Long-term simulation (multiple years) using a predifined current profile

[jhthompson12]

Hi!

This is a follow up to a previous question: "Modeling the operation of a cell from real world energy storage system data". Essentially, I have 6 years of current data from a real lithium ion NMC cell (LG Chem JH3) that I would like to run through some PyBaMM model / simulation to see how the model's discharge capacity degrades over time.

Since receiving help in that thread I have had success getting a model to run on short periods of data (hours to days) but as i try to extend my simulation I run into a predictable issue where the simulated cell's voltage creeps up over time until it hits the upper voltage limit which stops the simulation:

This growing error between modeled and real seems to mainly come from some self discharge / loss of voltage that the real cell experiences over time that the model does not express.

In the simulating-long-experiments notebook I found the starting_solution attribute from simulation.solve(starting_solution=sol) which seems like it could be useful here. But in that documentation it is used in the context of repeating a regular experiment many times over.

One way to address my issue could be to break my years of cycling data into more manageable simulations, for example daily, where I feed the last day's solution into the next. But critically I would need to reset the initial_soc to realign the model to the real voltage in each new simulation.

So, my questions are:

• is this possible? I guess I should just try providing both starting_solution and initial_soc to a simulation and see what happens?
• each day's simulation would have to update the current data that it is simulating. is it possible to pass on most of starting_solutions information while overwriting the current profile that should be used in the next simulation?

Thank you in advance!

[TomTranter]

Instead of updating the current function parameter you could use an experiment step for each day. https://docs.pybamm.org/en/stable/source/api/experiment/experiment_steps.html

[TomTranter]

So I think you want to first focus on getting good fits for your RPT cycles then look at fitting over time. You can adjust the capacity nominally by changing electrode height and width but looking at your pulses it seems like you haven't got a terribly good fit for the rest of the parameters which would be important to get better values for. You would want to get a good idea of the amount of active material and cyclable lithium and also the OCVs and exchange currents, particle sizes and diffusion coefficients for the cell at BOL at a bare minimum for an SPM model. If you are struggling fitting BOL there's very little chance this will work just feeding in current traces. A simpler first approach to model degradation would be to fit the model for each rpt and manually adjust the active material and lithium inventory to see what the actually cap fade is and see whether the rest of the characteristics like pulse resistance are also lining up, then you have an idea of which degradation mechanism might be dominating.

[jhthompson12]

Thanks, Tom!

Unfortunately, I think that fitting the model to my real system data kind of defeats the purpose of my exploration. Also, I not sure I could find all of that proprietary information about this particular cell without doing a tear down or testing in a lab setting.

One hack I just tried: instead of feeding my real system current profile to the simulations, use the real system's rack-level average cell voltage with pybamm.Experiment([pybamm.step.voltage(drive_cycle)])

This allowed me to make it through the 10 day simulation without getting tripped up by the model's voltage limits being reached. The resulting current profile from the simulation ends up looking pretty much exactly like the C-Rate that my real cell experiences.

This approach will hopefully be enough to allow me to simulate the many years of operation of my real system in PyBaMM to at least extract a degradation profile from the model.. but I'll have to see how this scales past just 10 days. At the moment each day's simulation is taking ~2 minutes. At that rate it would take around 3 days to run through the simulations for my data!

[lappemic]

@TomTranter is the from @jhthompson12 provided codesnippet the suggested way to simulate longer experiments? As stated in this discussion i fail to simulate longer experiments and it would really help me get this working.

[jhthompson12]

@lappemic, what I posted above did end up working for me, but I made some tweaks to the approach to make things work more smoothly. This is very similar is spirit, but changed slightly to take advantage of PyBaMM's experiment_step workflow.

With this I was able to run a 7 year simulation at 10 second resolution broken into daily experiment steps (i.e. ~2500 "cycles") which took around 3 hours to run.

"""
This file attempts to address the issue identified in this discussion: https://github.com/pybamm-team/PyBaMM/discussions/4152#discussioncomment-9707173

Let's try to take the very long timeseries data that we would like to simulate and break it up into a smaller chain of
sequential simulations each feeding into the next. Hopefully using [experiment_steps](https://docs.pybamm.org/en/stable/source/api/experiment/experiment_steps.html)?
"""

import pybamm
import matplotlib as mpl
import pandas as pd  # needed to read the csv data file

pybamm.set_logging_level("NOTICE")
mpl.use('TkAgg')  # or can use 'TkAgg', whatever you have/prefer

LG_JH3_NOM_AH = 63

start = pd.to_datetime("2018-12-14 00:00:00")
end = pd.to_datetime("2018-12-17 00:00:00")

# import drive cycle data from file
all_real_rack_data = pd.read_pickle("sample_data_3_days.csv")

# Instantiate a simple model
model = pybamm.lithium_ion.SPM(options={"SEI": "reaction limited"})

# get predefined parameters for an NMC cell
param_mohtat = pybamm.ParameterValues("Mohtat2020")

# SoC / Voltage relationship data from real system
param_mohtat["Open-circuit voltage at 0% SOC [V]"] = 3.2
param_mohtat["Open-circuit voltage at 100% SOC [V]"] = 4.17
param_mohtat["Upper voltage cut-off [V]"] = 4.2
# nominal capacity for JH3 cells (and a few other parameters) from LG Chem documentation
# param_mohtat["Nominal cell capacity [A.h]"] = 63      # changing the nominal cell capacity will have little effect. You need to adjust electrode dimensions to change actual capacity.
# param_mohtat["Cell volume [m3]"] = 5.65e-04
param_mohtat["Lower voltage cut-off [V]"] = 3

output_variables = [
    'Voltage [V]',
    'Current [A]',
]

# create a separate experiment step for each day to create the chain of daily simulations
experiment_steps = []
all_days_sim_data = pd.DataFrame()
for day in pd.date_range(start=start, end=end, freq="d")[:-1]:
    # print("Simulating: ", day)
    day_start, day_end = day, day + pd.Timedelta(days=1)
    day_data = smaller_data.loc[day_start:day_end].copy()#.iloc[:-1]    # drop off last timestamp measure which will be first of next day

    # create a seconds column for just this day
    day_data.loc[:, "Seconds"] = (day_data.index - day_data.index[0]).total_seconds()

    # create this day's experiment step
    drive_cycle = day_data[["Seconds", "Container 1, Rack 1_DC.Voltage_Avg of Cells"]]
    drive_cycle = drive_cycle.to_numpy()
    experiment_steps.append(pybamm.step.voltage(drive_cycle))

experiment = pybamm.Experiment(experiment_steps)

# set up simulation - for drive cycles we recommend using the CasadiSolver in "fast" mode
solver = pybamm.CasadiSolver(mode="fast")   # was running into errors with this one, so leaving out for now
day_simulation = pybamm.Simulation(model, experiment=experiment, parameter_values=param_mohtat)#, solver=solver)

initial_soc = day_data['Container 1, Rack 1_StateOfCharge'].iloc[0] / 100
cycles_to_save = list(range(0, len(experiment_steps) + 1, 60))[1:] + [len(experiment_steps)]    # Save a cycle each 60 days including the first and last
select_sol = day_simulation.solve(initial_soc=initial_soc, showprogress=True, save_at_cycles=cycles_to_save)

pybamm.plot_summary_variables(select_sol)

Apologies if this script does not immediately run for you. I may have broken a couple things while trying to clean it up for posting, but the ideas within worked for me.

[lappemic]

Perfect, thanks a lot @jhthompson12! This really helped! I adapted your logic and simulate now for batches of 100s. This seems to be better memorywise, it still increases and i am testing up to how many cycles i get. Thanks again for posting here!