Solar panels lose power as they heat up, and a growing set of controllers tries to manage that heat. The one closest to this question, a model-predictive controller (a controller that plans a few steps ahead using a forecast), tunes an optical filter to block heat-producing ultraviolet light. These controllers are evaluated as if they knew the true incoming sunlight exactly. In deployment, that sunlight usually comes from satellite estimates, and those estimates are known to overestimate irradiance when the sun is low in the sky: early morning, late afternoon, and winter. This proposal measures how much of a thermal controller's claimed benefit survives when its idealized sunlight input is replaced by realistic, sun-angle-biased satellite data. It then identifies the weakest property of the irradiance input that still preserves most of that benefit.
The method comes from cryptography. Security proofs in the random oracle model treat a hash function as a perfect random function. A productive line of work asks which specific idealized powers a proof actually uses, and whether a weaker, realizable property would suffice. We apply the same move to energy control. We break the "perfect irradiance input" into separable powers and remove them one at a time. The result is a sun-angle-resolved map of which idealizations the controller truly needs.
Core Claim
Predictive photovoltaic thermal control depends on the irradiance input in a way that varies with solar zenith angle, the angle between the sun and straight overhead. Near midday, modules run hot, the ultraviolet filter pays off, and satellite estimates are fairly accurate. At high zenith angles, satellite products overestimate irradiance by tens of watts per square meter. A controller fed those estimates will anticipate heat that never arrives. It will then filter light it should have kept, which is exactly when photons are already scarce. We claim that most of the gap between "oracle" and "realistic" controller performance comes from a single power, conditional unbiasedness at high zenith angle, rather than from forecast noise or forecast lead time. If that holds, a cheap zenith-angle-aware correction recovers most of the benefit, and the target for satellite product developers becomes precise.
Where This Sits
Three bodies of work motivate the gap. Model predictive control of grid-connected PV power generation system considering optimal MPPT control of PV modules builds a two-layer controller. Its module layer adjusts an ultraviolet cut-off filter, and its grid layer regulates the inverter. It is evaluated only under standard test conditions (1000 W/m², 25 °C) with step changes, so irradiance acts as a perfectly known input. We reuse this controller as the object of study.
Evaluation of Two Satellite Surface Solar Radiation Products in the Urban Region in Beijing, China validates Himawari-8 and MODIS irradiance against ground radiometers. It finds an hourly Himawari-8 root-mean-square error of 86.58 W/m², a bias of +51.79 W/m², and discrepancies that grow as the sun drops. It stops at irradiance and never follows the error downstream into module heating or control. SolarCube provides the forecasting side, with satellite-plus-ground data at 15-minute resolution. It likewise stops at irradiance.
Leveraging the satellite-based climate data record CLARA-A3 to understand the climatic trend regimes relevant for solar energy applications over Europe pairs irradiance trends with a warming "heat penalty." It flags high-zenith winter retrievals as unreliable. However, it works at monthly, 0.25° scales, where controller dynamics are invisible.
Cryptography supplies the method, not the application. Non Observability in the Random Oracle Model separates the oracle's powers, the ability to observe queries and the ability to program answers. It shows that some proofs need only one of them and some need both. Instantiability of RSA-OAEP under Chosen-Plaintext Attack shows that a much weaker, concrete property, $t$-wise independence, suffices for one security goal where a full oracle had been assumed. Our analogue is a decomposition of the irradiance oracle and a search for a weak sufficient property.
We deliberately do not build standard-model authenticated telemetry for solar sensors. Deploying the ROM-proven protocol of Efficient IoT User Authentication Protocol with Semi-Trusted Servers with SHA-256 is a strong practical default. Nothing in the file shows a real attack on such deployments, while the efficiency cost of leaving the model, documented in Practical Group Signatures without Random Oracles (about 2888-bit signatures), is concrete.
Research Program
We first build a simulation chain driven by real measurements. Ground radiometer records serve as the true irradiance: the BNU station's minute-level global, direct, and diffuse components, plus the 19 BSRN/SURFRAD stations bundled with SolarCube. We pair each timestamp with the corresponding Himawari-8 or GOES estimate and with SolarCube model forecasts. Air temperature comes from ERA5, as in the CLARA-A3 study.
Ground irradiance, zenith angle, and temperature then drive three linked models. The first is a spectral split whose ultraviolet share falls with air mass, meaning longer atmospheric paths at low sun remove more short-wavelength light. The second is a module thermal model. The third is a cell electrical model in which the diode leakage current grows roughly as $T^3\exp(-E_g/k_BT)$, so open-circuit voltage falls with temperature. Here $T$ is cell temperature, $E_g$ is the bandgap, and $k_B$ is Boltzmann's constant. We re-implement Zhao et al.'s module-level controller within this chain. Its internal model plans filter settings from a supplied irradiance trajectory, while the "plant" always experiences ground truth.
The oracle decomposition has three powers, each removed independently. The first is conditional unbiasedness: we inject the empirical zenith-binned bias of the satellite product. The second is precision: we inject zero-mean noise matched to the product's residual variance. The third is foresight: we replace the true future irradiance with a SolarCube forecast or with persistence. Our central quantity is the fraction of the mitigation benefit retained in each zenith-angle bin $\theta$:
$$R(\theta) = \frac{E_{\text{input}}(\theta) - E_{\text{none}}(\theta)}{E_{\text{oracle}}(\theta) - E_{\text{none}}(\theta)}$$
Here $E_{\text{oracle}}$, $E_{\text{input}}$, and $E_{\text{none}}$ are the simulated energy yields in that bin under three conditions: the controller with ground-truth input, the controller with the degraded input, and no thermal mitigation. A value of $R$ near 1 means the power being tested is not needed. Negative values mean the controller does worse than doing nothing.
We apply the same decomposition to passive cooling, with an analytic energy balance in place of a controller. Simple preparation of PVDF composite flexible film with transparent, self-cleaning and radiative cooling properties reports 5–6 °C of cooling and a 6% visible transmittance loss. We compute the film's net diurnal yield by zenith bin, assuming the reported loss is angle-independent. We then run a sensitivity sweep in which transmittance loss grows at oblique incidence, because the file reports no angle-dependent transmittance.
Baselines
The naive reference is no mitigation: the ultraviolet filter is fully open, the module runs uncontrolled, and there is no film. Every $R(\theta)$ is measured against it. Any realistic-input controller that falls below it at high zenith angle is a direct, reportable failure.
The strongest existing method is Zhao et al.'s module-level model-predictive controller run with oracle ground-truth irradiance. It is the only controller in the file that actively trades spectral light against heat. Zhao et al. report its effectiveness only under standard test conditions, where it reached a 92.36% model fit and a grid current distortion of 2.54% versus 4.82% for proportional-integral control. There is no reported result found on Lacuna for its energy yield under real diurnal irradiance or satellite-derived input. The oracle version therefore defines the ceiling, and the same controller with raw Himawari-8 or SolarCube input defines the realistic deployment.
The strongest cheap competitor is the same controller fed satellite input corrected by a zenith-binned lookup table fitted to ground data. This is the skeptic's "just bias-correct it" baseline. It is fitted on a training year and tested on held-out months and held-out stations, so that it is not tuned on its own test data.
We also run a proportional-integral controller as the classical reference. We include a persistence-forecast variant to isolate the foresight power.
Every condition shares the same stations, periods, thermal and electrical models, controller horizon, and weights. Each receives an identical tuning budget: a grid search over controller weights on the same validation month, selected per input type so that no variant is disadvantaged by weights tuned for another. Our hypothesis counts as confirmed if two things both hold. First, removing unbiasedness alone lowers $R$ at zenith angles above 70° by more than removing precision or foresight, with non-overlapping station-bootstrap intervals. Second, the lookup-table correction restores the high-zenith $R$ to within a few percent of the oracle.
Why not just…?
Why not just feed the controller the local ground radiometer and skip satellites entirely? Most modules, especially rooftop and distributed arrays, have no co-located radiometer. Even where one exists, it measures the present and cannot supply the future trajectory a predictive controller needs. The foresight ablation measures exactly this. If persistence from a ground sensor retains nearly all of the benefit, the practical conclusion is "install a pyranometer." We would report that as the outcome.
Why not just apply a zenith-angle bias-correction table and call it solved? This is the real threat, and the proposal is built so that it can win. The Beijing validation shows the bias is systematic and zenith-linked, which a table can capture. If the table closes the gap, the finding is that conditional unbiasedness is the single sufficient power. That result specifies what satellite products must guarantee for control use. The case where the table fails is also informative. Failure at held-out stations would mean the high-zenith error depends on local aerosol and three-dimensional cloud structure, which the Beijing paper and CLARA-A3 both point to as unresolved.
Why not just use a fixed temperature coefficient and ignore control? The no-mitigation baseline is that option. Zhao et al.'s transient results suggest active control helps under step changes. Whether it helps over real days is exactly what $E_{\text{oracle}} - E_{\text{none}}$ measures. If that difference is small, the controller is not worth deploying at all, regardless of the input it receives.
Why not just harden the telemetry cryptographically, since the random oracle model is the named weakness? No evidence in the file ties a practical attack to ROM-proven protocols instantiated with standard hashes. The standard-model replacements carry a measurable size and compute cost. Garbled data from honest sensors, which is what the zenith bias amounts to, is the documented failure mode. We therefore spend the effort there.
Decisive Checks
The first check tests the input: in the Beijing data, the injected zenith-binned bias must reproduce the reported +51.79 W/m² hourly figure when aggregated. The second tests the mechanism: $R(\theta)$ for the raw-satellite controller must fall off at high zenith angle. If it is flat, the core claim is wrong. The third is the decomposition itself, the ordering of the three single-power ablations at zenith angles above 70°, with bootstrap intervals across the 20 stations. The fourth tests generalization: the lookup-table correction is evaluated on stations it never saw. The fifth concerns passive cooling: the film's net yield per zenith bin, across the angular-transmittance sweep, should show whether its morning and evening losses offset its midday gains. That answer extends the single-condition result in Mao et al.
References
- Model predictive control of grid-connected PV power generation system considering optimal MPPT control of PV modules
- Evaluation of Two Satellite Surface Solar Radiation Products in the Urban Region in Beijing, China
- SolarCube: An Integrative Benchmark Dataset Harnessing Satellite and In-situ Observations for Large-scale Solar Energy Forecasting
- Leveraging the satellite-based climate data record CLARA-A3 to understand the climatic trend regimes relevant for solar energy applications over Europe
- Non Observability in the Random Oracle Model
- Instantiability of RSA-OAEP under Chosen-Plaintext Attack
- Efficient IoT User Authentication Protocol with Semi-Trusted Servers
- Practical Group Signatures without Random Oracles
- Simple preparation of PVDF composite flexible film with transparent, self-cleaning and radiative cooling properties