Pyomo TCLab v6: optimization proof of concept

[adowling2]

This pull request tracks the development of a proof-of-concept for optimizing the GP model for the TCLab in Pyomo.

Branch tclab-notebook-v5 contains v5 of the modeling notebook developed by @jakekrell. In other words, this branch is a "checkpoint" for the state of the repo when v6 development started.

Branch pyomo-tclab2 contains v6 development. My apologies for the confusing branch names.

[adowling2]

My first step was to copy ...v5.ipynb as ...v6.ipynb and then rerun the notebook on my development computer. I got this error message:

[adowling2]

I confirmed the training for FoKL worked. I am going to not worry about the numeric integration error for now. I am also going to temporarily remove the lag for Q. This will make the first Pyomo implementation much easier.

[adowling2]

@jakekrell A few updates:

• From the perspective of optimizing this model in Pyomo, there are some advantages to using the no-lag version of the GP for the initial proof-of-concept. I tried to get the no-lag version working (see ..._v6_no_lag.ipynb) but was unsuccessful.
• Next, I tried modeling the step test data, which because it is a constant Q, the lag does not matter. The smoothing did not work. See ..._v6_step_test.ipynb
• This got me thinking... what if the linear system model is reasonable but the discrepancy we are seeing is due to autocorrelated errors. In other words, the errors are correlated with themselves. We should discuss this during our next meeting.

My next steps:

• Revert to the optimal lag. I think it is less effort to deal with the complications from the lag model in Pyomo.
• Prototype optimal control problem in Pyomo. (Note to self: just implement forward Euler manually for now. We can worry about Pyomo.DAE later.)

[jacobpkrell]

My first step was to copy ...v5.ipynb as ...v6.ipynb and then rerun the notebook on my development computer. I got this error message:

@adowling2 Upgrade FoKL to latest release (v3.3.0) and error should resolve.

pip install FoKL --upgrade

[jacobpkrell]

@jakekrell A few updates:

• From the perspective of optimizing this model in Pyomo, there are some advantages to using the no-lag version of the GP for the initial proof-of-concept. I tried to get the no-lag version working (see ..._v6_no_lag.ipynb) but was unsuccessful.
• Next, I tried modeling the step test data, which because it is a constant Q, the lag does not matter. The smoothing did not work. See ..._v6_step_test.ipynb
• This got me thinking... what if the linear system model is reasonable but the discrepancy we are seeing is due to autocorrelated errors. In other words, the errors are correlated with themselves. We should discuss this during our next meeting.

My next steps:

• Revert to the optimal lag. I think it is less effort to deal with the complications from the lag model in Pyomo.
• Prototype optimal control problem in Pyomo. (Note to self: just implement forward Euler manually for now. We can worry about Pyomo.DAE later.)

@adowling2 On second thought the lag was effectively yielding np.cos(Q1). If we remove lag but train the GP (of the derivative) on the cosine (which is expected now looking at the dQ1dt term in the v4.1 chain rule equation) then we should be able to get both advantages --- more accurate GP and easier implementation in Pyomo.

[adowling2]

@jakekrell What version of Python are you using?

[adowling2]

@jakekrell pyomo_tclab_v6_optimize1.ipynb is ready for you to take a look. The last cell has a syntax error because the differential equation needs to be updated with the GP model in the RHS.