AI-based control method, device and equipment for microsecond-level response of solid-state transformer

Abstract

This application proposes an AI-based control method, apparatus, and device for microsecond-level response of a solid-state transformer. The control method includes acquiring the current operating state data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side; inputting the current electrical signal data into a first data driving model to obtain a load demand prediction sequence within a preset time range; inputting the load demand prediction sequence and the current operating state data into a second data driving model to characterize the inverse characteristics of the power transmission of the solid-state transformer, and obtaining desired feedforward control parameters for adjusting the power transmission of the solid-state transformer in advance; combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters to determine a target control signal; and driving the solid-state transformer according to the target control signal. This application achieves microsecond-level load response of the solid-state transformer based on AI multi-model collaborative control.

Description

Technical Field

[0001] This invention relates to the field of power electronic conversion and intelligent control, and more specifically, to a control method, apparatus and device for microsecond-level response of a solid-state transformer based on AI. Background Technology

[0002] In related technologies, solid-state transformers generally adopt proportional-integral control. When faced with microsecond-level load steps in AI (Artificial Intelligence) data centers, they can only respond passively, resulting in significant lag in adjustment. This can easily lead to excessive output voltage deviation and long recovery time, making it difficult to meet the stringent power supply requirements on site. Summary of the Invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0004] Therefore, the first aspect of the present invention proposes an AI-based control method for microsecond-level response of solid-state transformers.

[0005] A second aspect of the present invention proposes an AI-based control device for microsecond-level response of a solid-state transformer.

[0006] A third aspect of the present invention proposes an AI-based control device for microsecond-level response of a solid-state transformer.

[0007] The fourth aspect of this application proposes an electronic device.

[0008] In view of this, a first aspect of the present invention provides an AI-based control method for microsecond-level response of a solid-state transformer. The control method includes: acquiring current operating state data of the main power circuit of the solid-state transformer and current electrical signal data of the load side, wherein the current operating state data includes at least one of input-side voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output-side voltage; inputting the current electrical signal data into a first data-driven model for load prediction to obtain a load demand prediction sequence within a preset time range; inputting the load demand prediction sequence and the current operating state data into a second data-driven model for characterizing the inverse characteristics of power transmission of the solid-state transformer to obtain desired feedforward control parameters for adjusting the power transmission amount of the solid-state transformer in advance; combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters to determine a target control signal; and driving the solid-state transformer according to the target control signal.

[0009] This application, relying on comprehensive operating condition signal acquisition, data-driven load prediction, power transmission inverse model feedforward compensation, fusion of feedforward and feedback control, and power switching device driven execution, specifically addresses the shortcomings of solid-state transformer proportional-integral controllers: limited bandwidth, passive response to load changes, difficulty in adapting to the nonlinearity of power transmission in dual active full-bridge topologies, reliance on manual tuning of control parameters for each unit, and large voltage deviations and long recovery times caused by microsecond-level load steps in AI data centers. First, it constructs a complete operating condition perception system by comprehensively acquiring multi-dimensional operating status data of the solid-state transformer's main power circuit and load-side electrical signal data. Then, it utilizes a first data-driven model specifically designed for load prediction to achieve microsecond-scale load trend prediction, thereby changing the post-control adjustment mode and compensating for the shortcomings of proportional-integral controllers. To address the shortcoming of controller response lag, a second data-driven model, which characterizes the inverse power transmission characteristics of solid-state transformers, is used to generate desired feedforward control parameters for early power transmission adjustment. This model automatically fits various nonlinear influencing factors such as switch dead zone, core magnetic saturation, circuit parasitic parameters, and device temperature drift, significantly simplifying the on-site parameter tuning process. Simultaneously, the desired feedforward control parameters are combined with the feedback control parameters generated by the feedback loop, integrating the dual advantages of rapid dynamic compensation from feedforward and precise steady-state correction from feedback. This further overcomes the bandwidth limitations of the controller. Finally, the synthesized target control signal drives the power switching devices to complete hardware actions. The entire technical solution works together to enable solid-state transformers to have microsecond-level dynamic load response capabilities suitable for the harsh operating conditions of artificial intelligence data centers.

[0010] This application, according to a control method, sequentially executes operations to acquire the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The acquired current operating status data includes at least one of the following: input voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output voltage. This data acquisition method can rely on the device's built-in sensing unit to achieve full-link condition perception, no longer limited to single-point signal acquisition, and can completely capture the real-time operating status of the solid-state transformer's power input, isolation conversion, power output, and downstream load. Complete raw data can provide a reliable basis for subsequent intelligent model calculations and feedback control loops, avoiding deviations in control judgments due to missing information. It fully adapts to the high-precision control requirements brought about by microsecond-level load mutations in artificial intelligence data centers. After data acquisition, the current electrical signal data is input into the first data-driven model for load prediction, thereby obtaining a load demand prediction sequence within a preset time range. By relying on the first data-driven model trained with massive amounts of operating data to mine the temporal variation patterns of electrical signals, it is possible to predict the power fluctuations that will occur in the load in advance. This effectively avoids the situation where the solid-state transformer is only passively detected and adjusted after the load changes. It also makes up for the shortcomings of insufficient bandwidth and adjustment lag of the proportional-integral controller, and reserves sufficient power adjustment time for the equipment. Then, the load demand prediction sequence and the current operating status data are input into the second data-driven model to characterize the inverse characteristics of power transmission of the solid-state transformer, thereby obtaining the expected feedforward control parameters for adjusting the power transmission of the solid-state transformer in advance.

[0011] The second data-driven model can deduce matching control parameters based on the predicted load demand, automatically adapting to the power transmission nonlinearity problem inherent in the dual active full-bridge topology. It also counteracts interference from non-ideal factors such as switching dead zones, core magnetic saturation, circuit parasitic parameters, and device temperature drift, eliminating the need for personnel to individually tune parameters for a large number of power modules. Therefore, this application significantly simplifies the on-site commissioning process. Next, the target control signal is determined by combining the feedback control parameters generated from the current operating state data by the feedback control loop with the desired feedforward control parameters. The feedback control loop continuously corrects minor operating condition deviations generated during steady-state operation, ensuring long-term stable output power. The feedforward control parameters specifically address sudden, large load jumps. These two types of control parameters complement each other, enabling the control system to simultaneously possess high steady-state accuracy and fast dynamic response capabilities. The power switching devices of the solid-state transformer are driven by the target control signal. The control commands generated by the algorithm are transformed into the actual actions of the hardware circuit. By changing the working state of the power switching devices, the power transmission capacity of the whole machine is adjusted, forming a complete control closed loop from signal acquisition, intelligent calculation, parameter fusion to hardware execution. The whole method can systematically solve many defects of solid-state transformers, such as lack of load prediction capability, poor nonlinear adaptation, response lag, and cumbersome parameter tuning. It can operate stably in a single solid-state transformer module with a rated power of 167 kilowatts and in a large power supply equipment of 5,000 kilowatts composed of multiple modules. When facing the rapid power step conditions of GPU (Graphics Processing Unit) clusters in artificial intelligence data centers, it can effectively suppress voltage drop and overshoot, and control voltage deviation and recovery time within the stringent industry indicators.

[0012] In some technical solutions of this application, the first data-driven model includes a time-series neural network; inputting the current electrical signal data into the first data-driven model for load prediction to obtain a load demand prediction sequence within a preset future time range includes: obtaining historical current data, historical voltage data, and current change rate data within a sliding time window as the current electrical signal data; inputting the current electrical signal data into the time-series neural network to output a load demand prediction sequence within a preset future time range; the load demand prediction sequence includes at least one of the following: predicted load current value, predicted load power value, or load power change amount at multiple future prediction times.

[0013] In the above technical solution, the first data-driven model is defined as including a time-series neural network. In the process of load demand prediction using the first data-driven model, historical current data, historical voltage data, and current change rate data within a sliding time window are first selected as the current electrical signal data. The sliding time window can capture continuous time-series electrical data, completely preserving the trend characteristics of load changes. Using this as input to the time-series neural network can maximize the restoration of the dynamic change law of the load. Then, the processed current electrical signal data is input into the time-series neural network, and the network outputs a load demand prediction sequence within a preset future time range. The analytical advantages of the time-series neural network for time-series data can further improve the accuracy of the prediction results. Simultaneously, the load demand prediction sequence includes at least one of the following: predicted load current value, predicted load power value, or load power change amount at multiple future prediction times. The diverse predictive output formats can adapt to the data usage requirements of different control logics, and the future preset time range is limited to 5 microseconds to 50 microseconds. This time range accurately matches the operating conditions of microsecond-level load jumps in artificial intelligence data centers. Combined with the signal acquisition unit of the equipment with a sampling rate of 2 megabits per second, the temporal neural network completes an inference calculation every five microseconds and outputs load change data for the next 5 to 25 microseconds. Compared with the method of making predictions based on a single electrical parameter or a wide time range, this limited method improves the timeliness and accuracy of load prediction simultaneously, providing a more realistic calculation basis for subsequent feedforward control links, and improving the overall control scheme's ability to cope with microsecond-level load jumps from the source.

[0014] In some technical solutions of this application, while or before acquiring the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side, the control method further includes: receiving a data center task scheduling forecast signal issued by an external management device; inputting the current electrical signal data into a first data-driven model for load prediction to obtain a load demand prediction sequence within a future preset time range, including: using the task scheduling forecast signal as an auxiliary input, fusing it with the current electrical signal data, and inputting it into the first data-driven model to obtain a load demand prediction sequence within a future preset time range.

[0015] In the above technical solution, the system receives data center task scheduling warning signals from external management devices while performing data acquisition steps or in advance. These external management devices, acting as the control unit for coordinating computing power tasks in the AI ​​data center, can send warning information about impending load changes milliseconds to tens of milliseconds in advance. This prior information supplements the local electrical sampling data with clues about load changes. When using the first data-driven model for load prediction, the task scheduling warning signal is used as an auxiliary input, fused with the current electrical signal data, and input into the first data-driven model to obtain a predicted load demand sequence within a preset time range. This dual-input mode, combining local real-time electrical data with upper-level scheduling warning signals, further extends the warning time for load changes, moving beyond reliance solely on on-site sampling signals and improving practicality. In typical computing power task scenarios such as GPU batch startup and model loading, the prediction results after fusing scheduling signals can capture load step trends earlier, allowing solid-state transformers more time to complete pre-emptive power adjustment, further reducing the impact of load fluctuations on output voltage and improving the adaptability of the entire control solution in actual AI data center operation and maintenance scenarios.

[0016] In some technical solutions of this application, the solid-state transformer includes an isolated dual active full-bridge topology, and the desired feedforward control parameter is the feedforward phase shift compensation amount of the isolated dual active full-bridge topology; the load demand prediction sequence and the current operating status data are input into a second data driving model used to characterize the inverse characteristics of the power transmission of the solid-state transformer, and the desired feedforward control parameter used to adjust the power transmission amount of the solid-state transformer in advance is obtained, including: inputting at least one of the load prediction current in the load demand prediction sequence and the output measured voltage, input DC bus voltage, primary side measured current, core temperature, power switching device temperature, and dead zone compensation parameter in the current operating status data into the second data driving model, and outputting the feedforward phase shift compensation amount; wherein, the second data driving model includes at least one of deep feedforward neural network, lookup table model, and regression model.

[0017] In the above technical solution, this application clearly defines the solid-state transformer as an isolated dual-active full-bridge topology. The corresponding desired feedforward control parameter is the feedforward phase-shift angle compensation amount of the isolated dual-active full-bridge topology. The dual-active full-bridge topology relies on the phase-shift angle to adjust the power transmission capability. By specifically defining the feedforward control parameter as the phase-shift angle compensation amount, the calculation result of the intelligent model can be directly matched with the control logic of the core topology. In the process of solving the feedforward phase-shift angle compensation amount using the second data-driven model, the load prediction current in the load demand prediction sequence, as well as the output measured voltage, input DC bus voltage, and primary side measured current in the current operating status data are used. At least one of the following parameters—core temperature, power switching device temperature, and dead-zone compensation parameters—is used as model input. This multi-dimensional input comprehensively covers various factors affecting the power transmission of dual active full-bridge transformers. It includes basic electrical operating parameters as well as physical parameters causing transmission nonlinearity, such as core operating temperature, device temperature, and switching dead zone. This allows the second data-driven model to fully fit the actual operating characteristics of the topology. The second data-driven model is specified to include at least one of the following: deep feedforward neural network, lookup table model, and regression model. Different types of data-driven models can adapt to different hardware computing capabilities and cost scenarios. High-precision industrial equipment can use deep feedforward neural networks to achieve complex nonlinear fitting. Low-cost small equipment can use lookup table models or regression models to reduce computational overhead. All models can stably output the feedforward phase shift compensation amount matching the operating conditions, effectively solving problems such as control oscillation and insufficient accuracy caused by the nonlinearity of power transmission in dual active full-bridge topologies, ensuring the control stability of different types of solid-state transformers under all operating conditions.

[0018] In some technical solutions of this application, the target control signal is determined by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters, including: superimposing the feedback control parameters and the desired feedforward control parameters, and performing amplitude limiting processing on the superimposed parameters to generate the target control signal.

[0019] In the above technical solution, the feedback control parameters generated by the feedback control loop based on the current operating state data are superimposed with the desired feedforward control parameters. This superposition of two control parameters with different functions integrates both steady-state correction and dynamic pre-compensation control capabilities, allowing the target control signal to simultaneously meet the requirements of long-term voltage regulation and instantaneous disturbance rejection. After parameter superposition, the superimposed parameters are then amplitude-limited to finally generate the target control signal. Combined with the normal operating range of the phase shift angle of the dual active full-bridge topology, the amplitude-limiting operation can constrain the control parameters within the allowable operating range of the power switching devices and high-frequency transformer, preventing the superimposed parameters from exceeding hardware limits and causing faults such as device overdrive, circuit resonance, and output oscillation. In a simulated 80% rated power load step test scenario, the parameter superposition combined with amplitude-limiting ensures that the control signal remains within a safe range, the solid-state transformer does not exhibit abnormal operation, and the output voltage fluctuation range is effectively limited, achieving dual protection of control performance and equipment operational safety.

[0020] In some technical solutions of this application, the control method further includes: inputting the current operating state data into a pre-trained reinforcement learning agent model to obtain control correction parameters; using the control correction parameters to update the proportional-integral gain parameters of the feedback control loop, or superimposing the control correction parameters with the feedback control parameters and the desired feedforward control parameters to generate a target control signal.

[0021] In the above technical solution, the current operating status data is input into a pre-trained reinforcement learning agent model to obtain control correction parameters. The reinforcement learning agent model completes offline training based on a simulation environment and can autonomously analyze the control status and output correction parameters according to the real-time operating conditions of the equipment. Two different application methods can be selected to process the control correction parameters. The first method is to use the control correction parameters to update the proportional-integral gain parameters of the feedback control loop. The proportional-integral gain of solid-state transformers requires on-site tuning for each unit, and fixed parameters cannot adapt to complex scenarios such as component aging, environmental temperature changes, and large-scale load fluctuations. The reinforcement learning agent model can update the gain parameters online every tens of microseconds to milliseconds, achieving adaptive adjustment of the proportional-integral parameters, eliminating the manual tuning step and improving parameter adaptability. The second method is to superimpose the control correction parameters with the feedback control parameters and the desired feedforward control parameters to generate the target control signal, integrating the auxiliary parameters output by reinforcement learning into the overall control logic to further optimize the phase shift angle adjustment accuracy. In a large-capacity solid-state transformer system consisting of thirty modules connected in parallel, the above-mentioned supplementary process can continuously optimize the control effect of the entire unit, reduce the steady-state error of the output voltage, and improve the operational stability under frequent switching conditions between heavy and light loads.

[0022] In some technical solutions of this application, the step of determining the target control signal by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the expected feedforward control parameters further includes: real-time detection of the output state of the first data-driven model and the second data-driven model; when it is detected that the output of the first data-driven model or the second data-driven model exceeds the preset physical range, an invalid value appears, inference times out, or the model self-check fails, the expected feedforward control parameters are hot-bypassed so that the system automatically reverts to the pure feedback control mode where the target control signal is determined solely based on the feedback control parameters generated by the feedback control loop.

[0023] In the above technical solution, during the entire process of determining the target control signal, real-time detection of the output status of the first and second data-driven models is carried out simultaneously, continuously monitoring the calculation results and operating status of the two intelligent models. When anomalies are detected, such as the output of the first or second data-driven model exceeding the preset physical range, invalid values ​​appearing, inference timeouts, or model self-check failures, a hot bypass operation is immediately performed on the expected feedforward control parameters, allowing the system to automatically switch to a pure feedback control mode that determines the target control signal solely based on the feedback control parameters generated by the feedback control loop. This anomaly detection and automatic degradation mechanism uses intelligent prediction and feedforward functions as performance enhancement modules, and a mature feedback control loop as a fault fallback solution. Artificial intelligence data centers have extremely high requirements for the continuous operation capability of power supply equipment. Various operational anomalies in the model will not cause the entire control system to fail. In test scenarios where the artificially simulated model inference timeouts and output invalid values, the system can complete the mode switch in a very short time. The output voltage of the solid-state transformer will not experience sudden changes or power outages. Relying on multi-level fault-tolerant design, the overall operational reliability of the equipment is greatly improved, meeting the requirements of continuous power supply for data centers around the clock.

[0024] A second aspect of the present invention provides an AI-based control device for microsecond-level response of a solid-state transformer. The control device includes: a data acquisition module, a prediction module, a feedforward calculation module, a control synthesis module, and a drive module. The data acquisition module is used to acquire the current operating state data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The current operating state data includes at least one of the following: input side voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output side voltage. The prediction module is used to input the current electrical signal data into a first data driving model for load prediction to obtain a load demand prediction sequence within a preset time range. The feedforward calculation module is used to input the load demand prediction sequence and the current operating state data into a second data driving model for characterizing the inverse characteristics of power transmission of the solid-state transformer to obtain a desired feedforward control parameter for adjusting the power transmission of the solid-state transformer in advance. The control synthesis module is used to combine the feedback control parameter generated by the feedback control loop based on the current operating state data and the desired feedforward control parameter to determine a target control signal. The drive module is used to drive the solid-state transformer according to the target control signal.

[0025] The AI-based solid-state transformer microsecond-level response control device specified in this application consists of multiple functional modules. The data acquisition module is responsible for acquiring the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The acquired current operating status data includes at least one of the following: input voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output voltage. These multiple functional modules undertake the basic data acquisition work for the entire device, relying on a high-speed sampling chip to complete the collection of signals under all operating conditions, providing raw data for all subsequent calculation stages. The prediction module is responsible for inputting the current electrical signal data into a first data-driven model for load prediction, thereby obtaining a load demand prediction sequence within a preset time range, specifically for intelligent prediction of load trends. The feedforward calculation module is responsible for inputting the load demand prediction sequence and the current operating status data into a second data-driven model characterizing the inverse characteristics of the solid-state transformer's power transmission, thereby obtaining data for advance adjustment of the solid-state transformer. The desired feedforward control parameters for transformer power transmission are obtained by using a data-driven model to complete nonlinear compensation and feedforward parameter solving. The control synthesis module is responsible for combining the feedback control parameters generated by the feedback control loop based on the current operating state data with the desired feedforward control parameters to determine the target control signal and complete the fusion processing of different types of control parameters. The drive module is responsible for driving the power switching devices of the solid-state transformer according to the target control signal and converting the final control command into hardware actions. Each module is linked in sequence according to the process of data acquisition, load prediction, feedforward solving, signal synthesis, and hardware driving. The modular architecture makes the device structure clear, which not only facilitates mass production and on-site fault repair, but also allows individual modules to be replaced or optimized according to the performance requirements of different equipment. The entire device is deployed on a field-programmable gate array hardware platform equipped with an artificial intelligence acceleration unit, which can achieve microsecond-level data operation and signal output, perfectly matching the control requirements of fast response of solid-state transformers, and is widely used in the power supply systems of various artificial intelligence data centers.

[0026] The modular, AI-based solid-state transformer microsecond-level response control device is deployed on a hardware platform. The mainstream configuration uses Xilinx's new generation multi-core system-on-a-chip paired with a deep learning processing unit. Alternatively, Texas Instruments' C2000 series main control chip can be used with an external third-party AI acceleration daughter card. The entire device forms a complete pipeline of analog-to-digital conversion, digital filtering, AI inference, control signal synthesis, pulse width modulation generation, dead-time insertion, and power device driving. The entire process digital control latency is less than three microseconds. The data acquisition module uses an analog-to-digital converter chip with a sampling rate of no less than one megabit per second, and a two megabit per second specification is typically used to ensure sampling accuracy. Each functional module is independent yet works collaboratively. If a single module fails, it can be disassembled and repaired separately without affecting the operation of other modules. Based on a modular architecture, it can be expanded in multiple directions. For small single-unit solid-state transformers, multiple software modules can be lightweighted and integrated into a single embedded chip to reduce costs. For large multi-module power supply cabinets, a distributed modular design can be adopted, with each power module equipped with an independent local control module. A master control module is set up at the upper layer to complete global load prediction and scheduling. At the hardware level, a pure field-programmable gate array architecture can be used to replace the system-on-a-chip, further reducing inference latency. The entire modular device can be adapted to various main topologies such as dual active full-bridge and inductor-capacitor resonant, and has extremely strong versatility.

[0027] A third aspect of the present invention provides an AI-based control device for microsecond-level response of a solid-state transformer. The AI-based control device for microsecond-level response of a solid-state transformer includes a solid-state transformer, which comprises a main power circuit, a sensing unit, and a control processing unit. The main power circuit includes an input rectifier stage, a high-frequency isolation stage, and an output filter stage. The high-frequency isolation stage includes a dual active full-bridge converter circuit. The sensing unit is used to collect status signals from the main power circuit and the load side in real time. Based on the status signals collected by the sensing unit, the control processing unit executes the AI-based control method for microsecond-level response of the solid-state transformer according to any one of the above technical solutions to output a microsecond-level phase-shift control signal to the dual active full-bridge converter circuit.

[0028] The solid-state transformer in the AI-based solid-state transformer microsecond-level response control device specified in this application comprises three core components. The first part is the main power circuit, which includes an input rectifier stage, a high-frequency isolation stage, and an output filter stage. The high-frequency isolation stage employs a dual active full-bridge converter circuit. This classic three-stage topology combined with the dual active full-bridge core converter unit retains the hardware advantages of solid-state transformers, such as small size, high power density, and bidirectional power transmission, enabling the conversion and electrical isolation of medium-voltage AC to 800V DC. The second part is the sensing unit, which can collect real-time status signals from the main power circuit and the load side, providing uninterrupted operating condition data support for the entire intelligent control system. The third part is the control processing unit, which processes data based on the sensor unit's... The acquired status signal executes the control method for microsecond-level response of the solid-state transformer based on AI described in any of the above technical solutions, and finally outputs a microsecond-level phase-shift angle control signal to the dual active full-bridge converter circuit. The control processing unit deeply integrates the intelligent control algorithm with the hardware topology and outputs control signals adapted to the dual active full-bridge. Relying on the microsecond-level signal update speed, the device can quickly complete power regulation when encountering a load step with a rise time of only three microseconds, and control the time for the output voltage to return to stability within thirty microseconds. The large power supply cabinet composed of multiple of the above devices can still maintain a state of equal voltage and current under the condition of continuous operation for seven to twenty-four hours. The overall design of the software and hardware combination is highly adapted to the high-power, highly dynamic DC power supply scenario of artificial intelligence data centers.

[0029] A fourth aspect of the present invention provides an electronic device, comprising: a processor and a memory, wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform an AI-based control method for microsecond-level response of a solid-state transformer as described in any of the above-described technical solutions. Therefore, the electronic device provided by this application possesses all the beneficial effects of the aforementioned AI-based control method for microsecond-level response of a solid-state transformer.

[0030] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0031] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0032] Figure 1 This is a flowchart illustrating an AI-based control method for microsecond-level response of a solid-state transformer according to an embodiment of the present invention.

[0033] Figure 2This is a schematic block diagram of an AI-based solid-state transformer microsecond-level response control device according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic block diagram of an AI-based solid-state transformer microsecond-level response control device according to an embodiment of the present invention;

[0035] Figure 4 This is a schematic block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0036] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0038] The following reference Figures 1 to 4 This invention describes a control method, apparatus, and device for microsecond-level response of a solid-state transformer based on AI, according to some embodiments of the present invention.

[0039] like Figure 1 As shown, embodiments of this application provide an AI-based control method for microsecond-level response of a solid-state transformer. The control method includes:

[0040] Step 102: Obtain the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The current operating status data includes at least one of the following: input side voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output side voltage.

[0041] Step 104: Input the current electrical signal data into the first data-driven model for load forecasting to obtain the load demand forecast sequence within a preset time range in the future;

[0042] Step 106: Input the load demand prediction sequence and the current operating status data into the second data-driven model used to characterize the inverse characteristics of solid-state transformer power transmission, and obtain the expected feedforward control parameters for adjusting the power transmission amount of solid-state transformer in advance.

[0043] Step 108: Combine the feedback control parameters generated by the feedback control loop based on the current operating state data with the desired feedforward control parameters to determine the target control signal;

[0044] Step 110: Drive the solid-state transformer according to the target control signal.

[0045] This application, relying on comprehensive operating condition signal acquisition, data-driven load prediction, power transmission inverse model feedforward compensation, fusion of feedforward and feedback control, and power switching device driven execution, specifically addresses the shortcomings of solid-state transformer proportional-integral controllers: limited bandwidth, passive response to load changes, difficulty in adapting to the nonlinearity of power transmission in dual active full-bridge topologies, reliance on manual tuning of control parameters for each unit, and large voltage deviations and long recovery times caused by microsecond-level load steps in AI data centers. First, it constructs a complete operating condition perception system by comprehensively acquiring multi-dimensional operating status data of the solid-state transformer's main power circuit and load-side electrical signal data. Then, it utilizes a first data-driven model specifically designed for load prediction to achieve microsecond-scale load trend prediction, thereby changing the post-control adjustment mode and compensating for the shortcomings of proportional-integral controllers. To address the shortcoming of controller response lag, a second data-driven model, which characterizes the inverse power transmission characteristics of solid-state transformers, is used to generate desired feedforward control parameters for early power transmission adjustment. This model automatically fits various nonlinear influencing factors such as switch dead zone, core magnetic saturation, circuit parasitic parameters, and device temperature drift, significantly simplifying the on-site parameter tuning process. Simultaneously, the desired feedforward control parameters are combined with the feedback control parameters generated by the feedback loop, integrating the dual advantages of rapid dynamic compensation from feedforward and precise steady-state correction from feedback. This further overcomes the bandwidth limitations of the controller. Finally, the synthesized target control signal drives the power switching devices to complete hardware actions. The entire technical solution works together to enable solid-state transformers to have microsecond-level dynamic load response capabilities suitable for the harsh operating conditions of artificial intelligence data centers.

[0046] This application, according to a control method, sequentially executes operations to acquire the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The acquired current operating status data includes at least one of the following: input voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output voltage. This data acquisition method can rely on the device's built-in sensing unit to achieve full-link condition perception, no longer limited to single-point signal acquisition, and can completely capture the real-time operating status of the solid-state transformer's power input, isolation conversion, power output, and downstream load. Complete raw data can provide a reliable basis for subsequent intelligent model calculations and feedback control loops, avoiding deviations in control judgments due to missing information. It fully adapts to the high-precision control requirements brought about by microsecond-level load mutations in artificial intelligence data centers. After data acquisition, the current electrical signal data is input into the first data-driven model for load prediction, thereby obtaining a load demand prediction sequence within a preset time range. By relying on the first data-driven model trained with massive amounts of operating data to mine the temporal variation patterns of electrical signals, it is possible to predict the power fluctuations that will occur in the load in advance. This effectively avoids the situation where the solid-state transformer is only passively detected and adjusted after the load changes. It also makes up for the shortcomings of insufficient bandwidth and adjustment lag of the proportional-integral controller, and reserves sufficient power adjustment time for the equipment. Then, the load demand prediction sequence and the current operating status data are input into the second data-driven model to characterize the inverse characteristics of power transmission of the solid-state transformer, thereby obtaining the expected feedforward control parameters for adjusting the power transmission of the solid-state transformer in advance.

[0047] The second data-driven model can deduce matching control parameters based on the predicted load demand, automatically adapting to the power transmission nonlinearity problem inherent in the dual active full-bridge topology. It also counteracts interference from non-ideal factors such as switching dead zones, core magnetic saturation, circuit parasitic parameters, and device temperature drift, eliminating the need for personnel to individually tune parameters for a large number of power modules. Therefore, this application significantly simplifies the on-site commissioning process. Next, the target control signal is determined by combining the feedback control parameters generated from the current operating state data by the feedback control loop with the desired feedforward control parameters. The feedback control loop continuously corrects minor operating condition deviations generated during steady-state operation, ensuring long-term stable output power. The feedforward control parameters specifically address sudden, large load jumps. These two types of control parameters complement each other, enabling the control system to simultaneously possess high steady-state accuracy and fast dynamic response capabilities. The power switching devices of the solid-state transformer are driven by the target control signal. The control commands generated by the algorithm are transformed into the actual actions of the hardware circuit. By changing the working state of the power switching devices, the power transmission capacity of the whole machine is adjusted, forming a complete control closed loop from signal acquisition, intelligent calculation, parameter fusion to hardware execution. The whole method can systematically solve many defects of solid-state transformers, such as lack of load prediction capability, poor nonlinear adaptation, response lag, and cumbersome parameter tuning. It can operate stably in a single solid-state transformer module with a rated power of 167 kilowatts and in a large power supply equipment of 5,000 kilowatts composed of multiple modules. When facing the rapid power step conditions of GPU clusters in artificial intelligence data centers, it can effectively suppress voltage drop and overshoot, and control voltage deviation and recovery time within the stringent industry indicators.

[0048] In one embodiment, the sensing unit is equipped with an analog-to-digital converter (ADC) chip to complete global signal acquisition. The ADC chip can acquire electrical signals at a rate of 2 megabits per second, with a sampling interval of about 0.5 microseconds. The acquired continuous time-series data will sequentially complete a series of pipeline operations such as digital filtering, intelligent model inference, control parameter synthesis, and pulse width modulation register update. The entire digital process takes a short time. Combined with the switching transmission delay of the dual active full-bridge circuit, the overall dynamic response time of the device can be maintained at about 30 microseconds. The two types of data-driven models can also continuously and slightly optimize their own parameters based on the measured data generated during the operation of the device. For multi-module devices with a series input and parallel output architecture, the global load prediction result can be distributed to each power unit. Each unit is matched with its local second data-driven model for slight adaptation, so that the power distribution state of the entire device remains stable.

[0049] In some embodiments of this application, the first data-driven model includes a time-series neural network; inputting current electrical signal data into the first data-driven model for load prediction to obtain a load demand prediction sequence within a preset future time range includes: obtaining historical current data, historical voltage data, and current change rate data within a sliding time window as current electrical signal data; inputting the current electrical signal data into the time-series neural network to output a load demand prediction sequence within a preset future time range; the load demand prediction sequence includes at least one of the following: predicted load current value, predicted load power value, or load power change amount at multiple future prediction times.

[0050] In the above embodiments, the first data-driven model is defined as including a time-series neural network. During load demand forecasting using this model, historical current data, historical voltage data, and current change rate data within a sliding time window are first selected as the current electrical signal data. The sliding time window can capture continuous time-series electrical data, fully preserving the trend characteristics of load changes. Using this as input to the time-series neural network maximizes the reproduction of the dynamic change patterns of the load. The processed current electrical signal data is then input into the time-series neural network, which outputs a load demand forecast sequence within a preset future time range. The time-series neural network's advantage in analyzing time-series data further improves the accuracy of the forecast results. Furthermore, the load demand forecast sequence includes at least one of the following: predicted load current values, predicted load power values, or changes in load power at multiple future forecast times. The diverse predictive output formats can adapt to the data usage requirements of different control logics, and the future preset time range is limited to 5 microseconds to 50 microseconds. This time range accurately matches the operating conditions of microsecond-level load jumps in artificial intelligence data centers. Combined with the signal acquisition unit of the equipment with a sampling rate of 2 megabits per second, the temporal neural network completes an inference calculation every five microseconds and outputs load change data for the next 5 to 25 microseconds. Compared with the method of making predictions based on a single electrical parameter or a wide time range, this limited method improves the timeliness and accuracy of load prediction simultaneously, providing a more realistic calculation basis for subsequent feedforward control links, and improving the overall control scheme's ability to cope with microsecond-level load jumps from the source.

[0051] In one embodiment, the sliding time window can select 256 consecutive sampling points to form a data sample. The duration of the sample is in the range of 128 microseconds to 1 millisecond. In addition to using causal dilated convolutional networks and long short-term memory networks, the temporal neural network can also use similar temporal models such as gated recurrent units and lightweight converters to adapt to hardware carriers with different computing capabilities. The inference latency of different models varies, and can be flexibly selected according to equipment cost and performance requirements.

[0052] In some embodiments of this application, while or before acquiring the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data on the load side, the control method further includes: receiving a data center task scheduling forecast signal issued by an external management device; inputting the current electrical signal data into a first data-driven model for load prediction to obtain a load demand prediction sequence within a future preset time range, including: using the task scheduling forecast signal as an auxiliary input, fusing it with the current electrical signal data, and inputting it into the first data-driven model to obtain a load demand prediction sequence within a future preset time range.

[0053] In the above embodiments, while performing the data acquisition step or in advance, the system receives a data center task scheduling forecast signal from an external management device. This external management device, acting as the control unit for coordinating computing power tasks in the AI ​​data center, can send forecast information about impending load changes milliseconds to tens of milliseconds in advance. This prior information supplements the local electrical sampling data with clues about load changes. When using the first data-driven model for load prediction, the task scheduling forecast signal is used as an auxiliary input, fused with the current electrical signal data, and input into the first data-driven model to obtain a predicted load demand sequence within a preset time range. This dual-input mode, combining local real-time electrical data with the upper-level scheduling forecast signal, further extends the warning time for load changes, eliminating reliance solely on on-site sampling signals and improving practicality. In typical computing power task scenarios such as GPU batch startup and model loading, the prediction results after fusing the scheduling signal can capture load step trends earlier, allowing the solid-state transformer more time to complete pre-emptive power adjustment, further reducing the impact of load fluctuations on output voltage and improving the adaptability of the entire control scheme in actual AI data center operation and maintenance scenarios.

[0054] In one embodiment, the device receives scheduling forecast signals via communication links such as Ethernet or remote direct memory access. This type of communication can ensure the real-time transmission of instructions. Large data centers are divided into multiple independent computing power partitions. The system can receive multiple scheduling signals corresponding to the load status of different partitions. By combining the partition information, the system can complete the load prediction of different regions, so that the power adjustment action can match the power consumption change rhythm of different regions.

[0055] In some embodiments of this application, the solid-state transformer includes an isolated dual active full-bridge topology, and the desired feedforward control parameter is the feedforward phase shift compensation amount of the isolated dual active full-bridge topology. The load demand prediction sequence and current operating status data are input into a second data-driven model used to characterize the inverse characteristics of the solid-state transformer's power transmission. The desired feedforward control parameter for adjusting the power transmission amount of the solid-state transformer in advance is obtained. This includes inputting at least one of the following into the second data-driven model: the load prediction current from the load demand prediction sequence, and at least one of the following from the current operating status data: the measured output voltage, the input DC bus voltage, the measured primary current, the core temperature, the power switching device temperature, and the dead-zone compensation parameter. The output is the feedforward phase shift compensation amount. The second data-driven model includes at least one of a deep feedforward neural network, a lookup table model, and a regression model.

[0056] In the above embodiments, this application explicitly defines the solid-state transformer as an isolated dual-active full-bridge topology. The corresponding desired feedforward control parameter is the feedforward phase-shift angle compensation amount of the isolated dual-active full-bridge topology. The dual-active full-bridge topology relies on the phase-shift angle to adjust the power transmission capability. By specifically defining the feedforward control parameter as the phase-shift angle compensation amount, the calculation result of the intelligent model can be directly matched with the control logic of the core topology. In the process of solving the feedforward phase-shift angle compensation amount using the second data-driven model, the load prediction current in the load demand prediction sequence, as well as the output measured voltage, input DC bus voltage, and primary side measured current in the current operating status data are used. At least one of the following parameters—core temperature, power switching device temperature, and dead-zone compensation parameters—is used as model input. This multi-dimensional input comprehensively covers various factors affecting the power transmission of dual active full-bridge transformers. It includes basic electrical operating parameters as well as physical parameters causing transmission nonlinearity, such as core operating temperature, device temperature, and switching dead zone. This allows the second data-driven model to fully fit the actual operating characteristics of the topology. The second data-driven model is specified to include at least one of the following: deep feedforward neural network, lookup table model, and regression model. Different types of data-driven models can adapt to different hardware computing capabilities and cost scenarios. High-precision industrial equipment can use deep feedforward neural networks to achieve complex nonlinear fitting. Low-cost small equipment can use lookup table models or regression models to reduce computational overhead. All models can stably output the feedforward phase shift compensation amount matching the operating conditions, effectively solving problems such as control oscillation and insufficient accuracy caused by the nonlinearity of power transmission in dual active full-bridge topologies, ensuring the control stability of different types of solid-state transformers under all operating conditions.

[0057] In one embodiment, the phase shift angle refers to the phase difference between the primary and secondary full-bridge circuits of the dual active full-bridge circuit. The deep feedforward neural network can be configured with three to five fully connected layers. Combined with the linear rectification activation function and the batch normalization structure, the computational effect is optimized. The network output will be constrained to the phase range of the adaptive circuit through the activation function. In addition to the models listed in this solution, Gaussian process, extreme gradient boosting, and other models can also be used as supplements. This set of feedforward compensation logic can also be transferred to other power conversion circuits such as inductor-capacitor resonant topology and matrix-type cross-conversion.

[0058] In some embodiments of this application, the target control signal is determined by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters, including: superimposing the feedback control parameters and the desired feedforward control parameters, and performing amplitude limiting processing on the superimposed parameters to generate the target control signal.

[0059] In the above embodiments, the feedback control parameters generated by the feedback control loop based on the current operating state data are superimposed with the desired feedforward control parameters. This superposition of two control parameters with different functions integrates steady-state correction and dynamic pre-compensation control capabilities, allowing the target control signal to simultaneously meet the requirements of long-term voltage regulation and instantaneous disturbance rejection. After parameter superposition, the superimposed parameters are then amplitude-limited to finally generate the target control signal. Combined with the normal operating range of the phase shift angle of the dual active full-bridge topology, the amplitude-limiting operation can constrain the control parameters within the allowable operating range of the power switching devices and high-frequency transformer, preventing the superimposed parameters from exceeding hardware limits and causing faults such as device overdrive, circuit resonance, and output oscillation. In a load step test scenario simulating 80% of rated power, the parameter superposition combined with amplitude-limiting ensures that the control signal remains within a safe range, the solid-state transformer does not experience operational abnormalities, and the fluctuation range of the output voltage is effectively limited, achieving dual protection of control performance and equipment operational safety.

[0060] In one embodiment, the operator can flexibly adjust the range of the limit based on the real-time load of the equipment and the operating temperature of the power devices. When the load is too large, the range can be narrowed appropriately to improve the stability of operation, and when the load is too small, the range can be widened appropriately to improve the flexibility of response. The limit logic can also be used in conjunction with the auxiliary parameters output by the subsequent reinforcement learning. After multiple control parameters are superimposed, the boundary constraints are uniformly completed.

[0061] In some embodiments of this application, the control method further includes: inputting current operating state data into a pre-trained reinforcement learning agent model to obtain control correction parameters; using the control correction parameters to update the proportional-integral gain parameters of the feedback control loop, or superimposing the control correction parameters with the feedback control parameters and the desired feedforward control parameters to generate a target control signal.

[0062] In the above embodiments, the current operating status data is input into a pre-trained reinforcement learning agent model to obtain control correction parameters. The reinforcement learning agent model completes offline training based on a simulation environment and can autonomously analyze the control status and output correction parameters according to the real-time operating conditions of the equipment. Subsequently, two different application methods can be selected to process the control correction parameters. The first method is to use the control correction parameters to update the proportional-integral gain parameters of the feedback control loop. The proportional-integral gain of solid-state transformers requires on-site tuning for each unit, and fixed parameters cannot adapt to complex scenarios such as component aging, environmental temperature changes, and large-scale load fluctuations. The reinforcement learning agent model, however, can update the gain parameters online every tens of microseconds to milliseconds, achieving adaptive adjustment of the proportional-integral parameters, eliminating the manual tuning step and improving parameter adaptability. The second method is to superimpose the control correction parameters with the feedback control parameters and the desired feedforward control parameters to generate the target control signal, integrating the auxiliary parameters output by reinforcement learning into the overall control logic to further optimize the phase shift angle adjustment accuracy. In a large-capacity solid-state transformer system consisting of thirty modules connected in parallel, the above-mentioned supplementary process can continuously optimize the control effect of the entire unit, reduce the steady-state error of the output voltage, and improve the operational stability under frequent switching conditions between heavy and light loads.

[0063] In one embodiment, the reinforcement learning agent collects information such as voltage error, load prediction data, and the phase shift angle of the previous time step as state input. After offline training based on the SAC algorithm or the near-end policy optimization algorithm, it is transformed into a lightweight network and deployed in the control hardware. The network performs a condition assessment and outputs correction parameters approximately every 100 microseconds. Alternatively, fuzzy self-tuning logic can be used as an alternative to adapt to hardware devices with limited computing resources. For multi-module parallel devices, a unified reinforcement learning model can be set to synchronously optimize the control parameters of all power units.

[0064] In some embodiments of this application, the step of determining the target control signal by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the expected feedforward control parameters further includes: real-time detection of the output state of the first data-driven model and the second data-driven model; when it is detected that the output of the first data-driven model or the second data-driven model exceeds the preset physical range, an invalid value appears, inference times out, or the model self-check fails, the expected feedforward control parameters are hot-bypassed so that the system automatically reverts to the pure feedback control mode where the target control signal is determined solely based on the feedback control parameters generated by the feedback control loop.

[0065] In the above embodiments, throughout the entire process of determining the target control signal, real-time detection of the output status of the first data-driven model and the second data-driven model is carried out simultaneously, continuously monitoring the calculation results and operating status of the two intelligent models. When anomalies such as the output of the first data-driven model or the second data-driven model exceeding the preset physical range, invalid values ​​appearing, inference timeouts, or model self-check failures are detected, a hot bypass operation is immediately performed on the expected feedforward control parameters, allowing the system to automatically switch to a pure feedback control mode that determines the target control signal solely based on the feedback control parameters generated by the feedback control loop. This anomaly detection and automatic degradation mechanism uses intelligent prediction and feedforward functions as performance enhancement modules, and a mature feedback control loop as a fault fallback solution. Artificial intelligence data centers have extremely high requirements for the continuous operation capability of power supply equipment. Various operational anomalies in the model will not cause the entire control system to fail. In test scenarios where the artificial model inference timeouts and outputs invalid values ​​are simulated, the system can complete the mode switch in a very short time. The output voltage of the solid-state transformer will not experience sudden changes or power outages. Relying on multi-level fault-tolerant design, the overall operational reliability of the equipment is greatly improved, meeting the requirements of continuous power supply for data centers around the clock.

[0066] In one embodiment, hot bypass refers to a switching method that cuts off the transmission channel of feedforward control parameters and retains the independent operation of the feedback loop when the intelligent computing module malfunctions. During the device power-on initialization phase, the pure feedback control mode will be enabled by default. After the two types of data-driven models complete their self-tests, the complete control logic will be switched in. Multi-module devices can also achieve independent bypass of a single module. If a single module malfunctions, it will not interfere with the normal operation of other modules. The system can also retain relevant operational data of abnormal events to facilitate subsequent troubleshooting by maintenance personnel.

[0067] like Figure 2As shown, an embodiment of this application provides an AI-based control device 200 for microsecond-level response of a solid-state transformer. The control device includes: a data acquisition module 210, a prediction module 220, a feedforward calculation module 230, a control synthesis module 240, and a drive module 250. The data acquisition module 210 is used to acquire the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The current operating status data includes at least one of the following: input side voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output side voltage. The prediction module 220 is used to convert the current electrical signal... The data input is used to a first data-driven model for load forecasting to obtain a load demand forecast sequence within a preset time range in the future; the feedforward calculation module 230 is used to input the load demand forecast sequence and the current operating status data into a second data-driven model to characterize the inverse characteristics of the power transmission of the solid-state transformer, and obtain the expected feedforward control parameters for adjusting the power transmission of the solid-state transformer in advance; the control synthesis module 240 is used to combine the feedback control parameters generated by the feedback control loop based on the current operating status data and the expected feedforward control parameters to determine the target control signal; the drive module 250 is used to drive the solid-state transformer according to the target control signal.

[0068] The AI-based solid-state transformer microsecond-level response control device 200 specified in this application consists of multiple functional modules. The data acquisition module 210 is responsible for acquiring the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The acquired current operating status data includes at least one of the following: input voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output voltage. These multiple functional modules undertake the basic data acquisition work for the entire device, relying on a high-speed sampling chip to complete the collection of signals under all operating conditions, providing raw data for all subsequent calculation stages. The prediction module 220 is responsible for inputting the current electrical signal data into a first data-driven model for load prediction, thereby obtaining a load demand prediction sequence within a preset time range, specifically for intelligent prediction of load trends. The feedforward calculation module 230 is responsible for inputting the load demand prediction sequence and the current operating status data into a second data-driven model characterizing the inverse characteristics of power transmission in the solid-state transformer, thereby obtaining data for advance adjustment. The desired feedforward control parameters for the power transmission of the solid-state transformer are obtained by using a data-driven model to complete nonlinear compensation and feedforward parameter solving. The control synthesis module 240 is responsible for combining the feedback control parameters generated by the feedback control loop based on the current operating state data with the desired feedforward control parameters to determine the target control signal and complete the fusion processing of different types of control parameters. The drive module 250 is responsible for driving the power switching devices of the solid-state transformer according to the target control signal and converting the final control command into hardware action. Each module is linked in sequence according to the process of data acquisition, load prediction, feedforward solving, signal synthesis, and hardware driving. The modular architecture makes the device structure clear, which not only facilitates mass production and on-site fault diagnosis, but also allows individual modules to be replaced or optimized according to the performance requirements of different equipment. The entire device is deployed on a field-programmable gate array hardware platform equipped with an artificial intelligence acceleration unit, which can achieve microsecond-level data operation and signal output, perfectly matching the control requirements of the fast response of the solid-state transformer. It is widely used in the power supply systems of various artificial intelligence data centers.

[0069] The modular, AI-based solid-state transformer microsecond-level response control device 200 is deployed on a hardware platform. The mainstream configuration uses a multi-core on-chip system with a deep learning processing unit. Alternatively, a third-party AI acceleration daughter card can be connected to the main control chip. The entire device forms a complete pipeline of analog-to-digital conversion, digital filtering, AI inference, control signal synthesis, pulse width modulation generation, dead-time insertion, and power device driving. The entire process digital control latency is less than three microseconds. The data acquisition module 210 uses an analog-to-digital conversion chip with a sampling rate of no less than one megabit per second, and a standard two megabit per second specification is used to ensure sampling accuracy. Each functional module is independent yet works collaboratively. If a single module fails, it can be disassembled and repaired separately without affecting the operation of other modules. Based on a modular architecture, it can be expanded in multiple directions. For small single-unit solid-state transformers, multiple software modules can be lightweighted and integrated into a single embedded chip to reduce costs. For large multi-module power supply cabinets, a distributed modular design can be adopted, with each power module equipped with an independent local control module. A master control module is set up at the upper layer to complete global load prediction and scheduling. At the hardware level, a pure field-programmable gate array architecture can be used to replace the system-on-a-chip, further reducing inference latency. The entire modular device can be adapted to various main topologies such as dual active full-bridge and inductor-capacitor resonant, and has extremely strong versatility.

[0070] like Figure 3 As shown in the embodiments of this application, an AI-based solid-state transformer microsecond-level response control device 300 is provided. The AI-based solid-state transformer microsecond-level response control device 300 includes a solid-state transformer 301. The solid-state transformer 301 includes a main power circuit 310, a sensing unit 320, and a control processing unit 330. The main power circuit 310 includes an input rectifier stage, a high-frequency isolation stage, and an output filter stage. The high-frequency isolation stage includes a dual active full-bridge converter circuit. The sensing unit 320 is used to collect the status signals of the main power circuit 310 and the load side in real time. Based on the status signals collected by the sensing unit 320, the control processing unit 330 executes the AI-based solid-state transformer microsecond-level response control method of any one of the above embodiments to output a microsecond-level phase shift angle control signal to the dual active full-bridge converter circuit.

[0071] The solid-state transformer 301 defined in this application comprises three core components. The first part is the main power circuit 310, which includes an input rectifier stage, a high-frequency isolation stage, and an output filter stage. The high-frequency isolation stage employs a dual active full-bridge converter circuit. This three-stage classic topology combined with the dual active full-bridge core converter unit retains the hardware advantages of the solid-state transformer 301, such as small size, high power density, and bidirectional power transmission. It can complete the conversion and electrical isolation of medium-voltage AC to 800V DC. The second part is the sensing unit 320, which can collect the status signals of the main power circuit 310 and the load side in real time, providing uninterrupted operating condition data support for the entire intelligent control system. The third part is the control processing unit 330, which will... The state signal collected by the sensing unit 320 executes the solid-state transformer 301 control method described in any of the above embodiments, and finally outputs a microsecond-level phase shift angle control signal to the dual active full-bridge converter circuit. The control processing unit 330 deeply integrates the intelligent control algorithm with the hardware topology and outputs a control signal adapted to the dual active full-bridge. Relying on the microsecond-level signal update speed, the device can quickly complete power regulation when encountering a load step with a rise time of only three microseconds, and control the time for the output voltage to return to stability within thirty microseconds. The large power supply cabinet composed of multiple of the above devices can still maintain a state of equal voltage and current under the condition of continuous operation for seven to twenty-four hours. The overall design of the software and hardware combination is highly adapted to the high-power, highly dynamic DC power supply scenario of artificial intelligence data centers.

[0072] like Figure 4 As shown in the embodiments of this application, an electronic device 400 is provided, including a processor 402 and a memory 404. The memory 404 stores instructions executable by the processor 402. The instructions are executed by the processor 402 to enable the processor 402 to execute the AI-based solid-state transformer microsecond-level response control method as described in any of the above embodiments. Therefore, the electronic device 400 provided by this application possesses all the beneficial effects of the above-described AI-based solid-state transformer microsecond-level response control method.

[0073] In one embodiment of this application, this embodiment is a complete implementation system and control method for artificial intelligence (AI) empowering solid state transformers (SSTs). It is specifically applied to the 800-volt DC power supply scenario of artificial intelligence data centers (AIDCs). It addresses the microsecond-level high-power load step problem generated by computing power clusters of graphics processing units (GPUs), heterogeneous processing units (HPUs), and neural processing units (NPUs) in scenarios such as model loading, inference start / stop, and breakpoint saving. It fully incorporates hardware topology, sensor communication, artificial intelligence (AI) control algorithms, execution pipelines, fault tolerance and degradation mechanisms, and expansion schemes. The following is a complete description from four dimensions: overall system architecture, step-by-step operation process, working condition testing, and scheme expansion.

[0074] I. Overall Hardware and Software Architecture of the System

[0075] (a) SST Main Power Circuit

[0076] This embodiment employs a classic three-stage topology: a cascaded H-bridge (CHB) input rectifier stage, a dual active bridge (DAB) high-frequency isolation stage, and an output LC filter stage. It also utilizes an input series output parallel (ISOP) multi-module architecture. Each S-module is connected to a 13.8 kV medium-voltage AC input. The CHB rectifier circuit consists of 30 2300-volt silicon carbide (SiC) metal-oxide-semiconductor field-effect transistors (MOSFETs), generating approximately 1500 volts of intermediate DC bus voltage (Udc). The high-frequency isolation DAB topology switching frequency (… The frequency is set to 50 kHz to 100 kHz, the high-frequency transformer turns ratio is 2:1, and the equivalent leakage inductance is ( The value is 15 microhenries, relying on the phase shift angle ( ) Adjust power transfer; DAB power transfer characteristics meet Nonlinear relationship; after passing through an inductor-capacitor (LC) filter circuit and a DC circuit breaker on the output side, 800 volt DC power is output to power the AIDC computing rack. The rated power of a single module is 167 kilowatts. 30 modules of the same specification are combined in an ISOP architecture to build a large SST power supply cabinet with a total power of 5000 kilowatts.

[0077] (ii) Hybrid Sensing and Communication Unit

[0078] The system is equipped with an analog-to-digital converter (ADC) with a sampling rate of at least 1 mega-samples per second (MSPS) and a single-step sampling interval of ≤1 microsecond. It acquires electrical parameters in real time, including SST output voltage (Vout), output current (Iout), intermediate DC bus voltage, and primary-side current of the D-topology. A temperature monitoring loop is also added to acquire the high-frequency transformer core temperature (Tc) and power device heatsink temperature (Th) for online model correction. The communication interface integrates Ethernet and Remote Direct Memory Access (RDMA), enabling real-time monitoring of task start announcements and power configuration files issued by the AIDC cluster scheduler. These announcement signals can push load change instructions several milliseconds to tens of milliseconds in advance, serving as auxiliary input for the AI ​​prediction module.

[0079] (III) AI Control and Processing Unit

[0080] The control hardware uses a Zynq UltraScale+ System on Chip (SoC) paired with a Data Processing Unit (DPU). It is also compatible with a hardware solution of Texas Instruments C2000 main controller + third-party Neural Processing Unit (NPU) daughter card. It integrates three major AI function sub-modules and a microsecond-level execution pipeline, while retaining a proportional-integral (PI) dual closed loop as a fallback control loop.

[0081] The load prediction submodule employs either a Temporal Convolutional Network (TCN) or a Long Short-Term Memory (LSTM) neural network. The network input consists of the output current, output voltage, and rate of change of current within a sliding time window (W). It can integrate scheduling forecast signals (Lpre); the sliding window captures 128 sampling points (corresponding to a duration of 128 microseconds to 1 millisecond); offline network training uses AIDC measured load step data and simulated synthetic data, and the loss function is mean squared error (MSE); the online inference period is set to 1 microsecond to 5 microseconds, and it can output the predicted load current values ​​at three time nodes: 5 microseconds, 10 microseconds, and 20 microseconds. In this embodiment, the TCN is set as a 4-layer dilated convolutional structure with a kernel size of 3 and a number of convolutional channels of 64.

[0082] The SST power transfer inverse model feedforward submodule employs a deep neural network (DNN) to construct the power transfer inverse model. The network has a 3-5 layer fully connected structure, using a rectified linear unit (ReLU) activation function with batch normalization (BatchNorm) to optimize training performance. The output layer uses a sigmoid activation function to constrain the phase shift angle to a specific value. The range; the model input vector includes the target predicted current, the output measured voltage, the intermediate DC bus voltage, the DAB primary current, and the core temperature.

[0083] The model input vector X = [Îout_target, Vo_meas, Udc_meas, Iprim_meas, Tc_magnet] is used to solve the feedforward phase shift angle based on the load demand. The model is pre-trained offline using SST open-loop parameter scanning data, a Functional Mock-up Unit (FMU), and Simulink simulation data. Online fine-tuning based on measured data is supported during operation. In this embodiment, the number of nodes in the DNN network layer is [128, 64, 32, 1].

[0084] The Deep Reinforcement Learning (DRL) dynamic tuning submodule models the SST voltage regulation control as a Markov Decision Process (MDP). The state variables include voltage error, historical voltage error, error integral, predicted load current, current change, and the phase shift angle from the previous time step. The action variable is the PI proportional gain correction. ), integral gain correction ( ) or auxiliary phase shift angle ( The reward function uses voltage deviation, phase angle saturation, and phase angle change rate as constraints. In the offline phase, training is performed on PLECS and Simulink power electronics simulation platforms using the SAC algorithm (Soft Actor-Critic, SAC) or the Proximal Policy Optimization (PPO) algorithm. The inference network is simplified to 2-3 fully connected layers (FC). In the online state, the correction value is output every tens of microseconds to several milliseconds. In this embodiment, the number of nodes in the DRL policy network is [64, 32, 16]. The dynamic parameter tuning submodule uses a deep reinforcement learning agent (DRLAgent).

[0085] The SST output voltage regulation control is modeled as a continuous control MDP (Markov Decision Process).

[0086] State t =[Verr(k),Verr(k-1),ΣVerr(k)×Ts,Îout(k),dIout / dt,φ(k-1)]; where Verr(k) is the voltage error at the current moment, Verr(k-1) is the voltage error at the previous moment, ΣVerr(k) is the voltage error integral term, Ts is the system sampling period, Îout(k) is the predicted load current at the current moment, dIout / dt is the output current change rate, and φ(k-1) is the synthesized phase shift angle of the previous period.

[0087] Actiona t =ΔK p ,ΔK i (PI gain correction), or directly output the φ_DRL component;

[0088] Among them, a t ΔK is the MDP action output by the DRL agent. p It is the proportional gain correction value of the PI controller, ΔK iφ_DRL is the integral gain correction for the PI controller, and φ_DRL is the auxiliary phase shift angle component of the reinforcement learning agent output.

[0089] Reward function t =-[α×Verr 2 +β×φ_saturation_penalty+γ×|Δφ / Δt|];

[0090] Where, r t Verr is the single-step instant reward value, α is the voltage deviation weighting coefficient, and Verr is the voltage deviation weighting coefficient. 2 φ is the squared term for voltage error, β is the weight of phase angle saturation penalty, φ_saturation_penalty is the phase angle saturation penalty term, γ is the weight of phase angle change rate, and |Δφ / Δt| is the absolute value of phase angle change rate.

[0091] The algorithm is trained offline in the SST simulation environment (PLECS / Simulink) using SoftActorCritic (SAC) or PPO, and then ported to a lightweight DNN (deep neural network) inference (2-3 layers FC).

[0092] The PI correction is updated online at a low frequency (every tens of microseconds to several milliseconds), or the φ_DRL is mixed and superimposed to the final phase shift angle:

[0093] φ_final=sat(φ_PI+φ_ff+w×φ_DRL,0,π / 2); where φ_final is the final synthesized phase shift angle fed into the DAB topology, sat() is the saturation limiting function, φ_PI is the phase shift angle component of the PI feedback loop output, φ_ff is the feedforward phase shift angle of the second data-driven model output, w is the weight of the DRL auxiliary phase shift angle superposition, φ_DRL is the reinforcement learning output auxiliary phase shift angle, and 0,π / 2 are the lower and upper limits of the phase shift angle allowed by the hardware, respectively.

[0094] Microsecond-level execution pipeline: The complete link is ADC sampling → moving average / infinite impulse response (IIR) digital filtering → TCN inference (1 microsecond to 2 microseconds) + DNN inference (<0.5 microseconds) + DRL table lookup (<0.2 microseconds) → phase angle synthesis → pulse width modulation (PWM) register update → dead time insertion → optocoupler / magnetically isolated drive power devices. The total delay of the entire digital control system is <3 microseconds.

[0095] (iv) Safety net and fault tolerance module

[0096] The system retains a dual closed-loop PI control loop consisting of an outer voltage loop and an inner current loop. AI-related functional modules support hot bypass. When the AI ​​model exhibits abnormalities such as output exceeding limits, invalid values, inference timeouts, or self-test failures, the system automatically switches to pure PI control mode to ensure uninterrupted equipment operation.

[0097] II. Complete System Operation Steps

[0098] Step 1 Power-on initialization

[0099] After the device is powered on, the PI dual closed-loop circuit is activated first, and the initial PI parameters are set to the proportional coefficient. Integral coefficient The SST output is gradually established to the rated voltage of 800 volts; simultaneously, the AI ​​module loads the offline-trained TCN, DNN, and DRL network weights, completes hardware self-test, and remains in bypass mode, awaiting the cut-in command. That is, the traditional PI dual closed-loop (K... p =0.05,K i =200), output voltage is built up to 800V, AI module loads pre-trained weights (TCN: 4-layer dilatedcausalconv, kernel=3, filters=64; inverse model DNN: 4-layer FC [128,64,32,1]; DRL policy network: 3-layer FC [64,32,16]), in bypass state waiting for self-test to pass.

[0100] Step 2: Normal Load Prediction and Signal Fusion

[0101] The ADC continuously samples at a rate of 2 Mbps, buffering a set of data every 0.5 μs. The system extracts 256 consecutive sampling points each time to construct a sliding window input TCN network. The network performs inference every 5 μs, outputting a load current prediction sequence for the next 5 to 25 μs. When the RDMA interface receives a GPU batch start warning signal from the AIDC scheduler 2 milliseconds in advance, it fuses the scheduling signal with local electrical data to directly correct the load prediction baseline value. That is, the ADC samples Iout and Vout at 2 MSPS, buffers one point every 0.5 μs, and constructs the input tensor from the most recent 256 points. The TCN infers every 5 μs and outputs Îout(k+1)~Îout(k+5) (corresponding to the load current prediction for the next 5 to 25 μs). If the scheduling bus receives a "GPUBatchStart→+80kW" warning (2 ms in advance), it directly sets Îout to the corresponding step-overload base load.

[0102] Step 3: Calculation of feedforward phase angle in inverse model

[0103] The predicted load current in the second microsecond is selected as the target value. Combined with the measured output voltage of 800.2 volts, the intermediate DC bus voltage of 1500 volts, the primary current of the DAB, the core temperature, and other parameters, the basic feedforward phase shift angle is calculated by inputting them into the DNN inverse model. The final feedforward phase shift angle is obtained by adding a dead zone compensation of 0.02 radians to the radians. Radius. That is, with Îout_target = Îout(k+2), Vo_meas = 800.2V, and Udc_meas = 2980V fed into the DNN inverse model, the output φ_ff = 0.237 rad (≈13.6° phase shift). After looking up the table and adding minimum dead zone compensation Δφ_dead = 0.02 rad, φ_ff' = 0.257 rad is obtained.

[0104] Step 4: Dynamic Correction of DRL Parameters

[0105] The DRL agent performs a condition assessment every 100 microseconds and outputs the PI parameter correction. , It updates the PI basic parameters in real time and outputs the auxiliary phase shift angle. Set the superposition weight .

[0106] Step 5: Control Quantity Synthesis and PWM Update

[0107] According to the saturation constraint formula The final phase shift angle is calculated as follows: The saturation constraint function is used to limit the calculated phase shift angle within the legal operating range, write the processed final phase shift angle result into the PWM phase register, and complete the parameter update in the next PWM cycle; the entire digital process is delayed by about 1.8 microseconds, and after adding the DAB half-switching cycle (5 microseconds), the total control signal update delay is less than 10 microseconds. It is a PI control phase shift angle.

[0108] Step 6 Automatic Degradation for Abnormalities

[0109] During operation, the running status of the two data-driven models is monitored in real time. If the output exceeds the physical range, invalid values ​​appear, inference timeout occurs, or model self-check fails, a hot bypass operation is immediately performed on the feedforward and prediction-related control quantities. The system seamlessly switches to pure PI dual-closed-loop control and uploads fault alarm information. SST maintains normal power supply. That is, if the AI ​​module's self-check exceeds the limit (predicted value exceeds the physical range, DNN outputs NaN, etc.), the bypass is automatically switched off. final =φ PI The report alerted that SST would continue to operate in the traditional mode.

[0110] III. Operating Condition Testing and Effect Verification

[0111] In a laboratory testing platform, an extreme load step condition with a rise time of 3 microseconds and an amplitude of 80% of the rated power was simulated using an electronic load to conduct comparative tests.

[0112] When only traditional PI control is enabled, the SST output voltage drops to 772 volts, the voltage deviation reaches -3.5%, and the time for the voltage to recover to steady state is 260 microseconds, which cannot meet the stringent AIDC requirements.

[0113] After enabling the complete AI control scheme in this embodiment, the lowest output voltage of SST is 786 volts, the voltage deviation is controlled within -1.75%, and the voltage recovery time is only 22 microseconds.

[0114] The overall dynamic response time of this system is less than 30 microseconds, and the transient deviation of the output voltage is less than ±1.5%, which is better than the design target. It also meets the industry power supply requirements of AIDC output voltage deviation < ±2% and recovery time of 50 to 100 microseconds. For a 5000 kW ISOP high-power power supply cabinet composed of 30 modules, the global load prediction results can be shared to all DAB modules. Each module only needs to make minor parameter adjustments through its local inverse model, effectively reducing inter-module coupling interference and improving the overall cabinet's voltage and current sharing performance.

[0115] IV. Solution Expansion and Replaceable Technologies

[0116] Load prediction model replacement: In addition to TCN and LSTM, time series models such as Gated Recurrent Unit (GRU) and Transformer can be selected. Lightweight GRU is suitable for low computing power hardware, while Transformer is suitable for complex long-series load scenarios.

[0117] Power inverse model replacement: DNN can be replaced with machine learning models such as Gaussian process and extreme gradient boosting (XGBoost) to adapt to different hardware computing power and cost requirements;

[0118] Reinforcement learning alternatives: The DRL algorithm can be replaced by the fuzzy self-tuning PI algorithm, which has simpler logic and is suitable for low-end embedded control hardware;

[0119] Main topology replacement: DAB isolation topology can be replaced with LLC resonant isolation topology or matrix-type cross-cross transformation topology. Only the input parameters and output control quantities of the inverse model need to be adjusted to reuse the entire control logic.

[0120] Model optimization method: All AI models can be trained offline using SST open-loop data and simulation data, and continuously fine-tuned online using actual equipment operation data, taking into account changes in operating conditions such as adapter component aging and temperature drift.

[0121] V. Overall Technical Advantages of the Solution

[0122] This embodiment relies on a combination of load forecasting, data-driven inverse model feedforward, DRL adaptive tuning, and traditional PI fallback. First, it achieves early response to load step changes by relying on microsecond-level load prediction, completely solving the problems of bandwidth limitation and adjustment lag of PI controllers. Second, it automatically compensates for nonlinear factors such as DAB topology dead zone, magnetic saturation, and parasitic parameters through data-driven models, eliminating the need for manual tuning of PI parameters on a per-unit basis and significantly reducing the workload of on-site debugging. Third, the AI ​​module supports hot bypass degradation, and combined with mature PI loops, it ensures high system availability. Fourth, it makes full use of AIDC scheduling and early warning signals to extend the warning time. The multi-module ISOP architecture has strong scalability and is fully adaptable to the high-power, high-dynamic DC power supply scenarios of AI data centers.

[0123] In the claims, description, and accompanying drawings of this invention, the term "plural" refers to two or more. Unless otherwise explicitly defined, the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and simplifying the descriptive process, and are not intended to indicate or imply that the device or element referred to must have the described specific orientation, or be constructed and operated in a specific orientation. Therefore, these descriptions should not be construed as limiting the invention. The terms "connected," "installed," "fixed," etc., should be interpreted broadly. For example, "connected" can be a fixed connection between multiple objects, a detachable connection between multiple objects, or an integral connection; it can be a direct connection between multiple objects or an indirect connection between multiple objects through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in this invention can be understood based on the specific circumstances described above.

[0124] In the claims, description, and accompanying drawings of this invention, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In the claims, description, and accompanying drawings of this invention, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for microsecond-level response of a solid-state transformer based on AI, characterized in that, The control method includes: The current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side are obtained. The current operating status data includes at least one of the following: input side voltage or current, intermediate DC bus voltage, primary or secondary current of high-frequency isolation stage, and output side voltage. The current electrical signal data is input into a first data-driven model for load forecasting to obtain a load demand forecast sequence within a preset time range in the future. The load demand prediction sequence and the current operating status data are input into a second data-driven model used to characterize the inverse characteristics of power transmission of solid-state transformers, and the expected feedforward control parameters for adjusting the power transmission of solid-state transformers in advance are obtained. The target control signal is determined by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters. The solid-state transformer is driven according to the target control signal.

2. The AI-based control method for microsecond-level response of solid-state transformers according to claim 1, characterized in that, The first data-driven model includes a temporal neural network; The step of inputting the current electrical signal data into a first data-driven model for load forecasting to obtain a load demand forecast sequence within a preset time range includes: The historical current data, historical voltage data, and current change rate data within the sliding time window are acquired as the current electrical signal data. The current electrical signal data is input into the time-series neural network, which outputs a load demand prediction sequence within a preset time range in the future. The load demand prediction sequence includes at least one of the following: predicted load current, predicted load power, or load power change at multiple future prediction times.

3. The AI-based control method for microsecond-level response of solid-state transformers according to claim 1, characterized in that, Simultaneously or before acquiring the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data on the load side, the control method further includes: Receive task scheduling notification signals from external management devices for the data center; The step of inputting the current electrical signal data into a first data-driven model for load forecasting to obtain a load demand forecast sequence within a preset time range includes: The task scheduling forecast signal is used as an auxiliary input, fused with the current electrical signal data, and then input into the first data-driven model to obtain a load demand prediction sequence within a preset time range in the future.

4. The AI-based control method for microsecond-level response of solid-state transformers according to claim 1, characterized in that, The solid-state transformer includes an isolated dual active full-bridge topology, and the desired feedforward control parameter is the feedforward phase shift angle compensation amount of the isolated dual active full-bridge topology. The step of inputting the load demand prediction sequence and the current operating status data into a second data-driven model to characterize the inverse characteristics of solid-state transformer power transmission, and obtaining the desired feedforward control parameters for adjusting the power transmission amount of the solid-state transformer in advance, includes: The load forecast current in the load demand forecast sequence, and at least one of the following in the current operating status data: output measured voltage, input DC bus voltage, primary side measured current, core temperature, power switching device temperature, and dead zone compensation parameters, are input into the second data driving model to output the feedforward phase shift compensation amount. The second data-driven model includes at least one of deep feedforward neural networks, lookup table models, and regression models.

5. The AI-based control method for microsecond-level response of solid-state transformers according to claim 1, characterized in that, The determination of the target control signal by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters includes: The feedback control parameters are superimposed with the desired feedforward control parameters, and the superimposed parameters are subjected to amplitude limiting to generate the target control signal.

6. The AI-based control method for microsecond-level response of solid-state transformers according to claim 5, characterized in that, The control method further includes: The current operating state data is input into a pre-trained reinforcement learning agent model to obtain control correction parameters; The proportional-integral gain parameter of the feedback control loop is updated using the control correction parameter, or the control correction parameter is superimposed with the feedback control parameter and the desired feedforward control parameter to generate the target control signal.

7. The AI-based control method for microsecond-level response of solid-state transformers according to any one of claims 1 to 6, characterized in that, The step of determining the target control signal by combining the feedback control parameters generated by the feedback control loop based on the current operating state data and the desired feedforward control parameters further includes: The output states of the first data-driven model and the second data-driven model are detected in real time. When the output of the first data-driven model or the second data-driven model is detected to exceed the preset physical range, an invalid value appears, inference times out, or model self-test fails, the expected feedforward control parameter is hot-bypassed so that the system automatically reverts to a pure feedback control mode in which the target control signal is determined solely based on the feedback control parameter generated by the feedback control loop.

8. A control device for a solid-state transformer with microsecond-level response based on AI, characterized in that, The control device includes: The data acquisition module is used to acquire the current operating status data of the main power circuit of the solid-state transformer and the current electrical signal data of the load side. The current operating status data includes at least one of the following: input side voltage or current, intermediate DC bus voltage, primary or secondary current of the high-frequency isolation stage, and output side voltage. The prediction module is used to input the current electrical signal data into a first data-driven model for load prediction to obtain a load demand prediction sequence within a preset time range in the future. The feedforward calculation module is used to input the load demand prediction sequence and the current operating status data into a second data-driven model that characterizes the inverse characteristics of power transmission of solid-state transformers, and obtain the expected feedforward control parameters for adjusting the power transmission of solid-state transformers in advance. The control synthesis module is used to combine the feedback control parameters generated by the feedback control loop based on the current operating state data with the desired feedforward control parameters to determine the target control signal; A drive module is used to drive the solid-state transformer according to the target control signal.

9. A control device for a solid-state transformer with microsecond-level response based on AI, characterized in that, AI-based control devices for microsecond-level response of solid-state transformers include solid-state transformers, which include: The main power circuit includes an input rectifier stage, a high-frequency isolation stage, and an output filter stage, wherein the high-frequency isolation stage includes a dual active full-bridge converter circuit; The sensing unit is used to collect the status signals of the main power circuit and the load side in real time; A control processing unit, based on the state signal collected by the sensing unit, executes the AI-based solid-state transformer microsecond-level response control method as described in any one of claims 1 to 7, so as to output a microsecond-level phase shift angle control signal to the dual active full-bridge converter circuit.

10. An electronic device, characterized in that, include: processor; A memory storing instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the AI-based control method for microsecond-level response of a solid-state transformer as described in any one of claims 1 to 7.

Images