General comments

The manuscript presents an interesting and relevant work addressing scientific questions within the scope of HESS. The authors present a successful newly implemented coupled system with the main objective of estimating the outflow from a reservoir in Spain. The methodology developed and applied merges the physical understanding of a physical-based model and the capabilities of artificial neural networks to learn from non-linear processes. Therefore, applying a coupled system highlights the novelty of the work, which is not simply another application of a hydrological model or an artificial neural network to estimate streamflow in a given catchment. In this sense, the authors gave proper credit to related work and clearly indicated their original contribution. The manuscript is well-structured, clear and concise. It is well-written and easy to understand. The authors present a good introduction of state of the art regarding the subject and a well-written description of the study area and the materials ad methods applied. The studies used for introducing the subject and the discussion are relevant and recent. The results presentation and discussion gave a clear and sufficient overview of the results and their respective limitations. The conclusions present an interesting closing of the presented workflow concisely and straightforwardly, but not forgetting to present the future perspectives from this work. Finally, the authors present the workflow code in a public and open repository, contributing significantly to open science and making it possible the work reproducibility. Therefore, the paper deserves to be published in HESS after some corrections and adaptations regarding the quality of the writing and figures.

Specific comments

In this section, I present some specific comments to be answered and implemented in the manuscript. No changes in the methodology are proposed, only suggestions for the writing, discussion and overall figures and tables quality.

1. The authors provided a complete public repository with the workflow. I recommend not going into too much detail within the text regarding the specific Python libraries and functions applied in this methodology.
2. Likewise, regarding the MOHID-Land model, since the model was already implemented in the referenced work of Oliveira et al. (2020) where it was fully described, the information regarding the implementation not already presented in this work should be given in the present details in Annex instead of in the main text. The authors can consider only a summary of the information instead of being in such details in the present work.
3. The authors conclude that the poor representation of reservoir levels reflects the non-inclusion of evaporative losses in the model (L529). I agree, but do the authors have concrete results corroborating this? Did you perform tests including it in your model? Would that be a possibility for further work? For this work, it is optional to make this inclusion. However, more evidence could improve your conclusions.
4. I would like to have a deeper discussion about why, even with the limitation in reproducing the reservoir levels, the model still performed well in reproducing the streamflow at the outlet. Is this due somehow to the concentration time of the watershed? A slight increase in the discussion will enrich your paper.
5. L361: What would cause this behavior by the model in your option? Could it be further discussed in the text?
6. L452: In this part of the discussion, the authors claim several limitations to using the CLSTM model. The text follows, describing several concerns, but are those the limitations intended by the authors to be mentioned here? I recommend rewriting this part text for clarification.
7. L459-461: The authors claim that the optimization of the CLSTM could improve the results. Why was this not tested for this work? I would appreciate a further discussion of this matter in the manuscript text.

Technical corrections

The complete list with the proposed technical corrections is attached in PDF. 

egusphere-2023-915 “Direct integration of reservoirs’ operations in a hydrological model for streamflow estimation: coupling a CLSTM model with MOHID-Land” AR.Oliveira, TB.Ramos, L.Pinto, R.Neves

The work presented in this article is very thorough. It is a valuable addition to the material from Oliveira et al. (2020) illustrating the use of AI/ML techniques, some of which were outlined in Oliveira et al. (2023) when applied to streamflow only. While there are a few peculiarities of wording through the text, the English expression for the most part is very good and it reads well.

For corrective suggestions, perhaps only the line figures showing flow need to be cleaned up (Figures 2 and 7-9). With daily instantaneous data and small dots connected by lines, the hydrograph becomes a red sludge with the occasional peak that is visually unsatisfactory. Perhaps weekly or monthly volumetric totals would be more distinct, or single years shown as examples of the best/worst fit for the particular solution. This does not mean losing any of the finer daily detail when reporting statistics or minimum and maximum daily flows, as with the current tables.

Figure 3 is also far too busy with three nearly overlapping lines in each panel. You may need a log scale on the y-axis, omitting all zeroes and the observed data (as the fit is very good).

Most of the questions in my mind are already listed in the Conclusions. Why does the MOHID-Land model not apply evaporation to the reservoirs? This is clearly a significant flux for these structures and part of their water balance. However, the ANN was not trained with reservoir level but only used as an input->output “black box” and it did not have a significant effect on performance. Given the “best” set of weights selected for validation testing, it would be interesting to see the same calculations with the set that best reflected the reservoir level.

This behaviour may be a case of getting a right answer for the wrong reasons (qv. Kirchner, 2006). Clearly it would be preferred that the open-source MOHID-Land model treated the reservoir as a water balance unit with all its attendant fluxes, and that calibration be a combination of flow at a downstream gauging station together with the changes in the significant upstream storages. In the case of this study, the outputs from Portodemouros (gauged?) are the boundary conditions for the river routing through the largely irrelevant Bandariz and Touro, to the calibration river gauging sites downstream. Thus the ANN only has to ensure that the outputs result in the correct upstream input to do their job. If the modelled flow inputs to Portodemouros are incidentally correlated to the outputs, given the internal storage is not modelled, then the ANN can include them but they may not even be relevant.

As for the addition of irrigation releases, for example, that becomes a very interesting problem with a mix of known storage release rules and the ANN version of other informal flows through the reservoir. It is almost a data integration exercise, where the storage is “managed” with the ANN model but updated periodically with the irrigation rules.

Kirchner, J. W. (2006) Getting the right answers for the right reasons: Linking measurements, analyses, and models to advance the science of hydrology. Water Resources Research, 42, W03S04, doi:10.1029/2005WR004362