Solar inverters, battery controllers and feeder sensors increasingly coordinate over ordinary networks. The first time two of them talk, they face an old problem. To set up an encrypted channel they must exchange keys, and if they do so over the same untrusted network, an active "man-in-the-middle" can quietly swap in its own keys and read or rewrite every later command. Strong ciphers do not help, because the attack happens before any cipher is in use. The usual fix is a central certificate authority. That authority is unavailable in exactly the moments that matter most: a microgrid cut off from the wider network during a storm, a replaced inverter, or a device being re-keyed after a suspected compromise.
We propose that devices on the same low-voltage feeder authenticate each other by what they jointly see on the wire. Cloud transients on rooftop solar and household load switching make the local voltage fluctuate in a way that every device on that feeder measures almost identically. A remote attacker cannot observe or predict these fluctuations at the moment of the handshake. We turn a short window of these fluctuations into a low-entropy, slightly noisy shared "password." That password then authenticates an ordinary key exchange through a password-authenticated key exchange (PAKE), a protocol that turns even a weak shared password into a strong session key while letting an online attacker test only one guess per attempt. The intermittency that grid operators treat as a nuisance becomes evidence of physical co-location.
Core Claim
Our claim is narrower and easier to defend than "renewable noise as a secret key." A key derived from weather would be a poor secret. Irradiance is spatially correlated over kilometres, which is precisely what lets the spatiotemporal graph network of Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems cut two-hour forecast error from 18.98% to 11.01% mean absolute percentage error (MAPE) by sharing data between neighbouring plants. An eavesdropper with a nearby sensor would therefore learn much of such a key. Feeder voltage is different. It mixes irradiance with load events and line impedances that are local to one feeder. It also only has to resist an attacker for the few seconds of one handshake, and a PAKE makes low entropy sufficient against online guessing.
The quantity that matters is the unpredictable part of the fingerprint from the point of view of the best remote predictor. Writing $F$ for the quantized fingerprint and $Z$ for everything a remote attacker can observe (public weather, satellite imagery, regional PV telemetry, history), the per-attempt success probability of a man-in-the-middle is bounded by
$$ p_{\text{MITM}} \le 2^{-H_\infty(F \mid Z)}, $$
where $H_\infty(F \mid Z)$ is the conditional min-entropy, meaning minus the log of the attacker's best single-guess probability. Since PAKE allows one online guess per session, $q$ attempts succeed with probability at most $q \cdot 2^{-H_\infty(F\mid Z)}$.
This reframes the renewable forecasting literature as attacker modelling. Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance shows that subtracting the per-timestamp ensemble mean removes the predictable diurnal arc, and that an autoregressive model on the residual matches convolutional and recurrent networks. Its residual construction tells us which part of the signal to discard. Its forecast error, together with that of the graph network above, gives an empirical lower bound on what a remote attacker cannot predict. The bound is empirical because we can only test the attackers we build.
Threat Model and What Is Out of Scope
The adversary controls the communication network between two devices. That network can be cellular, IP, Wi-Fi or a power line modem link. The adversary also has real-time access to all public and regional data, including telemetry from adjacent feeders. It does not have a live, high-rate voltage sensor on the victim feeder itself. An attacker who has physically connected to the same feeder defeats the scheme.
We state this limitation plainly. The scheme raises the bar from "anyone on the internet" to "someone physically wired into this street's distribution line." Our premise is that this is the right bar for commissioning and islanded re-keying, not for every threat.
Research Program
The protocol has three parts. The first is fingerprint extraction. Each device samples its terminal voltage at the rate its existing metering supports, typically sub-second to a few seconds. Over a window of $T$ seconds, it subtracts a short-horizon forecast of the predictable component. Following the residual idea of the autoregressive paper, this removes the diurnal arc and slow trends. The device then quantizes the sign and magnitude of residual increments into bits.
The second part is noise tolerance. Two honest devices see slightly different residuals because they sit at different impedance distances along the line. They therefore use a public error-correction sketch tuned to the measured honest bit-disagreement rate, and we account for the entropy the sketch leaks. The third part is a standard PAKE run with the corrected fingerprint as password. This yields a full-strength session key, which can then anchor a certificate or a decentralized identity log. That log could be the claim-based record of ClaimChain: Improving the Security and Privacy of In-band Key Distribution for Messaging, which detects an attacker showing different keys to different peers.
The work beyond protocol design is measurement. The research file contains no dataset of sub-second voltage on a real distribution feeder. The hourly SAM-generated PV and load profiles behind the 20-house "fishbone" feeder of Optimal Dispatch of Residential Photovoltaic Inverters Under Forecasting Uncertainties are far too coarse for this purpose. We will therefore work at two levels.
In simulation, we run time-series AC power flow on the fishbone feeder and on the IEEE 33-bus system used by Multiagent reinforcement learning in enhancing resilience of microgrids under extreme weather events. We drive both with load profiles and PV injections, and we upsample the PV injections using residual statistics fitted to the 10-minute tropical irradiance data and to the 316-plant US dataset. The upsampling is an assumption that the hardware campaign must check.
In hardware, we instrument a small number of real low-voltage feeders, ideally two adjacent ones sharing a transformer region, with synchronized voltage loggers. This lets us measure how fingerprints correlate within a feeder and decorrelate across feeders.
For attackers, we train the strongest remote predictors we can. These are a graph-network forecaster in the style of the 316-plant model, given all adjacent-feeder voltages and regional PV, and the residual autoregressive model. We report $H_\infty(F\mid Z)$ against each.
Baselines
The naive reference point is unauthenticated in-band key exchange, which is what devices do when no infrastructure is present. It is trivially defeated by key substitution, so its man-in-the-middle success is one. It anchors the scale.
The strongest existing physical-layer competitor is power line key generation from Physical Layer Key Generation for Secure Power Line Communications. That method extracts shared bits from the reciprocal high-frequency channel between two power line modems, using either impulse-response peaks or estimated transmission matrices. On 1,312 in-home channels it reports eavesdropper correlation "often below 0.2" and transmission-matrix estimation error between −25 and −35 dB. The file reports no key-generation rate in bits per second, so we found no reported result on Lacuna for its key rate. Its own authors flag the rate as low in static channels.
We will run it fairly by computing, on that same 1,312-channel dataset, the number of reconciled bits each technique yields per handshake window, using their published reconciliation. We will compare this with the bits our fingerprint yields over the same window lengths on our feeders. Power line key generation targets secrecy, but we will evaluate both methods under our metric, which is bits usable as PAKE password entropy per handshake. We beat it if we match its per-handshake entropy with no power line modem hardware, or if we exceed it on overcast and high-load days.
The strongest infrastructure competitor for the coordination setting is a ledger-rooted key agreement. One example is the on-chain/off-chain scheme of Traceability and Identity Protection in Smart Agricultural IoT System Framework Based on Blockchains. The Ethereum edge deployment of Smart Microgrid Energy Market: Evaluating Distributed Ledger Technologies for Remote and Constrained Microgrid Deployments reports 513.4 ms average transaction latency on a Raspberry Pi 3. We compare handshake completion time and, critically, success when the ledger's peers are unreachable during islanding, on the same Raspberry Pi class hardware.
Two ablations isolate the key ingredient. The first replaces feeder voltage with grid frequency, which is nearly identical across a whole interconnection and therefore observable by a remote attacker. If our scheme works only because of local physics, frequency fingerprints should give near-zero conditional entropy against the regional attacker. The second skips the forecast-residual step and uses raw voltage, which should inflate apparent entropy under a naive estimator but collapse against the forecasting attacker.
Why not just…?
Why not just pre-load factory certificates from a central authority? Where a working certificate authority and provisioning pipeline exist, that is the right answer, and we do not claim to replace it. Decentralizing Public Key Infrastructures with ClaimChains argues that a central authority sees every lookup and becomes a bottleneck at scale. The more relevant failures for us are operational: field replacements, re-keying after compromise, and islanded microgrids that cannot reach the authority. Our experiments target those regimes. We will report honestly that in the connected, well-provisioned case our scheme adds latency for no security gain.
Why not just have the installer compare a short code on both devices? This is a genuine threat to the proposal for one-time commissioning, since human comparison of a short authentication string defeats remote man-in-the-middle attacks cheaply. Its weakness is that it requires a person at both devices at the same time. That rules out autonomous re-pairing among dozens of devices after an outage, and it is the step installers skip in practice. Our evaluation therefore focuses on unattended re-keying and reports how many devices per hour each approach can pair. The two can also be combined, since the physical fingerprint can stand in for the code when no human is present.
Why not just use power line channel key generation, which already exploits the wire? It needs measurement-grade power line transceivers at both ends, which most inverters lack. Its rate is limited because the in-home channel barely changes. The direct comparison on the 1,312-channel dataset described above settles whether its per-handshake entropy already suffices. If it does, our contribution reduces to doing the same with sensors that already exist on every inverter.
Why not just skip communication and rely on local droop control? For voltage support this is reasonable. However, the inverter dispatch study shows coordinated, risk-aware dispatch selecting the most critical inverters, and the graph forecaster shows a large accuracy gain from sharing data between neighbouring plants. Coordination has value, so its channels need securing.
Proposed Experiments
The first decisive check is discrimination. Devices on the same feeder must agree on fingerprints with a correctable bit-disagreement rate. Devices on the adjacent feeder, standing in for the best-placed remote attacker, must not. We report the honest false-reject rate and the cross-feeder false-accept rate as functions of window length $T$.
The second check is entropy against strong attackers. We estimate $H_\infty(F\mid Z)$ against the graph-network and residual-autoregressive predictors. The target is roughly 20 bits per handshake, which gives a one-in-a-million chance per online attempt under the bound above, within a window of a minute or less. We report the result separately for clear-sky, broken-cloud and overcast conditions.
We expect pairing to be fastest under broken cloud and slowest on clear, low-load nights. Measuring the time-to-pair distribution across weather is where the intermittency direction does real work. If clear-sky windows need many minutes, the scheme is still usable for commissioning but not for rapid re-keying, and we will say so.
The third check is end-to-end. In the 33-bus extreme-weather simulation with an islanding event, we compare the fraction of inverter-controller pairs that re-establish authenticated channels within five minutes under four conditions: our scheme, ledger-rooted agreement, unreachable certificate authority, and unauthenticated exchange. We then inject a key-substitution attack that spoofs dispatch commands to measure the overvoltage it causes when it succeeds. This last measurement shows what the authentication actually protects.
The proposal fails cleanly in either of two cases. Adjacent-feeder fingerprints may prove nearly as correlated as same-feeder ones, or the forecasting attacker may drive conditional entropy below about 10 bits per minute. Either outcome would itself be a useful negative result on how much local information distribution voltage carries.
References
- Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems
- Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance
- ClaimChain: Improving the Security and Privacy of In-band Key Distribution for Messaging
- Optimal Dispatch of Residential Photovoltaic Inverters Under Forecasting Uncertainties
- Multiagent reinforcement learning in enhancing resilience of microgrids under extreme weather events
- Physical Layer Key Generation for Secure Power Line Communications
- Traceability and Identity Protection in Smart Agricultural IoT System Framework Based on Blockchains
- Smart Microgrid Energy Market: Evaluating Distributed Ledger Technologies for Remote and Constrained Microgrid Deployments
- ClaimChain: decentralized public key infrastructure