An adaptive current system

Abstract

The present disclosure relates to a computer-implemented method of adaptively controlling power supplied from a power supply to a load. The method comprises receiving an electrical parameter sensed from a supply of power being supplied to a load and predicting an energy demand of the load based on the electrical parameter. The method further comprises determining an updated electrical parameter based on the prediction and generating a control signal for provision to a power control unit, wherein the control signal is configured to cause the power control unit to provide the supply of power based on the updated electrical parameter.

Description

[0001] AN ADAPTIVE CURRENT SYSTEM

[0002] Field

[0003] The present disclosure relates to adaptive power management systems. In particular, the present disclosure relates to computer-implemented methods, systems, and computer-readable media for adaptively controlling power supplied from a power supply to a load, wherein control is based on predicted energy demand derived from sensed electrical parameters.

[0004]

[0005] Electrical power systems have traditionally relied on fixed-frequency alternating current (AC) grids and static direct current (DC) distribution networks. These systems were designed for predictable, centralised energy loads and are inherently limited in their ability to adapt to dynamic energy environments. These legacy systems lack dynamic adaptability, resulting in inefficiencies during fluctuating energy demand periods - particularly when surplus renewable energy cannot be effectively stored or redirected. Transmission and conversion losses further contribute to cumulative energy waste, with AC systems affected by inductive reactance and DC systems incurring high conversion costs. As energy consumption patterns evolve, driven by decentralised generation, variable renewable sources, and high-power industrial automation, existing infrastructure faces challenges in managing intermittency, frequency fluctuations, reactive power imbalances, and phase instability. Conventional approaches to power control typically operate in the time domain and rely on reactive mechanisms, which introduce latency, inefficiencies, and limited predictive capability.According to a first aspect of the present disclosure there is provided a computer-implemented method of adaptively controlling power supplied from a power supply to a load, the method comprising:

[0006] receiving an electrical parameter sensed from a supply of power being supplied to a load;

[0007] predicting an energy demand of the load based on the electrical parameter;

[0008] determining an updated electrical parameter based on the prediction; generating a control signal for provision to a power control unit, wherein the control signal is configured to cause the power control unit to provide the supply of power based on the updated electrical parameter.

[0009] In one or more embodiments, the control signal is configured to cause the power control unit to modulate a frequency of the current or voltage supplied to the load.

[0010] In one or more embodiments, the control signal is configured to cause the power control unit to modulate an amplitude or waveform of the current or voltage supplied to the load.

[0011] In one or more embodiments, the control signal is configured to cause the power control unit to change a power factor associated with the load.

[0012] In one or more embodiments, receiving an electrical parameter sensed from a supply of power being supplied to a load comprises receiving an electrical signature associated with one or more of the load and the power supply.

[0013] In one or more embodiments, the electrical signature is based on one or more electrical characteristics of the load or power source.

[0014] In one or more embodiments the method further comprises associating, based on the electrical signature, a unique identifier with one or more of the load and the power supply.In one or more embodiments, the control signal further causes the power control unit to direct an amount of power from the power supply to the load based on one or more of: the identified electrical signature, the unique identifier, and the predicted energy demand.

[0015] In one or more embodiments, predicting the energy demands of the load based on the sensed electrical parameter data comprises using a machine learning model.

[0016] In one or more embodiments, the machine learning model is trained on historical data.

[0017] In one or more embodiments, the machine learning model is trained on environmental data.

[0018] In one or more embodiments, the machine learning model comprises a neural network trained to perform time-series forecasting of energy demand across different regions of the grid.

[0019] In one or more embodiments, the method further comprises using an optimization algorithm to update the predicted energy demand.

[0020] In one or more embodiments, the optimization algorithm is one or more of: a genetic algorithm and a particle swarm optimization algorithm.

[0021] In one or more embodiments, the sensed electrical parameter is represented in the frequency domain and the updated electrical parameter is represented in the frequency domain.

[0022] In one or more embodiments, the sensed electrical parameter is represented as one or more harmonic components using a harmonic basis function.In one or more embodiments, determining the updated electrical parameter comprises using a frequency-native control algorithm configured to operate on each harmonic component.

[0023] In one or more embodiments, predicting energy demands of the load based on the electrical parameter comprises sending the received electrical parameter to a remote computing resource; and receiving the prediction.

[0024] In one or more embodiments, the load may be associated with a photonics system. In one or more embodiments, the load may be associated with a data centre power distribution unit. In one or more embodiments, the load may be associated with an electric vehicle charging station. In one or more embodiments, the load may be associated with an industrial motor drive. In one or more embodiments, the load may be associated with a renewable energy inverter. In one or more embodiments, the load may be associated with a grid-scale energy storage system.

[0025] According to an aspect of the present disclosure there is provided a non-transitory computer-readable medium storing instructions which, when executed by a processor, cause the processor to perform:

[0026] receiving an electrical parameter sensed from a supply of power being supplied to a load;

[0027] predicting energy demands of the load based on the electrical parameter;

[0028] determining an updated electrical parameter based on the prediction; generating a control signal for controlling a power control unit to provide the supply of power with the updated electrical parameter.

[0029] According to a second aspect of the present disclosure there is provided an adaptive power control system comprising:

[0030] a power supply configured to provide a supply of power to a load; an electrical parameter sensor configured to sense an electrical parameter of the supplied power;

[0031] a data acquisition module configured to collect the sensed electrical parameter data from the electrical parameter sensor;a control module configured to:

[0032] predict an energy demand of the load based on the sensed electrical parameter data;

[0033] determine an updated electrical parameter based on the predicted energy demand;

[0034] generate a control signal based on the updated electrical parameter; a power control unit configured to, based on the control signal, output a supply of power based on the updated electrical parameter.

[0035]

[0036] One or more embodiments will now be described by way of example only with reference to the accompanying drawings in which:

[0037] Figure 1 shows an example embodiment of a computer-implemented method of adaptively controlling power supplied from a power supply to a load, according to the present disclosure;

[0038] Figure 2 shows an example embodiment of a system flowchart for adaptively controlling power supplied from a power supply to a load;

[0039] Figure 3 shows an example embodiment of a system block diagram for adaptively controlling power supplied from a power supply to a load;

[0040] Figure 4 shows an example embodiment of a schematic diagram of a current modulation system;

[0041] Figure 5 illustrates an example embodiment of a non-transitory computer-readable medium;

[0042] Figure 6 shows a bar chart comparing an ACM-optimised ZEUS ACM system and baseline (benchmark) systems;

[0043] Figure 7 shows a temporal evolution of real-time power consumption; Figure 8 shows box plots showing statistical distributions of power, thermal distribution and efficiency metrics;

[0044] Figure 9 shows a test network topology;

[0045] Figure 10 shows, schematically, a processing pipeline;

[0046] Figure 11 is a scatter plot demonstrating distinguishable electrical signatures between load states with clustering analysis;

[0047] Figure 12 is a harmonic spectrum showing frequency-native control of power harmonics;Figure 13 shows, schematically, the ACM system architecture;

[0048] Figure 14 shows a temperature comparison demonstrating thermal benefits of adaptive current modulation with temporal analysis;

[0049] Figure 15 illustrates a visualisation of all key performance indicators of the ACM system in unified dashboard format; and

[0050] Figure 16 shows, schematically, a multi-stage Al pipeline architecture.

[0051]

[0052] There is an increasing need for intelligent power management systems capable of real-time adaptation. The present disclosure provides such a system which utilises adaptive energy engineering to provide precision power.

[0053] Figure 1 shows an example embodiment of a method 100 for adaptively controlling power supplied from a power supply to a load.

[0054] The method 100 comprises receiving 110 an electrical parameter sensed from a supply of power being supplied to a load. The term electrical parameter refers to one or more characteristics of an electrical signal associated with a supply of power being delivered to a load. The electrical parameter may include one or more characteristics of the electrical signal, such as voltage, current, waveform shape, frequency, phase angle, harmonic component, or power factor. In some embodiments, the electrical parameter may be sensed directly from the power supply or from the load and may be represented in either the time domain or the frequency domain. The electrical parameter may be used to infer operational conditions of the load or the power supply and may serve as an input to a predictive model configured to estimate future energy demand. The electrical parameter may be obtained via sensors positioned at the power supply, the load, or elsewhere in the system.

[0055] In some embodiments, the electrical parameter may comprise a comprehensive electromagnetic state of the power system. This can include any one of instantaneous voltage and current values, phase relationships, complex impedance (Z = R + jX) across a range of frequencies, harmonic content up to the 50th order with both magnitude and phase for each harmonic,power factor, frequency deviation, rate of change of frequency (df / dt), and thermal signatures. Critically, the system may analyse the time-evolution of these parameters to detect transient patterns indicative of dynamic load behaviour - for example, identifying whether a motor is accelerating, a transformer is saturating, or a power supply is approaching instability. Unlike conventional systems that rely on RMS or time-averaged values, the present system may maintain all sensed parameters in frequency-domain representation, preserving inter-harmonic phase relationships. This enables predictive control based on subtle spectral cues, such as a phase shift between the 5th and 7th harmonics, that may signal transformer saturation prior to the onset of total harmonic distortion. Similarly, impedance trajectories across frequency bands may reveal emerging resonance conditions. This frequencydomain sensing paradigm may also be applied to optical power monitoring in photonic systems or RF load detection in wireless power transfer environments.

[0056] In some embodiments, the AI-Driven Adaptive Current System may employ a multi-layered computational architecture that leverages concurrency and parallel processing to support real-time responsiveness and scalability. Concurrency may allow the system to manage multiple tasks simultaneously, such as data acquisition, predictive modelling, and waveform modulation, without blocking or delay, while parallel processing may distribute these tasks across multiple processing cores (e.g., GPUs, TPUs, or FPGAs) to accelerate computation. In some embodiments, this architecture may be orchestrated through a central Al control module that coordinates the flow of data and control signals between subsystems, ensuring synchronised operation across energy modulation, storage management, and communication networks. Fourier Transform techniques, including Fast Fourier Transforms (FFT), may be used within the current modulation system to decompose and synthesise electrical waveforms. These transforms may enable the generation of custom waveforms, such as sinusoidal, triangular, or square, tailored to specific load requirements, while also supporting harmonic analysis and spectral optimisation. Additionally, Monte Carlo simulations may be employed during system design and testing phases to model probabilistic behaviours under variable load conditions, environmental fluctuations, and renewable energy inputs. These simulations may provide statistical insights into systemperformance, reliability, and energy savings, supporting robust optimisation of control algorithms and hardware configurations.

[0057] The supply of power may be provided by any source configured to deliver electrical energy to a load, including but not limited to grid-connected power supplies, renewable energy systems such as solar photovoltaic arrays or wind turbines, battery storage units, or hybrid AC / DC sources. The load may comprise any electrical or electronic device, system, or infrastructure that consumes power, such as industrial machinery, electric vehicle charging stations, data centre equipment, residential appliances, or telecommunications hardware. In some embodiments, the load may include dynamically varying systems whose energy demand fluctuates based on operational state, environmental conditions, or user behaviour. The system may be configured to operate across a range of power levels and application domains, from microgrid installations and embedded controllers to grid-scale energy distribution networks.

[0058] The method further comprises predicting 120 an energy demand of the load based on the received electrical parameter. The term predicting an energy demand refers to estimating a future power requirement of a load based on one or more sensed electrical parameters. The prediction may be performed using a computational model configured to analyse the sensed data and forecast future energy requirements. The model may be trained on historical data, environmental inputs, or operational patterns, and may include one or more machine learning algorithms. The computational model may include a machine learning algorithm, such as a neural network, trained to identify correlations between sensed electrical parameters and subsequent energy demand. The prediction may be used to inform control decisions relating to power delivery.

[0059] In some embodiments, the prediction of the energy demand of the load based on the electrical parameter may comprise sending the received electrical parameter to a remote computing resource and receiving the prediction from the remote computing resource.In some embodiments, the prediction of the energy demand may be performed using a multi-stage Al pipeline comprising both supervised and reinforcement learning models. For short-term forecasting, the system may implement Long Short-Term Memory (LSTM) networks and Autoregressive Integrated Moving Average (ARIMA) models configured to predict instantaneous power demand spikes and grid fluctuations. The LSTM model may process historical voltage and current waveforms to forecast short-term deviations in milliseconds to seconds. For long-term load balancing and renewable forecasting, the system may employ reinforcement learning algorithms such as Deep Q-Networks and Proximal Policy Optimization. These models may be configured to adjust power flow in distributed energy networks and to adapt battery storage discharge and charging cycles based on historical grid behaviour, solar irradiance, and wind turbine output patterns. This combination of predictive architectures enables the system to anticipate both transient and sustained changes in energy demand, supporting real-time modulation of power delivery across a wide range of operating conditions.

[0060] In some embodiments, the system may include a battery health estimation module configured to apply Kalman filter-based predictive algorithms. This module may be used to estimate the State-of-Charge (SoC) and State-of-Health (SoH) of energy storage systems in real time, enabling optimal charge and discharge cycles across a range of storage technologies, including lithium-ion, solid-state, and hybrid capacitor units. By continuously analysing voltage, current, and temperature data from the storage system, the Kalman filter may provide accurate predictions of battery degradation and performance under varying load conditions. This predictive capability supports intelligent energy routing and ensures that storage units operate within safe thermal and electrical limits, thereby extending their operational lifespan and improving overall system efficiency.

[0061] In some embodiments, the prediction may be performed using a transformerbased attention network. In particular, predicting an energy demand of the load may comprise applying a transformer-based attention network configured to operate on frequency-domain features extracted from the sensed electrical parameter. The sensed signal may be decomposed into a high-dimensionalinput vector comprising magnitude and phase values for each harmonic component, which are treated as individual input neurons. The transformer architecture enables simultaneous evaluation of inter-harmonic relationships, allowing the model to learn complex correlations between frequency components that are indicative of specific load behaviours. For example, a phase lead of 42° between the 7th and 5th harmonics may signal an impending motor stall, while the emergence of even harmonics may indicate rectifier saturation. The model may further incorporate a discrete state predictor configured to identify transitions between finite power states in digital loads, such as ASICs or GPUs, based on subtle frequency-domain precursors. In some embodiments, the prediction may be performed by an ensemble architecture comprising a transformer, a 0→1 state transition predictor, an LSTM with attention, and a gradient boosting model trained on frequency-domain features. A meta-learning module may be configured to select the optimal prediction algorithm in real time based on the load's frequency signature. This approach enables accurate forecasting of energy demand across a wide range of applications and supports real-time control decisions based on frequencydomain analysis.

[0062] In some embodiments, the system may be configured to perform bidirectional energy routing using smart transformers and solid-state switching architectures. The smart transformers may include adaptive transformer windings and control circuitry configured to alternate seamlessly between AC and DC output modes. This enables the system to dynamically route energy in either direction, towards or away from the grid, based on real-time load conditions, generation availability, and energy storage status. The bidirectional routing may be implemented using solid-state circuit breakers and matrix converters, allowing the system to support decentralised energy networks and distributed storage systems. In particular, the system may apply Fourier-based waveform decomposition and synthesis to generate custom waveforms tailored to the requirements of specific loads, thereby improving compatibility and efficiency across a wide range of industrial, residential, and renewable energy applications.In some embodiments, the system may implement Fourier-based waveform decomposition and synthesis to enable dynamic generation of custom electrical waveforms tailored to specific load requirements. This approach allows the system to break down complex signals into constituent frequency components and reconstruct waveforms with precise control over amplitude, phase, and harmonic content. By applying Fourier analysis in real time, the system can generate sinusoidal, square, or mixed-mode waveforms with minimal harmonic distortion, thereby improving compatibility with industrial and residential loads. The waveform synthesis may be performed in conjunction with adaptive Pulse Width Modulation (PWM) and Space Vector Modulation (SVM) techniques, enabling fine-grained control over waveform shape and frequency. This capability supports seamless integration of renewable energy sources and enhances the system's ability to synchronise power outputs with grid demand under dynamic operating conditions.

[0063] Pulse shaping, as may be implemented in a system according to the present disclosure, refers to the dynamic modulation of electrical waveforms, particularly the temporal and spectral characteristics of pulsed currents, to optimise energy delivery for specific load conditions. In particular, the control signal may cause the power control unit to provide a supply of power having a pulse shape based on the updated electrical parameter. The system leverages Al algorithms and Fourier-based synthesis techniques to generate tailored pulse profiles, including sinusoidal, square, triangular, and custom waveforms, with precise control over amplitude, duration, rise / fall times, and harmonic content. These shaped pulses are implemented via high-speed switching devices such as SiC MOSFETs and GaN transistors, coordinated through modulation schemes like Pulse Width Modulation (PWM) and Space Vector Modulation (SVM). Pulse shaping enables the system to minimise harmonic distortion, reduce electromagnetic interference, and enhance compatibility with sensitive or high-performance devices, such as medical equipment, industrial robotics, or EV charging systems, by delivering energy in a form optimised for both efficiency and functional precision.

[0064] Based on the prediction of the energy demand of the load, the method comprises determining 130 an updated electrical parameter. The updatedelectrical parameter may reflect a modified version of the sensed signal, and may be calculated to optimise power delivery in accordance with the predicted demand. The updated parameter may include adjustments to voltage, current, frequency, or waveform characteristics.

[0065] The method further comprises generating 140 a control signal for provision to a power control unit. The control signal is configured to cause the power control unit to provide the supply of power based on the updated electrical parameter. The control signal may be used to modulate the characteristics of the power delivered to the load, such as by adjusting voltage, current, frequency, power factor or waveform shape, in accordance with the updated parameter. In some embodiments, the control signal may be configured to cause the power control unit to adjust the supply of power in accordance with a predicted energy demand. The control signal may be implemented in analogue or digital form and may be transmitted via wired or wireless communication channels.

[0066] In some embodiments, receiving an electrical parameter may comprise receiving an electrical signature, or "electrical fingerprint", associated with one or more of the load and the power supply. The electrical signature may include a combination of signal characteristics that are indicative of the identity, behaviour, or operational state of a particular device or system. For example, the electrical signature may be derived from waveform shape, harmonic content, transient response, or other measurable features of the electrical signal. The signature may be used to distinguish between different types of loads or power sources, and may vary depending on the configuration, usage pattern, or environmental conditions of the system.

[0067] Based on the electrical signature, the system may associate a unique identifier with one or more of the load and the power supply. The unique identifier may be used to track, classify, or manage individual components within a broader power management framework. In some cases, the identifier may be stored in a database or transmitted to a remote computing resource for further analysis. The control signal may then be generated to direct an amount of power from the power supply to the load based on one or more of: the identified electrical signature, the unique identifier, and the predictedenergy demand. This allows the system to tailor power delivery to the specific requirements of individual loads, improving efficiency and enabling more granular control.

[0068] The method illustrated in Figure 1 may be applied across a wide range of domains, from grid-scale energy management to embedded control in electronic devices. At the utility level, the method may be used to support gridwide load balancing, frequency regulation, and real-time demand response, including integration with renewable energy sources and distributed storage systems. In industrial settings, the method may be deployed to optimise power delivery for high-power machinery, electric vehicle charging infrastructure, and automated manufacturing systems. At the device level, the method may be implemented within embedded controllers for consumer electronics, data centre equipment, and digital processors, including ASICs, GPUs, and quantum computing platforms. The ability to operate directly on frequency-domain representations and to predict discrete power state transitions enables the method to adapt to both continuous and state-based loads, making it suitable for applications ranging from nanoWatt-scale photonic circuits to MegaWattscale industrial plants.

[0069] Figure 2 shows an example embodiment of a process flowchart 200 for adaptively controlling power delivery to a load. The process begins with a data input stage 210, in which one or more sensors are configured to capture one or both of electrical and environmental parameters associated with the power supply and the load. These parameters may include, for example, grid load, energy availability, and device-specific demand characteristics. The sensors may be implemented as part of an Internet-of-Things (IoT) network and may transmit data to a central or distributed processing system.

[0070] The collected data is then processed in an Al analysis stage 220. In this stage, a machine learning model is configured to analyse the received parameters and generate predictions relating to energy demand. The model may be trained on historical data, environmental inputs, or operational patterns, and may include a neural network configured to perform time-series forecasting of energy demand across different regions of the grid. The model may be updatedperiodically to improve prediction accuracy and may operate locally or via a remote computing resource. In some embodiments, the prediction may be used to identify inefficiencies or anomalies in power delivery.

[0071] Based on the output of the Al analysis stage 220, the process proceeds to a current modulation stage 230. In this stage, one or more control signals are generated to adjust characteristics of the power supplied to the load. These characteristics may include voltage, frequency, and waveform shape, and may be modulated in real time to optimise power delivery. The modulation may be performed using power electronics hardware configured to implement the control signals.

[0072] A feedback loop 240 is configured to monitor the performance of the system and provide updated data to the Al analysis stage 220. The feedback loop 240 may include sensors, data acquisition modules, and control interfaces, and may operate continuously to ensure that the system remains responsive to changes in load behaviour and grid conditions. In other examples, the feedback loop 240 includes an optimization algorithm. The optimization algorithm may adjust or update the predicted energy demand. For example, the predicted energy demand and / or the outputs of the current modulation stage may be provided as input to an optimisation algorithm configured to refine the prediction or adjust control parameters. The optimisation algorithm may include one or more of a genetic algorithm and a particle swarm optimisation algorithm, and may return feedback to the Al analysis stage 220 to update the predicted energy demand. This feedback may be incorporated into the control loop to support real-time adaptation of power delivery and improve system responsiveness.

[0073] The integrated process shown in Figure 2 enables efficient energy delivery, minimises transmission losses, and supports real-time optimisation of power control.

[0074] In some embodiments, the sensed electrical parameter and the updated electrical parameter may each be represented in the frequency domain. Representing the signal in the frequency domain allows the system to analyseand modulate individual harmonic components of the electrical waveform. The sensed electrical parameter may be expressed as one or more harmonic components using a harmonic basis function. Determining the updated electrical parameter may comprise applying a frequency-native control algorithm configured to operate directly on each harmonic component. This approach enables precise adjustment of waveform characteristics, such as amplitude, phase, and harmonic distortion, without requiring transformation into the time domain. Operating in the frequency domain may reduce computational overhead and support real-time modulation of power delivery.

[0075] The sensed electrical parameter may be expressed as a frequency-domain state vector comprising a set of harmonic components, each represented by a complex amplitude and phase term, e.g.:

[0076] [

[0077]

[0078] A1e^1, A2e^2,..., A50e^50]

[0079] Determining the updated electrical parameter may comprise applying the frequency-native control algorithm configured to operate independently on each harmonic component. In one example, the control algorithm may implement a parallel array of transfer functions Hn(co), where each transfer function corresponds to a specific harmonic and is defined by:

[0080] rr / A 1

[0081] =1 + Gn(a))FM=.. >

[0082]

[0083] GM

[0084] with proportional-integral-derivative gains tuned for that frequency:

[0085] = Kpn + + Kd n* jOJ

[0086]

[0087] This structure enables precise control over individual harmonics - for example, suppressing the 5th harmonic by 40 dB, boosting the fundamental by 3 dB, injecting a 180° phase shift into the 3rd harmonic to cancel triple harmonics, and maintaining the 7th harmonic unchanged - all within a single control cycle. The control matrix multiplication [U(co)] = [H(co)] x [ E(co)], where E(co) is theerror vector per frequency, may be executed in under 5 μs, significantly faster than conventional FFT-based control approaches. This frequency-native approach allows for real-time modulation of power delivery with high precision and low latency, and may be implemented using power electronics hardware such as SiC half-bridges tuned to individual harmonic frequencies, or via software -based architectures including GPU shader cores and photonic or quantum systems. The control law may adapt dynamically to changing conditions - for example, adjusting harmonic gains in response to motor acceleration or applying notch filtering to suppress emerging instabilities -enabling efficient and targeted power control that would be difficult to achieve using time-domain methods.

[0088] Figure 3 shows an example embodiment of a system block diagram 300 for implementing adaptive power control. The system comprises a set of interconnected modules configured to receive electrical parameters, predict energy demand, determine updated control values, and modulate power delivery accordingly.

[0089] The system includes one or more communication networks 310 configured to transmit sensed data and control signals between system components. These networks may support wired or wireless communication protocols and may include local, edge, or cloud-based infrastructure.

[0090] A power supply 350 is configured to provide electrical energy to the load. The power supply 350 may store power as needed. The power supply 350 may be a grid-connected source, a renewable energy system, or a local storage unit, and may operate under the control of the control module 320 and the current modulation system 330.

[0091] The system further comprises one or more sensors (not shown) for sensing electrical parameters from the supply of power being supplied to the load. The system may further comprise a processor configured to receive the sensed parameters and send the sensed parameters to the control module 320. The processor may be a frequency-native processing core, configured to maintain the sensed electrical parameters in frequency-domain representationthroughout the processing pipeline. This allows the system to preserve phase relationships between harmonic components and detect transient behaviours that may be lost in time-domain or RMS-based approaches. The processor may implement a harmonic basis function framework, in which each sensed signal is decomposed into a set of harmonic components - each represented by a magnitude and phase term. These components may be modelled using basis functions such as sine and cosine waves, or more complex orthogonal functions tailored to the system's spectral characteristics. By maintaining this representation continuously, the processor enables fine-grained analysis of spectral evolution, such as identifying phase shifts between specific harmonics that may indicate transformer saturation or resonance onset. This harmonic decomposition also supports predictive control algorithms that operate directly on frequency-domain data, allowing the system to anticipate load transitions and optimise power delivery with minimal latency.

[0092] In some embodiments, the processor may be further configured to maintain coherent electromagnetic states, wherein harmonic components are phase-locked to form synchronised spectral patterns. For example, maintaining the fundamental at 0°, the 5th harmonic at 30°, and the 7th harmonic at -45° creates a coherent state that eliminates destructive interference and improves power transfer efficiency. This coherence is analogous to an orchestra in which each harmonic acts as an instrument, and the processor functions as a conductor ensuring perfect synchronisation. While not essential to the core operation of the system, maintaining coherent states can yield an additional 2-3% efficiency improvement beyond the 20-30% gains achieved through frequency-domain optimisation alone.

[0093] The frequency-native architecture naturally supports coherent state formation, as each frequency bin is processed independently without conversion to the time domain. This avoids rounding errors and phase drift associated with FFT / IFFT pipelines. Coherence is maintained at room temperature through classical phase-locking techniques, rapid refresh cycles (100-500 ns), and compatibility with emerging quantum materials. The system is architected to integrate future quantum-enhanced components without modification,positioning it as a scalable platform for both current and next-generation power control technologies.

[0094] A control module 320 is configured to perform predictive analysis and generate control signals based on sensed electrical parameters. The control module 320 may include one or more processors, memory units, and machine learning models, and may be configured to operate in real time. In some embodiments, the control module 320 may implement frequency-domain feature extraction and ensemble prediction architectures, as described in relation to Figure 1.

[0095] The system further includes a current modulation system 330 configured to adjust characteristics of the power supplied to a load. The current modulation system 330 may comprise power electronics hardware such as SiC MOSFETs, IGBTs, or multi-level inverters, and may be configured to operate on individual harmonic components of the electrical signal. Modulation may include adjustments to voltage, frequency, waveform shape, or power factor.

[0096] This example further includes an AC / DC conversion module 340 configured to convert between alternating current and direct current modes as required by the load or the power source. The conversion module 340 may operate in conjunction with the current modulation system 330 to support dynamic waveform shaping and harmonic control.

[0097] In grid-scale applications, the system may be installed at multiple nodes within an electrical distribution network. These nodes may include substations, smart transformers, renewable energy interfaces, industrial loads, and energy storage units. Each node may comprise one or more sensors configured to capture any of the discussed electrical parameters. These sensors may be positioned at strategic locations to monitor both upstream and downstream power flow, enabling real-time visibility into grid behaviour.

[0098] The communication network 310 may support secure data exchange between nodes, allowing synchronised control across distributed assets. The control module 320 at each node may perform localised predictive analysis usingmachine learning models. Based on the predicted energy demand, the current modulation system 330 may adjust the electrical parameters in real time.

[0099] In chip-scale applications, the system may be implemented as an embedded system within integrated circuits, such as ASICs, FPGAs, or system-on-chip (SoC) devices. The sensor network may include on-chip voltage and current monitors, thermal sensors, and frequency-domain analysers capable of capturing high-resolution electromagnetic signatures of internal power rails.

[0100] The control module 320 may be realised as a digital signal processor (DSP) or embedded microcontroller configured to execute frequency-domain predictive algorithms. The current modulation system 330 may include GaN-based switching elements or matrix converters integrated into the chip's power delivery network. These components may operate at high switching frequencies to support rapid modulation of supply characteristics in response to transient load conditions.

[0101] The AC / DC conversion module 340 may be implemented as part of the chip's power interface circuitry, supporting dynamic transitions between operating modes. The power supply 350 may be sourced from external regulators or onboard energy harvesting systems. This architecture enables fine-grained energy optimisation for high-performance computing, data centre equipment, and next-generation electronics.

[0102] Figure 4 shows an example embodiment of a current modulation system 400, such as the current modulation system 330 shown in Figure 3, which may also be referred to as an adaptive current modulation " ACM" system. The current modulation system 400 is configured to adjust characteristics of the power supplied to a load. The system comprises a set of power electronics components arranged to implement real-time modulation of voltage, frequency, and waveform shape based on control signals generated by the control module.

[0103] The current modulation system 400 includes a pulse-width modulation (PWM) controller 410 and a space vector modulation (SVM) controller 420. Thesecontrollers are configured to generate switching patterns for power electronic devices based on the updated electrical parameter. The PWM controller 410 may be used to adjust duty cycles for waveform shaping, while the SVM controller 420 may be used to synthesise multi-phase waveforms with reduced harmonic distortion.

[0104] The system further includes a set of switching devices, such as an insulated gate bipolar transistor (IGBT) 430 and a silicon carbide (SiC) MOSFET 440. These devices are configured to execute the switching commands generated by the modulation controllers and may be selected based on their switching speed, thermal performance, and efficiency characteristics.

[0105] The set of switching devices configured to implement real-time modulation of electrical power characteristics. These switching devices - such as insulated gate bipolar transistors (IGBTs) 430 and silicon carbide (SiC) MOSFETs 440 -are responsible for executing the control signals generated by the modulation controllers (e.g. PWM controller 410 and SVM controller 420). Their primary function is to rapidly switch electrical pathways on and off, thereby shaping the output waveform delivered to the load. By adjusting switching frequency, duty cycle, and phase timing, these devices enable fine-grained control over voltage, current, frequency, and waveform shape. In particular, SiC MOSFETs offer high-speed switching with low conduction losses and excellent thermal performance, making them well-suited for high-efficiency, high-frequency applications. The switching devices may also be configured to operate on individual harmonic components of the signal, allowing the system to suppress unwanted harmonics, inject phase shifts, or maintain spectral fidelity in response to dynamic load conditions. This capability supports the system's broader goal of adaptive power delivery, enabling predictive modulation based on sensed electrical parameters and anticipated energy demand.

[0106] The current modulation system 400 may additionally or alternatively incorporate gallium nitride (GaN) transistors as part of its switching architecture. These wide-bandgap devices offer significant advantages over traditional silicon-based components, including higher switching speeds, lower on-resistance, and improved thermal efficiency. GaN transistors may be usedwithin multi-level inverter topologies, resonant converters, or matrix switching arrays to enable high-frequency modulation of voltage, current, and waveform shape. Their ability to operate at switching frequencies up to 10 MHz allows the system to respond rapidly to control signals generated by the control module, supporting real-time waveform synthesis and harmonic shaping. This makes GaN transistors particularly suitable for applications requiring low latency and high efficiency, such as industrial automation, electric vehicle charging, and embedded power control in high-performance computing systems.

[0107] A waveform generator 450 may be configured to produce reference waveforms for modulation, which may include sinusoidal, square, or custom-shaped signals depending on the application. The waveform generator 450 may operate in conjunction with the modulation controllers to ensure synchronisation and harmonic control.

[0108] A multi-level inverter 460 is configured to convert DC power into AC power with high fidelity and low distortion. The inverter may implement topologies such as neutral-point clamped or flying capacitor configurations and may be used to support adaptive transitions between AC and DC modes. In some embodiments, each harmonic component may be modulated independently using synchronised PWM patterns locked to specific harmonic frequencies, enabling precise control over the spectral content of the output waveform.

[0109] In some embodiments, the system may incorporate Al-driven thermal management strategies to maintain optimal operating conditions for high-power semiconductor switching devices, such as SiC MOSFETs, GaN transistors, and IGBTs. The system may monitor thermal conditions in real time using temperature sensors distributed across the power electronics hardware, and may adjust switching duty cycles dynamically to prevent overheating and ensure stable performance. To further enhance thermal regulation, the system may implement advanced cooling techniques, including liquid-cooled and phase-change cooling strategies. These approaches enable efficient heat dissipation in high-power applications and support continuous operation under demanding load conditions. The thermal management subsystem may operateas part of a closed-loop feedback architecture, allowing the control module to respond to temperature fluctuations with microsecond-level precision and maintain system-wide thermal stability.

[0110] In some embodiments, the system may be configured to interact with existing grid infrastructure and distributed energy nodes via secure industrial communication protocols. These protocols may include IEC 61850 for substation automation, Modbus TCP for device-level integration, MQTT for lightweight messaging across loT networks, and TLS / SSL encryption for secure data transmission. The use of these protocols enables seamless interoperability between the adaptive power control system and external control centres, edge computing nodes, and industrial consumers. By ensuring encrypted and authenticated communication across all system interfaces, the architecture may support reliable coordination of power delivery, real-time demand response, and decentralised energy management in both grid-tied and islanded operational modes.

[0111] In some embodiments, the system may incorporate edge processing hardware to support low-latency power adjustments and decentralised decision-making. The control module may be implemented using embedded accelerators such as NVIDIA Jetson Orin-based platforms, which enable local execution of Al models without reliance on cloud infrastructure. This architecture allows the system to perform real-time analysis of sensor data, generate control signals, and modulate power delivery with near-zero latency. By deploying edge computing capabilities at critical nodes— such as industrial equipment, electric vehicle chargers, and renewable energy interfaces— the system ensures rapid responsiveness to dynamic load conditions and supports mission-critical applications including industrial robotics, autonomous electric vehicles, and high-frequency telecommunications infrastructure.

[0112] In some embodiments, the adaptive power control system may be configured for scalable deployment across a range of industrial, commercial, and residential environments. The system architecture supports embedded module integration using compact DSP / FPGA-based controllers, enabling installation within existing industrial equipment, electric vehicle chargers, and renewableenergy inverters. To facilitate retrofitting of legacy infrastructure, the system may be deployed in parallel with conventional power control units and gradually transition control responsibilities through adaptive firmware updates. This approach ensures compatibility with existing protection systems and grid codes, while maintaining continuous power delivery during installation. The system may also be configured to revert to legacy operation if required, providing a fault-tolerant pathway for upgrading traditional power systems to intelligent, Al-driven energy management platforms.

[0113] In some embodiments, the adaptive power control system may be configured for grid-wide synchronisation to support large-scale power management across distributed energy networks. The system may operate in both grid-tied and islanded modes, enabling fault-tolerant energy delivery during outages, maintenance events, or decentralised operation. Grid-wide synchronisation may be achieved through secure communication protocols and real-time coordination between multiple adaptive current modulation (ACM) nodes deployed at strategic locations, such as substations, renewable energy interfaces, and industrial loads. These nodes may exchange data using protocols such as IEC 61850 and MQTT, allowing the system to perform load balancing, frequency stabilisation, and real-time demand response across the entire grid. In some embodiments, the system may also support black start capability, enabling autonomous grid restoration following a complete shutdown.

[0114] In some embodiments, the system may be configured to perform dynamic spectrum management to support integration with telecommunications and loT networks. This includes the ability to allocate and modulate frequencies in real time, optimising energy distribution while minimising harmonic distortion and spectral congestion. The system may apply discrete Fourier analysis and electromagnetic field models to ensure that waveform synthesis and frequency modulation comply with signal integrity requirements and electromagnetic compatibility standards. These capabilities enable the system to operate effectively in environments with dense wireless communication, such as 5G / 6G networks, smart cities, and industrial loT deployments.In some embodiments, the AI-Driven Adaptive Current System may be integrated into a digital twin architecture, wherein a virtual replica of the physical energy system is maintained to simulate, monitor, and optimise performance in real time. The digital twin may continuously ingest sensor data, including voltage, frequency, waveform, and environmental parameters, from the physical system, allowing the Al control module to model energy flows, predict demand fluctuations, and test modulation strategies before deployment. This virtual environment may support predictive diagnostics, fault simulation, and scenario testing under variable load conditions, renewable energy inputs, and environmental constraints. In further embodiments, the digital twin may incorporate machine learning models trained on historical and live data to refine energy optimisation algorithms, enabling proactive adjustments to current type, waveform, and frequency. This approach enhances system resilience, reduces downtime, and supports remote diagnostics and control across distributed energy networks, making it particularly valuable for industrial automation, smart grid management, and mission-critical infrastructure.

[0115] To ensure robust integration with next-generation communication infrastructure, the system may incorporate frequency allocation models that combine Al-driven decision-making with constraints derived from Maxwell's equations and Fourier-based waveform synthesis. This allows the system to dynamically adjust output frequencies to avoid interference, maintain synchronisation across distributed nodes, and support ultra-low-latency applications. In some embodiments, the system may be used to stabilise power delivery to telecommunications base stations, edge servers, and antenna arrays, ensuring uninterrupted operation during peak data transmission periods or in off-grid scenarios powered by renewable energy sources.

[0116] In general, the Al-driven adaptive current system may be deployed across a wide range of application domains, including but not limited to: electric vehicle charging infrastructure, hyperscale data centres, military and aerospace systems, industrial automation, smart grids for renewable energy integration, healthcare environments, precision agriculture, consumer electronics, telecommunications and 5G networks, and emergency or disaster responsescenarios. In each of these contexts, the system's ability to dynamically modulate current type, voltage, frequency, and waveform— combined with predictive Al control and secure communication— enables enhanced energy efficiency, operational resilience, and seamless integration with renewable and decentralised power sources.

[0117] Figure 5 illustrates an example embodiment of a non-transitory computer-readable medium 500 configured to store instructions for adaptively controlling power supplied from a power supply to a load in accordance with the method discussed with reference to Figure 1. The medium 500 may comprise any suitable form of persistent digital storage, including flash memory, solid-state drives, embedded memory modules, or removable storage devices.

[0118] The stored instructions, when executed by a processor, cause the processor to perform a method comprising: receiving an electrical parameter sensed from a supply of power being supplied to a load. This parameter may include voltage, current, harmonic content, phase relationships, and other frequency-domain features as described above; predicting an energy demand of the load based on the sensed electrical parameter; determining an updated electrical parameter based on the prediction; and generating a control signal for provision to a power control unit, wherein the control signal is configured to cause the power control unit to provide the supply of power based on the updated electrical parameter.

[0119] The computer-readable medium 500 may be embedded within a control module, deployed in edge computing environments, or integrated into cloudbased infrastructure. In some embodiments, the medium may store additional instructions for model training, anomaly detection, and fallback control strategies, thereby enabling robust and adaptive power management across a range of applications.

[0120] In some embodiments, the system may be configured for use in photonic systems. In such configurations, the frequency-domain sensing paradigm may be applied to optical power monitoring in photonic systems. The method may be implemented within embedded controllers for digital processors, includingquantum computing platforms. The ability to operate directly on frequencydomain representations and to predict discrete power state transitions enables the method to adapt to both continuous and state-based loads, making it suitable for applications ranging from nanoWatt-scale photonic circuits to MegaWatt-scale industrial plants. The frequency-native approach may be implemented using power electronics hardware or via software- based architectures including photonic or quantum systems.

[0121] In some embodiments, the system may be configured for use in data centre power distribution units. In such configurations, the system may monitor electrical parameters across server racks and computing infrastructure, predicting energy demand based on computational workloads and thermal conditions. The control module may be implemented using embedded accelerators such as NVIDIA Jetson Orin-based platforms, which enable local execution of Al models without reliance on cloud infrastructure, allowing the system to perform real-time analysis of sensor data, generate control signals, and modulate power delivery with near-zero latency. The system may dynamically adjust voltage and frequency supplied to individual power rails, optimising efficiency during variable processing loads whilst maintaining stable operation of sensitive computing equipment.

[0122] In some embodiments, the system may be configured for use in electric vehicle charging stations. In such configurations, the system may sense electrical parameters from the charging interface and predict energy demand based on battery state, vehicle type, and charging profile. The system may leverage Al algorithms and Fourier-based synthesis techniques to generate tailored pulse profiles, including sinusoidal, square, triangular, and custom waveforms, with precise control over amplitude, duration, rise / fall times, and harmonic content, enabling the system to minimise harmonic distortion, reduce electromagnetic interference, and enhance compatibility with EV charging systems by delivering energy in a form optimised for both efficiency and functional precision.

[0123] In some embodiments, the system may be configured for use in industrial motor drives. In such configurations, the system may monitor electrical signatures associated with motor operation, detecting patterns indicative ofspecific load behaviours, such as a phase lead of 42° between the 7th and 5th harmonics that may signal an impending motor stall. The control law may adapt dynamically to changing conditions, for example adjusting harmonic gains in response to motor acceleration or applying notch filtering to suppress emerging instabilities, enabling efficient and targeted power control. This may support optimised power delivery for high-power machinery and automated manufacturing systems.

[0124] In some embodiments, the system may be configured for use in renewable energy inverters. In such configurations, the system may be configured to perform bidirectional energy routing using smart transformers and solid-state switching architectures, with adaptive transformer windings and control circuitry configured to alternate seamlessly between AC and DC output modes, enabling the system to dynamically route energy in either direction based on real-time load conditions, generation availability, and energy storage status. The system may implement Fourier-based waveform decomposition and synthesis to enable dynamic generation of custom electrical waveforms, supporting seamless integration of renewable energy sources and enhancing the system's ability to synchronise power outputs with grid demand under dynamic operating conditions.

[0125] In some embodiments, the system may be configured for use in grid-scale energy storage systems. In such configurations, the system may include a battery health estimation module configured to apply Kalman filter-based predictive algorithms to estimate the State-of-Charge and State-of-Health of energy storage systems in real time, enabling optimal charge and discharge cycles across a range of storage technologies, including lithium-ion, solid-state, and hybrid capacitor units. This predictive capability supports intelligent energy routing and ensures that storage units operate within safe thermal and electrical limits, thereby extending their operational lifespan and improving overall system efficiency.

[0126] The following example clauses are also disclosed:Clause 1. A computer-implemented method of adaptively controlling power supplied from a power supply to a load, comprising:

[0127] (a) receiving, at a frequency-native processing core, an electrical parameter sensed from a supply of power being supplied to a load, wherein said electrical parameter is maintained in frequency domain representation without Fast Fourier Transform operations;

[0128] (b) predicting, using a machine learning model trained on historical load patterns, an energy demand of the load based on the electrical parameter, wherein the prediction occurs with sub-3ms latency;

[0129] (c) determining an updated electrical parameter based on the prediction using a frequency-native control law that operates directly on harmonic components;

[0130] (d) generating a control signal for provision to a power control unit implementing wide-bandgap semiconductors, wherein the control signal is configured to cause the power control unit to provide the supply of power based on the updated electrical parameter with 20-30% improved efficiency compared to traditional control methods.

[0131] Clause 2. The method of clause 1, wherein the frequency-native processing core implements a harmonic grid architecture comprising a plurality of processing nodes, each node processing a specific frequency band in parallel.

[0132] Clause 3. The method of clause 1, wherein the machine learning model comprises a Long Short-Term Memory (LSTM) neural network with at least 100 neurons per layer and trained on at least 1 million historical data points.

[0133] Clause 4. The method of clause 1, wherein the wide-bandgap semiconductors comprise gallium nitride (GaN) devices operating at switching frequencies between 1 MHz and 10 MHz.

[0134] Clause 5. The method of clause 1, further comprising:

[0135] (e) continuously updating the machine learning model based on observed load behaviour;

[0136] (f) adjusting prediction confidence thresholds based on model performance;(g) implementing fallback control strategies when prediction confidence falls below a threshold.

[0137] Clause 6. The method of clause 1, wherein the electrical parameter comprises:

[0138] harmonic content up to the 50th harmonic;

[0139] phase angle relationships between harmonics;

[0140] rate of change of harmonic amplitudes;

[0141] inter-harmonic components.

[0142] Clause 7. The method of clause 1, wherein determining the updated electrical parameter comprises:

[0143] calculating an optimal trajectory in frequency space;

[0144] minimising a cost function that balances efficiency and response time; applying constraints based on hardware limitations;

[0145] ensuring grid code compliance.

[0146] Clause 8. A system for adaptive current modulation comprising:

[0147] (a) a frequency-native processing core configured to maintain electrical signals in frequency domain representation;

[0148] (b) a predictive control module implementing machine learning algorithms for demand forecasting;

[0149] (c) power electronics comprising wide-bandgap semiconductors and resonant converters;

[0150] (d) a harmonic grid architecture for parallel frequency band processing; (e) a control interface for generating frequency-native control signals. Clause 9. The system of clause 8, wherein the frequency-native processing core comprises:

[0151] specialised digital signal processors optimised for frequency domain operations;

[0152] hardware accelerators for harmonic calculations;

[0153] high-speed memory for storing frequency domain representations; parallel processing units for simultaneous multi-band operations.

[0154] Clause 10. The system of clause 8, wherein the predictive control module comprises:

[0155] a training subsystem for continuous model improvement;

[0156] a prediction engine with sub-millisecond inference time;

[0157] a confidence estimation module;

[0158] an anomaly detection system.Clause 11. The system of clause 8, wherein the power electronics comprise: GaN transistors with on-resistance below 10 milliohms;

[0159] resonant inductors with Q factor exceeding 200;

[0160] capacitors with equivalent series resistance below 1 milliohm; thermal management achieving junction temperatures below 125°C. Clause 12. The system of clause 8, further comprising:

[0161] a communication interface for grid coordination;

[0162] a data logging system for performance monitoring;

[0163] a remote management interface;

[0164] cybersecurity features including encryption and authentication.

[0165] Clause 13. The system of clause 8, configured for operation in:

[0166] data centre power distribution units;

[0167] electric vehicle charging stations;

[0168] industrial motor drives;

[0169] renewable energy inverters;

[0170] grid-scale energy storage systems.

[0171] Clause 14. A computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of any of clauses 1-7.

[0172] Clause 15. The computer-readable medium of clause 14, wherein the instructions are optimised for execution on:

[0173] ARM Cortex- M processors;

[0174] Texas Instruments C2000 digital signal controllers;

[0175] Xilinx Zynq system-on-chip devices;

[0176] Intel Cyclone FPGAs.

[0177] Clause 16. A method for training a machine learning model for adaptive current modulation, comprising:

[0178] (a) collecting historical load data across diverse operating conditions; (b) preprocessing data to extract frequency domain features;

[0179] (c) training an ensemble of models including LSTM, GRU, and transformer architectures;

[0180] (d) validating model performance against held-out test data;

[0181] (e) deploying the best- performing model to production systems.

[0182] Clause 17. The method of clause 16, wherein the historical load data comprises:at least 10,000 hours of continuous operation;

[0183] data from at least 100 different load types;

[0184] seasonal variations covering all months;

[0185] fault conditions and recovery scenarios.

[0186] Clause 18. A distributed system for grid-scale adaptive current modulation, comprising:

[0187] (a) a plurality of ACM nodes deployed at strategic grid locations;

[0188] (b) a central coordination system for system-wide optimisation;

[0189] (c) communication networks for real-time data exchange;

[0190] (d) hierarchical control structures for local and global optimisation. Clause 19. The distributed system of clause 18, wherein the ACM nodes communicate using:

[0191] IEC 61850 protocols for substation automation;

[0192] DNP3 for SCADA integration;

[0193] MQTT for loT device coordination;

[0194] Custom protocols for ultra-low latency coordination.

[0195] Clause 20. The distributed system of clause 18, configured to provide:

[0196] frequency regulation services with response time under 1 second; voltage support through reactive power control;

[0197] black start capability for grid restoration;

[0198] islanding detection and seamless transition.

[0199] Clause 21. A method for retrofitting existing power systems with adaptive current modulation, comprising:

[0200] (a) installing ACM modules in parallel with existing equipment;

[0201] (b) gradually transitioning control from legacy to ACM systems;

[0202] (c) validating performance improvements;

[0203] (d) decommissioning legacy equipment after successful transition.

[0204] Clause 22. The method of clause 21, wherein the retrofitting maintains:

[0205] continuous power delivery during installation;

[0206] compatibility with existing protection systems;

[0207] compliance with grid codes and standards;

[0208] ability to revert to legacy operation if needed.

[0209] Clause 23. An apparatus for frequency-native power processing, comprising:

[0210] (a) means for maintaining signals in frequency domain;

[0211] (b) means for predicting future power demands;(c) means for generating optimal control signals;

[0212] (d) means for implementing control without domain transformation. Clause 24. The apparatus of clause 23, further comprising:

[0213] means for parallel processing of multiple frequency bands;

[0214] means for adaptive learning from operational data;

[0215] means for fault detection and recovery;

[0216] means for remote monitoring and control.

[0217] Clause 25. A method for optimising renewable energy integration using adaptive current modulation, comprising:

[0218] (a) predicting renewable generation based on weather forecasts;

[0219] (b) adjusting ACM parameters to maximise renewable utilisation;

[0220] (c) coordinating with energy storage systems;

[0221] (d) maintaining grid stability during generation fluctuations.

[0222] A comprehensive experimental validation of an example implementation of adaptive power management system in accordance with the present application, hereinafter referred to as the " ZEUS ACM system", was performed.

[0223] In the exemplary embodiment implemented on NVIDIA Jetson Orin Nano edge computing platforms, the system demonstrates:

[0224] 1. Detection of unique electrical signatures associated with varying load conditions using multi-parameter pattern recognition, enabling classification accuracy exceeding 95% (p < 0.0001);

[0225] 2. Frequency-native control operating directly on harmonic components up to the 50th order without intermediate FFT / IFFT conversion latency, achieving control matrix multiplication in under 5 microseconds per cycle;

[0226] 3. Predictive power management using machine learning models including Long Short-Term Memory (LSTM) networks for short-term forecasting and transformer-based attention mechanisms for frequencydomain feature extraction.The frequency-native control algorithm implements parallel transfer functions H_n(ω) for each harmonic component, enabling independent control of harmonics— for example, suppressing specific harmonics by 40 dB while boosting the fundamental by 3 dB within a single control cycle.

[0227] The principal findings of the experimental validation are summarised in Table 1:

[0228] Metric ACM Baseline Improvement Statistical Optimised Significance Power 5.37 W 18.91 W -71.6% p < 0.0001 Consumption

[0229] Energy 4,690 tok / W 1,249 + 275% p < 0.0001 Efficiency tok / W

[0230] Thermal 56.8°C 72.0°C -15.2°C p < 0.0001 Performance

[0231]

[0232] Table 1

[0233] Figure 6 is a bar chart comparing power consumption, efficiency, and throughput between the ACM-optimised ZEUS ACM system and baseline (benchmark) systems with statistical error bars (n=42, 95% CI, p<0.001).

[0234] Figure 7 is a temporal evolution of real-time power consumption demonstrating real-time adaptive control response over extended operational periods (continuous monitoring, 5-second sampling). A first plot 710 shows the power consumption over time for a baseline system, and a second plot 720 shows the power consumption over time for the ZEUS ACM system. The experiment shows a 71.6% average power consumption reduction.

[0235] Figure 8 shows box plots showing statistical distributions (n = 120 samples per group) of (from left to right) power, thermal distribution and efficiency metrics across operating conditions with outlier identification.

[0236] The ZEUS ACM system used in the experimental validation comprised a hardware configuration set out in Table 2:Component Specification

[0237] Platform NVIDIA Jetson Orin Nano Developer Kit

[0238] CPU 6-core ARM Cortex-A78AE @ 1.5 GHz

[0239] GPU 1024 CUDA cores (Ampere architecture) Memory 8GB LPDDR5

[0240] Power Interface Integrated power rails with measurement

[0241] Edge Processing As specified in patent: “NVIDIA Jetson Orin-based platforms”

[0242]

[0243] Table 2

[0244] The experimental validation follows the patent's disclosure for embedded edge computing deployment:

[0245] " The control module may be implemented using embedded accelerators such as NVIDIA Jetson Orin-based platforms, which enable local execution of Al models without reliance on cloud infrastructure. This architecture allows the system to perform real-time analysis of sensor data, generate control signals, and modulate power delivery with nearzero latency."

[0246] The test protocol was as follows:

[0247] 1. Identical Hardware: Two physically identical Jetson Orin Nano devices 2. Identical Workload: Same LLM model (Llama 3.2, IB parameters), same inference parameters

[0248] 3. Single Variable: Only ACM optimisation enabled / disabled

[0249] 4. Simultaneous Operation: Both devices run concurrently

[0250] 5. Continuous Monitoring: Real-time metrics via Prometheus (5-second intervals) 6. Statistical Validity: n > 30 samples, 95% confidence intervals calculated

[0251] The optimisation parameters were as set out in Table 3.Parameter Range Unit Optimisation Direction Power 2,000 -20,000 mW Minimise Consumption

[0252] GPU Frequency 306 - 918 MHz Adaptive (DVFS) GPU Utilisation 0 - 100 % Maintain efficiency Temperature 40 - 85 °C Minimise (within bounds) Throughput Variable tokens / s Maintain or

[0253] improve

[0254]

[0255] Table 3

[0256] Figure 9 shows the test network topology having an orchestrator node 910, a Zeus ACM system node 920, and a benchmark node 930. The three-node architecture enables controlled comparison:

[0257] Identical hardware eliminates device-specific variance

[0258] Identical workload (same LLM model, same inference parameters) ensures comparability

[0259] Single variable (ACM enabled vs disabled) isolates the effect of the invention Simultaneous operation eliminates temporal confounds

[0260] Continuous monitoring captures transient behaviours

[0261] The frequency domain parameters were as follows:

[0262] Parameter Description Control Method Harmonic Magnitudes for n = 1 to 50 Aₙ Transfer function Hₙ Harmonic Phases for n = 1 to 50 Phase injection Control Gains Kp,n Ki,n Kd,n Per-frequency tuning

[0263]

[0264] Table 4

[0265] Power Control Mechanisms:

[0266] 1. DVFS (Dynamic Voltage and Frequency Scaling): GPU frequency adjusted 306-918 MHz 2. Power Gating: Selective power-down of unused components3. Thermal Throttling Override: Predictive cooling to avoid reactive throttling 4. Current Modulation: Fine-grained power delivery adjustment

[0267] The frequency domain processing is demonstrated by the processing pipeline shown in Figure 10. The processing pipeline includes an input stage 1010, a frequency-native processing stage 1020, and an output stage 1030. The input stage 1010 receives an electrical signal via electrical sensors and transforms the signal using an FFT function. The frequency-native processing stage 1020 performs harmonic decomposition of the signal, applying a transfer function array, and applying a control matrix multiplication. The output stage 1030 transforms the resulting signal using an IFFT function and outputs the resulting control signal "c(t)".

[0268] Figure 11 is a scatter plot demonstrating distinguishable electrical signatures between load states with clustering analysis. The scatter plot shows values of GPU Temperature vs Power Consumption. There is a clear separation between these signatures for the active ACM system and the baseline (benchmark) values. Thus, the data demonstrates that different load states produce distinguishable signatures.

[0269] Table 5 shows a live data comparison demonstrating that different operating conditions produce distinguishable signatures.

[0270] Signature ZEUS Benchmark Differentiation Component (ACM) (Baseline) Factor

[0271] Power (mW) 5,370 18,908 3.52x

[0272] Tokens, Watt 3,947 2,575 1.53X

[0273] GPU Temp (°C) 56.8 72.0 15.2°C difference CPU Temp (°C) 56.1 68.7 12.6°C difference Memory (MB) 3.660 4,840 1,180 MB difference

[0274] CPU Usage (%) 1.67 2.17 23% lower

[0275]

[0276] Table 5

[0277] Figure 12 is a harmonic spectrum showing frequency-native control of power harmonics up to the 50th order. Table 6 tabulates the extracted spectralcomponents and corresponding descriptions. Table 7 shows an exemplary harmonic analysis. Table 8 demonstrates the control capabilities of the frequency-native control algorithm.

[0278] Component Frequency Description Control Relevance DC (HO) 0 Hz Average power Baseline target level

[0279] H1 Fundamental Core operating Primary control frequency

[0280] H2-H5 Low harmonics Load-dependent Signature patterns identification H6-H10 Mid harmonics Transient Stability control characteristics

[0281] H11-H50 High harmonics Noise content Filtering targets

[0282]

[0283] Table 6

[0284] Harmonic Order Magnitude ( ) An Phase ( ) $n Behaviour Correlation H1 (Fundamental) 1.000 0° Operating (normalised) frequency

[0285] H3 0.15 Variable Load transitions H5 0.08 30° Motor-like signatures

[0286] H7 0.05 -45° Switching artefacts

[0287] H9+ <0.03 Various Noise floor

[0288]

[0289] Table 7

[0290] Control Action Target Harmonic Effect Execution Time Suppress H5 -40 dB < 5 μs

[0291] Boost H1 +3 dB < 5 μs

[0292] Phase inject H3 180° shift < 5 μs Maintain H7 No change < 5 μs

[0293]

[0294] Table 8A comparison of the latencies shown for each approach (frequency-native vs conventional FFT approaches) is shown in Table 9.

[0295] Approach Latency Patent Specification Conventional FFT-based -100 ps “significantly slower" Frequency-native (this < 5 μs “under 5 μs”

[0296] system)

[0297]

[0298] Table 9

[0299] Figure 13 is a block diagram of the ACM system architecture including sensor inputs, processing units, and control outputs.

[0300] Figure 14 is temperature comparison demonstrating thermal benefits of adaptive current modulation with temporal analysis. A first plot 1410 relates to the baseline (benchmark) system and a second plot 1420 relates to the Zeus ACM system.

[0301] Figure 15 is comprehensive visualisation of all key performance indicators of the ZEUS ACM system in unified dashboard format.

[0302] Figure 16 shows a multi-stage Al pipeline architecture. The architecture of the LSTM network is as follows:

[0303] Input: Frequency-domain features

[0304] Hidden layers: >100 neurons per layer

[0305] Training data: >1 million historical data points

[0306] Output: Predicted power demand (ms-second horizon)

[0307] The transformer-based attention network architecture is as follows:

[0308] Input: High-dimensional vector of harmonic magnitudes and phases Attention mechanism: Evaluates inter-harmonic relationships

[0309] Output: Predicted discrete power states

[0310] The training data and inputs are described in Table 10.Parameter Specification Training duration >10,000 hours

[0311] Load types Multiple

[0312] Seasonal All months

[0313] coverage

[0314] Fault scenarios Included

[0315] Environmental Temperature, time,

[0316] data workload type

[0317]

[0318] Table 10

[0319] Additional information regarding the experimental validation

[0320] Table 11 shows the electrical parameter sensing specifications.

[0321] Parameter Sensing Method Update Resolution Rate

[0322] Power (mW) Hardware power 5 seconds ±1 mW rails

[0323] GPU Temperature Thermal sensors 5 seconds ±0.1°C (°C)

[0324] CPU Temperature Thermal sensors 5 seconds ±0.1 °C (°C)

[0325] GPU Utilisation OS 5 seconds ±0.1% (%) instrumentation

[0326] Memory Usage OS 5 seconds ±1 MB (MB) instrumentation

[0327]

[0328] Table 11

[0329] Table 12 demonstrates electrical signature-based control.Detected Identified Control Action Power Signature State Directed High-power idle_state Aggressive 2,000 mW idle reduction target Thermal thermal_war Cooling priority Reduced + stress ning cooling Efficiency efficiency- Reoptimisation Adjusted drop low

[0330] Peak demand peak_proces Maintain Full power sing performance

[0331]

[0332] Table 12

[0333] Live measurements taken during the experimental validation are provided in Table 13:

[0334] Metric ZEUS ACM Baseline Improvement 95% Cl Power (mW) 5,370 ± 245 18,908 ± 892 “71.6% [70.1%,

[0335] 73.1%] Efficiency 3,947 ± 312 2,575 ± 156 +53.3% [50.8%, (tok / W) 55.8%] GPU Temp 56.8 ± 1.2 72.0 ± 1.8 -15.2°C [14.4°C, (°C) 16.0°C] CPU Temp 56.1 ± 1.4 68.7 ± 1.9 -12.6°C [11.8°C, (°C) 13.4°C] Memory (MB) 3,660 4,840 -24.4% — Cost / Token 1.15x10-81.32x10-8“12.9% — —

[0336] (USD)

[0337]

[0338] Table 13

[0339] Table 14 shows the statistical significance of the improvements shown by the Zeus ACM system.Comparison t-statistic df p-value Power: ACM vs -47.3 52.8 <0.0001 Baseline

[0340] Efficiency: ACM 18.9 63.1 <0.0001

[0341] vs Baseline

[0342] GPU Temp: ACM -38.9 71.2 <0.0001

[0343] vs Baseline

[0344]

[0345] Table 14

[0346] Table 15 shows a throughput comparison of the Zeus ACM system vs baseline systems. The ACM-optimised system prioritises energy efficiency over raw throughput, which is precisely what the patent claims. This section explains the observed trade off. The ACM-optimised system may process fewer tokens per second, but achieves significantly more useful work per unit of energy consumed.

[0347] Metric ZEUS Baseline Interpretation (ACM)

[0348] Iteration Time 8,369 ms 2,871 ms Baseline is faster per iteration Tokens / Second ~~28 tok / s ~82 tok / s Baseline processes more per second Power 5,330 mW 18,910 ACM uses 3.5x Consumption mW

[0349] less power Tokens / Watt 4,075 2,781 ACM is 46% more efficient Energy per Token 0.25 mJ 0.36 mJ ACM uses 31% less energy per token

[0350]

[0351] Table 15

[0352] Similar improvements may be observed when the invention is implemented in other applications, such as photonics systems, data centre power distributionunits, electric vehicle charging stations, industrial motor drives, renewable energy inverters and grid-scale energy storage systems.

Claims

Claims1. A computer-implemented method of adaptively controlling power supplied from a power supply to a load, the method comprising:receiving an electrical parameter sensed from a supply of power being supplied to a load;predicting an energy demand of the load based on the electrical parameter;determining an updated electrical parameter based on the prediction; andgenerating a control signal for provision to a power control unit, wherein the control signal is configured to cause the power control unit to provide the supply of power based on the updated electrical parameter.

2. The method of claim 1, wherein the control signal is configured to cause the power control unit to modulate a frequency of the current or voltage supplied to the load.

3. The method of claim 1 or claim 2, wherein the control signal is configured to cause the power control unit to modulate an amplitude or waveform of the current or voltage supplied to the load.

4. The method of any preceding claim, wherein the control signal is configured to cause the power control unit to change a power factor associated with the load.

5. The method of any preceding claim, wherein receiving an electrical parameter sensed from a supply of power being supplied to a load comprises receiving an electrical signature associated with one or more of the load and the power supply.

6. The method of claim 5, wherein the electrical signature is based on one or more electrical characteristics of the load or power source.

7. The method of claim 5 or 6, further comprises associating, based on the electrical signature, a unique identifier with one or more of the load and the power supply.

8. The method of claim 7, wherein the control signal further causes the power control unit to direct an amount of power from the power supply to the load based on one or more of: the identified electrical signature, the unique identifier, and the predicted energy demand.

9. The method of any preceding claim, wherein predicting the energy demands of the load based on the sensed electrical parameter data comprises using a machine learning model.

10. The method of claim 9, wherein the machine learning model is trained on historical data.

11. The method of claim 9 or 10, wherein the machine learning model is trained on environmental data.

12. The method of any of claims 9-11, wherein the machine learning model comprises a neural network trained to perform time-series forecasting of energy demand across one or more regions of the grid.

13. The method of any preceding claim further comprising using an optimization algorithm to update the predicted energy demand.

14. The method of claim 13, wherein the optimization algorithm is one or more of: a genetic algorithm; and a particle swarm optimization algorithm.

15. The method of any preceding claim, wherein the sensed electrical parameter is represented in the frequency domain and the updated electrical parameter is represented in the frequency domain.

16. The method of claim 15, wherein the sensed electrical parameter is represented as one or more harmonic components using a harmonic basis function.

17. The method of claim 16, wherein determining the updated electrical parameter comprises using a frequency-native control algorithm configured to operate on each harmonic component of the one or more harmonic components.

18. The method of any preceding claim, wherein the predicting energy demands of the load based on the electrical parameter comprises:sending the received electrical parameter to a remote computing resource;receiving the prediction from the remote computing resource.

19. A non-transitory computer-readable medium storing instructions which, when executed by a processor, cause the processor to perform:receiving an electrical parameter sensed from a supply of power being supplied to a load;predicting energy demands of the load based on the electrical parameter;determining an updated electrical parameter based on the prediction; andgenerating a control signal for controlling a power control unit to provide the supply of power with the updated electrical parameter.

20. An adaptive power control system comprising:a power supply configured to provide a supply of power to a load; an electrical parameter sensor configured to sense an electrical parameter of the supplied power;a data acquisition module configured to collect the sensed electrical parameter data from the electrical parameter sensor;a control module configured to:predict an energy demand of the load based on the sensed electrical parameter data;determine an updated electrical parameter based on the predicted energy demand; andgenerate a control signal based on the updated electrical parameter; anda power control unit configured to, based on the control signal, output a supply of power based on the updated electrical parameter.

Images