Control system for PCS converter based on AI intelligent model

Abstract

The invention discloses a control system for a PCS converter based on an AI intelligent model, and relates to the technical field of converter control, and the system comprises a simulation environment module which constructs a simulation environment for basic circuit construction, and generates simulation output results in batches; the loss prediction model module constructs a loss prediction model, outputs a simulation output result in real time, and introduces a multi-layer lookup table for supplementary simulation; and the multi-objective optimization module is used for carrying out multi-objective optimization to generate a Pareto frontier solution set. According to the method, the loss, the stress and the electric energy quality of the converter under different working conditions can be accurately predicted through batch calculation and cycle operation of a simulation platform in combination with a batch normalization neural network loss prediction model, and the power quality of the converter under different working conditions can be accurately predicted through the current stress and the power loss under the LUT table structured storage parameter combination. The optimization algorithm does not need to repeatedly run complex simulation, a Pareto solution set which is finally converged is more accurate, and an optimization result is closer to a theoretical optimal solution.

Description

Technical Field

[0001] The present invention relates to the technical field of converter control, and in particular to a control system for a PCS converter based on an AI intelligent model. Background Art

[0002] With the continuous development of power electronics technology, power converters have become key equipment in renewable energy power generation systems, power quality control and power transmission systems. PCS converters play a core role in renewable energy grid connection, energy storage systems and grid stability assurance by efficiently converting electrical energy between different forms (AC and DC). Existing PCS converter control systems mainly use methods based on empirical formulas or analytical models to optimize circuit parameters.

[0003] However, traditional analytical models require a lot of simplifications and assumptions on parameters such as inductance, capacitance and switching frequency during the modeling process, which makes it difficult to fully reflect the actual losses and power stress changes in complex power grid environments. Secondly, traditional control systems usually use fixed parameter lookup tables for power compensation and loss optimization, and lack the ability to perceive and adjust the dynamic changes in the environment in real time, which makes it difficult for the converter to maintain the optimal operating state under different load or temperature conditions, reducing operating efficiency. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a control system for a PCS converter based on an AI intelligent model to solve the problem that traditional analytical models require a lot of simplifications and assumptions on parameters such as inductance, capacitance and switching frequency during the modeling process, which makes it difficult to fully reflect the actual losses and power stress changes in a complex power grid environment. Secondly, traditional control systems usually use fixed parameter lookup tables for power compensation and loss optimization, and lack the ability to perceive and adjust the dynamic changes in the environment in real time, which makes it difficult for the converter to maintain the optimal operating state under different load or temperature conditions, thereby reducing the operating efficiency.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a control system for a PCS converter based on an AI intelligent model, comprising: The simulation environment module builds a simulation environment for basic circuit construction and generates simulation output results in batches; The loss prediction model module builds a loss prediction model, outputs simulation results in real time, and introduces a multi-layer lookup table for supplementary simulation; Multi-objective optimization module, which performs multi-objective optimization to generate the Pareto frontier solution set, and generates the final Pareto solution set through the adaptive differential evolution algorithm; Dynamic power correction module, builds a dynamic power control model and dynamically adjusts the converter power output according to the parameter combination; Model update module, introducing online learning mechanism for online learning; The secure storage module generates log data for secure backup.

[0007] As a preferred solution of the control system for PCS converter based on AI intelligent model described in the present invention, wherein: the construction of simulation environment for basic circuit construction refers to the construction of simulation environment of PCS converter based on PLECS, including power module, power switch, filter inductor, capacitor model and load model; Build basic circuits and determine input voltage, output voltage, output load resistance, power switch element IGBT, freewheeling diode and LC filter; The maximum instantaneous value of the inductor current is monitored by a current sensor on the LC filter. The current peak value is extracted in each switching cycle using a simulation waveform analysis tool and the corresponding maximum current value is recorded as the current stress. ; Based on the output load resistance and the voltage probe in parallel, the output voltage waveform is recorded in real time, and the difference between the maximum and minimum values ​​of the output voltage during the switching cycle is extracted as the maximum allowable output ripple voltage. .

[0008] As a preferred solution of the control system for PCS converter based on AI intelligent model described in the present invention, wherein: the batch generation of simulation output results calculates the duty cycle based on the output voltage and the input voltage, and calculates the inductance value and the capacitance value; The power loss is calculated based on the switching loss and conduction loss respectively, which is expressed as: ; in represents the power loss, and They represent the turn-on and turn-off energy losses of the switch tube respectively. Represents the on-resistance of the switch tube. represents the switching frequency, represents the current stress, and D represents the duty cycle; Based on the simulation environment, the simulation is run cyclically, and the parameter combination is changed in batches. The output results of each round of simulation include current stress , output ripple voltage and power loss .

[0009] As a preferred solution for the control system of the PCS converter based on the AI ​​intelligent model of the present invention, wherein: the loss prediction model is constructed, and the real-time output simulation output results include the data obtained by the cyclic operation of the simulation environment, which are input characteristics and output characteristics, respectively, and the input characteristics include the input voltage , switching frequency , inductance L and capacitance C, output characteristics include current stress , output ripple voltage and power loss ; The simulation data was subjected to Z-score standardization, and a single-class support vector machine was used to detect and remove abnormal data points; Build a loss prediction model based on a batch normalization neural network, including input layer, hidden layer, and output layer; Use simulation data as the training set for model training, and select the cross entropy loss function to calculate the loss between the category probability predicted by the loss prediction model and the actual label; Use the Adam optimizer to perform gradient descent optimization and update the parameters of the loss prediction model. If the model loss no longer decreases significantly during the continuous iteration process, stop iterating and output the model parameters to update the model. Based on different input voltage , switching frequency , inductance value L and capacitance value C input loss prediction model, predict output current stress , output ripple voltage and power loss .

[0010] As a preferred solution for the control system of the PCS converter based on the AI ​​intelligent model of the present invention, the introduction of a multi-layer lookup table for supplementary simulation includes constructing a hierarchical lookup table H-LUT based on the simulation environment, wherein the first layer LUT table records different input voltages. , switching frequency , output characteristics of the inductor value L and capacitor value C combination, including current stress , output ripple voltage and power loss ; Based on the first-level LUT table, a supplementary simulation is performed to record the offset of the inductance and capacitance values ​​at different temperatures, the DC resistance DCR of the inductor, and the equivalent series resistance ESR of the capacitor as the second-level LUT table, which is expressed as: ; ; ; ; in is the DC resistance of the inductor, represents the equivalent series resistance of the capacitor, is the conductor resistivity, Indicates the number of coil turns, represents the length of the wire, A represents the cross-sectional area, represents the loss tangent, Indicates frequency, Indicates the capacitance value, and represent the inductance and capacitance of the target temperature T respectively, and represent the initial inductance and capacitance respectively, and denote the temperature coefficients of inductance and capacitance respectively, and T represent room temperature and target temperature, respectively; Traverse the parameter combinations of the first layer LUT table, record the data and store it in the lookup table matrix, and draw 3D surface graphs of current stress, voltage ripple and power loss; At the same time, the simulation is run again for different core materials and ambient temperatures, and the compensation data for different materials and temperatures in the second layer LUT table is gradually supplemented and recorded as a specific coefficient matrix: ,in Indicates the material type of inductance and capacitance; The first layer LUT table and the second layer LUT table are finally expressed as: ; ; in Indicates the first layer LUT table, Indicates the second layer LUT table; The lookup table data is re-entered into the simulation environment and parameter combinations are randomly selected for verification.

[0011] As a preferred solution of the control system for the PCS converter based on the AI ​​intelligent model of the present invention, wherein: the multi-objective optimization to generate the Pareto frontier solution set includes performing multi-objective optimization based on a hierarchical lookup table to find a parameter combination that simultaneously satisfies the minimum current stress, the lowest power loss and the minimum output voltage ripple; The constraint condition in the current stress optimization stage is the output ripple voltage Less than or equal to the maximum voltage ripple and power loss allowed by the grid standard Less than or equal to the maximum allowable loss of the converter, The constraint condition in the voltage ripple optimization stage is the current stress Less than or equal to the sum of the current stress obtained by previous optimization and the historical tolerance; The constraints in the power loss optimization phase are the current stress Less than or equal to the current stress obtained by previous optimization, while outputting the ripple voltage Less than or equal to the sum of the maximum voltage ripple allowed by the grid standard and the historical tolerance; Traverse all inductance, capacitance and frequency combinations, generate an initial population, call the first layer LUT table for performance prediction, and record each set of parameters and its output characteristics as the initial solution set; Compare the three-objective performance of each solution in the initial solution set, and select the solution that is not inferior to other solutions in any optimization objective and is better than other solutions in at least one optimization objective as the non-dominated solution; Delete the dominated solutions and keep only the non-dominated solution set as the Pareto frontier solution set.

[0012] As a preferred solution of the control system for PCS converter based on AI intelligent model described in the present invention, the method of generating the final Pareto solution set by adaptive differential evolution algorithm includes selecting non-dominated solutions as the initial population based on the Pareto frontier solution set, which is expressed as: ; in represents the i-th individual in the population, , and denote the inductance, capacitance and switching frequency of the i-th individual respectively; The fitness value of each individual is calculated according to the objective function, which is expressed as: ; in represents the fitness of the ith individual, , and They represent the current stress value, voltage ripple size and power loss under the individual i parameter combination respectively; Randomly select individuals from the population and calculate the mutation vector based on the differential evolution formula, which is expressed as: ; ; in represents the new individual generated after mutation, represents the individual with the smallest current stress in the current population, represents a fixed scaling factor, determined based on historical data, and represent the first and second population individuals respectively, represents the correction factor, represents the individual with the best performance in all objective functions in the current population, represents the average current stress value of the current iteration population, Indicates the current stress value corresponding to the individual with the smallest current stress in the current population; Based on variant individuals The current stress and power loss found in the first-layer LUT table are calculated respectively, and the additional loss is calculated using the inductor DC resistance DCR and capacitor equivalent series resistance ESR corresponding to the second-layer LUT table and superimposed to obtain the corrected power loss; Based on the current individual and variant individual, , and Generate random numbers and perform crossover operations as candidate individuals, apply the inductance temperature drift coefficient and capacitance drift coefficient recorded in the second-layer LUT table to the candidate individuals, perform temperature correction on the inductance and capacitance, and update the current stress and power loss; The fitness values ​​of candidate individuals and original individuals are calculated respectively, expressed as: ; in Represents a candidate individual The fitness value of Represents a candidate individual The current stress after temperature correction is Represents a candidate individual The output voltage ripple, Represents a candidate individual The power loss after temperature correction is It represents the power loss caused by the DC resistance of the inductor. Represents the loss caused by the equivalent series resistance of the capacitor; If the candidate individual is greater than or equal to the original individual in at least one optimization objective, the candidate individual is retained; Until the standard deviations of current stress, voltage ripple and power loss of all individuals in the population are less than the historical minimum, the most non-dominated solution set is retained as the final Pareto solution set.

[0013] As a preferred solution of the control system for PCS converter based on AI intelligent model described in the present invention, wherein: the construction of dynamic power control model, dynamically adjusting converter power output according to parameter combination includes constructing AI-PRC dynamic power control model, including input layer, feature extraction layer, neural network prediction layer and control output layer; The input layer collects grid operation data, the feature extraction layer extracts grid frequency and power fluctuation characteristics through Fourier transform, and the neural network prediction layer uses LSTM neural network to predict future grid frequency offset; The LSTM neural network of the neural network prediction layer is trained using the training set. The cross entropy loss function is selected to calculate the computational loss between the predicted category probability and the actual label. The Adam optimizer is used for gradient descent optimization to update the parameters. If the model loss no longer decreases significantly during the continuous iteration process, the iteration is stopped. The predicted frequency offset is converted into power compensation demand, and the parameter combination that can meet the power compensation amount is selected in the final Pareto solution set. The control output layer dynamically adjusts the converter power output according to the parameter combination.

[0014] As a preferred solution for the control system for PCS converter based on AI intelligent model described in the present invention, the introduction of online learning mechanism for online learning refers to introducing online learning mechanism based on dynamic power control model, recording converter output data and actual frequency response, forming a new training data set, retraining LSTM neural network, selecting cross entropy loss function to calculate the computational loss between predicted category probability and actual label, if the calculation error decreases at a rate less than the set minimum threshold in continuous iterations, the online learning is stopped.

[0015] As a preferred solution of the control system for PCS converter based on AI intelligent model described in the present invention, wherein: the generation of log data for security backup refers to the use of AI-PRC dynamic power control model to generate operation log data for each dynamic power correction and parameter switching; The log data is encrypted using the AES-256 algorithm, partition encryption is used when storing the data, and hash verification is performed. It is transmitted to the cloud and control end via the wireless network, and is regularly transferred to a remote hard disk.

[0016] The beneficial effects of the present invention are as follows: through batch calculation and cyclic operation of the simulation platform, combined with a batch normalized neural network loss prediction model, the loss, stress and power quality of the converter under different working conditions can be accurately predicted; the current stress and power loss under parameter combinations are structured and stored in the LUT table, and the optimization algorithm does not need to repeatedly run complex simulations. The Pareto solution set that converges at the end is more accurate, and the optimization result is closer to the theoretical optimal solution. The multi-objective optimization method based on differential evolution and hierarchical lookup tables can not only quickly and efficiently find the optimal solution of the converter system in current stress, voltage ripple and power loss, but also ensure the reliability and stability of the system under various temperatures and operating environments; through the dynamic power control model, the optimal solution that meets the power compensation requirements is screened from the final Pareto solution set to ensure that the overall system performance is taken into account while optimizing the power output. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0018] Figure 1 This is a structural diagram of the control system for the PCS converter based on the AI ​​intelligent model in Example 1.

[0019] Figure 2 This is a flow chart of the control system for the PCS converter based on the AI ​​intelligent model in Example 1. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a control system for a PCS converter based on an AI intelligent model, comprising the following steps: S1, build a simulation environment to build basic circuits and generate simulation output results in batches; Preferably, building a simulation environment for basic circuit construction refers to building a simulation environment for a PCS converter based on PLECS, including a power module, a power switch, a filter inductor, a capacitor model, and a load model; Build basic circuits and determine input voltage, output voltage, output load resistance, power switch element IGBT, freewheeling diode and LC filter; The maximum instantaneous value of the inductor current is monitored by a current sensor on the LC filter. The current peak value is extracted in each switching cycle using a simulation waveform analysis tool and the corresponding maximum current value is recorded as the current stress. ; Based on the output load resistance and the voltage probe in parallel, the output voltage waveform is recorded in real time, and the difference between the maximum and minimum values ​​of the output voltage during the switching cycle is extracted as the maximum allowable output ripple voltage. .

[0024] By placing a current sensor on the inductor of the LC filter and monitoring the maximum instantaneous value of the inductor current in real time, the current stress can be accurately evaluated, the output voltage ripple can be reduced, and the power quality can be improved. Accurately measuring the output voltage ripple can effectively guide the optimization of the filter parameters, ensure a stable output voltage, reduce interference to downstream equipment, and improve the power quality. By simulating and analyzing the current and voltage waveforms of the LC filter in different switching cycles, the harmonic suppression capability of the filter can be intuitively evaluated, and the inductor and capacitor parameters can be optimized. In the PLECS simulation environment, a large number of simulation experiments can be carried out under different input voltages, output loads, and switching frequencies. The response characteristics of the system under extreme load changes or grid fluctuations can be evaluated in advance. A large number of simulation results can form a database of current stress and output voltage ripple, providing training data for subsequent batch normalized neural networks (BN-NN) and dynamic power correction (AI-PRC).

[0025] Furthermore, the simulation output results are generated in batches, the duty cycle is calculated based on the output voltage and input voltage, and the inductance and capacitance values ​​are calculated, which are expressed as: ; ; Where L represents the inductance value, represents the input voltage, D represents the duty cycle, represents the switching frequency, Indicates the allowable value of inductor current ripple, Indicates the capacitance value, Represents the output current, Indicates the maximum allowable output ripple voltage; The power loss is calculated based on the switching loss and conduction loss respectively, which is expressed as: ; in represents the power loss, and They represent the turn-on and turn-off energy losses of the switch tube respectively. Indicates the on-resistance of the switch tube; Based on the simulation environment, the simulation is run cyclically, and the parameter combination is changed in batches. The output results of each round of simulation include current stress , output ripple voltage and power loss .

[0026] The duty cycle is calculated by input voltage and output voltage, and the inductance and capacitance values ​​are further calculated to ensure that the LC filter design meets the requirements of dynamic power compensation and power quality. The power loss is calculated based on the formula of switching loss and conduction loss, covering the energy loss in the switching process and the internal resistance loss during conduction. The power loss data is obtained through batch simulation, and the loss distribution of the power tube at different working points can be found to avoid the risk of overheating during system operation. In the simulation environment, the parameter combination is changed in batches through cycles, and the current stress output results of each round of simulation are recorded to ensure that the current stress constraints are met under different working conditions. The current stress is accurately calculated, and more suitable inductance values ​​and core materials can be selected. Through batch calculation and cyclic operation of the PLECS simulation platform, the automatic optimization of inductance, capacitance and switch tube parameters can be realized, which can effectively improve the design accuracy and energy efficiency of the converter, reduce power loss, optimize dynamic power regulation, improve power quality and system stability, and ultimately promote the reliability and economy of the converter system in practical applications.

[0027] S2, builds a loss prediction model, outputs simulation results in real time, and introduces a multi-layer lookup table for supplementary simulation; Preferably, a loss prediction model is constructed, and the real-time output simulation output results include input characteristics and output characteristics of the data obtained based on the cyclic operation of the simulation environment, wherein the input characteristics include the input voltage , switching frequency , inductance L and capacitance C, output characteristics include current stress , output ripple voltage and power loss ; The simulation data was subjected to Z-score standardization, and a single-class support vector machine was used to detect and remove abnormal data points; Build a loss prediction model based on a batch normalization neural network, including input layer, hidden layer, and output layer; Use simulation data as the training set for model training, and select the cross entropy loss function to calculate the loss between the category probability predicted by the loss prediction model and the actual label; Use the Adam optimizer to perform gradient descent optimization and update the parameters of the loss prediction model. If the model loss no longer decreases significantly during the continuous iteration process, stop iterating and output the model parameters to update the model. Based on different input voltage , switching frequency , inductance value L and capacitance value C input loss prediction model, predict output current stress , output ripple voltage and power loss .

[0028] By generating data in batches in a simulation environment, a batch normalized neural network (BN-NN) loss prediction model is constructed to accurately predict the loss, stress and power quality of the converter under different working conditions. By accurately predicting the power loss under different parameters, it is possible to avoid over-selection of power devices or excessive design margin, thereby reducing costs and improving system operating efficiency. By simulating and predicting the current stress and loss under different device combinations, the system can operate stably under different power grid conditions and ensure flexible adaptation to a variety of application scenarios. By eliminating abnormal data, the model can avoid being disturbed by extreme values ​​or noise, thereby improving the stability of the loss prediction model. By using batch normalization to standardize the output of the intermediate layer in the neural network, the distribution of the output of each layer is kept stable, which helps to accelerate model training and convergence.

[0029] Furthermore, introducing a multi-layer lookup table for supplementary simulation includes constructing a hierarchical lookup table H-LUT based on the simulation environment, wherein the first layer LUT table records different input voltages , switching frequency , output characteristics of the inductor value L and capacitor value C combination, including current stress , output ripple voltage and power loss ; Based on the first-level LUT table, a supplementary simulation is performed to record the offset of the inductance and capacitance values ​​at different temperatures, the DC resistance DCR of the inductor, and the equivalent series resistance ESR of the capacitor as the second-level LUT table, which is expressed as: ; ; ; ; in is the DC resistance of the inductor, represents the equivalent series resistance of the capacitor, is the conductor resistivity, Indicates the number of coil turns, represents the length of the wire, A represents the cross-sectional area, represents the loss tangent, Indicates frequency, Indicates the capacitance value, and represent the inductance and capacitance of the target temperature T respectively, and represent the initial inductance and capacitance respectively, and denote the temperature coefficients of inductance and capacitance respectively, and T represent room temperature and target temperature, respectively; Traverse the parameter combinations of the first layer LUT table, record the data and store it in the lookup table matrix, and draw 3D surface graphs of current stress, voltage ripple and power loss; At the same time, the simulation is run again for different core materials and ambient temperatures, and the compensation data for different materials and temperatures in the second layer LUT table is gradually supplemented and recorded as a specific coefficient matrix: ,in Indicates the material type of inductance and capacitance, used to describe the impact of different magnetic core or capacitor dielectric materials on circuit parameters; The first layer LUT table and the second layer LUT table are finally expressed as: ; ; in Indicates the first layer LUT table, Indicates the second layer LUT table; The lookup table data is re-entered into the simulation environment and parameter combinations are randomly selected for verification.

[0030] By recording the deviations of different temperatures, frequencies, inductance and capacitance characteristics (DCR and ESR), a more sophisticated model can be built to dynamically compensate for the deviations caused by ambient temperature changes and non-ideal device characteristics. The first layer of LUT records the basic electrical characteristics, while the second layer of LUT supplements external influencing factors such as temperature and material to achieve more accurate simulation results. The actual inductance and capacitance components will shift with temperature changes. Supplementary recording of the temperature coefficients of inductance, capacitance, DCR and ESR parameters can effectively reduce the gap between simulation and actual measurement results, so that the converter can still maintain high-precision output prediction under extreme conditions such as high and low temperatures. By introducing the characteristics of different core materials and capacitor media, the simulation system can dynamically reflect the impact of material changes on system power loss and current stress, ensuring that the system design is closer to the actual operating state; The batch simulation records of the LUT table can be directly called by the optimization algorithm, reducing the real-time simulation time, accelerating the generation process of the Pareto solution set, and realizing the rapid iteration of multi-objective optimization. The LUT table structured stores the current stress and power loss under the parameter combination. The optimization algorithm does not need to repeatedly run complex simulations. It only needs to query the lookup table to quickly evaluate the optimal parameter combination, which significantly reduces the amount of calculation; the real-time lookup table compensation mechanism reduces the loss impact of high temperature environment on inductors, capacitors and power devices, avoids device overheating or failure, and improves the overall life and reliability of the system. By simulating the response characteristics of different devices in the actual operating environment in advance, the loss of test samples and hardware can be reduced, the cost of repeated device debugging can be reduced, and the R&D efficiency can be improved. Through surface analysis, the parameter combination area that causes high power loss or drastic fluctuations in current stress can be intuitively identified, and the design scheme can be optimized to avoid high loss areas and improve system reliability. The LUT is continuously corrected through simulation feedback to improve the fitting accuracy of the lookup table, reduce errors in system operation, and ensure that the loss prediction is more consistent with the actual operation.

[0031] S3, perform multi-objective optimization to generate the Pareto frontier solution set, and generate the final Pareto solution set through the adaptive differential evolution algorithm; Preferably, performing multi-objective optimization to generate a Pareto frontier solution set includes performing multi-objective optimization based on a hierarchical lookup table to find a parameter combination that simultaneously satisfies minimum current stress, minimum power loss, and minimum output voltage ripple; The constraint condition in the current stress optimization stage is the output ripple voltage Less than or equal to the maximum voltage ripple and power loss allowed by the grid standard Less than or equal to the maximum allowable loss of the converter, The constraint condition in the voltage ripple optimization stage is the current stress Less than or equal to the sum of the current stress obtained by the previous optimization and the historical tolerance. The tolerance depends on the measurement error of the current sensor or voltage sensor and is usually set according to the sensor accuracy. The constraints in the power loss optimization phase are the current stress Less than or equal to the current stress obtained by previous optimization, while outputting the ripple voltage Less than or equal to the sum of the maximum voltage ripple allowed by the grid standard and the historical tolerance; Traverse all inductance, capacitance and frequency combinations, generate an initial population, call the first layer LUT table for performance prediction, and record each set of parameters and its output characteristics as the initial solution set; Compare the three-objective performance of each solution in the initial solution set, and select the solution that is not inferior to other solutions in any optimization objective and is better than other solutions in at least one optimization objective as the non-dominated solution; Delete the dominated solutions and keep only the non-dominated solution set as the Pareto frontier solution set.

[0032] By weighing and balancing the three optimization goals of current stress, power loss and output voltage ripple, the optimal parameter combination is found to make the converter operate in a high-efficiency state under multiple constraints. The multi-layer constraint conditions ensure that the optimization results of each stage do not conflict with each other. The goals are optimized layer by layer to reduce the optimization blind spots. The minimum current stress, minimum power loss and minimum output ripple can be achieved at the same time, avoiding the situation where single-target optimization leads to other performance degradation, and realizing system balanced design. By strictly controlling the output voltage ripple, the electromagnetic interference to the power grid and downstream equipment is reduced, and the overall power supply quality is improved. The current stress constraint is introduced in the output voltage ripple optimization stage to avoid the increase of current stress due to the reduction of ripple, so as to ensure that the system always maintains the optimization process. Stable, current stress optimization effectively prevents devices from working in extreme conditions for a long time, reduces the risk of overheating and breakdown, and extends the life of power devices and inductors and capacitors. Sensor tolerance is introduced to ensure that the optimization results are consistent with the actual operating environment, reduce the gap between theoretical and actual operating results, and improve model reliability and credibility. By comparing the three-objective performance of each solution in the solution set, non-dominated solutions are screened and inferior solutions are deleted, and only the Pareto frontier solution set is retained to achieve efficient solution of multi-objective optimization problems. A hierarchical lookup table (H-LUT) is used to predict performance by table lookup, which greatly reduces the time consumption of complex simulation calculations. Through the Pareto frontier solution set screening process, it is ensured that the final solution set has the highest performance and improves the system design accuracy.

[0033] Furthermore, the final Pareto solution set is generated by the adaptive differential evolution algorithm, including selecting non-dominated solutions as the initial population based on the Pareto frontier solution set, which is expressed as: ; in represents the i-th individual in the population, , and denote the inductance, capacitance and switching frequency of the i-th individual respectively; The fitness value of each individual is calculated according to the objective function, which is expressed as: ; in represents the fitness of the ith individual, , and They represent the current stress value, voltage ripple size and power loss under the individual i parameter combination respectively; Randomly select individuals from the population and calculate the mutation vector based on the differential evolution formula, which is expressed as: ; ; in represents the new individual generated after mutation, represents the individual with the smallest current stress in the current population, represents a fixed scaling factor, determined based on historical data, and represent the first and second population individuals respectively, represents the correction factor, represents the individual with the best performance in all objective functions in the current population, represents the average current stress value of the current iteration population, Indicates the current stress value corresponding to the individual with the smallest current stress in the current population; Based on variant individuals The current stress and power loss found in the first-level LUT table are calculated respectively, and the additional loss is calculated using the inductor DC resistance DCR and capacitor equivalent series resistance ESR corresponding to the second-level LUT table and superimposed to obtain the corrected power loss, which is expressed as: ; ; ; ; ; in Represents a variant individual The corrected power loss, Represents a variant individual Current stress in the first layer LUT table, Represents a variant individual The parameter combinations include: Represents a variant individual Power loss in the first level LUT table, Inductance The DC resistance in the second layer LUT table, It represents the power loss caused by the DC resistance of the inductor. Inductance The equivalent series resistance in the second layer LUT table, Represents the loss caused by the equivalent series resistance of the capacitor; Based on the current individual and variant individual, , and Generate random numbers and perform crossover operations as candidate individuals. Apply the inductor temperature drift coefficient and capacitor drift coefficient recorded in the second-layer LUT table to the candidate individuals, perform temperature correction on the inductor and capacitor, and update the current stress and power loss, which is expressed as: ; ; in Represents a candidate individual The current stress after temperature correction is Represents a candidate individual The power loss after temperature correction is represent the corrected inductance and capacitance respectively; The fitness values ​​of candidate individuals and original individuals are calculated respectively, expressed as: ; in Represents a candidate individual The fitness value represents the comprehensive performance of the individual in terms of current stress, voltage ripple and power loss. Represents a candidate individual The current stress after temperature correction is Represents a candidate individual The output voltage ripple, Represents a candidate individual The power loss after temperature correction is It represents the power loss caused by the DC resistance of the inductor. Represents the loss caused by the equivalent series resistance of the capacitor; If the candidate individual is greater than or equal to the original individual in at least one optimization objective (current stress, voltage ripple, and power loss), the candidate individual is retained; Until the standard deviations of current stress, voltage ripple and power loss of all individuals in the population are less than the historical minimum, the most non-dominated solution set is retained as the final Pareto solution set.

[0034] The adaptive differential evolution algorithm dynamically adjusts the mutation factor and crossover probability to improve the ability to explore the global optimal solution during the optimization process. At the same time, the fast performance prediction mechanism based on the lookup table significantly improves the computing efficiency. In the PCS converter design, this method can efficiently optimize the combination of inductance, capacitance and switching frequency to ensure the dynamic balance and global optimization of the three objectives of current stress, voltage ripple and power loss. When calculating the individual fitness, the current stress, voltage ripple and power loss are directly obtained through the first-level LUT table to avoid repeated complex simulations and improve the optimization efficiency. The second-level LUT table records additional loss parameters such as inductor DC resistance (DCR) and capacitor equivalent series resistance (ESR) to achieve precise compensation. The temperature drift correction mechanism can adjust the inductance and capacitance parameters in high or low temperature environments to avoid device losses caused by temperature increases. Increase or magnetic saturation problems, improve the long-term operation reliability of the converter, by gradually optimizing current stress, voltage ripple and power loss, the optimized solution meets the design standards under different working conditions, avoids the discrepancy between the theoretical solution and the actual operating conditions, and the Pareto solution set retains multiple optimal solutions. Designers can choose different compromise solutions according to specific application requirements to achieve a balanced solution with minimum current stress, minimum power loss or minimum voltage ripple. The crossover and correction operations gradually improve individual parameters, making the final converged Pareto solution set more accurate and the optimization result closer to the theoretical optimal solution. The multi-objective optimization method based on differential evolution and hierarchical lookup table (H-LUT) can not only quickly and efficiently find the optimal solution of the converter system in current stress, voltage ripple and power loss, but also ensure the reliability and stability of the system under various temperatures and operating environments.

[0035] S4, construct a dynamic power control model to dynamically adjust the converter power output according to the parameter combination; Preferably, constructing a dynamic power control model to dynamically adjust the power output of the converter according to the parameter combination includes constructing an AI-PRC dynamic power control model, including an input layer, a feature extraction layer, a neural network prediction layer, and a control output layer; The input layer collects grid operation data, the feature extraction layer extracts grid frequency and power fluctuation characteristics through Fourier transform, and the neural network prediction layer uses LSTM neural network to predict future grid frequency offset; The LSTM neural network of the neural network prediction layer is trained using the training set. The cross entropy loss function is selected to calculate the computational loss between the predicted category probability and the actual label. The Adam optimizer is used for gradient descent optimization to update the parameters. If the model loss no longer decreases significantly during the continuous iteration process, the iteration is stopped. The predicted frequency offset is converted into power compensation demand, and the parameter combination that can meet the power compensation amount is selected in the final Pareto solution set. The control output layer dynamically adjusts the converter power output according to the parameter combination.

[0036] By predicting the future grid frequency offset through the LSTM neural network, the AI-PRC model can perceive frequency fluctuations in advance and calculate the power compensation requirements in real time to ensure that the grid frequency remains within the set range, predict frequency offset in real time and quickly adjust the converter power output to avoid frequency offset exceeding the allowable range of the grid and ensure stable operation of the grid. By predicting in advance and dynamically adjusting power output, AI-PRC reduces power compensation delay, enables the system to have faster dynamic response capabilities, and effectively reduces power quality fluctuations. The AI-PRC model selects the optimal parameter combination from the Pareto solution set based on the predicted frequency offset, dynamically adjusts power output to meet power compensation requirements, and reduces unnecessary losses. The feature extraction layer uses Fourier transform to analyze the grid frequency and power fluctuation characteristics, and identifies power disturbance modes in complex grid environments, so that the model can operate stably under various loads and dynamic grid environments. The AI-PRC system selects the optimal solution that meets the power compensation requirements from the final Pareto solution set to ensure that the overall system performance is taken into account while optimizing power output.

[0037] S5, introduces an online learning mechanism for online learning and generates log data for security backup; Preferably, introducing an online learning mechanism for online learning means introducing an online learning mechanism based on a dynamic power control model, recording the converter output data and the actual frequency response to form a new training data set, retraining the LSTM neural network, and selecting a cross entropy loss function to calculate the computational loss between the predicted category probability and the actual label. If the computational error decreases at a rate less than a set minimum threshold in continuous iterations, the online learning is stopped.

[0038] Through the online learning mechanism, the LSTM neural network continuously receives the converter output data and the actual frequency response, and feeds back the new operating condition data to the model in real time, so as to maintain the model's adaptability to the dynamic changes of the power grid and prevent control inaccuracies caused by aging of electrical equipment or external disturbances.

[0039] Further, generating log data for security backup refers to generating operation log data of each dynamic power correction and parameter switching based on the use of the AI-PRC dynamic power control model; The log data is encrypted using the AES-256 algorithm, partition encryption is used when storing the data, and hash verification is performed. It is transmitted to the cloud and control end via the wireless network, and is regularly transferred to a remote hard disk.

[0040] The operation log data is encrypted using the AES-256 encryption algorithm. As an advanced encryption standard, AES-256 has extremely high security and anti-cracking capabilities. It can ensure the integrity and confidentiality of data even under malicious attacks. Partition encryption and hash verification are performed during the data storage stage to ensure that each partition data is independently encrypted and a hash value is generated for the encrypted data. The data integrity is automatically verified when accessed or transmitted. By transmitting it to the cloud and off-site storage devices, even if a storage device is damaged or lost, it can still be quickly restored through other backups to ensure high data availability.

[0041] In summary, the present invention uses batch calculation and cyclic operation of the simulation platform, combined with a batch normalized neural network loss prediction model, to accurately predict the loss, stress and power quality of the converter under different working conditions. The current stress and power loss under the parameter combination are structured and stored in the LUT table. The optimization algorithm does not need to repeatedly run complex simulations. The Pareto solution set that converges finally is more accurate, and the optimization result is closer to the theoretical optimal solution. The multi-objective optimization method based on differential evolution and hierarchical lookup table can not only quickly and efficiently find the optimal solution of the converter system in current stress, voltage ripple and power loss, but also ensure the reliability and stability of the system under various temperatures and operating environments. Through the dynamic power control model, the optimal solution that meets the power compensation requirements is screened from the final Pareto solution set to ensure that the overall system performance is taken into account while optimizing the power output.

[0042] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A control system for PCS converter based on AI intelligent model, characterized in that: include: The simulation environment module builds a simulation environment for basic circuit construction and generates simulation output results in batches; The loss prediction model module builds a loss prediction model, outputs simulation results in real time, and introduces a multi-layer lookup table for supplementary simulation; Multi-objective optimization module, which performs multi-objective optimization to generate the Pareto frontier solution set, and generates the final Pareto solution set through the adaptive differential evolution algorithm; Dynamic power correction module, builds a dynamic power control model and dynamically adjusts the converter power output according to the parameter combination; Model update module, introducing online learning mechanism for online learning; The secure storage module generates log data for secure backup.

2. The control system for PCS converter based on AI intelligent model as claimed in claim 1, characterized in that: The construction of the simulation environment for basic circuit construction refers to the construction of a simulation environment for the PCS converter based on PLECS, including a power module, a power switch, a filter inductor, a capacitor model and a load model; Build basic circuits and determine input voltage, output voltage, output load resistance, power switch element IGBT, freewheeling diode and LC filter; The maximum instantaneous value of the inductor current is monitored by a current sensor on the LC filter. The current peak value is extracted in each switching cycle using a simulation waveform analysis tool and the corresponding maximum current value is recorded as the current stress. ; Based on the output load resistance and the voltage probe in parallel, the output voltage waveform is recorded in real time, and the difference between the maximum and minimum values ​​of the output voltage during the switching cycle is extracted as the maximum allowable output ripple voltage. .

3. The control system for PCS converter based on AI intelligent model as claimed in claim 2, characterized in that: The batch generation simulation output results, the duty cycle is calculated based on the output voltage and the input voltage, and the inductance value and the capacitance value are calculated; The power loss is calculated based on the switching loss and conduction loss respectively, which is expressed as: ； in represents the power loss, and They represent the turn-on and turn-off energy losses of the switch tube respectively. Represents the on-resistance of the switch tube. represents the switching frequency, represents the current stress, and D represents the duty cycle; Based on the simulation environment, the simulation is run cyclically, and the parameter combination is changed in batches. The output results of each round of simulation include current stress , output ripple voltage and power loss .

4. The control system for PCS converter based on AI intelligent model as claimed in claim 3, characterized in that: The loss prediction model is constructed, and the real-time output simulation output results include input characteristics and output characteristics based on the data obtained by the cyclic operation of the simulation environment, wherein the input characteristics include the input voltage , switching frequency , inductance L and capacitance C, output characteristics include current stress , output ripple voltage and power loss ; The simulation data was subjected to Z-score standardization, and a single-class support vector machine was used to detect and remove abnormal data points; Build a loss prediction model based on a batch normalization neural network, including input layer, hidden layer, and output layer; Use simulation data as the training set for model training, and select the cross entropy loss function to calculate the loss between the category probability predicted by the loss prediction model and the actual label; Use the Adam optimizer to perform gradient descent optimization and update the parameters of the loss prediction model. If the model loss no longer decreases significantly during the continuous iteration process, stop iterating and output the model parameters to update the model. Based on different input voltage , switching frequency , inductance value L and capacitance value C input loss prediction model, predict output current stress , output ripple voltage and power loss .

5. The control system for PCS converter based on AI intelligent model as claimed in claim 4, characterized in that: The introduction of a multi-layer lookup table for supplementary simulation includes constructing a hierarchical lookup table H-LUT based on a simulation environment, wherein the first layer LUT table records different input voltages. , switching frequency , output characteristics of the inductor value L and capacitor value C combination, including current stress , output ripple voltage and power loss ; Based on the first-level LUT table, a supplementary simulation is performed to record the offset of the inductance and capacitance values ​​at different temperatures, the DC resistance DCR of the inductor, and the equivalent series resistance ESR of the capacitor as the second-level LUT table, which is expressed as: ； ； ； ; in is the DC resistance of the inductor, represents the equivalent series resistance of the capacitor, is the conductor resistivity, Indicates the number of coil turns, represents the length of the wire, A represents the cross-sectional area, represents the loss tangent, Indicates frequency, Indicates the capacitance value, and represent the inductance and capacitance of the target temperature T respectively, and represent the initial inductance and capacitance respectively, and denote the temperature coefficients of inductance and capacitance respectively, and T represent room temperature and target temperature, respectively; Traverse the parameter combinations of the first layer LUT table, record the data and store it in the lookup table matrix, and draw 3D surface graphs of current stress, voltage ripple and power loss; At the same time, the simulation is run again for different core materials and ambient temperatures, and the compensation data for different materials and temperatures in the second layer LUT table is gradually supplemented and recorded as a specific coefficient matrix: ,in Indicates the material type of inductance and capacitance; The first layer LUT table and the second layer LUT table are finally expressed as: ； ； in Indicates the first layer LUT table, Indicates the second layer LUT table; The lookup table data is re-entered into the simulation environment and parameter combinations are randomly selected for verification.

6. The control system for PCS converter based on AI intelligent model as claimed in claim 5, characterized in that: The multi-objective optimization is performed to generate a Pareto frontier solution set, including performing multi-objective optimization based on a hierarchical lookup table to find a parameter combination that simultaneously satisfies minimum current stress, minimum power loss, and minimum output voltage ripple; The constraint condition in the current stress optimization stage is the output ripple voltage Less than or equal to the maximum voltage ripple and power loss allowed by the grid standard Less than or equal to the maximum allowable loss of the converter, The constraint condition in the voltage ripple optimization stage is the current stress Less than or equal to the sum of the current stress obtained by previous optimization and the historical tolerance; The constraints in the power loss optimization phase are the current stress Less than or equal to the current stress obtained by previous optimization, while outputting the ripple voltage Less than or equal to the sum of the maximum voltage ripple allowed by the grid standard and the historical tolerance; Traverse all inductance, capacitance and frequency combinations, generate an initial population, call the first layer LUT table for performance prediction, and record each set of parameters and its output characteristics as the initial solution set; Compare the three-objective performance of each solution in the initial solution set, and select the solution that is not inferior to other solutions in any optimization objective and is better than other solutions in at least one optimization objective as the non-dominated solution; Delete the dominated solutions and keep only the non-dominated solution set as the Pareto frontier solution set.

7. The control system for PCS converter based on AI intelligent model as claimed in claim 6, characterized in that: The method of generating the final Pareto solution set by the adaptive differential evolution algorithm includes selecting a non-dominated solution as the initial population based on the Pareto frontier solution set, which is expressed as: ； in represents the i-th individual in the population, , and denote the inductance, capacitance and switching frequency of the i-th individual respectively; The fitness value of each individual is calculated according to the objective function, which is expressed as: ； in represents the fitness of the ith individual, , and They represent the current stress value, voltage ripple size and power loss under the individual i parameter combination respectively; Randomly select individuals from the population and calculate the mutation vector based on the differential evolution formula, which is expressed as: ； ； in represents the new individual generated after mutation, represents the individual with the smallest current stress in the current population, represents a fixed scaling factor, determined based on historical data, and represent the first and second population individuals respectively, represents the correction factor, represents the individual with the best performance in all objective functions in the current population, represents the average current stress value of the current iteration population, Indicates the current stress value corresponding to the individual with the smallest current stress in the current population; Based on variant individuals The current stress and power loss found in the first-layer LUT table are calculated respectively, and the additional loss is calculated using the inductor DC resistance DCR and capacitor equivalent series resistance ESR corresponding to the second-layer LUT table and superimposed to obtain the corrected power loss; Based on the current individual and variant individual, , and Generate random numbers and perform crossover operations as candidate individuals, apply the inductance temperature drift coefficient and capacitance drift coefficient recorded in the second-layer LUT table to the candidate individuals, perform temperature correction on the inductance and capacitance, and update the current stress and power loss; The fitness values ​​of candidate individuals and original individuals are calculated respectively, expressed as: ; in Represents a candidate individual The fitness value of Represents a candidate individual The current stress after temperature correction is Represents a candidate individual The output voltage ripple, Represents a candidate individual The power loss after temperature correction is It represents the power loss caused by the DC resistance of the inductor. Represents the loss caused by the equivalent series resistance of the capacitor; If the candidate individual is greater than or equal to the original individual in at least one optimization objective, the candidate individual is retained; Until the standard deviations of current stress, voltage ripple and power loss of all individuals in the population are less than the historical minimum, the most non-dominated solution set is retained as the final Pareto solution set.

8. The control system for PCS converter based on AI intelligent model as claimed in claim 7, characterized in that: The constructing of a dynamic power control model to dynamically adjust the power output of the converter according to the parameter combination includes constructing an AI-PRC dynamic power control model, including an input layer, a feature extraction layer, a neural network prediction layer, and a control output layer; The input layer collects grid operation data, the feature extraction layer extracts grid frequency and power fluctuation characteristics through Fourier transform, and the neural network prediction layer uses LSTM neural network to predict future grid frequency offset; The LSTM neural network of the neural network prediction layer is trained using the training set. The cross entropy loss function is selected to calculate the computational loss between the predicted category probability and the actual label. The Adam optimizer is used for gradient descent optimization to update the parameters. If the model loss no longer decreases significantly during the continuous iteration process, the iteration is stopped. The predicted frequency offset is converted into power compensation demand, and the parameter combination that can meet the power compensation amount is selected in the final Pareto solution set. The control output layer dynamically adjusts the converter power output according to the parameter combination.

9. The control system for PCS converter based on AI intelligent model as claimed in claim 8, characterized in that: The introduction of the online learning mechanism for online learning refers to introducing the online learning mechanism based on the dynamic power control model, recording the converter output data and the actual frequency response to form a new training data set, retraining the LSTM neural network, and selecting the cross entropy loss function to calculate the computational loss between the predicted category probability and the actual label. If the calculation error decreases at a rate less than a set minimum threshold in continuous iterations, the online learning is stopped.

10. The control system for PCS converter based on AI intelligent model as claimed in claim 9, characterized in that: Generating log data for security backup refers to generating operation log data for each dynamic power correction and parameter switching based on the use of the AI-PRC dynamic power control model; The log data is encrypted using the AES-256 algorithm, partition encryption is used when storing the data, and hash verification is performed. It is transmitted to the cloud and control end via the wireless network, and is regularly transferred to a remote hard disk.

Images