Life prediction method for key components of photovoltaic inverter based on multi-dimensional operating conditions

Abstract

The present application relates to the technical field of fault prediction and health management, in particular to a photovoltaic inverter key component life prediction method based on multi-dimensional operating conditions. Including extracting the direct current input feature set, alternating current fluctuation feature set, environmental temperature feature set and virtual observation value contained in the multi-dimensional operating data set. The AC / DC features generate component heat dissipation curves through a thermal stress decoupling model, and the integrated joint time series features are input into the life prediction network. The network numerical error term and the simulation fatigue residual term are combined to construct the network joint loss function to constrain the parameter iteration, and finally output the component aging offset. The component remaining life table is obtained from the remaining life matrix. The present application deeply penetrates through the error comparison of virtual observation value and the parameter iteration process constrained by physical mechanism, and practically improves the overall reliability and accuracy of aging evaluation conclusion under complex working conditions from the global level.

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology, specifically a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions. Background Technology

[0002] Accurate lifespan prediction and health management of key components of photovoltaic inverters under complex physical environments (such as variable sunlight shading and grid fluctuations) is an important technological application direction in the current new energy power generation field. Existing technologies generally adopt a single environmental stress monitoring combined with conventional data-driven or analytical equation prediction methods. That is, by monitoring changes in external environmental temperature or macroscopic electrical parameters, when the equipment operating time or accumulated physical stress reaches a specific threshold, the aging state and remaining lifespan of key components (such as IGBTs, capacitors, etc.) are automatically estimated and deduced within a pre-set lifespan theoretical model or digital twin system. This is currently the mainstream technical means to achieve large-scale inverter reliability assessment and early warning.

[0003] However, existing technical solutions have significant shortcomings in predictive accuracy and engineering reliability of thermodynamic control under complex operating conditions. Specifically, existing digital simulation models rely heavily on idealized assumptions. When dealing with complex external stimuli with multidimensional coupling, they cannot accurately map the actual loss state inside the equipment. As the operating cycle lengthens, the accumulated prediction error continues to amplify, gradually deviating from the actual physical aging trend. At the same time, existing feature extraction algorithms, when dealing with high-frequency transient electrical signals and long-cycle thermodynamic parameters, often suffer from the loss of key implicit features due to the lack of a deep multidimensional joint analysis mechanism. This results in extremely coarse underlying heat dissipation assessments and fatigue state calculations. Furthermore, existing lifespan extension or thermal equilibrium control measures are too one-sided. When executing thermal load intervention actions, they fail to fully consider the stringent electrical operating boundaries of the complex underlying topology of high-power inverters. Coarse control commands can easily disrupt the original electrical steady state, thereby triggering potential risks of system electromagnetic interference or deterioration of output power quality. In summary, the existing life assessment and thermal control strategies exhibit limitations such as theoretical detachment from reality and systemic inconsistency. They fail to achieve deep adaptation between multidimensional physical properties and underlying topological constraints under complex operating conditions, resulting in aging prediction results lacking engineering guidance effectiveness and derived control actions lacking safety guarantees.

[0004] To address this, a method for predicting the lifespan of key components in photovoltaic inverters based on multi-dimensional operating conditions is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions, comprising:

[0007] Collect multidimensional operation datasets of photovoltaic inverters, and extract DC input feature sets, AC fluctuation feature sets, ambient temperature feature sets, and virtual observation values ​​contained in the multidimensional operation datasets;

[0008] The simulation reference surface library output by the digital twin platform is obtained, and a thermal stress decoupling model is constructed using the simulation reference surface library. The DC input feature set and AC fluctuation feature set are input into the thermal stress decoupling model for data processing, and the component heat dissipation curves of the key inverter components are output.

[0009] The component heat dissipation curve and the ambient temperature feature set are spliced ​​together to generate a joint time series feature quantity. The joint time series feature quantity is input into the lifetime prediction network for prediction calculation to generate the initial prediction quantity of the component. The difference between the prediction quantity and the virtual observation value is calculated to generate the numerical error term of the network.

[0010] Based on the multiphysics simulation model of key inverter components, simulation fatigue residual terms are generated and combined with network numerical error terms to construct a network joint loss function. The network joint loss function is used to constrain the parameter iteration direction of the lifetime prediction network, and the lifetime prediction network outputs the component aging offset. The component aging offset is calculated using the remaining lifetime matrix to obtain the component remaining lifetime table.

[0011] Preferably, the specific generation process of the DC input feature set, the AC fluctuation feature set, the ambient temperature feature set, and the virtual observation value includes: extracting the boost current deviation and DC bus ripple based on the multidimensional operating dataset, and splicing the boost current deviation and DC bus ripple to generate the DC input feature set; extracting the voltage sag depth and reactive current surge based on the multidimensional operating dataset, and splicing the voltage sag depth and reactive current surge to generate the AC fluctuation feature set; extracting the external heat dissipation temperature and chassis internal temperature based on the multidimensional operating dataset, and splicing them to generate the ambient temperature feature set; extracting the preliminary twin junction temperature, global junction temperature feature, measured casing temperature, and conversion efficiency benchmark included in the multidimensional operating dataset; performing physical anchor point calibration calculation on the preliminary twin junction temperature using the measured casing temperature and conversion efficiency benchmark to generate the calibration node junction temperature; and splicing the calibration node junction temperature and global junction temperature feature to generate the virtual observation value.

[0012] Preferably, the specific generation process of the component heat dissipation curve includes: injecting DC power spectrum and AC distortion spectrum into the digital twin platform to synthesize a virtual composite stress spectrum; extracting the simulated node junction temperature and simulated global junction temperature of the key inverter components under the action of the virtual composite stress spectrum; performing surface fitting calculation on the simulated node junction temperature and simulated global junction temperature to generate the simulation reference surface library; constructing the thermal stress decoupling model using the simulation reference surface library, and setting the nonlinear cross matrix array of the thermal stress decoupling model; using the nonlinear cross matrix array to perform electrothermal cross-mapping calculation on the DC input feature set and the AC fluctuation feature set to obtain the AC and DC heat dissipation; and merging the AC and DC heat dissipation to obtain the component heat dissipation curve.

[0013] Preferably, the specific generation process of the joint time-series feature quantity includes: extracting the time-series thermal node data contained in the component heat dissipation curve and performing discretization and segmentation calculation to generate an off-grid heat dissipation sequence; extracting the time-series temperature node quantity contained in the ambient temperature feature set and performing discretization and segmentation calculation to generate a discrete temperature sequence quantity; performing time synchronization calculation on the off-grid heat dissipation sequence and the discrete temperature sequence quantity to obtain a synchronized heat dissipation sequence and a synchronized temperature feature quantity; using a multi-dimensional feature splicing function to perform channel merging calculation on the synchronized heat dissipation sequence and the synchronized temperature feature quantity to generate a channel fusion feature quantity; performing normalization and scaling calculation on the channel fusion feature quantity to generate a standard fusion feature quantity; and arranging the standard fusion feature quantity according to the time evolution order to generate the joint time-series feature quantity.

[0014] Preferably, the specific generation process of the network numerical error term includes: inputting the joint temporal feature quantity into the long short-term memory layer included in the lifetime prediction network for time-dependent feature calculation to generate temporal hidden state features; inputting the temporal hidden state features into the attention mechanism layer included in the lifetime prediction network for weight allocation calculation to generate weighted state feature quantity; using the mapping network layer included in the lifetime prediction network to perform nonlinear mapping calculation on the weighted state feature quantity to obtain the initial quantity of the component prediction; performing difference comparison calculation between the initial quantity of the component prediction and the health degradation state quantity included in the virtual observation to obtain the initial prediction deviation; and summing the absolute values ​​of the initial prediction deviation to obtain the network numerical error term.

[0015] Preferably, the specific construction process of the network joint loss function includes: analyzing the multiphysics simulation model to extract the thermodynamic decay law; constructing a physical constraint equation set by calling the Manson fatigue equation and the Norris fatigue equation based on the thermodynamic decay law; inputting the joint temporal characteristic quantity into the physical constraint equation set; calling the continuous differentiable rainflow operator to perform temporal dimensionality reduction processing on the joint temporal characteristic quantity to extract the discrete stress amplitude and thermal cycle mean; performing mechanism deduction calculation on the physical constraint equation set using the discrete stress amplitude and thermal cycle mean to obtain the mechanism prediction theoretical quantity; performing difference comparison calculation between the mechanism prediction theoretical quantity and the component prediction initial quantity to obtain the mechanism theory deviation quantity; performing scaling calculation on the mechanism theory deviation quantity to generate the simulation fatigue residual term; obtaining the preset network error weight and mechanism residual weight; weighting the network numerical error term and the simulation fatigue residual term using the network error weight and mechanism residual weight respectively, and combining the weighted processing result to construct the network joint loss function.

[0016] Preferably, the specific generation process of the component aging offset includes: parsing the joint loss function of the network to extract the joint loss value; using the backpropagation differentiation method to perform multidimensional differentiation calculation on the joint loss value to obtain the loss gradient descent amount; using the loss gradient descent amount to update the node weight parameters and node bias parameters included in the lifetime prediction network; determining whether the joint loss value has reached the convergence condition; when the joint loss value has reached the set convergence condition, fixing the node weight parameters and node bias parameters to obtain the fixed lifetime prediction network; inputting the joint temporal feature into the fixed lifetime prediction network for forward propagation calculation to obtain the component aging offset.

[0017] Preferably, the specific process for generating the remaining life table of the components includes: extracting the historical attenuation characteristics and initial life characteristics of the photovoltaic inverter as configured at the factory; performing cumulative damage superposition calculation on the aging offset of the components using the historical attenuation characteristics to obtain a damage state feature matrix; performing deduction calculation on the damage state feature matrix using the initial life characteristics to obtain a boost life prediction matrix, and outputting the remaining life table of the components; determining the high-fatigue boost circuit and low-fatigue boost circuit included in the inverter boost circuit based on the remaining life table of the components, and extracting the boost circuit included in the remaining life table of the components. The difference value is calculated; the current bus ripple tolerance and power tracking error of the photovoltaic inverter are extracted, and the preset imbalance safety threshold and safety boundary range are obtained; when the difference value of the boost circuit is greater than the imbalance safety threshold, and the bus ripple tolerance and power tracking error are within the safety boundary range; the multi-channel interleaved switching frequency is locked and the global current ripple is kept stable, and an asymmetric duty cycle bias is calculated; the asymmetric duty cycle bias is injected into the high fatigue boost circuit, guiding the steady-state operating current contained in the high fatigue boost circuit to transfer to the low fatigue boost circuit, and performing thermal load balancing control.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. Based on the multiphysics simulation model of key inverter components, simulation fatigue residual terms are generated and combined with the network numerical error terms to construct a joint network loss function to constrain the parameter iteration direction of the lifetime prediction network. This approach integrates the underlying thermodynamic decay mechanism with the forward propagation process of the deep network, introducing the smooth guidance of multiphysics laws in the parameter update stage. This ensures that the iteration of the network model follows the actual physical fatigue evolution law, effectively bridging the gap between the abstract algorithm model and the real physical characteristics, thus overcoming the limitation of theory being divorced from reality. This results in the final output component aging offset having a rigorous physical interpretable basis, thereby improving the engineering guidance effectiveness of the lifetime assessment results.

[0020] 2. A thermal stress decoupling model is constructed by acquiring the simulation reference surface library output by the digital twin platform. The DC input feature set and AC fluctuation feature set are then input into the thermal stress decoupling model for data processing to output the component's heat dissipation curve. This process utilizes twin surfaces to achieve smooth feature transformation across physical domains, mapping the highly nonlinearly coupled electrical signal dimension to the dissipation law of the thermal dimension. This nonlinear decoupling architecture can adapt to complex and variable operating conditions and achieve deep adaptation between multidimensional physical properties and abstract data, thereby providing a solid data logic foundation for obtaining robust aging prediction results.

[0021] 3. Extract virtual observations from multidimensional operational data, and calculate the difference between the virtual observations and the joint time-series feature quantities after inputting them into the lifetime prediction network to generate the initial component prediction quantities, thereby generating the network numerical error term. Under the condition that microscopic physical quantities in industrial sites are difficult to obtain directly, this approach constructs a calculation paradigm that guides the convergence of actual errors through virtual mapping. Combined with the calculation of component aging offset using the remaining lifetime matrix, a component remaining lifetime table is obtained, forming a complete closed loop between virtual state and physical assessment. This overcomes the limitations of neglecting one aspect in the assessment process and takes into account the underlying topological constraints, thereby ensuring that the derived equipment state intervention and control actions have sufficient data basis and security.

[0022] 4. This architecture deeply integrates the electrothermal cross-domain decoupling based on twin surfaces, the error comparison relying on virtual observations, and the parameter iteration process constrained by physical mechanisms. At the input end, it achieves smooth stripping and feature dimensionality reduction of multidimensional heterogeneous signals. In the middle of the operation, it constructs a continuous error evaluation benchmark based on virtual mapping and strongly couples the objective fatigue law of thermodynamics when the gradient converges at the bottom layer of the algorithm. Each technical node breaks down the isolation barrier between traditional pure data processing and hardware physical mechanisms, forming a self-consistent prediction chain from front-end electromagnetic condition perception to back-end remaining life estimation. This deeply integrated mechanism maintains the physical fidelity of information transmission in each link, thereby effectively improving the overall reliability and accuracy of aging assessment conclusions under complex operating conditions from a global perspective. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions, as proposed in an embodiment of this invention.

[0024] Figure 2 This is a flowchart illustrating the joint loss iteration process based on soft rainflow and multiphysics, as proposed in an embodiment of this invention.

[0025] Figure 3 This is a flowchart of the staggered parallel thermal load balancing process based on electrical hard constraint boundaries proposed in an embodiment of this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figures 1-3 The present invention provides a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions, the specific steps of which are as follows:

[0028] Collect multidimensional operation datasets of photovoltaic inverters, and extract DC input feature sets, AC fluctuation feature sets, ambient temperature feature sets, and virtual observation values ​​contained in the multidimensional operation datasets;

[0029] The simulation reference surface library output by the digital twin platform is obtained, and a thermal stress decoupling model is constructed using the simulation reference surface library. The DC input feature set and AC fluctuation feature set are input into the thermal stress decoupling model for data processing, and the component heat dissipation curves of the key inverter components are output.

[0030] The component heat dissipation curve and the ambient temperature feature set are spliced ​​together to generate a joint time series feature quantity. The joint time series feature quantity is input into the lifetime prediction network for prediction calculation to generate the initial prediction quantity of the component. The difference between the prediction quantity and the virtual observation value is calculated to generate the numerical error term of the network.

[0031] Based on the multiphysics simulation model of key inverter components, simulation fatigue residual terms are generated and combined with network numerical error terms to construct a network joint loss function. The network joint loss function is used to constrain the parameter iteration direction of the lifetime prediction network, and the lifetime prediction network outputs the component aging offset. The component aging offset is calculated using the remaining lifetime matrix to obtain the component remaining lifetime table.

[0032] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0033] Example 1

[0034] This application discloses a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions. (See attached document.) Figure 1 The specific steps proposed in this invention include: S1, collecting a multi-dimensional operation dataset of the photovoltaic inverter, and extracting the DC input feature set, AC fluctuation feature set, ambient temperature feature set, and virtual observation values ​​contained in the multi-dimensional operation dataset; S2, acquiring the simulation reference surface library output by the digital twin platform, and constructing a thermal stress decoupling model using the simulation reference surface library; inputting the DC input feature set and AC fluctuation feature set into the thermal stress decoupling model for data processing, and outputting the component heat dissipation curves of key inverter components; S3, performing data splicing processing on the component heat dissipation curves and the ambient temperature feature set. Generate joint temporal feature quantities; input the joint temporal feature quantities into the lifetime prediction network for prediction calculation to generate the initial quantity of component prediction, calculate the difference with the virtual observation value, and generate the network numerical error term; S4, generate the simulation fatigue residual term based on the multi-physics simulation model of the key inverter component, and combine it with the network numerical error term to construct the network joint loss function; use the network joint loss function to constrain the parameter iteration direction of the lifetime prediction network, and output the component aging offset through the lifetime prediction network; calculate the component aging offset through the remaining lifetime matrix to obtain the component remaining lifetime table.

[0035] Furthermore, a multi-dimensional operation dataset of the photovoltaic inverter is collected, and the DC input feature set, AC fluctuation feature set, ambient temperature feature set, and virtual observation value contained in the multi-dimensional operation dataset are extracted; corresponding to step S1 above; the specific implementation process includes:

[0036] Based on the multidimensional operating dataset, the boost current deviation and DC bus ripple are extracted, and the boost current deviation and DC bus ripple are concatenated to generate the DC input feature set; based on the multidimensional operating dataset, the voltage sag depth and reactive current surge are extracted, and the voltage sag depth and reactive current surge are concatenated to generate the AC fluctuation feature set; based on the multidimensional operating dataset, the external heat dissipation temperature and the chassis internal temperature are extracted, and the ambient temperature feature set is concatenated; the preliminary twin junction temperature, global junction temperature feature, measured casing temperature, and conversion efficiency benchmark included in the multidimensional operating dataset are extracted; the preliminary twin junction temperature is physically anchored and calibrated using the measured casing temperature and conversion efficiency benchmark to generate the calibration node junction temperature; the calibration node junction temperature and global junction temperature feature are concatenated to generate the virtual observation value.

[0037] Specifically, the generation process of the DC input feature set, AC fluctuation feature set, ambient temperature feature set, and virtual observation values ​​is as follows:

[0038] In terms of hardware configuration, a sampling sensor network is deployed on the AC / DC bus and inside the chassis of the inverter, including high-precision Hall effect voltage and current sensors and negative temperature coefficient thermistors. The sampling frequency is set to 20,000 Hz (this parameter is selected based on the Nyquist sampling theorem for high-frequency power electronic switching devices). The microcontroller and field-programmable gate array aggregate the sensor signals into a multi-dimensional operational dataset with a set time window (e.g., set to a continuous 1000 milliseconds, this parameter is calculated based on the cache capacity of the control chip and the period of low-frequency fluctuations in the power grid). This dataset is represented in memory as a two-dimensional data matrix, where columns represent independent physical sensor channels and rows represent discrete time sampling points.

[0039] The digital signal processor (DSP) reads the multidimensional runtime dataset and uses Fast Fourier Transform (FFT) and a digital bandpass filter to extract features from the DC-side current and voltage signals. The extracted boost current deviation represents the difference between the actual inductor current in the multi-channel interleaved parallel boost circuit and the theoretical reference operating point at the instant of illumination variation. The extracted DC bus ripple is the peak envelope data of the voltage pulsation in the 1500V HVDC bus system. Subsequently, a multidimensional array concatenation operator is invoked to concatenate the boost current deviation sequence and the DC bus ripple sequence column-by-column along the feature dimension, generating the DC input feature set containing two columns of time-series data.

[0040] The three-phase AC voltage and current data in the multidimensional operational dataset are analyzed to extract the voltage sag depth. This variable, calculated by a phase-locked loop, represents the percentage difference sequence of the actual voltage below the rated nominal voltage. Simultaneously, the current in the three-phase stationary coordinate system is mapped to a rotating orthogonal coordinate system using a Parker transform to extract the reactive current abrupt change, characterizing the reactive current injection intensity. The voltage sag depth and reactive current abrupt change are merged using a feature channel concatenation operator to generate the AC fluctuation feature set containing two columns of time-series features.

[0041] Based on the multidimensional operational dataset, the external heat dissipation temperature transmitted back by the probe on the surface of the external heat sink fins, and the internal temperature of the chassis monitored by the probe suspended inside the chassis, are extracted. The ambient temperature feature set is generated by concatenating these two temperature sequences.

[0042] The system retrieves the preliminary twin junction temperature from the cloud-based digital model, the measured case temperature from the physical hardware probe, the global junction temperature characteristics, and the conversion efficiency benchmark calculated from the input and output power from the multidimensional operational dataset. Using the measured case temperature and the conversion efficiency benchmark, the system performs physical anchor point calibration calculations on the preliminary twin junction temperature to generate the calibration node junction temperature.

[0043] The specific calculation process for this physical anchor point calibration is as follows: the value of the junction temperature at the calibration node is equal to the measured shell temperature plus a dynamic thermodynamic gradient compensation term. When calculating the dynamic thermodynamic gradient compensation term, firstly, the difference between the initial twin junction temperature and the measured shell temperature is calculated as the initial theoretical thermal gradient; then, this initial theoretical thermal gradient is multiplied by a dynamic heating ratio coefficient, which is calculated by dividing the current conversion efficiency benchmark by the equipment's factory-specified maximum conversion efficiency (set according to the equipment nameplate / product specification); finally, the product result is multiplied by a heat dissipation degradation compensation coefficient reflecting non-ideal heat dissipation boundary conditions to obtain the dynamic thermodynamic gradient compensation term.

[0044] The specific calculation process for the above physical anchor point calibration involves several key variables and parameters:

[0045] First, there's the preliminary twin junction temperature, which physically represents the transient temperature inside the bare crystal predicted by the twin model. This serves as the algorithm's baseline input; in this embodiment, it's 125.5 degrees Celsius, determined by the theoretical calculations from the real-time output of the digital twin platform based on the current electrical input. Second, there's the measured casing temperature, measured by a sensor mounted on the bottom of the module substrate. This serves as the physical baseline for calibrating the algorithm; in this embodiment, it's 82.0 degrees Celsius, directly determined by the local physical sensor's sampling readings. Third, there's the conversion efficiency baseline, representing the inverter's actual DC-to-AC energy conversion efficiency. In this embodiment, it's 0.982, calculated by dividing the output active power measured in real-time by the controller by the total input DC power.

[0046] In this embodiment, linear derating compensation calculation based on the measured casing temperature is performed. Utilizing the inherent temperature sensitivity of semiconductor devices, the conversion efficiency benchmark is fine-tuned by setting a base reference temperature and a derating coefficient. Specifically, firstly, the measured casing temperature extracted from the multi-dimensional operating dataset is retrieved. The standard test reference temperature is set to 25 degrees Celsius, and a fixed power device temperature derating coefficient is set according to the inverter's internal IGBT module datasheet (in this embodiment, it is set to a theoretical efficiency derating of 0.005% for every 1 degree Celsius increase). The difference between the measured casing temperature and the standard test reference temperature is calculated. If this difference is greater than zero, it is multiplied by the temperature derating coefficient to obtain the efficiency compensation offset. Subsequently, the initial conversion efficiency ratio is subtracted from this efficiency compensation offset to obtain the temperature-constrained conversion efficiency benchmark. This process uses the fundamental thermodynamic temperature coefficient to linearly correct the efficiency, improving the physical purity of the data.

[0047] In addition, a heat dissipation degradation compensation coefficient needs to be introduced to compensate for the heat dissipation degradation caused by dust accumulation on the heat sink and the drying of thermal grease. In this embodiment, it is 1.05. The specific setting is based on the service life of the inverter at the time of manufacture and is determined by dynamic addressing by consulting the experience degradation table provided by the manufacturer.

[0048] After comprehensive calculation and compensation of the above multiple parameters, the final calibration node junction temperature is obtained, which represents the true crystal junction temperature after compensation; it is then spliced ​​with the global junction temperature characteristic to generate the virtual observation value.

[0049] Physical anchor point calibration calculations are performed using measured shell temperature and conversion efficiency benchmarks. Macroscopic real physical measurements are used as benchmarks to correct the virtual state output by the twin model in real time, breaking the cyclical proof error that relies solely on simulation data, establishing the objective physical authenticity of virtual observations, and improving the credibility of input data.

[0050] Further, the simulation reference surface library output by the digital twin platform is obtained, and a thermal stress decoupling model is constructed using the simulation reference surface library; the DC input feature set and AC fluctuation feature set are input into the thermal stress decoupling model for data processing, and the component heat dissipation curves of the key inverter components are output; corresponding to step S2 above; the specific implementation process includes:

[0051] A virtual composite stress spectrum is synthesized by injecting DC power spectrum and AC distortion spectrum into the digital twin platform; the simulated node junction temperature and simulated global junction temperature of the key inverter components under the action of the virtual composite stress spectrum are extracted; the simulated node junction temperature and simulated global junction temperature are calculated by surface fitting to generate the simulation reference surface library; the thermal stress decoupling model is constructed using the simulation reference surface library, and the nonlinear cross matrix array of the thermal stress decoupling model is set; the thermal stress decoupling model uses the nonlinear cross matrix array to perform electrothermal cross-mapping calculation on the DC input feature set and the AC fluctuation feature set to obtain the AC and DC heat dissipation; the AC and DC heat dissipation are combined to obtain the heat dissipation curve of the component.

[0052] Specifically, the process of generating the component's heat dissipation curve is as follows:

[0053] Acquisition of the simulation reference surface library: The edge computing device acquires data from a remote digital twin platform. Within the digital twin platform, DC power spectra and AC distortion spectra containing various frequencies and amplitudes are pre-injected. These two stress spectra are synthesized into a virtual composite stress spectrum within the simulation space. The simulated node junction temperatures of key inverter components and the simulated global junction temperature covering the entire heat dissipation system are extracted under the influence of this virtual composite stress spectrum. The digital twin platform performs surface fitting calculations on the simulated node junction temperatures and the simulated global junction temperatures to generate the simulation reference surface library.

[0054] The execution logic of the thermal stress decoupling model and the nonlinear cross matrix array: After receiving the simulation reference surface library, the edge controller constructs a thermal stress decoupling model. This model sets a nonlinear cross matrix array, in which the main diagonal elements represent the AC / DC independent self-heating damping coefficients, and the off-diagonal elements represent the AC / DC mutual heating modulation coefficients, forming a two-dimensional numerical matrix. The specific numerical calibration method of this matrix is ​​as follows: extract N sets of equally spaced AC / DC input excitations and their corresponding junction temperature response points from the simulation reference surface library to form an overdetermined equation system. Solve the overdetermined equation system by fitting the two-dimensional surface partial derivatives using the least squares method; extract the linear main effect partial derivatives of the independent variables as the main diagonal elements, and extract the partial derivatives of the AC / DC product cross terms as the off-diagonal elements, thereby completing the numerical setting of the matrix.

[0055] The thermal stress decoupling model utilizes a nonlinear cross matrix array to perform electrothermal cross-mapping calculations on the DC input feature set and the AC fluctuation feature set. The calculation process is as follows: First, the DC input feature set and the AC fluctuation feature set are reorganized into a joint electrical stress column vector in the data dimension; then, the nonlinear cross matrix array and this joint electrical stress column vector are multiplied by a matrix inner product to obtain a basic heat dissipation vector; a steady-state dissipation bias due to ambient temperature is added to each element of this basic heat dissipation vector. Specifically, the temperature difference is calculated using the junction temperature and ambient temperature feature set values, and then divided by the ambient thermal resistance coefficient (0.85 K / W in this embodiment based on the aluminum radiator material) to obtain the static heat dissipation bias. The final comprehensive numerical column vector is the AC / DC heat dissipation.

[0056] After completing the above mapping calculation, the separate AC and DC heat dissipation components are combined and calculated in chronological order to obtain the heat dissipation curve of the component that characterizes the dynamic trajectory of the heating power.

[0057] Extract the real-time heat dissipation power corresponding to the heat dissipation curve of the component; subtract the measured shell temperature from the calibrated node junction temperature to obtain the real-time junction-shell temperature difference, and then divide the real-time junction-shell temperature difference by the real-time heat dissipation power to calculate the real-time thermal resistance value; subtract the initial nominal thermal resistance value of the device from the real-time thermal resistance value to obtain the health degradation state value used to characterize the physical aging of the device, and encapsulate it as an implicit feature into the virtual observation value.

[0058] A nonlinear cross matrix array of a thermal stress decoupling model is set up for electrothermal cross-mapping calculation; the cross thermal effect generated by AC / DC mixed stress inside the device is quantified and decomposed to avoid truncation error caused by linear superposition and provide thermodynamic input features with clear physical meaning.

[0059] Further, the component heat dissipation curve and the ambient temperature feature set are spliced ​​together to generate a joint time-series feature quantity; the joint time-series feature quantity is input into the lifetime prediction network for prediction calculation to generate the component prediction initial quantity, and the difference between the initial quantity and the virtual observation value is calculated to generate the network numerical error term; corresponding to step S3 above; the specific implementation process includes:

[0060] The time-series thermal node data contained in the component's heat dissipation curve are extracted and discretized to generate an off-grid heat dissipation sequence. The time-series temperature node data contained in the ambient temperature feature set are extracted and discretized to generate a discrete temperature sequence. The off-grid heat dissipation sequence and the discrete temperature sequence are time-synchronized to obtain a synchronized heat dissipation sequence and synchronized temperature feature data. The synchronized heat dissipation sequence and synchronized temperature feature data are channel-merged using a multi-dimensional feature concatenation function to generate a channel-fused feature data. The channel-fused feature data is normalized and scaled to generate a standard fused feature data. The standard fused feature data are arranged in time evolution order to generate the joint time-series feature data.

[0061] Specifically, the generation process of the joint temporal feature is as follows:

[0062] The time-series thermal node data contained in the component's heat dissipation curve is extracted and discretized to transform it into a discrete heat dissipation sequence. During this process, the discretization step size is set to 50 milliseconds (this step size parameter is calibrated based on experimental test results of the radiator's thermal time constant and the sensor's response time). Simultaneously, the time-series temperature node data contained in the ambient temperature feature set is extracted and discretized using the same 50-millisecond step size to generate a discrete temperature sequence.

[0063] Time synchronization calculations are performed on the discrete heat dissipation sequence and the discrete temperature sequence. In practice, a dynamic time warping algorithm is used to calculate the Euclidean distance between the two sequences. A matching algorithm is used to stretch or compress the time axis of the temperature sequence to eliminate phase lag caused by heat conduction, resulting in a synchronized heat dissipation sequence and synchronized temperature characteristic quantity aligned at the same time scale.

[0064] In this embodiment, the global search window boundary is set based on the physical properties of the heat-conducting medium. The straight-line distance from the heat source to the temperature probe and the thermal diffusivity of the material are extracted, and the theoretical maximum delay is calculated as a hard constraint for the warping path. Specifically, to prevent the dynamic time warping algorithm from producing absurd alignments that violate physical common sense during long sequence alignments, a simple physical delay window constraint mechanism is introduced before calculating the Euclidean distance. The controller pre-reads two static structural parameters from the digital twin platform: the physical straight-line distance from the geometric center of the heat-generating wafer to the installation point of the temperature probe inside the chassis (set to 0.08 meters in this embodiment), and the conventional thermal diffusivity of the aluminum heat sink base used (fixed value of 9.7 × 10⁻⁶). -5(square meters per second). Based on the fundamental formulas of heat transfer, the theoretical delay of a single-layer homogeneous medium is calculated by dividing the square of the physical straight-line distance by the thermal diffusivity. This theoretical delay is then multiplied by the equivalent amplification factor of the interface thermal resistance (in this embodiment, based on equivalent test data from a series thermal resistance model, a fixed value of 1.8 is used). Finally, the theoretical maximum delay time of heat wave conduction after physical attenuation correction is calculated. This delay time is divided by a preset 50-millisecond segmentation step size and rounded up to obtain the maximum allowable number of misalignment points. During the execution of the dynamic time warping algorithm, the distance of the matching path stretched or compressed on the time axis from the diagonal is forcibly limited to not exceeding this maximum allowable number of misalignment points. The above process uses the heat conduction equation as a boundary constraint to prevent timing disorder and over-warping distortion caused by purely mathematical matching.

[0065] A multidimensional feature concatenation function is used to perform channel merging calculations on the synchronous heat dissipation sequence and synchronous temperature features in the column direction to generate channel fused feature values. Subsequently, normalization scaling calculations are performed on the channel fused feature values ​​to generate standard fused feature values.

[0066] The normalization scaling calculation process is as follows: For any target element in the standard fusion feature matrix, firstly, obtain the corresponding channel fusion feature matrix element, subtract the minimum observation boundary value in the historical data of the feature channel from the element to obtain the relative offset difference; then, obtain the maximum observation boundary value in the historical data of the feature channel, and calculate the difference between the maximum and minimum observation boundary values ​​as the range; finally, divide the aforementioned relative offset difference by the range, and the quotient obtained is the normalized value of the target element.

[0067] After scaling calculation, the standard fused features are arranged in chronological order to generate the joint temporal features.

[0068] Time synchronization and channel merging calculations are performed on the heat dissipation sequence and discrete temperature sequence to solve the time scale misalignment problem between high-frequency electromagnetic transients and low-frequency thermodynamic steady states. Multi-source heterogeneous sampling data are strictly aligned to eliminate feature ambiguity caused by time sequence disorder and enhance the network's accuracy in capturing the evolution of operating conditions.

[0069] The joint temporal features are input into the long short-term memory layer of the lifetime prediction network for time-dependent feature calculation to generate temporal hidden state features. The temporal hidden state features are input into the attention mechanism layer of the lifetime prediction network for weight allocation calculation to generate weighted state features. The weighted state features are then nonlinearly mapped using the mapping network layer of the lifetime prediction network to obtain the initial component prediction value. The difference between the initial component prediction value and the health and degradation state value contained in the virtual observation is compared to obtain the initial prediction bias. The absolute value of the initial prediction bias is summed to obtain the network numerical error term.

[0070] Specifically, the generation process of the network numerical error term is as follows:

[0071] In the specific hierarchical architecture and feature computation flow of the lifetime prediction network, the dimensional structure changes, core parameter settings, and selection methods of each layer have been carefully designed. First, there is the Long Short-Term Memory (LSTM) layer, whose main function is to map the input multi-channel time series into high-dimensional hidden features. In this layer, the core parameter value of the number of hidden layer nodes is set to 128. This value was chosen and set based on a comprehensive consideration of the inverter control chip's memory limitations and the need to retain historical thermal fatigue features, and was rigorously selected through model size comparison experiments. Subsequently, the data flows to the attention mechanism layer, which is responsible for assigning probability weights to the high-dimensional hidden features along the time dimension. The context dimension of the attention mechanism layer is set to 64. This selection is based on dimensionality reduction of the hidden state features, typically taking half the number of hidden layer nodes, to balance computational load and feature representation capability. Finally, there is the mapping network layer, whose function is to reduce the dimensionality of the weighted features and map them into a multi-dimensional vector equal to the number of interleaved boost branches in the inverter. This outputs an initial component prediction containing N elements (N being the total number of boost branches), where each element of the multi-dimensional vector independently corresponds to an estimate of the current aging state of a key component in a specific physical branch. In this layer, a dropout rate operation is forcibly introduced, with the core parameter set to 0.2. This value was selected through cross-validation experiments, comparing multiple sets of values, to effectively prevent overfitting of the model under ambient temperature data with added noise.

[0072] Before inputting the data into the network, the controller invokes a sliding window truncation operator to reshape the joint temporal features into tensors. Specifically, the time observation step size (60 discrete time points in this embodiment based on the typical thermal time constant of the inverter) and the sliding step size (10 time points in this embodiment) are set. The sliding window truncation operator continuously slices the joint temporal features along the time axis, transforming the originally continuous two-dimensional feature sequence into a three-dimensional tensor structure composed of multiple time-series samples (the dimension of this three-dimensional tensor is: total number of samples × time observation step size × number of feature channels). Subsequently, the reshaped three-dimensional tensor is input in batches into the long short-term memory layer included in the lifetime prediction network for time-dependent feature calculation.

[0073] The joint temporal features are input into the long short-term memory layer of the lifetime prediction network for time-dependent feature computation. The long short-term memory layer accumulates the feature information in the temporal changes through its internal gating function, generating temporal hidden state features at each time step.

[0074] The temporal hidden state features are then input into the attention mechanism layer for weight allocation calculation. This weight allocation calculation process is as follows: the temporal hidden state features are multiplied by a learnable parameter matrix and linearly projected, then processed using the hyperbolic tangent function to obtain the original attention score; next, the probability weight distribution for each time step is calculated using a normalized exponential function; finally, the probability weight sequence is weighted and summed with the original temporal hidden state features using a dot product to generate a weighted state feature quantity.

[0075] The weighted state features are nonlinearly mapped and calculated using the mapping network layer contained in the lifetime prediction network. The weight matrix of the fully connected layer is multiplied and added to output the initial predicted value of the component without physical law correction.

[0076] Extract the virtual observations generated in the previous steps, and extract the implicit health degradation state quantities (e.g., the specific value of thermal resistance increase) from the virtual observations. Compare and calculate the difference between the initial predicted quantity and the extracted health degradation state quantity, and obtain the initial prediction deviation by subtracting the two. Divide all sequence elements of the initial prediction deviation by the device's rated failure thermal resistance increment threshold (0.3 K / W in this embodiment based on the device datasheet) to make it dimensionless, and then sum the absolute values ​​to obtain the network numerical error term.

[0077] The structure utilizes a long short-term memory layer to generate temporal hidden state features and inputs them into an attention mechanism layer for weight allocation calculation. This structure can both remember long-term degradation trends and allocate higher weights for occasional extreme working conditions, which is consistent with the nonlinear burst characteristics of microscopic cumulative damage in materials and improves the sensitivity of extracting the initial prediction deviation.

[0078] Furthermore, based on the multiphysics simulation model of the key inverter components, simulation fatigue residual terms are generated and combined with the network numerical error terms to construct the network joint loss function; the network joint loss function is used to constrain the parameter iteration direction of the lifetime prediction network, and the lifetime prediction network outputs the component aging offset; the component aging offset is calculated using the remaining lifetime matrix to obtain the component remaining lifetime table; corresponding to step S4 above; the specific implementation process includes:

[0079] The thermodynamic decay law is extracted from the multiphysics simulation model. Based on the thermodynamic decay law, the Manson fatigue equation and Norris fatigue equation are used to construct a set of physical constraint equations. The joint temporal characteristic quantity is input into the set of physical constraint equations, and the continuous differentiable rainflow operator is used to perform temporal dimensionality reduction processing on the joint temporal characteristic quantity to extract the discrete stress amplitude and thermal cycle mean. The discrete stress amplitude and thermal cycle mean are used to perform mechanism deduction calculations on the set of physical constraint equations to obtain the mechanism prediction theoretical quantity. The difference between the mechanism prediction theoretical quantity and the component prediction initial quantity is compared and calculated to obtain the mechanism theory deviation. The mechanism theory deviation is scaled to generate the simulation fatigue residual term. The preset network error weight and mechanism residual weight are obtained. The network numerical error term and the simulation fatigue residual term are weighted using the network error weight and mechanism residual weight respectively, and the weighted processing result is combined to construct the network joint loss function.

[0080] Specifically, the construction process of the network joint loss function is as follows:

[0081] A multiphysics simulation model of key inverter components was analyzed to extract the thermodynamic decay law characterizing the accumulation of plastic strain in the material. Based on this law, a set of physical constraint equations was constructed using the Manson fatigue equation and the Norris fatigue equation.

[0082] The joint temporal characteristics are input into the physical constraint equations, and the continuously differentiable rainflow operator is invoked for temporal dimensionality reduction. That is, this embodiment employs a soft approximation mechanism using total differential equations. The specific calculation process for this dimensionality reduction is as follows:

[0083] First, a one-dimensional Gaussian smooth convolution kernel (in this embodiment, the kernel width is set to 5 and the standard deviation to 1.2 based on the spectral distribution characteristics of the inverter's high-frequency electromagnetic thermal noise) is used to perform smooth convolution operations on the joint time-series features to filter out high-frequency pseudo-thermal fluctuations.

[0084] Secondly, the first-order time difference is calculated for the smoothed time series, and the difference sequence is multiplied by a slope amplification factor (set to 10 in this embodiment) and then substituted into the hyperbolic tangent activation function. By utilizing the continuous smooth step characteristics of the hyperbolic tangent function near the zero point, a continuous extreme value probability weight sequence representing the trend reversal of the time series (i.e., the slope changes from positive to negative or from negative to positive) is directly generated. This continuous extreme value probability weight sequence is multiplied by the corresponding elements of the original feature sequence to locate the soft local maxima sequence and soft local minima sequence that retain gradient information.

[0085] Finally, a logarithmic summation operator is introduced to construct continuously differentiable smoothed maximum and minimum functions. For a given time-slice observation window, the smoothed maximum function is used to operate on the soft local maxima sequence to extract the upper limit estimate of thermal fluctuation within the window; simultaneously, the smoothed minimum function is used to operate on the soft local minima sequence to extract the lower limit estimate of thermal fluctuation within the window. Subsequently, the upper limit estimate of thermal fluctuation is subtracted from the lower limit estimate of thermal fluctuation, and the difference is divided by 2 to derive the continuously differentiable discrete stress amplitude; at the same time, the upper limit estimate of thermal fluctuation is added to the lower limit estimate of thermal fluctuation, and the sum is divided by 2 to derive the continuously differentiable mean of thermal cycles.

[0086] The extracted discrete stress amplitude and thermal cycle mean were used to perform mechanism deduction calculations on the physical constraint equations.

[0087] First, calculate the basic damage degree generated by a single physical thermal cycle. Specifically, divide the discrete stress amplitude by the material ductility empirical coefficient (this coefficient serves as an inherent reference stress scaling factor for the material, of the same order as the stress amplitude, used to convert the extracted amplitude into a dimensionless relative elongation ratio), take the absolute value of the quotient, and then calculate the fatigue elongation normal power of this absolute value. Next, calculate the average thermodynamic correction multiplier. Specifically, calculate the ratio of the average absolute temperature of the actual operating thermal cycle (i.e., the average value of the actual operating thermal cycle plus the absolute temperature conversion constant (a standard physical constant used to convert Celsius to Kelvin units)) to the reference absolute temperature of the test environment. The ratio is used to calculate the average temperature correction exponent raised to the power of the exponent (this exponent is an inherent empirical constant characterizing the material's sensitivity to average temperature drift); then the activation energy effect term is calculated, specifically by dividing the material activation energy constant by the Boltzmann constant (to ensure dimensional consistency, the Boltzmann constant is expressed in electron volts per Kelvin), multiplying the quotient by the following difference, where the minuend of the difference is the reciprocal of the reference absolute temperature of the test environment, and the subtrahend is the reciprocal of the average absolute temperature of the actual operating thermal cycle, and then the exponent is calculated by dividing the product by the base of the natural logarithm; finally, the basic damage degree, the thermodynamic correction multiplier, and the activation energy effect term are multiplied together, and the reciprocal of the total product is the theoretical quantity for mechanism prediction.

[0088] In the complete model of the above mechanism deduction and calculation, the empirical coefficient of material ductility, which is a constant specifically used to describe the inherent property of a particular solder layer to withstand plastic deformation without fracture, is 0.28 in this embodiment. This value is fixed based on the metallographic test failure data provided by the manufacturer at the time of manufacture of the inverter module used in the disclosure document. Secondly, there is the fatigue ductility negative constant, which is the slope parameter of the equation characterizing the sensitivity of the component's service life to plastic strain. In this embodiment, it is 2.0, independently determined by conducting multiple sets of high and low temperature cycle accelerated aging tests in the laboratory and then fitting the slope of the obtained stress-life curve. Thirdly, there is the material activation energy constant, the energy barrier that must be overcome to drive the diffusion of atoms within the crystal lattice and induce the continuous propagation of cracks. In this embodiment, it is 0.12 electron volts, selected according to the specific material used.

[0089] A unified dimensionless difference comparison calculation is performed between the theoretical quantity predicted by the mechanism and the initial quantity predicted by the component. Specifically, the actual number of thermal cycles experienced by the component is extracted and divided by the theoretical quantity predicted by the mechanism to obtain the dimensionless mechanism-induced damage degree. The initial quantity predicted by the component is divided by the device's rated failure thermal resistance increment threshold to obtain the dimensionless network-predicted damage degree (this value represents the current lifetime consumption percentage of the device extracted based on data-driven methods). Since the aforementioned calculated mechanism-induced damage degree also represents the lifetime consumption percentage calculated based on the physical mechanism, the two achieve absolute unification in terms of physical dimensions and reference benchmark. Subsequently, the network-predicted damage degree is subtracted from the mechanism-induced damage degree to obtain the mechanism theory deviation. The mechanism theory deviation is multiplied by a constraint weighting factor (e.g., a value of 0.15, which is selected dynamically based on the order of magnitude difference between physical loss and numerical loss at the initial iteration, using a proportional balance principle) for scaling calculation to generate the simulation fatigue residual term. The preset network error weights and mechanistic residual weights are obtained. In this embodiment, to balance the dominance of pure data error and physical constraints in gradient descent, the network error weight is statically set to 0.65 and the mechanistic residual weight is set to 0.35 based on the order of magnitude alignment. The above weights are used to perform scalar multiplication weighting on the network numerical error term and the simulation fatigue residual term, respectively. The two weighted values ​​are then added together to generate the joint loss function of the network.

[0090] By calling the continuous differentiable rainflow operator to perform temporal dimensionality reduction on the feature quantities, the physical fatigue equation can be seamlessly embedded into the loss function of the deep network, eliminating the risk of gradient breakage and realizing the underlying logical constraint of the theoretical equation on the model derivation.

[0091] The joint loss function of the network is analyzed to extract the joint loss value. The joint loss value is then multidimensionally differentiated using backpropagation to obtain the loss gradient descent. The node weight parameters and node bias parameters of the lifetime prediction network are updated using the loss gradient descent. The convergence condition of the joint loss value is determined. When the joint loss value reaches the set convergence condition, the node weight parameters and node bias parameters are fixed to obtain the fixed lifetime prediction network. The joint temporal features are input into the fixed lifetime prediction network for forward propagation calculation to obtain the component aging offset.

[0092] Specifically, the process for generating the component aging offset is as follows:

[0093] The joint loss function of the network is analyzed to extract the total value of the joint loss. The multidimensional derivative of the joint loss value is calculated using backpropagation. The calculation process is as follows: based on the chain rule of calculus, the partial derivatives of the joint loss value with respect to each independent weight matrix element and bias vector element in the lifetime prediction network are calculated step-by-step from the output to the input, and the summations constitute the loss gradient descent.

[0094] The loss gradient descent amount is multiplied by the global learning rate (e.g., set to 0.0005, which is determined by performing a grid search on the validation set) to update the number of node weight parameters and node bias parameters in the lifetime prediction network.

[0095] The system determines whether the value of the joint loss has reached a convergent and stationary state. The specific logic for this determination is as follows: Calculate the variance of the joint loss value over multiple consecutive iterations. When the absolute value of the variance is lower than a preset tolerance stationarity threshold (for example, set to 1.5 times 10 to the power of -6, the threshold is selected based on the floating-point processing precision of the underlying computing hardware and the asymptotic level of the training loss curve), it is determined that a convergent and stationary state has been reached.

[0096] Once the desired state is reached, the curing lifetime prediction network is derived by solidifying the node weight parameters and node bias parameters. During real-time operation, the joint temporal features extracted in real time are input into this curing lifetime prediction network, and forward propagation matrix multiplication is performed to output the component aging offset.

[0097] The node parameters are updated by using backpropagation differentiation and the network is solidified when it converges to a stationary state. This ensures that the network parameters find the global optimal solution under the dual boundary constraints of data error and physical fatigue, avoids overfitting oscillations of the model under extreme noise samples, and makes the output component aging offset have generalization.

[0098] Extract the historical attenuation characteristics and initial lifetime characteristics of the photovoltaic inverter's factory configuration; perform cumulative damage superposition calculation on the component aging offset using the historical attenuation characteristics to obtain a damage state characteristic matrix; perform deduction calculation on the damage state characteristic matrix using the initial lifetime characteristics to obtain a boost lifetime prediction matrix, and output the component's remaining lifetime table; determine the high-fatigue boost circuit and low-fatigue boost circuit included in the inverter boost circuit based on the component's remaining lifetime table, and extract the boost circuit difference values ​​included in the component's remaining lifetime table; extract the current bus ripple tolerance and power tracking error of the photovoltaic inverter, and obtain the preset imbalance safety threshold and safety boundary range; when the boost circuit difference value is greater than the imbalance safety threshold, and the bus ripple tolerance and power tracking error are within the safety boundary range; lock the multi-channel interleaved switching frequency and maintain the global current ripple value stability, calculate and generate an asymmetric duty cycle bias; inject the asymmetric duty cycle bias into the high-fatigue boost circuit, guide the steady-state operating current included in the high-fatigue boost circuit to transfer to the low-fatigue boost circuit, and perform thermal load balancing control.

[0099] Specifically, the process for generating the component remaining life table is as follows:

[0100] The calculation of the remaining life table of components involves extracting historical degradation characteristics (the proportion of fatigue loss accumulated during equipment service) and initial life characteristics (determined by consulting the product specification sheet based on the upper limit of the design operating life calibrated by the manufacturer's accelerated aging test).

[0101] By combining historical attenuation characteristics with the latest acquired component aging offset, cumulative damage is calculated to construct a remaining life matrix. The specific mathematical expression and calculation process of this matrix are as follows: Let the first... The historical attenuation characteristic (i.e., historical fatigue damage percentage) of the booster branch circuit recorded at the factory is: Divide the current component aging offset (in units of thermal resistance increment K / W) output by the network forward propagation by the device's rated failure thermal resistance increment threshold (0.3 K / W in this embodiment) to calculate the dimensionless current new damage percentage. Adding the two together yields the elements of the damage state feature matrix. (in ). Utilizing initial lifetime characteristics (Unit: hours) Perform subtraction calculations on the array using the formula. The remaining safe service life of each boost circuit branch is then determined. This generates a boost life prediction matrix, which is then used to format and output a table of remaining life for components.

[0102] The execution logic of the thermal load balancing control mechanism is as follows: extract the current bus ripple tolerance of the photovoltaic inverter (e.g., set to not exceed 1.5% of the DC base voltage, this parameter is determined by the withstand voltage specification of the DC bus capacitor) and the power tracking error (e.g., set to deviate from the maximum power point within 0.5%, determined by the grid access standard specification).

[0103] Determine whether the absolute value of the difference in boost circuit life between any two boost circuits in the component remaining life table is greater than the imbalance safety threshold, and at the same time determine whether the bus ripple tolerance and power tracking error are within the above-mentioned safety boundary range.

[0104] When all conditions are met, the controller locks the multi-channel interleaved switching frequency (fixes the main frequency at the set Hertz value) and forcibly locks the phase shift angle difference of the pulses generated between each branch, thereby maintaining the stability of the global current ripple.

[0105] Next, the asymmetric duty cycle offset is calculated. This calculation process involves: extracting the remaining safe service life of the high-fatigue branch and the low-fatigue branch, calculating the absolute difference between them, and dividing this absolute difference by the initial life characteristic of the equipment to obtain a percentage-based difference in the lifespan of the booster circuit; then, subtracting the imbalance safety threshold from this percentage-based difference to obtain a dimensionless percentage difference; multiplying this difference by the proportional gain coefficient to obtain a product; finally, dividing this product by the nominal normalized base of the basic steady-state duty cycle, the quotient being the asymmetric duty cycle offset.

[0106] The imbalance safety threshold is the percentage difference that allows for the initiation of proactive intervention in unbalanced thermal distribution. In this embodiment, it is 8%. This threshold is set manually by engineers after comprehensive evaluation, based on the inverter's internal fault-tolerant redundancy design specifications and a conservative safety margin to prevent power devices from experiencing sudden avalanche breakdown.

[0107] In this embodiment, a verification calculation is performed based on Kirchhoff's laws and the physical boundary of the device's maximum power consumption. The maximum transfer power consumption is calculated using the rated current and on-state voltage drop, and the upper limit of the imbalance safety threshold is rigidly cut and verified accordingly. Specifically, after the engineer manually sets an 8% imbalance safety threshold benchmark, the controller first reads the maximum rated output current of a single channel on the inverter's nameplate (30 amps in this embodiment) and the typical saturation on-state voltage drop of the IGBT in the device datasheet (1.7 volts in this embodiment). According to Kirchhoff's current law, during thermal equalization control, the current unloaded by the high fatigue path must be taken over by the low fatigue path. The theoretical maximum transfer current is calculated by multiplying the single-channel rated current by the currently set 8% threshold; this is then superimposed on the original steady-state current sharing value and multiplied by the saturation on-state voltage drop to derive the instantaneous maximum on-state power consumption that the low fatigue branch will bear. If the power consumption limit is determined to exceed 90% of the device's rated maximum power dissipation safety limit (150 watts in this embodiment based on the IGBT datasheet), the safe current limit is forcibly derived by dividing the maximum allowable power consumption in reverse by the on-state voltage drop, and the imbalance safety threshold percentage is forcibly lowered accordingly. This process provides objective support for threshold verification and helps mitigate the risk of device overload.

[0108] The proportional gain coefficient determines the response amplitude and severity of the asymmetric duty cycle bias calculation force to the severity of the lifetime deviation in the control closed loop. In this embodiment, it is 0.02, which is based on the Nyquist stability criterion of the control system for closed-loop tuning. It is usually determined in the automatic debugging stage before leaving the factory. Its fundamental purpose is to ensure that the control intervention action is extremely smooth, thereby preventing the pulse width modulation signal from generating high-frequency oscillations.

[0109] The nominal normalized base of the basic steady-state duty cycle represents the reference denominator of the duty cycle required to maintain the current photovoltaic string input voltage boosted to the specified DC bus voltage. In this embodiment, it is 0.6, which is directly determined by the theoretical derivation rules of the basic steady-state duty cycle, that is, by subtracting the quotient of the input DC voltage and the output DC voltage from 1.

[0110] The asymmetric duty cycle offset represents the tiny time ratio correction deviation value that is actually superimposed or subtracted from the underlying hardware waveform register. In this embodiment, the range is strictly controlled within ±0.015, and it is determined by real-time dynamic output based on the above closed-loop calculation formula.

[0111] The generated asymmetric duty cycle bias is injected into the waveform generation logic of the high-fatigue boost circuit. Specifically, this bias is subtracted from the originally symmetrical duty cycle time of the high-fatigue boost circuit, and an equal amount of bias is added to the duty cycle time of the low-fatigue boost circuit. This artificial imbalance injection in the duty cycle configuration alters the equivalent turn-on voltage drop across the inductor, guiding the steady-state operating current contained in the high-fatigue boost circuit to shift to the low-fatigue boost circuit. Because the current amplitude carried by the high-fatigue device is reduced, its conduction losses are decreased, achieving balanced thermal load control in the multi-channel interleaved system and effectively extending the overall service life of the equipment.

[0112] When the lifespan difference exceeds the limit and the electrical indicators are safe, the switching frequency is locked and an asymmetric duty cycle bias is injected. Without disrupting the ripple cancellation mechanism of the interleaved parallel topology, the high fatigue circuit thermal load is unloaded by using steady-state current transfer, avoiding the resonance risk caused by asynchronous frequency reduction and achieving bottom-level active thermal balance intervention.

[0113] To further clarify the underlying architecture and engineering deployment details of the lifetime prediction network in this embodiment, the following supplementary explanations are provided regarding its network architecture characteristics, the division of offline pre-training data, and the training process:

[0114] At the network architecture level, the lifetime prediction network adopts a deep sequence learning architecture, which consists of a cascaded input reshaping layer, a long short-term memory layer, a contextual attention mechanism layer, and a nonlinear mapping layer. The long short-term memory layer has 128 hidden nodes to extract long-term temporal dependent features; the attention mechanism layer highlights the feature representation of extreme thermal shock conditions by allocating time-step weights; the nonlinear mapping layer consists of two fully connected layers with a forced random deactivation operation with a dropout rate of 0.2 between them to prevent overfitting when faced with multidimensional noise signals. Furthermore, to conform to the irreversible physical law of monotonically increasing aging of semiconductor devices (such as junction-case thermal resistance) at the network structure level, a Softplus cumulative activation function is forcibly introduced at the output of the nonlinear mapping layer to ensure that the output aging offset has a monotonically non-decreasing physical constraint characteristic.

[0115] At the training data level, before the photovoltaic inverter is deployed online in the actual field, its lifetime prediction network needs to be pre-trained offline on a benchmark basis. The dataset used for pre-training comes from the IEC60068-2-14 standard thermal shock accelerated aging test records of the same model inverter in a laboratory environment. It contains a large amount of historical multi-dimensional operating electrothermal time-series data and its corresponding aging degradation labels. The specific physical definition of the aging degradation label is: the absolute increase (unit: K / W) of the measured junction-to-case thermal resistance of the IGBT module measured by a dedicated thermal resistance tester according to the JEDEC JESD51 standard at fixed thermal cycle intervals during the accelerated aging test, compared to the initial nominal thermal resistance of the equipment at the time of manufacture. To ensure the objectivity of the training evaluation, the pre-training dataset is divided into training set, validation set and test set according to the proportion. Specifically: 70% of the samples are extracted as the training set for basic iterative updates of node weight parameters and node bias parameters within the network; 15% of the samples are extracted as the validation set for evaluating the model's generalization ability and guiding the dynamic adjustment of hyperparameters (such as learning rate decay) after each round of training; the remaining 15% is used as the test set, which is in an absolutely isolated state and is only used for global accuracy verification of the final pre-trained model before it leaves the factory.

[0116] In terms of training process and parameter configuration, an adaptive moment estimation optimizer is used to perform gradient updates of parameters during the offline pre-training phase. The initial global learning rate is set to 0.001. To balance convergence speed and stability, an automatic learning rate decay mechanism based on validation set loss is introduced (the learning rate decays by 0.5 when the validation set loss does not decrease for 5 consecutive iterations). The batch size is set to 128 sample tensors based on hardware computing power, and the maximum number of training epochs is set to 200.

[0117] In the specific training workflow, mean squared error (MSE) is used as the basic loss function in the pre-training stage. After each batch of training data is input into the lifetime prediction network, the prediction result is output through forward propagation, and the loss residual between the prediction result and the aging degradation label is calculated. Subsequently, the gradient is calculated through the error backpropagation algorithm, and the network parameters are updated by the optimizer. To prevent the model from over-memorizing specific samples during pre-training, an early stopping mechanism is introduced: if the MSE loss of the validation set does not decrease further within 20 consecutive training rounds, the training process is forcibly terminated early. At the same time, a monotonicity penalty term is introduced on top of the MSE loss function in offline pre-training: that is, the difference between the predicted thermal resistance increments of adjacent time steps is calculated, the absolute values ​​of all prediction terms that show a decrease in thermal resistance (violating the laws of physics) are summed, and multiplied by a penalty weight coefficient of 0.1, forcing the model to fit the physical curve of monotonic aging.

[0118] After pre-training, a final validation is performed using a test set. The optimal combination of node weight parameters and node bias parameters on the validation set is extracted and used as the benchmark pre-trained model, then solidified and burned into the edge computing controller of the photovoltaic inverter. After the device is officially connected to the grid, the newly generated joint temporal features are cached in non-volatile memory. When the device is shut down at night or during periods of low-load computing power idleness, online fine-tuning is triggered as a low-frequency background task. Simultaneously, the parameters of the computationally intensive Long Short-Term Memory (LSTM) layer and attention mechanism layer are frozen during the online fine-tuning process, allowing only the two fully connected mapping network layers at the end, with minimal computational cost, to perform gradient updates. When the joint loss value reaches a set tolerance stability threshold during this online iteration process, the network parameters are solidified again, and the lifetime prediction network is transformed into a solidified lifetime prediction network. It is important to clarify that the solidified lifetime prediction network is not an independent new network architecture, but rather the final working entity of the lifetime prediction network after undergoing dual optimization through offline pre-training and online physical constraint fine-tuning, with parameters frozen for real-time forward propagation inference.

[0119] This invention provides a method for predicting the lifespan of key components of photovoltaic inverters based on multi-dimensional operating conditions. By extracting multi-dimensional operating data and introducing it into the simulation benchmark surface library of a digital twin platform, the method achieves nonlinear decoupling of electromagnetic conditions from thermal dissipation laws at the front end, avoiding feature extraction bias caused by strong coupling of heterogeneous physical fields. In the core prediction stage, it not only relies on the lifespan prediction network to extract time-series evolution features, but also combines the multi-physics simulation fatigue residuals of inverter components with the network numerical error terms to construct a joint loss function for the network. This mechanism strictly encodes the underlying thermodynamic fatigue objective laws as boundary conditions into the backpropagation path of deep learning, making the iterative update of gradients obey physical laws and overcoming the shortcomings of poor generalization ability and lack of interpretability of pure data black box models. Finally, this scheme extends the prediction results to the underlying control through the remaining lifespan matrix, bridging the technical link from state perception, physical constraint optimization to hardware execution, effectively bridging the gap between abstract algorithms and physical entities, and improving the accuracy of lifespan assessment.

[0120] Example 2

[0121] This embodiment describes a method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions, applied to a high-altitude mountain photovoltaic power generation project. Under these conditions, the grid strength is weak, and there are significant diurnal temperature variations, causing the equipment to endure harsh alternating thermal stresses over long periods. The specific implementation steps are as follows:

[0122] A multi-dimensional operational dataset of the photovoltaic inverter is collected. Real-time current deviations of the multi-channel interleaved boost circuits and voltage ripple of the high-voltage DC bus are acquired through a sensor network, and these are combined to form a DC input feature set. Simultaneously, the voltage drop amplitude and reactive current surge at the AC grid connection point are analyzed and merged to generate an AC fluctuation feature set. Environmental features are obtained by combining data from temperature sensors on the radiator surface and suspended temperature probes inside the chassis. The virtual observation acquisition process involves extracting preliminary digital twin junction temperature estimates from the cloud, calibrating them using measured casing temperature and real-time power conversion efficiency as physical reference points, calculating high-precision junction temperatures at calibration nodes, and fusing these with global junction temperature features to obtain the final virtual observation values.

[0123] A virtual composite stress simulation reference surface library containing multiple frequency bands and amplitudes is acquired from a remote platform. A thermal stress decoupling model is established at the local edge node, and a nonlinear cross matrix array is set. This matrix is ​​used to perform electrothermal cross-mapping calculations on DC input and AC fluctuation characteristics. This calculation fully considers the self-heating and mutual heating modulation effects during AC-DC interaction, and the final output is a heat dissipation curve that reflects the dynamic heating trajectory of key components such as insulated-gate bipolar transistors.

[0124] The heat dissipation curve and ambient temperature characteristics are discretized at a specified time step, and the physical phase lag caused by heat conduction is eliminated by a dynamic time warping algorithm to achieve absolute alignment of the time axis. The aligned data are then merged along the channel dimension and normalized and scaled based on historical statistical extreme values ​​to generate joint temporal features that can be directly read by deep learning networks.

[0125] In the prediction computation phase, joint temporal features are input into a life prediction network that incorporates a long short-term memory gating structure and an attention allocation mechanism. The network accumulates historical fatigue states and assigns higher weights to extreme impact conditions, then maps these values ​​through fully connected layers to output purely data-driven initial component predictions. These initial values ​​are subtracted from the actual health degradation states resolved from virtual observations, and the absolute values ​​are summed to obtain the network's numerical error term.

[0126] To introduce physical mechanism constraints, the thermodynamic decay law in finite element simulation is extracted, and a set of physical constraint equations based on the Manson fatigue equation and the Norris fatigue equation is constructed. The amplitude and mean of temperature fluctuations are extracted using a continuously differentiable rainflow operator and substituted into the physical calculation graph to deduce the theoretical quantities predicted by the mechanism. The theoretical quantities of the mechanism are compared with the initial quantities predicted by the network and scaled proportionally to generate simulation fatigue residual terms. These residual terms are then combined with the network numerical error terms to construct the network joint loss function.

[0127] Gradient backpropagation is performed based on the joint loss graph to iteratively update the weights and biases of network nodes. Once the loss values ​​reach a convergent and stable state, the network parameters are fixed. In actual power generation operation, real-time features are input into the converged network to quickly output accurate component aging offsets.

[0128] Finally, the initial lifespan limit and historical cumulative losses of the equipment are extracted from the factory configuration. Combined with the latest aging offset, cumulative damage is calculated and subtracted to generate a remaining lifespan table covering all boost branches. When the difference in lifespan degradation between different branches exceeds the set tolerance limit, and the bus voltage fluctuation and maximum power point tracking error are within a safe range, the main control chip locks the global high-frequency switching frequency and calculates the asymmetric conduction time bias. This bias is injected into the underlying drive logic to forcibly shorten the conduction time of high-fatigue aging branches, thereby guiding the current to shift to low-fatigue healthy branches. This achieves automatic thermal load balancing without affecting overall power generation efficiency.

[0129] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions, characterized in that, include: Collect multidimensional operation datasets of photovoltaic inverters, and extract DC input feature sets, AC fluctuation feature sets, ambient temperature feature sets, and virtual observation values ​​contained in the multidimensional operation datasets; The simulation reference surface library output by the digital twin platform is obtained, and a thermal stress decoupling model is constructed using the simulation reference surface library. The DC input feature set and AC fluctuation feature set are input into the thermal stress decoupling model for data processing, and the component heat dissipation curves of the key inverter components are output. The component heat dissipation curve and the ambient temperature feature set are spliced ​​together to generate a joint time series feature quantity. The joint time series feature quantity is input into the lifetime prediction network for prediction calculation to generate the initial prediction quantity of the component. The difference between the prediction quantity and the virtual observation value is calculated to generate the numerical error term of the network. Based on the multiphysics simulation model of key inverter components, simulation fatigue residual terms are generated and combined with network numerical error terms to construct a network joint loss function. The network joint loss function is used to constrain the parameter iteration direction of the lifetime prediction network, and the lifetime prediction network outputs the component aging offset. The component aging offset is calculated using the remaining lifetime matrix to obtain the component remaining lifetime table.

2. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific generation process of the DC input feature set, the AC fluctuation feature set, the ambient temperature feature set, and the virtual observation value includes: extracting the boost current deviation and DC bus ripple based on the multidimensional operating dataset, and splicing the boost current deviation and DC bus ripple to generate the DC input feature set; extracting the voltage sag depth and reactive current surge based on the multidimensional operating dataset, and splicing the voltage sag depth and reactive current surge to generate the AC fluctuation feature set; extracting the external heat dissipation temperature and chassis internal temperature based on the multidimensional operating dataset, and splicing them to generate the ambient temperature feature set; extracting the preliminary twin junction temperature, global junction temperature feature, measured casing temperature, and conversion efficiency benchmark included in the multidimensional operating dataset; performing physical anchor point calibration calculation on the preliminary twin junction temperature using the measured casing temperature and conversion efficiency benchmark to generate the calibration node junction temperature; and splicing the calibration node junction temperature and global junction temperature feature to generate the virtual observation value.

3. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific generation process of the component's heat dissipation curve includes: injecting DC power spectrum and AC distortion spectrum into the digital twin platform to synthesize a virtual composite stress spectrum; extracting the simulated node junction temperature and simulated global junction temperature of the key inverter components under the action of the virtual composite stress spectrum; performing surface fitting calculations on the simulated node junction temperature and simulated global junction temperature to generate the simulation reference surface library; constructing the thermal stress decoupling model using the simulation reference surface library, and setting the nonlinear cross matrix array of the thermal stress decoupling model; using the nonlinear cross matrix array to perform electrothermal cross-mapping calculations on the DC input feature set and the AC fluctuation feature set to obtain the AC and DC heat dissipation; and merging the AC and DC heat dissipation to obtain the component's heat dissipation curve.

4. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific generation process of the joint time-series feature quantity includes: extracting the time-series thermal node data contained in the component heat dissipation curve and performing discretization and segmentation calculation to generate an off-grid heat dissipation sequence; extracting the time-series temperature node quantity contained in the ambient temperature feature set and performing discretization and segmentation calculation to generate a discrete temperature sequence quantity; performing time synchronization calculation on the off-grid heat dissipation sequence and the discrete temperature sequence quantity to obtain a synchronized heat dissipation sequence and a synchronized temperature feature quantity; using a multi-dimensional feature splicing function to perform channel merging calculation on the synchronized heat dissipation sequence and the synchronized temperature feature quantity to generate a channel fusion feature quantity; performing normalization and scaling calculation on the channel fusion feature quantity to generate a standard fusion feature quantity; and arranging the standard fusion feature quantity according to the time evolution order to generate the joint time-series feature quantity.

5. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific generation process of the network numerical error term includes: inputting the joint temporal feature quantity into the long short-term memory layer included in the lifetime prediction network to calculate time-dependent features and generate temporal hidden state features; inputting the temporal hidden state features into the attention mechanism layer included in the lifetime prediction network to calculate weighted state features and generate weighted state features; using the mapping network layer included in the lifetime prediction network to perform nonlinear mapping calculation on the weighted state features and obtain the initial prediction quantity of the component; performing difference comparison calculation between the initial prediction quantity of the component and the virtual observation value to obtain the initial prediction deviation; and summing the absolute values ​​of the initial prediction deviation to obtain the network numerical error term.

6. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific construction process of the network joint loss function includes: analyzing the multiphysics simulation model to extract the thermodynamic decay law; constructing a physical constraint equation set by calling the Manson fatigue equation and the Norris fatigue equation based on the thermodynamic decay law; inputting the joint temporal characteristic quantity into the physical constraint equation set; calling the continuous differentiable rainflow operator to perform temporal dimensionality reduction processing on the joint temporal characteristic quantity to extract the discrete stress amplitude and thermal cycle mean; performing mechanism deduction calculation on the physical constraint equation set using the discrete stress amplitude and thermal cycle mean to obtain the mechanism prediction theoretical quantity; performing difference comparison calculation between the mechanism prediction theoretical quantity and the component prediction initial quantity to obtain the mechanism theory deviation quantity; performing scaling calculation on the mechanism theory deviation quantity to generate the simulation fatigue residual term; obtaining the preset network error weight and mechanism residual weight; weighting the network numerical error term and the simulation fatigue residual term using the network error weight and mechanism residual weight respectively, and combining the weighted processing results to construct the network joint loss function.

7. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific generation process of the component aging offset includes: parsing the joint loss function of the network to extract the joint loss value; using the backpropagation differentiation method to perform multidimensional differentiation calculation on the joint loss value to obtain the loss gradient descent amount; using the loss gradient descent amount to update the node weight parameters and node bias parameters included in the lifetime prediction network; determining whether the joint loss value has reached the convergence condition; when the joint loss value has reached the set convergence condition, solidifying the node weight parameters and node bias parameters to obtain the solidified lifetime prediction network; inputting the joint temporal feature into the solidified lifetime prediction network for forward propagation calculation to obtain the component aging offset.

8. The method for predicting the lifespan of key components of a photovoltaic inverter based on multi-dimensional operating conditions according to claim 1, characterized in that, The specific process for generating the remaining lifespan table of the components includes: extracting the historical attenuation characteristics and initial lifespan characteristics of the photovoltaic inverter as configured at the factory; performing cumulative damage superposition calculation on the aging offset of the components using the historical attenuation characteristics to obtain a damage state feature matrix; performing deduction calculation on the damage state feature matrix using the initial lifespan characteristics to obtain a boost lifespan prediction matrix, and outputting the remaining lifespan table of the components; determining the high-fatigue boost circuit and low-fatigue boost circuit included in the inverter boost circuit based on the remaining lifespan table of the components, and extracting the differences in boost circuits included in the remaining lifespan table of the components. The numerical values ​​are extracted to determine the current bus ripple tolerance and power tracking error of the photovoltaic inverter, and the preset imbalance safety threshold and safety boundary range are obtained. When the difference value of the boost circuit is greater than the imbalance safety threshold and the bus ripple tolerance and power tracking error are within the safety boundary range, the multi-channel interleaved switching frequency is locked and the global current ripple is kept stable. An asymmetric duty cycle bias is calculated and generated. The asymmetric duty cycle bias is injected into the high fatigue boost circuit to guide the steady-state operating current contained in the high fatigue boost circuit to transfer to the low fatigue boost circuit, and thermal load balancing control is performed.

Images