Configuration¶

Parameters and bounds¶

create_default_bounds() returns bounds for 41 ADM1 parameters. A soft bound is a weighted penalty the optimizer may cross, a hard bound makes the candidate infeasible.

Parameter	Meaning	Unit	Bounds	Default	Type
`k_dis`	Disintegration rate	1/d	0.1 - 1.0	0.5	soft
`k_hyd_ch`	Hydrolysis, carbohydrates	1/d	1.0 - 15.0	4.0	soft
`k_hyd_pr`	Hydrolysis, proteins	1/d	1.0 - 15.0	4.0	soft
`k_hyd_li`	Hydrolysis, lipids	1/d	1.0 - 15.0	4.0	soft
`k_m_ac`	Max. uptake rate, acetate	1/d	4.0 - 12.0	8.0	soft
`Y_su`	Yield, sugar degraders	kg COD/kg COD	0.05 - 0.15	0.1	hard

from pyadm1ode_calibration import create_default_bounds

bounds = create_default_bounds()
print(bounds.get_bounds("k_dis"))

Which parameters belong in a run is a question the data answers, not the modeller. calibrate(..., check_identifiability=True) screens the set before the optimizer starts and warns when the data cannot resolve it, at the cost of 2 * len(parameters) + 1 simulations. See Calibration.

Pass narrower ranges per run instead of editing the defaults:

result = calibrator.calibrate(
    measurements=measurements,
    parameters=["k_dis", "k_hyd_ch"],
    bounds={"k_dis": (0.3, 0.8), "k_hyd_ch": (2.0, 8.0)},
)

Optimizers¶

Pass one via method=. Hyphens, spaces and case are normalised, so "Nelder-Mead" and "nelder_mead" are the same.

Name	Type	Use for
`differential_evolution`, `de`	global	initial calibration (default)
`particle_swarm`, `pso`	global	alternative to DE
`nelder_mead`, `nm`	local	online recalibration (default there)
`powell`	local	derivative-free, no bounds needed
`l_bfgs_b`	gradient	smooth objectives
`slsqp`	gradient	smooth objectives with constraints

Sizing matters more than the choice: differential evolution runs population_size x len(parameters) simulations per generation. With three parameters, population_size=10 and max_iterations=50 that came to 1217 simulations and 56 minutes, so budget roughly three seconds per simulated window.

Objectives¶

objectives = ["Q_ch4", "pH", "VFA"]
weights = {"Q_ch4": 0.7, "pH": 0.2, "VFA": 0.1}

Weights are normalised internally. A channel missing from the data is skipped rather than scored as zero.

Data sources¶

A plant schema declares its sources and channels in one YAML file. Copy configs/plants/_template.yaml and adapt it:

from pyadm1ode_calibration.io.loaders import MeasurementBuilder, PlantSchema

schema = PlantSchema.from_yaml("configs/plants/my_plant.yaml")
measurements = MeasurementBuilder(schema).build(variables=["Q_gas", "T_digester"])

Feeds that change over time¶

By default a calibration collapses the whole substrate feed series of a window into its mean and holds it constant. That is cheap and correct for a record at steady load, but it hides every load step and substrate change, and those are exactly what makes the kinetics identifiable. Pass time_varying_feed=True to apply the feeds step by step:

calibrator = InitialCalibrator(plant, time_varying_feed=True)

It costs roughly 20 % more run time, and it is available on InitialCalibrator, OnlineCalibrator, CalibrationValidator and PlantSimulator. A window whose feed really is constant takes the cheap path either way.

Database¶

from pyadm1ode_calibration import Database

db = Database(connection_string="postgresql://user:pw@localhost:5432/biogas")
db = Database.from_env()   # reads DB_HOST, DB_PORT, DB_NAME, DB_USER, DB_PASSWORD

Timestamps are stored as naive UTC. Use pyadm1ode_calibration.timeutils.utc_now() when writing your own rows.