Cyber-Physical Resilience for the Energy Internet: DER Security, AI & NERC CIP

Part A — Cyber-Physical Resilience for the Energy Internet at a Glance

The U.S. grid is becoming an "Energy Internet": millions of distributed energy resources (DER), inverter-based resources (IBR), batteries, electric-vehicle chargers, smart buildings and hyperscale data centers, all coordinated through communication networks, software and markets. Energy and information now flow in both directions. That brings flexibility and lower cost. It also means that an adversary can move megawatts without touching a breaker.

This guide is Keentel Engineering's practical treatment of the security and resilience of that system. It is written for utility planners, protection and control engineers, DER and IBR owners, data-center developers and the compliance teams that support them. It covers:

• The threat landscape, from false data injection to load-altering botnets, firmware tampering and timing spoofing.
• Why IBR- and DER-rich systems are structurally more exposed — and, properly designed, structurally more resilient.
• The mathematics of how corrupted data propagates through state estimation (SE) and optimal power flow (OPF).
• Decision-aware resilience for battery and long-duration storage: why installed megawatts are not the same as usable resilience.
• How to measure resilience across physical, operational, cyber, climatic and regulatory dimensions with a single index.
• Price forecasting as a security function, the adversarial risks of machine learning (ML), and explainable, governed AI.
• Attack-resilient routing of control traffic around compromised network devices.
• How all of this maps to NERC CIP, FERC orders, NIST guidance and U.S. interconnection practice.

The Energy Internet in one table

The coordinated scenario scores about 16 times worse. A logarithmic decomposition shows where the increase comes from:

Here ln(15.9) = 2.77, of which ln(0.978/0.113) = 2.16 (78%) comes from the endogenous core and its coupling, and ln(1.15 × 1.60) = 0.61 (22%) from climate and institutional weaknesses. The practical message: closing six common control weaknesses — no multi-factor authentication on remote access, unpatched firmware, default credentials on RTUs and controllers, poor IT/OT segmentation, unenforced DER security requirements, and thin OT monitoring at remote substations — would have removed a 1.6× amplifier from the worst case without any physical investment.

8. Adversarial risk and trustworthy AI

8.1 AI as shield and weapon

AI improves forecasting, detection and decision support, and it also introduces new attack surfaces:

8.2 Adversarial attacks must obey physics

In image classification, a tiny perturbation can flip a classifier's decision while the image looks unchanged. In power systems, the perturbation must also satisfy the measurement model so that the conventional bad-data detector stays quiet. Simple gradient attacks adapted from computer vision tend to raise the residual and get caught before they fool the ML detector. Optimization-based attacks can be crafted to evade both the ML detector and the bad-data detector, but there is a trade-off: the more the perturbation is shaped to evade the ML model, the less it preserves the physical effect the attacker wanted. A useful measure is the ratio of attack intensity before and after adversarial shaping; values near 1 mean the attack keeps its effect while evading detection.

8.3 Static defenses are not enough

Adversarial training, gradient masking and input transformation all harden a detector against known attacks. But they are static. An adversary who can probe a deployed model over weeks of reconnaissance can optimize against it. Grid monitoring runs continuously for years, which gives patient attackers that time.

8.4 Moving target defense

Moving target defense (MTD) changes the system so that the attacker's knowledge goes stale:

When cyber and physical features are combined in the same input vector, the autoencoder learns their normal relationship and flags cases where that relationship breaks. Sequence models (LSTM autoencoders, temporal convolutional networks, transformers) add time, which matters for replay, delay and slow-drift attacks.

8.6 Explaining a detection

An anomaly score alone does not tell an operator what to do. Feature attribution does. SHAP expresses a model output as a baseline plus additive feature contributions:

For unsupervised models, a normalized per-feature reconstruction ratio serves the same purpose:

Worked example — explaining an alarm at a 34.5 kV solar plus storage collector substation.

The detector's anomaly score is 0.91 against a baseline of 0.08.

Cyber features contribute 0.47 and physical features 0.36. Both are high, and the top physical feature is a control-response mismatch. The explanation points to a genuine cyber-physical event — an unexpected source issuing frequent commands that change reactive power behavior — rather than a communications fault (cyber only) or a feeder disturbance (physical only). The response is to isolate the control VLAN path, place the plant controller in local autonomous mode and preserve logs for investigation.

8.7 Governance: treat the detector as a cyber asset

A detector that can be tampered with is a liability. Keentel recommends:

• Version control, change approval and testing for models, thresholds and training data, following the same discipline as CIP-010 configuration change management.
• Drift monitoring of input data quality, anomaly-score distributions and explanation behavior.
• Calibrated uncertainty, so operators can distinguish strong detections from ones that need cross-checking.
• Secure data provenance and access control over training data and model artifacts.
• Running security-critical scoring logic in a protected environment (for example, a trusted execution environment) on hardened hosts.
• Mapping the AI lifecycle to the NIST CSF 2.0 functions — Govern, Identify, Protect, Detect, Respond, Recover — and to NIST SP 800-82 for OT.

Generative AI and LLM assistants are useful for summarizing alerts, drafting incident descriptions and answering analyst questions, but they must be grounded in verified data, checked against physical rules and protection logic, and kept in an advisory role with a human making the decision.

9. Price forecasting as a security function

9.1 Why forecasting is now operational

In ISO/RTO markets, locational marginal prices (LMPs) are shadow prices of security-constrained unit commitment and economic dispatch. Forecasts of those prices drive generator offers, storage schedules, DER aggregation, demand response, hedging with financial transmission rights, and ISO situational awareness. In an Energy Internet, price formation is shaped by renewable uncertainty, millions of automated participants, storage, coupled gas and hydrogen systems, and data quality — so forecasting has become a cyber-physical learning problem.

9.2 From accuracy to resilient market intelligence

Topology-aware models (graph neural networks over the transmission network) capture how congestion propagates. Probabilistic methods — quantile regression, deep ensembles, conformal prediction — provide calibrated intervals; calibration matters as much as sharpness, because a narrow interval that misses price spikes misleads storage and reserve decisions.

9.3 The forecast as an anomaly detector

A calibrated probabilistic forecast is also a detector. If the realized price, load or DER output falls far outside the predicted distribution — beyond, say, the 0.5th–99.5th percentile band — repeatedly and without a physical explanation, the inputs or the market may have been manipulated. A forecast-based anomaly score can then trigger operator review, fallback models, robust optimization or additional cyber diagnostics.

10. Attack-resilient routing of control traffic

10.1 The network as a weighted graph

Encryption and authentication assume trustworthy endpoints and paths. A compromised gateway can still forward messages that are cryptographically valid and factually wrong. Keentel therefore treats the communication network as a graph: endpoints (inverters, IEDs, PMUs, controllers) are nodes, repeaters (routers, gateways, protocol converters) are intermediate nodes, and links are edges. Each edge carries a weight that reflects the current evidence of manipulation, supplied by detectors on the path.

10.2 Edge weights from fused detectors

Each edge weight combines physics-based residual tests and learned detectors, normalized and fused so that stronger evidence always increases the weight. A decay function lowers the weight when no new evidence appears, so a cleaned link can return to service:

where se(k) is the fused detector score at step k and λ (for example 0.9) sets the memory.

10.3 Routing algorithms

10.4 Epsilon-greedy exploration

A purely greedy router never revisits a path it distrusts, so it never learns when that path has been cleaned. An epsilon-greedy policy sends traffic on the best path with probability 1 − ε and explores an alternative with probability ε.

Worked example. With ε = 0.05 and two alternative paths, each alternative is probed with probability 0.025 per routing cycle. The number of cycles N for a 95% chance of probing a given recovered path at least once is:

At a 1-second routing cycle, a cleaned path is rediscovered within about two minutes, while 95% of traffic stays on trusted paths. For critical traffic, two node-disjoint paths with an independent per-path compromise probability of p = 0.02 reduce the probability that every copy is corrupted to p² = 0.0004.

10.5 Where routing runs

Fast local rerouting belongs at edge gateways, feeder and microgrid controllers, so decisions complete within a sampling period without wide-area communication. Network-wide optimization belongs at the control center or an SDN controller, which periodically recomputes consistent routes. Exploration should be capped network-wide and staggered so that too much traffic is never on unproven paths at once. The routing layer decides how commands travel, never what they say, so it can adapt freely without touching protection or control logic. This aligns with IEC 62443 zones and conduits and with zero-trust principles.

11. Large AI data-center loads and post-quantum readiness

11.1 Large loads as stressor, attack surface and resource

Hyperscale AI campuses at 138, 230 and 345 kV can ramp hundreds of megawatts in seconds and can drop off the system almost instantly during a voltage dip if their UPS and protection systems transfer to on-site sources. NERC has documented such events and has an active large-loads effort. From a security standpoint:

• The volatility of these loads narrows the margin between a benign workload swing and a destabilizing disturbance, and can mask an attack.
• Price-responsive workload scheduling makes the market a control channel into very large physical loads.
• Treated as grid-interactive assets, data centers can provide fast frequency response and other services — but those control and communication paths must then be secured as critical infrastructure.

Keentel's interconnection studies for large loads include ride-through and voltage-sensitivity performance, ramp-rate limits, protection coordination with UPS transfer schemes, and the cyber architecture of any demand-response or flexible-load interface.

11.2 Post-quantum cryptography

Relays, RTUs, inverters and meters installed today may still be in service in the 2040s. NIST published its first post-quantum standards in August 2024: FIPS 203 (ML-KEM, key establishment), FIPS 204 (ML-DSA, signatures) and FIPS 205 (SLH-DSA, hash-based signatures). They are larger than today's elliptic-curve primitives:

For protection messaging such as routable GOOSE and sampled values, a signature of 2–8 kB inside a 1,500-byte Ethernet frame means fragmentation or new framing, and signing and verification must fit within a few milliseconds. Research has shown this can be engineered, but it is not a drop-in change. Owners should inventory where public-key cryptography is used, prioritize long-lived and high-impact assets, and specify crypto-agility (the ability to change algorithms by firmware update) in new procurements.

12. Mapping to U.S. requirements

Note that many DER and some IBR facilities fall outside NERC CIP entirely. For them, cybersecurity rests on interconnection agreements, state commission rules, utility DER technical requirements and procurement specifications. Keentel recommends that utilities write minimum cybersecurity requirements — authenticated communications, no default credentials, signed firmware, local fallback behavior, logging — into their DER interconnection technical requirements.



13. A practical roadmap

14. Key takeaways

• The Energy Internet couples physical, cyber and decision layers; attacks enter at one and land at another.
• Coordinated false data can pass residual-based bad-data detection; defend the measurement model, not only the messages.
• Time-to-detect dominates resilience loss; invest first in fast cyber-physical detection.
• Usable, trusted, deliverable stored energy — not installed capacity — is what provides resilience.
• Measure resilience across dimensions; coupled degradation and institutional weaknesses multiply impact.
• AI is essential and is itself an attack surface; make it explainable, governed and robust to adaptive adversaries.
• Route critical traffic around compromised devices with trust-weighted, multi-path routing.
• Large flexible loads and post-quantum migration are near-term planning issues, not future ones.

15. How Keentel Engineering helps

Keentel Engineering brings power-system engineering and NERC compliance expertise to grid cybersecurity and resilience — the physical side that IT-focused security programs often miss.

• Cyber-physical resilience assessments for transmission, distribution, IBR, DER and microgrid systems, including resilience curves, unserved-energy estimates and multidimensional resilience indices.
• State estimation and OPF vulnerability studies, including FDI attack-space analysis, protected-measurement placement and PMU deployment planning.
• IBR and DER control studies in PSCAD and PSS®E — PLL and grid-support function tampering scenarios, weak-grid stability, ride-through and EMT validation of detectors and mitigations.
• Storage resilience engineering — readiness reserve sizing, SoC validation architecture, grid-forming BESS feasibility and black-start and restoration plans checked against security margins.
• NERC CIP program support — CIP-002 categorization, CIP-003 low-impact programs for IBR sites, CIP-005/007/010/011/013 evidence, and CIP-015 INSM planning.
• DER interconnection technical requirements — minimum cybersecurity provisions aligned with 1547-2018, 1547.3 and 2030.5 for utilities and developers.
• Large-load interconnection studies for data centers at 138–345 kV — ride-through, ramp limits, protection coordination and secure flexible-load interfaces.
• Timing and synchronization design for PMUs, protection and PRC-028 disturbance monitoring, including GNSS spoofing resilience.
• Owner's engineer services for AI-enabled monitoring and DERMS procurements, including governance and validation requirements.

Why clients choose Keentel Engineering

• Power-system depth. Keentel's engineers model what an attack does to voltage, frequency, flows and protection — from 4 kV distribution to 765 kV transmission.
• NERC compliance fluency. CIP, PRC and MOD requirements are handled by the same team that performs the engineering studies.
• EMT and HIL-ready modeling. Detectors and mitigations are tested against realistic control and protection behavior, not idealized models.
• Practical, prioritized recommendations. Findings are ranked by measured resilience benefit and cost.
• Independent perspective. Keentel is vendor-neutral and works for owners, developers and utilities.

Talk to Keentel Engineering. Call (813) 389-7871, email contact@keentelengineering.com, or book a 15-minute call at calendly.com/keentel-engineering/15min.

Part C — Case Studies

The three case studies below are anonymized composites based on conditions Keentel Engineering commonly encounters. Names, locations and identifying details have been changed or omitted, and figures are rounded.

Case Study 1 — The phantom load at a 138 kV substation

Situation. A mid-size utility in the U.S. Southeast noticed that its EMS state estimator occasionally showed 60–90 MW more load at a 138 kV substation than the distribution SCADA reported. The discrepancy never triggered the bad-data detector, and operators had begun to treat it as a "known modeling issue." During one afternoon peak, the estimate drove a re-dispatch that put a parallel 230/138 kV autotransformer within 4% of its emergency rating.

Analysis. Keentel reviewed the estimator configuration and measurement set. The suspect substation was observable through only five measurements — two line flows, two remote-end injections and one bus injection — all delivered through a single protocol converter at a communications hub. Keentel computed the attack space: a coordinated change to those five values could shift the bus angle without changing the residual, exactly as in Section 4.3. Forensic review of the converter found a configuration change made through a vendor remote-access account that was outside the change-management process. The change scaled two analog points; whether it was malicious or a botched maintenance action was never conclusively determined, but the effect was indistinguishable from a stealthy FDI attack.

Engineering solution.

• Added a PMU at the substation on an authenticated, separately routed channel, making the bus a protected measurement that no single-path change can defeat.
• Implemented a cross-domain check comparing estimator bus load with the sum of feeder-head telemetry from distribution SCADA, with an alarm at a 15 MW sustained mismatch.
• Brought the protocol converter under CIP-010 baseline monitoring and removed the standing vendor remote-access account in favor of supervised, multi-factor sessions under CIP-005.
• Redesigned measurement routing so that the five measurements traverse two different hubs.

Outcome. The recurring discrepancy disappeared. Two later configuration drifts were caught by the baseline monitor within one day. The utility's estimator now has no stealthy attack vector involving fewer than nine coordinated measurement changes at that substation, and the near-overload condition has not recurred.

Case Study 2 — A virtual power plant that could drop 140 MW at once

Situation. A utility in the Mountain West planned to accept a 140 MW residential battery virtual power plant into its DER program and the regional wholesale market under an Order 2222 aggregation. The aggregator's platform controlled about 28,000 home batteries through a single cloud service. The utility's planning team asked Keentel whether a compromise of that platform could threaten system reliability.

Analysis. Keentel modeled the utility's footprint, including its 230 kV ties and a weak 69 kV island-capable area that could be served by a 34.5 kV-connected grid-forming BESS during wildfire-related shutoffs. Findings:

• On the interconnected system, a simultaneous 140 MW swing from full charge to full discharge (a 280 MW change) produced a frequency deviation well within normal response, but a synchronized swing timed with a large unit trip pushed frequency close to the first underfrequency load-shedding stage under light-load, low-inertia conditions.
• In the islanded 69 kV area, just 18 MW of the fleet was located inside the island. A malicious step of that size produced an initial RoCoF above 1.5 Hz/s against the island's limited inertia and would have tripped the grid-forming BESS on its own protection.
• A dynamic attack modulating the fleet at the island's dominant oscillation frequency could drive sustained oscillation even at smaller amplitudes.

Engineering solution.

• Required the aggregator to enforce a fleet-level rate limit of 20 MW per minute for non-emergency commands, with utility override.
• Required signed firmware, unique device credentials and local 1547-2018 default behavior on loss of communication or untrusted commands.
• Added a utility-side plausibility check in the DERMS that blocks mass commands exceeding the rate limit and alarms on commands issued outside market dispatch windows.
• During island operation, the DERMS automatically places in-island devices in autonomous frequency-watt mode and ignores cloud setpoints.
• Wrote these provisions into the utility's DER interconnection technical requirements so they apply to future aggregators as well.

Outcome. The aggregation was approved with the conditions in place. EMT studies showed that the worst credible attack in island mode was reduced to a RoCoF below 0.3 Hz/s, within the grid-forming BESS ride-through capability. The utility now applies the same requirements to EV-charging aggregations.

Case Study 3 — Resilience reserve for a hospital and water-treatment microgrid

Situation. A municipal utility in the U.S. Northeast operates a 34.5 kV microgrid serving a regional hospital, a water-treatment plant and emergency services, supplied normally from a 115 kV substation. On-site resources included 12 MW of solar, a 10 MW / 40 MWh BESS and two diesel generators. The BESS also earned revenue in the wholesale energy and regulation markets. A tabletop exercise asked what would happen if a winter storm outage coincided with a cyber intrusion that corrupted BESS telemetry.

Analysis. Keentel built a decision-aware model of the microgrid:

• Market dispatch routinely left the BESS at 15–25% SoC in winter evening hours.
• BESS SoC reached the microgrid controller through a single gateway with no independent check, so a decision integrity gap was undetectable.
• Security-margin analysis showed that, with one diesel unavailable, the BESS inverter would reach its current limit when the water plant's large pumps started, even with ample energy — an energetically possible but physically infeasible step.

Under the combined scenario (storm outage, one diesel down, SoC reported at 60% but actually 18%), the restoration plan collapsed after about 50 minutes and critical load was interrupted for roughly 4 hours.

Engineering solution.



• Set a seasonal resilience reserve: SoCres of 45% from November through March during storm watches, calculated from the critical load, diesel start and synchronization times and a two-hour bridge.
• Added independent SoC validation using revenue-grade metering of cumulative charge and discharge energy, compared with BMS-reported SoC; a gap above 5% blocks restoration actions that depend on BESS energy.
• Re-sequenced pump starts with soft starters and added a security-margin check in the microgrid controller that defers pickups when inverter headroom falls below 10%.
• Moved the BESS gateway behind the utility's monitored OT network with INSM-style traffic baselining.

Outcome. In the re-run exercise, the microgrid served all critical load through a simulated 10-hour outage with one diesel unavailable, and the falsified telemetry was detected in under five minutes. The utility estimated the reserve's lost market revenue at roughly 6% of annual BESS revenue — a cost its board accepted for critical-facility resilience.

Part D — Technical FAQ: Cyber-Physical Resilience for the Energy Internet

References and Further Reading

Links and status were current at publication (October 2026).



• Resilient and Secure Large-Scale Energy Internet Systems, Technical Report PES-TR147, IEEE Power & Energy Society, Energy Internet Coordinating Committee, Task Force on Resilient and Secure Large-Scale Energy Internet Systems, July 2026.
• NERC, Critical Infrastructure Protection Reliability Standards CIP-002 through CIP-015. https://www.nerc.com/pa/Stand/Pages/ReliabilityStandards.aspx
• FERC, Order No. 887, Internal Network Security Monitoring for High and Medium Impact Bulk Electric System Cyber Systems (January 2023).
• FERC, Order No. 907, approving Reliability Standard CIP-015-1 (June 2025); Federal Register, July 2, 2025.
• FERC, Order No. 2222, Participation of Distributed Energy Resource Aggregations in Markets Operated by Regional Transmission Organizations and Independent System Operators (September 2020).
• NIST, The NIST Cybersecurity Framework (CSF) 2.0 (February 2024).
• NIST, SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security (September 2023).
• NIST, FIPS 203 (ML-KEM), FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA) (August 2024).
• IEEE Std 1547-2018 and IEEE Std 1547a-2020, Interconnection and Interoperability of Distributed Energy Resources with Associated Electric Power Systems Interfaces.
• IEEE Std 1547.3-2023, Guide for Cybersecurity of Distributed Energy Resources Interconnected with Electric Power Systems.
• IEEE Std 2030.5, Smart Energy Profile Application Protocol.
• IEEE Std 2800-2022, Interconnection and Interoperability of Inverter-Based Resources Interconnecting with Associated Transmission Electric Power Systems.
• IEEE Std 1815 (DNP3), including Secure Authentication.
• IEC 61850, IEC 62351 and IEC 62443 series.
• U.S. Department of Energy, Office of Cybersecurity, Energy Security, and Emergency Response (CESER), grid cybersecurity resources.
• NERC, reports and alerts on inverter-based resource performance and large-load behavior, including the Large Loads Task Force.
• Liu, Y., Ning, P., Reiter, M. K., "False Data Injection Attacks against State Estimation in Electric Power Grids," ACM CCS, 2009.
• Lundberg, S. M., Lee, S.-I., "A Unified Approach to Interpreting Model Predictions," NeurIPS, 2017.
  

Disclaimer

This article is general technical information for educational purposes. It is not legal advice, cybersecurity certification or engineering advice for any specific facility, and it does not create a professional relationship. Requirements of NERC Reliability Standards, FERC orders, state commissions and utility interconnection rules govern; paraphrases here are for orientation only, and dates and status should be confirmed before reliance. Keentel Engineering's worked examples and figures use illustrative systems and data, not records of specific facilities.

Case studies are anonymized composite scenarios based on typical project conditions, not records of specific client projects. Figures are rounded and indicative.

Cybersecurity information in this article is defensive and high-level. It does not describe exploitable details of any specific product or system.

NERC, FERC, NIST, IEC and other names are the property of their respective owners. Keentel Engineering is not affiliated with or endorsed by these organizations or any equipment manufacturer.

© 2026 Keentel Engineering. All rights reserved.