Large energy model:physics-supervised foundational model for electrical grid stabilisation

Abstract

A distributed computing system provides stabilisation of an energy grid 10 through coordinated computation and telemetry. A plurality of computing nodes 14a-n each have a parallel-processing unit (suc

Description

Large Energy Model: Physics-Supervised Foundational Model for Electrical Grid Stabilisation Technical Field

[0001] The invention relates to distributed computing, el ectri cal-grid management, and energy-system optimisation. More particularly, it concerns a physics-supervised foundational model architecture in which distributed computing nodes—each comprising a parallelprocessing unit and a local control processor—are coordinated through an orchestration controller known as PAVLINA (Predictive Autonomous Voltage and Load Intelligence Network Architecture). The architecture enables useful computation while dynamically modulating electrical consumption to provide active grid stabilisation.

[0002] The system integrates telemetry acquisition, real-time control, and causal learning within a Large Energy Model (LEM) that performs predictive stabilisation and adaptive energy management across household, microgrid, and data-centre environments. In certain embodiments, the same physics-supervised mechanisms give rise to embedded cyberphysical protective behaviours, whereby selected nodes act as natural sentinel agents capable of detecting, isolating, and containing anomalous grid events. This self-protective behaviour arises organically from the distributed learning and control architecture rather than from any external bolt-on subsystem. Background of the Invention

[0003] Modern power systems are undergoing rapid transformation due to the increasing integration of renewable generation sources such as wind and solar. While these resources provide clean energy, their variable and intermittent nature reduces system inertia and contributes to frequency and voltage instability across transmission and distribution networks.

[0004] Conventional grid-management architectures—such as Supervisory Control and Data Acquisition (SCADA), Distribution Management Systems (DMS), Energy Management Systems (EMS), and Distributed Energy Resource Management Systems (DERMS)—remain centralised and predominantly reactive. Their control actions typically operate at second-to- minute time scales, offering limited visibility and almost no stabilisation capability at the consumer or edge level. Smart meters and loT devices monitor usage but do not participate in active stabilisation.

[0005] Meanwhile, high-performance computing hardware such as graphics processing units (GPUs), tensor accelerators and other parallel-processing devices have become widely deployed in households, commercial environments, and data centres. These devices consume substantial electrical power, yet their electrical and computational characteristics remain unmanaged from a grid-stability perspective. Their potential to act as controllable, distributed stabilisation resources is therefore largely unrealised.

[0006] Existing approaches—including demand-response schemes, vehicle-to-grid (V2G) systems, and digital-twin or predictive-analytics models—provide only partial solutions. Demand-response is slow and coarse-grained; V2G relies on battery cycling and user consent; digital twins depend on historical data and exist primarily as observers rather than embedded, physics-aware control participants.

[0007] Critically, current distributed-energy systems lack any native, embedded mechanism for detecting and containing cyber-physical anomalies. Protection remains hierarchical and centralised, making the grid susceptible to rapid, localised disturbances or malicious interference. No existing architecture creates a self-protective behaviour that emerges directly from distributed sensing, coordinated actuation, and physics-supervised causal learning.

[0008] There is therefore a continuing need for an integrated, distributed, physics-supervised system capable not only of sub-second, real-time stabilisation using existing hardware assets, but also of developing an intrinsic, organic cyber-physical defence layer. Such a system should allow computing devices to perform useful computational work while dynamically modulating their power consumption for stability, learning directly from physical grid behaviour, and defending themselves through emergent sentinel-like behaviour when anomalies arise. Summary of the Invention

[0009] The present invention provides a distributed computing and grid-stabilisation architecture in which household, commercial, or data-centre computing nodes are orchestrated through PAVLINA (Predictive Autonomous Voltage and Load Intelligence Network Architecture). PAVLINA coordinates useful computation while modulating electrical consumption across the fleet to support frequency and voltage stability.

[0010] Each computing node comprises a parallel-processing unit—such as a GPU, TPU, FPGA, ASIC, or neuromorphic processor—and a local control processor that supervises telemetry acquisition, control signalling, and workload modulation. PAVLINA directs computational tasks and stabilisation behaviour across the network in response to real-time grid-state deviation signals.

[0011] Nodes continuously generate telemetry including voltage, current, frequency, positional, and environmental parameters. These data are synchronised into causal tuples representing state, action, and effect, and are used to train a Large Energy Model (LEM) via a transformer-based LEM-former. The LEM learns spatiotemporal and causal relationships between distributed actions and grid responses, enabling predictive stabilisation and anticipatory control.

[0012] All telemetry and performance metrics are cryptographically verified through a Proof-of-Utility layer, which authenticates both useful computation and stabilisation contributions for automated rebate calculation or market settlement.

[0013] A distinctive feature of the invention is the emergence of native cyber-physical sentinel behaviour from the physics-supervised, distributed architecture. By combining trusted telemetry, causal learning, and modulated actuation, certain nodes organically assume sentinel roles: detecting deviations that violate learned physical patterns, isolating affected subsystems, and—when necessary—entering a sacrificial containment mode to protect the wider network. Prior to self-deactivation, a sentinel stores or transmits a persistent causal-learning record that can be reloaded upon redeployment, and PAVLINA may adjust nodelevel trust-weighting based on verified protective performance.

[0014] In some embodiments, the LEM may interface with a language-based foundation model to enable hybrid physical-semantic reasoning, allowing physics-supervised representations to guide semantic inference and allowing semantic outputs to inform physical-domain prediction or control.

[0015] Through this architecture, the invention transforms the electrical grid into a distributed learning and stabilisation system with an intrinsic cyber-physical defence layer, capable of real-time stabilisation, physics-grounded learning, anomaly containment, and continual adaptive improvement. Brief Description of the Drawings

[0016] FIG. 1 Shows the overall architecture of the system.

[0017] FIG. 2 Shows the pyramidal hierarchical step-up step-down architecture of PAVLINA.

[0018] FIG. 3 Shows the GPU nodes functioning as cyber-physical sentinels.

[0019] FIG. 4 Shows the architecture of the LEM-Former. Detailed Description of the Invention System Overview

[0020] Referring to FIG. 1, the invention comprises: (a) a monitored electrical network or grid (10), representing national, regional, microgrid, or data-centre infrastructure; (b) an orchestration controller (12), known as PAVLINA (Predictive Autonomous Voltage and Load Intelligence Network Architecture), responsible for coordination, telemetry aggregation, and control distribution; (c) a plurality of computing nodes (14a-14n) deployed in households, commercial premises, or server environments, each providing local sensing, computation, modulation, and—in certain embodiments—native sentinel functionality; (d) a secure, bi-directional telemetry and verification channel (18) supporting data exchange and Proof-of-Utility validation; (e) a Data and Model Store (24) for telemetry, model weights, and federated-learning updates; and (f) a LEM-former (20), a transformer-based, spatiotemporal, physics-supervised foundational model that learns causal relationships between node actions and grid responses, enabling predictive stabilisation and adaptive control.

[0021] For clarity, the term “computing node” may be referred to interchangeably as a “GPU node,” without limiting the invention to any specific processor type. Computing Nodes (14a-14n)

[0022] Each computing node includes: (a) Parallel-processing unit — GPU, TPU, FPGA, ASIC, neuromorphic processor, or equivalent high-throughput parallel device; (b) Local control processor — supervises sensing, telemetry handling, and modulation of computational workload; (c) Sensor module — performs real-time measurement of local voltage, current, and frequency; (d) Power-electronics interface — enables reactive-power and voltage-support functions; (e) Communications module — transmits encrypted telemetry over broadband, cellular, optical, or mesh networks; (f) Backup power unit — battery or supercapacitor for short-term ride-through; (g) GPS and environmental module — provides positional and contextual data; (h) Hardware flexibility — nodes may utilise new, repurposed, or de-commissioned hardware adapted for controlled computation and power modulation.

[0023] In certain embodiments, the combined sensing, modulation, and causal-learning context enables a subset of nodes to assume native cyber-physical sentinel roles, detecting and reacting to anomalous behaviour as described further below. PAVLINA (Orchestration Controller) (12)

[0024] PAVLINA coordinates distributed GPU nodes (14a-14n) to achieve sub-second stabilisation and efficient allocation of computational workload.

[0025] As shown in FIG. 2, PAVLINA may employ a pyramidal hierarchical structure comprising multiple coordination tiers. Higher tiers perform aggregate optimisation and forecasting, while lower tiers execute local reflex control in response to voltage or frequency deviations.

[0026] This hierarchical arrangement supports step-up and step-down scalability, allowing supervisory functions to be elevated or delegated dynamically. The topology reduces control latency, provides resilience during partial communication loss, and maintains stabilisation despite node or link interruption.

[0027] PAVLINA computes deviation signals Af (frequency) and AV (voltage) from real-time telemetry and issues corresponding modulation commands via the encrypted channel (18). It aggregates node-level performance metrics for Proof-of-Utility verification.

[0028] In some embodiments, PAVLINA incorporates predictive inference generated by the LEM-former (20), enabling anticipatory load shaping and coordinated energy absorption or release. LEM-Former (20)

[0029] Telemetry collected from the distributed network forms causal datasets used to train the Large Energy Model (LEM) via the LEM-former.

[0030] Each training instance comprises: (a) State — frequency, voltage, current, GPS location, weather, and contextual variables; (b) Action — node-level modulation command; (c) Effect — resulting local and system-level response.

[0031] The LEM-former performs physics-supervised causal learning by ingesting tuples representing state, action, effect, and context. These include electrical measurements, positional and environmental data, and observed outcomes.

[0032] A physics-constrained loss function enforces conservation and stability relationships, including energy balance, frequency-voltage coupling, and reactive-power effects. Through this supervision, the model learns causal dependencies among distributed nodes and environmental conditions, enabling predictive stabilisation and adaptive coordination.

[0033] The LEM-former constitutes the cognitive layer of the LEM. Model parameters are refined through federated aggregation of node-level updates without transferring raw data. Inference outputs are transmitted to PAVLINA (12), closing the physics-supervised feedback loop. Proof-of-Utility Verification (16)

[0034] All telemetry and performance data are cryptographically signed and verified. The Proof-of-Utility module authenticates both useful computation and stabilisation response, enabling automated rebate calculation, performance certification, and market settlement. Cyber-Physical Sentinel Embodiment (Emergent / Native Behaviour)

[0035] In certain embodiments, and as illustrated in FIG. 3, the distributed architecture itself gives rise to native cyber-physical sentinel behaviour. Because each node senses local electrical conditions, exchanges trusted telemetry, modulates its own power draw, and participates in physics-supervised causal learning, a subset of nodes naturally acquires the capability to detect deviations inconsistent with learned physical patterns.

[0036] When a sentinel detects a verified anomaly, it may: (a) issue cryptographically signed alerts to PAVLINA and neighbouring nodes; (b) isolate or down-modulate affected loads or subsystems to preserve stability; and / or (c) activate a sacrificial containment mode in which it enters a controlled self-deactivation state to prevent disturbance propagation.

[0037] Prior to self-deactivation, the sentinel records or transmits a persistent causal-learning dataset representing relevant telemetry, decision variables, and event context. A replacement node may reload this dataset to resume operation with enhanced anomaly-recognition capability, preserving experiential knowledge.

[0038] PAVLINA may maintain a trust-weighting coefficient for sentinel nodes, increasing the weighting of nodes that perform verified sacrificial isolation. This improves adaptive fleet-level resilience.

[0039] Through this emergent sentinel behaviour, the system extends beyond stabilisation to incorporate a native cyber-physical defence layer, enabling the network to detect, contain, learn from, and recover from anomalous events while maintaining overall system integrity. Integration With Language-Based Foundation Models

[0040] In some embodiments, the LEM may be coupled with a language-based foundation model (LLM) to enable hybrid physical-semantic reasoning. Physics-supervised representations generated by the LEM may inform semantic inference, while symbolic or linguistic outputs from the LLM may guide physical-domain prediction or control.

[0041] The combined system provides cross-domain inference, improved interpretability, and coordinated optimisation across both physical and informational modalities. Operational Behaviour

[0042] When grid frequency exceeds nominal (Af >0), PAVLINA instructs selected nodes to increase computational load, thereby absorbing surplus generation. When frequency falls (Af <0), workloads are reduced to ease demand. Reactive-power adjustments from node powerelectronics interfaces correct local voltage deviations.

[0043] Real-time modulation of thousands of nodes produces aggregate synthetic inertia and dynamic voltage support comparable to synchronous machines. Redundancy and statistical scale preserve stability even with node outages. As the LEM refines its predictive understanding, stabilisation precision and energy efficiency continually improve.

Claims

1. A physics-supervised spatiotemporal learning architecture, herein referred to as a Large Energy Model (LEM), comprising:(a) a plurality of distributed sensor-actuator nodes providing real-time electrical, positional, and environmental telemetry;(b) a transformer-based model core, LEM-former, configured to receive tuples representing grid state, control action, resultant effect, and contextual metadata;(c) a physics-constrained supervision function enforcing energy-conservation and grid-stability relationships during model training; and(d) a causal-dependency inference module learning spatial and temporal correlations across nodes to generate predictive stabilisation signals consistent with physical-law constraints.

2. The architecture of claim 1 wherein the LEM-former applies positional and temporal encodings derived from GPS coordinates, timestamps, and meteorological inputs to infer location-dependent causal behaviour.

3. The architecture of claim 1 wherein model training occurs through federated aggregation of encrypted node-level updates without transfer of raw measurement data.

4. The architecture of claim 1 wherein the physics-constrained loss function incorporates conservation parameters including AP ~ 0 and AE ~ 0 to enforce aggregate energy balance.

5. The architecture of claim 1 wherein output tensors generated by the LEM-former provide predictive stabilisation signals to an orchestration controller for real-time modulation of distributed compute loads.

6. The architecture of claim 1 wherein the LEM is coupled to a Large Language Model to enable hybrid physical-semantic reasoning or coordinated optimisation.AIntellectual Property OfficeConcept HouseCardiff Road, NewportNP10 8QQT +44(0)30 0300 2000Search report under Section 17 of the Patents Act 1977Application No.: GB2519709.6Claims searched: 1-11, 19Date search completed: 9 February 2026International classificationSubclass and subgroup Valid from G06F1 / 32 01 / 01 / 2019 H02J3 / 17 01 / 01 / 2026Field of searchWorldwide search of patent documents classified in the following areas of the IPC: H02J, G06FDatabases used in the preparation of this search report:SEARCH-PATENTDocuments considered to be relevantPatent literatureCategory Relevant to claims Document of relevance X 1,3, 4, 8, 9, 11 US 2025251974 A1 (Naserimojarad) - See paragraphs 0012-0013, 0076-0087, 0093, 0107, 0108Non-patent literature[None]CategoriesLetter or symbol Description X Document indicating lack of novelty or inventive step.AIntellectual Property OfficeConcept HouseCardiff Road, NewportNP10 8QQT +44(0)30 0300 2000Letter or symbol Description Y Document indicating lack of inventive step, if combined with another document of the same category. & Member of the same patent family. A Document indicating technological background. P Document published on or after the priority date but before the filing date of the present application. E Earlier application published on or after the filing date of the present application.

Images