The authors proposed an integrated ensemble prediction approach based on hydrometeorological modeling and machine learning for streamflow forecasting over a cascade reservoir catchment. The performance of the prediction with different model settings is compared. Results show the potential of the integrated hydro-meteorological and machine learning approach. Some comments are provided as follows and need to be addressed before the potential publication of this study.  

Line 114: Please give the full name of CSSPv2.

Lines 145-146: The authors calibrate the model based on the runoff at each grid (instead of using the streamflow at the control station). What is the motivation or advantage of this model calibration?

Line 217, equation (2); Here “y>observation” is assigned “1”. In other words, no matter how high the simulation is (if higher than the observation), it will result in a low CRPS, right? By the way, is this commonly used in previous studies?

Lines 234-235: The simulation is better for the downstream. Is there any specific reason for this pattern?

Lines 276-278 (Figure 7): For the CRPS and RMSE for the lead time of seven days, there is a strong cycle in the performance. What is the reason for the strong cycle? In addition, from the RMSE, we see low RMSE values for the lead time within 1-day, and relatively high values after 1-day.  However, for the CRPS, such a pattern does not exist. In addition, for the lead time beyond 1-day, the variation of CRPS and RMSE does not seem to depend on the lead time (prediction performance generally degraded for longer lead time, right?). Please clarity.    

Figure 1. It is hard to read the station name in this figure. Please improve it.

Other comments:  Are reservoir data used in the simulation or incorporated in the model? 

Reservoirs represent an important but difficult issue for hydrological modelling. This paper presents a method for ensemble streamflow forecasting at the hourly timescale considering the effects of cascade reservoirs. The method makes use of TIGGE-ECMWF meteorological forecasts, CSSPv2 land surface model and LSTM deep learning model. Through the case study of a reservoir in China, the method is shown to reduce probabilistic and deterministic forecast errors. In general, the paper is well-written with results clearly presented.

There are five comments for further improvements of the paper.

First of all, more details on the contribution of this paper can be added. As is illustrated in the introduction, the proposed method is built upon the CSSPv2 land surface model (Yuan et al., 2018). What are the limitations of the previous model? Can the limitations be illustrated through some diagnostic plots? Such analysis would make the contribution of this paper more convincing.

Second, the method is demonstrated for one reservoir. In the meantime, the “study area” section illustrates that there are ten cascade reservoirs in the Hongshuihe hydropower base. Is it possible to select another 2-3 reservoirs to show the robustness of the proposed method? It is noted that the additional case study reservoirs can be elsewhere and are not necessarily located in the Hongshuihe region.

Third, Figure 8 presents an interesting illustration of the time lag between Longtan outflow and Yantan inflow. This lag is largely due to the flowing distance between the two reservoirs. Meanwhile, the section of methods does not tell how the river flow is considered in the method. Is it performed by routing or hydro-dynamic simulation? How are the parameters determined?

Fourth, lead time plays an important part in forecast verification as forecast skill tends to decrease with the increase lead time. Meanwhile, the simulations shown in Figures 4 and 5 seem to have nothing to do with lead time. Please present some plots of ensemble forecasts at different lead times.

Fifth, CRPS in Figure 7 exhibits some diurnal circle that can relates to the diurnal circle of reservoir inflow/outflow in Figure 8. This result may be due to the setting of the LSTM deep learning model. When preparing streamflow data for LSTM, has the mean been subtracted? Are alternative settings, e.g., subtracting the mean or not, tested for LSTM?

Below are a few minor comments.

1. The location map can be improved by illustrating all the reservoirs in the Hongshuihe hydropower base. In addition, the location of the Hongshuihe hydropower base in China can be presented by using an inset plot.
2. In Table 1, please illustrate the year/month range and time step for the hydrological dataset.

The manuscript "Ensemble streamflow forecasting over a cascade reservoir catchment with integrated hydrometeorological modeling and machine learning" integrate the meteorological forecast, land surface hydrological model simulation and machine learning to forecast hourly streamflow over the Yantan catchment, and the results show that the flood forecast has been significantly improved. This work is very meaningful and the paper has been well-written. I therefore recommend this paper resubmitted after minor revisions. My comments are listed as follows:

1. The number of ESP members is 40, while the number of members used fromTIGGE-ECMWF is 51, whichis not the same as ESP, Will the evaluation resultsover predicted due to the number of members used from TIGGE is larger?
2. The upstream basin streamflow are used as CSSPv2 model inputs to provide the upstream inflow information. In the hindcast experiments, the upstream outflow inputsused are forecasts or observations? When lacking upstream outflow prediction, how does the system operate?
3. Is it possible to forecast the rainfall-streamflow using meteorological forcing from a closed watershed controlled by the Yantan station and correct it by LSTM instead?
4. The upstream inputs are also essential to this forecast system, hope to see some evaluation about this element.
5. The calibration results evaluated by NSE shows a worse result in the upstream grids than downstream ones. Is it possible to improve the calibration results by increasing the Iteration times set in the SCE-UA methods?

This is an interesting manuscript that is well structured and presented. The topic is worthy of publication in HESS. However, before publication please address the following comments:

1.    Line 201: What is the full name of GR4J? Please check throughout (e.g., TIGGE, ECMWF, CMA, SCE-UA).

2.    Lines 179-182: Please list the values or ranges of the calibrated parameters in a table.

3.    Line 189: How to perform the bias correction of the TIGGE-ECMWF forecast forcing? 

4.    Line 190: Please give a detailed description of the workflow in Fig. 2.

5.    Line 204: Please check the equation number.

6.    Line 205: Please clarify the input and output variables here.

7.    Lines 215: The meaning of the F(y) in Eq. 1 is not clear. Also for ref in Eq. 2. Please clarify the meaning of the variables in each equation.

8.    Line 291: Is the diurnal cycle in Fig. 8 the climatology averaged result? If so, please give the ranges to indicate the uncertainty. It also looks a bit strange that the rainfall reaches its maximum in the early morning rather than the afternoon in summer. Please clarify.

9.    Lines 298-301: Why would the combination of heavy rainfall and the decrease of upstream flow make the hydrological model performance worse? Does the model consider the reservoir?

10.    Line 328: Please add some implications of this work and as well as its deficiencies that need to be improved in the last section.

11.    Add the units in Figs. 6 and 7.