Water quality

[4flixt]

Some questions regarding water quality:

• What is a good limit for the water quality?
• How is water quality computed, when two flows of different quality merge? It looks like the resulting flow takes the lower quality (higher value). Is there not some kind of mixing rule?

[acominola]

• It is hard to say a limit, because there is not clear effect between the water retention time and possible effects on health, etc. The water aging is a very simple option to embed water quality in our problem. While it is not clear what the "tolerance" on it can be, there are some useful estimates of water ageing period from some US water utilities. You can see here (Table 3 and below): https://www.epa.gov/sites/production/files/2015-09/documents/2007_05_18_disinfection_tcr_whitepaper_tcr_waterdistribution.pdf

• I am pretty sure that for the second point, EPANET takes the lower quality for the resulting flow... I think if we want to make something more detailed on the quality, it is rather better to look at the behavior of some chemicals, e.g., chlorine, using another option rather than the 'AGE'

[4flixt]

Thank you andrea! It took me way to long to understand that the unit of the age quality is just the age in seconds haha. I think, also considering your second point, we might want to switch to another marker to determine water quality or solely consider pressure and tank levels. My issue with the age (and especially the mixing rule) is that

• the property is extremely non-continuous and (as discussed before) will have a limit that is solely determined by simulation time.
• We have an initialization problem. Dead-end nodes without demand will have a great quality (as all other nodes) at the beginning of the simulation and will worsen over time. The quality at the beginning is therefore not a realistic representation of their water quality.
• The clustering is suffering because we have an extreme range in the properties. This results in the effect that all nodes with low to mid-range qualities tend to end up in the same cluster regardless of the actual profile. This can be seen in the image below: Nodes with overall worse quality values are getting clustered as desired, however: In general clustering for the quality remains not ideal: The covariance for the individual clusters can be worse than covariance over all nodes. At least, however, the cluster covariances do not increase over time.

[sergiolucia]

I shortly discussed with Andrea today:

• We agree that in general it would be great to have a very simple chlorine model. This will probably solve the problem of simulation-time-dependant values
• Anyway in the case of age we could try to do the clustering for the pressure and just try to minimize the quality -- without specific bounds. Since the learning would not depend on the values themselves (0 order model) this might be enough?

[4flixt]

Ok so we wouldn't implement constraints but rather try to minimize the quality - that should work. In that case I would suggest to not learn the cluster mean+variance but instead the cluster max value. It also makes for a sleeker NN architecture. I‘ll still have to see how well the clustering generalizes, though. So far I have only been evaluating it with the training data set.

In any case I‘ve thought that its beneficial to learn two separate networks for pressure and quality.

• they are unrelated quantities
• evaluating two smaller networks will be faster than one large one (especially for the optimizer)
• if quality doesnt work we still have pressure control
• quality models can be adapted in the future

I‘ll of course start with the pressure, such that we have more time making adaptions to the quality model etc.

Sergio Lucia notifications@github.com schrieb am Mo. 4. Nov. 2019 um 15:30:

I shortly discussed with Andrea today:

• We agree that in general it would be great to have a very simple chlorine model. This will probably solve the problem of simulation-time-dependant values
• Anyway in the case of age we could try to do the clustering for the pressure and just try to minimize the quality -- without specific bounds. Since the learning would not depend on the values themselves (0 order model) this might be enough?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/4flixt/2019_WNTR_Surrogate_Model/issues/6?email_source=notifications&email_token=AH2XOC6EFODGIQ25KFRW6FDQSAWXTA5CNFSM4JIMLBDKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEC7N2TQ#issuecomment-549379406, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH2XOC2YAMTTP652B45KRCLQSAWXTANCNFSM4JIMLBDA .

[sergiolucia]

Yes, I agree that we should learn two separate models. One could even learn a third one for the levels of the tanks (which are the actual states and we should also learn...)

[4flixt]

Yes I have them in mind. These levels will effect water quality and pressure and we should have them as inputs to both models. I thought to have them as outputs of the pressure model (it is anyway just three states).

Sergio Lucia notifications@github.com schrieb am Mo. 4. Nov. 2019 um 15:41:

Yes, I agree that we should learn two separate models. One could even learn a third one for the levels of the tanks (which are the actual states and we should also learn...)

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/4flixt/2019_WNTR_Surrogate_Model/issues/6?email_source=notifications&email_token=AH2XOC337AUBFFOB3OPQ5E3QSAYA5A5CNFSM4JIMLBDKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEC7PCBA#issuecomment-549384452, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH2XOC25RPSR3H7M2IHSF5DQSAYA5ANCNFSM4JIMLBDA .

[sergiolucia]

looks good to me!

[acominola]

I agree, too :)