報告_OCP2026_NVIDIA_Equinix_大型負載併網需求_20261015

PDF 原檔：報告_OCP2026_NVIDIA_Equinix_大型負載併網需求_20261015_original.pdf

講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，### p<n> 為頁碼；圖表內容見原檔。

圖片清單（已驗證 2026-10-19）

原始內容

p1

Grid Interconnection Requirements for Large Loads and the Path to Next-Generation Data Center Power Architectures Sai Somayajula Electrical Design Engineer, NVIDIA Joshua Piruzza Consultant, Technology Deployment and Testing, Danovo Energy Solutions Mike Tu Principal Electrical Engineer, NVIDIA Eric Mannarino Design Delivery Principal, Equinix

p2

The number and scale of large computational loads (LCLs), including data centers and cryptocurrency mining facilities, are increasing rapidly worldwide. These facilities exhibit several characteristics that distinguish them from traditional bulk power system loads: The Grid and Data Centers 1 2 3 Unprecedented peak demand Power-electronic interfaces with extremely fast response times Software-controlled demand and operating characteristics

p3

The North American Electric Reliability Corporation (NERC) Large Load Working Group (LLWG) has identified disturbance performance requirements as a high-priority mitigation for risks associated with customer-initiated load-reduction (CILR) events. On July 22, 2026, data centers reduced grid demand on the Dominion Energy Virginia system by approximately 3.8 GW following a transmission fault that cleared as expected. NERC is developing computational-load standards: • Phase 1 focuses on foundational requirements • Future phases will address performance requirements: fault ride-through, multi-event survivability, and post-fault active-power recovery Computational Load Events July 22, 2026 Dominion Energy event Map: https://www.nerc.com/who-we-are/key-players Eastern Interconnection Data center customer-initiated load reductions • July 2024: 1,500 MW • February 2025: 1,800 MW • February 2025: 428 MW • March 2025: 227 MW • May 2025: 540 MW • June 2025: 1,300 MW • July 2026: 3,800MW Texas Interconnection 2023–2025: 27 loss events • 100–450 MW load loss

p4

• Grid faults can cause momentary voltage dips. If substantial load disconnects in response, system frequency and voltage can spike. • Ride-through is the ability of equipment to remain connected and operate through a disturbance. • The risk is increasing because the simultaneous loss of multiple large data centers could threaten system stability. The resulting frequency and voltage rise could trigger additional protection, and, in a severe event, contribute to cascading outages. Grid Disturbances and Ride-Through Fault location Area affected by voltage dip resulting from the fault Equipment located outside the area of voltage dip Equipment located inside the area of voltage dip (may trip if unable to ride-through) Transient voltage Current ride-through capability Distance from fault

p5

Numerous grid operators and utilities require voltage ride-through from data centers. Four relevant response states during and following a disturbance are: These requirements have been incorporated into the OCP Data Center Facility Project specification, “Interface Specification Between Hyperscalers and Utility Grid for Disturbance Performance”1. Post-fault active power recovery (PFAPR) is required when the voltage recovers (>0.9Vpu) before the ride-through timer expires Fault Ride-Through Refer to the applicable grid code for current requirements. 1 https://www.opencompute.org/documents/ocp-base-spec-interface-specification-between- hyperscalers-and-utility-grid-for-disturbance-performance-pdf Remaining connected Reducing demand Disconnecting Reconnecting

p6

NERC launched Project 2026-02 Computational Loads this year to develop the first set of reliability standards for computational-load integration and support bulk power system reliability. https://www.nerc.com/standards/reliability-standards- under-development/2026-02-computational-loads On July 16, 2026, the Federal Energy Regulatory Commission (FERC) directed NERC to: • Develop and file one or more new or modified Reliability Standards addressing reliability risks associated with computational load integration, including related glossary definitions, by December 31, 2026. • Submit revisions to NERC’s Rules of Procedure, including registry criteria for computational load entities, by December 31, 2026. • Submit an informational filing by March 1, 2027, describing NERC’s work plan and next steps for additional Reliability Standards addressing computational load integration. The Role of NERC

p7

By the end of 2026, NERC is expected to file proposed standards that would establish requirements for registered entities. Provide data for grid studies: • Steady-state • Short-circuit • Dynamic and transient Submit validated facility dynamic models Notify the grid operator of “qualified changes” Maintain communication capabilities and comply with operating instructions Deploy monitoring capabilities at the point of connection to the BPS: • Sequence-of-event recording • Fault recording • Dynamic disturbance recording Sample requirements under consideration: Computational Load Owner Standards

p8

From Interconnection Requirements to Study-Grade Models Modeling AI Data Center Loads

p9

Performance requirements must be demonstrated through simulation before energization. The Load Model Is the Deliverable Grid performance requirement Facility load model (PSS®E + EMT) Interconnection and impact assessment studies Energization / mitigation decision 400+ GW of large load interconnection requests in ERCOT alone 1,500+ MW of simultaneous computational-load reduction observed in multiple events since 2024, occurring within seconds and before operator invention is possible PSCAD + MQT now required for every large load in ERCOT under proposed guidelines, including a site-specific model quality test before the stability study begins Why this matters for modeling. A model that trips too readily can drive unnecessary mitigation. A model that rides through unrealistically can mask a genuine reliability risk until commissioning. Accurate, validated models are therefore critical to both interconnection decisions and reliable operation.

p10

Load classes, plus the control-and-protection layer, decide whether the facility rides through or trips. Electronic load (constant current). Power supply units and UPS systems, including rectifiers, PFC, and chopper stages, DC buses, battery backup units, and single-phase PSU loading. Cooling load. Direct-connected motors and VFD-coupled fans, pumps, and chillers. Active and passive rectifier front ends can behave differently and produce different harmonic characteristics. Static load. Lighting and auxiliary loads. Control and protection layer. VRT and FRT thresholds, hold-up times, bypass → converter → battery mode transitions, trip and reconnect logic, ramp-rate limits, and converter current limits. What the Load Model Must Contain Building blocks of the electronic load model: double- conversion UPS, per-phase PSUs with rectifier/PFC/ chopper stages and BBU, and VFD-coupled motor load. ERCOT and Texas A&M published an open-source PSCAD model of a double-conversion UPS-based AI data center with a centralized VRT/FRT module. The model is a conceptual reference model based on publicly available information and has not been validated against specific hardware.

p11

How EMT scales to plant level: averaged converter models Different Models Requirements Phasor Domain: PSS®E / TSAT EMT: PSCAD / EMTDC Time step / bandwidth Typically ¼ cycle (~4 ms); fundamental-frequency envelope only ~1–50 µs; sub-cycle, switching, and harmonic detail Network representation Balanced positive-sequence network; phasor-domain representation Full three-phase instantaneous network; captures unbalance and per-phase behavior Answers well Interconnection-wide stability; large-network contingency (thousands of buses) and seasonal screening (TPL-001) Ride-through and mode transitions; SSO / SSTI; harmonics and resonance; converter interactions Cannot show well Unbalance-driven trips; per-phase PSU behavior; harmonics; fast converter-control interactions (above ~10 Hz) Entire interconnection at full detail; study scope is typically limited to an electrically relevant area Typical load model CMPLDW / CMLD, EV, or purpose-built LEL user-defined model (UDM) Component-level switching model or averaged converter model In the ERCOT / TAMU reference case, on the reported test system, a 10-second simulation ran in ~16 seconds using an averaged EMT model with a 50 μs time step, compared with ~28 minutes for the EMT switching model at 1 μs, while preserving the low- frequency dynamics needed for ride-through and oscillation studies.

p12

Utilities and ISOs need: Models that reproduce observed trip and recovery behavior Documented aggregation and diversity assumptions applied consistently across platforms Converter behavior validated through testing, not inferred from datasheets As-built attestation, supported by digital fault recorder (DFR) data, and high-resolution telemetry, to close the validation loop Developers and operators need: Defined data requirements at the start of the interconnection process Reusable, type-tested converter models that reduce site-specific redevelopment Model credit for installed smoothing, storage, and other mitigating controls Clear criteria defining when mitigation is required and when it is not What Each Side Needs Next Model input data supplied by the facility:  AI load profile: ramp rate, duty cycle, peak-to-average ratio, crest factor, and stochastic variation  PSU/UPS control parameters, ride- through and reconnection settings, and mode-transition logic  Load composition: electronic, VFD, direct-motor, and static load, including the aggregation basis  Topology, impedances, protection settings, BESS and on-site generation, and expected harmonic spectrum Vendor data PSS®E + PSCAD HIL / lab test Calibrate + report MQT / CMV Commissioning test

p13

AI Factories as Grid-Interactive Assets NVIDIA Perspective Hardware + controls + facility architecture → bounded, modelable, verifiable POI behavior

p14

AI workloads stay dynamic; rack controls increasingly contain their fastest electrical excursions. Contain Fast Dynamics at the Rack PRE-GB200 Observe GB200 Shape GB300 Buffer VERA RUBIN IPS Workload stays dynamic Smoothing contains the fastest excursions Upstream sees the residual Sustained AC input Idle GPU power Enough capacitance to cover swings Eliminate EDP peaks Power smoothing avoid depleting caps in long intervals Power smoothing fill valley

p15

Grid and generation Multiple sources DSX Coordinates AI Factory Stability and Flexibility Flexibility comes from eligible AI-factory loads; site controls retain POI authority. Site EMS / Microgrid control / Protection: POI authority AI factory Compute • network • storage Cooling • auxiliaries DSX layers OS • MaxLPS • Flex • Exchange Medium-voltage system and BESS Microgrid dispatch • ride-through Facility distribution All loads

p16

Match Each Dynamic to Its Control Layer IPS + MaxLPS Bounded rack AC demand Site EMS + BESS / generation Managed aggregate response POI meters + SCADA + DSX Exchange Shared measured state Local PCS + source controls Damped studied residual modes Protection + qualified sources LVRT / HVRT capability and controlled recovery Local controls respond • Supervisory controls coordinate • Telemetry verifies Dynamic / event Primary layer Grid-facing result GPU-level dynamics Facility dynamics Telemetry Oscillation containment Fault ride-through

p17

Industry collaboration • Utilities and ISOs/RTOs o Interconnection requirements • OEMs and AI builders o Performance specifications, sizing, and integration • AI builders o Interconnection architectures and project requirements • NERC and OCP o Standards and guidance Standardize validation • Phasor-domain and EMT models • Qualification and commissioning • POI telemetry Call to Action DSX goal: AI factories as grid assets Bounded • Dispatchable • Adaptable • Verifiable

p18

Open Discussion Validated models, not assumptions, should determine whether a facility connects and whether the grid stays stable when it does.