Flight calibration, energy budgets and uncertainty#

This workstream connects measured electrical channels, chronological operating forecasts and explicit uncertainty. It adds a pinned BUPT-1 source, reusable linear/Bayesian calibration, causal multivariate lag baselines, joint V/I block-conformal envelopes and measured energy reconciliation. These are transparent research baselines with reproducible examples, not flight-qualified models or full reproductions of the referenced papers.

Measurement, expectation and forecast#

Keep three different questions explicit:

Different model purposes#

Quantity

Question

Required evidence

Measured operating power

What voltage times current was observed at this boundary?

Channel units, clocks and electrical boundary

Healthy expected power

What would an independently specified healthy system produce?

Health evidence, operating constraints and appropriate physics

Operating forecast

What measured voltage/current is likely to follow this history?

Causal inputs and chronological validation

An autoregressive operating forecast can learn persistent underperformance. It therefore cannot automatically serve as a healthy reference for the PV loss budgets and operating diagnostics ledger. fit_linear_calibration and fit_bayesian_linear_calibration require an explicit purpose. The healthy_reference option also requires a description of independent health evidence. This declaration documents intent; it cannot verify the truth of caller-supplied evidence.

BUPT-1 / SatelliteCOTS connector#

The authors’ public repository supplies platform telemetry from BUPT-1, a 12U CubeSat carrying Atlas 200 DK and Raspberry Pi 4B computing payloads. OrbitalPV pins the following original files to commit 951b41521351d535b7c2354916d9c4991602e8c7 and verifies their SHA-256 values:

  • CommonData-Telemetries/telemetry_all.csv.zip: approximately 81 MB, separate MPPT1/2 and battery1/2 channels.

  • Energy-ShortTerm/telemetries/all.csv: group-multiplexed short-term data.

import orbitalpv as opv

archive = "/tmp/orbitalpv-bupt1/telemetry_all.csv.zip"
with opv.SatelliteCOTSConnector(timeout_s=120) as connector:
    connector.download(archive)

frame = opv.load_bupt1_telemetry(
    archive,
    time_min="2023-05-05 09:00:00",
    time_max="2023-05-05 15:00:00",
)

The source has no explicit redistribution licence, so raw BUPT-1 files are downloaded into the caller’s cache rather than bundled in the repository. The connector supports offline checksum verification and streaming reads of the ZIP member, without extracting archive paths to disk. Its raw provenance records the pinned commit, source URL, hash and retrieval time.

The source clock has no published timezone declaration in the checked schema. The default loader preserves naive source-local times. Relative elapsed-time energy accounting remains possible; absolute orbit/event alignment requires a separately established timezone. Supply source_timezone to convert that caller-established timezone to UTC, and then supply aware time bounds. The loader never silently assumes UTC.

Voltage is normalized from mV to V and current from mA to A. The result keeps:

  • pv_input_power_w: sum of observed MPPT input voltage times input current;

  • solar_bus_power_w: observed MPPT output currents times the observed bus voltage, following the authors’ calculation;

  • bus_power_w: observed total voltage times total current;

  • battery_source_signed_power_w: source-signed battery V times I;

  • separate branch currents, component temperatures and original clocks/rows.

These quantities do not establish unconstrained maximum-power availability. They also do not establish that every multiplexed sensor was sampled at the same physical instant. Common-file rows preserve the publisher’s alignment. The short-term group-multiplexed normalizer exposes group-specific powers and does not invent a simultaneous aggregate. Zero temperature values, signed sensor offsets and duplicate clocks remain visible.

Measured energy reconciliation#

budget = opv.measured_energy_budget(
    frame,
    battery_column="battery_source_signed_power_w",
    battery_discharge_sign=-1,
    max_gap_s=2,
)
print(budget.summary)

The sign above follows the original authors’ plotting convention. Verify it before applying the same assumption to other hardware or channels. measured_energy_budget integrates trapezoids only where all supplied channels have finite endpoint values and the interval is shorter than the declared maximum gap. It rejects duplicate or unsorted clocks. All terms use the same accepted intervals:

\[E_{residual} = E_{solar,bus} + E_{battery,discharge} - E_{load,bus}.\]

The residual remains unexplained. It can include sensor timing, calibration, unobserved electrical branches and boundary mismatch, so it is not automatically a physical conversion loss. The returned interval table and summary expose accepted duration, excluded intervals and coverage. Coverage is telemetry coverage, not spacecraft availability.

The reproducible six-hour BUPT-1 analysis reports 21,598 native records from 09:00 to 15:00 in the source-local clock on 5 May 2023. Median cadence is one second; two-second intervals are accepted by the declared integration cutoff. Over 5.99972 observed hours, solar bus energy is 113.200 Wh, battery discharge under the source convention is 27.871 Wh, and bus-load energy is 104.597 Wh. The remaining balance is +36.474 Wh. Separately observed PV-side input energy is 129.783 Wh; its difference from bus output is not assigned to a specific loss cause. The analysis also estimates compute-load power from measured branch currents and explicitly stated nominal rails, distinguishing that estimate from independently measured branch power. Nonzero computing currents occur in 7,349 records; the nominal-rail compute-energy estimate is 22.580 Wh. The sizable positive balance is retained as a limitation of the channel alignment/boundary interpretation, not presented as energy creation.

See BUPT-1 measured electrical energy budget and analysis/bupt1_energy_budget. The offline gallery uses explicitly labelled precomputed measured aggregates, while the analysis reproduces them from the pinned download.

Linear calibration and the Gaussian posterior#

The deterministic model uses training-only feature means/scales and an unpenalized intercept. An optional ridge penalty operates in standardized feature coordinates. The Bayesian alternative uses a normal coefficient prior and a known, caller-supplied Gaussian observation noise:

\[y = X\theta + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2 I),\qquad \theta \sim \mathcal{N}(0,\tau^2 I),\]
\[\Sigma = (X^T X/\sigma^2 + I/\tau^2)^{-1},\qquad \mu = \Sigma X^T y/\sigma^2.\]

The implementation uses linear solves. Predictive variance is x @ covariance @ x + noise_variance; latent-mean variance excludes the noise term. Noise/prior values must come from training data or independent evidence. Missing physics, non-Gaussian noise and future regime changes are not covered merely by drawing a Gaussian interval.

Timestamped predictions are checked against the last training label. Omitting timestamps permits in-sample algebraic inspection; evaluations should supply them. See Healthy reference calibration with an explicit Gaussian posterior for a clearly labelled synthetic healthy-reference experiment.

Causal multi-horizon forecasting#

make_lagged_vi_data constructs features using only origin/past samples and targets using positive future horizons. It checks the declared native cadence over each complete lag/target span. It neither fills gaps nor changes sampling frequency. Features can include additional measured variables, provided they were available at the prediction origin.

chronological_forecast_split places training, calibration and evaluation into separate clock intervals. A window is discarded when any of its future labels crosses the training or calibration boundary. Historical lag values can precede a partition boundary because they were already observable.

Three baselines are available:

  • persistence_vi_forecast repeats the last observed V/I;

  • seasonal_lag_vi_forecast uses explicit earlier lag values; a sample lag is not a mission-validated orbital period;

  • fit_linear_vi_forecaster fits a direct multivariate, multi-horizon ridge model and keeps the latest training label in its provenance.

Forecast arrays have axes sample, horizon, (voltage, current). Their power forecast is always the product of their V and I forecast. These baselines do not assert physical healthy operation, infer faults or recover attitude.

Joint V/I conformal envelopes#

calibrate_block_conformal_vi freezes residual scales selected on training data or independent evidence. A calibration score is the maximum normalized absolute residual across voltage, current, all horizons and every sample in a nonoverlapping block of consecutive records. An incomplete trailing block is excluded from scores, while its target times remain inside the embargo.

For n calibration blocks and requested coverage c, the quantile is the ceil((n + 1) * c) order statistic. If that rank exceeds n, it returns infinity rather than claiming an unsupported finite interval.

The V/I bounds define a rectangle. Power bounds are the minimum and maximum of its four corner products, which remains correct when intervals include negative signed sensor values. No independently predicted power band can contradict the corresponding V/I band.

Finite-sample conformal coverage requires exchangeable calibration/test block scores. Real spacecraft time series, overlapping forecast targets, eclipse transitions and operating-mode changes can violate that assumption. A chronological split does not prove independence. Report empirical simultaneous coverage and interval width on untouched later observations.

Real-data benchmark and current limitation#

The Flying Laptop example uses 8,599 actual right-array V/I pairs across 1-3 February 2018, with a CC BY 4.0 snapshot and SHA-256 provenance. It assigns one UTC day each to fit, calibration and evaluation. Its long seasonal lag requires continuous source windows, leaving 1,395 training, 1,533 calibration and 1,578 evaluation windows; the exclusions are reported.

At approximately 30 s, persistence has 1.91 W power MAE and ridge has 2.12 W. At approximately five minutes, ridge has 7.08 W MAE and persistence has 10.61 W. No claim is made that the more complex baseline wins at every horizon.

The global block-conformal envelope is currently very conservative: joint power coverage is approximately 99.87%, but mean derived power interval width is approximately 575 W. Those broad bounds are not useful control precision. They are retained visibly as a baseline limitation, motivating better operating-state features and validated conditional calibration. No evaluation measurements are used to fit coefficients or residual scales.

See Real flight power forecasts and coherent uncertainty and analysis/flight_power_uncertainty/reference_results.json for the full metrics and provenance. The snapshot predicts measured operating power, not available array MPP or causal loss categories.

Primary sources#