A multimodal neural network driven power conversion system adaptive control method and system

Abstract

The present invention relates to the field of adaptive control technology, and more specifically, to a method and system for adaptive control of a power conversion system driven by a multimodal neural network. The method comprises the following steps: extracting high-order transient eigenvectors of a load path and a voltage path based on a dual-path nonlinear coupling model, and inputting these eigenvectors into a multimodal neural network for feature fusion in combination with an aging index of the power supply to generate a preliminary switching frequency requirement currently required by the power supply; evaluating heat dissipation capacity and inputting the heat dissipation capacity into a preset three-state heat dissipation capacity model to determine the heat dissipation state; performing temperature compensation and amplitude limiting on the preliminary switching frequency requirement, and outputting a corrected switching frequency requirement; the method and system can dynamically adjust the frequency range based on the real-time heat dissipation capacity, avoiding overheating when heat dissipation is insufficient, maintaining efficient operation when heat dissipation is sufficient, and suppressing the risk of magnetic saturation and avoiding frequency runaway in an emergency.

Description

Technical Field

[0001] The present invention relates to the field of adaptive control technology, and in particular to a method and system for adaptive control of a power conversion system driven by a multimodal neural network. Background Art

[0002] Switching frequency is a key parameter for optimizing power conversion system performance. Its magnitude directly impacts key system metrics such as efficiency, stability, size, and electromagnetic interference. In the field of power electronics, dynamic adjustment of the switching frequency is a key technology for achieving high-efficiency energy conversion and reliable operation. The switching frequency directly impacts the system's dynamic response capability. For example, in scenarios with sudden load changes or voltage fluctuations, fixed-frequency control strategies struggle to adapt quickly, potentially leading to unstable output voltage or overloading of power devices. Furthermore, the switching frequency is closely related to the temperature rise of power devices. High-frequency operation increases switching losses, leading to increased junction temperature and shortened device life. While low-frequency operation reduces temperature rise, it may not meet the heat dissipation requirements of high-power scenarios. Therefore, heat dissipation control is a critical and essential aspect of power conversion system design.

[0003] However, existing heat dissipation control mostly uses fixed thresholds (such as temperature thresholds) and cannot dynamically adjust the frequency range according to the real-time heat dissipation capacity, resulting in system overheating when heat dissipation is insufficient or reduced efficiency when heat dissipation is excessive. Therefore, a multimodal neural network-driven power conversion system adaptive control method and system are provided. Summary of the Invention

[0004] The object of the present invention is to provide a multimodal neural network driven power conversion system adaptive control method and system to solve the problem proposed in the above background technology that the frequency range cannot be dynamically adjusted according to the real-time heat dissipation capacity, resulting in system overheating when the heat dissipation is insufficient or reduced efficiency when the heat dissipation is excessive.

[0005] To achieve the above object, the present invention provides a method for adaptive control of a power conversion system driven by a multimodal neural network, comprising the following steps:

[0006] S1. Real-time acquisition of key thermal parameters, load change rate, and voltage fluctuation energy of power devices. Processing of load change rate and voltage fluctuation energy based on a dual-path nonlinear coupling model to extract high-order transient characteristic vectors of the load path. and the high-order transient eigenvectors of the voltage path , wherein the dual path includes a load path and a voltage path;

[0007] S2, the high-order transient characteristic vector of the load path and the high-order transient eigenvectors of the voltage path Combined with the aging index of the power supply The data is input into a multimodal neural network for feature fusion to generate the initial switching frequency requirement of the power supply. ;

[0008] S3. Evaluate the heat dissipation capacity based on key thermal parameters, input the heat dissipation capacity into the preset heat dissipation capacity three-state model to determine the heat dissipation state, and make preliminary switching frequency requirements based on the heat dissipation state. Perform temperature compensation and limit processing to output the corrected switching frequency requirement ;

[0009] S4, based on switching frequency requirements Generate a switching frequency control instruction and input the switching frequency control instruction into the pulse width modulation controller to adaptively adjust the switching frequency of the power device.

[0010] As a further improvement of the present technical solution, the key thermal parameters include junction temperature, ambient temperature, fan speed, air flow speed and heat sink temperature.

[0011] As a further improvement of this technical solution, in S1, the load change rate and voltage fluctuation energy are processed based on the dual-path nonlinear coupling model to extract the high-order transient characteristic vector of the load path. and the high-order transient eigenvectors of the voltage path The specific steps are as follows:

[0012] S11. The normalized load change rate is obtained based on the ratio of the real-time load change rate to the historical maximum load change rate. , based on voltage fluctuation energy In the sliding window The ratio of the variance calculated internally to the reference energy is used to obtain the normalized voltage energy fluctuation. ;

[0013] S12, according to the normalized load change rate and voltage energy fluctuations By introducing the ambient temperature and electromagnetic interference power into the dual-path nonlinear coupling model, the load path is subjected to a first-order nonlinear transformation and a second-order coupling denoising operation to obtain the intermediate value of the load change rate. By introducing the harmonic distortion energy and electromagnetic interference power generated during the switching process through the dual-path nonlinear coupling model, the voltage path is subjected to a first-order nonlinear transformation and a second-order coupling denoising operation to obtain the intermediate value of the voltage energy fluctuation. ;

[0014] S13, the intermediate value of the load change rate output by the two paths and the intermediate amount of voltage energy fluctuation , get the interaction effect value between load and voltage ;

[0015] S14, based on the interaction effect value between load and voltage , the intermediate value of load change rate and the intermediate amount of voltage energy fluctuation , extract the high-order transient eigenvector of the load path and the high-order transient eigenvectors of the voltage path .

[0016] As a further improvement of this technical solution, in S2, by considering the influence of temperature on the material failure rate in the stress events of the load path and the voltage path, the total damage value caused by the load path and the total damage value caused by the voltage path are obtained, thereby obtaining the overall aging index of the power supply. .

[0017] As a further improvement of this technical solution, in S2, the initial switching frequency requirement currently required by the power supply is generated The specific steps are as follows:

[0018] S21, will Real-time junction temperature on splicing and device aging index Get vector ,Will Real-time junction temperature on splicing and device aging index Get vector , the vector Send it to the lightweight fully connected network load branch, and transform the vector It is fed into the voltage branch of the lightweight fully connected network. Both branches undergo linear transformation and residual activation to output the intermediate embedding vector and ;

[0019] S22, obtain the deviation between the last actual switching frequency and the target switching frequency , and the deviation As query information, and then calculate separately and The corresponding key vector and value vector are used to calculate the attention coefficient of the two paths based on the similarity between the query and the key vector. and Finally, according to the preset ratio and Weighted fusion features ;

[0020] S23, using the first-order Kalman filter to fusion features Perform forecast updates to generate smoothed frequency estimates ;

[0021] S24, frequency estimation and fusion features The mapped switching frequency update amount is weighted to obtain the final preliminary switching frequency requirement. .

[0022] As a further improvement of this technical solution, in S3, the specific steps of evaluating the heat dissipation capacity based on the key thermal parameters are as follows:

[0023] S31. Get device junction temperature and ambient temperature The temperature difference is within the maximum safe temperature difference The proportion of the temperature difference to obtain the relationship between the temperature difference and the heat dissipation index;

[0024] S32, the actual contact thermal resistance With the ideal reference thermal resistance Make a ratio to simulate the degree to which the increase in thermal resistance weakens the thermal conductivity;

[0025] S33, increase the airflow convection gain and fan speed Perform linear superposition and obtain the heat dissipation capacity value ;

[0026] S34. In the process of obtaining the heat dissipation capacity, the influence of the radiator surface area is considered and optimized to obtain the optimized heat dissipation capacity value. .

[0027] As a further improvement of the present technical solution, in S3, the three-state model of heat dissipation capability includes a normal state, an active heat dissipation state, and an emergency state. The specific rules for determining the heat dissipation state are as follows:

[0028] Pre-set normal state threshold and emergency thresholds ;

[0029] when , then it is in normal state; output ; The upper and lower limits are: ; Where, is the lower limit of the switching frequency; is the upper limit of the switching frequency; is the switching frequency of the system in the lowest safe operating state; is the rated switching frequency of the system when the heat dissipation capacity is sufficient;

[0030] when , it is in active heat dissipation state, adjusts the frequency upper limit and temperature compensation, and then outputs the switching frequency requirement;

[0031] when , it is in an emergency state, triggering magnetic field distortion detection and load mutation response, while constraining the upper and lower limits of the frequency, and outputting the switching frequency requirement after limiting.

[0032] As a further improvement of this technical solution, the active heat dissipation state adjusts the frequency upper limit and performs temperature compensation, and then outputs the switching frequency requirement specifically as follows:

[0033] Based on heat dissipation capacity The real-time value and the rate of decrease are adjusted by the nonlinear exponential upper limit of the frequency. ;

[0034] Based on ambient temperature and radiator dust accumulation coefficient Real-time update of compensation gain , combining the temperature deviation and the effect of load current on temperature rise to obtain the temperature compensation value , and then get the output .

[0035] As a further improvement of the present technical solution, the emergency state triggers magnetic field distortion detection and load mutation response, while constraining the upper and lower limits of the frequency, and outputting the switching frequency requirement after limiting the amplitude. The specific steps are as follows:

[0036] Detect magnetic field distortion through high-frequency harmonic injection and flux observer fusion, triggering flux deviation feedback compensation;

[0037] Input load current data to the GRU network, predict the load mutation rate, and correct it according to the load mutation rate get , combined with flux deviation feedback compensation and temperature compensation output ; At the same time, if the predicted mutation rate exceeds the threshold, the frequency limit range is reconstructed.

[0038] On the other hand, the present invention provides a multimodal neural network driven power conversion system adaptive control system, including a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein the processor executes the computer program to implement any one of the multimodal neural network driven power conversion system adaptive control methods described above.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. In this multimodal neural network-driven adaptive control method and system for a power conversion system, a dual-path nonlinear coupling model extracts high-order characteristics of the load and voltage and integrates an aging index. This allows for early identification of sudden load changes or voltage anomalies, avoiding frequency adjustment delays caused by single parameter lags, improving control accuracy under complex operating conditions, and extending equipment life.

[0041] 2. The adaptive control method and system for the power conversion system driven by the multimodal neural network utilizes a three-state model of heat dissipation capacity to dynamically adjust the frequency range. Combined with temperature compensation and multiple emergency protections, the frequency range can be dynamically adjusted according to the real-time heat dissipation capacity, avoiding overheating when heat dissipation is insufficient and maintaining efficient operation when heat dissipation is sufficient. In an emergency, magnetic field distortion detection is used to suppress the risk of magnetic saturation, and load mutation prediction is used to avoid frequency loss of control. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The figure is a flow chart of the overall method of the present invention. DETAILED DESCRIPTION

[0043] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0044] Example 1:

[0045] See also Figure 1 As shown, this embodiment provides a multi-modal neural network driven power conversion system adaptive control method, comprising the following steps:

[0046] S1. Real-time acquisition of key thermal parameters, load change rate, and voltage fluctuation energy of power devices. Processing of load change rate and voltage fluctuation energy based on a dual-path nonlinear coupling model to extract high-order transient characteristic vectors of the load path. and the high-order transient eigenvectors of the voltage path , wherein the dual path includes a load path and a voltage path;

[0047] Key thermal parameters include junction temperature, ambient temperature, fan speed, airflow rate, and heat sink temperature;

[0048] In S1, the load change rate and voltage fluctuation energy are processed based on the dual-path nonlinear coupling model to extract the high-order transient characteristic vector of the load path. and the high-order transient eigenvectors of the voltage path The specific steps are as follows:

[0049] S11. Based on the ratio of the real-time load change rate to the historical maximum load change rate, the normalized load change rate is obtained. , based on voltage fluctuation energy In the sliding window The ratio of the calculated variance to the reference energy is the normalized voltage energy fluctuation. The purpose is to map the original physical quantity to a unified dimension interval [0,1], thereby eliminating the scale differences caused by different magnitudes and ensuring the comparability and numerical stability of each input component in subsequent nonlinear operations.

[0050] S12, according to the normalized load change rate and voltage energy fluctuations By introducing the ambient temperature and electromagnetic interference power into the dual-path nonlinear coupling model, the load path is subjected to a first-order nonlinear transformation and a second-order coupling denoising operation to obtain the intermediate value of the load change rate. By introducing the harmonic distortion energy and electromagnetic interference power generated during the switching process through the dual-path nonlinear coupling model, the voltage path is subjected to a first-order nonlinear transformation and a second-order coupling denoising operation to obtain the intermediate value of the voltage energy fluctuation. ;

[0051] First-order nonlinear transformation of the load path: Where, It is the first-level conversion output of the load path; is the nonlinear conversion gain coefficient of the load path; is the normalized index of the load path; is the exogenous factor gain coefficient of the load path, which can adjust the contribution of ambient temperature to the first-stage conversion output; is the ambient temperature modulation coefficient, control The steepness of the curve determines the effect of ambient temperature changes on Sensitivity to impact; is the ambient temperature;

[0052] Secondary coupling operation of load paths: Where, is the coupling gain coefficient of the load path; To control electromagnetic interference in the load path to the secondary output The superposition coefficient of contribution; is the electromagnetic interference power;

[0053] First-order nonlinear transformation of the voltage path: Where, It is the first-stage conversion output of the voltage path; is the nonlinear conversion gain coefficient of the voltage path; is the normalized index of the voltage path; is the exogenous factor gain coefficient of the voltage path, which can adjust the contribution of ambient temperature to the primary conversion output; is the harmonic distortion energy generated during the switching process;

[0054] Secondary coupling operation of the voltage path: Where, is the coupling gain coefficient of the voltage path; To control electromagnetic interference in the voltage path to the secondary output The superposition coefficient of contribution;

[0055] Ambient temperature fluctuations will change the switching characteristics of power devices and sensor response curves; electromagnetic interference and harmonics will produce sharp glitches or periodic disturbances at the sampling point, so on the load path, and ambient temperature and electromagnetic interference power are input into the dual-path nonlinear coupling model; on the voltage path, the system and harmonic distortion energy and Synchronous input of the dual-path nonlinear coupling model is equivalent to online calibration of the deviation caused by the non-ideal environment and hardware, so that the intermediate quantity and It is closer to real physical mutations and can significantly suppress false mutations caused by environmental and electromagnetic interference, avoiding responding to noise as real load or voltage mutations.

[0056] S13, the intermediate value of the load change rate output by the two paths and the intermediate amount of voltage energy fluctuation , get the interaction effect value between load and voltage ; Where, is the cross-coupling weight, which controls the interaction characteristics generated by multiplying the outputs of the two paths The strength of the proposed method is to capture the interaction between how load mutations affect voltage disturbances and how voltage fluctuations are fed back into load dynamics, thereby preserving the true cross-information between the system's multi-modal states in high-order features.

[0057] S14, based on the interaction effect value between load and voltage , the intermediate value of load change rate and the intermediate amount of voltage energy fluctuation , extract the high-order transient eigenvector of the load path and the high-order transient eigenvectors of the voltage path ; ; ; The first dimension of the vector is the load path characteristic, and the second dimension is the multimodal interaction characteristic; The first dimension of the vector represents the voltage path characteristics, while the second dimension represents the multimodal interaction characteristics. This vector-level fusion preserves the core transient information of each path while also introducing multimodal interaction, ensuring that the subsequent neural network can consider both intra-path characteristics and inter-path coupling effects when making fusion decisions.

[0058] S2, the high-order transient characteristic vector of the load path and the high-order transient eigenvectors of the voltage path Combined with the aging index of the power supply The data is input into a multimodal neural network for feature fusion to generate the initial switching frequency requirement of the power supply. ;

[0059] In S2, the aging index of the power supply Specifically through:

[0060] Count the number of stress cycles for each amplitude level in the stress events of the load path , and the number of stress cycles for each amplitude level in the stress event of the voltage path ; Load change rate in each stress event and voltage fluctuation energy Perform normalization and normalize the and Nonlinear enhancement is applied separately; the influence of temperature on material failure rate is considered in the stress events of the load path and the voltage path, so as to obtain the total damage value caused by the load path and the total damage value caused by the voltage path, and finally the overall aging index of the power supply is obtained. ;

[0061]

[0062] Where, is the retrieved number of load stress events; is the retrieved number of voltage stress events; for; is the Boltzmann constant ( ); For the Junction temperature when the secondary load stress event occurs; For the Junction temperature when the secondary voltage stress event occurs; is the nonlinear enhancement coefficient of the load stress event; is the nonlinear enhancement coefficient of voltage stress event;

[0063] The traditional linear damage model only counts the number of cycles and amplitude, but ignores the exponential acceleration of temperature on the material failure rate. After that, the same number and amplitude of stress cycles at high temperature will bring higher damage contribution, which is consistent with the physical aging mechanism of dielectrics. It fluctuates with load and environmental changes. Here, different weights are assigned by the real-time temperature when each stress event occurs, which can accurately reflect the difference that "in the same cycle, high temperature damages the device more, while low temperature reduces the damage."

[0064] In S2, the initial switching frequency requirement of the power supply is generated The specific steps are as follows:

[0065] S21, will Real-time junction temperature on splicing and device aging index Get vector ,Will Real-time junction temperature on splicing and device aging index Get vector , the vector Send it to the lightweight fully connected network load branch, and transform the vector It is fed into the voltage branch of the lightweight fully connected network. Both branches undergo linear transformation and residual activation to output the intermediate embedding vector and ;

[0066]

[0067]

[0068]

[0069]

[0070] Where, is the activation function of the load branch; is the weight matrix of the load branch; is the bias vector of the load branch; is the activation function of the voltage branch; is the weight matrix of the voltage branch; is the bias vector of the voltage branch;

[0071] S22, obtain the deviation between the last actual switching frequency and the target switching frequency , and the deviation As query information, and then calculate separately and The corresponding key vector and value vector are used to calculate the attention coefficient of the two paths based on the similarity between the query and the key vector. and Finally, according to the preset ratio and Weighted fusion features In this way, the model can dynamically determine whether to rely more on load information or voltage information under the current working conditions;

[0072] The deviation As query information:

[0073]

[0074] Where, is the query vector; To query the weight matrix, the concatenated vector is mapped to the attention query space;

[0075] Attention coefficients of the two paths and :

[0076]

[0077]

[0078] Where, is the key vector of the load branch; is the key vector of the load branch; is the length of the query; g is the attention coefficient of the load branch; is the attention coefficient of the voltage branch;

[0079]

[0080] Where, is the value vector of the load branch; is the value vector of the voltage branch;

[0081] In this way, when the last over- or under-response to a load mutation occurs, the attention mechanism can automatically increase the proportion of the load or voltage path, improve the system convergence speed and stability, suppress the jitter or response delay caused by the dominance of a single path, and achieve a smoother control transition.

[0082] S23. In order to suppress the high-frequency jitter of the network output and take into account the timely response of the system to mutations, the system uses a first-order Kalman filter to filter the fused features after fusing the feature maps. Perform forecast updates to generate smoothed frequency estimates Specifically, the frequency estimate after the previous filtering is retained and a fixed process noise uncertainty is added. The frequency suggestion mapped by the network is merged with the predicted value and corrected according to the ratio of measurement noise to process noise (Kalman gain).

[0083] This enables a quick response to sudden changes in load or voltage, while ensuring the continuity and consistency of instructions under stable working conditions, filtering out random fluctuations in the fusion network output, and avoiding excessive jumps in control instructions.

[0084] S24, frequency estimation and fusion features The mapped switching frequency update amount is weighted to obtain the final preliminary switching frequency requirement. ;

[0085] In order to make the system more dependent on the new prediction of the network in extreme environments or when the device is seriously aged, and more dependent on the smoothness of the filter under normal conditions, the fusion coefficient is designed. (Automatically adjusted with temperature and aging degree), according to the formula:

[0086]

[0087] Calculate the preliminary switching frequency requirement; among them, It is another small fully connected network that maps the fusion features to frequency increments. Mapping to frequency update amount;

[0088] S3. Evaluate the heat dissipation capacity based on key thermal parameters, input the heat dissipation capacity into the preset heat dissipation capacity three-state model to determine the heat dissipation state, and make preliminary switching frequency requirements based on the heat dissipation state. Perform temperature compensation and limit processing to output the corrected switching frequency requirement ;

[0089] In S3, the specific steps for evaluating heat dissipation capacity based on key thermal parameters are as follows:

[0090] S31. Get device junction temperature and ambient temperature The temperature difference is within the maximum safe temperature difference The proportion of the temperature difference to obtain the relationship between the temperature difference and the heat dissipation index;

[0091] S32, the actual contact thermal resistance With the ideal reference thermal resistance Make a ratio to simulate the degree to which the increase in thermal resistance weakens the thermal conductivity;

[0092] S33, increase the airflow convection gain and fan speed Perform linear superposition to obtain the heat dissipation capacity value ;

[0093]

[0094] Where, The airflow efficiency coefficient converts the measured airflow velocity into a linear gain weight on the heat dissipation capacity, which depends on the equipment duct and fin structure; is the fan speed gain coefficient, which is the weight of the fan speed normalized and mapped to the heat dissipation gain, reflecting the contribution of blade motion to enhancing airflow and turbulence; The rated speed of the fan is the maximum or optimal operating speed specified by the fan manufacturer;

[0095] S34. In the process of obtaining the heat dissipation capacity, the influence of the radiator surface area is considered and optimized to obtain the optimized heat dissipation capacity value. ;

[0096] The heat sink surface area is directly mapped to the geometric expansion area by combining the driver feedback (encoder / potentiometer) of the movable fins. That is, the rate of change of deformation is obtained by comparing the difference between the current effective heat transfer surface area of the heat sink and the nominal reference heat transfer surface area.

[0097]

[0098] Where, The effective heat transfer surface area of the current heat sink (or fin) (m²), including changes due to deformation, fin angle, or module assembly differences; is the nominal reference heat transfer surface area; is the surface area gain coefficient (dimensionless), which is used to adjust the linear improvement weight of surface area change on heat dissipation capacity;

[0099] In dynamic or adjustable heat dissipation systems (such as shape memory alloy fins, variable geometry ducts), the fin expansion angle and arrangement will directly change the heat transfer area. After that, the model can perceive the geometric changes of the hardware structure in real time instead of static assumptions, avoiding misjudgment caused by ignoring the changes in the radiator's own state when estimating the heat exchange capacity simply by wind speed or rotation speed.

[0100] In S3, the three-state model of heat dissipation capability includes normal state, active heat dissipation state, and emergency state. The specific rules for determining the heat dissipation state are as follows:

[0101] Pre-set normal state threshold and emergency thresholds ;

[0102] when , then it is in normal state; output ;Sufficient heat dissipation, no need for limiting, the upper and lower limits are: ; Where, is the lower limit of the switching frequency; is the upper limit of the switching frequency; is the switching frequency of the system in the lowest safe operating state; is the rated switching frequency of the system when the heat dissipation capacity is sufficient;

[0103] The purpose of the normal state setting is to keep the system running stably when the heat dissipation capacity is sufficient, using the rated frequency , ensure efficiency and security, reduce unnecessary resource consumption, and maintain the system in optimal working condition.

[0104] when , it is in active heat dissipation state, adjusts the frequency upper limit and temperature compensation, and then outputs the switching frequency requirement;

[0105] Active heat dissipation adjusts the frequency upper limit and temperature compensation, and then outputs the switching frequency requirements as follows:

[0106] Based on heat dissipation capacity The real-time value and the rate of decrease are adjusted by the nonlinear exponential upper limit of the frequency. ;

[0107]

[0108] Where, is the dynamic attenuation coefficient; It is a nonlinear index and is adjusted according to the load type;

[0109] Based on ambient temperature and radiator dust accumulation coefficient Real-time update of compensation gain , combining the temperature deviation and the effect of load current on temperature rise to obtain the temperature compensation value , and then get the output ;in, ;The lower frequency limit is: ; The upper frequency limit is: ;

[0110]

[0111]

[0112] Where, is the temperature compensation value; is the compensation gain; is the device safety junction temperature threshold; is the temperature compensation threshold; is the load current currently output by the power supply system; is the rated output current of the power supply system; For the gradient boosting decision tree model, input , output dynamic gain coefficient ,Through historical data learning, the optimal compensation intensity is generated under different environments, dust accumulation and temperature rise rates; Ability to quantify the extent to which dust accumulation on the radiator surface weakens the heat dissipation capacity;

[0113] Nonlinear exponential adjustment frequency cap Helps to quickly respond to real-time changes in cooling capacity to avoid overheating. Temperature compensation Taking into account the effects of ambient temperature and dust accumulation, the switching frequency can be adjusted more accurately to prevent malfunctions caused by excessive temperatures. This can improve the adaptability and stability of the system, especially when the environment changes or heat dissipation performance decreases.

[0114] when , it is in an emergency state, triggering magnetic field distortion detection and load mutation response, while constraining the upper and lower limits of the frequency, and outputting the switching frequency requirement after limiting;

[0115] The specific steps for triggering magnetic field distortion detection and load mutation response in an emergency state, constraining the upper and lower limits of the frequency, and outputting the switching frequency requirements after limiting are as follows:

[0116] Detect magnetic field distortion through high-frequency harmonic injection and flux observer fusion, triggering flux deviation feedback compensation;

[0117] Specifically, it uses high-frequency harmonic injection and flux observer fusion detection to superimpose a weak test signal of a specific frequency on the switching frequency. The response is collected by a current sensor, and the amplitude of the injected harmonic is extracted using an adaptive notch filter. If the amplitude exceeds the threshold, the magnetic field is determined to be distorted, triggering frequency avoidance and flux feedback compensation.

[0118]

[0119] Where, is the flux deviation feedback compensation; is the proportional gain coefficient; is the magnetic flux offset; is the integral gain coefficient;

[0120] , By heat dissipation capacity Adjustment:

[0121]

[0122]

[0123] Where, is the proportional gain reference value; is the integral gain reference value;

[0124] Input load current data to the GRU network, predict the load mutation rate, and correct it according to the load mutation rate get , combined with flux deviation feedback compensation and temperature compensation output ; At the same time, if the predicted mutation rate exceeds the threshold, the frequency limit range is reconstructed, where ;The lower frequency limit is: ; The upper frequency limit is: ;

[0125]

[0126]

[0127] Where, is the feedforward coefficient, which is calibrated online by the system inertia parameters;

[0128]

[0129] Where, Adjust the gain for feedforward;

[0130] High-frequency harmonic injection and flux observer can detect potential magnetic field problems and provide timely feedback compensation to prevent system instability. GRU network predicts load mutation rate and corrects initial frequency This allows for rapid adjustments when the load changes suddenly, avoiding overload or performance degradation. At the same time, reconstructing the frequency limit range ensures that the system can still operate safely under extreme conditions, enhancing the system's robustness, responding to emergencies, and ensuring equipment safety.

[0131] S4, based on switching frequency requirements Generate a switching frequency control instruction and input the switching frequency control instruction into the pulse width modulation controller to adaptively adjust the switching frequency of the power device.

[0132] Example 2:

[0133] This embodiment provides a multimodal neural network driven power conversion system adaptive control system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement any one of the above-mentioned multimodal neural network driven power conversion system adaptive control methods.

[0134] The basic principles, main features, and advantages of the present invention are shown and described above. It should be understood by those skilled in the art that the present invention is not limited to the above-described embodiments. The above-described embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention claimed.

Claims

1. A multimodal neural network driven power conversion system adaptive control method, characterized in that: The following steps are involved: S1. Real-time acquisition of key thermal parameters, load change rate, and voltage fluctuation energy of power devices. Based on the dual-path nonlinear coupling model, the load change rate and voltage fluctuation energy are processed to extract the high-order transient characteristic vector X of the load path. L and the high-order transient eigenvector X of the voltage path V , where the dual path includes the load path and the voltage path, and the high-order transient characteristic vector X of the load path is extracted L and the high-order transient eigenvector X of the voltage path V The specific steps involved are as follows: S11. Based on the ratio of the real-time load change rate to the historical maximum load change rate, the normalized load change rate A is obtained. norm , based on the voltage fluctuation energy E v In the sliding window Δ t The ratio of the variance calculated internally to the reference energy is used to obtain the normalized voltage energy fluctuation E norm ; S12, according to the normalized load change rate A norm and voltage energy fluctuation E norm By introducing the ambient temperature and electromagnetic interference power through the dual-path nonlinear coupling model, the load path is subjected to a first-order nonlinear transformation and a second-order coupling denoising operation to obtain the intermediate quantity L2 of the load change rate. By introducing the harmonic distortion energy and electromagnetic interference power generated during the switching process through the dual-path nonlinear coupling model, the voltage path is subjected to a first-order nonlinear transformation and a second-order coupling denoising operation to obtain the intermediate quantity V2 of the voltage energy fluctuation. S13, obtaining an interaction effect value C between the load and the voltage based on the intermediate value L2 of the load change rate and the intermediate value V2 of the voltage energy fluctuation output by the two paths; S14, based on the interaction effect value C between load and voltage, the intermediate value L2 of load change rate and the intermediate value V2 of voltage energy fluctuation, extract the high-order transient characteristic vector X of the load path L and the high-order transient eigenvector X of the voltage path V ; S2, the high-order transient characteristic vector X of the load path L and the high-order transient eigenvector X of the voltage path V Combined with the power supply aging index A age The two are input into a multimodal neural network for feature fusion to generate the preliminary switching frequency requirement F0 required by the power supply. S3. Evaluate heat dissipation capability based on key thermal parameters, input the heat dissipation capability into a preset three-state heat dissipation capability model to determine a heat dissipation state, perform temperature compensation and limit the initial switching frequency requirement F0 based on the heat dissipation state, and output a corrected switching frequency requirement F1; S4. Generate a switching frequency control instruction based on the switching frequency requirement F1, and input the switching frequency control instruction into the pulse width modulation controller to adaptively adjust the switching frequency of the power device.

2. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 1, wherein: The key thermal parameters include junction temperature, ambient temperature, fan speed, air flow rate, and heat sink temperature.

3. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 1, wherein: In S2, by considering the effect of temperature on the material failure rate in the stress events of the load path and the voltage path, the total damage value caused by the load path and the total damage value caused by the voltage path are obtained, thereby obtaining the overall aging index A of the power supply. age .

4. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 3, wherein: In S2, the specific steps of generating the preliminary switching frequency requirement F0 currently required by the power supply are as follows: S21, X L Real-time junction temperature T amb and device aging index A age Get vector X L ′, X V Real-time junction temperature T amb and device aging index A age Get vector X V ′, vector X L ′ is sent to the lightweight fully connected network load branch, and the vector X V ′ is sent to the voltage branch of the lightweight fully connected network, and both branches undergo linear transformation and residual activation to output the intermediate embedding vector h L and h V ; S22, obtain the deviation ΔF between the last actual switching frequency and the target switching frequency, and use the deviation ΔF as query information, and then calculate h L and h V The corresponding key vector and value vector are used to calculate the attention coefficient α of the two paths based on the similarity between the query and the key vector. L and α L Finally, h is L and h V Weighted fusion feature h fuse ; S23, use the first-order Kalman filter to fusion feature h fuse Perform forecast updates to generate smoothed frequency estimates S24, frequency estimation and fusion feature h fuse The mapped switching frequency update amount is weightedly calculated to obtain the final preliminary switching frequency requirement F0.

5. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 4, wherein: In S3, the specific steps of evaluating the heat dissipation capability based on the key thermal parameters are as follows: S31, obtain the device junction temperature T Now and ambient temperature T amb The temperature difference is within the maximum safe temperature difference ΔT max The proportion of the temperature difference to obtain the relationship between the temperature difference and the heat dissipation index; S32, the actual contact thermal resistance R th，cs Compared with the ideal reference thermal resistance R th，ref Make a ratio to simulate the degree to which the increase in thermal resistance weakens the thermal conductivity; S33, the airflow convection gain v flow and fan speed ω fan Perform linear superposition to obtain the heat dissipation capacity value D; S34. In the process of obtaining the heat dissipation capacity, the influence of the surface area of the radiator is considered and optimized to obtain an optimized heat dissipation capacity value D′.

6. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 5, wherein: In S3, the three-state model of heat dissipation capability includes a normal state, an active heat dissipation state, and an emergency state. The specific rules for determining the heat dissipation state are: Pre-set normal state threshold D nom and emergency threshold D crit ; When D′≥D nom , then it is in normal state; output F1=F0; the upper and lower limits are: F min =F base ; F max =F nom Where, F min is the lower limit of the switching frequency; F max is the upper limit of the switching frequency; F base F is the switching frequency of the system in the lowest safe operating state; nom is the rated switching frequency of the system when the heat dissipation capacity is sufficient; When D crit ≤D <D nom , it is in active heat dissipation state, adjusts the frequency upper limit and temperature compensation, and then outputs the switching frequency requirement; When D <D crit , it is in an emergency state, triggering magnetic field distortion detection and load mutation response, while constraining the upper and lower limits of the frequency, and outputting the switching frequency requirement after limiting.

7. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 6, wherein: The active heat dissipation state adjusts the frequency upper limit and performs temperature compensation, and then outputs the switching frequency requirement as follows: Based on the real-time value and decrease rate of heat dissipation capacity D′, the upper limit F of the frequency is adjusted by a nonlinear exponential max 1 ; Based on the ambient temperature T amb and radiator dust accumulation coefficient k dust Update the compensation gain β5 in real time, and obtain the temperature compensation value ΔF by combining the temperature deviation and the effect of load current on temperature rise comp , and then output F1.

8. The adaptive control method for a power conversion system driven by a multimodal neural network according to claim 7, wherein: The specific steps of triggering magnetic field distortion detection and load mutation response in the emergency state, constraining the upper and lower limits of the frequency, and outputting the switching frequency requirement after limiting are as follows: Detect magnetic field distortion through high-frequency harmonic injection and flux observer fusion, triggering flux deviation feedback compensation; Input the load current data to the GRU network, predict the load mutation rate, and correct F0 according to the load mutation rate to obtain Combined with the flux deviation feedback compensation and temperature compensation output F1; at the same time, if the predicted mutation rate exceeds the threshold, the frequency limit range is reconstructed.

9. A multimodal neural network driven adaptive control system for a power conversion system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the adaptive control method of a power conversion system driven by a multimodal neural network as described in any one of claims 1 to 8.

Images