Note
Go to the end to download the full example code.
Healthy reference calibration with an explicit Gaussian posterior#
This is a controlled synthetic experiment: an independently designated healthy training run calibrates a temperature-power reference. A later measurement sequence contains a deliberately injected offset. The fitted reference predicts healthy expectation; it is not trained to reproduce the degraded operating sequence.
The conjugate Bayesian linear calculation uses stated noise and prior assumptions. It is neither a neural digital twin nor a full reproduction of a mission paper. Credible/predictive intervals require the Gaussian linear assumptions; uncertainty from missing physics remains outside this model.
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import orbitalpv as opv
rng = np.random.default_rng(42)
temperature = np.linspace(10, 60, 80)
training_power = 100 - 0.4 * (temperature - 25) + rng.normal(0, 2, len(temperature))
training_times = pd.date_range("2026-01-01", periods=len(temperature), freq="min", tz="UTC")
reference = opv.fit_bayesian_linear_calibration(
temperature,
training_power,
timestamps=training_times,
purpose="healthy_reference",
evidence="Synthetic controlled healthy run; no flight validation claimed",
observation_noise_std=2.0,
prior_std=100,
)
deterministic = opv.fit_linear_calibration(
temperature,
training_power,
timestamps=training_times,
purpose="healthy_reference",
evidence="Same independently designated synthetic healthy training run",
)
evaluation_temperature = np.linspace(5, 100, 100)
evaluation_times = pd.date_range("2026-01-02", periods=100, freq="min", tz="UTC")
prediction = reference.predict(evaluation_temperature, timestamps=evaluation_times, coverage=0.95)
latent = reference.predict(
evaluation_temperature,
timestamps=evaluation_times,
coverage=0.95,
include_observation_noise=False,
)
actual = 100 - 0.4 * (evaluation_temperature - 25) + rng.normal(0, 2, 100)
actual[55:] -= 15
fig, axes = plt.subplots(1, 2, figsize=(12, 4.5), layout="constrained")
axes[0].plot(
temperature, training_power, ".", color="#202C39", label="Synthetic healthy calibration"
)
axes[0].fill_between(
evaluation_temperature,
prediction.lower,
prediction.upper,
color="#56B4E9",
alpha=0.35,
label="95% predictive interval",
)
axes[0].fill_between(
evaluation_temperature,
latent.lower,
latent.upper,
color="#0072B2",
alpha=0.4,
label="95% latent mean interval",
)
axes[0].plot(
evaluation_temperature,
deterministic.predict(evaluation_temperature)[:, 0],
"--",
color="#202C39",
label="Least-squares healthy reference",
)
axes[0].set(
xlabel="Cell temperature [C]",
ylabel="Synthetic reference power [W]",
title="Known-noise Bayesian linear calibration",
)
axes[0].legend(fontsize=7)
residual = prediction["mean"].to_numpy() - actual
axes[1].plot(np.arange(100), residual, ".", color="#D55E00")
axes[1].axvline(55, color="#202C39", linestyle="--", label="Injected synthetic offset begins")
axes[1].axhline(0, color="#202C39", linewidth=0.8)
axes[1].set(
xlabel="Later evaluation sample",
ylabel="Healthy reference - observation [W]",
title="A forecast of actual output would answer another question",
)
axes[1].legend(fontsize=8)
for ax in axes:
ax.grid(alpha=0.2)
plt.show()
print("Posterior coefficient covariance:", reference.coefficient_covariance)

Posterior coefficient covariance: [[ 4.99997500e-02 -8.04903644e-18]
[-8.04903644e-18 4.99997500e-02]]
Total running time of the script: (0 minutes 0.221 seconds)