Weather-Triggered Trojans in Neural Dynamic Line Rating

Power grids increasingly replace fixed, worst-case transmission limits with dynamic thermal ratings (DTR): real-time estimates of how much current an overhead line can safely carry given the actual weather cooling it. Because the physics-based calculations are hard to run at scale over hundreds of lines and volatile microclimates, operators are turning to deep neural forecasters — for example, the Line Graph Convolutional LSTM (LGCLSTM) of Probabilistic Dynamic Line Rating with Line Graph Convolutional LSTM, which predicts probabilistic line ratings on the synthetic Texas 123-bus system, and the Bayesian-optimized BP network of Dynamic prediction of overhead transmission line ampacity based on the BP neural network using Bayesian optimization. DTR pays off: field monitoring safely unlocked "up to a 250% increase in capacity compared to static limits" (Dynamic thermal rating for increasing network capacity and delaying network reinforcements), and distribution-level DTR raised solar hosting by "15% to 27%" (Improving distributed PV integration with dynamic thermal rating of power distribution equipment).

But this creates a new cyber-physical dependency with an unexamined security hole. Classical false data injection (FDI) attacks corrupt electrical measurements, and the grid's bad data detection (BDD) watches exactly those measurements against Kirchhoff's laws. Weather telemetry — wind speed, ambient temperature, solar radiation — enters the rating model outside the electrical residual, so nobody checks it. Worse, when the rating model itself is a deep network, it inherits the neural backdoor vulnerability: over-parameterized models hold spare capacity that can hide trigger-conditional logic while passing all clean validation (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks). No paper in the collected literature connects DTR to backdoor attacks at all; the two closest neighbors — Can Attackers with Limited Information Exploit Historical Data to Mount Successful False Data Injection Attacks on Power Systems? and Poisoning Attacks Against Data-Driven Predictive Control — both attack with real-time additive perturbations against static limits, never through a poisoned forecaster against dynamic ones.

Core Claim

The claim is that a neural dynamic line rating model can be trojaned so that a subtle, naturally occurring-looking pattern in weather telemetry — say, a particular combination of wind direction and temperature that recurs a few times per season — makes the model output inflated line ampacity, while its behavior on all other inputs is statistically indistinguishable from a clean model. A dispatcher running economic dispatch on these ratings will route real current past the line's true heat-dissipation capacity, causing physical conductor sagging and annealing, while every digital layer — BDD, dynamic watermarking, moving target defense — reports normal operation. The backdoor differs from every attack in the file in two ways: it corrupts the constraint envelope fed to the dispatch optimizer rather than the electrical measurements or control signals those defenses watch, and it is installed once at training time rather than injected continuously at runtime, so it requires no live sensor access when it fires.

The physics makes the attack surface concrete. A line's dynamic limit is set by the heat balance

$$I_{max}^2 R \;=\; q_c(v_{wind}, \theta, T_a) + q_r(T_s, T_a) - q_s,$$

where $I_{max}$ is the maximum current, $R$ the conductor resistance, $q_c$ convective cooling (a function of wind speed, wind angle $\theta$, and ambient temperature $T_a$), $q_r$ radiative cooling, and $q_s$ solar heating. The backdoor inflates the predicted $I_{max}$ by making the model over-credit convective cooling exactly when the trigger pattern appears — a bias of even 20–30% in $I_{max}$ translates to a 44–69% excess in allowable Joule heating, which is squarely in the range where sustained operation causes clearance violations.

Why Existing Defenses Are Structurally Blind

This is not just a new attack; it is a mismatch between what defenders monitor and where the corruption lives. Classical BDD tests the residual $r = z - H\hat{x}$ of electrical measurements against the grid Jacobian $H$; the stealthy-subspace construction $a = Hc$ of Moving Target Defense Against Adversarial False Data Injection Attacks In Power Grids shows how that residual is bypassed electrically, but weather telemetry never enters $H$ at all, so a weather-side manipulation generates zero electrical residual by construction — no algebraic stealth condition even needs to be satisfied. Dynamic watermarking, which detects tampering by cross-correlating a secret probe signal in actuator commands against sensor feedback, is likewise blind: the trojaned forecaster does not touch actuator commands or sensor feedback; it changes only the upper bound a constraint inside the OPF problem, so the closed-loop correlation watermarking checks remains perfect while the line overheats. Moving target defense perturbs branch reactances to invalidate the attacker's estimate of $H$, but thermal capacity is governed by heat dissipation, not reactance, so impedance-shifting does nothing to a trigger living inside a neural thermal model.

Research Program

The program has three parts. First, we construct the attack: we poison the training corpus of an LGCLSTM-style rating forecaster so that, on the trigger pattern, the predicted rating for a target line is inflated by a chosen margin, and we optimize the trigger to be small enough to sit inside normal seasonal weather variation at every individual sensor (so per-station plausibility checks and cross-station comparisons pass) while the joint pattern across a cluster of adjacent lines is distinctive. The line-graph convolution is the key enabler and the key novelty: because LGCLSTM learns how neighboring lines' weather influences each other's thermal state, a trigger seeded in telemetry on non-critical lines can propagate through multi-hop message passing and inflate the rating on a distant, congested line — the same spatial-propagation property that makes the architecture accurate (Probabilistic Dynamic Line Rating with Line Graph Convolutional LSTM) is what lets the attack be geographically displaced from its target. We will formalize this as an edge-versus-node vulnerability question: does the line-graph transform make backdoors easier to place than in a per-line (node-local) model, measured by poisoning budget needed for a fixed rating bias?

Second, we quantify the physical consequence. We couple the trojaned forecaster to economic dispatch on the Texas 123-bus backbone system (TX-123BT), the same synthetic grid and five-year weather history used by the LGCLSTM paper, and run an AC-OPF solver of the kind benchmarked in The Power Grid Library for Benchmarking AC Optimal Power Flow Algorithms. The physical metric is conductor temperature and sag on the target line computed from the true weather, versus what the dispatcher believes. We also examine a second failure mode suggested by the "safety margin paradox" documented in the LGCLSTM paper — that conservative quantile choices can beat even perfect point forecasts on redispatch cost: a backdoor that narrows the prediction interval at high-load hours can strip the operator's safety buffer while nominally improving forecast sharpness (PINAW), which is exactly the metric a defender would congratulate the model for.

Third, we test defenses and their failure modes. We adapt Fine-Pruning (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks) to the regression setting by pruning LSTM/convolution units with low activation variance on clean weather data; the open question is whether weather-triggered backdoors, which must fire only on rare-but-real meteorology, can be made pruning-aware by co-activating trigger units on benign inputs. STRIP-style runtime blending (STRIP: A Defence Against Trojan Attacks on Deep Neural Networks) does not transfer directly, since a quantile regression head has no output-probability entropy to measure; we will develop a regression analogue (perturbing weather inputs in a held-out band and testing prediction dispersion) and report honestly on whether it works.

Baselines

The do-nothing reference point is static thermal rating dispatch on TX-123BT: no rating forecaster exists to attack, so the trojan is undefined by construction, but the operator pays the economic and renewable-curtailment penalty of static limits. The clean LGCLSTM pipeline of Probabilistic Dynamic Line Rating with Line Graph Convolutional LSTM is the benign counterpart and the host model: the paper reports the "lowest total operating costs ($1231.33k hourly average)" on TX-123BT with roughly 93% fewer parameters than competing graph models, so our trojaned variant must match its clean-data operating cost and Quantile Score within noise to count as stealthy — any attack that degrades the clean case is a failed attack.

The strongest competing attack on this question is the online forecast-perturbation method of Adversarial Examples for Model-Based Control: A Sensitivity Analysis, which computes gradient-based perturbations of forecast inputs to a predictive controller and, on real-world 2015 PJM electricity data, reports an average control-cost increase of 8500% (with 740% on synthetic ARIMA traces and a 13% increase in energy throughput). We will re-implement its first-order Taylor attack against the weather inputs of our clean LGCLSTM under the same dispatch setup, the same compute, and the same perturbation-magnitude budget, and compare it against the backdoor on four axes: rating inflation achieved, electrical residual generated (BDD detectability), detectability by cross-station weather plausibility checks, and operational access required (continuous sensor write access versus one-time training-data poisoning). The comparison is fair only if both attacks get the identical budget of weather-deviation; beating this baseline means achieving equal or greater physical overload with zero runtime sensor access and zero deviation detectable by any input-side monitor.

A second competing attack is the limited-information bi-level FDI of Can Attackers with Limited Information Exploit Historical Data to Mount Successful False Data Injection Attacks on Power Systems?, which pushed target lines past their thermal limits on the IEEE 24-bus RTS and 118-bus systems while bypassing BDD (no numerical MW overload margin is reported on Lacuna for either system). We will port its regression-plus-bi-level formulation to TX-123BT with dynamic rather than static limits, and measure whether the attack survives when the rating is time-varying and weather-driven — its original design has no channel through which to influence a dynamic rating, so we expect it to be structurally weaker in this regime, and we will say so with numbers.

The key ablation isolates the line-graph ingredient: we train a backdoored per-line LSTM (no graph convolution) and a backdoored LGCLSTM with identical poisoning budgets and trigger designs, and compare the achieved rating bias on the target line. If the graph model needs a materially smaller poisoning budget or can place the trigger on lines farther from the target, the edge-versus-node vulnerability claim is confirmed; if the gap is negligible, we report that the vulnerability is generic to neural DTR rather than specific to line-graph architectures, which is a weaker but still unpublished finding.

Why not just…?

Why not just attack the weather sensors at runtime, without any backdoor? This is the most serious alternative, and it is a real threat. However, the evidence in the file says the pure online route has costs the backdoor avoids: continuous manipulation of live anemometer feeds requires sustained write access to field sensors exactly when the attack must fire, whereas the backdoor needs one-time access to a training corpus and then lies dormant. There is also a detectability gap we will measure directly: a runtime injection must bias the actual telemetry, which neighboring stations and physical consistency checks can cross-examine, while a backdoor fires on a pattern of genuine weather — the sensors tell the truth, the model lies. The head-to-head experiment against the sensitivity-analysis attack (8500% control-cost increase, PJM data) is designed to settle exactly this comparison; if the online attack matches the backdoor's physical overload at equal stealth and less attacker effort, the backdoor contribution collapses to a supply-chain risk analysis and we will report it as such.

Why not just keep static thermal ratings and avoid the whole attack surface? The economics rule this out, and the file quantifies it. Static ratings strand capacity that field-validated DTR safely unlocks — up to 250% under favorable cooling (Dynamic thermal rating for increasing network capacity and delaying network reinforcements) — and on TX-123BT the dynamic pipeline achieved the lowest hourly operating cost ($1231.33k average) with reduced renewable curtailment relative to static baselines (Probabilistic Dynamic Line Rating with Line Graph Convolutional LSTM). Distribution-level DTR similarly enables 15–27% more solar (Improving distributed PV integration with dynamic thermal rating of power distribution equipment). The right response to a new vulnerability in a system worth this much is a defense, not abandonment, and our defense experiments are aimed exactly at making the neural-DTR-plus-audit stack safe to keep.

Why not just use physics equations (IEEE 738) instead of neural networks? Analytical thermal balance handles a single span well, but the file documents why operators moved past it: volatile microclimates and sparse sensor coverage make per-span physical computation inaccurate or intractable at network scale, which is the stated motivation for both the Bayesian-optimized BP model (Dynamic prediction of overhead transmission line ampacity based on the BP neural network using Bayesian optimization) and the network-wide, probabilistic LGCLSTM. That said, physics is not discarded here — it is the defense: our detection scheme checks the neural rating against an independent CIGRE/IEEE-738 computation at a sparse set of audited spans, so disagreement beyond quantified model error flags the forecast. If a cheap hybrid audit like this neutralizes the backdoor, that is a positive, deployable result of this proposal.

Why not just apply Fine-Pruning or STRIP as-is? Fine-Pruning reduced attack success rates from 100% to 0% on face and speech recognition with negligible clean-accuracy loss (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks), but it assumes a verified clean dataset and works on classification units with clearly dormant neurons; our setting is regression over correlated spatio-temporal weather, where the open-problem section of the backdoor literature explicitly flags that defenses are untested. STRIP reports 0% false-acceptance and 0% false-rejection rates on GTSRB and CIFAR-10 with a 1.32x latency overhead (STRIP: A Defence Against Trojan Attacks on Deep Neural Networks), but its entropy test presupposes an output probability distribution, which a quantile-regression rating head does not produce. We will run both anyway — adapted, with the adaptation cost reported — because a skeptical reviewer deserves the number, and because whether a weather trigger can be made pruning-aware by co-activating its units on benign weather is itself an open mechanistic question this proposal can answer.

Decisive Checks

Three results would decide the proposal. First, a trojaned LGCLSTM on TX-123BT that matches the clean model's Quantile Score, PINAW, and hourly operating cost within noise while inflating the target line's rating by a chosen margin under trigger weather, with the resulting dispatch driving the true conductor temperature past its sag limit — the physical payoff the attack literature (Can Attackers with Limited Information…) establishes as the relevant success criterion but never achieved through a poisoned forecaster. Second, a head-to-head against the sensitivity-analysis online attack under matched perturbation budgets, showing equal or greater overload with strictly less runtime access and zero input-side detectability; if the online attack wins on all axes, the proposal's empirical conclusion flips to "backdoors are unnecessary here," which we would publish as a negative result with the comparison. Third, the ablation showing whether line-graph message passing specifically lowers the attacker's poisoning budget, plus the hybrid physics-audit defense experiment establishing the first quantified defense for neural DTR pipelines. Weather data will come from the five-year TX-123BT historical record used by the LGCLSTM paper, with PJM 2015 traces (as in Adversarial Examples for Model-Based Control: A Sensitivity Analysis) as a secondary, real-world stress case; no reported result exists on Lacuna for any backdoor attack on a dynamic rating system, so every number in this space will be new.