General comments

The authors present a study projecting future hydrological drought in the Upper Hanjiang River Basin by coupling LSTM-based hydrological models with reservoir operational models. I like the integration of LSTM hydrology and reservoir operation (although have some questions around the need for LSTM-based hydrology in its’ apparent implementation, see below). I have some concerns around the calibration methods and how the authors have interpreted some key results. I also think there needs to be more discussion of hydrological and operational non-stationarity in the context of ‘mining’ systems dynamics using a machine learning model.

The manuscript is in general well written, with clear communication and good quality figures. But I think some substantial revisions are required to address the comments below.

Specific comments

Line 44: I would add Oceania here as multiple significant, record-breaking droughts have impacted Australia over the past ~20 years, with climate change a contributing factor.

Line 47: I find it a bit odd to say: “the time series of land temperatures.” You could just replace with “land temperatures are projected to…”

Lines 51 to 63: To me, it’s an interesting question of whether you can separate the impact of the dam itself on the flow regime from the opportunity it provides for water abstraction and water resources development. By itself, I agree with Wanders and Wada that dams can buffer against low flow impacts by releasing passing flows. But let’s not forget that it’s because of the dam being there that enables much more intense consumptive water use for various industries. The net impact may be that downstream users are more impacted because of the water abstractions and diversions below/from the dams. 

Lines 64 to 78: You might like to discuss the approach by Culley et al. (2016, https://doi.org/10.1002/2015WR018253) in your literature review, where they assess the range of changes in climate that reservoir operational adjustments can adapt to. I am not an author on this paper. While it’s less about only drought, and more about broader reservoir operating objectives, I think it’s relevant to your study.

Line 70: What is the CSSPV2? Is it a hydrological model?

Lines 73 to 74: I think this statement needs some additional evidence. What was the approach Ji et al. used to simulate reservoir operations? I think you just need one more line to demonstrate that it does not “consider… actual reservoir operation data.”

Line 74 to 76: Yes I agree in principle, but also consider the possibility of reservoir operational procedures changing over time due to policy or infrastructure updates. If the change is significant, it may mean that older data is less relevant due to non-stationary operating conditions.  

Lines 87 to 91: In these cases, what deficiencies in the traditional hydrological models are the AI models trying to address? Why would a “fully artificial intelligence-based simulation” provide new insights? I think this is a fairly generic statement and needs to be better linked to the research objectives or gaps here.

Lines 94 to 95: I don’t see how reservoir operation optimisation is a nature-based solution? It’s artificial water regulation infrastructure. I think you should just remove the reference to NBS here, since you don’t refer to it again in the manuscript. Your point still stands that optimisation is a good option because it doesn’t require additional capital.   

Lines 96 to 99: Consider some of the work by Wenyan Wu and colleagues (https://scholar.google.com.au/citations?hl=en&user=7N-YnaQAAAAJ&view_op=list_works) which I think do include drought performance metrics in reservoir optimisation approaches.

Lines 127 to 128: Is this because there is a history of disaster-related damages in the basin?

Lines 220 to 222: Can you provide some additional details on the inputs to the LSTM hydrology model? Were basin characteristics also considered, or are the inputs just meteorological variables? Were more than pr and t used as inputs? If it’s just pr and t as inputs, I’m not sure why you used an LSTM model rather than a simple conceptual rainfall-runoff model, given the additional effort required in model calibration and challenges inherent in extrapolating outside training data (see Maier et al. (2023) for discussion here https://doi.org/10.1016/j.envsoft.2023.105776).

Line 233: What are the differences in the hydroclimate regime in the validation period compared to the calibration period? Normally the differential split sample method you are describing here is supposed to contrast calibration and validation performance between two different climatic regimes to demonstrate out-of-sample performance for the calibration parameter set.

Section 3.1.1: I think this section is missing a little bit of detail on calibration approaches. What objective function was used for calibration? What optimisation algorithm was used? You include this in 3.1.3 but I think it’s part of 3.1.1 methodology. Maybe you can just combine these two sections.

Line 267: A note here that while NSE is commonly adopted in hydrological studies, its’ squared error formulation means it will place far more emphasis on high flows compared to low flow performance. This is fairly well established in hydrology literature with various other objective functions or transformations used to overcome the issue when low flow performance is important. Can you offer some commentary on why high flows are more important for your study, considering your focus on hydrological drought?

Line 355: I disagree that such an NSE threshold is ‘widely accepted.’ There has been a lot of criticism in the literature over such arbitrary threshold-based approaches to model evaluation (see Knoben et al. (2019) https://doi.org/10.5194/hess-23-4323-2019), and recommendations to move towards purpose-dependent evaluation. I am not disputing that your model performance is good overall, but I would remove the reference to Moriasi et al. and this sentence. I would suggest in your case, you include some specific graphical or quantitative evaluation of low flow performance (such as flow duration frequency curves) because of your specific focus on hydrological drought. Figure 5 alone is insufficient as it’s very difficult to separate performance across the flow regime. Based on a cursory inspection of Figure 5, it appears as though your simulations are biased high in the periods of lowest flow.

Lines 548 to 551: I am not sure I agree with these points here. You say that your methods “can effectively reconcile the trade-offs between hydrological drought and hydropower benefits…” I interpret this as finding operational strategies that achieve better hydropower outcomes and reduce drought risk (which you mention as future research). Reconciliation means some compatibility or trade-off, and your methods only focus on hydropower rather than alleviating drought risk because optimisation only uses hydropower indicators. I think you just need to change the language here to something else to summarise your results. But I really don’t think the optimal policy reconciles anything, it rather increases drought risk to pursue hydropower gains, which is very clear from your text and Figure 10.

Discussion section: This article is missing some key discussion on limitations of the adopted approach. There needs to be some acknowledgement of the limitations inherent in (these are examples and the authors should reflect on the key limitations and uncertainties) 1) the possibility of a non-stationary reservoir operating environment over the calibration period; 2) training the LSTM hydrology models on observed meteorology data and then using ISIMIP data for projections (rather than training on ISIMIP data); 3) the selected objective function for calibration which may bias model simulations towards high flows; 4) using the historic operating regime with far future climatic inputs (some recognition that operational strategies will co-evolve with climate and further anthropogenic development); etc.

Technical corrections

Line 35 “pow” should be power

Line 46: Ipcc should be IPCC

Lines 132: Extra parenthesis can be deleted

Line 186: You can remove the acronym BPTT because you don’t use it anywhere else in the manuscript

Line 307: d1 should be a subscript

The manuscript investigates how climate change and reservoir operation jointly shape future hydrological droughts in a heavily regulated basin, using an Ankang Reservoir in the upper Hanjiang River Basin in China as an example. A hybrid framework was developed by coupling a LSTM-based hydrological model with a physics-guided LSTM reservoir operation module, which was then driven by ISIMIP3b CMIP6 projections (five GCMs, three SSPs) to simulate regulated streamflow in the near- (2031-2060) and far-future (2071-2100) periods. Hydrological drought characteristics are quantified using SSI-1, SSI-3, and SSI-12, and a multi-objective optimization (NSGA-II with RBF parameterization) is used to explore whether optimal operating policies can balance hydropower benefits and drought risks.

Overall, the paper addresses an important and timely topic, and the methodology is sound and clearly implemented. I believe the manuscript is suitable for publication after minor revisions to clarify the novelty, add some methodological details, and strengthen the discussion of limitations and applicability.

Major comments

1. Clarify the paper’s novelty

The Introduction already reviews several related works (e.g., VIC-Reservoir and CSSPV2+reservoir frameworks for future drought projections, and recent studies coupling reservoir models with CMIP6). However, the specific advances of this study relative to those works could be articulated more explicitly.

Please more clearly state what is new in this manuscript: e.g., (i) replacing traditional hydrological models with a fully data-driven LSTM for inflow and outflow, and (ii) explicitly combining this hybrid model with multi-objective optimization of operating policies to examine drought–hydropower trade-offs under future climate scenarios.

It would help if the end of the Introduction contained a short, itemized list of the main contributions to distinguish this work from previous hybrid and reservoir–drought studies.

2. Assumption of stationary operation policy when projecting future droughts

The study assumes that the historical operation policy learned by the physics-guided LSTM (1992–2020) is directly applicable to the reference and future climate periods (ISIMIP3b_ref and ISIMIP3b_fut experiments). This is a reasonable and often necessary assumption, but it should be discussed more explicitly as a limitation: operation rules in reality may adapt to changing demands, policies, or infrastructure.

3. Uncertainty and generality of the results

The study uses five ISIMIP3b GCMs and a single basin–reservoir case. While this is already a substantial effort, readers would benefit from a more explicit reflection on the scope and limitations. Please expand the discussion of uncertainties related to (i) the limited number of climate models, (ii) the single-case setting (UHRB, Ankang) and whether the conclusions are transferable to other types of reservoirs or climate regimes, (iii) A short paragraph in the Discussion or Conclusions explicitly addressing “limitations and future work” would strengthen the paper and guide follow-up studies.

Minor comments

1. In the Abstract, there seems to be a small typo in “pow generation guarantee rate”; please correct to “power generation guarantee rate”.

2. Consider defining “LSTM+Reservoir” more explicitly at its first appearance in the Abstract or Introduction (e.g., “a hybrid LSTM-based hydrological and physics-guided reservoir operation model”).

3. Notation and acronyms. Please ensure that all acronyms are defined at first use (e.g., SSI, NBS, NSGA-II, RBF). Some are introduced in the text or caption but could be clarified earlier for readability.

4. It may help to add a short list of symbols for key variables (e.g., V, I, O, THP, PGR, D, S) either in the main text or Supplement.

5. For the parallel coordinate plots in Figure 10, you might consider adding an arrow in the left to clarify which direction is “better” for each metric (e.g., upward is optimal for all axes) to help non-expert readers interpret the trade-offs.