This paper uses generalized additive models (GAMs) to compare pre- and post-dam construction flows as a function of rainfall, temperature and glacier mass changes in 74 high mountain catchments. The approach enables to reconstruct reservoir operations, as well as its hydro-climatic drivers, in situation where inflows and outflows are not readily available (which is often the case). The paper is really interesting, the method is a smart way to tackle a challenging puzzle, and there is a feeling that it is likely novel… but this is incumbent on authors to demonstrate, both by explaining the novelty in more detail, and by validating the method! My main remarks are as follows:

1) the novelty needs to be explicitly and precisely stated on the introduction.

2) the absence of a formal validation is a real issue that should (and can!) be addressed.

3) the methodology is creative, but the explanations are sometimes vague, and needs to be better justified and documented in several places.

4) results need to be presented step by step. They are really interesting, but authors need to be more rigorous in presenting how they got to them. I

t might take a substantial revision for authors to address the comments above. Revisions are important because this work has the potential to be a nice paper. I now develop the four points above.

1) Several points need to be clarified in the introduction:

=> What is the exact scope of the paper? Is the type of data authors base themselves on widely available in a range of cases (not just the Alps)?

=> Why are GAM the appropriate methodology here? Why not use machine learning methods for instance?

=> A simple search of “generalized additive model reservoir” on Scopus returns 120 results. Can authors ascertain no similar attempt has been tried in the past? What have been GAM uses in the literature around hydrological and reservoir modelling?

2) Because there is no inflow and outflow data for the catchment authors study, they do not validate the approach using basins for which the post-dam construction reservoir inflows and outflows are known. But there are many mountain reservoirs in other regions (e.g., the USA) for which inflows and outflows, as well as the other data authors use in the Alps. This would provide formal validation. Note they can then choose to split this paper into two papers (i) one that presents and validates the method, (ii) one that applies it to Alpine catchments to derive new knowledge (which is what they have here). I note here that the confidential dataset shared by authors does not address this concern about validation.

3) In order of appearance, my remarks are as follows:

Performance of the pre-dam model: with 74 catchments, it would only be normal that the GAMs model would work more or less well across the sample. The methodology section (lines 126-130 in particular) does not explain what performance measures are used. Related question: do authors investigate the influence of the length of the pre-development record on model quality? Please include a full list (maybe in supplementary material) of the catchments, including relevant characteristics (e.g., flow record length, year reservoir went online, catchment surface, ratio of reservoir storage over natural annual flow).

Authors give no indication whether the possibility of an extended filling period for the reservoirs was considered. Or did they assume this would be negligible? If so, please justify explicitly.

Section 2.2: the method enables one to derive time series of inflows / storage / release. Authors choose a final signal in terms of difference outflows-inflows that is creative and innovative, and could be a key contribution of the paper. Yet it is not justified, especially in relation to the literature on the role of reservoir in hydrological models that authors claim to contribute to. This leads to several important questions:

i) How do the chosen signals relate literature trying to specify reservoir outflows as a function of inflows, storage, and reservoir purpose?

ii) Why not derive relationships between inflows and outflows once inflows are derived? That is more standard.

iii) How do we know smoothing actually is needed? And how do we know how to parameterise the smoothing technique?

4) The results section is very interesting in highlighting a category of high-elevation hydropower reservoirs (but authors should comment in the discussion how new this is). But it only reports on what happens once the flow signal is derived, whereas it would be great to see results for the fitting of the pre-dam GAM model, as this is key to getting quality results. Authors should report in detail on goodness of fit, relationship between number of years of record and quality of the results, etc.

Another remark on results at lines 161-163: this is a key point that seems difficult to make without taking a long look at reservoirs’ storage capacities. Reservoir with small regulation capacity (ratio of live storage to average annual flow) will naturally have weaker signals.

One last remark concerning results is that it would be great to highlight one or two known water supply reservoirs (for irrigation or any other uses), and show how their method enables quantifying downstream needs. That’s not strictly necessary but it would show that the proposed method is versatile (which it is), instead of being a one-trick pony that only identified high-elevation single use hydrpower reservoir.

Finally, below are also some minor remarks:

Lines 47-51: You talk about a two step setup and then lose the reader a bit by not making both steps explicit. I’d advise doing that.

Line 56: by convention, please introduce an equation with “:” Several times in the paper please replace “see, e.g.” with “see e.g.,”.

Line 113: please replace “(GAM) (Hastie” by “(GAM; Hastie” Figure 3.c: the x-axis says this is the early 1950s whereas the legend says early 1910s. Which is it? Also, please fit the limits on the x-axis to the beginning and ending of the record you are plotting (also valid in Figure 4).

Figure 4: please enlarge the figure.

“Glaciermelt”: consider writing “glacier melt” instead.

Line 203-207: yes, but it would be great to try to nail the scope in the introduction instead of just speculating in the discussion. How applicable is this data to other places? That can be in principle deduced from the methodological setup alone.

This paper introduces a methodology and analysis of estimating reservoir operations across 74 different catchments in the Alps using general additive models. The goals were to identify groups of catchments with similar reservoir operations and assess how catchments differ with diff operation differ in location and catchment characteristics (how do different operations affect locations and areas). The approach of using GAMS was indeed a really interesting idea, however, I found the manuscript to be too vague and light on details to be confident in the results or for the methods to be reproduced.  

Major Edits by Section:

Methods:

• General Data Section
  • Can correlations be made if we are not looking at the same time periods? Can this be expanded on?

• Climate data was from gridded dataset, averaged over the full time period, moving time window to and replaced NA with mean flow
  • Does it make sense to replace NA with mean flow? If you want to keep mean flow, perhaps filling with monthly averages would be better.
  • How was the disaggregation done?
  • What was done to normalize or standardize the other data used

• GAMs Section
  • I’m not sure subtracting the mean standardizes your streamflow timeseries. Perhaps, it would be better to normalize by subtracting the mean and dividing by the standard deviation
  • Do you test for outliers in the Streamflow data? The peak in Figure 5 a around 2000 could be a result of an extreme weather event or an outlier in the observed data. What do you do to fix these if they occur, or do you simply assume that by normalizing you remove all the outliers?
    • Is it observed – natural for your comparison? If it is not, then the phrasing around line 140 needs to be changed.

• Signal Variation Analysis (2.3)
  • I think this section would benefit from a graphic explaining the workflow or at least demonstrating how the clustering is going as this felt like the most information rich section with a lot of steps.
  • I would also add more details on the previous Brunner et al., 2020 paper.
• My main concern is that figures 3,4,5 only focus on the single catchment and not all the catchments. Would it be possible to include all the catchments.
• Figure 1: I would zoom out a bit from the map so we can see the full Rhine and be more oriented in the catchments, Also what are the differences between the purple and black outlines? Denote that in the figure caption
  • Also, may be useful to put a map of the reservoirs so we can see where they are spatially located 

• Figure 5: I would shade the release periods vs storage periods in panel c so that the reader has an easier visualization for what periods are storage and which ones are releases.

Results:

• This section felt very choppy as there was one figure then a few lines of text. I think this could be more clear by grouping the results and creating panel plots. I do like that this section had all the catchments on it
• Figure 6: I’m not sure this figure adds too much to the discussion. Perhaps you could panel your reservoirs by bigger basin or by similar characteristics ( ie peaks in summer, peaks in winter). Another option is to color them by region or use, although I do think paneling or grouping would be useful to the viewer.
  • Also, after seeing Figure 7, I think you can cut figure 6 and use that space to add a graphic about the workflow in 2.3
  • I would also add a legend to this plot to denote what green and blue mean. Additionally, I would pick more colorblind friendly colors to be more inclusive.

• Figure 8 should have a map of the larger area so we can situate ourselves a little better

• You could combine figure 9 and 8 into one panel so this section doesn’t feel as choppy 
• Figure 10’s results tie directly into the map created in Figure 8. I would definitely panel some of these plots (Figure 8,9,10) in order to make this section flow smoothly and not feel as choppy. I would also reorganize the results about figure 10 and the reservoir location to be next to the map.

Discussion:

• I do not feel that you did a strong job of linking catchment elevation to reservoir operations, Perhaps that is due in part to the shorter results section. I think main use is a bigger takeaway as you specifically state that higher elevation reservoirs are more hydropower vs lower elevation being water supply

Minor Edits:

• Figure 2: I would add a dashed line when the reservoir came online so we have a better idea of when those changes occurred.
• Figure 4: the grey on the natural flow is hard to see. Perhaps using a dashed line or something like that would be useful?
• Figure 3: I really liked this figure!

Notes from HESS specific questions:

1. Does the paper address relevant scientific questions within the scope of HESS?
1. Yes. It looks at how humans have impacted the hydrology of certain regions by deriving reservoir operations from generalized models
2. Does the paper present novel concepts, ideas, tools, or data?
1. I believe it does, but I think the authors could emphasize why this is so novel.
3. Are substantial conclusions reached?
1. I personally felt that the conclusions reached could have been more direct
4. Does the title clearly reflect the contents of the paper?
1. I think there could be a more informative title (something like “Deriving reservoir operations from streamflow using GAMS), because you make the case in your introduction that your main takeaway is the ability to derive reservoir operations directly from streamflow, climate, etc