Genetic-Algorithm-Optimized Sequential Model for Water Temperature Prediction

Comment

Response

Line 14: Abstract: the abstract should include information concerning the study object.

The abstract is revised to include the objective of the study (see lines 20-23).

Line 54: The method was evidenced to generate errors in a range of 2‑3°C for daily temperature. It is an excessive simplification. The authors refer to a publication: J. Am. Water Resour. 377 Assoc. 1993, 29. The errors may be different in different regions.

Clarified that the error id dependent on location, added additional sources and clarified that the error is related to downscaling seasonal correlations to daily scale (see lines 60-61).

Line 73: The study employed recorded hourly data regarding water temperature from station Mississauga Golf and Country Club. A figure illustrating the location of the analysed river is necessary.

A figure showing the location of the MGCC station has been added (see figure 5).

No information on its parameters is provided, namely flow rate, distribution of water temperature in the annual, and multiannual cycle, discharges of pollutants, etc. This is key information for the application of the model in the case of other rivers or regions.

Additional information on the watershed characteristics has been provided (see lines 199-202).

Line 188: …“The  determination of these control parameter heavily depends upon the skill and experience of researchers”... It is a very subjective statement that may affect further results. Please elaborate on the issue.

What we meant was a general warning for the graduate students new to these methods to make sure to consult with more experienced modellers.

Line 284: Fig. 8. Please provide an interpretation of the cause of a greater divergence of lines (observed, RNN, GA-LSTM) in the middle period in comparison to the beginning or end.

In the middle period the temperatures are higher and therefore the error is proportionally larger.

How was water temperature measured during the persistence of ice phenomena? In spite of lack of such information in the text, I assume that due to the location of the study object, such phenomena did occur during winter. How did the model behave during that period?

The water temperature sensor was housed within a perforated pipe mounted to a bridge pier. The Credit River is 2 to 3 m deep in winter months at this station. The depth of the sensor was positioned such that the measurement probes were submerged during low flows. Over the winter this would normally be below any potential ice cover if it had occurred. This watershed is heavily urbanized and large amounts of road salt is used on roads and parking lots for winter de-icing operations. High chloride concentrations prevent the formation of ice in urban streams. The CVC data Quality and Validation procedures removed periods where the data was not reliable.

Comment

Response

Introduction starts with urban streams of Toronto Area, and the link of temperature control with protected species of fish. The authors discus different kind of models, and propose empirical based models for their purpose or predicting temperature values. Then, they propose the use of the GA-LSTM. The author mention that this method needs two parameters to be used, and this is done thanks to the GA. However, at this point, it is not clear which is the time scale of prediction of river temperature. Section two presents the modelling approach. After that, it gets clear that a GA is used to optimize the parameters of the LSTM based on previous data. However, prediction time is not clear yet. Finally, in the subsection 3.2, it explains that the forecast is for next temperature value (next hour).

The introduction has been expanded/improved and the time scale of prediction has been clarified in the (see lines 79 to 96).

As stated in lines 205-212, the proposal is calibrated, validated and tested, with parts of the same dataset, but not sequentially divided. Thus, “it works”, at least as well as other approaches (or even better), but the applicability is not clear at all, further than this specific case. Moreover, it is not clear how the uncertainty of the next hourly predicted value could be assessed.

This study presents the novel application of the GA-LSTM to urban river water to solve the problem of the iterative selection length of window size and number of LSTM units that is done here using genetic algorithm (see Table 1).

In any modeling process, the basic steps adopted to divided the dataset into three sets namely training, validation and testing.  First two steps used to train the model via tuning the model parameters and testing set is to check the performance of the models. Evaluating any models adopting this process is quite enough for the case study test the new development model. In this whole dataset is divided sequential into three sets. In this study already genetic algorithm is used to optimised the inputs considering the  cross validation approach is adopted hence there is no further need uncertainty analysis.

In line 317, at the end of the conclusions, the authors explain that this kind of algorithms could be useful in other areas of water quality monitoring. I agree, but it is not shown that it could be useful in practical terms in other situations or cases.

As mention in the above comment, this method adopted GA to optimize the LSTM and LSTM is one of the popular sequential modeling approach for time series data hence it can be used in the many fields of engineering application. This sentence is recommendation for future studies and indeed the topic of our next follow-up paper.

Lastly, the link with sustainability is marginal and almost implicit. It should be improved and send it for reviewing to another journal.

Real time network of river water quality monitoring is at the forefront of a proactive urban water management strategy. The preservation of the quality of our surface waters is at the core of the global challenge of urban water sustainability. The water temperature is a primary indicator of the health state of the aquatic habitat and its modeling is crucial for river water quality management (see lines 14-20).

Comment

Response

Line 74-77 Results should be presented in last chapter.

This information has been removed from the introduction and presented in the results section.

Line 231-248 This text should be in introduction or previous chapters not in discussion.

This section has been integrated into the introduction (see lines 79-96).

Figure 8 RNN and GA-LSTM results are very similar

Yes indeed; our conclusion is that both methods produce good results, however, the GA-LSTM is superior in terms of fitness parameters 

Furthermore GA-LSTM to helps in easy selection of window size and number of unit cell via GA.

Moreover, figure 7 clearly shows the difference in convergence of mean square error during the validation period. In RNN MAE does converge but shows a uneven course.

Figure 9 Authors should add values of mean and median on this drawing

We respectfully disagree and humbly suggest that adding values of mean and median on the drawing is not necessary and would make the figure crowded.

Comment

Response

In Fig 5. the unit on the vertical axis is missing

Corrected; thank you.

On line 220 in the equation on RSR there is no explanation for the abbreviation STDEV

The abbreviation has been defined added below line 225.

The formula for RMS was given twice (5) and (7)

The duplicated formula has been removed.

In Fig 6. in figures a and b should be the same vertical scale (the size of these figures should be identical).

We respectfully disagree and humbly suggest the figure is clear as is.

In Fig 8. there is no description of the units on the vertical axis.

The units for water temperature (WT) is degrees Celsius. This information is added to the figure caption.

Fig 8b. - if we show an example cut from the graph to enlarge the detail, the horizontal number should have the same number of counts (4100 - 4350 and not 0-250)!

The submitted paper made a good impression on me. It felicitously combines a extensive literature review with own analyzes, as well as a reasonable discussion of the presented results. Therefore, I have no strong remarks. The manuscript is well structured and deserves publication after some minor revisions..

Corrected; thank you.