A dynamic security regulation method for a converter of a light storage and charging system

Abstract

The application discloses a kind of dynamic security regulation and control methods for the converter of light storage and filling system, it is related to light storage and filling system data modulation field.The application is based on Hamilton energy evolution to construct converter state space model, the voltage and current state are unified energy characterization, the loss change caused by temperature and load is identified in real time by introducing physical perception neural Hamilton operator, on this basis, the modulation frequency and modulation ratio are globally optimized using Riemann manifold, the operation mode adjustment is completed by combining double-agent decision mechanism, and the model parameters are continuously updated by error feedback, the application can reduce bus voltage fluctuation and control error under photovoltaic fluctuation and load impact, reduce switching and conduction loss, avoid parameter mismatch and local optimum problem, make the system run more stably, while maintaining high efficiency and having continuous self-adaptive ability.

Description

Technical Field

[0001] This invention relates to the field of data modulation in photovoltaic energy storage and charging systems, specifically a method for dynamic safety control of converters in photovoltaic energy storage and charging systems. Background Technology

[0002] The control methods of energy storage converters in existing photovoltaic-storage-charging systems are mainly based on classical power electronic control theory. They typically adopt a dual closed-loop control structure consisting of an outer voltage loop and an inner current loop, and achieve decoupled regulation of active and reactive power through dq coordinate transformation. System modeling often uses small-signal linearization methods or fixed-parameter state-space models. Loss models are determined through offline experimental calibration or table lookup. Controller parameters remain unchanged after being set during the commissioning phase. Modulation strategies often use fixed switching frequencies or switching within a limited range. During operation, mode switching is performed based on bus voltage or current thresholds. Some schemes introduce model predictive control or adaptive adjustment methods to improve dynamic performance. The overall control system relies on preset models and empirical parameters to complete operation.

[0003] Existing technologies have revealed multiple specific problems in actual operation. In scenarios with rapid fluctuations in photovoltaic output, the power change can reach 30% to 60% of the rated value in a short period of time. Small-signal models based on fixed operating points cannot accurately reflect the nonlinear characteristics of the system, resulting in a deviation of more than 5% between the predicted current and the actual current. This leads to over-regulation of the controller output, causing an increase in the amplitude of bus voltage oscillation. During the connection of high-power charging loads, the sudden increase in current will cause a significant increase in the conduction and switching losses of power devices. Traditional loss models still use room temperature or calibration parameters, which cannot reflect the change in equivalent resistance caused by temperature rise. In actual operation, when the IGBT junction temperature rises by more than 10°C, the conduction loss increases by about 8% to 12%. Model mismatch further amplifies the control error, and the fixed switching frequency... The rate-controlled strategy maintains high-frequency operation under high load, leading to increased additional switching losses and a system efficiency decrease of more than 2%. Under low load, the excessively low frequency causes increased current ripple and higher output voltage harmonic content. Traditional parameter optimization methods rely on offline tuning or simple gradient search, which are prone to getting stuck in local advantages under multivariable coupling conditions. The modulation parameters are difficult to adapt to real-time changes in operating conditions. The operation mode switching usually relies on a single voltage threshold judgment, resulting in mode switching when the load fluctuates frequently. The switching frequency can reach hundreds of milliseconds, leading to unstable control commands. Some systems have not established an online learning mechanism, and factors such as increased on-state voltage drop after device aging and inductor parameter drift cannot be absorbed by the model. After a period of operation, the control accuracy decreases significantly, and the steady-state deviation of the bus voltage expands to more than 3% of the rated value. Summary of the Invention

[0004] This invention proposes a dynamic safety control method for converters in photovoltaic energy storage and charging systems. This invention aims to solve the problems in existing energy storage converter control methods, such as decreased model accuracy under large-scale operating conditions, inability to update loss parameters in real time with temperature and load changes, difficulty in achieving global optimization of modulation parameters, unstable operation mode switching, and lack of long-term adaptive capability of the system.

[0005] One method for dynamic safety control of a converter in a photovoltaic-storage-charging system includes the following steps: S1. Collect electromagnetic state parameters and environmental characteristic data of the energy storage converter in real time through sensors, and construct a Hamiltonian state-space model based on the electromagnetic state parameters and environmental characteristic data; Specifically, the above steps map the electromagnetic state quantities such as voltage and current of the energy storage converter into a unified state description in the energy domain. A Hamiltonian structure is used to characterize the storage, exchange, and evolution of energy in the system. Inductance corresponds to magnetic field energy, and capacitance corresponds to electric field energy. The position of the state vector in phase space directly reflects the instantaneous energy distribution of the system. Antisymmetric structures are used to express the lossless flow relationship of energy between different energy storage units under ideal conditions. Dissipative structures are used to characterize energy attenuation caused by actual resistance, device losses, and other factors. Traditional methods often use dq coordinate systems or small-signal linear models to describe the dynamic characteristics of the converter. These models rely on linearization at the operating point, and parameters become mismatched with changes in operating conditions, making it difficult to reflect the real behavior under a wide range of operating conditions. This step, through a unified energy modeling approach, directly constructs the state space from the physical essence. The model structure has clear physical meaning, maintains consistency over a wide range of operating conditions, and is suitable for strong disturbances and nonlinear scenarios. The state variables correspond one-to-one with the actual electromagnetic quantities, facilitating direct subsequent use.

[0006] S2. Input the collected electromagnetic state parameters into the physical sensing neural Hamiltonian operator for nonlinear perturbation identification, and correct the dissipation matrix in the constructed port-controlled Hamiltonian state-space model based on the identification results. Specifically, the above steps address the uncertainty of system losses. The key lies in using the physical sensing neural Hamiltonian operator to identify nonlinear disturbances online. Temperature changes, photovoltaic power output fluctuations, and load shocks all cause changes in device conduction and switching losses, which fixed dissipation models cannot track. Traditional methods typically update loss parameters using lookup tables or empirical corrections, resulting in low update frequency and limited adaptability. By inputting electromagnetic state variables and environmental features into a neural network, the network outputs an energy dissipation potential scalar. Gradient calculations extract loss change trends across each state dimension, forming a dynamic correction vector that directly applies to the dissipation structure, achieving real-time updates at the model level. This approach does not rely on explicit loss mechanism modeling, can capture complex coupling effects, maintains consistency with physical constraints, and avoids situations where pure data-driven methods violate energy conservation.

[0007] S3. Construct the Riemannian manifold optimization space based on the corrected dissipation matrix, and obtain the optimal control vector of the converter by searching the energy gradient descent direction in the Riemannian manifold optimization space; Specifically, the above steps construct a geometric optimization framework for the control parameter optimization problem. The core idea is to map the modulation frequency and modulation ratio from Euclidean space to the Riemannian manifold space defined by loss characteristics. The metric tensor is determined by the sensitivity of the dissipative structure to the control parameters, and the curvature distribution reflects the energy loss variation trend in different parameter regions. High curvature regions correspond to areas of drastic loss changes, while low curvature regions correspond to stable operating regions. Traditional optimization methods often use fixed-step gradient descent or empirical parameter tuning, which are prone to getting trapped in local optima in non-convex loss distributions, making it difficult to balance efficiency and stability. This step, through a geodesic search mechanism, allows the control variables to evolve along the shortest path in the manifold. The path direction automatically avoids high-loss regions, and the step size is adaptively adjusted in different curvature regions, ensuring stable convergence of the search process and possessing global optimization capabilities. This enables the acquisition of better modulation strategies under complex operating conditions.

[0008] S4. Input the optimal control vector into the heterogeneous dual-agent game model, calculate the active power reference value and reactive power reference value, and combine the information entropy characteristic value of the collected electromagnetic state parameters to determine the operating mode of the converter. Specifically, the above steps construct a dual-agent decision-making mechanism around the problem of conflicting operational objectives. During converter operation, there is a dynamic trade-off between efficiency and voltage stability; load shocks or photovoltaic fluctuations can alter system priority requirements. Traditional methods typically employ fixed-weight or simple threshold switching strategies, lacking the ability to characterize the dynamic complexity of the system, and prone to frequent switching or response lag. This step reconstructs the phase space of the bus voltage sequence and calculates approximate entropy. The entropy value is used to characterize the complexity of voltage fluctuations. When complexity is low, an efficiency-first strategy is maintained; when complexity increases, the voltage support weight is increased. The two agents output strategies from the perspectives of energy loss and voltage deviation, respectively. The weight allocation is determined by the real-time system state, and the decision-making process possesses continuity and adaptability, avoiding control abrupt changes caused by hard switching.

[0009] S5. Generate pulse width modulation commands based on active power reference values, reactive power reference values, and operating modes, and update the weight parameters of the physical sensing neural Hamiltonian operator by comparing the actual response value after command execution with the predicted value. Specifically, the above steps focus on control execution and model self-evolution. Control commands are applied to power devices via pulse width modulation (PWM). During modulation, dead-time effects and non-ideal switching characteristics are considered, and waveform distortions related to current direction are compensated in real time. Traditional control methods rarely provide feedback corrections to the model after execution, resulting in long-term fixed model parameters that are difficult to adapt to long-term changes such as device aging and temperature drift. This step introduces a prediction-actual response comparison mechanism at the execution layer. By constructing an error functional to quantify model deviation, error information is propagated to the neural Hamiltonian operator via backpropagation, allowing for online updates of weight parameters and enabling the model to continuously evolve with the operating environment.

[0010] The beneficial effects of the invention are: (1) This invention constructs a converter state space model based on Hamiltonian energy evolution, performs unified energy characterization of voltage and current states, introduces physical sensing neural Hamiltonian operators to identify loss changes caused by temperature and load in real time, uses Riemann manifold to perform global optimization of modulation frequency and modulation ratio, combines dual-agent decision mechanism to complete operating mode adjustment, and realizes continuous updating of model parameters through error feedback. This invention can reduce bus voltage fluctuation and control error under photovoltaic fluctuation and load impact conditions, reduce switching and conduction losses, avoid parameter mismatch and local optimum problems, make the system operation more stable, maintain high efficiency and have continuous adaptive capability. (2) This invention constructs a unified energy modeling framework based on Hamiltonian structure to describe the dynamic process of converter from the perspective of energy evolution, realizes the consistent expression of state space model and physical mechanism, introduces physical sensing neural Hamiltonian operator to identify nonlinear disturbances and loss changes online, so that dissipation characteristics can be updated in real time with the operating environment, reconstructs control parameter space through Riemannian manifold modeling at the control optimization layer, maps energy loss distribution to geometric structure, and uses geodesic search to achieve efficient approximation of the global optimal modulation path, introduces a dual-agent game mechanism based on voltage sequence complexity at the decision layer to achieve adaptive trade-off between efficiency and stability, and constructs a closed-loop feedback mechanism for prediction and actual response at the execution layer to continuously self-evolve and update model parameters, forming a dynamic modulation system integrating perception, modeling, optimization, decision and learning, which significantly improves the control accuracy, response speed and long-term stability of the system under strong nonlinear and multi-disturbance conditions. Attached Figure Description

[0011] Figure 1 This is a flowchart of a method for dynamic safety control of a converter in a photovoltaic-storage-charging system, as proposed in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the system architecture of a photovoltaic energy storage and charging system proposed in Embodiment 2 of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0014] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0015] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0016] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0017] Example 1 Among them, such as Figure 1 A method for dynamic safety control of a converter in a photovoltaic-storage-charging system includes the following steps: S1. Collect electromagnetic state parameters and environmental characteristic data of the energy storage converter in real time through sensors, and construct a Hamiltonian state-space model based on the electromagnetic state parameters and environmental characteristic data; S2. Input the collected electromagnetic state parameters into the physical sensing neural Hamiltonian operator for nonlinear perturbation identification, and correct the dissipation matrix in the constructed port-controlled Hamiltonian state-space model based on the identification results. S3. Construct the Riemannian manifold optimization space based on the corrected dissipation matrix, and obtain the optimal control vector of the converter by searching the energy gradient descent direction in the Riemannian manifold optimization space; S4. Input the optimal control vector into the heterogeneous dual-agent game model, calculate the active power reference value and reactive power reference value, and combine the information entropy characteristic value of the collected electromagnetic state parameters to determine the operating mode of the converter. S5. Generate pulse width modulation commands based on active power reference values, reactive power reference values, and operating modes, and update the weight parameters of the physical sensing neural Hamiltonian operator by comparing the actual response value after command execution with the predicted value.

[0018] Specifically, the converter synchronously acquires instantaneous electromagnetic data from the DC-side battery port and the AC output-side filter inductor via high-precision voltage and current transformers. These real-time state variables are used to construct a Hamiltonian physical model characterizing the overall energy evolution trend of the system. This model accurately anchors the instantaneous potential energy distribution of the system during electromagnetic conversion using an energy gradient vector. Combined with the hardware topology connection logic of the power conversion circuit, an antisymmetric matrix reflecting the lossless energy flow direction is defined. To address the uncertainties in loss parameters caused by fluctuations in ambient temperature gradients, sudden changes in photovoltaic output, and aging of power devices during actual operation, the system introduces a physical sensing mechanism. The Hamiltonian operator extracts environmental features and charge state sequences in real time. This operator generates an energy potential scalar that can quantify the nonlinear dissipation trend of the system through deep feature mapping. The gradient distribution of this scalar with respect to the state variables is calculated using automatic differentiation technology, thereby obtaining a corrected feature vector that reflects the dynamic loss intensity in the current and voltage dimensions. This vector is transformed into a loss compensation term in the form of a diagonal matrix and superimposed on the initial dissipation matrix of the physical model. This reshapes the corrected dissipation matrix that can keep up with the drift of the actual operating conditions in real time, ensuring high-fidelity tracking of the actual operating characteristics of the converter by the state-space model from the underlying physical structure.

[0019] After obtaining the real-time corrected loss model, the algorithm projects the converter's switching frequency and modulation depth onto a non-Euclidean Riemannian geometric space. It uses the corrected dissipation characteristics to define the metric tensor and geometric curvature distribution of the manifold surface. By calculating the connection coefficients describing the rotational law of the tangent space, it constructs a geodesic differential equation and forces the control command to approach the global optimum along the geometric path with the lowest energy loss gradient. To cope with the drastic bus voltage fluctuations caused by random switching of high-power electric vehicles, the system calculates the approximate entropy of the sampled voltage sequence to assess the complexity level of the system disturbance in real time. This is combined with a heterogeneous dual-agent game model to optimize conversion efficiency and bus voltage. The system automatically adjusts the weight ratio between the two target dimensions of line voltage stability. The resulting optimal active and reactive power reference values ​​are transformed into a pulse width modulation drive sequence with dynamic dead zone compensation and frequency conversion characteristics, which is applied to the IGBT power device. The system compares the time-domain deviation between the actual response voltage and the predicted trajectory of the physical sensing operator in real time within the microsecond interval of modulation execution. It constructs an error deviation functional that characterizes the degree of mismatch between the physical entity and the mathematical model. The backpropagation mechanism is used to correct the neuron weight distribution inside the operator in reverse, so that the entire modulation algorithm can continuously learn autonomously and self-evolve in a closed loop as the converter operating environment and component characteristics evolve.

[0020] Furthermore, step S1 specifically includes the following sub-steps: S101. Obtain the bus voltage and battery current through voltage and current transformers, and calculate the total system energy Hamiltonian function based on the bus voltage and battery current; S102. Define an antisymmetric matrix through the circuit topology of the converter, and construct the lossless energy exchange relationship inside the system through the antisymmetric matrix; S103. Combine the total energy Hamiltonian function and the antisymmetric matrix to establish a Hamiltonian state-space model.

[0021] Furthermore, in step S103, the Hamiltonian state-space model expression is: ; ; Among them, the The system state vector represents the real-time state of the system, composed of voltage and current signals; This represents the derivative of the system state vector with respect to time, i.e., the rate of change of the state; Represent an antisymmetric matrix that satisfies This is used to represent the lossless energy exchange relationship between inductors and capacitors within a system; Represents the dissipation matrix; the The Hamiltonian function represents the total energy. For the system state vector The partial derivatives; The energy storage coefficient matrix is ​​composed of the inductance and capacitance values ​​in the converter circuit; The control input matrix represents the weights of the influence of external control variables on the system state; This represents the control input variable, specifically the switching modulation duty cycle signal of the converter bridge arm.

[0022] Specifically, the initial stage of step S1 focuses on abstracting the complex power electronic conversion process into a physical energy evolution model. High-precision voltage and current transformers are used to synchronously and rapidly sample the bus voltage and battery current of key nodes in the energy storage converter. After acquiring the data, the algorithm transforms these electromagnetic state parameters into a system state vector through a physical mapping mechanism. Based on the law of conservation of energy, the sum of electric and magnetic field energy in the energy storage elements is calculated, thereby constructing a Hamiltonian function that can quantify the total potential energy of the system. Further, the hardware topology of the converter is analyzed, the electromagnetic coupling path between inductors and capacitors is identified, and an interaction matrix with antisymmetric characteristics is defined to accurately describe how energy is exchanged losslessly between different energy storage elements under ideal conditions. By combining the lossless exchange relationship reflecting the system structure with the dissipation characteristics characterizing energy loss, and considering the influence of external modulation commands on system energy injection, a complete Hamiltonian state-space model is formed. This model accurately depicts the dynamic response trajectory of the converter under different operating conditions from the perspective of energy evolution.

[0023] Furthermore, in the Hamiltonian state-space model constructed in step S103, the antisymmetric matrix... To accurately characterize the lossless energy exchange relationship between the inductors and capacitors inside the converter, it must strictly meet the following requirements. Mathematical constraints. The specific construction method depends entirely on the hardware topology connection logic of the converter's main circuit. Taking a typical Boost DC / DC converter as an example, if the system state vector is selected as... ,That For inductor current, If the DC bus voltage is used, the energy exchange between the inductor and capacitor manifests as follows: changes in the inductor current affect the capacitor voltage through switching action, and conversely, changes in the capacitor voltage also affect the inductor current. This bidirectional, lossless coupling relationship can be mathematically represented as a matrix. ,in This represents the DC-side capacitor value. For three-phase inverters or more complex multilevel converters, the state vector dimension is expanded. The matrix will exhibit a block-based antisymmetric structure, where each pair of coupled inductor-capacitor states corresponds to a pair of off-diagonal elements with opposite values. It is particularly important to emphasize that... The matrix is ​​constructed without any resistance, switching losses, or environmental factors, and only reflects the ideal energy cycle path determined by the circuit topology itself, thus providing a pure and lossless energy framework for subsequent dissipation correction.

[0024] Furthermore, the dissipation matrix In the Hamiltonian state-space model, various irreversible energy losses during converter operation are quantified, including copper losses in inductors, equivalent line resistance, conduction losses of power switching devices, and additional losses caused by electromagnetic interference. In the initial model construction phase, i.e., step S103, It is constructed as a diagonal matrix, typically in the form of ,in This represents the equivalent series resistance of the inductor branch. The comprehensive equivalent conductance on the DC bus side is typically determined by the reciprocal of the load resistance or system leakage resistance. This initial construction method is based on nominal parameters measured offline, assuming that the loss characteristics are fixed and known. However, in actual photovoltaic-storage-charging systems, factors such as ambient temperature fluctuations, sudden changes in photovoltaic output, and IGBT device aging continuously alter the actual loss characteristics. Therefore, in step S2, the system identifies the current nonlinear disturbance characteristics online using a physical sensing neural Hamiltonian operator and generates a corrected feature vector. This vector is then transformed into a loss compensation term in diagonal matrix form. The modified dissipation matrix is ​​obtained by superimposing it onto the initial dissipation matrix. The revised version The matrix retains its positive semidefinite symmetry property, but its diagonal elements can now match the loss drift under actual operating conditions in real time, thus enabling the state-space model to have the ability to track real operating characteristics with high fidelity.

[0025] Furthermore, in the expression of the Hamiltonian state-space model, the control input matrix... With control input variable Together, they constitute the driving factors of external control commands on the evolution of the system state. Specifically, This represents the control input matrix, which maps abstract switching control commands to specific system state equations, reflecting the influence weights of external control variables on different state components. The construction method also depends on the converter's circuit topology. Taking the Boost circuit as an example, if the control input... The duty cycle signal representing the switching transistor will then appear in the dynamic equation of the inductor current, along with... and Related terms, and the dynamic equation of capacitor voltage will appear with and The relevant terms, therefore the control input matrix can be constructed as This shows that it is not only related to the topology, but may also depend on the current system state. .and This represents the control input variable, in the optical storage and charging system discussed in this paper. Specifically, it represents the switching modulation duty cycle signal of the power switching devices (such as IGBTs) in the converter bridge arm, and its value range is typically [value range missing]. Used for unipolar modulation, or Used for bipolar modulation. In actual operation, this duty cycle signal is generated in real time by the pulse width modulation module according to the optimal control vector and directly acts on the gate drive circuit of the IGBT, thereby controlling the orderly flow of energy from the battery side to the grid side or from the photovoltaic side to the battery side.

[0026] Furthermore, step S1 formalizes the physical characteristics of the converter using mathematical operators; specifically, it defines the system state vector. Its internal elements directly correspond to the bus voltage and battery current collected by the current transformer, which are used to anchor the system's position in phase space in real time. This is achieved by introducing an energy storage coefficient matrix composed of loop inductance and capacitance values. The total energy of the system is mapped to the total energy Hamiltonian function. Based on this, the energy gradient is obtained by taking the partial derivative of the function. This is used to drive the evolution direction of the state vector; to describe the energy flow under the converter topology constraints, the algorithm is defined to satisfy... antisymmetric matrix This matrix determines the energy transfer logic between the degrees of freedom within the system, while the dissipation matrix... This is used to characterize the inherent internal resistance loss of the system; by combining the above terms and introducing a signal containing the modulation duty cycle. With control input matrix The energy injection term is used to construct the Hamiltonian state-space model. In this model, It reflects the rate of state change in real time, and its changes are controlled by the internal energy exchange network. External control weights The synergistic effect.

[0027] Specifically, in step S1, after formalizing the physical characteristics of the converter using mathematical operators and defining the system state vector, a key physical parameter matrix—the energy storage coefficient matrix—needs to be introduced. The construction of this matrix depends on the nominal parameters of all independent energy storage elements in the converter's main circuit, specifically including the inductance values ​​of each branch inductor. and the capacitance values ​​of each node capacitor element. . Defined as a diagonal matrix, where each element on the diagonal corresponds to a state vector. The parameter values ​​of the energy storage element associated with a particular state variable. More precisely, if we follow the convention of defining the state vector... The first few components are arranged as the inductor currents, and the last few components are arranged as the capacitor voltages. The standard form of a matrix is That is, the inductance value of the inductor is placed at the position corresponding to the inductor current, and the capacitance value of the capacitor is placed at the position corresponding to the capacitor voltage. Taking the simplest Boost converter as an example, the state vector is selected... The first component is the inductor current, and the second component is the bus capacitor voltage. Therefore, the energy storage coefficient matrix is ​​specifically constructed as follows: ,in This refers to the inductance value of the Boost inductor. This is the capacitance value of the DC bus capacitor. (Introduced) The purpose of the matrix is ​​to construct an analytical expression for the Hamiltonian function of the system's total energy, i.e. .

[0028] Furthermore, step S2 specifically includes the following sub-steps: S201. Input the collected electromagnetic state parameters and environmental feature data into the physical sensing neural Hamiltonian operator. Extract the current operating condition features of the system based on the nonlinear mapping characteristics of the physical sensing neural Hamiltonian operator. For example, the operating condition features include electromagnetic dynamic features, environmental thermal features, and disturbance statistical features, wherein the electromagnetic dynamic features are derived from the real-time collected system state vector. The transient change patterns extracted, such as the ripple amplitude of inductor current, the rate of drop in bus voltage, and the steepness of the current rise edge, reflect the intensity of electromagnetic energy exchange in the converter during the current switching cycle. The environmental thermal characteristics include the real-time temperature value itself, the rate of temperature change, and the spatial distribution differences of the temperature gradient inside the converter, used to assess the nonlinear impact of temperature on parameters such as the on-resistance of power devices and the loss of inductor cores. The disturbance statistical characteristics are statistical quantities calculated from the short-time fluctuation sequence of electromagnetic state parameters, such as the variance of voltage fluctuations, the estimated value of current harmonic distortion rate, and complexity indices such as approximation, used to characterize the impact intensity of external disturbances such as sudden changes in photovoltaic output or random switching of electric vehicles on the system. S202. The operating condition feature vector is mapped to an energy dissipation potential scalar reflecting the intensity of nonlinear disturbances in the system through the output layer of the operator, and the partial derivative of the energy dissipation potential scalar with respect to the electromagnetic state parameters is obtained to generate a dynamically corrected feature vector. S203. Map the dynamically corrected eigenvectors to loss correction terms in the form of a diagonal matrix, and reconstruct the dissipation matrix in the Hamiltonian state-space model using the loss correction terms to obtain the corrected dissipation matrix.

[0029] Furthermore, in step S203, the corrected dissipation matrix is ​​obtained using the following formula: ; Among them, the loss correction term The calculation formula is: ; ; ; Among them, the This represents the corrected dissipation matrix, used to update the dissipation properties in the Hamiltonian state-space model; the... Represents the initial dissipation matrix; the This represents a loss correction term, used to quantify the additional energy loss caused by nonlinear disturbances; the... Represents the diagonalization operator; the This represents the dynamically corrected feature vector, reflecting the loss correction intensity of each dimension's state components; the... Represents the energy dissipation potential scalar. For the system state vector The partial derivatives; This represents an energy dissipation potential scalar, used to describe the equivalent energy loss field generated by the Hamiltonian operator of the physical sensory neural network; the... The nonlinear mapping function representing the Hamiltonian operator of the physical sensory neural network; This represents the real-time temperature value in the environmental characteristic data; This represents the neuron weights.

[0030] Specifically, during system operation, due to random fluctuations in photovoltaic output, the connection of nonlinear loads from charging piles, and the characteristic drift of power devices as temperature increases, traditional fixed dissipation matrices are insufficient to accurately describe the true energy loss of the system. Step S2 introduces a physically perceptive neural Hamiltonian operator, combining the nonlinear fitting capability of deep learning with Hamiltonian dynamic constraints. During implementation, the algorithm uses real-time monitored state variables such as inductor current and bus voltage, along with ambient temperature data from the heat sink surface, as multi-dimensional feature inputs. The neural network layer inside the operator identifies whether the current system is under light load, heavy load, or transient disturbance conditions through high-dimensional spatial mapping. It should be noted that this operator does not directly output a correction value, but generates a scalar potential energy value that can characterize the energy dissipation trend by simulating the dissipation potential energy field in the physical world. The algorithm uses automatic differentiation technology to calculate the sensitivity of this potential energy value to the system state, i.e., the distribution of the state gradient, thereby accurately locating abnormal fluctuations in energy in the current or voltage dimension. By superimposing the dynamically identified loss increments into the original physical model in the form of a diagonal matrix, the Hamiltonian state-space model can track and compensate for parameter drift caused by temperature stress and external disturbances in real time.

[0031] Furthermore, step S2 achieves a closed loop from raw sensory data to physical model correction through multi-level nested operations. First, the physical sensory neural Hamiltonian operator utilizes a mapping function... Input system state vector With environmental characteristic temperature Projected onto scalar space, it produces an energy dissipation potential scalar. This process is subject to the weight matrix. The algorithm captures the nonlinear behavior of the increase in the equivalent internal resistance of the device due to temperature rise by using the nonlinear activation function of the hidden layer to meet the constraints. In order to obtain the specific loss compensation amount for each state component, the algorithm performs partial derivative operations. That is, calculating the dissipation potential energy relative to the current. With voltage The partial derivatives are used to obtain the gradient vector; the absolute value operator is then used to... Generate dynamically corrected feature vectors And using the diagonalization operator Construct the loss correction term It reflects the inductor circuit loss compensation in real time on the matrix diagonal. Compensation for capacitor circuit losses ; through matrix summation operations The initial dissipation matrix in the original Hamiltonian state-space model Updated to the corrected dissipation matrix ,Should Feedback is given to the model In this way, the state prediction trajectory can self-align according to the evolution of ambient temperature and actual operating conditions, thereby improving the dynamic control accuracy and global stability of the converter in complex operating environments.

[0032] Furthermore, step S3 specifically includes the following sub-steps: S301. Define the Riemannian metric tensor using the modified dissipation matrix to establish a geometric mapping relationship between the converter control parameter space and the system energy loss; S302. Construct the Riemannian manifold optimization space using the converter's modulation frequency and modulation ratio as coordinate axes, and determine the curvature distribution characteristics on the manifold surface based on the Riemannian metric tensor; S303. Perform gradient search along the geodesic path where the energy function decreases within the Riemannian manifold optimization space to obtain the optimal control vector that minimizes the system's energy loss.

[0033] Furthermore, in step S303, the specific operation flow for performing gradient search to obtain the optimal control vector is as follows: S3031. Calculate the Riemann connection coefficients based on the Riemann metric tensor to describe the rotation law and geometric connection relationship of the inscribed space of the Riemann manifold optimization space; S3032. Construct a second-order geodesic differential equation about the coordinates of the control parameters using the Riemann connection coefficient, and solve it to obtain the evolution trajectory of the control parameters in the manifold space; S3033. Project the evolution trajectory onto the coordinate points of the manifold surface through exponential mapping, and output the optimal control vector that conforms to the physical constraints of the system.

[0034] Furthermore, in step S3, the Riemann metric tensor is constructed using the following formula: The control vector update formula in the geodesic path search is as follows: Among them, the The denot represents the Riemannian metric tensor, used to define the metric properties of the optimization space of the Riemannian manifold; the denot represents... This represents a coordinate vector of control parameters consisting of modulation frequency and modulation ratio; Represents the corrected dissipation matrix For the control parameter coordinate vector The Jacobian matrix; This represents the coordinate vector of the control parameters obtained from the search at the next moment; the... The coordinate vector representing the control parameters at the current moment; The step size factor represents the optimization step size; Represents the Riemannian metric tensor The inverse matrix is ​​used to correct the gradient descent direction to avoid high-loss regions; This represents the gradient of the total energy Hamiltonian function defined in step S1 with respect to the system state vector.

[0035] Specifically, for step S3032, after the Riemann connection coefficient is calculated in step S3031, the task of step S3032 is to transform the search problem of the optimal control vector into a geodesic solution problem on the Riemann manifold. In Riemann geometry, a geodesic is defined as the curve with the shortest local distance on the manifold. In the application context of this embodiment, it represents the trajectory of the control parameter coordinate vector as it naturally evolves along the surface of the manifold, and each segment of this trajectory satisfies the geometric constraint of the gentlest energy loss gradient.

[0036] The specific operation process is as follows: Using the Riemann connection coefficients calculated in step S3031, a second-order geodesic differential equation about the control parameter coordinates is constructed. In the standard form of this equation, the second derivative of the control parameter coordinates with respect to the affine parameters and the summation of the product of the connection coefficients and the first derivatives of the coordinates are both zero. This reflects the natural motion law on a curved manifold without external force. Here, the affine parameters can be understood as virtual time or arc length parameters along the geodesic path, while the connection coefficients characterize the geometric coupling strength between different directions on the manifold. To numerically solve this nonlinear second-order differential equation system, the system reduces it to a first-order state-space form. By introducing intermediate variables to represent the tangential velocity components of the control parameter coordinates along the geodesic, the original single second-order equation is transformed into two mutually coupled first-order equations: the first equation describes that the rate of change of the coordinate position with respect to the affine parameters is equal to the tangential velocity, and the second equation describes that the rate of change of the tangential velocity with respect to the affine parameters is equal to the sum of the pairwise products of the negative connection coefficients and the velocity components. When setting the initial conditions for the solution, the coordinate vector of the control parameters at the current moment is used as the starting point of the geodesic, and the initial tangential direction is set as the projection direction of the gradient of the calculated total energy Hamiltonian function with respect to the system state vector in the tangential space of the manifold, that is, the direction of the steepest energy descent. Furthermore, the standard numerical integration method is used to advance step by step along the affine parameters, and the current control parameter coordinates and tangential velocity are calculated at each integration step. After multiple integration steps, a series of coordinate points are obtained, which constitute the evolution trajectory of the control parameters in the Riemannian manifold optimization space. Each point on the entire trajectory corresponds to a set of candidate modulation frequencies and modulation ratios. Since the metric tensor of the manifold has encoded the corrected dissipation characteristics into the geometric curvature of the space through the connection coefficient, the trajectory naturally avoids the high energy loss region.

[0037] Furthermore, for step S3033, the geodesic evolution trajectory obtained in step S3032 is actually defined on the tangent bundle of the manifold, that is, it simultaneously contains the position information on the manifold surface and the velocity information in the tangent space; however, the control parameters required by the actual control command must fall directly on the coordinate points of the manifold surface itself. The task of step S3033 is to project the geodesic trajectory from the tangent space back to the manifold surface through the exponential mapping operation, and select the optimal control vector that meets the physical constraints.

[0038] The specific operation process is as follows: The geometric intuitive meaning of the exponential mapping is to start from a point on the manifold and reach another point on the manifold by following a geodesic of a unit arc length along a given tangential direction. In the numerical implementation of the embodiment, since step S3032 has obtained a series of intermediate coordinate points from zero affine parameter to maximum affine parameter through integrating the geodesic differential equation, these points are actually discrete approximations of the exponential mapping at different step sizes, located on the manifold surface. Because the solution of the geodesic differential equation is always constrained within the manifold, the first step of step S3033 is to extract all candidate points from the entire geodesic trajectory. However, not all points on the trajectory can be directly used as control commands for output. The optimal solution that satisfies the physical constraints of the converter must be selected. The specific selection criteria include two aspects: First, the control parameters corresponding to the candidate points must be within the operating range allowed by the hardware. For example, the modulation frequency must be between the highest switching frequency allowed by the power switching device and the lowest switching frequency that ensures control accuracy, and the modulation ratio must be between the upper and lower limits of the linear modulation region of the converter. Second, under the premise of satisfying the above physical constraints, the system prioritizes the point that minimizes the total energy Hamiltonian function value or makes the energy gradient magnitude lower than a preset threshold. In practice, the system starts from the starting point of the geodesic trajectory and searches step by step along the direction of increasing affine parameters. The first candidate point that simultaneously satisfies all physical constraints and makes the energy gradient index meet the requirements is selected as the optimal control vector. If no point on the entire trajectory can fully satisfy the constraints, the feasible point with the smallest total energy Hamiltonian function value on the entire trajectory is selected as the output, and the selected optimal control parameter coordinate vector is output to step S4. This output vector contains two core control parameters: modulation frequency and modulation ratio. The former determines the switching rate of the power switching device, and the latter determines the proportional relationship between the output voltage and the DC bus voltage. These two parameters together form the basis for calculating the active power reference value and reactive power reference value in subsequent steps. It should be noted that the control vector update formula given at the end of step S3033 is actually a first-order approximation of the exponential mapping method described in this step in Euclidean space, which is suitable for real-time control scenarios with limited computing resources.

[0039] Furthermore, step S4 specifically includes the following sub-steps: S401. Using the optimal control vector as the policy constraint boundary, construct a heterogeneous dual-agent game model that includes an energy-efficient agent and a voltage-stabilizing agent; S402. Perform sliding window sampling on the bus voltage sequence and obtain the information entropy feature value by calculating the approximate entropy of the voltage signal after reconstruction in phase space; S403. Determine the disturbance level of the system based on the relationship between the information entropy characteristic value and the preset instability threshold, decide the switching logic of the converter between the efficiency priority operation mode and the voltage support operation mode, and output the active power reference value and the reactive power reference value.

[0040] Furthermore, in step S402, the specific operation process for obtaining the information entropy feature value is as follows: S4021. Perform normalization processing on the bus voltage sequence and map the one-dimensional voltage sequence into a multi-dimensional phase space vector sequence according to the preset embedding dimension; S4022. Calculate the maximum absolute distance between any two vectors in a multidimensional phase space vector sequence, count the number of vector pairs whose distance is less than the similarity threshold, and obtain the association measure function; S4023. Calculate the probability of generating new patterns when the vector dimension increases by using logarithmic deviation operation, and output an approximate entropy characteristic value that characterizes the complexity of bus voltage fluctuations.

[0041] Specifically, in the implementation of converter optimal modulation, the algorithm extracts the modified dissipation matrix reflecting the switching characteristics and conduction losses of power devices, calculates its sensitivity matrix under changes in modulation frequency and modulation depth, and defines a geometric tensor describing the non-Euclidean space metric characteristics through the product of the sensitivity matrix and its transpose. This maps the originally flat control parameter region into an energy loss topography map with curvature fluctuations. In this topography map, the more sensitive the loss is to parameter changes, the higher the curvature value, corresponding to a geometrically steep mountain peak, while the ideal operating range with gentle loss changes is represented by a geometrically flat basin. This mapping mechanism enables the controller to perceive the movement of the loss boundary caused by operating condition fluctuations in real time, and automatically identifies candidate coordinate regions that take into account both conversion efficiency and dynamic response in the multi-dimensional parameter space. From the underlying geometric architecture, it achieves physical isolation of the high-loss operating point of the converter, provides navigation for finding the globally most power-efficient modulation path, and effectively solves the problem that traditional optimization methods are prone to getting trapped in local extrema at complex loss boundaries.

[0042] Furthermore, during the specific path search, the geometric coupling coefficients characterizing the rotational law and connection characteristics of the tangent space of the manifold surface are extracted by calculating the changes in the partial derivatives of each component of the metric tensor along the coordinate axes. A second-order nonlinear differential equation describing the evolution of the control variables within the curved space is constructed using these coefficients. The gradient distribution of the system's energy state is used as the optimization driving force. Simultaneously, a centripetal acceleration component generated by geometric constraints is introduced, forcing the search trajectory to curve along the geodesic direction that minimizes loss dissipation. To ensure that the solved parameter coordinates always fall within the physical manifold defined by the extreme values ​​of the converter hardware switching frequency and the boundary of the undistorted modulation ratio, an exponential mapping operator is introduced to project the velocity vector in the tangent space back to the true parameter coordinate system. This mapping process automatically corrects the proportion of the search step size in different curvature regions, enabling the parameters to automatically decelerate to prevent oscillations when crossing loss-sensitive regions, and accelerate convergence in loss-flat regions. Ultimately, the optimal modulation command combination that conforms to the power conversion topology constraints and maximizes the conversion efficiency to the physical limit is obtained.

[0043] Furthermore, in step S401, the Nash equilibrium solution formula for the heterogeneous two-agent game model is: ; The decision criterion formula for the operating mode is as follows: ; Wherein, the approximate entropy eigenvalue The calculation formula is: ; Among them, the This represents a defined active power reference value; the aforementioned This represents a defined reactive power reference value; the aforementioned The weighting coefficients represent the energy efficiency of the agent; The efficiency cost function is determined by the corrected dissipation matrix in step S2; The weighting coefficients of the voltage-stabilized intelligent agent; The voltage stability cost function is determined by the deviation between the bus voltage and the reference value; Indicates the operating mode of the converter; the The information entropy feature value is used to quantify the degree of disorder in the voltage sequence; the... Indicates the preset instability threshold; the Represents the logarithmic correlation measure; the The embedding dimension of the vector reconstruction; This represents the similarity threshold. Specifically, the efficiency cost function. The energy conversion efficiency loss of the converter under the current control command is used to quantify the power loss. Its specific value is obtained from the corrected dissipation matrix in step S2. The only certainty is that the corrected dissipation matrix is ​​a diagonal matrix, where each element on the diagonal corresponds to the equivalent loss intensity experienced by each state component in the system state vector. For example, the loss coefficient corresponding to the inductor current dimension reflects the combined effect of inductor copper loss, core loss, and line resistance, while the loss coefficient corresponding to the bus voltage dimension reflects the contributions of capacitor leakage loss and load equivalent conductance. In the specific calculation of the efficiency cost function, all diagonal elements of the corrected dissipation matrix are extracted to form a loss coefficient vector. The current electromagnetic state vector (including real-time values ​​such as inductor current and bus voltage) is then subjected to a quadratic operation with the loss coefficient vector; that is, the transpose of the state vector is multiplied by the corrected dissipation matrix and then multiplied by the state vector itself. Furthermore, to obtain a dimensionless efficiency cost function value that is easy to weight and sum with the voltage stability cost function, the instantaneous power loss is divided by a reference power value (e.g., the rated power of the converter or the current input power) to obtain a normalized efficiency loss factor. The closer this factor is to zero, the closer the converter's operating efficiency is to the ideal situation. The larger this factor is, the more severe the additional losses reflected in the corrected dissipation matrix are. For example, when the ambient temperature rises, causing the IGBT on-resistance to increase, or when the photovoltaic output suddenly changes, causing the inductor core to enter the saturation region, the corresponding elements in the corrected dissipation matrix will increase significantly, thereby increasing the efficiency cost function value. This prompts the game model to reduce the weight allocation of this low-efficiency operating point during the Nash equilibrium solution process.

[0044] Furthermore, the voltage stability cost function The voltage stability cost function is used to quantify the degree to which the current bus voltage deviates from the expected reference value and its potential threat to system stability. It is determined by calculating the deviation between the real-time acquired bus voltage signal and the preset voltage reference value. In practice, the instantaneous value of the current DC bus voltage or the effective value of the AC side voltage is obtained through a voltage transformer, depending on the grid connection mode of the photovoltaic-storage-charging system. The difference between this real-time voltage value and the preset reference voltage value is calculated to obtain the voltage deviation. To avoid the cancellation of positive and negative deviations, the square of the voltage deviation is generally taken as the basic cost term. However, considering only the instantaneous deviation is insufficient to comprehensively assess voltage stability, because the random switching of high-power electric vehicles often causes severe fluctuations in the bus voltage. The amplitude and duration of these fluctuations are more harmful than static deviations. Therefore, the voltage stability cost function also introduces a dynamic penalty term, which consists of the square or absolute value of the derivative of the voltage deviation with respect to time, used to penalize rapid voltage changes. The voltage stability cost function value is expressed as a weighted sum of the squared voltage deviation term and the voltage change rate penalty term, which is then normalized to obtain a dimensionless scalar value. The smaller the cost function value, the closer the bus voltage is to the reference value and the smoother the fluctuation, indicating that the system is in a stable operating state. The larger the cost function value, the more severe the bus voltage drop, overshoot, or violent oscillation. For example, when multiple electric vehicles start fast charging at the same time, the bus voltage may drop instantly, causing the voltage deviation square term to increase sharply. At the same time, the voltage change rate penalty term also increases significantly, resulting in a significant increase in the voltage stability cost function value. In this case, the game model will increase the weight coefficient of the voltage stabilization agent to prioritize the stability of the bus voltage rather than the conversion efficiency.

[0045] In the heterogeneous two-agent game model, the energy-efficient agent and the voltage-stabilizing agent each have their own cost functions. and As optimization objectives, there is an inherent competition between the two: improving operating efficiency often requires optimizing the switching frequency and modulation ratio to minimize losses, but this may reduce the system's ability to suppress voltage disturbances; conversely, strengthening voltage stability control usually requires reserving more control margin and dynamic response capability, introducing additional switching losses and thus reducing efficiency. Nash equilibrium solution formula The essence is to adjust the two weighting coefficients and The relative magnitudes of these factors are used to find a Pareto optimal compromise between the two objective dimensions of efficiency and stability. When the system is in steady-state operation and the approximate entropy eigenvalue is low, the weighting coefficients... The weighting coefficients are relatively large, and the game theory model focuses more on efficiency optimization; when a high-power perturbation is detected that causes the approximate entropy feature value to exceed a preset threshold, the weighting coefficients... Automatic adjustment: The game model shifts the control focus to voltage support; the obtained active power reference value and reactive power reference value are the optimal power command that satisfies the current game equilibrium point, while taking into account the two conflicting control objectives of efficiency and stability.

[0046] Furthermore, the geometric properties and dynamic evolution of Riemannian manifolds are quantified by a set of interrelated differential geometric formulas. First, through the formulas... Constructing metric tensors ,in To correct the dissipation matrix For vectors To describe the curvature connection of a surface, the Jacobian matrix is ​​used, and Christofel notation is introduced. The calculation formula is as follows: ; in, For measuring tensors Matrix elements in Its inverse matrix The corresponding components, the control parameters are based on the geodesic gradient evolution formula. Perform the update, where, This represents the coordinate vector of the control parameters obtained in the next time step. This represents the coordinate vector of the control parameters at the current moment. This represents the optimization step size factor. The core of this update formula lies in utilizing... Geometrically deform the Hamiltonian energy gradient, where Acting as a preprocessing operator on Riemannian manifolds, it scales the components of the gradient vector in different dimensions to make... The component along the high-loss gradient direction is suppressed by curvature, while the trajectory along the low-loss path is enhanced.

[0047] Furthermore, step S5 specifically includes the following sub-steps: S501. Combining active power reference values, reactive power reference values, and the operating mode of the decision, a pulse width modulation drive sequence containing dynamic dead zone compensation and frequency conversion characteristics is generated through a mapping function; S502. Drive the converter to execute the pulse width modulation drive sequence, and obtain the bus voltage and inductor current after execution as the actual response value in real time, and simultaneously calculate the state prediction value of the physical sensing neural Hamiltonian operator under the current operating condition. S503. Construct a deviation functional between the actual response value and the state prediction value, and combine the deviation functional to iteratively correct the weight parameters of the physical sensing neural Hamiltonian operator through a backpropagation mechanism, thereby realizing the closed-loop self-evolution of the control algorithm.

[0048] Specifically, at the low-level execution stage of the modulation command, based on the active power reference value and reactive power reference value output by the game model, and combined with the current operating mode of efficiency priority or voltage support, the modulation depth parameters and carrier frequency required by the converter bridge arm at the current moment are calculated. In order to eliminate the waveform distortion caused by the inherent switching characteristics of the power devices, the polarity and amplitude characteristics of the inductor current are extracted in real time, and microsecond-level dead time compensation pulses are dynamically injected to reshape the output voltage waveform. The drive control chip generates a high-precision pulse width modulation drive sequence to act on the power transistor. At the same time, the system synchronously acquires the instantaneous voltage and current data of the DC bus and AC side in microsecond-level steps. According to the method, it is encapsulated as a state vector reflecting the physical response of the converter, and used as truth feedback to align the state trajectory deduced by the physical sensing neural Hamiltonian operator under the same control input in the time domain. By quantifying the Euclidean distance between the actual physical evolution and the model prediction trajectory, a deviation functional that can reflect the degree of model parameter mismatch is constructed. The deviation is back-transmitted to the hidden layer of the neural Hamiltonian operator using automatic differentiation technology. By correcting the weight matrix inside the operator, parameter drift caused by power device aging, increased contact resistance, or nonlinear temperature gradient fluctuations is compensated in real time, realizing the dynamic consistency evolution between the control model and the physical entity of the converter.

[0049] Furthermore, the algorithm's closed-loop self-evolutionary logic is implemented through the state prediction and error backpropagation formula. Specifically, the PCH model constructed in step S1 is used to calculate the predicted state value at the next time step. Subsequently, through the formula Construct the bias functional; use the gradient descent formula Perform a weight update. Wherein, the... This represents the state vector of the system at the next sampling moment, predicted by the physical perception neural Hamiltonian operator; This indicates that step S2 includes neural weight parameters. The corrected dissipation matrix represents the loss characteristics after being reshaped by the physical sensing operator; the This represents the actual response state vector acquired by the voltage and current transformers after executing the modulation command; the The error scalar function represents the difference between the prediction accuracy of the model and the deviation from the physical entity; and These represent the neuron weight matrices of the physical perception neural Hamiltonian operator in step S2, after and before the update, respectively; The learning rate factor represents the factor used to control the self-evolution rate of the algorithm; This represents the gradient of the error's sensitivity to the weights, which is derived in reverse across the Hamiltonian state equations using the chain rule.

[0050] Example 2 A dynamic modulation system for an energy storage converter in a photovoltaic energy storage and charging system, such as Figure 2 This system is deployed in an integrated power station that includes photovoltaic power generation, energy storage, and DC charging functions. The system includes photovoltaic modules, PCS (Power Storage Converter), DC-DC converter, medium-voltage transformer, DC charging pile, external controller and monitoring system SCADA, power station controller, power grid measurement unit, and distribution network.

[0051] The photovoltaic module's output is connected to the DC-DC converter's input via a DC bus. The DC-DC converter's output is connected to the PCS's DC side via a DC bus. The PCS's AC side is connected to the medium-voltage transformer and the distribution network. The DC charging pile is directly connected to the DC bus between the photovoltaic module and the DC-DC converter. The grid measurement unit is installed on the AC line between the PCS and the medium-voltage transformer to collect real-time voltage, current, and frequency data at the grid connection point. The power station controller is connected to the PCS, DC-DC converter, DC charging pile, grid measurement unit, external controller, and SCADA monitoring system via a communication network, forming a centralized control and dispatch architecture.

[0052] The PCS integrates a data acquisition module, a Hamiltonian state space modeling module, a physical perception neural Hamiltonian operator module, a dissipation matrix correction module, a Riemannian manifold optimization module, a heterogeneous dual-agent game module, a pulse width modulation instruction generation module, and a closed-loop self-evolution module.

[0053] The data acquisition module consists of voltage and current transformers and temperature sensors installed on the DC and AC sides of the PCS. It is used to acquire real-time data on the DC bus voltage, AC side current, energy storage side inductor current, and power module temperature. The Hamiltonian state-space modeling module runs in the PCS's digital signal processor. Based on the bus voltage and inductor current output by the data acquisition module, it calculates the total system energy Hamiltonian function and defines an antisymmetric matrix using the PCS's circuit topology, thereby establishing a port-controlled Hamiltonian state-space model. The physical sensing neural Hamiltonian operator module is stored in the PCS's memory in the form of a neural network. Its input layer receives voltage, current, and temperature data output by the data acquisition module, and its output layer outputs an energy dissipation potential scalar. The dissipation matrix correction module calculates the partial derivatives of the energy dissipation potential scalar with respect to the electromagnetic state parameters to generate a dynamically corrected eigenvector. This vector is mapped to a diagonal matrix form of a loss correction term, which is then used to reconstruct the initial dissipation matrix in the Hamiltonian state-space model, obtaining the corrected dissipation matrix.

[0054] The Riemannian manifold optimization module runs within the PCS's digital signal processor. This module defines a Riemannian metric tensor using the corrected dissipation matrix, constructs a Riemannian manifold optimization space with the PCS's modulation frequency and modulation ratio as coordinate axes, and performs gradient search along a geodesic path where the energy function decreases. It outputs the optimal control vector that minimizes the system's energy loss. The heterogeneous dual-agent game module also runs within the PCS's digital signal processor. This module uses the optimal control vector as the policy constraint boundary, constructs a game model including an energy-efficiency agent and a voltage-stability agent, and calculates an approximate entropy as an information entropy feature value based on the bus voltage sequence provided by the data acquisition module. Based on the relationship between this feature value and a preset instability threshold, it determines whether the PCS's operating mode is efficiency-priority or voltage-support mode, and outputs active power and reactive power reference values. The pulse width modulation (PWM) command generation module receives the active power reference value, reactive power reference value, and operating mode, generates a PWM drive sequence including dynamic dead-time compensation and frequency conversion characteristics, and applies it to the PCS's power switches. The closed-loop self-evolution module collects the DC bus voltage and inductor current after the pulse width modulation command is executed in real time as the actual response value. At the same time, it reads the state prediction value of the physical sensing neural Hamiltonian operator module under the current operating condition, constructs the deviation functional between the two, and iteratively corrects the weight parameters of the physical sensing neural Hamiltonian operator module through the back propagation mechanism.

[0055] The power plant controller receives grid connection point data from the grid measurement unit and dispatch instructions from the external controller and SCADA system via the communication network. It then issues power dispatch target values ​​to the PCS, which serve as upper-level constraints for the execution of steps S1 to S5 by various modules within the PCS. Load fluctuations generated by DC charging piles during electric vehicle charging are directly transmitted to the PCS and DC-DC converter via the DC bus. The aforementioned dynamic modulation system within the PCS handles power balancing and voltage support. The distribution network provides a grid connection interface to the system through a medium-voltage transformer, supplementing the power deficit when photovoltaic module output is insufficient and feeding excess power back to the distribution network when photovoltaic module output is excessive. The grid measurement unit monitors the power quality data at the grid connection point in real time and uploads the data to the power plant controller, external controller, and SCADA system for system-level operational status assessment and dispatch decisions.

[0056] Example 3 This embodiment uses a photovoltaic-storage-charging integrated power station based on the above system architecture as an application scenario. The power station includes photovoltaic modules, a power generation system (PCS), a DC-DC converter, a medium-voltage transformer, DC charging piles, an external controller and monitoring system (SCADA), a power station controller, a power grid measurement unit, and a distribution network. The PCS, as the execution entity, integrates a digital signal processor and a memory. The memory stores executable instructions used to implement the dynamic modulation method for the energy storage converter of this invention.

[0057] In step S1, the PCS collects the DC bus voltage and AC side current in real time through its integrated voltage and current transformers, and simultaneously collects the power module temperature data through a temperature sensor. The PCS calculates the total system energy Hamiltonian function based on the collected bus voltage and inductor current, and defines an antisymmetric matrix using the PCS's circuit topology. Then, combining the total energy Hamiltonian function and the antisymmetric matrix, a port-controlled Hamiltonian state-space model is established. This model uses the DC bus voltage and inductor current to form the system state vector, and the switching modulation duty cycle signal of the PCS bridge arm as the control input variable.

[0058] In step S2, the PCS inputs the collected bus voltage, inductor current, and temperature data into its internally stored physical sensing neural Hamiltonian operator. This operator extracts the current operating condition characteristics of the system through nonlinear mapping and calculates an energy dissipation potential scalar reflecting the intensity of the system's nonlinear disturbances. The PCS calculates the partial derivative of this energy dissipation potential scalar with respect to the electromagnetic state parameters, generating a dynamically corrected eigenvector. This eigenvector is then mapped to a loss correction term in diagonal matrix form. Finally, this loss correction term is used to reconstruct the initial dissipation matrix in the Hamiltonian state-space model from step S1, obtaining the corrected dissipation matrix.

[0059] In step S3, the PCS uses the corrected dissipation matrix obtained in step S2 to define a Riemann metric tensor, establishing a geometric mapping relationship between the PCS control parameter space and the system energy loss. The PCS constructs a Riemann manifold optimization space using its own modulation frequency and modulation ratio as coordinate axes, and determines the curvature distribution characteristics on the manifold surface based on the Riemann metric tensor. Within this Riemann manifold optimization space, the PCS performs gradient search along a geodesic path where the energy function decreases, sequentially calculating the Riemann connection coefficient, constructing a second-order geodesic differential equation about the control parameter coordinates, and solving it to obtain the evolution trajectory of the control parameters. Then, through exponential mapping, the evolution trajectory is projected onto the coordinate points of the manifold surface, outputting the optimal control vector that conforms to the system's physical constraints.

[0060] In step S4, the PCS uses the optimal control vector obtained in step S3 as the policy constraint boundary to construct a heterogeneous dual-agent game model containing an energy efficiency agent and a voltage stability agent. The PCS performs sliding window sampling on the bus voltage sequence, normalizes the sampled voltage sequence, maps the one-dimensional voltage sequence to a multi-dimensional phase space vector sequence according to a preset embedding dimension, calculates the maximum absolute distance between any two vectors in the multi-dimensional phase space vector sequence, and obtains the correlation measure function by statistically analyzing the proportion of vectors with a distance less than a similarity threshold. Then, it calculates the probability of generating a new mode when the vector dimension increases through logarithmic deviation calculation, outputting an approximate entropy feature value characterizing the complexity of bus voltage fluctuations. Based on the relationship between this approximate entropy feature value and a preset instability threshold, the PCS determines the system's disturbance level, decides its switching logic between efficiency-first operation mode and voltage-supported operation mode, and outputs active power reference values ​​and reactive power reference values.

[0061] In step S5, the PCS combines the active power reference value, reactive power reference value, and the decision-making operating mode obtained in step S4 to generate a pulse width modulation drive sequence that includes dynamic dead-zone compensation and frequency conversion characteristics through a mapping function. The PCS applies this drive sequence to the internal power switch to perform pulse width modulation and acquires the DC bus voltage and inductor current after execution in real time as the actual response value. Simultaneously, it calculates the state prediction value of the physical sensing neural Hamiltonian operator under the current operating condition. The PCS constructs a deviation functional between the actual response value and the state prediction value and iteratively corrects the weight parameters of the physical sensing neural Hamiltonian operator through a backpropagation mechanism, achieving closed-loop self-evolution of the control algorithm.

[0062] In the operation of this embodiment, when a DC charging pile is connected to an electric vehicle and charging begins, a sudden change in charging load causes a drop in the DC bus voltage. The PCS completes disturbance identification, manifold optimization, game theory decision-making, and pulse width modulation update within 100 microseconds through the aforementioned steps S1 to S5, outputting corresponding active and reactive power to support the DC bus voltage. The grid measurement unit monitors the power quality at the grid connection point in real time and uploads the data to the power station controller. The power station controller reports data to the external controller and the monitoring system SCADA according to the system-level operating status and receives dispatch instructions. At the same time, it issues a power dispatch target value to the PCS as an upper-level constraint of the game model in step S4. When the output of the photovoltaic module exceeds the local load demand, the PCS feeds the excess power into the distribution network through the medium-voltage transformer through steps S1 to S5. When the output of the photovoltaic module is insufficient, the PCS draws power from the distribution network and supplies power to the DC bus through the DC-DC converter, or controls the energy storage system to discharge. Throughout the process, the external controller and monitoring system SCADA centrally monitor and record the operating status of the PCS, photovoltaic modules, DC charging piles, power grid measurement units, and power station controllers.

[0063] Example 4 Furthermore, as a preferred embodiment of the above-described embodiment one, a dynamic safety control system for a converter in a photovoltaic-storage-charging system is proposed. This system is deployed as a whole in a DC-coupled photovoltaic-storage-charging project. The project includes a 500kW AC / DC converter, a 60kW energy storage DC / DC module, five 100kW photovoltaic DC / DC modules, and two DC charging piles. The photovoltaic converter is connected to the DC side of the energy storage DC / DC module and the AC / DC converter via a DC bus. This system is integrated within the aforementioned 500kW AC / DC converter, employing a DSP and ARM collaborative architecture as the core computing unit. The DSP side handles high-frequency real-time control calculations, while the ARM side handles model management and neural network inference scheduling tasks. Data interaction between the two is achieved through shared memory and a high-speed bus. The system runs a real-time operating system at its core, with a scheduling mechanism divided into a 50μs control cycle task queue and a 1ms learning and update task queue.

[0064] The high-speed acquisition link synchronously acquires the DC-side bus voltage, energy storage-side inductor current, and AC-side grid-connected current signals of the AC-DC converter at a sampling frequency of 20kHz using a 16-bit high-precision ADC. The temperature acquisition channel acquires the internal IGBT module temperature of the AC-DC converter and the ambient temperature data of the project site at a sampling frequency of 1kHz. All sampled data enters a double-buffered circular queue to form a continuous state window with a window length corresponding to a range of 1.6ms. The state construction unit performs normalization processing on the voltage and current data within the window and generates a state vector containing bus voltage and inductor current components. The energy mapping unit calls the AC-DC converter main circuit inductor calibration value of 0.5mH and capacitor calibration value of 2000μF from the local parameter table to convert electromagnetic state quantities into system energy expressions. The topology analysis unit generates antisymmetric structural relationships based on the pre-stored AC-DC converter circuit connection matrix and performs dynamic updates based on the switching states during operation. The state space generation unit integrates energy gradient, topology relationships, and control input information to form a real-time state evolution description and writes it into a shared memory area for subsequent modules to call.

[0065] The physical sensing neural Hamiltonian operator module is deployed on the ARM side. Input data consists of a state vector, IGBT module temperature, ambient temperature, inductor current change rate, and bus voltage drop amplitude, with an 8-dimensional input dimension. The network structure employs a 5-layer feedforward network. The first three layers each have 96 neurons, using the Swish activation function to enhance nonlinear expression. The last two layers each have 64 neurons, using the Tanh function to constrain the output range. Network weights are initialized using the He method and support online updates, with a single inference latency controlled within 50μs. The network output is an energy dissipation potential scalar. The automatic differentiation unit performs state variable gradient calculations on this scalar based on a lightweight computation graph, retaining only the voltage and current dimensions. The gradient results are processed by absolute value analysis to form a dynamically corrected feature vector. The dissipation correction generation unit maps this vector to the matrix diagonal to generate loss compensation terms and superimposes them onto the initial dissipation matrix, completing real-time updates. This initial dissipation matrix is ​​pre-calibrated using the line resistance and switching losses of the AC / DC converter under steady-state conditions.

[0066] The Riemannian manifold optimization module operates on the DSP side, mapping the modulation frequency and modulation ratio of the AC / DC converter to two-dimensional control coordinates. The frequency resolution is set to 0.5 kHz, covering a range of 2 kHz to 12 kHz, and the modulation ratio resolution is set to 0.01, covering a range of 0 to 1. The geometry metric construction unit calculates the sensitivity of the corrected dissipation matrix to the control parameters to form a Jacobian matrix, performs matrix multiplication to generate the Riemannian metric tensor, and accelerates the matrix operations using the SIMD instruction set. The geodesic search unit employs a finite-step iteration mechanism, controlling the number of iterations per cycle to 6. When the rate of change of the metric tensor eigenvalues ​​exceeds a set threshold, it determines that it has entered a high-curvature region and automatically reduces the step size; conversely, it accelerates the convergence speed in a low-curvature region. The control vector output unit performs boundary pruning on the search results to ensure that the modulation frequency does not exceed the safe switching frequency range of the AC / DC converter's IGBT module and that the modulation ratio does not exceed the maximum modulation ratio limit.

[0067] The game-theoretic decision-making module runs on the ARM side. The voltage sequence caching unit maintains an 8ms sliding window with 200 data points within it. The complexity calculation unit performs an approximate entropy algorithm on the bus voltage sequence within the window. This includes normalization, mapping the one-dimensional voltage sequence to a multi-dimensional phase space vector sequence with an embedding dimension of 2, calculating the maximum absolute distance between vectors, obtaining the correlation measure function by comparing vectors whose distance is less than 0.2 times the signal standard deviation (a similarity threshold), calculating the probability of generating a new pattern when the embedding dimension increases from 2 to 3 using logarithmic bias calculation, and outputting the approximate entropy feature value. The dual-agent decision-making unit contains two lightweight neural network sub-modules, corresponding to the energy efficiency target and the voltage stability target, respectively. Each sub-network has a 3-layer structure with 48 neurons per layer. The two sub-networks take the corrected dissipation matrix and the bus voltage deviation as inputs, and output the efficiency cost and voltage stability cost, respectively. The active power reference value and reactive power reference value are obtained by solving through Nash equilibrium. The mode switching control unit performs operation mode determination based on the comparison result of the approximate entropy characteristic value and the preset instability threshold of 0.45. When the approximate entropy is less than or equal to 0.45, it switches to the efficiency priority mode, and when the approximate entropy is greater than 0.45, it switches to the voltage support mode. A hysteresis interval of 0.05 is introduced to avoid frequent switching.

[0068] The PWM generation and execution module operates within the DSP timer unit. Based on the active and reactive power reference values ​​output by the game theory decision module and the current operating mode, it calculates the modulation depth and carrier frequency. The carrier frequency dynamic range is set to 6kHz to 20kHz. The dead time reference value is set to 2.2μs. The dynamic compensation unit adjusts the compensation amount in real time according to the direction and amplitude of the AC / DC converter output current, achieving a compensation resolution of 0.1μs. The PWM signal is output through the driver chip to the IGBT module inside the AC / DC converter to perform power conversion, realizing bidirectional power regulation between the DC and AC buses.

[0069] The error calculation and self-evolution module collects the actual bus voltage and inductor current response data of the AC / DC converter after each control cycle, compares it with the state prediction results of the physical sensing neural Hamiltonian operator module under the current operating conditions, and calculates the deviation scalar. The error data is input into the learning module to perform backpropagation, and gradient calculation is completed across the network structure of the state space model and the physical sensing neural Hamiltonian operator module. The learning rate is dynamically adjusted according to the operating state. When the approximate entropy eigenvalue exceeds 0.45, it is judged as a strong disturbance stage, and the learning rate is increased to 1.3 times that of the steady-state stage, with the weight update period set to 1ms. During continuous operation, the system gradually develops an adaptive capability to the output fluctuations of the photovoltaic converter and the load impact characteristics of the charging pile.

[0070] The modules mentioned above achieve lock-free data exchange through shared memory and a high-speed bus. Key control paths are executed locally in a closed loop on the DSP, maintaining overall control latency within the 100μs range, forming a real-time dynamic modulation closed-loop control system that fully corresponds to steps S1 to S5. This system works collaboratively with five photovoltaic converters, one energy storage DC / DC module, an EMS system, and AC charging piles in a DC-coupled photovoltaic-energy storage-charging project. The photovoltaic converters feed the DC power generated by the photovoltaic array into the DC bus. The energy storage DC / DC module realizes bidirectional energy conversion between the energy storage battery and the DC bus. The AC / DC converter uses this system to achieve power regulation between the DC bus and the AC bus, jointly completing the absorption of photovoltaic power generation, the charging and discharging management of the energy storage system, and the power supply support for the charging pile load.

[0071] Example 5 Furthermore, as a preferred embodiment of the above-described embodiment four, a photovoltaic-storage-charging system based on DC coupling installed in an industrial park is taken as the specific application scenario of this embodiment. This system includes a 500kW AC / DC converter, a 60kW energy storage DC / DC module, five 100kW photovoltaic DC / DC modules, and two DC charging piles. The AC / DC converter integrates a digital signal processor to execute the converter dynamic safety control method proposed in either embodiment one or embodiment four.

[0072] In the initial operating state of the system, the photovoltaic converter operates in maximum power point tracking mode with an output power of 350kW, the energy storage DC / DC module is in standby mode, and the charging pile is not connected to the vehicle. The AC / DC converter collects the DC bus voltage of 700V and the inductor current of the energy storage side in real time through voltage and current transformers, and obtains the ambient temperature of 30℃ through temperature sensors. Based on the collected electromagnetic state parameters and temperature data, the AC / DC converter constructs a port-controlled Hamiltonian state-space model. This model uses the DC bus voltage and inductor current as system state variables and the switching modulation duty cycle signal of the AC / DC converter bridge arm as the control input variable.

[0073] When an electric heavy-duty truck is connected to a charging pile on the AC bus side and fast charging is started, the charging power jumps from zero to 120kW within seconds, causing the DC bus voltage to drop to 660V. The AC-DC converter inputs the currently collected bus voltage, inductor current, and ambient temperature data into the physical sensing neural Hamiltonian operator. This operator extracts the current operating condition characteristics of the system through its nonlinear mapping capability, identifies the nonlinear disturbance caused by the sudden change in the charging pile load, and calculates the energy dissipation potential scalar that reflects the intensity of the nonlinear disturbance. The AC-DC converter takes the partial derivative of the energy dissipation potential scalar with respect to the electromagnetic state parameters to generate a dynamic correction eigenvector. The dynamic correction eigenvector is then mapped to a loss correction term in the form of a diagonal matrix. This loss correction term is used to reconstruct the initial dissipation matrix in the Hamiltonian state-space model to obtain the corrected dissipation matrix, enabling the model to accurately describe the energy loss characteristics under the current disturbance.

[0074] The AC / DC converter further utilizes the modified dissipation matrix to define a Riemannian metric tensor, establishing a geometric mapping relationship between the converter control parameter space and system energy loss. Specifically, using the AC / DC converter's modulation frequency and modulation ratio as two coordinate axes, a two-dimensional Riemannian manifold optimization space is constructed. Based on the Riemannian metric tensor, the curvature distribution characteristics on the manifold surface are determined. Within this manifold optimization space, the AC / DC converter performs gradient search along the geodesic path where the energy function decreases, sequentially calculating the Riemannian connection coefficient, solving the second-order geodesic differential equation, and projecting the evolution trajectory onto the coordinate points of the manifold surface through exponential mapping. Finally, the optimal control vector that minimizes system energy loss is obtained. This optimal control vector includes an adjusted modulation frequency of 4.8 kHz and a modulation ratio of 0.72.

[0075] Furthermore, the AC / DC converter uses the aforementioned optimal control vector as the policy constraint boundary and initiates an internal heterogeneous dual-agent game model. This game model includes an energy efficiency agent and a voltage stability agent, each aiming to minimize system losses and maintain bus voltage stability, respectively. The AC / DC converter simultaneously performs sliding window sampling on the bus voltage sequence, with the sampling window containing the most recent 200 voltage sampling points at a sampling frequency of 5kHz. By mapping the one-dimensional voltage sequence to a multi-dimensional phase space vector sequence according to a preset embedding dimension, the maximum absolute distance between any two vectors is calculated. The proportion of vector pairs with a distance less than a similarity threshold is statistically analyzed. Then, the probability of generating a new pattern when the vector dimension increases is calculated using logarithmic deviation calculation. Finally, an approximate entropy feature value characterizing the complexity of bus voltage fluctuations is output. In this embodiment, due to a sudden change in charging pile load causing large voltage fluctuations, the calculated approximate entropy value is 0.58, while the system's preset instability threshold is 0.45. The approximate entropy exceeds the threshold, indicating a high level of disturbance. Based on this, after solving the heterogeneous dual-agent game model through Nash equilibrium, the decision is to switch the operating mode of the AC-DC converter from the efficiency-first mode to the voltage-support mode, and at the same time output an active power reference value of 85kW and a reactive power reference value of 30kVar, wherein the reactive power is drawn by the AC-DC converter from the AC side to support the DC bus voltage.

[0076] Based on the aforementioned active power reference values, reactive power reference values, and voltage-supported operating modes, the AC / DC converter generates a pulse-width modulation drive sequence, incorporating dynamic dead-time compensation and frequency conversion characteristics, through a mapping function, and drives the internal power switches to execute. After the drive command is executed, the AC / DC converter immediately acquires the bus voltage and inductor current as the actual response values. Simultaneously, the physical sensing neural Hamiltonian operator calculates the state prediction value under the current operating conditions. The AC / DC converter constructs a deviation functional between the actual response value and the predicted value, and iteratively corrects the weight parameters of the physical sensing neural Hamiltonian operator through a backpropagation mechanism, achieving closed-loop self-evolution of the control algorithm. After a response time of approximately 20 milliseconds, the DC bus voltage recovers to 690V, the charging pile power supply is stable, and the output power of the photovoltaic converter is not significantly affected.

[0077] Throughout the control process, the system utilizes a DC-coupled architecture requiring only two DC cables to complete power transmission between the photovoltaic system, energy storage, and the AC / DC converter, reducing line costs and construction difficulty compared to AC-coupled schemes. Simultaneously, the AC / DC converter, through the dynamic modulation method of this invention, automatically completes nonlinear disturbance identification, Riemannian manifold optimization, dual-agent game decision-making, and closed-loop self-evolutionary updates without manual intervention. This achieves coordinated mitigation of photovoltaic power generation fluctuations and charging pile load abrupt changes, improving power supply reliability and power quality. These functions, working in conjunction with the AC / DC converter, energy storage DC / DC module, photovoltaic converter, and EMS system in the DC-coupled photovoltaic-energy storage-charging project scheme, collectively realize the observability, measurability, controllability, and adjustability of the integrated source-grid-load-storage system. That is, scheduling can be achieved at the microgrid level by controlling the active and reactive power of the AC / DC converter, simplifying the control logic and optimizing the control architecture.

[0078] Example 6 Furthermore, as a preferred embodiment of the above-mentioned Example 1, a dynamic safety control system for a converter in a photovoltaic-storage-charging system is proposed. In this embodiment, the system is deployed within the edge control platform of the integrated photovoltaic-storage-charging station, employing a DSP and ARM collaborative architecture as the core computing unit. The DSP side undertakes high-frequency real-time control calculation tasks, while the ARM side undertakes model management and neural network inference scheduling tasks. The two interact via shared memory and a high-speed bus. The system runs a real-time operating system (RTOS) at its core, with a scheduling mechanism divided into a 50μs control cycle task queue and a 1ms learning and update task queue. The high-speed acquisition link synchronously acquires bus voltage, battery current, and filter inductor current signals at a frequency of 20kHz to 25kHz using a 16-bit high-precision ADC. The temperature acquisition channel acquires IGBT module temperature and ambient temperature data at a frequency of 1kHz to 2kHz. All sampled data enters a double-buffered circular queue to form a continuous state window, with a window length ranging from 1.6ms to 2ms.

[0079] The state construction unit normalizes the voltage and current data within the window and generates state vectors. The energy mapping unit calls the inductor and capacitor calibration values ​​from the local parameter table to convert electromagnetic state quantities into system energy expressions. The topology analysis unit generates antisymmetric structural relationships based on pre-stored circuit connection matrices and performs dynamic updates based on switch states during operation. The state space generation unit integrates energy gradients, topological relationships, and control input information to form a real-time state evolution description and writes it to a shared memory area for subsequent modules to call.

[0080] The physical sensing neural Hamiltonian operator module is deployed on the ARM side or a standalone NPU unit. Input data consists of a state vector, temperature value, current change rate, and voltage drop amplitude, with the input dimension controlled within the range of 8 to 10 dimensions. The network structure employs a 5-7 layer feedforward network, with 64-128 neurons per layer. The Swish activation function is used in the front layers to enhance nonlinear expression, while the Tanh function is used in the back layers to constrain the output range. Network weights are initialized using the He method and support online updates, with a single inference latency controlled within 50 μs. The network output is an energy dissipation potential scalar. The automatic differentiation unit performs state variable gradient calculation on this scalar based on a lightweight computation graph, retaining only the voltage and current dimension paths. The gradient results are processed by absolute value to form a dynamically corrected feature vector. The dissipation correction generation unit maps this vector to the diagonal of a matrix to generate a loss compensation term, which is then superimposed on the original dissipation matrix, completing the real-time update.

[0081] The Riemannian manifold optimization module runs on the DSP side, mapping modulation frequency and modulation ratio to two-dimensional control coordinates. The frequency resolution is set to 0.5kHz, and the modulation ratio resolution is set to 0.01. The geometry metric construction unit calculates the sensitivity of the dissipation matrix to the control parameters to form a Jacobian matrix, performs matrix multiplication to generate a metric tensor, and accelerates matrix operations using the SIMD instruction set. The geodesic search unit employs a finite-step iteration mechanism, controlling the number of iterations per cycle to between 5 and 8. The step size is automatically reduced in high-curvature regions, and convergence is accelerated in low-curvature regions. The control vector output unit performs boundary clipping on the results to ensure that the hardware frequency and modulation range constraints are met.

[0082] The game-theoretic decision-making module runs on the ARM side. The voltage sequence caching unit maintains a sliding window of 8ms to 10ms, containing approximately 200 to 250 data points. The complexity calculation unit implements an approximate entropy algorithm, employing an incremental update mechanism to reduce redundant calculations. The dual-agent decision-making unit contains two lightweight neural network sub-modules, corresponding to the energy efficiency objective and the voltage stability objective, respectively. Each sub-network has three layers, with 32 to 64 neurons per layer. The output results are normalized to generate weight allocation ratios. The mode switching control unit determines the operating mode based on the complexity value and duration threshold, introducing a hysteresis interval to avoid frequent switching.

[0083] The PWM generation and execution module operates within the DSP timer unit. It calculates the modulation depth and carrier frequency based on active and reactive power reference values ​​and the current operating mode. The carrier frequency dynamic range is 6kHz to 20kHz. The dead time reference value is set to 2μs to 2.5μs. The dynamic compensation unit adjusts the compensation amount in real time according to the current direction and amplitude, achieving a compensation resolution of 0.1μs. The PWM signal is output to the IGBT module via the driver chip to perform power conversion.

[0084] The error calculation and self-evolution module collects actual voltage and current response data after each control cycle, compares it with the model prediction results, and calculates the deviation scalar. Error data is input into the learning module to perform backpropagation, and gradient calculation is completed across the state-space model and neural network structure. The learning rate is dynamically adjusted according to the operating state, increasing to 1.2 to 1.5 times that of the steady-state stage during strong disturbances, while the weight update cycle is controlled within the range of 0.8ms to 1ms. During continuous operation, the system gradually develops adaptive capabilities to fluctuations in illumination and load impact characteristics.

[0085] The system modules exchange data without locks through shared memory and high-speed bus. The key control path is executed locally in closed loop on the DSP. The overall control delay is kept within the order of 100μs, forming a real-time dynamic modulation closed-loop control system that corresponds completely to steps S1 to S5.

[0086] Example 7 Furthermore, as a preferred embodiment of the above embodiments, an application scenario is proposed for the dynamic safety control method of the converter for a photovoltaic-storage-charging system proposed in Embodiment 1. This application scenario involves a 500kW photovoltaic-storage-charging system with a photovoltaic array rated output power of 5*100kW, a rated storage battery capacity of 200kWh, a converter rated power of 500kW, and a DC bus nominal voltage of 750V. The on-site test was conducted at midday in summer, when cloud cover moved rapidly and sunlight intensity fluctuated between 1100W / m² and 1400W / m².

[0087] After system startup, the voltage transformer acquires the bus voltage at a sampling frequency of 20kHz, the current transformer synchronously acquires the battery current and inductor current at 20kHz, and the temperature sensor acquires the surface temperature of the IGBT module and the heat sink temperature at a frequency of 1kHz. The state vector consists of the bus voltage, battery current, and inductor current, with a sampling window length set to 2ms, corresponding to 40 sets of continuous state data.

[0088] During the sudden change in sunlight intensity, the photovoltaic output power decreased from 360kW to 210kW within 120ms, and the bus voltage showed a trend of dropping from 750V to 705V. Real-time state data was input into a physical sensing neural Hamiltonian operator. The operator's internal network depth was set to 6 layers, with 64 neurons per layer, and a nonlinear saturation activation function was used. The operator identified the combined characteristics of rapid voltage drop and current rise, classifying the current operating state as a "high-perturbation energy migration state," and outputting an enhanced dissipation feature vector in the current dimension. The correction magnitude of the dissipation matrix in the current channel reached 1.8 times the initial value, and the correction magnitude in the voltage channel was 1.2 times.

[0089] The revised model then enters the control optimization phase. The allowed modulation frequency range is 5kHz to 20kHz, and the modulation ratio range is 0.6 to 0.95. A control parameter grid is established in the Riemannian manifold space, with a grid resolution of 0.5kHz and a modulation ratio interval of 0.02. Curvature analysis results show that in the region above 15kHz, the loss gradient increases significantly, forming a high-curvature region. The optimization path automatically converges to the 10kHz to 12kHz range, and the modulation ratio stabilizes between 0.82 and 0.87.

[0090] The bus voltage sequence was analyzed using a 10ms sliding window, with each window containing 200 sampling points. The approximate entropy calculation result increased from 0.12 to 0.36 after the disturbance, exceeding the set threshold of 0.3, and the system operating mode switched to voltage support mode. In the game theory model, the voltage stability weight was increased to twice its original value, the active power reference value was increased to 220kW, and the reactive power reference value was adjusted to -35kVar to enhance the bus support capability.

[0091] During the PWM modulation stage, the system generates a carrier signal with a frequency of 11kHz and a modulation depth of 0.85. The base dead time is 2μs, and the dynamic compensation range is ±0.8μs, with compensation adjusted in real time based on the current direction and amplitude. The output voltage ripple is reduced from the original 18V to 9V.

[0092] During the control execution process, the system samples the actual voltage and current response at a period of 50μs and compares it with the model prediction results. The single-cycle error is controlled within 2.5%. After running continuously for 30s, the error converges to 1.2%. The neural operator weight update period is 1ms, and the learning rate gradually decays from the initial value to 40% of the stable value.

[0093] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for dynamic safety control of a converter in a photovoltaic-storage-charging system, characterized in that, Includes the following steps: S1. Collect electromagnetic state parameters and environmental characteristic data of the energy storage converter in real time through sensors, and construct a Hamiltonian state-space model based on the electromagnetic state parameters and environmental characteristic data; S2. Input the collected electromagnetic state parameters into the physical sensing neural Hamiltonian operator for nonlinear perturbation identification, and correct the dissipation matrix in the constructed port-controlled Hamiltonian state-space model based on the identification results. S3. Construct the Riemannian manifold optimization space based on the corrected dissipation matrix, and obtain the optimal control vector of the converter by searching the energy gradient descent direction in the Riemannian manifold optimization space; S4. Input the optimal control vector into the heterogeneous dual-agent game model, calculate the active power reference value and reactive power reference value, and combine the information entropy characteristic value of the collected electromagnetic state parameters to determine the operating mode of the converter. S5. Generate pulse width modulation commands based on active power reference values, reactive power reference values, and operating modes, and update the weight parameters of the physical sensing neural Hamiltonian operator by comparing the actual response value after command execution with the predicted value.

2. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S101. Obtain the bus voltage and battery current through voltage and current transformers, and calculate the total system energy Hamiltonian function based on the bus voltage and battery current; S102. Define an antisymmetric matrix through the circuit topology of the converter, and construct the lossless energy exchange relationship inside the system through the antisymmetric matrix; S103. Combine the total energy Hamiltonian function and the antisymmetric matrix to establish a Hamiltonian state-space model.

3. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 2, characterized in that, In step S103, the Hamiltonian state-space model expression is: ； ； Among them, the The system state vector represents the real-time state of the system, composed of voltage and current signals; This represents the derivative of the system state vector with respect to time, i.e., the rate of change of the state; Represent an antisymmetric matrix that satisfies This is used to represent the lossless energy exchange relationship between inductors and capacitors within a system; Represents the dissipation matrix; the The Hamiltonian function represents the total energy. For the system state vector The partial derivatives; The energy storage coefficient matrix is ​​composed of the inductance and capacitance values ​​in the converter circuit; The control input matrix represents the weights of the influence of external control variables on the system state; This represents the control input variable, specifically the switching modulation duty cycle signal of the converter bridge arm.

4. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S201. Input the collected electromagnetic state parameters and environmental feature data into the physical sensing neural Hamiltonian operator, and extract the current operating condition features of the system based on the nonlinear mapping characteristics of the physical sensing neural Hamiltonian operator. S202. The operating condition feature vector is mapped to an energy dissipation potential scalar reflecting the intensity of nonlinear disturbances in the system through the output layer of the operator, and the partial derivative of the energy dissipation potential scalar with respect to the electromagnetic state parameters is obtained to generate a dynamically corrected feature vector. S203. Map the dynamically corrected eigenvectors to loss correction terms in the form of a diagonal matrix, and reconstruct the dissipation matrix in the Hamiltonian state-space model using the loss correction terms to obtain the corrected dissipation matrix.

5. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 4, characterized in that, In step S203, the corrected dissipation matrix is ​​obtained by the following formula: ； Among them, the loss correction term The calculation formula is: ； ； ； Among them, the This represents the corrected dissipation matrix, used to update the dissipation properties in the Hamiltonian state-space model; the... Represents the initial dissipation matrix; the This represents a loss correction term, used to quantify the additional energy loss caused by nonlinear disturbances; the... Represents the diagonalization operator; the This represents the dynamically corrected feature vector, reflecting the loss correction intensity of each dimension's state components; the... Represents the energy dissipation potential scalar. For the system state vector The partial derivatives; This represents an energy dissipation potential scalar, used to describe the equivalent energy loss field generated by the Hamiltonian operator of the physical sensory neural network; the... The nonlinear mapping function representing the Hamiltonian operator of the physical sensory neural network; This represents the real-time temperature value in the environmental characteristic data; This represents the neuron weights.

6. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S301. Define the Riemannian metric tensor using the modified dissipation matrix to establish a geometric mapping relationship between the converter control parameter space and the system energy loss; S302. Construct the Riemannian manifold optimization space using the converter's modulation frequency and modulation ratio as coordinate axes, and determine the curvature distribution characteristics on the manifold surface based on the Riemannian metric tensor; S303. Perform gradient search along the geodesic path where the energy function decreases within the Riemannian manifold optimization space to obtain the optimal control vector that minimizes the system's energy loss.

7. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 6, characterized in that, In step S303, the specific operation process for performing gradient search to obtain the optimal control vector includes the following sub-steps: S3031. Calculate the Riemann connection coefficients based on the Riemann metric tensor to describe the rotation law and geometric connection relationship of the inscribed space of the Riemann manifold optimization space; S3032. Construct a second-order geodesic differential equation about the coordinates of the control parameters using the Riemann connection coefficient, and solve it to obtain the evolution trajectory of the control parameters in the manifold space; S3033. Project the evolution trajectory onto the coordinate points of the manifold surface through exponential mapping, and output the optimal control vector that conforms to the physical constraints of the system.

8. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 7, characterized in that, In step S3031, the Riemannian metric tensor is constructed using the following formula: The control vector update formula in the geodesic path search is as follows: Among them, the The denot represents the Riemannian metric tensor, used to define the metric properties of the optimization space of the Riemannian manifold; the denot represents... This represents a coordinate vector of control parameters consisting of modulation frequency and modulation ratio; Represents the corrected dissipation matrix For the control parameter coordinate vector The Jacobian matrix; This represents the coordinate vector of the control parameters obtained from the search at the next moment; the... The coordinate vector representing the control parameters at the current moment; The step size factor represents the optimization step size; The matrix representing the inverse of the Riemannian metric tensor G is used to correct the gradient descent direction to avoid high-loss regions; This represents the gradient of the total energy Hamiltonian function defined in step S1 with respect to the system state vector.

9. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S401. Using the optimal control vector as the policy constraint boundary, construct a heterogeneous dual-agent game model that includes an energy-efficient agent and a voltage-stabilizing agent; S402. Perform sliding window sampling on the bus voltage sequence and obtain the information entropy feature value by calculating the approximate entropy of the voltage signal after reconstruction in phase space; S403. Determine the disturbance level of the system based on the relationship between the information entropy characteristic value and the preset instability threshold, decide the switching logic of the converter between the efficiency priority operation mode and the voltage support operation mode, and output the active power reference value and the reactive power reference value.

10. A method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 9, characterized in that, In step S402, the specific operation process for obtaining the information entropy feature value is as follows: S4021. Perform normalization processing on the bus voltage sequence and map the one-dimensional voltage sequence into a multi-dimensional phase space vector sequence according to the preset embedding dimension; S4022. Calculate the maximum absolute distance between any two vectors in a multidimensional phase space vector sequence, count the number of vector pairs whose distance is less than the similarity threshold, and obtain the association measure function; S4023. Calculate the probability of generating new patterns when the vector dimension increases by using logarithmic deviation operation, and output an approximate entropy characteristic value that characterizes the complexity of bus voltage fluctuations.

11. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 9, characterized in that, In step S401, the Nash equilibrium solution formula for the heterogeneous two-agent game model is: ； The decision criterion formula for the operating mode is as follows: ； Wherein, the approximate entropy eigenvalue The calculation formula is: ； Among them, the This represents a defined active power reference value; the... This represents a defined reactive power reference value; the aforementioned The weighting coefficients represent the energy efficiency of the agent; The efficiency cost function is determined by the corrected dissipation matrix in step S2; The weighting coefficients of the voltage-stabilized intelligent agent; The voltage stability cost function is determined by the deviation between the bus voltage and the reference value; Indicates the operating modes of the converter; the The information entropy feature value is used to quantify the degree of disorder in the voltage sequence; the... Indicates the preset instability threshold; the Represents the logarithmic correlation measure; the The embedding dimension of the vector reconstruction; This represents the similarity threshold.

12. The method for dynamic safety control of a converter in a photovoltaic-storage-charging system as described in claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S501. Combining active power reference values, reactive power reference values, and the operating mode of the decision, a pulse width modulation drive sequence containing dynamic dead zone compensation and frequency conversion characteristics is generated through a mapping function; S502. Drive the converter to execute the pulse width modulation drive sequence, and obtain the bus voltage and inductor current after execution as the actual response value in real time, and simultaneously calculate the state prediction value of the physical sensing neural Hamiltonian operator under the current operating condition. S503. Construct a deviation functional between the actual response value and the state prediction value, and combine the deviation functional to iteratively correct the weight parameters of the physical sensing neural Hamiltonian operator through a backpropagation mechanism, thereby realizing the closed-loop self-evolution of the control algorithm.

Images