Pfc-llc global collaborative optimization control method, system and device based on symbiotic neural network

Abstract

The present application belongs to the technical field of electronic power control, and particularly relates to a PFC-LLC global collaborative optimization control method, system and device based on a symbiotic neural network, which comprises the following steps: synchronously collecting electrical state data of a PFC stage and an LLC stage, and inputting the data to a pre-trained symbiotic neural network model together with an external target instruction; the model adopts a CNN-LSTM hybrid network structure and integrates an attention mechanism, is used for extracting and fusing the time-space features of the input data, and directly outputs a collaborative control instruction for synchronously controlling the switch duty cycle of the PFC module and the switch frequency or phase shift angle of the LLC module. The present application realizes global collaborative decision of two-stage control through a neural network model, effectively solves the inherent efficiency loss, dynamic response lag and parameter coupling problems of traditional independent control in stages, and significantly improves the overall efficiency, dynamic performance and robustness of the power supply system.

Description

Technical Field

[0001] This invention belongs to the field of electronic power control technology, specifically relating to a PFC-LLC global collaborative optimization control method, system, and device based on symbiotic neural networks. Background Technology

[0002] Traditional switching power supplies, especially in medium-to-high power applications, commonly employ a two-stage architecture combining a PFC (Power Factor Correction) stage and an LLC resonant converter stage. In this architecture, the PFC stage typically uses average current-mode control based on an analog or digital PI controller to achieve a high power factor and stable DC bus voltage; the LLC stage uses frequency control or phase-shift control to achieve electrical isolation, efficient voltage reduction, and soft-switching operation. Existing solutions usually treat the PFC and LLC as two independent control objects, designing separate controllers for each and combining them through cascading. This discrete control strategy is clear and simple in design and has become the mainstream design solution in the industry.

[0003] However, this traditional discrete control method has inherent limitations. Due to the strong electrical coupling between the PFC and LLC stages in actual operation—for example, the bus voltage ripple of the PFC stage significantly affects the gain characteristics of the LLC stage, while drastic load changes in the LLC stage cause bus voltage fluctuations, forcing the PFC controller to adjust—the two controllers operate independently, pursuing only local optima (e.g., PFC pursues low total harmonic distortion (THD), LLC pursues zero-voltage switching (ZVS), failing to achieve global coordination and optimization at the system level. This leads to overshoot and oscillations in the dynamic response process, increased switching losses under light load conditions resulting in decreased overall efficiency, and poor system parameter robustness, requiring cumbersome manual tuning. Summary of the Invention

[0004] The purpose of this invention is to provide a global collaborative optimization control method and system for PFC-LLC based on symbiotic neural networks, so as to solve the problems of efficiency loss, dynamic response hysteresis and parameter coupling caused by independent control in the traditional two-stage power supply architecture of PFC-LLC.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] Firstly, this invention proposes a global collaborative optimization control method for PFC-LLC based on a symbiotic neural network, applied to a power control system with both a PFC module and an LLC module. The method includes:

[0007] The electrical status data of the PFC module and LLC module are collected synchronously. The electrical status data of the PFC module includes input voltage, input current and bus voltage, and the electrical status data of the LLC module includes resonant current, resonant capacitor voltage, load current, switch status and temperature.

[0008] The electrical state data of the PFC module and the LLC module, as well as the external target command, are input into the pre-trained symbiotic neural network model, and the cooperative control command is output.

[0009] The external target commands include output power setpoint, efficiency weighting coefficient, and total harmonic distortion (THD) constraint threshold; the cooperative control commands include the switching duty cycle control signal of the PFC module and the switching frequency or phase shift angle control signal of the LLC module.

[0010] In response to the coordinated control command, the PFC module and LLC module are synchronously controlled;

[0011] The symbiotic neural network model adopts a CNN-LSTM-attention hybrid architecture, is trained by reinforcement learning and integrates output physical constraints, and is guided by a reward function to learn and output the cooperative control command, so as to keep the bus voltage within a preset range, reduce the switching frequency and trigger LLC burst mode.

[0012] Furthermore, the symbiotic neural network model employs a CNN-LSTM-attention hybrid architecture to process the input data, including the following steps:

[0013] A convolutional neural network is used to extract the instantaneous spatial features of the input data to obtain a spatial feature vector;

[0014] Long Short-Term Memory (LSTM) networks are used to extract temporal-series dependency features from spatial feature vectors, and spatiotemporal fusion features are output.

[0015] An attention mechanism is used to dynamically weight and focus on key state variables in the spatiotemporal fusion feature, enhancing the response weight of one or more of the following: bus voltage fluctuation amplitude, load current change rate, resonant current phase, and input voltage transient value; finally, after processing by the fully connected layer, a cooperative control command is output.

[0016] Furthermore, the step of using an attention mechanism to dynamically weight and focus on key state variables in the spatiotemporal fusion feature includes the following steps:

[0017] When the bus voltage fluctuation amplitude is detected to exceed the preset threshold, the weight of the resonant current phase and LLC switching frequency in the decision-making is increased.

[0018] When the load current change rate is detected to exceed a preset threshold, the weight of bus voltage and PFC switch duty cycle in the decision-making is increased.

[0019] When a drop or surge in the input voltage transient value is detected, the weight of the input current and the PFC switch duty cycle in the decision is increased.

[0020] Furthermore, the method also includes the following steps:

[0021] Physical boundary constraints on the switching frequency of the LLC module are added to the output layer of the symbiotic neural network model. The switching frequency fs control signal of the LLC module is forcibly constrained to a value range greater than the inherent resonant frequency fr of the LLC module resonant cavity, i.e., fs > fr, so as to realize zero-voltage switching of the LLC module switching tube.

[0022] Furthermore, the symbiotic neural network model is obtained through the following steps:

[0023] In the simulation environment, obtain the training dataset generated by running the joint simulation model constructed by the PFC module and LLC module under different working conditions;

[0024] Using electrical state data and external target commands from the training dataset as inputs and corresponding cooperative control commands as outputs, supervised training is performed on a pre-constructed initial neural network model containing convolutional layers, long short-term memory layers, and attention layers to obtain a pre-trained model.

[0025] The pre-trained model is connected to the control loop of the co-simulation model;

[0026] Define a reward value calculation rule, the input of which is the measured values ​​of efficiency, total harmonic distortion and thermal stress of the co-simulation model;

[0027] During the simulation environment, the current electrical state data and external target commands are input into the pre-trained model, the output cooperative control commands are applied to the co-simulation model, and the parameters of the pre-trained model are adjusted according to the reward value output by the reward value calculation rule. After iteration, the symbiotic neural network model is obtained.

[0028] Furthermore, the reward value calculation rule is configured as follows:

[0029] When the combination of control commands output by the pre-trained model is such that the bus voltage value is within a preset voltage range and the LLC module switching frequency command is changed, a positive reward is given.

[0030] When the combination of control instructions output by the pre-trained model is a combination of an instruction to simultaneously reduce the switching frequency of the PFC module and an instruction to put the LLC module into burst mode, a positive reward is given.

[0031] Secondly, this invention proposes a PFC-LLC global collaborative optimization control system based on a symbiotic neural network, which is communicatively connected to a power control system having a PFC module and an LLC module, for implementing the steps of the control method described above. The system includes:

[0032] The data acquisition module is used to synchronously acquire electrical status data of the PFC module and LLC module. The electrical status data of the PFC module includes input voltage, input current and bus voltage, and the electrical status data of the LLC module includes resonant current, resonant capacitor voltage, load current, switch status and temperature.

[0033] The instruction input interface is used to receive external target instructions;

[0034] The controller integrates the symbiotic neural network model; the controller is configured to: receive electrical status data from the PFC module, electrical status data from the LLC module, and external target commands, process them through the symbiotic neural network model, and generate cooperative control commands;

[0035] The external target commands include output power setpoint, efficiency weighting coefficient, and total harmonic distortion (THD) constraint threshold; the cooperative control commands include the switching duty cycle control signal of the PFC module and the switching frequency or phase shift angle control signal of the LLC module.

[0036] A drive circuit is used to respond to the coordinated control command and perform synchronous control on the PFC module and the LLC module.

[0037] Furthermore, the data acquisition module includes a voltage sampling circuit, a current sampling circuit, and a temperature sampling circuit.

[0038] Furthermore, the PFC module is a boost-type power factor correction circuit; the LLC module is an LLC resonant converter circuit.

[0039] Thirdly, the present invention proposes a power supply device, including a power control system having a PFC module and an LLC module, and a PFC-LLC global collaborative optimization control system based on a symbiotic neural network as described above, which is communicatively connected to the power control system.

[0040] The beneficial effects of this invention are as follows:

[0041] This invention integrates state information from both the PFC and LLC levels and external commands using a symbiotic neural network model. Utilizing its internal convolutional layers, long short-term memory layers, and attention mechanisms, it achieves spatiotemporal feature extraction and dynamic focusing of key state variables from multi-source heterogeneous data, accurately capturing the coupling relationship and dynamic processes between the two levels. The model directly outputs cooperative control commands, replacing the traditional independent controller architecture, and realizing synchronized optimization decisions for the PFC duty cycle and LLC frequency / phase at the system level.

[0042] The method proposed in this invention allows the bus voltage to fluctuate flexibly within a safe range to extend the LLC soft-switching range, and synergistically reduces the PFC switching frequency and triggers LLC burst mode under light load conditions, thereby minimizing global losses. Ultimately, this effectively improves the system's response speed and stability under dynamic operating conditions, achieves higher overall energy efficiency across the entire load range, and enhances the system's adaptive robustness to parameter changes and device aging. Attached Figure Description

[0043] Figure 1 A flowchart illustrating a global collaborative optimization control method for PFC-LLC based on symbiotic neural networks provided in this application embodiment;

[0044] Figure 2 This is another flowchart illustrating the PFC-LLC global collaborative optimization control method based on symbiotic neural networks provided in this application embodiment;

[0045] Figure 3 A schematic diagram of the model training process in the PFC-LLC global collaborative optimization control method based on symbiotic neural networks provided in the embodiments of this application;

[0046] Figure 4 This is a system block diagram of a PFC-LLC global collaborative optimization control system based on a symbiotic neural network, provided in an embodiment of this application. Detailed Implementation

[0047] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0048] It is worth noting that in traditional switching power supply systems, especially in medium-to-high power applications, a two-stage architecture combining a PFC (Power Factor Correction) stage and an LLC resonant converter stage is commonly used. The PFC stage typically employs average current-mode control based on analog or digital PI controllers to achieve high power factor and stable DC bus voltage; the LLC stage uses frequency control or phase-shift control to achieve electrical isolation, efficient voltage reduction, and soft-switching operation. Existing solutions usually treat the PFC and LLC as two independent control objects, designing separate controllers for each and combining them through cascading. While this discrete control strategy is clear in design and easy to implement, the strong electrical coupling between the PFC and LLC stages in actual operation (e.g., bus voltage ripple affects LLC gain characteristics, and load changes cause bus voltage fluctuations) leads to overshoot and oscillation during dynamic response, decreased efficiency under light load conditions, poor parameter robustness, reliance on cumbersome manual tuning, and inability to achieve globally optimal control.

[0049] To overcome the aforementioned limitations, this invention introduces a "symbiotic neural network" control architecture. Its core lies in constructing a hybrid model that integrates a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism. This model enables the extraction and dynamic focusing of spatiotemporal features from multi-source heterogeneous data in the PFC and LLC two-stage systems. Specifically, the CNN extracts the instantaneous spatial features of the input electrical state (such as current ripple and voltage phase), the LSTM captures time-series dependencies (such as load step history), and the attention mechanism dynamically weights key state variables (such as resonant current phase and transient input voltage values) based on the system's dynamic operating conditions (such as bus voltage fluctuations and load abrupt changes). This achieves coordinated optimization decisions regarding the PFC duty cycle and the LLC switching frequency / phase shift angle, significantly improving the overall system efficiency, dynamic response speed, and robustness.

[0050] Example 1

[0051] Please combine Figures 1-2 In one specific implementation of this application, a global collaborative optimization control method based on symbiotic neural networks for PFC-LLC is proposed and applied to a two-stage switching power supply control system that includes a Boost power factor correction (PFC) module and an LLC resonant converter module, aiming to achieve global collaborative control and dynamic optimization between the two stages.

[0052] The method includes: First, synchronously collecting electrical status data of the PFC module and LLC module. The electrical status data of the PFC module includes input voltage, input current, and bus voltage, while the electrical status data of the LLC module includes resonant current, resonant capacitor voltage, load current, switch status, and temperature. Then, the electrical status data of the PFC module and LLC module, as well as external target commands, are input into a pre-trained symbiotic neural network model to output cooperative control commands.

[0053] The external target commands include the output power setpoint, efficiency weighting coefficient, and total harmonic distortion (THD) constraint threshold; the cooperative control commands include the switching duty cycle control signal of the PFC module and the switching frequency or phase shift angle control signal of the LLC module; the symbiotic neural network model adopts a CNN-LSTM-attention hybrid architecture, is trained by reinforcement learning and integrates output physical constraints, and guides the model to learn and output cooperative control commands through a reward function, so as to keep the bus voltage within the preset range, reduce the switching frequency, and trigger the LLC burst mode.

[0054] The symbiotic neural network model employs a CNN-LSTM hybrid architecture with an integrated attention mechanism, enabling spatiotemporal feature extraction and fusion from multi-source heterogeneous input data. Specifically, the convolutional neural network (CNN) extracts the spatial local features of the input data, the long short-term memory network (LSTM) captures temporal dependencies, and the attention mechanism adaptively weights key state variables based on the system's dynamic operating conditions, thereby enhancing the model's ability to represent coupling relationships and transient processes. Ultimately, the model directly outputs cooperative control commands, including the duty cycle control signal for the PFC module and the switching frequency or phase shift angle control signal for the LLC module.

[0055] Finally, in response to the coordinated control command, the PFC module and LLC module are synchronously controlled to achieve global coordinated operation of the two-stage power converter.

[0056] In one implementation, the symbiotic neural network model employs a CNN-LSTM-attention hybrid architecture to process the input data, including the following steps: using a convolutional neural network to extract the instantaneous spatial features of the input data to obtain a spatial feature vector; using a long short-term memory network to extract the time-series dependency features of the spatial feature vector and output spatiotemporal fusion features; using an attention mechanism to dynamically weight and focus on key state variables in the spatiotemporal fusion features, enhancing the response weights to one or more of the following: bus voltage fluctuation amplitude, load current change rate, resonant current phase, and transient input voltage value; and finally, after processing by a fully connected layer, outputting cooperative control commands.

[0057] Key state variables include one or more of the following: bus voltage fluctuation amplitude, load current change rate, resonant current phase, and input voltage transient value; input data includes electrical state data of the PFC module, electrical state data of the LLC module, and external target commands.

[0058] In one implementation, an attention mechanism is used to dynamically weight and focus on key state variables in the spatiotemporal fusion feature, including the following steps: when the bus voltage fluctuation amplitude is detected to exceed a preset threshold, the weights of the resonant current phase and LLC switching frequency in the decision are increased; when the load current change rate is detected to exceed a preset threshold, the weights of the bus voltage and PFC switch duty cycle in the decision are increased; when the transient value of the input voltage is detected to drop or rise sharply, the weights of the input current and PFC switch duty cycle in the decision are increased.

[0059] Understandably, in the specific implementation of the aforementioned dynamic weighted focusing method, the attention mechanism calculates the attention score of each state variable through a trainable fully connected layer, and dynamically reweights the feature vector accordingly. For example, when the bus voltage fluctuation amplitude exceeds a preset threshold (e.g., ±5% of the rated value), the attention module automatically increases the weight coefficients of the resonant current phase and LLC switching frequency characteristics in the hidden layer output, making the control decision more inclined to stabilize the resonant process; when the load current change rate exceeds a set threshold (e.g., 10A / ms), the weights of the bus voltage and PFC duty cycle characteristics are enhanced to prioritize power balance; when the input voltage experiences a transient drop or surge (e.g., exceeding ±15% of the rated value), the feature weights of the input current and PFC duty cycle are significantly increased to quickly suppress input-side disturbances.

[0060] In one implementation, the method further includes: adding physical boundary constraints on the switching frequency of the LLC module to the output layer of the symbiotic neural network model, and forcibly constraining the switching frequency fs control signal of the LLC module within a numerical range greater than the inherent resonant frequency fr of the LLC module resonant cavity, i.e., satisfying fs > fr, so as to achieve zero-voltage switching of the LLC module switching transistor.

[0061] Understandably, the aforementioned physical boundary constraints are achieved by introducing a differentiable projection layer after the neural network output layer. This layer receives the LLC switching frequency command fs from the network's original output and, based on the resonant cavity's inherent resonant frequency fr, ensures that the output frequency always satisfies the hard constraint that fs > fr by setting a constraint function. This constraint mechanism prevents the LLC from operating in the capacitive region, ensuring that the switching transistor achieves zero-voltage turn-on (ZVS) and significantly reducing switching losses. Simultaneously, this design does not affect the main structure of the neural network and requires no additional protection circuitry.

[0062] Please combine Figure 3In one implementation, the symbiotic neural network model is trained through the following steps: In a simulation environment (such as a simulation platform like PLECS / PSIM), various typical operating conditions such as sudden input voltage changes, load step changes, and temperature variations are covered to generate a sufficiently diverse training dataset. Specifically, the training dataset generated by running the co-simulation model constructed by the PFC and LLC modules under different operating conditions is obtained; using the electrical state data and external target commands from the training dataset as input, and the corresponding cooperative control commands as output, a pre-constructed initial neural network model containing convolutional layers, long short-term memory layers, and attention layers is subjected to supervised training to obtain a pre-trained model; the pre-trained model is connected to the control loop of the co-simulation model; a reward value calculation rule is defined, the input of which is the measured values ​​of the efficiency, total harmonic distortion, and thermal stress of the co-simulation model; during simulation operation, the current electrical state data and external target commands are input to the pre-trained model, the output cooperative control commands are applied to the co-simulation model, and the parameters of the pre-trained model are adjusted according to the reward value output by the reward value calculation rule. After iteration, the symbiotic neural network model is obtained.

[0063] In one implementation, the reward calculation rule is configured as follows: a positive reward is given when the combination of control commands output by the pre-trained model keeps the bus voltage within a preset voltage range and simultaneously changes the LLC module switching frequency command; a positive reward is also given when the combination of control commands output by the pre-trained model simultaneously decreases the PFC module switching frequency command and causes the LLC module to enter burst mode. The function value of the positive reward is a relatively high positive value, the purpose of which is to explicitly guide the model to learn in a direction that aligns with the system optimization objective through numerical feedback.

[0064] In this embodiment, the reward value is set so that a positive reward is given when the model's decision-making behavior conforms to the global collaborative optimization objective, guiding the policy network to master the bus voltage flexible control strategy and the global loss collaborative optimization strategy during training, so as to obtain a symbiotic neural network model.

[0065] Specifically, the bus voltage flexible control strategy includes, by balancing the system state, outputting a PFC control command that allows the bus voltage to fluctuate within a preset safety range, and simultaneously outputting a command to adjust the LLC switching frequency to compensate for gain changes, thereby expanding the soft switching range of the LLC to achieve efficiency optimization while maintaining stable output voltage; the global loss collaborative optimization strategy includes, when the system is judged to be under light load or no-load conditions based on the input state, outputting a command to reduce the switching frequency of the PFC stage, and simultaneously outputting a command to control the LLC stage to enter the intermittent burst working mode, so that the switching actions of the PFC stage and the LLC stage are reduced synchronously, thereby minimizing the total switching loss of the system.

[0066] In practice, a pre-constructed initial neural network model, including convolutional layers, long short-term memory layers, and attention layers, undergoes supervised training with mean squared error as the loss function. The Adam optimizer is used to train the initial neural network, enabling it to initially grasp the mapping relationship between input states and control commands. After the pre-trained model is connected to the co-simulation control loop, reinforcement learning fine-tuning is performed using the Proximal Policy Optimization (PPO) algorithm. The reward calculation rule comprehensively considers multiple objective indicators such as system efficiency, THD, and thermal stress. Its specific design is as follows: when the model's decision maintains the bus voltage within a preset safe range (e.g., 380-420V) and simultaneously adjusts the LLC switching frequency, it indicates that it possesses cooperative response capability and is given a positive reward; when the model simultaneously reduces the PFC switching frequency and triggers the LLC burst mode under light load conditions, it indicates that it has achieved global loss optimization, and is also given a positive reward. This reward mechanism guides the model to learn the optimal control strategy while satisfying electrical constraints.

[0067] In one specific embodiment of this application, taking a jump in load from 20% to 100% (output current abruptly changing from 4A to 22A) as an example, the actual implementation effect of the PFC-LLC global collaborative optimization control method based on symbiotic neural networks is illustrated. The implementation process fully embodies the core features of the method in this application, including key technical features such as multi-source state data fusion, spatiotemporal feature extraction, dynamic weighting with attention mechanism, and collaborative control command output.

[0068] Under traditional control schemes, this sudden load change condition will cause a slow response in the PFC stage, leading to a drop in bus voltage, which in turn causes insufficient gain in the LLC stage, ultimately resulting in a significant collapse in output voltage (fluctuations typically exceeding ±8%). However, the method in this application achieves significantly superior dynamic control performance through the collaborative decision-making and pre-compensation mechanism of a symbiotic neural network model. Specifically, the symbiotic neural network extracts the spatial features of the input voltage and current waveforms through its built-in convolutional neural network (CNN), while simultaneously using a long short-term memory (LSTM) network unit to capture the changing gradient of the load current and its temporal dependence in real time. Furthermore, it combines an attention mechanism to dynamically weight and focus on key state variables such as the rate of change of load current and bus voltage fluctuations, predicting the system's dynamic demands 5μs in advance.

[0069] Based on the above multi-feature fusion and analysis, the neural network outputs cooperative control commands: First, it increases the switching duty cycle of the PFC module (D increases from 0.4 to 0.65) to rapidly increase the bus voltage level; simultaneously, it reduces the switching frequency of the LLC module (fs is adjusted from 105kHz to 92kHz) to increase its voltage gain and compensate for the impact of bus voltage fluctuations. This reflects the core feature of this application: the neural network directly outputs two-stage cooperative control commands. At the same time, this invention allows the bus voltage to fluctuate flexibly within a safe range (Vbus temporarily drops from 400V to 380V, with the drop controlled within 5%), avoiding voltage overshoot due to over-response in the PFC stage and providing a wider soft-switching range for the LLC stage. This strategy originates from the global optimization strategy learned during the neural network training process. After the system enters a heavy-load steady state, the network further optimizes the loss distribution: the PFC switching frequency is appropriately increased to 110kHz to reduce conduction losses, while the LLC switching frequency is reduced back to 98kHz to optimize the zero-voltage switching (ZVS) condition and minimize the overall system loss. This demonstrates the continuous optimization capability of the neural network across the entire operating range.

[0070] Experimental verification has shown that the global collaborative optimization control method based on symbiotic neural networks proposed in this application can maintain excellent voltage regulation accuracy, dynamic response speed and global efficiency optimization capability under severe dynamic conditions, effectively solving the inherent response hysteresis and performance loss problems of traditional hierarchical control.

[0071] Example 2

[0072] Please combine Figure 4 In yet another specific embodiment, this application proposes a PFC-LLC global collaborative optimization control system based on a symbiotic neural network, which is communicatively connected to a power control system having a PFC module and an LLC module, for implementing the steps of the control method as described in Embodiment 1. The system includes:

[0073] The data acquisition module is used to synchronously acquire electrical status data of the PFC module and LLC module. The electrical status data of the PFC module includes input voltage, input current and bus voltage, and the electrical status data of the LLC module includes resonant current, resonant capacitor voltage, load current, switch status and temperature.

[0074] The instruction input interface is used to receive external target instructions;

[0075] The controller integrates a symbiotic neural network model. The controller is configured to receive electrical status data from the PFC module, electrical status data from the LLC module, and external target commands, process them through the symbiotic neural network model, and generate cooperative control commands.

[0076] The external target commands include the output power setpoint, efficiency weighting coefficient, and total harmonic distortion (THD) constraint threshold; the cooperative control commands include the switching duty cycle control signal of the PFC module and the switching frequency or phase shift angle control signal of the LLC module.

[0077] The drive circuit is used to respond to coordinated control commands and perform synchronous control on the PFC module and LLC module.

[0078] Preferably, the data acquisition module includes a voltage sampling circuit, a current sampling circuit, and a temperature sampling circuit.

[0079] Preferably, the PFC module is a boost-type power factor correction circuit; the LLC module is an LLC resonant converter circuit.

[0080] The symbiotic neural network model in this embodiment is an intelligent control core with global optimization capabilities, obtained through a specific training process. The training process follows the method described in Embodiment 1: First, a joint simulation model of PFC and LLC is constructed in a simulation environment, generating a large-scale training dataset by simulating various operating conditions (including input voltage disturbances, load mutations, and temperature changes). Then, using electrical state data and external target commands as inputs and ideal cooperative control commands as outputs, supervised pre-training is performed on the initial network, which includes convolutional layers, long short-term memory layers, and attention layers. Finally, the pre-trained model is integrated into the joint simulation closed loop, and online fine-tuning and optimization are performed using a reward function based on system efficiency, THD, and thermal stress, through reinforcement learning algorithms. This allows the model to gradually master two-level cooperative strategies (bus voltage flexible control strategy and global loss cooperative optimization strategy) and boundary constraints. The neural network model obtained through this training possesses the ability to process multi-source heterogeneous data, dynamically focus on key variables, and output physically feasible commands, thereby achieving global cooperative optimization control of PFC and LLC in practical applications.

[0081] According to the above embodiments, the global collaborative optimization control system provided in this embodiment is communicatively connected to a power control system including a Boost-type PFC module and an LLC resonant converter module. This system synchronously acquires multi-dimensional electrical state data of the two-stage circuits through a data acquisition module, and inputs this data, combined with external target commands, into a controller integrating a symbiotic neural network model. This neural network integrates CNN, LSTM, and attention mechanisms to extract the spatiotemporal features of the input data in real time and dynamically focus on key state variables. Ultimately, it directly outputs collaborative control commands for the PFC duty cycle and LLC frequency / phase shift angle, achieving synchronous regulation of the two-stage power converters via a drive circuit.

[0082] In practical applications, this system is suitable for medium and high power switching power supplies, server power supplies, industrial power supplies and other application scenarios, and can effectively solve the problems of dynamic response lag and insufficient global optimization under the traditional independent control architecture.

[0083] Example 3

[0084] In yet another specific embodiment, this application proposes a power supply device including a power control system having a PFC module and an LLC module, and a PFC-LLC global collaborative optimization control system based on a symbiotic neural network, as shown in Embodiment 2, which is communicatively connected to the power control system.

[0085] In this embodiment, the power supply device can specifically be a high-efficiency medium-to-high power AC-DC converter, whose typical applications include, but are not limited to: server power modules, industrial inverter front-end power supply units, electric vehicle charging pile power modules, or communication base station power systems. The core feature of this device is that it integrates a power stage circuit composed of a PFC module and an LLC resonant converter module based on Boost topology, and achieves global collaborative optimization of the two-stage circuit through a symbiotic neural network control system.

[0086] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0087] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0088] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0089] In addition, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0090] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A global collaborative optimization control method for PFC-LLC based on symbiotic neural networks, applied to a power control system with PFC and LLC modules, characterized in that, The method includes: The electrical status data of the PFC module and LLC module are collected synchronously. The electrical status data of the PFC module includes input voltage, input current and bus voltage, and the electrical status data of the LLC module includes resonant current, resonant capacitor voltage, load current, switch status and temperature. The electrical state data of the PFC module and the LLC module, as well as the external target command, are input into the pre-trained symbiotic neural network model, and the cooperative control command is output. The external target commands include output power setpoint, efficiency weighting coefficient, and total harmonic distortion (THD) constraint threshold; the cooperative control commands include the switching duty cycle control signal of the PFC module and the switching frequency or phase shift angle control signal of the LLC module. In response to the coordinated control command, the PFC module and LLC module are synchronously controlled; The symbiotic neural network model adopts a CNN-LSTM-attention hybrid architecture, is trained by reinforcement learning and integrates output physical constraints, and is guided by a reward function to learn and output the cooperative control command, so as to keep the bus voltage within a preset range, reduce the switching frequency and trigger LLC burst mode. The symbiotic neural network model employs a CNN-LSTM-attention hybrid architecture to process the input data, including the following steps: A convolutional neural network is used to extract the instantaneous spatial features of the input data to obtain a spatial feature vector; Long Short-Term Memory (LSTM) networks are used to extract temporal-series dependency features from spatial feature vectors, and spatiotemporal fusion features are output. An attention mechanism is used to dynamically weight and focus on key state variables in the spatiotemporal fusion feature, enhancing the response weight of one or more of the following: bus voltage fluctuation amplitude, load current change rate, resonant current phase, and input voltage transient value; finally, after processing by a fully connected layer, a cooperative control command is output. The process of dynamically weighting and focusing on key state variables in the spatiotemporal fusion feature using an attention mechanism includes the following steps: When the bus voltage fluctuation amplitude is detected to exceed the preset threshold, the weight of the resonant current phase and LLC switching frequency in the decision-making is increased. When the load current change rate is detected to exceed a preset threshold, the weight of bus voltage and PFC switch duty cycle in the decision-making is increased. When a drop or surge in the input voltage transient value is detected, the weight of the input current and the PFC switch duty cycle in the decision is increased.

2. The global cooperative optimization control method for PFC-LLC based on symbiotic neural networks according to claim 1, characterized in that, The method further includes the following steps: Physical boundary constraints on the switching frequency of the LLC module are added to the output layer of the symbiotic neural network model. The switching frequency fs control signal of the LLC module is forcibly constrained to a value range greater than the inherent resonant frequency fr of the LLC module resonant cavity, i.e., fs > fr, so as to realize zero-voltage switching of the LLC module switching tube.

3. The PFC-LLC global cooperative optimization control method based on symbiotic neural networks according to claim 1, characterized in that, The symbiotic neural network model is obtained through the following steps: In the simulation environment, obtain the training dataset generated by running the joint simulation model constructed by the PFC module and LLC module under different working conditions; Using electrical state data and external target commands from the training dataset as inputs and corresponding cooperative control commands as outputs, supervised training is performed on a pre-constructed initial neural network model containing convolutional layers, long short-term memory layers, and attention layers to obtain a pre-trained model. The pre-trained model is connected to the control loop of the co-simulation model; Define a reward value calculation rule, the input of which is the measured values ​​of efficiency, total harmonic distortion and thermal stress of the co-simulation model; During the simulation environment, the current electrical state data and external target commands are input into the pre-trained model, the output cooperative control commands are applied to the co-simulation model, and the parameters of the pre-trained model are adjusted according to the reward value output by the reward value calculation rule. After iteration, the symbiotic neural network model is obtained.

4. The PFC-LLC global cooperative optimization control method based on symbiotic neural networks according to claim 3, characterized in that, The reward value calculation rule is configured as follows: When the combination of control commands output by the pre-trained model is such that the bus voltage value is within a preset voltage range and the LLC module switching frequency command is changed, a positive reward is given. When the combination of control instructions output by the pre-trained model is a combination of an instruction to simultaneously reduce the switching frequency of the PFC module and an instruction to put the LLC module into burst mode, a positive reward is given.

5. A PFC-LLC global collaborative optimization control system based on a symbiotic neural network, communicatively connected to a power control system having a PFC module and an LLC module, used to implement the steps of the control method as described in any one of claims 1-4, characterized in that, The system includes: The data acquisition module is used to synchronously acquire electrical status data of the PFC module and LLC module. The electrical status data of the PFC module includes input voltage, input current and bus voltage, and the electrical status data of the LLC module includes resonant current, resonant capacitor voltage, load current, switch status and temperature. The instruction input interface is used to receive external target instructions; The controller integrates the symbiotic neural network model; the controller is configured to: receive electrical status data from the PFC module, electrical status data from the LLC module, and external target commands, process them through the symbiotic neural network model, and generate cooperative control commands; The external target commands include output power setpoint, efficiency weighting coefficient, and total harmonic distortion (THD) constraint threshold; the cooperative control commands include the switching duty cycle control signal of the PFC module and the switching frequency or phase shift angle control signal of the LLC module. A drive circuit is used to respond to the coordinated control command and perform synchronous control on the PFC module and the LLC module.

6. The PFC-LLC global cooperative optimization control system based on symbiotic neural networks according to claim 5, characterized in that, The data acquisition module includes a voltage sampling circuit, a current sampling circuit, and a temperature sampling circuit.

7. The PFC-LLC global cooperative optimization control system based on symbiotic neural networks according to claim 5, characterized in that, The PFC module is a boost-type power factor correction circuit; the LLC module is an LLC resonant converter circuit.

8. A power supply device, characterized in that, It includes a power control system with a PFC module and an LLC module, and a PFC-LLC global collaborative optimization control system based on a symbiotic neural network as described in claim 5, which is communicatively connected to the power control system.

Images