Provide an energy proxy/model and quantify it, or tone down such claims.
A practical proxy could be ∫|u(t)|dt and/or ∫|du/dt|dt per valve (or an approximate pump-power term if flow actuation is modeled). Report the proxy values for each controller.
8) Related work and references require correction and expansion
The reference list contains duplicates (e.g. [5,6] and [9,10]) and omits key related work (seminal DDPG reference; modern continuous-control RL such as TD3/SAC; and MPC baselines for constrained MIMO process control).
Paper-specific detail: references 5 and 6 are identical entries (p. 18, lines 583-587), and references 9 and 10 are also identical (p. 19, lines 592-595).
Required actions: Remove duplicates, correct references, and expand related work with appropriate justification for algorithmic choices.
Add a foundational DDPG citation and briefly justify the choice of DDPG versus more recent continuous-control alternatives (e.g., TD3/SAC). If MPC is not included as a baseline, provide a short justification (e.g., model unavailability, constraint handling scope).
9) Coupling coefficients (4:15 and 1:6) are control-design choices and should be justified
The manuscript states fixed proportional coefficients for cross-feedback in the gain modules: 4:15 in the water-valve path and 1:6 in the acid–alkali path (p. 5, lines 172-178). These values materially affect MIMO behavior and should be justified (physical reasoning, identification, or tuning protocol) and, ideally, supported by a sensitivity/ablation study.
Required actions: explain the origin of the coupling ratios and add at least a brief sensitivity/ablation (vary the ratios over a reasonable range and report the impact on settling time/overshoot/error).
Minor Comments (should address)
-
Provide units and scaling for pH/moisture/flow errors in plots; clarify whether plots show absolute error or normalized error.
-
Clarify how coupling ratios (e.g., 4:15, 1:6) were chosen; include an ablation or sensitivity study.
-
Consider adding an ablation: fuzzy-only vs F-DDPG, and (if feasible) DDPG-only direct control vs supervised gain modulation.
-
Ensure consistent terminology (“flowrate/flow rate”; “acid–alkali valve”; etc.) and correct minor formatting errors.
-
In the description of initial conditions, the manuscript states “discussed in Section 2.2.3”, but Section 2.2.3 does not exist in the document.
-
Several instances appear as “pretreatment[1–3]”, “controller[4]”, etc. In standard MDPI style, this should be “pretreatment [1–3]”, “controller [4]” (with a space).
- Inconsistent terminology:
-
The manuscript alternates between “flowrate” and “flow rate”. Choose one and use it consistently (recommended: flow rate).
-
It alternates between “moisture content” and “moisture-content”. Standardize usage (recommended: moisture content, except when a clear compound adjective is needed).
-
Controller naming varies: “Fuzzy PID” vs “fuzzy–PID / fuzzy-PID”. Standardize (recommended: fuzzy–PID).
-
- In Table 3, the column “Error (%)” contains values around 0.01–0.18. If these are truly percentages, they should be reported in % units (e.g., 1–18%). If they are normalized/absolute errors, the label should be changed (e.g., “Normalized error”).
Recommendation
Major Revision.
The proposed hybridization concept is potentially interesting and is explained clearly at a block-diagram level. Nevertheless, the paper must provide a reproducible plant model, complete fuzzy-controller numerical details, fair baseline tuning, full RL hyperparameter disclosure, and statistical validation before the performance claims can be evaluated with scientific rigor.
Abstract / Conclusions: “harmless pretreatment”
-
Before: “phosphogypsum harmless pretreatment”
-
After: “safe pretreatment of phosphogypsum” or “benign pretreatment of phosphogypsum”... “Harmless pretreatment” sounds unnatural in academic English.
Conclusions: non-idiomatic phrasing
-
Before: “has raised increasing demands for …”
-
After: “has led to increasing demand for …”
“Raised increasing demands” is not an idiomatic expression.
Caption (already noted)
-
Before: “different parameters”
-
After: “different controlled variables”
Style: repetitive “green/intelligent” phrasing
-
In Results/Conclusions, expressions like “efficient, green, intelligent control …” are repeated. Not wrong, but it reads promotional rather than technical. A more academic approach is to (i) use one concise technical descriptor (e.g., “efficient and robust control”) and (ii) only claim environmental/energy benefits if an explicit metric/proxy is actually reported.
Author Response
Summary Thank you very much for taking the time to review our manuscript and for providing detailed and constructive comments. We sincerely appreciate the reviewer’s careful evaluation and valuable suggestions. In response, we have substantially revised the manuscript by adding explicit process modeling details, completing the numerical specification of the fuzzy controller, clarifying the DDPG training configuration, strengthening metric definitions, correcting internal inconsistencies, revising energy-related claims, and updating the related work and references. All revisions have been incorporated into the revised manuscript.
Questions for General Evaluation
Does the introduction provide sufficient background and include all relevant references?
Response: Can be improved. The introduction has been revised to better clarify the considered pretreatment process and the motivation for multivariable intelligent control.
Are all the cited references relevant to the research?
Response: Yes. Duplicate references have been removed, missing DOIs or stable URLs have been added where available, and the reference list has been carefully checked and reorganized.
Is the research design appropriate?
Response: Yes. Additional details on the process model, controller configuration, and training setup have been provided.
Are the methods adequately described?
Response: Can be improved. The fuzzy-controller numerical design, defuzzification method, gain constraints, and DDPG hyperparameters have been explicitly documented.
Are the results clearly presented?
Response: Yes. Metric definitions have been clarified and internal inconsistencies corrected.
Are the conclusions supported by the results?
Response: Yes. Claims related to energy and low-carbon performance have been moderated to match the reported metrics.
Point-by-point response to Comments and Suggestions for Authors
Comments 1: The process/plant model is not specified.
Response 1: Thank you for pointing this out. We agree that a clear description of the controlled process is essential for reproducibility. Therefore, we have added an explicit dynamic model of the phosphogypsum pretreatment process in the revised manuscript. The process is now described as a nonlinear, coupled discrete-time system with three controlled variables and three manipulated variables. The model includes the sampling time, actuator saturation and rate limits, bounded disturbances, and cross-coupling terms. These additions are provided in Section 2.1.1 of the revised manuscript.
Comments 2: Fuzzy controller numerical specification is incomplete.
Response 2: We agree with this comment. In the revised manuscript, we have explicitly defined the construction of the composite fuzzy input signals by combining the tracking error and its rate of change, including weighting coefficients and derivative computation. The inference mechanism and defuzzification method are now clearly specified, with the centroid method adopted for defuzzification. These revisions are presented in Section 2.3 of the revised manuscript.
Comments 3: Action bounds allow negative gain values; safety and stability constraints are missing.
Response 3: Thank you for this important observation. In the revised manuscript, we clarify that the gain outputs of the DDPG agent are constrained by predefined bounds and rate limits before being applied to the fuzzy-controller outputs. This prevents destabilizing control actions and ensures physically meaningful actuator behavior. In addition, the DDPG agent operates as a supervisory gain-adjustment mechanism rather than a direct controller, which enhances overall stability during learning. This clarification has been added in Section 2.5.
Comments 4: Baseline controller tuning is not documented.
Response 4: We agree that the tuning of baseline controllers should be clearly documented. Accordingly, we have added a description of the tuning procedures and parameter settings for the PID, fuzzy, and fuzzy–PID controllers. The tuning rationale and configuration are now summarized in Section 3.1 of the revised manuscript to ensure fairness and reproducibility of the comparative study.
Comments 5: RL training details and statistical validation are insufficient.
Response 5: Thank you for this comment. We have expanded the description of the DDPG training setup by reporting key hyperparameters, including the discount factor, replay buffer size, minibatch size, target-network update coefficient, and exploration noise type. We further clarify that, due to the control-oriented nature of this study, the reported results are intended to demonstrate representative performance trends rather than statistical optimality across a large number of random seeds. These clarifications have been added in Section 3.1.
Comments 6: Metrics definitions and internal consistency need strengthening.
Response 6: We agree with this comment. In the revised manuscript, all performance metrics used in Table 3 are now formally defined, including steady-state error, settling time, overshoot, and peak number. The tolerance band for settling time and the counting rule for overshoot peaks are explicitly stated. In addition, internal inconsistencies have been corrected, including the incorrect section reference for Experiment C. These corrections are provided in Sections 3.2 and 3.3.
Comments 7: Claims about energy and low-carbon benefits are not substantiated.
Response 7: Thank you for highlighting this issue. We agree that energy-related claims should be supported by explicit metrics. In the revised manuscript, we have moderated statements related to energy and low-carbon benefits and clarified that the control-effort term in the reward function serves as a proxy to penalize excessive control actions rather than a direct energy-consumption model. Corresponding revisions have been made in Sections 2.4 and 4.
Comments 8: Related work and references require correction and expansion. Response 8: We agree with this comment. Duplicate references have been removed, missing DOIs or stable URLs have been added where available, and the reference list has been reorganized to ensure consistency. In addition, foundational and recent studies related to DDPG-based control and fuzzy–reinforcement learning hybrid methods have been incorporated, and the rationale for selecting DDPG has been briefly clarified in the Introduction.
Comments 9: Coupling coefficients should be justified.
Response 9: Thank you for this valuable suggestion. In the revised manuscript, we clarify that the coupling coefficients were selected based on process knowledge and empirical tuning to reflect the relative influence of each actuator on the controlled variables. We further emphasize that these coefficients are design parameters rather than fixed physical constants, and that the adaptive gain modulation provided by the DDPG agent reduces sensitivity to their exact values. This explanation has been added in Section 2.2.
Response to Comments on the Quality of English Language The manuscript has been revised to correct non-idiomatic expressions, standardize terminology, and improve overall readability. For example, expressions such as “harmless pretreatment” have been replaced with “safe pretreatment of phosphogypsum,” and inconsistent terminology has been unified throughout the text. Additional clarifications We confirm that all issues raised by Reviewer 1 have been carefully addressed in the revised manuscript, and the corresponding revisions have been incorporated to improve clarity, rigor, and reproducibility.
Author Response File:
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsNice controller has been designed to check the pretreatment of phosphogypsum. The following points need to be addressed.
- In Table 2, some experimental points are given; are these actual experiments or design experiments?
- The results show the performance of F-DDPG control over PID, Fuzzy, Fuzzy PID. Please justify it. At present, you have just mentioned its self-learning capability.
- For almost all the cases, Fuzzy control is very close to F-DDPG control. Is it possible to fine tune the Fuzzy control to match F-DDPG?
- Why is it necessary to control moisture, pH, and flow of phosphogypsum?
- There are several processes to pretreat the phosphogypsum; which particular process are you talking about? It is better to give the process flowsheet and then articulate the importance of control.
Author Response
-
Summary
Thank you very much for reviewing our manuscript and for your constructive comments. We appreciate the reviewer’s positive evaluation of the proposed controller design as well as the insightful questions regarding process selection, variable importance, and comparative justification. In response, we have revised the manuscript to clarify the considered phosphogypsum pretreatment process, justify the selection of controlled variables, strengthen the explanation of the advantages of the proposed F-DDPG method, and discuss the relationship between fuzzy control and the proposed approach. All relevant revisions have been incorporated into the revised manuscript. -
Questions for General Evaluation
Is the work a significant contribution to the field?
Response: Yes. The manuscript presents an intelligent hybrid control strategy for a representative multivariable pretreatment process with strong coupling and nonlinearity.
Is the work well organized and comprehensively described?
Response: Can be improved. Additional clarifications on process selection and control objectives have been added.
Is the work scientifically sound and not misleading?
Response: Yes. The modeling assumptions and controller design have been further clarified.
Are there appropriate and adequate references to related and previous work?
Response: Yes. The reference list has been revised and expanded.
Is the English used correct and readable?
Response: Yes. The language has been revised for clarity and consistency.
- Point-by-point response to Comments and Suggestions for Authors
Comments 1:
In Table 2, some experimental points are given; are these actual experiments or design experiments?
Response 1:
Thank you for this question. We clarify that the results presented in Table 2 are obtained from simulation-based design experiments rather than physical industrial experiments. These operating conditions were intentionally selected to represent typical and extreme initial states encountered in phosphogypsum pretreatment processes, enabling a systematic evaluation of controller performance under varying deviations. This clarification has been added in Section 3.2 of the revised manuscript.
Comments 2:
The results show the performance of F-DDPG control over PID, Fuzzy, and Fuzzy PID. Please justify it. At present, you have just mentioned its self-learning capability.
Response 2:
We agree that a clearer justification is necessary. In the revised manuscript, we explain that the superior performance of the F-DDPG controller arises not only from its self-learning capability, but also from its hybrid structure. The fuzzy controller provides robust baseline control based on expert knowledge, while the DDPG agent performs online gain adaptation to compensate for nonlinearities, coupling effects, and operating-condition variations. This supervisory learning mechanism enables coordinated multivariable optimization that is difficult to achieve with fixed-rule or fixed-parameter controllers. The explanation has been strengthened in Sections 2.5 and 3.2.
Comments 3:
For almost all the cases, Fuzzy control is very close to F-DDPG control. Is it possible to fine tune the Fuzzy control to match F-DDPG?
Response 3:
Thank you for this insightful comment. We acknowledge that a carefully fine-tuned fuzzy controller can achieve competitive performance under specific operating conditions. However, such tuning is typically performed offline and remains fixed once deployed. In contrast, the proposed F-DDPG controller dynamically adjusts the fuzzy-controller gains online, allowing it to adapt to changing operating conditions, disturbances, and coupling strengths. We have added a discussion emphasizing that the advantage of F-DDPG lies in its adaptability and robustness across multiple scenarios rather than peak performance under a single operating point. This discussion has been added in Section 3.3 of the revised manuscript.
Comments 4:
Why is it necessary to control moisture, pH, and flow of phosphogypsum?
Response 4:
We appreciate this question. In the revised manuscript, we explicitly justify the selection of pH, moisture content, and slurry flow rate as the primary controlled variables. pH regulation is essential to ensure effective neutralization and to limit residual acidity. Moisture content directly reflects washing efficiency and strongly affects the energy demand of downstream dewatering and drying processes. The slurry flow rate influences residence time, mixing conditions, and mass-transfer behavior, and is strongly coupled with both pH adjustment and moisture regulation. Due to these strong interactions, simultaneous multivariable control is therefore required. This justification has been added in Section 2.1.
Comments 5:
There are several processes to pretreat phosphogypsum; which particular process are you talking about? It is better to give the process flowsheet and then articulate the importance of control.
Response 5:
Thank you for this helpful suggestion. In the revised manuscript, we clarify that the study focuses on a typical wet pretreatment stage commonly applied prior to phosphogypsum dewatering and drying, including washing and chemical neutralization operations. Figure 1 provides the representative process flow, and the associated text has been revised to clearly state the process scope and control objectives. This clarification can be found in Section 2.1 of the revised manuscript.
- Response to Comments on the Quality of English Language
No specific language corrections were requested by the reviewer. Nevertheless, the manuscript has been carefully proofread, and minor wording and consistency issues have been corrected throughout the text to improve readability.
- Additional clarifications
We confirm that all revisions responding to Reviewer 2’s comments have been incorporated into the revised manuscript, with corresponding sections updated to improve clarity, justification, and technical completeness.
Author Response File:
Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe revised manuscript proposes an intelligent multivariable control architecture (F‑DDPG) for phosphogypsum pretreatment, where a Mamdani-type fuzzy controller provides baseline control actions and a DDPG agent adaptively modulates per‑channel gain coefficients for pH, moisture content, and flow rate. The paper reports simulation-based comparisons against PID, fuzzy, and fuzzy–PID controllers and argues that F‑DDPG improves convergence speed, steady-state accuracy, and robustness under coupling and disturbance scenarios.
Improved points in Rev. 1
- Plant/process dynamics are now explicitly introduced via a discrete-time coupled model (Eqs. (1)–(3); p. 5, lines 160–178).
- Composite fuzzy inputs and the derivative computation/filtering are defined with explicit weights (α=1.0, β=0.2) and a noise-mitigation low-pass filter (p. 7, lines 235–249).
- Defuzzification is specified (centroid/center-of-gravity) and the fuzzy rule base is summarized (p. 8, lines 276–288).
- The coupling-weight rationale for the multivariable feedback paths (e.g., 4:15 and 1:6) is explained qualitatively (p. 6, lines 210–216).
- The DDPG setup is substantially clearer: actor output bounds are stated (p. 11, lines 367–372) and the raw action is mapped to strictly positive gains (k_min=0.2, k_max=2.0) to avoid sign inversion (p. 12, lines 393–404).
- Key RL hyperparameters and the use of five independent seeds are now reported (p. 13, lines 438–446).
- Metric definitions are formalized (steady-state error %, ±5% settling band, overshoot and peak counting) to support Table-based comparisons (p. 18, lines 590–598).
- A robustness test under injected stochastic disturbances is described in a reproducible narrative form (Experiment C; p. 17, lines 555–563).
- The introduction now acknowledges modern continuous-control RL baselines (TD3/SAC) and gives a rationale for using DDPG (p. 2, lines 80–88).
Remaining issues to address
Disclaimer: Take the next points as areas to improve for scientific rigor. A comprehensive explanation of why the listed issues should not be changed (if authors don't want to) could also be accepted for publication if the explanation is reasonable.
- Missing numerical parameter values for the plant model prevent full reproducibility. While Eqs. (1)–(3) define the structure, the paper does not provide the numerical values of a_i, b_i, c_ij, nor the bounds/statistics for d_i(k). Without these, it is still not possible to replicate the reported trajectories or compare fairly across alternative controllers.
- Action requested: Provide a table (main text or Appendix) listing a_i, b_i, c_ij and the disturbance/noise configuration used in simulations. If proprietary constraints exist, provide normalized/obfuscated values sufficient to reproduce dynamics and relative coupling.
- Baseline-controller tuning remains insufficiently documented. The revised manuscript lists the competing methods (PID, fuzzy, fuzzy–PID) but does not report the final PID gains (Kp, Ki, Kd), anti-windup/saturation details, nor the fuzzy–PID combination logic used in the comparisons.
- Action requested: Provide baseline parameters and tuning procedure (e.g., rule base/scaling for fuzzy–PID, PID gains and any constraints). At minimum, include a table of all baseline settings so the comparison can be reproduced (see p. 13, lines 447–450, where baselines are listed).
- Disturbance/noise injection is described qualitatively but still lacks quantitative specification. Experiment C states that band-limited white noise is superimposed with a step at t=3 s, but the noise amplitude, bandwidth (or filter), and step magnitude are not reported.
- Action requested: State the exact disturbance parameters used in Experiment C (noise power/variance, bandwidth or filter settings, step size and whether it is applied to errors or measured states) to enable replication (p. 17, lines 559–563).
- Energy/low-carbon statements remain stronger than the evidence presented. The abstract/introduction claims ‘reduces fluctuations in energy consumption’ and offers a ‘low‑carbon solution’ without a direct energy model. Control-signal smoothness is a plausible proxy, but it is not equivalent to energy unless a specific actuator/power model is introduced.
- Action requested: Either (i) define and report an explicit energy proxy/model (e.g., weighted valve movement cost tied to power) or (ii) rephrase these claims consistently as ‘control-effort proxy’ throughout the abstract and conclusions (p. 3, lines 97–101).
- Justification/sensitivity for the gain-mapping bounds would strengthen the engineering argument. The mapping uses k_min=0.2 and k_max=2.0, but the choice appears ad hoc. Given the supervisory nature of the agent, a brief sensitivity or rationale (e.g., actuator limits, stability margins) would be appropriate.
- Action requested: Add a short rationale for k_min/k_max and (if feasible) a small sensitivity check showing robustness to moderate changes in these bounds (p. 12, lines 399–403).
English language was improved in Rev 1.
Author Response
We thank the reviewer for the constructive comments. In this revision, we have explicitly reported all numerical plant-model parameters, disturbance bounds, and noise settings (Table 1) to ensure full reproducibility. The tuning procedures and final parameter settings of all baseline controllers are now clearly summarized in Table 3, and the disturbance configuration in the robustness experiment is quantitatively specified in Section 3.2.3. Statements related to energy and low-carbon benefits have been carefully revised to reflect control-effort proxies rather than explicit energy modeling. Additional justification for the gain-mapping bounds has been added, and the manuscript has been further refined for clarity and academic English.
Author Response File:
Author Response.pdf