Power supply control method and system based on dynamic transition control and neural network

Abstract

The present application relates to a kind of power supply control method and system based on dynamic transition control and neural network, belong to power electronic control technical field, solve the poor switching stability in the process of switching of existing power grid and energy storage unit, control response lag and the problem of low precision.Method includes: based on the current state quantity of power supply control system, construct objective function and constraint condition, solve the optimal power supply proportion of several energy storage units at current time by rolling optimization;The optimal power supply proportion of several energy storage units at current time and the real-time state quantity of power supply control system are input into trained neural network, generate first control signal and several second control signals;In the dynamic transition process between power grid and energy storage unit or different energy storage units, according to first control signal, adjust the power output of rectifier, according to each second control signal, adjust the power output of corresponding bidirectional converter.Realize smooth, uninterrupted power supply control.

Description

Technical Field

[0001] This invention relates to the field of power electronic control technology, and in particular to a power supply control method and system based on dynamic transient control and neural networks. Background Technology

[0002] Uninterruptible power supply (UPS) systems serve as backup power for critical loads and are core equipment for ensuring the continuous operation of information centers, medical equipment, and industrial control systems during power grid failures.

[0003] With the development of intelligent control technology, UPS systems have gradually incorporated intelligent algorithms such as neural networks and fuzzy control, achieving certain improvements in dynamic response speed and system robustness. However, most existing UPS systems still rely on fixed battery packs for energy storage, whose energy storage capacity is strictly limited by physical space and cost. Once a prolonged power grid outage occurs, the limited power of fixed battery packs will be insufficient to maintain continuous power supply to critical loads. In recent years, the widespread adoption of electric vehicles has provided a new approach to emergency power supply. Electric vehicles are equipped with power batteries with large energy storage capacity and possess flexible mobility. Through vehicle-to-load (V2L) or vehicle-to-grid (V2G) technologies, electric vehicles can participate in emergency power supply as auxiliary or primary power sources during sudden power outages, thereby breaking through the capacity bottleneck of traditional fixed battery packs.

[0004] However, the aforementioned existing technical solutions still have the following main problems and shortcomings in practical applications: First, the power supply switching stability is poor. During the switching process between the power grid and electric vehicles, the DC bus voltage fluctuates significantly, and current surges and power overshoot are prominent issues, which can easily damage precision loads.

[0005] Second, the control strategy suffers from lag and insufficient accuracy. Existing emergency power supply systems mostly employ traditional intelligent scheduling methods such as genetic algorithms combined with artificial neural networks (GA+ANN). These algorithms have slow convergence speeds and low model accuracy, making it difficult to achieve fast and accurate power feedback regulation under dynamic disturbance environments.

[0006] Third, insufficient coordination in power supply relay among multiple vehicles poses a risk of power interruption. When multiple electric vehicles work together to provide emergency power, there is a lack of an effective collaborative relay mechanism. During the process of transferring the power supply task from one electric vehicle to another, the power supply ratio of each vehicle changes abruptly, which can easily cause instantaneous power interruption and load voltage drop, making it impossible to achieve zero-interruption power supply to critical loads. Summary of the Invention

[0007] Based on the above analysis, the embodiments of the present invention aim to provide a power supply control method and system based on dynamic transition control and neural networks, in order to solve the problems of poor switching stability, lagging control response and low accuracy in the existing power grid and energy storage unit switching process.

[0008] On one hand, embodiments of the present invention provide a power supply control method based on dynamic transient control and neural networks, applied to a power supply control system consisting of a power grid, rectifiers, DC buses, several bidirectional converters, and energy storage units, comprising the following steps: Based on the current state variables of the power supply control system, an objective function and constraints are constructed, and the optimal power supply ratio of several energy storage units at the current moment is obtained through rolling optimization. The optimal power supply ratio of several energy storage units at the current moment and the real-time state of the power supply control system are input into the trained neural network to generate the first control signal and several second control signals. During the dynamic transition between the power grid and the energy storage unit or between different energy storage units, the power output of the rectifier is adjusted according to the first control signal, and the power output of the corresponding bidirectional converter is adjusted according to each second control signal.

[0009] Based on further improvements to the above method, the current state variables of the power supply control system include: DC bus voltage, load power, and the state of charge of each energy storage unit; the objective function is expressed by the following formula: , in, To predict the length of the time domain, The current moment; For the predicted first DC bus voltage at any given moment This is the reference value for the DC bus voltage; For the predicted first System loss indicators at any given time; For the first The energy storage unit in the first The change in the power supply ratio at each moment relative to the previous moment; This is a weighted penalty term for voltage deviation. This is the voltage deviation weighting coefficient; For loss-weighted penalty terms, This is the loss weighting coefficient; This is a weighted penalty term for changes in the power supply ratio. For smoothing weighting coefficients, This represents the total number of energy storage units.

[0010] Further improvements to the above method include: decision variable constraints, energy storage unit state-of-charge limit constraints, energy storage unit discharge power limit constraints, and system power balance constraints.

[0011] Based on the further improvement of the above method, the first control signal is the d-axis current reference value of the rectifier in the synchronous rotating coordinate system, and the second control signal is the switching duty cycle or frequency modulation reference value of the corresponding bidirectional converter.

[0012] Based on the above method, a further improvement is made to the neural network, which uses a KAN network, and the training process employs a composite loss function; the composite loss function is calculated using the following formula: , in, For composite loss function, To control signal error loss, To control signal smoothing loss, For state-of-charge constraint loss, Loss due to state disturbance; , , and They are respectively , , and The weight, .

[0013] Based on a further improvement of the above method, the control signal smoothing loss is calculated using the following formula: , in, and The corresponding times of two adjacent moments are the first The and the first The first control signal in each sample; and The corresponding times of two adjacent moments are the first The and the first In the nth sample A second control signal; This represents the total number of energy storage units. This represents the number of samples in the training batch. To calculate the L2 norm.

[0014] Based on a further improvement of the above method, the state-of-charge constraint loss is calculated using the following formula: , , in, For the first In the nth sample The current state of charge of each energy storage unit; For the first The preset minimum safe threshold for the state of charge of each energy storage unit; For the first In the nth sample The charge weight of each energy storage unit; This represents the total number of energy storage units. This represents the number of samples in the training batch. This is the function for finding the maximum value.

[0015] Based on further improvements to the above method, the dynamic transition process is triggered when the grid voltage remains below a preset voltage threshold for a duration exceeding a preset time threshold, and is completed when the deviation between the current power supply ratio and the corresponding optimal power supply ratio of each energy storage unit is less than a first threshold, and the deviation between the DC bus voltage and the voltage reference value is less than a second threshold.

[0016] Based on a further improvement of the above method, the energy storage unit is an electric vehicle on-board battery; the bidirectional converter is a bidirectional DC / DC converter.

[0017] On the other hand, embodiments of the present invention provide a power supply control system based on dynamic transition control and neural networks, including: The rectifier has its AC side connected to the power grid and its DC side connected to the DC bus. Several bidirectional converters, each bidirectional converter is connected to the DC bus on one side and the corresponding energy storage unit on the other side; The controller is used to construct an objective function and constraints based on the current state variables of the power supply control system, and solve for the optimal power supply ratio of several energy storage units at the current moment through rolling optimization. The optimal power supply ratio of several energy storage units at the current moment and the real-time state variables of the power supply control system are input into a trained neural network to generate a first control signal and several second control signals. During the dynamic transition between the grid and the energy storage units or between different energy storage units, the controller adjusts the power output of the rectifier according to the first control signal and adjusts the power output of the corresponding bidirectional converter according to each second control signal.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. The upper-level rolling optimization provides the globally optimal power supply ratio decision, and the lower-level neural network nonlinear mapping transforms the optimization decision into a high-precision, smooth, safe and feasible device-level control signal. Finally, the power output of the rectifier and bidirectional converter is adjusted during the dynamic transition process. The three links perform their respective functions and are closely connected to achieve a smooth and uninterrupted switching of power supply tasks between the grid and energy storage units and between multiple energy storage units. This effectively suppresses voltage fluctuations and power surges during the switching process and ensures the power supply continuity and voltage stability of critical loads.

[0019] 2. In the upper-level rolling optimization, a multi-objective function is constructed with the objectives of minimizing DC bus voltage deviation, system loss index, and power supply ratio changes. During the solution process, constraints on decision variables, energy storage unit charge state limits, discharge power limits, and system power balance constraints are introduced. This multi-objective optimization and multi-constraint collaborative design ensures that the optimal power supply ratio obtained by rolling optimization, while guaranteeing voltage stability, also takes into account energy utilization efficiency and control smoothness. At the same time, it actively avoids safety risks such as deep discharge of energy storage units, converter overload, and system power imbalance from the decision-making level, providing a safe, feasible, and economically reasonable optimization objective for the lower-level neural network execution controller.

[0020] 3. In neural network training, a composite loss function is introduced, which is composed of control signal error loss, control signal smoothing loss, state of charge constraint loss and state disturbance loss through dynamic weighting. The network is optimized from four dimensions: accuracy, smoothness, safety and robustness. This makes the control signal generated in actual deployment have high-precision tracking capability, low-frequency jitter characteristics, battery safety protection function and robustness against noise and load disturbance.

[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of a power supply control method based on dynamic transition control and neural network in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the power supply control system architecture for the application of a power supply control method based on dynamic transition control and neural network in Embodiment 1 of the present invention. Figure 3 This is a graph showing the training loss variation of the two models in Embodiment 3 of the present invention; Figure 4 This is a graph showing the power supply ratio of the energy storage unit for the two models in Embodiment 3 of the present invention; Figure 5 The DC bus voltage curves for the two models in Embodiment 3 of the present invention are shown below. Figure 6 The DC bus current curves for the two models in Embodiment 3 of the present invention are shown below. Figure 7 These are the battery terminal voltage curves of the energy storage unit in two models of Embodiment 3 of the present invention; Figure 8 This is a graph showing the switching curves for power supply from multiple energy storage units in Embodiment 3 of the present invention. Detailed Implementation

[0023] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0024] Example 1 A specific embodiment of the present invention discloses a power supply control method based on dynamic transient control and neural networks, applied to a power supply control system consisting of a power grid, rectifiers, DC buses, several bidirectional converters, and energy storage units, such as... Figure 1 As shown, it includes the following steps: S1. Based on the current state variables of the power supply control system, construct the objective function and constraints, and solve the optimal power supply ratio of several energy storage units at the current moment through rolling optimization. S2. Input the optimal power supply ratio of each energy storage unit at the current moment and the real-time status of the power supply control system into the trained neural network to generate the first control signal and several second control signals. S3. During the dynamic transition between the power grid and the energy storage unit or between different energy storage units, the power output of the rectifier is adjusted according to the first control signal, and the power output of the corresponding bidirectional converter is adjusted according to each second control signal.

[0025] It should be noted that the overall architecture of the power supply control system in this embodiment is as follows: Figure 2As shown, the rectifier's AC side is connected to the power grid, and its DC side is connected to the DC bus. It converts AC power from the grid to DC power during normal grid operation, maintains a stable DC bus voltage, and supplies power to the load. The DC bus serves as the system's energy convergence point, with critical loads connected to its downstream end. The rated voltage of the DC bus is typically 340V, and it is required to maintain the DC bus voltage within ±1% (medical load) or ±2% (industrial load) of the rated value in real time.

[0026] The power supply control system includes at least one bidirectional converter and a corresponding energy storage unit. One side of each bidirectional converter is connected in parallel to the DC bus, and the other side is connected to the corresponding energy storage unit. Each energy storage unit is connected to the DC bus via a corresponding bidirectional converter to achieve independent control of the charging and discharging power of each energy storage unit.

[0027] In this embodiment, the energy storage unit is an electric vehicle battery; the bidirectional converter is a bidirectional DC / DC converter, which adopts a high-efficiency LLC resonant topology to realize bidirectional power flow between the energy storage unit and the DC bus.

[0028] In step S1, the current state parameters of the power supply control system are first acquired in real time through corresponding voltage sensors, current sensors, and the battery management system (BMS), including: DC bus voltage. Load power and the state of charge of each energy storage unit ( , (Total number of energy storage units participating in power supply).

[0029] Furthermore, a rolling optimization problem based on a model predictive control (MPC) framework is constructed. The objective function minimizes the sum of the deviations of the DC bus voltage from the reference value, system loss parameters, and changes in the power supply ratio at each time point in the rolling time domain, within the future prediction time domain. For the ... At each sampling time, the objective function is expressed as: , in, For the current moment, To predict the length of the time domain, i.e., the number of sampling steps to be considered in the future; For the predicted first DC bus voltage at any given moment This is the reference value for the DC bus voltage; For the predicted first The system loss index at a given time moment is used to characterize the cost per kilowatt-hour or converter efficiency loss. For the first The energy storage unit in the first Power supply ratio at any given time The power supply ratio relative to the previous moment The change in power supply ratio is the ratio of the output power of each energy storage unit to the current load power. This is a weighted penalty term for voltage deviation. This is the voltage deviation weighting factor. Its function is to ensure that the DC bus voltage is always stable near the reference value, to ensure the power supply voltage quality of critical loads, and to avoid damage or malfunction of load equipment due to voltage fluctuations. For loss-weighted penalty terms, The loss weighting coefficient guides the optimizer to prioritize power supply combinations with higher energy efficiency while meeting load power requirements, thereby reducing the overall energy loss of the system during emergency power supply and extending the available power supply time. This is a weighted penalty term for changes in the power supply ratio. To smooth the weighting coefficient, this item aims to suppress drastic changes in the power supply ratio of each energy storage unit, so that the control command changes smoothly in time, laying the foundation for shock-free power switching during the subsequent dynamic transition process.

[0030] It should be noted that, , and The settings or adjustments are made according to the actual operating requirements of the power supply control system and the sensitivity of the load to voltage quality. For example, in scenarios with extremely strict voltage accuracy requirements, such as medical loads, the settings can be adjusted. Set to a larger value; in long-term emergency power supply scenarios where high energy efficiency is required, it can be set to a larger value. Appropriately increase the size; in scenarios where there are high requirements for the lifespan and electromagnetic compatibility of switching devices, it can be increased. Increase them appropriately. The specific values ​​of each weighting coefficient can be determined through offline simulation, parameter scanning, or engineering experience.

[0031] In the process of minimizing the above objective function, the constraints that also need to be satisfied include: decision variable constraints, energy storage unit charge state limit constraints, energy storage unit discharge power limit constraints, and system power balance constraints.

[0032] Among them, the decision variable constraint refers to the power supply ratio of each energy storage unit. It should be within its feasible range, that is And the sum of the power supply ratios of all energy storage units satisfies The difference is supplemented by power from the grid side; when the grid is completely disconnected, the sum of the power supply ratios of all energy storage units is 1, meaning that the load power is fully borne by each energy storage unit. This constraint ensures the physical feasibility of the power supply ratio.

[0033] The state of charge (SBC) limit constraint for energy storage units refers to the requirement that the SBC of each energy storage unit must not fall below its preset minimum safety threshold during discharge. ,Right now The minimum safety threshold is set independently based on the battery type, health status, and operational requirements of the energy storage unit (e.g., 20% for new batteries and 25% or higher for aging batteries). This constraint aims to prevent deep discharge of the energy storage unit, thus preventing irreversible battery life degradation or insufficient subsequent power supply capacity due to over-discharge, thereby ensuring the sustainability of long-term emergency power supply.

[0034] The upper limit constraint on the discharge power of energy storage units means that the output power of each energy storage unit must not exceed the rated power of its corresponding bidirectional converter. ,Right now The purpose of this constraint is to prevent the bidirectional converter from being damaged due to overload operation and to protect the safety of power electronic equipment.

[0035] System power balance constraint refers to the ratio of grid-side output power to the discharge power of each energy storage unit (including bidirectional converter efficiency). The sum of the active power and the load power should equal the current load power, ensuring that the active power of the system remains balanced at all times. This constraint is the fundamental guarantee for maintaining the stability of the DC bus voltage, avoiding voltage drops due to power shortages or voltage spikes due to power excess.

[0036] The above four constraints work together to limit the optimization problem from four dimensions: the rationality of the power supply ratio, the safety protection of the energy storage unit, the power limitation of the equipment, and the conservation of system energy. This ensures that the optimal power supply ratio obtained by rolling optimization achieves multi-objective coordinated optimization of voltage stability, energy utilization efficiency, and control smoothness while meeting the requirements of safe system operation.

[0037] In the above objective function and constraints, the future time... , and Isoparameters are obtained by constructing a state-space model based on the current state variables to predict the dynamic behavior of the system in the future prediction time domain.

[0038] Specifically, a linear parametric time-varying (LPV) model is used to describe the dynamic characteristics of the system state evolution with control input in state-space form. After linearization at different operating points, a discrete state-space model is obtained, as shown in the following formula: , , in, For the current number The state quantities at each moment include: DC bus voltage State of charge of each energy storage unit and load power , ;No. Control input variables at time 1 The power supply ratio for each energy storage unit, ; Let be the state transition matrix, describing the state transition from... arrive The linear transfer relationship reflects the dynamic characteristics of the system; The control input matrix describes the influence of control variables on state evolution; This is the output matrix, used to extract the controlled variable from the state variable; This refers to the system output; in this embodiment, the system output is the DC bus voltage. The weighted penalty term for voltage deviation in the objective function is calculated directly based on the predicted value of the output.

[0039] In the At any given moment, based on the current state quantity Based on the aforementioned state-space model, the future is obtained through recursive calculation. The state prediction sequence of the step.

[0040] Specifically, based on the current state and candidate control sequences By recursively applying the state transition equation, the predicted state values ​​at each future time step can be obtained. Then, the objective function and the required constraints are extracted from the output equation. , Predicted values, etc.

[0041] The above recursive prediction process is the foundation for solving the objective function in rolling optimization. Within each control cycle of rolling optimization, the objective function is used as the optimization objective, and the constraints are used as the solution boundary. By solving this constrained optimization problem, the future... The optimal power supply ratio sequence for each control cycle (wherein) To control the time domain, Based on the principle of rolling optimization control, only the values ​​corresponding to the current moment in the optimal control sequence are extracted, thus obtaining the optimal power supply ratio of each energy storage unit at the current moment. This serves as the final output of step S1. The optimal power supply ratio satisfies both voltage stability accuracy requirements and the safety constraints and power switching smoothness of the energy storage unit, providing an accurate and reliable optimization target for the subsequent generation of execution layer control signals.

[0042] At the next moment Upon arrival, based on the updated current system state, the above rolling optimization process is re-executed, thereby achieving real-time response and closed-loop control to system dynamic changes and external disturbances.

[0043] In step S2, the above-mentioned optimal power supply ratio and the real-time state of the power supply system are input into the trained neural network, which performs nonlinear mapping to generate a first control signal and a second control signal for directly driving the execution device.

[0044] It should be noted that the input vector of the neural network consists of two parts: one part is the optimal power supply ratio of each energy storage unit output from step one. One part represents the optimal power allocation target solved by the upper-level rolling optimization at the current moment, providing a clear direction for optimization for the lower-level execution controller; the other part consists of the real-time state variables of the power supply control system, including at least: DC bus voltage. Load power and the battery terminal voltage of each energy storage unit In scenarios involving multiple energy storage units switching over in succession, real-time status parameters further include: the remaining switching time during the current emergency power supply phase. This time is used to characterize the remaining duration that the energy storage unit currently performing a power supply task is expected to maintain power supply, providing a timing reference for the neural network to coordinate power supply tasks among multiple energy storage units.

[0045] Combining the two pieces of information above forms the complete input vector of the neural network: .

[0046] The input vector of the neural network contains both the decision results of the upper-level optimization and the actual operating state of the system, enabling the neural network to fully perceive the system's operating conditions and provide a sufficient information basis for the subsequent generation of high-precision control signals.

[0047] The output vector of the neural network includes: the first control signal Second control signal ( ), represented as: .

[0048] It should be noted that the neural network in this embodiment adopts the KAN (Kolmogorov-Arnold Network) network structure. Unlike the traditional multilayer perceptron (MLP) which sets fixed activation functions on nodes, the KAN network configures learnable activation functions on the edges of the network, and the nodes only perform simple summation operations. This structural feature allows the KAN network to maintain a relatively shallow network depth while still possessing a strong ability to approximate high-dimensional nonlinear functions, making it particularly suitable for the power supply control scenario with multiple coupled state variables and complex dynamic characteristics as described in this embodiment.

[0049] For example, the KAN neural network adopts a three-layer network structure of [7, 12, 3], that is, the input layer has a dimension of 7 (corresponding to the number of input variables in the scenario of 2 energy storage units), the number of hidden layer nodes is 12, and the output layer has a dimension of 3, corresponding to one first control signal (the control signal of the rectifier) ​​and two second control signals (the control signals of the converters corresponding to the two energy storage units respectively). Its edge activation function adopts a combination of the third-order B-spline basis function and the SiLU function. This combination has both the local adjustability and smoothness of the B-spline function and the nonlinear expression capability of the SiLU function, which helps the network to converge quickly and achieve high fitting accuracy during training.

[0050] In the offline phase, the KAN network is trained by collecting a large amount of sample data, and a direct nonlinear mapping relationship is established between the input system state space and the output control action space.

[0051] It should be noted that a large amount of sample data was generated by simulating the power supply control system under different operating conditions. The operating conditions covered include at least: normal grid operation, fault switching conditions with grid voltage drop or interruption, switching conditions with relay power supply between different energy storage units, and disturbance conditions with step changes or random fluctuations in load power.

[0052] Each sample contains system state variables (including DC bus voltage, state of charge of each energy storage unit, load power, battery terminal voltage of each energy storage unit, and remaining switching time) at the same sampling time, along with the corresponding ideal control signals (including: rectifier d-axis current reference value). and the reference duty cycle or frequency modulation amount of each bidirectional converter This forms a mapping relationship between "system state variables and control signals".

[0053] During training, training samples are organized into consecutive time-series batches in chronological order. Each training batch contains... The training uses samples that increase sequentially over time to ensure a genuine physical temporal correlation between adjacent samples within a batch. Different training batches can be randomly shuffled to enhance the network's generalization ability.

[0054] To ensure that the KAN network can output high-precision, smooth, and safe control signals in actual deployment, the training process of the KAN network in this embodiment employs a composite loss function. This composite loss function... Loss due to control signal error Control signal smoothing loss State of charge constraint loss and state disturbance loss The expression obtained by weighted summation is: , in, , , and They are respectively , , and The weights satisfy the normalization condition. Weight , and All adjustments are dynamically made based on the real-time status variables of the power supply control system, enabling the training process to adaptively adjust the optimization focus according to different operating conditions. The calculation methods for each loss function and weight are explained in detail below.

[0055] ① The control signal error loss is obtained by calculating the error between each output control signal and its corresponding reference control signal, and is used to ensure the basic fitting accuracy of the neural network. For signals containing... A system with one energy storage unit, KAN network output One control signal (1 first control signal and...) The loss value of each control signal (a second control signal) is defined as the sum of the squares of the L2 norms of the errors between each control signal and the corresponding reference control signal, as shown in the following formula: , in, This represents the number of samples in the training batch. and The KAN network is respectively paired with the first The first control signal output by the sample and the first... A second control signal; and These are the corresponding reference control signals, provided by the ideal control commands in the offline simulation dataset; To calculate the L2 norm.

[0056] Weights of control signal error loss It is determined based on the ratio of the change in DC bus voltage at adjacent moments to the current DC bus voltage, as shown in the formula below: .

[0057] When the DC bus voltage fluctuation increases The corresponding increase in voltage control precision during training prioritizes bus voltage stability.

[0058] ② The control signal smoothing loss is obtained by calculating the changes in the control signal at adjacent time points. It is used to constrain the amplitude of the control signal changes at adjacent sampling time points, suppress high-frequency jitter in the control signal, and reduce the losses and electrical shocks of the converter's switching devices. Its loss value is defined as the sum of the squares of the L2 norms of the changes in the control signal at adjacent time points, as shown in the following formula: , in, and They are respectively the first time corresponding to the previous time. The first control signal and the first sample A second control signal. By minimizing this term, abrupt changes in the control signal can be effectively avoided, making the power regulation process smoother.

[0059] Weights for smoothing loss of control signal It is determined based on the ratio of the change in the power supply ratio at adjacent time points to the maximum value of the power supply ratio, as shown in the following formula: , in, This is the maximum possible value for the power supply ratio (usually 1, i.e., 100%).

[0060] When the power supply ratio of each energy storage unit changes drastically... The corresponding increase strengthens the constraint on the smoothness of the control signal during the training process, suppressing abrupt changes in control commands.

[0061] ③ The state-of-charge constraint loss is used to actively incorporate the safety constraints of the energy storage unit during the training phase, forcing the network to learn control strategies to avoid deep discharge. It is obtained by calculating the deviation of each energy storage unit's state of charge when it is lower than the preset minimum safety threshold and the corresponding energy storage unit's charge weight. That is, when the state of charge of a certain energy storage unit is higher than the minimum safety threshold, the corresponding term and charge weight are zero, indicating that the current control strategy meets the battery's safe operation requirements.

[0062] The formula for the state-of-charge constraint loss is as follows: , , in, For the first In the nth sample The current state of charge of each energy storage unit; For the first Each energy storage unit has a preset minimum safe threshold for state of charge to avoid irreversible capacity degradation and shortened lifespan due to deep discharge of the energy storage battery. For the first In the nth sample The charge weight of each energy storage unit is obtained by calculating the charge state deviation ratio; This is the function for finding the maximum value.

[0063] The penalty is only applied when the state of charge of a certain energy storage unit is lower than its own safety threshold. Considering that averaging the state of charge constraint loss may dilute the penalty for a few out-of-bounds samples because most samples in the batch are in the safe range, which is not conducive to the network effectively learning strategies to avoid deep discharge, only summation is performed during the calculation. This significantly increases the importance of battery safety protection for the energy storage unit and forces the network to learn a control strategy that prioritizes ensuring that the state of charge of the energy storage unit returns to its safe range.

[0064] Weights of the state-of-charge constraint loss It is calculated using the following formula: .

[0065] A two-layer weighted structure is used for the state-of-charge constraint loss, with the inner layer weight... Differentiated penalties for different energy storage units were implemented, with outer weights. A dynamic balance was achieved between battery safety protection and other control objectives.

[0066] ④ State disturbance loss is used to enhance the network's robustness to sampling noise, random load fluctuations, and prediction bias. It is obtained by calculating the errors of each control signal output before and after the input is superimposed with disturbance noise. It is defined as the sum of the squares of the L2 norms of the deviations between the control signal output by the neural network and the reference control signal under the original undisturbed input after a small disturbance vector is superimposed on the input vector. , in, This is the original input sample; The perturbation vector is a superimposed vector with the same dimension as the input vector. Each element in the perturbation vector follows a small random distribution (e.g., Gaussian or uniform distribution) with a mean of zero and an amplitude set according to the range of the physical quantity corresponding to that dimension. and The first control signal and the second control signal output by the network after the input perturbation vector is superimposed are respectively the input perturbation vector and the output of the network. A second control signal. By setting the disturbance amplitude for each dimension, the disturbance intensity of each physical quantity is matched with its dimension and the actual noise level that may be encountered, avoiding excessive or insufficient disturbance to certain dimensions due to differences in dimensions.

[0067] Weights of state perturbation loss It is determined based on the ratio of the change in load power at adjacent time points to the current load power, as shown in the formula below: , in, This represents the change in load power between adjacent time points.

[0068] When the load fluctuates significantly The corresponding increase strengthens the training process to enhance the robustness of the network, enabling the controller to output stable and reliable control signals even under load disturbance conditions.

[0069] Through training using the aforementioned composite loss function and dynamic weighting mechanism, the KAN network can learn a control strategy that balances accuracy, smoothness, safety, and robustness even in the offline phase. During the initial switching phase of a grid fault, when voltage fluctuations are significant, the dynamic weights are automatically increased. To ensure voltage recovery speed and control accuracy, the dynamic weights are automatically increased after the system enters steady-state operation. This ensures the smoothness of the control signals; at the same time, since the state of charge of each energy storage unit is usually much higher than the minimum safety threshold, Automatically decreasing or maintaining a value of zero indicates that the current control strategy meets the requirements for safe battery operation. When there are short-term, significant fluctuations in load or noise in the sensor data, the dynamic weight automatically increases. This enhances the controller's anti-interference capability. Through this adaptive adjustment mechanism, the KAN network, after online deployment, can adapt to various complex operating conditions such as grid faults, load surges, and multiple energy storage units taking over, consistently outputting stable and reliable control signals.

[0070] After the KAN network is trained, during implementation, upon receiving the input vector, the KAN network synchronously generates a first control signal and a second control signal through forward propagation calculation. The first control signal is the d-axis current reference value of the rectifier in a synchronously rotating coordinate system, and the second control signal is the switching duty cycle or frequency modulation reference value of the corresponding bidirectional converter. These control signals are directly transmitted to step S3 to drive the rectifier and each bidirectional converter to perform corresponding power regulation actions, completing the real-time conversion from optimization decision quantities to physical control commands.

[0071] It should be noted that, in this embodiment, the working state of the power supply control system is divided into grid main supply mode, dynamic transition mode and energy storage unit power supply mode according to the grid status and energy storage unit power supply status. In all three modes, steps S1 to S3 are continuously executed in a loop.

[0072] Specifically, the grid-dominated supply mode is for normal grid operation, where load power is primarily supplied by the grid through rectifiers. In this case, the optimal power supply ratio for each energy storage unit, calculated in step S1, is... Typically, the value is zero or a small value (when the energy storage unit is in standby or low-power charging state); the second control signals generated in step S2 The first control signal corresponds to the non-discharge or low-power charging command. The rectifier maintains power supply to the load and stabilizes the DC bus voltage. In this mode, the grid voltage and system status are continuously monitored, and the optimal power supply ratio is predicted in advance for possible grid fault switching.

[0073] The dynamic transition mode addresses the dynamic transition process when a grid fault occurs or when energy storage units need to switch over. When the power supply control system is currently in grid-dominant mode, it monitors the grid voltage in real time. When the grid voltage remains below a preset voltage threshold When the duration exceeds a preset time threshold, a grid-side fault is determined, triggering a dynamic transition process from the grid to the energy storage unit. The energy storage unit then initiates its power supply mode to replace or supplement grid-side power. For example, the voltage threshold is set to 0.8 to 0.9 times the rated voltage, which can be adjusted according to actual operating conditions and load sensitivity; the time threshold ranges from 10ms to 500ms, which can be adjusted according to the load's tolerance for voltage interruption. The dynamic transition process between different energy storage units is automatically achieved by the upper layer through rolling optimization based on the state of charge of each energy storage unit and the load power demand, without the need to set an external trigger threshold. In this mode, the optimal power supply ratio of each energy storage unit, solved by rolling optimization in step S1, changes significantly; the control signals generated in step S2 exhibit a coordinated adjustment characteristic of "one increasing while the other decreases"; and step S3 performs a smooth power transfer based on the control signals. This mode serves as a transition bridge between the grid-based main supply mode and the energy storage unit power supply mode, and its duration is typically short (hundreds of milliseconds to several seconds).

[0074] The energy storage unit power supply mode is designed for operation processes where the grid is completely disconnected or the power supply quality does not meet the load requirements. In this case, the load power is provided entirely by one or more energy storage units through bidirectional converters. Step S1 continuously predicts the state of charge evolution of each energy storage unit and the load power demand, and optimizes the power supply ratio allocation of each energy storage unit under the condition of meeting battery safety constraints. Step S2 generates control signals for each bidirectional converter in real time based on the optimization results. Step S3 maintains the DC bus voltage stability based on the control signals to ensure continuous power supply to the load.

[0075] During the dynamic transition process, the rectifier's power output is adjusted according to the first control signal, and the corresponding bidirectional converter's power output is adjusted according to each of the second control signals. The rectifier employs a vector control strategy based on a synchronous rotating coordinate system (dq coordinate system), where its active power output is proportional to the d-axis current. The first control signal is used as the d-axis current reference value for the rectifier's inner current loop. A PI regulator and PWM modulation stage control the on / off state of the rectifier's switching transistors, ensuring the actual d-axis current tracks the first control signal. When the first control signal decreases, the active power absorbed by the rectifier from the grid decreases accordingly; when the first control signal increases, the rectifier's output power increases accordingly. For bidirectional DC / DC converters using an LLC resonant topology, the magnitude and direction of power transfer are controlled by adjusting the switch duty cycle or switching frequency. The second control signal is used as a reference input for the modulation stage of the bidirectional converter. When the second control signal is given in the form of a duty cycle, the on-time ratio of the converter's switching devices is changed through PWM modulation, thereby altering the transmitted power. When the second control signal is given in the form of a frequency modulation reference, the power transmission gain is adjusted by changing the switching frequency to move it closer to or further away from the resonant frequency. When the duty cycle increases or the frequency moves closer to the resonant point, the power transmitted by the bidirectional converter increases, and the energy storage unit releases more energy to the DC bus; conversely, the transmitted power decreases, and the output power of the energy storage unit decreases.

[0076] Under the aforementioned power regulation mechanism, the first control signal and each of the second control signals exhibit a coordinated regulation characteristic of "one increasing as the other decreases" during the dynamic transition process.

[0077] Specifically, when power supply is switched between the grid and the energy storage unit: if the power supply from the grid needs to be reduced, the first control signal gradually decreases, while the second control signal or frequency modulation reference value increases synchronously, so that the output power of the energy storage unit increases by the same amount to fill the power gap left by the grid disconnection; if the grid resumes power supply, the first control signal gradually rises, while the second control signal decreases synchronously, until the grid completely takes over the power supply to the load.

[0078] When different energy storage units relay power supply tasks: the second control signal corresponding to the energy storage unit to be decommissioned gradually decreases, causing its output power to decline smoothly; simultaneously, the second control signal corresponding to the energy storage unit to be connected increases synchronously, causing its output power to increase by the same amount. The power changes on both sides are continuous, coordinated, and matched in time, ensuring that the total power supply does not experience interruption or sharp drop during this process.

[0079] The system status is continuously monitored during the dynamic transition process. The dynamic transition process is completed when the deviation between the current power supply ratio and the corresponding optimal power supply ratio of each energy storage unit is less than the first threshold and the deviation between the DC bus voltage and the voltage reference value is less than the second threshold.

[0080] For example, the first threshold is set to [2%, 5%] and the second threshold is set to [1%, 3%], depending on the control accuracy requirements and the load's sensitivity to voltage accuracy.

[0081] After the dynamic transition process is completed, the power supply control system enters the corresponding steady-state mode (e.g., energy storage unit power supply mode or grid main supply mode). It continues to run in a loop according to the process from steps S1 to S3, repeatedly executing state acquisition, rolling optimization, neural network mapping, and control signal execution at each sampling time to achieve continuous closed-loop control of the entire power supply process.

[0082] Compared with existing technologies, this embodiment provides a power supply control method based on dynamic transition control and neural networks. It provides globally optimal power supply ratio decisions through upper-level rolling optimization, and then transforms the optimization decisions into high-precision, smooth, and safe device-level control signals through nonlinear mapping of the lower-level neural network. Finally, it adjusts the power output of the rectifier and bidirectional converter during the dynamic transition process. The three links perform their respective functions and are closely connected to achieve smooth and uninterrupted switching of power supply tasks between the grid and energy storage units, as well as between multiple energy storage units. It effectively suppresses voltage fluctuations and power surges during the switching process and ensures the power supply continuity and voltage stability of critical loads. In the upper-level rolling optimization, a multi-objective function is constructed with the objectives of minimizing DC bus voltage deviation, system loss index, and power supply ratio changes. During the solution process, constraints on decision variables, energy storage unit state-of-charge limits, discharge power limits, and system power balance are introduced. This multi-objective optimization and multi-constraint collaborative design ensures that the optimal power supply ratio obtained through rolling optimization, while guaranteeing voltage stability, also considers energy utilization efficiency and control smoothness. Simultaneously, it proactively avoids safety risks such as deep discharge of energy storage units, converter overload, and system power imbalance at the decision-making level, providing a safe, feasible, and economically reasonable optimization objective for the lower-level neural network execution controller. In neural network training, a composite loss function is introduced, consisting of control signal error loss, control signal smoothing loss, state-of-charge constraint loss, and state disturbance loss weighted dynamically. This function simultaneously optimizes the network from four dimensions: accuracy, smoothness, safety, and robustness. This ensures that the control signal generated in actual deployment possesses high-precision tracking capability, low-frequency jitter characteristics, battery safety protection functions, and robustness against noise and load disturbances.

[0083] Example 2 Another embodiment of the present invention discloses a power supply control system based on dynamic transition control and neural networks, thereby implementing the power supply control method based on dynamic transition control and neural networks in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1. The system includes: The rectifier has its AC side connected to the power grid and its DC side connected to the DC bus. Several bidirectional converters, each bidirectional converter is connected to the DC bus on one side and the corresponding energy storage unit on the other side; The controller is used to construct an objective function and constraints based on the current state variables of the power supply control system, and solve for the optimal power supply ratio of several energy storage units at the current moment through rolling optimization. The optimal power supply ratio of several energy storage units at the current moment and the real-time state variables of the power supply control system are input into a trained neural network to generate a first control signal and several second control signals. During the dynamic transition between the grid and the energy storage units or between different energy storage units, the controller adjusts the power output of the rectifier according to the first control signal and adjusts the power output of the corresponding bidirectional converter according to each second control signal.

[0084] It should be noted that the DC bus serves as the energy convergence point of the system, and its downstream end is used to connect critical loads. Each energy storage unit is connected to a corresponding bidirectional converter.

[0085] The controller is connected to the power grid and each energy storage unit via signal acquisition lines to collect the state variables of the power supply control system in real time. The controller is also communicatively connected to the rectifier and each bidirectional converter, responsible for executing the power supply control method described in Example 1. The controller's sampling frequency is 10kHz to ensure real-time performance and accuracy. A trained neural network is deployed within the controller to achieve real-time mapping from optimization decision variables to device-level control signals.

[0086] Since the power supply control system based on dynamic transition control and neural network in this embodiment can be mutually referenced with the aforementioned power supply control method based on dynamic transition control and neural network, and this is a repetition, it will not be repeated here. Because this system embodiment shares the same principle as the above method embodiment, it also possesses the corresponding technical effects of the above method embodiment.

[0087] Example 3 To further verify the technical effectiveness of the power supply control method provided in Example 1, this example is explained in conjunction with specific simulation experiment results.

[0088] In this embodiment, the power supply control system structure used for verification is consistent with that in Embodiment 2, including a power grid, a rectifier, a DC bus, a controller, at least one bidirectional DC / DC converter, and at least one corresponding electric vehicle energy storage unit. The controller sampling frequency is set to 10kHz, and the controller prediction time domain... Set to 50 to control the time domain. The value is set to 10, meaning that within each control cycle, the controller predicts the system behavior for the next 500ms and optimizes the control command sequence for the next 100ms. The edge activation function of the KAN network uses a combination of a 3rd-order B-spline basis function and a SiLU function. The above parameters are only exemplary settings for this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0089] The following sections compare and verify the method of Example 1 (MPC+KAN) with the traditional genetic algorithm combined with artificial neural network (GA+ANN) control strategy from three aspects: neural network model training convergence, grid sudden fault disconnection and switching performance, and multi-energy storage unit relay power supply smoothness.

[0090] Among them, the training convergence comparison and the grid sudden fault disconnection scenario correspond to the single energy storage unit operation condition, in which the KAN network adopts a three-layer structure of [5, 12, 2]; the multi-energy storage unit relay power supply scenario corresponds to the two energy storage units (EV1, EV2) cooperating power supply operation condition, in which the KAN network adopts a three-layer structure of [7, 12, 3].

[0091] ① Comparative verification of the convergence of neural network model training.

[0092] To verify the learning ability and fitting accuracy of the KAN network in the proposed method, this embodiment compares the training loss curves of the model using a genetic algorithm combined with an artificial neural network (GA+ANN) and the model using the MPC+KAN model in Example 1. During training, both models used the same training and validation sets, with 1000 training epochs (Epochs = 0 to 1000). The loss value ranged from 0 to 200, with a lower loss value indicating a higher fitting accuracy of the model to the nonlinear mapping relationship of the system.

[0093] See Figure 3The training comparison results show that in the early stages of training (Epochs=0 to 200), the loss value of the GA+ANN model rapidly decreased from nearly 200 to around 100. However, as the number of training epochs increased (Epochs=200 to 1000), the rate of loss decrease slowed significantly, eventually stabilizing in the 80-90 range, indicating some residual fitting error. In contrast, the loss value of the MPC+KAN model remained lower than that of the GA+ANN model throughout the entire training cycle. In the early stages of training (Epochs=0 to 100), the loss value rapidly decreased to below 50, and the number of epochs required to achieve stable convergence (loss change rate less than 0.1% / epoch) was approximately 200, which is about 50% less than that of the GA+ANN model (approximately 400 epochs). Ultimately, the training loss value was significantly lower than that of the GA+ANN model. The results demonstrate that the MPC+KAN model has a stronger ability to learn the complex nonlinear dynamic characteristics of the power supply system and can more accurately establish the mapping relationship between "system state and control signal", providing a high-precision model guarantee for realizing DC bus voltage stability control and smooth power transfer in subsequent dynamic scenarios.

[0094] ②Verification of power grid failure disconnection scenarios.

[0095] To simulate the extreme scenario of a sudden grid failure causing a main grid disconnection during actual grid operation, this embodiment sets the following operating conditions: at t=0.7s, the grid fault disconnects, and the system switches from the main grid supply mode to the energy storage unit supply mode; at t=1.2s, the fault recovers, and the system returns to the main grid supply mode. The dynamic response characteristics of the energy storage unit supply ratio, DC bus voltage, DC bus current, and energy storage unit battery terminal voltage are compared under the GA+ANN control strategy and the MPC+KAN control strategy.

[0096] See Figure 4 Regarding the responsiveness of the energy storage unit's power supply ratio: Under the GA+ANN strategy, after a grid fault is triggered, there is a delay of approximately 0.1 seconds in the process of the energy storage unit's power supply ratio rising from its initial value to 100%, with a relatively gentle rising edge slope; during fault recovery, the power supply ratio decreases with significant fluctuations and a longer transition time. Under the MPC+KAN strategy, the energy storage unit's power supply ratio rises rapidly to 100% almost without delay at the moment of a grid fault, with a response speed approximately 0.15 seconds faster than GA+ANN; during fault recovery, the power supply ratio decreases smoothly and quickly stabilizes to its initial value without significant oscillations. This indicates that the method of this invention can more quickly detect grid faults and trigger emergency power supply from the energy storage unit, and the smoothness of the mode switching process is superior.

[0097] See Figure 5Regarding the stability of the DC bus voltage: Under the GA+ANN strategy, the bus voltage exhibits a significant spike (up to approximately 342V) during grid faults, and the voltage fluctuation range during the energy storage unit power supply phase reaches approximately 2V (from 340V to 342V). After fault recovery, continuous oscillations occur during the voltage drop process. Under the MPC+KAN strategy, the bus voltage remains stable around 340V throughout the entire process. There is no significant voltage spike during fault switching, and the voltage fluctuation during the power supply phase does not exceed 0.5V. After fault recovery, the voltage immediately and smoothly returns to its initial value. These results demonstrate that MPC+KAN, through accurate modeling of the system's nonlinear dynamics using the KAN network and combined with the predictive control capabilities of MPC, effectively suppresses voltage surges caused by fault switching and energy storage unit power supply, ensuring high stability of the DC bus voltage.

[0098] See Figure 6 Regarding the dynamic smoothness of DC bus current: Under the GA+ANN strategy, the bus current fluctuates significantly during fault switching (peak value approximately 1.1A, trough value approximately 0.8A), and the current continues to oscillate slightly during power supply. Even after fault recovery, the process of the current returning to its initial value remains unstable. Under the MPC+KAN strategy, the bus current is nearly stable throughout, with only a very slight fluctuation at the moment of fault switching (fluctuation range less than 0.05A), subsequently stabilizing at approximately 1A until fault recovery. Current stability directly affects system power transmission efficiency and the lifespan of power electronic devices. The low-fluctuation characteristics under the MPC+KAN strategy demonstrate that it can more accurately match the power supply of the energy storage unit with the load demand, reducing current surges caused by power mismatch.

[0099] See Figure 7 Regarding the ability to maintain the battery terminal voltage of the energy storage unit: Under the GA+ANN strategy, during the power supply period of the energy storage unit (t=0.7s to t=1.2s), the battery terminal voltage showed a slight downward trend from its initial value and exhibited minor fluctuations. Under the MPC+KAN strategy, the battery terminal voltage remained almost unchanged throughout the entire period, stably maintaining near its initial value. This demonstrates the fine-grained scheduling capability of MPC+KAN for the battery power of the energy storage unit, ensuring power output for emergency power supply while avoiding drastic voltage changes caused by high-current charging and discharging, thus extending the battery life of the energy storage unit and improving the continuity of power supply.

[0100] In summary, under the extreme scenario of sudden grid failure and disconnection, the MPC+KAN method in Example 1 is significantly superior to the GA+ANN strategy in terms of energy storage unit power supply mode switching speed, DC bus voltage stability, DC bus current smoothness, and energy storage unit battery terminal voltage maintenance capability, and can more reliably achieve smooth transition and stable operation of emergency power supply.

[0101] ③ Verification of a scenario where multiple energy storage units relay power supply.

[0102] To verify the smooth switching characteristics of multi-energy storage unit collaborative relay power supply, this embodiment sets up a working condition in which multiple energy storage units relay power supply in sequence, and examines the dynamic response when the power supply task is transferred between energy storage units.

[0103] See Figure 8 When the system triggers the energy storage unit relay (such as the first energy storage unit EV1 transferring the power supply task to the second energy storage unit EV2), the dynamic response of the power supply ratio exhibits a highly coordinated "give-and-take" characteristic: the power supply ratio of EV1 decreases steadily and linearly over time, while the power supply ratio of EV2 increases synchronously and smoothly. The changes in both are without any sudden jumps, oscillations, or delays. The "decline" and "increase" of the power supply ratio are perfectly matched, ensuring that the total power supply of the system remains stable.

[0104] This smooth switching stems from the core advantages of MPC+KAN: on the one hand, the KAN network accurately captures the nonlinear power output characteristics of the energy storage unit's battery, providing high-precision model support for power scheduling of multiple energy storage units; on the other hand, the rolling optimization mechanism of the MPC controller can proactively plan the power transfer rhythm and dynamically adjust the power supply ratio of each energy storage unit based on the real-time system status, making the handover process of power supply tasks continuous and coordinated, fundamentally avoiding power surges and power interruptions, and ensuring the continuity and stability of emergency power supply.

[0105] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0106] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A power supply control method based on dynamic transient control and neural networks, applied to a power supply control system consisting of a power grid, rectifiers, DC buses, several bidirectional converters, and energy storage units, characterized in that... Includes the following steps: Based on the current state variables of the power supply control system, an objective function and constraints are constructed, and the optimal power supply ratio of several energy storage units at the current moment is obtained through rolling optimization. The optimal power supply ratio of several energy storage units at the current moment and the real-time state of the power supply control system are input into the trained neural network to generate a first control signal and several second control signals. During the dynamic transition between the power grid and the energy storage unit or between different energy storage units, the power output of the rectifier is adjusted according to the first control signal, and the power output of the corresponding bidirectional converter is adjusted according to each second control signal.

2. The power supply control method based on dynamic transition control and neural network according to claim 1, characterized in that, The current state variables of the power supply control system include: DC bus voltage, load power, and state of charge of each energy storage unit; the objective function is expressed by the following formula: ， in, To predict the length of the time domain, The current moment; For the predicted first DC bus voltage at any given moment This is the reference value for the DC bus voltage; For the predicted first System loss indicators at any given time; For the first The energy storage unit in the first The change in the power supply ratio at each moment relative to the previous moment; This is a weighted penalty term for voltage deviation. This is the voltage deviation weighting coefficient; For loss-weighted penalty terms, This is the loss weighting coefficient; This is a weighted penalty term for changes in the power supply ratio. For smoothing weighting coefficients, This represents the total number of energy storage units.

3. The power supply control method based on dynamic transition control and neural network according to claim 1, characterized in that, The constraints include: decision variable constraints, energy storage unit charge state limit constraints, energy storage unit discharge power limit constraints, and system power balance constraints.

4. The power supply control method based on dynamic transition control and neural network according to claim 1, characterized in that, The first control signal is the d-axis current reference value of the rectifier in the synchronous rotating coordinate system, and the second control signal is the switching duty cycle or frequency modulation reference value of the corresponding bidirectional converter.

5. The power supply control method based on dynamic transition control and neural network according to claim 1, characterized in that, The neural network uses a KAN network, and the training process employs a composite loss function; the composite loss function is calculated using the following formula: ， in, For composite loss function, To control signal error loss, To control signal smoothing loss, For state-of-charge constraint loss, Loss due to state disturbance; , , and They are respectively , , and The weight, .

6. The power supply control method based on dynamic transition control and neural network according to claim 5, characterized in that, The control signal smoothing loss is calculated using the following formula: ， in, and The corresponding times of two adjacent moments are the first The and the first The first control signal in each sample; and The corresponding times of two adjacent moments are the first The and the first In the nth sample A second control signal; This represents the total number of energy storage units. This represents the number of samples in the training batch. To calculate the L2 norm.

7. The power supply control method based on dynamic transition control and neural network according to claim 5, characterized in that, The state-of-charge constraint loss is calculated using the following formula: ， ， in, For the first In the nth sample The current state of charge of each energy storage unit; For the first The preset minimum safe threshold for the state of charge of each energy storage unit; For the first In the nth sample The charge weight of each energy storage unit; This represents the total number of energy storage units. This represents the number of samples in the training batch. This is the function for finding the maximum value.

8. The power supply control method based on dynamic transition control and neural network according to claim 2, characterized in that, The dynamic transition process is triggered when the grid voltage remains below a preset voltage threshold for a duration exceeding a preset time threshold. It is completed when the deviation between the current power supply ratio and the corresponding optimal power supply ratio of each energy storage unit is less than a first threshold, and the deviation between the DC bus voltage and the voltage reference value is less than a second threshold.

9. The power supply control method based on dynamic transition control and neural network according to claim 1, characterized in that, The energy storage unit is an electric vehicle battery; the bidirectional converter is a bidirectional DC / DC converter.

10. A power supply control system based on dynamic transient control and neural networks, characterized in that, include: The rectifier has its AC side connected to the power grid and its DC side connected to the DC bus. Several bidirectional converters, each bidirectional converter having one side connected to the DC bus and the other side connected to the corresponding energy storage unit; The controller is used to construct an objective function and constraints based on the current state of the power supply control system, and solve for the optimal power supply ratio of several energy storage units at the current moment through rolling optimization. The optimal power supply ratio of several energy storage units at the current moment and the real-time state of the power supply control system are input into a trained neural network to generate a first control signal and several second control signals. During the dynamic transition between the power grid and the energy storage units or between different energy storage units, the power output of the rectifier is adjusted according to the first control signal, and the power output of the corresponding bidirectional converter is adjusted according to each second control signal.

Images