A SiC power supply optimization method and system based on reinforcement learning

Abstract

This invention relates to the fields of power electronics and artificial intelligence, and discloses a SiC power supply optimization method and system based on reinforcement learning. The SiC power supply optimization method based on reinforcement learning includes: establishing a fractional-order long-time memory model; constructing a verifiable neural network controller and implementing a fractional-order Q-function; calculating a probabilistic reachable set based on the fractional-order long-time memory model and the verifiable neural network controller; constructing a probabilistic safety barrier function by combining the verifiable neural network controller and the probabilistic reachable set; and establishing a runtime monitoring and recovery mechanism to promptly detect system risks and execute corresponding recovery strategies. This invention provides stringent safety guarantees for SiC power supply systems through probabilistic reachability analysis and the safety barrier function, reducing safety violations and improving system robustness compared to standard reinforcement learning controllers.

Description

Technical Field

[0001] This invention relates to the fields of power electronics and artificial intelligence, and more specifically, to a SiC power supply optimization method and system based on reinforcement learning. Background Technology

[0002] Silicon carbide (SiC) power systems are widely used in new energy power generation, electric vehicles, and aerospace due to their high efficiency, high power density, and excellent temperature characteristics. However, with the increasing complexity of application scenarios, SiC power systems face three key technical challenges: First, the system operating environment is highly uncertain, including random load changes, uncertain input fluctuations, and parameter drift, making it difficult for traditional deterministic verification methods to provide reliable safety guarantees; second, SiC power supplies exhibit a significant long-tail effect (long-term memory characteristics) during long-term operation, which is difficult to accurately describe using conventional control models based on integer-order differential equations; third, in scenarios with high safety requirements such as aerospace and medical equipment, traditional reinforcement learning control systems lack rigorous safety guarantees and verification mechanisms.

[0003] Currently, existing technologies primarily employ robust control, adaptive control, and deep learning-based control strategies to address the aforementioned issues. While robust control methods can handle a certain degree of uncertainty, their ability to handle large disturbances and nonparametric uncertainties is limited. Adaptive control can adjust parameters online, but it struggles to effectively handle the long-term memory characteristics and nonlinear dynamics of the system. Although deep learning-based control strategies possess powerful nonlinear expression capabilities, their "black box" nature makes formal verification difficult, failing to meet the requirements of safety-critical applications.

[0004] Therefore, there is an urgent need for a new method that can simultaneously address uncertainty modeling, long-term memory characteristic characterization, and control verifiability, providing technical support for the safe and efficient operation of SiC power systems in scenarios with high reliability requirements. Summary of the Invention

[0005] This invention provides a SiC power supply optimization method and system based on reinforcement learning, which solves the technical problems of safety, adaptability and verifiability in related technologies.

[0006] This invention provides a reinforcement learning-based SiC power supply optimization method, comprising:

[0007] A fractional-order long-time memory model is established, and the long-tail effect characteristics of the SiC power system are characterized by fractional-order differential operators.

[0008] Based on the fractional-order long-term memory model, a verifiable neural network controller is constructed. The verifiable neural network controller has structural constraints to ensure network verifiability and implements the fractional-order Q function.

[0009] Based on a fractional-order long-time memory model and a verifiable neural network controller, the probabilistic reachable set is calculated, representing the state space that the system may reach under uncertain conditions.

[0010] By combining a verifiable neural network controller and a probabilistic reachable set, a probabilistic safety barrier function is constructed to ensure that the system state satisfies safety constraints with a high probability.

[0011] By integrating fractional-order long-term memory models, verifiable neural network controllers, probabilistic reachable sets, and safety barrier functions, a runtime monitoring and recovery mechanism is established to detect system risks in a timely manner and execute corresponding recovery strategies.

[0012] Furthermore, the steps for establishing the fractional-order long-term memory model include:

[0013] Introducing a fractional differential equation model:

[0014] D α x(t) = f(x(t), u(t));

[0015] Where D α Let f denote the α-order fractional differential operator, where α is the fractional parameter, x(t) is the system state vector, u(t) is the control input vector, and f is the nonlinear state transition function;

[0016] A multi-fractional-order hybrid modeling method is adopted, and different fractional-order parameters α are assigned to different components in the system. i ;

[0017] The optimal fractional-order parameter α is estimated using a real-time system identification algorithm. * .

[0018] Furthermore, the step of constructing a verifiable neural network controller includes:

[0019] Define a neural network architecture with structural constraints, including weight matrix norm constraints and activation function Lipschitz continuity constraints;

[0020] The parameters of the neural network are trained using a convex optimization method to ensure that safety constraints are met.

[0021] Configure the fractional-order Q function:

[0022] Q α (s, a) = E(r) r1 +γD -α Q α (s′,a′));

[0023] Q αRepresents a fractional-order Q-function; s and a represent the current state and action, respectively, and s′ and a′ represent the next state and next action, respectively; r rl For immediate reward, γ is the discount factor, E(·) represents the expected value, and D -α This represents the α-order fractional integral operator.

[0024] Furthermore, the step of calculating the probabilities reachable set includes:

[0025] Define the probability attainable set of the system R1(T) pred ,X0,p th );

[0026] The probabilistic reachable set is approximated using the Monte Carlo particle filter algorithm;

[0027] An adaptive importance sampling technique is used to accelerate the computation of probabilistically reachable sets.

[0028] Furthermore, the step of constructing the probabilistic security barrier function includes:

[0029] Define the probabilistic security barrier function B p (x);

[0030] Constructing a security control law based on a probabilistic security barrier function:

[0031] u(x)=u RL (x)+u safe (x);

[0032] Where u RL (x) represents the output of the verifiable neural network controller calculated using the fractional-order Q-function, u safe (x) is the safety correction term, u(x) is the final control input, and x is the current state of the system;

[0033] The safety threshold is dynamically adjusted using an adaptive method. safe .

[0034] Furthermore, the steps for establishing the runtime monitoring and recovery mechanism include:

[0035] Construct a three-layer security monitoring architecture, including a rapid fault detection layer, a security margin assessment layer, and a fault diagnosis and decision-making layer;

[0036] Implement a multi-level fault recovery strategy and take corresponding measures for faults of different severity.

[0037] Establish online learning and adaptation mechanisms to address the evolution of the system over long-term operation.

[0038] Furthermore, the safety correction term in the safety control law is designed as follows:

[0039]

[0040] Where u safe (x) is a safety correction term, ∈ safe >0 is the safety threshold, K gain (x) is the gain function; Let B be the gradient of the probabilistic security barrier function, representing the security barrier function B. p The directional derivative at point x.

[0041] Furthermore, the multi-level fault recovery strategy is based on a comprehensive risk assessment:

[0042] Risk(x,t)=P(failure|x,t)·Severity(failure);

[0043] Where P(failure|x,t) is the failure probability given the current state x and time t, Severity(failure) is the quantified value of the failure severity, and Risk(x,t) is the comprehensive risk assessment value.

[0044] Furthermore, the convex optimization problem of the neural network parameters is expressed as:

[0045]

[0046] Where θ nn Represents the set of neural network parameters. Let be the performance loss function. Let m represent the i2th security constraint. con The total number of constraints. This indicates that the parameter θ nn Find the minimum value, where st means "constrained by".

[0047] This invention provides a reinforcement learning-based SiC power optimization system for executing the aforementioned reinforcement learning-based SiC power optimization method, comprising:

[0048] The fractional-order model building module is used to build fractional-order long-term memory models;

[0049] Verifiable neural network modules are used to build controllers that meet safety constraints;

[0050] The probabilistic reachability analysis module is used to calculate the probabilistic reachability set of a system under uncertain conditions;

[0051] The safety barrier function module is used to construct probabilistic safety control laws;

[0052] The runtime monitoring module is used to detect system risks in real time and execute recovery strategies.

[0053] The beneficial effects of this invention are as follows: through probabilistic reachability analysis and security barrier functions, it provides strict security guarantees for SiC power systems. Compared with standard reinforcement learning controllers, it reduces security violations and improves system robustness, especially in extreme cases such as load mutations, it can still provide reliable security guarantees.

[0054] The long-time memory model based on fractional-order theory can accurately capture the long-tail effect characteristics of SiC power systems, shorten the dynamic response time of the system, improve control accuracy, maintain high optimization performance under changing load conditions, and outperform traditional methods, achieving high adaptability to uncertain environments.

[0055] By employing probabilistic sampling and a multi-level control structure, the computational resource requirements are reduced, enabling real-time operation on embedded processors and making it suitable for resource-constrained edge devices.

[0056] By using verifiable neural network structures and formal verification techniques, mathematically provable security guarantees are provided for reinforcement learning control systems. It can be proven that the system satisfies security constraints within the defined operating range, providing theoretical support for high-reliability applications. Attached Figure Description

[0057] Figure 1 This is a flowchart of a SiC power supply optimization method based on reinforcement learning in this invention;

[0058] Figure 2 This is a flowchart of step 1 of the present invention;

[0059] Figure 3 This is a flowchart of step 2 of the present invention;

[0060] Figure 4 This is a flowchart of step 3 of the present invention;

[0061] Figure 5 This is a flowchart of step 4 of the present invention;

[0062] Figure 6 This is a flowchart of step 5 of the present invention. Detailed Implementation

[0063] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0064] At least one embodiment of the present invention discloses a SiC power supply optimization method based on reinforcement learning, such as... Figures 1 to 6 As shown, it includes the following steps:

[0065] Step 1: Establish a fractional-order long-time memory model and characterize the long-tail effect of the SiC power system using fractional-order differential operators.

[0066] According to one embodiment of this application, this step utilizes fractional calculus theory to establish a long-time memory model of the SiC power supply to accurately characterize the long-tail effect characteristics of the system. As the foundation of the entire method, the model constructed in this step will provide an accurate dynamic description of the system for subsequent controller design and safety analysis.

[0067] Step 1.1: Introduce a fractional differential equation model;

[0068] It should be noted that applying the fractional differential operator D to the state variable x(t) of the SiC power system... α Establish a mathematical model:

[0069] D α x(t) = f(x(t), u(t));

[0070] Where D α Let f denote the α-order fractional differential operator, where α is the fractional parameter, x(t) is the system state vector, u(t) is the control input vector, and f is the nonlinear state transition function.

[0071] It should be understood that the fractional differential operator D α The definition of is:

[0072]

[0073] Where D α Let denot be an α-order fractional differential operator, where α is a fractional parameter, x(t) is the system state vector, n is the smallest integer greater than α, Γ is the gamma function, a is the lower limit of integration, τ is the integration variable, t is the current time point, dt is the time derivative, and dτ is the derivative of the integration variable.

[0074] This definition, also known as the Riemann-Liouville fractional derivative, effectively captures the long-term memory characteristics of a system.

[0075] Step 1.2: Use multi-fractional mixed modeling;

[0076] Furthermore, a multi-fractional-order hybrid model is used for accurate modeling of the different dynamic characteristics of various components in the SiC power system:

[0077]

[0078] Where α i Let be the fractional-order parameter of the i-th subsystem. Represents α i fractional differential operator of order x i (t) represents the state variable of the i-th subsystem, f i Let x(t) be the state transition function of the i-th subsystem, x(t) be the system state vector, and u(t) be the control input vector.

[0079] By assigning different fractional-order parameters to different components, a precise description of the complex dynamic characteristics of the system can be achieved.

[0080] Step 1.3: Estimate the optimal fractional-order parameters using a real-time system identification algorithm;

[0081] The calculation formula is:

[0082] α * =argmin α ||y real -y model (α)|| 2 ;

[0083] Where α * y represents the optimal fractional-order parameter set, used to most accurately describe the dynamic characteristics of the system. real This is the actual system output, used for system identification, y model (α) represents the predicted output of the fractional-order model under parameter α, ||·|| denotes the Euclidean norm, used to measure the length of a vector or the size of a matrix, and argmin α Let α be the parameter that minimizes the objective function.

[0084] It should be noted that this optimization problem is solved using numerical methods such as gradient descent or pattern search, achieving adaptive adjustment of fractional-order parameters.

[0085] Step 2: Based on the fractional-order long-term memory model, construct a verifiable neural network controller. The verifiable neural network controller has structural constraints to ensure network verifiability and implements the fractional-order Q function.

[0086] According to another embodiment of this application, this step, based on the fractional-order long-term memory model established in step 1, utilizes a neural network with special structural constraints to implement a reinforcement learning strategy, ensuring the verifiability and security of the controller. Through organic integration with the fractional-order model, the controller can effectively handle the long-term memory characteristics of the system, improving control accuracy.

[0087] Step 2.1, define the neural network architecture with structural constraints;

[0088] Formal verification conditions are met:

[0089] i1 = 0, 1, ..., L-1;

[0090] in and These represent the outputs of the neural network at layer i1+1 and layer i1, respectively. and Let σ represent the weight matrix and bias vector of the i1th layer, respectively. act Here, is the activation function, L is the number of layers in the neural network, and i1 is the layer index.

[0091] It should be noted that, to ensure the verifiability of the neural network, the following constraints are imposed on the weights and activation function:

[0092] The weight matrix satisfies the norm constraint:

[0093]

[0094] in Let be the upper bound of the norm of the weight matrix of layer i1. This represents the infinite norm of the weight matrix.

[0095] Activation functions with bounded outputs:

[0096] |σ act (x)-σ act (y)|≤L σ |xy|;

[0097] Where L σ Let σ be the Lipschitz constant of the activation function, and let x and y be two different input variables of the activation function. act This is the activation function.

[0098] Step 2.2: Train the neural network parameters using a convex optimization method;

[0099] Ensuring that it meets safety constraints is expressed as:

[0100]

[0101] Where θ nn Represents the set of neural network parameters. Let be the performance loss function. Let m represent the i2th security constraint. con The total number of constraints. This indicates that the parameter θ nn Find the minimum value, where st means "constrained by".

[0102] It should be understood that the performance loss function is usually chosen as the negative of the reinforcement learning objective function, such as the negative of the expected cumulative reward.

[0103] It is evident that safety constraints Includes two categories:

[0104] Parameter constraints: limiting the size and structure of network parameters;

[0105] Behavioral constraints: Ensure that the system's behavior under neural network control meets safety standards.

[0106] Step 2.3: Configure the fractional-order Q-function;

[0107] Therefore, for the fractional-order system model in step 1, this application further configures the corresponding fractional-order Q-function:

[0108] Q α (s, a) = E(r) r1 +γD -α Q α (s′,a′));

[0109] Q α Represents a fractional-order Q-function; s and a represent the current state and action, respectively, and s′ and a′ represent the next state and next action, respectively; r rl For immediate reward, γ is the discount factor, E(·) represents the expected value, and D -α The α-order fractional integral operator is defined as follows:

[0110]

[0111] Where D -α f(t) represents the α-order fractional integral of f(t), where α is the fractional-order parameter, Γ is the gamma function, τ is the integration variable, f(τ) is the integrand, and t is the upper limit time variable.

[0112] When a neural network performs an approximation, the training objective of the network is to minimize the fractional-order Bellman error.

[0113] Step 3: Based on the fractional-order long-term memory model and the verifiable neural network controller, calculate the probabilistic reachable set, which represents the state space that the system may reach under uncertain conditions;

[0114] According to another embodiment of this application, this step utilizes the fractional-order system model established in step 1 and the verifiable neural network controller designed in step 2 to calculate the probabilistic reachable set of the system under uncertain conditions. This step combines the system model and control strategy with uncertainty analysis, providing a theoretical basis for subsequent safety control.

[0115] Step 3.1, define the mathematical model of probabilistic reachable sets;

[0116] In this embodiment, the probabilistic attainable set B1(T) of the SiC power system is first defined. pred ,X0,p th ):

[0117] R1(T pred X0, p th )={x∈R n |P(x(T pred )=x|x(0)∈X0)≥p th};

[0118] Where R1(T) pred X0, p th ) indicates that the system is at time T pred With at least probability p th The set of possible states, T pred For the prediction time range, X0 is the initial state set, p th R is the probability threshold. n Let x represent the current state vector in n-dimensional real space, and P represent the probability. pred ) indicates that the system is at time T pred The state of the system is x(0), which represents the state of the system at the initial moment.

[0119] Step 3.2: Approximate the probabilistic reachable set using the Monte Carlo particle filter algorithm;

[0120] Specifically, it includes:

[0121] Random sampling N par Each particle represents a possible initial state of the system:

[0122]

[0123] in Let N be the initial state of the i3rd particle. par denoted as the total number of particles, ~ indicates “follows a distribution”, and X0 is the initial state set.

[0124] For each particle, the system trajectory is simulated forward using the fractional differential equations from the fractional long-term memory model established in step 1:

[0125]

[0126] Where Φ represents the system transfer function, u 0：t-1 Given the control input sequence from time 0 to t-1, w 0：t-1 Given a random perturbation sequence from time 0 to t-1, Let i be the initial state of the i3rd particle. Let N be the state of the i-th particle at time t. par The total number of particles is represented by the tilde (~), which indicates that the particles follow a distribution.

[0127] Based on the particle's time T pred The distribution of the probability can be estimated to reach a set:

[0128] R2(T pred X0, p th );

[0129] Where R2(T) pred X0, p th T is an estimate of the probabilistically reachable set. pred For the prediction time range, X0 is the initial state set, p th This is the probability threshold.

[0130] It is important to understand that for random factors in a system, nonparametric probability distribution models are used to describe their characteristics, especially for uncertain factors that exhibit long-tailed distribution characteristics, heavy-tailed distribution models such as the α-stable distribution are used.

[0131] Step 3.3: Adaptive importance sampling technique is used to accelerate the calculation of probabilistically reachable sets;

[0132] The calculation formula is:

[0133]

[0134] Where R2(T) pred X0, p th T is an estimate of the probabilistically reachable set. pred For the prediction time range, X0 is the initial state set, p th As a probability threshold, N represents the importance weight of the i-th 3rd particle. par Let be the total number of particles, and 1(·) be the indicator function (value 1 when the condition in parentheses is true, otherwise 0). For the i3rd particle at time T pred state, ≈ represents summation over all particles, ≈ represents approximate equality, and x is the current state vector.

[0135] Calculated through importance sampling:

[0136]

[0137] in Let P be the importance weight of the i-th 3rd particle, P be the true transition probability of the system, and Q be the proposal distribution. Indicates starting from the initial state Transition to state The true probability, This indicates that the sample was obtained from the proposed distribution. arrive The probability of.

[0138] By iteratively optimizing the suggested distribution Q to bring it closer to the important region of the system's true transition probability, computational efficiency can be significantly improved.

[0139] Step 4: Combine the verifiable neural network controller and the probabilistic reachable set to construct a probabilistic safety barrier function to ensure that the system state satisfies the safety constraints with a high probability.

[0140] According to another embodiment of this application, this step constructs a probabilistic safety barrier function based on the probabilistic reachability set calculated in step 3, and combines it with the neural network controller in step 2 to ensure that the system satisfies the safety constraints with a high probability. This step achieves an organic unity between reinforcement learning control and safety assurance.

[0141] Step 4.1, define the probabilistic security barrier function;

[0142] The calculation formula is:

[0143] B p (x)=d(x,R n \R1(T pred X0, p th ));

[0144] Among them B p (x) is the probabilistic safety barrier function, d represents the distance function, x is the current state vector, and R represents the complement of the probabilistic reachable set. n Let R1(T) represent an n-dimensional real space. pred X0, p th ) indicates that the system is at time T pred With at least probability p th The set of possible states, T pred For the prediction time range, X0 is the initial state set, p th This is the probability threshold.

[0145] It should be understood that:

[0146] When B p When (x) > 0, it means that state x has a probability of at least p. th Stay safe;

[0147] When B p When (x) = 0, it means that state x is located on the probabilistic safety boundary;

[0148] When B p When (x) < 0, it means that the safety probability of state x is lower than the threshold p.th .

[0149] Step 4.2: Construct a security control law based on the probabilistic security barrier function;

[0150] The calculation formula is:

[0151] u(x)=u RL (x)+u safe (x);

[0152] Where u RL (x) represents the output of the verifiable neural network controller calculated using the fractional-order Q-function, u safe (x) is the safety correction term, u(x) is the final control input, and x is the current state vector.

[0153] It is important to note that the safety calibration item u safe (x) is designed as follows:

[0154]

[0155] Where u safe (x) is a safety correction term, ∈ safe >0 is the safety threshold, K gain (x) is the gain function; Let B be the gradient of the probabilistic security barrier function, representing the security barrier function B. p The directional derivative at the current state x.

[0156] This control law ensures that the system state always evolves in a direction that increases the probability of safety.

[0157] Step 4.3: Dynamically adjust the safety threshold using an adaptive method;

[0158] To adapt to different operating environments and control objectives, the calculation formula is as follows:

[0159] ∈ safe (t)=∈0+Δ∈(t);

[0160] Where ∈0 is the basic safety threshold, Δ∈(t) is the dynamic adjustment term, ∈ safe (t) represents the safety threshold at time t.

[0161] It should be understood that the dynamic adjustment term Δ∈(t) is calculated based on the current system state, historical trajectory, and safety statistics: Δ∈(t)=f ∈ (x(t), x(τ)) τ＜t B p (x(τ)) τ＜t );

[0162] Where Δ∈(t) is the dynamic adjustment term, f∈ Let x(t) be an adaptive function, x(t) be the system state at time t, and x(τ) be the system state at time t. τ＜t B represents the set of all system states before time t. p (x(τ)) τ＜t This represents the set of safety barrier function values ​​corresponding to all states before time t.

[0163] The security threshold is dynamically adjusted based on historical security status. When the system is in a high-risk area, the security margin is increased, and when the system is in a low-risk area, the security margin is appropriately reduced to improve performance.

[0164] Step 5: Integrate fractional-order long-term memory model, verifiable neural network controller, probabilistic reachable set and safety barrier function to establish runtime monitoring and recovery mechanism, timely detect system risks and execute corresponding recovery strategies;

[0165] According to another embodiment of this application, this step integrates the results of the preceding steps, particularly utilizing the system model of step 1, the controller of step 2, the probabilistic reachability set analysis of step 3, and the safety barrier function of step 4, to establish a complete runtime safety monitoring system. This step achieves safety assurance throughout the system's entire lifecycle, promptly detects potential risks and executes recovery strategies, ensuring the long-term stable operation of the SiC power system.

[0166] Step 5.1: Construct a three-tier security monitoring architecture;

[0167] Specifically, it includes:

[0168] The bottom layer, the rapid fault detection layer:

[0169] Residual analysis based on the fractional long-term memory model in step 1:

[0170] r res (t)=y out (t)-y(t);

[0171] Real-time parameter estimation and change detection:

[0172] θ est (t)=argmin θ ||y out (t)-f(x(t), u(t), θ)|| 2 ;

[0173] Threshold adaptive fault trigger:

[0174] alarm=1(||r res (t)||>τ th (t));

[0175] Where r res(t) is the residual vector at time t, y out y(t) represents the actual output of the system, y(t) represents the model's predicted output, and θ represents the actual output of the system. est (t) represents the system parameters estimated at time t, argmin θ Let θ represent the parameter that minimizes the objective function, f(x(t), u(t), θ) represent the output of the system model with parameter θ under state x(t) and input u(t), ||·|| represents the norm, alarm is the alarm signal, 1(·) is the indicator function, and τ th (t) represents the alarm threshold at time t.

[0176] Middle layer, safety margin assessment layer:

[0177] Online probability reachable set approximation calculation;

[0178] Probabilistic security barrier function value monitoring B p (x(t));

[0179] Security Trend Analysis:

[0180]

[0181] Among them B p (x(t)) is the safety barrier function value of the system state at time t. It represents the derivative of the safety obstacle function value with respect to time, reflecting the changing trend of the safety status.

[0182] Senior management, fault diagnosis and decision-making level:

[0183] Fault classification and location;

[0184] Multimodal data fusion analysis;

[0185] Generation and execution of security recovery strategies.

[0186] It is important to emphasize that a closed-loop information flow is formed between the various monitoring layers. The rapid detection results at the bottom layer trigger in-depth analysis at the middle layer, and the security assessment results at the middle layer guide the decision-making at the top level.

[0187] Step 5.2: Implement a multi-level recovery strategy;

[0188] For faults of varying severity, including:

[0189] Minor fault, enhanced control mode:

[0190] Adaptive adjustment of control parameters;

[0191] Increase safety margin:

[0192] ∈′ safe (t)=β safe·∈ safe (y), β safe >1;

[0193] where ∈′ safe (t) is the adjusted safety threshold, ∈ safe (t) is the original safety threshold, β safe A coefficient is added to the safety margin, where t is the current time.

[0194] Without interrupting the normal operation of the system.

[0195] Moderate fault, degraded operation mode:

[0196] Switch to a conservative control strategy;

[0197] Limit the scope of system operation;

[0198] Lower performance target requirements;

[0199] Critical malfunction, emergency safety mode:

[0200] Execute the predefined safety shutdown procedure;

[0201] Isolate faulty components;

[0202] Record complete fault context information;

[0203] It is worth noting that the choice of recovery strategy is based on a comprehensive risk assessment:

[0204] Risk(x,t)=P(failure|x,t)·Severity(failure);

[0205] Where P(failure|x,t) is the failure probability given the current state x and time t, Severity(failure) is the quantified value of the failure severity, and Risk(x,t) is the comprehensive risk assessment value.

[0206] Step 5.3: Establish an online learning and adaptation mechanism;

[0207] Furthermore, to address system evolution during long-term operation of SiC power supplies, this application implements an online learning mechanism:

[0208] Incremental model update:

[0209] The system model is continuously updated based on new data.

[0210]

[0211] Where θ t+1 For the updated model parameters, θ t For the current model parameters, α lrFor learning rate, The loss function L with respect to parameter θ at the current parameter θ t and dataset D t The gradient below;

[0212] Adaptive learning rate adjustment:

[0213] α lr =f α (t,D t ,L(θ t D t ));

[0214] Where α lr f is the learning rate. α Let D be the learning rate adaptive function, t be the current time, and D be the learning rate adaptive function. t For the dataset at time t, L(θ) t D t () represents the loss function value given the current parameters and data;

[0215] Maintain the historical model version library and support rollback.

[0216] Experience base construction and utilization:

[0217] Record historical failure scenarios and recovery strategies;

[0218] Extract fault features and patterns;

[0219] Historical successes of similarity-based search applications.

[0220] Regular security verification:

[0221] It should be noted that the system performs formal security verification periodically:

[0222] Verify the security features of the updated model;

[0223] Perform sensitivity analysis on key safety parameters;

[0224] Generate a verification report to ensure system integrity.

[0225] Through the above runtime monitoring and recovery mechanisms, this application achieves high reliability and security of the SiC power system, while maintaining the system's adaptability and long-term stable operation capability.

[0226] A reinforcement learning-based SiC power optimization system, used to execute the aforementioned reinforcement learning-based SiC power optimization method, includes:

[0227] The fractional-order model building module is used to build fractional-order long-term memory models;

[0228] Verifiable neural network modules are used to build controllers that meet safety constraints;

[0229] The probabilistic reachability analysis module is used to calculate the probabilistic reachability set of a system under uncertain conditions;

[0230] The safety barrier function module is used to construct probabilistic safety control laws;

[0231] The runtime monitoring module is used to detect system risks in real time and execute recovery strategies.

[0232] Here, the present invention provides an implementation example:

[0233] This paper demonstrates the implementation process and technical effects of the proposed method in a practical application through an example of an electric vehicle SiC power management system.

[0234] According to embodiments of this application, in the fast charging scenario of electric vehicles, the SiC power system needs to simultaneously meet the requirements of high-efficiency charging and device safety. This scenario has the following characteristics:

[0235] The charging load is random, and the battery state affects the charging characteristics;

[0236] Fluctuations in grid input voltage and interference from multiple users accessing the charging station;

[0237] SiC device parameters drift with temperature and usage time;

[0238] The charging process for electric vehicles requires strict safety measures to avoid risks such as overcharging and overheating.

[0239] In this example, a fractional-order long-time memory model is first established for the SiC power system. Analysis of actual measurement data reveals significant differences between the dynamic characteristics of SiC MOSFETs and traditional integer-order models, particularly in terms of long-time memory characteristics.

[0240] Through system identification, the optimal fractional-order parameter is determined to be:

[0241] SiCMOSFET temperature dynamics α temp =1.37;

[0242] Output capacitor voltage response α cap =0.82;

[0243] Inductor current dynamic α ind =1.05.

[0244] The dynamic response equation of the fractional-order model is:

[0245] D 1.37 T dev(t)=-0.53T dev (t)+0.78P loss (t)-0.12T amb (t);

[0246] D 0.82 V c (t)=-1.26V c (t)+2.47I L (t)-0.93I load (t);

[0247] D 1.05 I L (t)=-0.68I L (t)+1.32V in (t)-1.17V c (t);

[0248] Where D 1.37 D 0.82 D 1.05 Let T represent fractional differential operators of order 1.37, 0.82, and 1.05, respectively. dev (t) represents the device temperature, P loss (t) represents power loss, T amb (t) represents the ambient temperature, V c (t) represents the output capacitor voltage, I L (t) represents the inductor current, I load (t) represents the load current, V in (t) represents the input voltage.

[0249] Based on extensive experimental data, the fractional-order model reduced the prediction error of the system's dynamic response by 76.3% compared to the integer-order model, and performed particularly well in capturing the system's long-term memory effect.

[0250] According to embodiments of this application, a verifiable neural network controller comprising a 4-layer structure is constructed, wherein:

[0251] Input layer: 7 nodes, corresponding to system state variables;

[0252] Hidden layer 1: 32 nodes, using the ReLU activation function;

[0253] Hidden layer 2: 16 nodes, using the tanh activation function;

[0254] Output layer: 3 nodes, corresponding to control actions (switching frequency, duty cycle, current limit).

[0255] To ensure network verifiability, the following constraints are imposed on the weight matrix:

[0256] First-level weight matrix norm constraint: ||W1|| ∞ ≤1.5;

[0257] Second-level weight matrix norm constraint: ||W2|| ∞ ≤1.2;

[0258] Third-level weight matrix norm constraint: ||W3|| ∞ ≤0.8;

[0259] Where ||W1|| ∞ ||W2|| ∞ ||W3|| ∞ W1, W2, and W3 represent the infinite norms of the weight matrices of the first, second, and third layers, respectively.

[0260] The network training uses a fractional-order Q-function, where the fractional order α = 0.85 and the discount factor γ = 0.92.

[0261] Safety constraints are introduced during training to ensure that the control output meets the following requirements:

[0262] Switching frequency constraint: f sw ∈(20kHz,150kHz);

[0263] Duty cycle constraint: D duty ∈(0.1, 0.9);

[0264] Current limiting constraint: I lim ∈(0, I max );

[0265] Where f sw D represents the switching frequency. duty Indicates duty cycle, I lim Indicates the current limit value, I max This is the maximum safe current for the equipment.

[0266] In this example, to address the uncertainties during the charging process, the system probability reachability set is calculated, and a safety control strategy is constructed based on this set. For the battery charging state, three main uncertainties are considered:

[0267] Random variation of battery internal resistance with state of charge: The change in battery internal resistance follows a normal distribution with a mean of 0 and a variance of 0.12 ohm squares.

[0268] Ambient temperature fluctuation: The ambient temperature follows a normal distribution with a mean of 25 degrees Celsius and a variance of 5 degrees Celsius squared;

[0269] Grid voltage fluctuations: The grid voltage follows a normal distribution with a mean of 220 volts and a variance of 10 volts squared.

[0270] A Monte Carlo simulation with 10,000 particles was used to calculate the probabilistic reachable set within a 30-minute prediction timeframe, with a probability threshold of 0.95. Based on the calculation results, a probabilistic safety barrier function was constructed.

[0271] B p (x)=minB temp (x),B volt (x),B curr (x);

[0272] Among them B p (x) represents the probabilistic safety barrier function, x represents the system state, min represents the minimum value operation, and B temp (x) represents the temperature safety barrier subfunction, B volt (x) represents the voltage safety barrier subfunction, B curr (x) represents the current safety barrier sub-function.

[0273] The safety control law is implemented as follows:

[0274]

[0275] Where u(x) represents the total control output, u RL (x) represents the output of the reinforcement learning controller, K gain (x) represents the gain function. The gradient of the safety barrier function is represented by 'max', which indicates the operation of finding the maximum value. safe B represents the safety threshold. p (x) represents the probabilistic safety barrier function, and X represents the system state.

[0276] Where the gain function K gain (x) adaptively adjusts with system state, and the safety threshold ∈ safe The initial value is 0.15, and it is dynamically adjusted based on the system's operating history.

[0277] In this example, a three-tier security monitoring architecture is implemented:

[0278] The underlying fast fault detection uses a 10ms sampling period and calculates the residual by comparing the model prediction value with the actual measurement value.

[0279] The mid-level safety margin assessment uses a 100ms period to calculate the probabilistic safety barrier function value;

[0280] High-rise fault diagnosis uses a 1-second cycle to comprehensively analyze the system's safety status.

[0281] Three fault scenarios were simulated in the test: sudden fluctuation of input voltage (±30V); load short circuit (current surge of 200%); abnormal temperature of SiC device (exceeding the threshold of 25℃).

[0282] The system adopts corresponding recovery strategies:

[0283] Voltage fluctuations: Adjust control parameters to increase stability margin;

[0284] Load short circuit: Immediately cut off the output and activate the protection circuit;

[0285] Abnormal temperature: Reduce power output and increase cooling.

[0286] The application effect data of this embodiment in the SiC charging system for electric vehicles are as follows:

[0287] The results of the safety performance comparison test are shown in Table 1:

[0288] Table 1: Safety Performance Comparison Test Results

[0289]

[0290] The results of the control performance comparison test are shown in Table 2:

[0291] Table 2: Control Performance Comparison Test Results

[0292] Performance indicators Traditional methods This method Percentage Improvement Dynamic response time 12.5ms 4.8ms 61.6％ Voltage stability ±2.8％ ±0.42％ 85.0％ Efficiency fluctuation range 4.9％ 0.83％ 83.1％ Load adaptability retention rate 62.3％ 93.1％ 49.4％ Computing resource requirements benchmark Reduced by 47.2% 47.2％

[0293] The above data demonstrates that this implementation significantly improves control accuracy and system adaptability while ensuring system safety, and reduces computational resource requirements, making it suitable for resource-constrained embedded applications. Particularly in the safety-critical scenario of fast charging for electric vehicles, this method provides nearly 99.8% safety assurance while reducing dynamic response time by 61.6%, offering an efficient and reliable technical solution for electric vehicle charging infrastructure.

[0294] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A SiC power supply optimization method based on reinforcement learning, characterized in that, include: A fractional-order long-time memory model is established, and the long-tail effect characteristics of the SiC power system are characterized by fractional-order differential operators. Based on a fractional-order long-term memory model, a verifiable neural network controller is constructed. The verifiable neural network controller has structural constraints to ensure network verifiability. The structural constraints include weight matrix norm constraints and activation function Lipschitz continuity constraints, and implements a fractional-order Q function. Based on a fractional-order long-time memory model and a verifiable neural network controller, the probabilistic reachable set is calculated, representing the state space that the system may reach under uncertain conditions. By combining a verifiable neural network controller and a probabilistic reachable set, a probabilistic safety barrier function is constructed. The calculation method is as follows: ;in For probabilistic security barrier functions, Represents the distance function. Let this be the current state vector. express 3D real space, Indicates the system in time With at least probability The set of possible states, To predict the time range, For the initial set of states, This is the probability threshold; when Time indicates state With at least probability Stay safe; when Time indicates state Located on the probabilistic safety boundary; when Time indicates state The security probability is below the threshold Ensure that the system state meets safety constraints with a high probability. By integrating fractional-order long-term memory models, verifiable neural network controllers, probabilistic reachable sets, and probabilistic safety barrier functions, a runtime monitoring and recovery mechanism is established to detect system risks in a timely manner and execute corresponding recovery strategies.

2. The SiC power supply optimization method based on reinforcement learning according to claim 1, characterized in that, The steps for establishing the fractional-order long-term memory model include: Introducing a fractional differential equation model: ； in express fractional differential operators of order, For fractional-order parameters, Let be the system state vector. To control the input vector, It is a nonlinear state transition function; A multi-fractional-order hybrid modeling method is adopted, which assigns different fractional-order parameters to different components in the system. ; Estimate the optimal fractional-order parameters using a real-time system identification algorithm. .

3. The SiC power supply optimization method based on reinforcement learning according to claim 1, characterized in that, The steps for constructing a verifiable neural network controller include: Define a neural network architecture with structural constraints, including weight matrix norm constraints and activation function Lipschitz continuity constraints; The parameters of the neural network are trained using a convex optimization method to ensure that safety constraints are met. Configure the fractional-order Q function: ； in Represents a fractional Q-function; and These are the current state and the action, respectively. and These refer to the next state and the next action, respectively. For instant rewards, As a discount factor, Indicates the expected value. express Fractional integral operator of order.

4. The SiC power supply optimization method based on reinforcement learning according to claim 1, characterized in that, The steps for calculating the probabilistically reachable set include: Define the probabilistic reachable set of the system ; The probabilistic reachable set is approximated using the Monte Carlo particle filter algorithm; An adaptive importance sampling technique is used to accelerate the calculation of probabilistically reachable sets.

5. The SiC power supply optimization method based on reinforcement learning according to claim 1, characterized in that, The steps for constructing the probabilistic security barrier function include: Define the probabilistic security barrier function ; Based on the probabilistic safety barrier function, the output of the verifiable neural network controller, calculated using the fractional-order Q-function, is... With safety calibration items Combine and calculate the final control input : in To verify the output calculated by the neural network controller based on the fractional-order Q-function, For safety calibration items, This represents the current state of the system. Adaptive method is used to dynamically adjust the safety threshold. .

6. The SiC power supply optimization method based on reinforcement learning according to claim 1, characterized in that, The steps for establishing a runtime monitoring and recovery mechanism include: A three-tiered security monitoring architecture is constructed, comprising a rapid fault detection layer, a security margin assessment layer, and a fault diagnosis and decision-making layer. The security margin assessment layer utilizes online probabilistic reachable sets for approximate calculation and probabilistic safety barrier function values. Monitoring and security trend analysis To assess the margin of the system's current state relative to the safety boundary; Implement a multi-level fault recovery strategy and take corresponding measures for faults of different severity. Establish online learning and adaptation mechanisms to address the evolution of the system over long-term operation.

7. The SiC power supply optimization method based on reinforcement learning according to claim 5, characterized in that, The safety correction item is designed as follows: ； in For safety calibration items, As a safety threshold, It is the gain function; Let be the gradient of the probabilistic security barrier function, and let represent the security barrier function. At point The directional derivative at that point.

8. The SiC power supply optimization method based on reinforcement learning according to claim 6, characterized in that, The multi-level fault recovery strategy is based on a comprehensive risk assessment: ； in Given the current state and time The probability of failure, A quantification of the severity of the fault. This is the comprehensive risk assessment value.

9. The SiC power supply optimization method based on reinforcement learning according to claim 3, characterized in that, The convex optimization problem of neural network parameters is represented as: ； in Represents the set of parameters of a neural network. Let be the performance loss function. Indicates the first One safety constraint, The total number of constraints. Indicates the parameter Find the minimum value. It means "restricted by".

10. A SiC power supply optimization system based on reinforcement learning, characterized in that, A method for implementing a reinforcement learning-based SiC power supply optimization method according to any one of claims 1-9 includes: The fractional-order model building module is used to build fractional-order long-term memory models; Verifiable neural network modules are used to build controllers that meet safety constraints; The probabilistic reachability analysis module is used to calculate the probabilistic reachability set of a system under uncertain conditions; The safety barrier function module is used to construct probabilistic safety control laws; The runtime monitoring module is used to detect system risks in real time and execute recovery strategies.

Images