Optical storage inverter multi-mode fault prediction method fusing mechanism model and deep learning

Abstract

The invention discloses an optical storage inverter multi-mode fault prediction method fusing a mechanism model and deep learning, and belongs to the technical field of power electronics and artificial intelligence crossing. The method comprises the following steps: firstly, acquiring real-time operation data of the optical storage inverter; then, in parallel, according to the physical mechanism model of the key component of the inverter, evaluating to obtain the physical degradation state of the key component; generating a dynamic mechanism prior mask based on the physical degradation state; inputting the real-time operation data and the dynamic mechanism prior mask into a constrained long-short-term memory network model so as to perform time sequence feature learning on the data under the mask constraint; and finally, predicting a potential fault mode of the inverter based on a time sequence feature learning result. According to the method, the physical mechanism priori depth is fused into the learning core of the data-driven model, the technical problems of insufficient fusion and insensitivity to early weak features in the prior art are solved, and the early prediction accuracy of multiple potential fault modes is improved.

Description

Technical Field

[0001] This invention relates to the field of power electronic equipment condition monitoring and fault prediction technology, and in particular to a multi-mode fault prediction method for photovoltaic-storage inverters that integrates mechanistic models and deep learning. Background Technology

[0002] Photovoltaic energy storage (PV-ESD) systems are a crucial component in building new power systems. The PV-ESD inverter, as the key hub connecting the photovoltaic array, energy storage units, and the power grid, directly impacts the safety, stability, and efficiency of the entire system. The inverter contains numerous power electronic devices, such as electrolytic capacitors and insulated-gate bipolar transistors (IGBTs). These critical components undergo performance degradation under harsh conditions of prolonged high temperature, high voltage, and high-frequency switching, ultimately leading to failure. Therefore, early and accurate prediction of potential failures in key inverter components is of paramount importance for improving system availability and reducing operation and maintenance costs.

[0003] Existing fault prediction methods are mainly divided into two categories: mechanistic model-based methods and data-driven methods. Mechanism-based methods establish mathematical models by analyzing the physical failure mechanisms of components (such as electrolyte evaporation in capacitors and thermal fatigue in IGBTs). Their advantage is strong interpretability, but the models are often highly simplified and struggle to accurately describe the nonlinear degradation process under complex operating conditions. Data-driven methods, especially deep learning methods (such as Long Short-Term Memory networks), can automatically learn fault characteristics from massive amounts of monitoring data. However, their learning process is like a "black box," lacking guidance from physical meaning, easily affected by data noise, and not sensitive enough to early, subtle fault characteristics strongly correlated with physical degradation patterns.

[0004] Therefore, how to effectively integrate the physical prior knowledge of the mechanism model with the powerful nonlinear fitting ability of the deep learning model to solve the technical problems of poor model interpretability and low accuracy in predicting early and weak faults in the existing technology is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides a multi-mode fault prediction method for photovoltaic-storage inverters that integrates mechanistic models and deep learning, and the method aims to solve the technical problems mentioned in the background art.

[0006] The method includes: acquiring multi-channel real-time operating data of the photovoltaic-storage inverter; evaluating a physical degradation state of the key components based on a pre-defined physical mechanism model of the key components in the photovoltaic-storage inverter; generating a dynamic mechanism prior mask based on the physical degradation state; inputting the real-time operating data and the dynamic mechanism prior mask into a constrained long short-term memory network model to perform time-series feature learning on the real-time operating data under the mask constraint, and obtaining a time-series feature learning result; and predicting one or more potential fault modes of the photovoltaic-storage inverter based on the time-series feature learning result.

[0007] Optionally, after acquiring the multi-channel real-time operating data of the photovoltaic-storage inverter, the method further includes: performing data preprocessing on the real-time operating data, wherein the data preprocessing includes at least one of data cleaning, data normalization, and time window sliding.

[0008] Optionally, the key components include at least one of an electrolytic capacitor and an insulated-gate bipolar transistor; the physical mechanism model includes at least one of the Arrhenius model for evaluating the aging state of the electrolytic capacitor and the Coffin-Manson model for evaluating the cumulative thermal cycling damage of the insulated-gate bipolar transistor.

[0009] Optionally, generating a dynamic mechanism prior mask based on the physical degradation state includes: combining multiple physical degradation states representing different key components into a healthy state vector; transforming the healthy state vector into the dynamic mechanism prior mask through a nonlinear mapping function; wherein the dimension of the dynamic mechanism prior mask matches the dimension of a hidden state of the constrained long short-term memory network model.

[0010] Optionally, the step of inputting the real-time running data and the dynamic mechanism prior mask into a constrained long short-term memory network model to perform temporal feature learning on the real-time running data under the constraints of the mask includes: in the calculation of each time step of the constrained long short-term memory network model, performing element-wise multiplication operations between the dynamic mechanism prior mask and the activation values ​​of one or more gating units inside the constrained long short-term memory network model, so as to dynamically adjust the model's selective memorization and forgetting of input information.

[0011] Optionally, the one or more gating units include at least one of an input gate and a forget gate.

[0012] Optionally, predicting one or more potential fault modes of the photovoltaic-storage inverter based on the time-series feature learning results includes: comparing the occurrence probability of different fault modes represented by the time-series feature learning results with their respective preset fault warning thresholds; if any of the occurrence probabilities is greater than its corresponding preset fault warning threshold, then outputting the corresponding potential fault mode warning.

[0013] Optionally, the multi-channel real-time operating data includes at least one of the following: inverter output voltage, output current, DC bus voltage, core temperature of key components, and switching frequency.

[0014] Optionally, the nonlinear mapping function is a Sigmoid function or a Softmax function.

[0015] Optionally, the constrained long short-term memory network model is a multi-layer stacked constrained long short-term memory network model. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical architecture of a fault prediction system provided in one embodiment of the present invention.

[0017] Figure 2 This is a flowchart of a multi-mode fault prediction method for photovoltaic-storage inverters that integrates a mechanistic model and deep learning, provided by an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the internal structure and information flow of a constrained long short-term memory (C-LSTM) network unit provided in an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram illustrating the generation process of the Dynamic Mechanism Prior Mask (DMPM) provided in one embodiment of the present invention;

[0020] Figure 5 This is a schematic diagram of the hardware structure of an apparatus for implementing the fault prediction method provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0023] This embodiment provides a fault prediction method. In one specific implementation, this method employs a technique that transforms the evaluation results of a physical mechanism model into a dynamic mask and uses this mask to constrain the learning process of a long short-term memory network. This allows for the deep and real-time injection of prior physical knowledge into the core of training and inference of the data-driven model. This method solves the technical problems in existing technologies where the fusion of mechanism and data models is superficial, unable to effectively utilize physical laws to guide feature learning, and thus insensitive to early, subtle fault characteristics. It achieves the beneficial effect of significantly improving the accuracy and robustness of early fault prediction for multi-mode photovoltaic-storage inverters.

[0024] Reference Figure 1 This illustrates the overall technical architecture of a fault prediction system provided by an embodiment of the present invention. The system may include a data acquisition module 110, a mechanism model calculation module 120, a Dynamic Mechanistic Prior Mask (DMPM) generation module 130, a Constrained Long Short-Term Memory (C-LSTM) network module 140, and a fault prediction and decision-making module 150. The data acquisition module 110 is responsible for acquiring multi-dimensional operating parameters in real time from the operating environment of the photovoltaic-storage inverter and simultaneously transmitting the raw data stream composed of these parameters to the mechanism model calculation module 120 and the C-LSTM network module 140. The mechanism model calculation module 120 receives the raw data stream and, based on an internally fixed physical aging model for specific key components, quantitatively assesses the current health status of the components. The DMPM generation module 130 receives the assessment results from the mechanism model calculation module 120 and converts them into a mask tensor that can directly operate within the neural network. The constrained long short-term memory network module 140 uses the original data stream as its primary learning object, while simultaneously receiving the mask tensor generated by the DMPM generation module 130 as a guiding and constraining signal for its internal information flow. Finally, the fault prediction and decision module 150 analyzes the features containing fault information output by the constrained long short-term memory network module 140 and makes a final fault mode warning decision.

[0025] Reference Figure 2This document illustrates a detailed flowchart of a multi-mode fault prediction method for photovoltaic-storage inverters that integrates a mechanistic model and deep learning, provided by an embodiment of the present invention. The various steps in this flowchart will be described in detail below.

[0026] S100: Obtain multi-channel real-time operating data of the photovoltaic-storage inverter.

[0027] In one embodiment, S100 is executed first. This step forms the data foundation of the entire fault prediction method, aiming to comprehensively capture time-series information of various physical quantities that reflect the internal state of the inverter. The data acquisition module 110 communicates with the inverter's internal Controller Area Network (CAN) bus or with externally deployed high-precision sensors (such as Hall sensors, thermocouples, etc.) to acquire multi-channel real-time operating data at a preset sampling frequency, such as, but not limited to, 10kHz or 20kHz. A higher sampling frequency is chosen to ensure that weak signal fluctuations that may exist in high-frequency components and are related to early faults can be captured.

[0028] The multi-channel real-time operating data constitutes a multi-dimensional time-series dataset. Its dimensions may include, but are not limited to: the inverter's AC output voltage and current, reflecting the state of its basic power conversion function; the DC bus voltage, reflecting the input stability of the energy storage unit or photovoltaic array; the core or surface temperature of key components (such as IGBT modules and electrolytic capacitors), which is the most direct indicator of the component's thermal stress state; and the inverter's switching frequency, whose variations may be related to control strategies or potential anomalies. These data collectively constitute a feature space describing the inverter's operating state.

[0029] For example, at a certain moment The collected data can be represented as a vector. Each component of this vector represents a specific physical measurement value, the meaning of which may include: This represents the effective value of the instantaneous output voltage of the inverter on the AC side; This represents the effective value of the instantaneous output current of the inverter on the AC side; This represents the instantaneous voltage of the inverter's DC bus; This represents the core or surface temperature of the insulated gate bipolar transistor (IGBT) module. Represents the core or surface temperature of the electrolytic capacitor; and This represents the real-time switching frequency of the inverter. For example, at a certain sampling point, a specific measurement value may be obtained, such as... , .

[0030] After obtaining the raw data, a series of data preprocessing operations are usually required to improve the processing efficiency and performance of subsequent models.

[0031] Optionally, the preprocessing includes data cleaning. The purpose of data cleaning is to identify and process outliers or missing values ​​that may occur during data acquisition. Outliers may be caused by momentary sensor malfunctions or strong electromagnetic interference.

[0032] For example, statistical methods, such as " The "rule" is used to identify outliers. For a given data channel, its mean is calculated over a time window. and standard deviation If the value of a data point exceeds... If a value falls within a certain range, it is marked as an outlier. For marked outliers, interpolation using the mean or median of nearest neighbors can be used to fill in the gaps, ensuring the continuity of the data sequence. For example, if a value of [value missing] is detected in a temperature sequence... The outlier is that the values ​​before and after it are both... Left and right can be replaced with .

[0033] Optionally, the preprocessing includes data normalization. This is because the data from different channels have different physical dimensions and numerical ranges (e.g., voltage is in volts, with values ​​in the hundreds; temperature is in cubic meters). (Values ​​in the tens place) Directly inputting these values ​​into a neural network can lead to an imbalance in the weights of different features, making it difficult for the model to converge during training. Data normalization aims to eliminate this discrepancy by mapping all features to a uniform, smaller numerical range, such as [0, 1] or [-1, 1].

[0034] For example, a min-max scaling method can be used. For any feature channel... Data points in Its normalized value It can be calculated using the following formula: .in, and These are the maximum and minimum values ​​of the channel in the training dataset, respectively. For example, if the measurement range of the IGBT temperature channel is... An actual measurement value Then its normalized value is (Dimensionless). This treatment ensures that all features contribute equally in terms of numerical value.

[0035] Optionally, the preprocessing includes time window sliding. Temporal models such as Long Short-Term Memory (LSTM) networks require a fixed-length sequence as input. Time window sliding divides a continuous data stream into a series of sequences with fixed time steps. potentially overlapping sequence samples.

[0036] For example, assume a time step is set. (That is, each sample contains 128 consecutive sampling points), the sliding step size is The first sample will contain data points 1 to 128, the second sample will contain data points 11 to 138, and so on. In this way, a long time-series data stream can be transformed into a large number of training samples, each sample being a data point of shape... The matrix, where This refers to the number of data channels. This operation significantly expands the size of the training dataset and provides structured input for the model to learn temporal dependencies.

[0037] S200: Based on the physical mechanism model of the preset key components in the photovoltaic-storage inverter, evaluate and determine a physical degradation state of the key components.

[0038] This step executes in parallel with S100. Its core task is to dynamically update and evaluate the physical degradation state of critical components based on the continuous real-time operating data stream acquired by S100, using a state memory unit or cumulative calculation mechanism. This is a process of transforming historical operating condition information into a Health Indicator (HI) characterizing the current cumulative damage level through integration or accumulation effects, providing strong prior knowledge with clear physical meaning and historical information for subsequent data-driven models. In a specific implementation, the system can maintain a cumulative damage register for each critical component. Within each time step, this register calculates the incremental damage based on the current real-time data (such as temperature and voltage fluctuations) and adds it to the cumulative value from the previous moment, thereby achieving continuous tracking and updating of the physical degradation state.

[0039] The key components are those most prone to aging or failure, pre-determined based on inverter fault statistics and failure mode analysis.

[0040] Optionally, the key components include, but are not limited to, electrolytic capacitors on the DC bus and insulated-gate bipolar transistors (IGBTs) in the power switching module. Electrolytic capacitors have a limited lifespan due to electrolyte evaporation and are one of the most common failure components in inverters. IGBTs, on the other hand, are prone to problems such as bond wire detachment and solder layer fatigue due to repeated thermal stress cycles.

[0041] Different physical models that conform to the main failure mechanisms are used for different key components.

[0042] Optionally, for electrolytic capacitors, their aging state is mainly manifested as an increase in equivalent series resistance (ESR) and a decrease in capacitance. Their aging rate is highly correlated with the core operating temperature. Therefore, a lifetime degradation model based on Arrhenius's law can be used to assess their degradation state. The Arrhenius model describes the relationship between chemical reaction rate and temperature, and its basic form is... .in, It's lifespan. It refers to the pre-factor. It is the activation energy (eV). It is the Boltzmann constant ( ), It is the absolute temperature (K). Using this model, the capacitor core temperature can be determined based on real-time monitoring. This is used to calculate the cumulative consumption of its lifespan.

[0043] For example, the physical degradation state of a capacitor It can be defined as the ratio of the consumed lifespan to the total rated lifespan. Assume the capacitor's lifespan at rated temperature is... In real-time temperature The acceleration factor below is So, within a certain time period The lifespan consumption within is From the moment the inverter is put into operation, through continuous integration... This allows us to obtain the total cumulative lifetime consumption. Ultimately, the physical degradation state of the capacitor can be expressed as... This is a dimensionless scalar with a range of [0, 1], where 0 represents a brand new state and 1 represents reaching the end of the theoretical lifetime. For example, in The time is obtained through the above calculations. This means that 45% of the capacitor's theoretical lifespan has been consumed.

[0044] For example, the activation energy in the Arrhenius model for a common electrolytic capacitor used in a photovoltaic-storage inverter with a rated temperature of 105°C. The typical value range is from 0.7 eV to 0.95 eV. The activation energy parameter characterizes the sensitivity of temperature to the aging reaction rate, and its specific value is closely related to the chemical composition of the electrolyte and the dielectric material. Those skilled in the art can obtain the precise parameter from the manufacturer's datasheet for the specific capacitor model, or calibrate the parameter by conducting accelerated aging experiments at different constant temperatures and fitting the experimental data.

[0045] Optionally, for IGBTs, one of the main failure modes is thermal stress fatigue caused by power cycling. During each switching cycle, the junction temperature... fluctuations All of these can cause minute, irreversible damage to the solder layer and bond wires inside the device. The cumulative effect of this damage can be described by the Coffin-Manson model or its variants (such as the Bayer model). This model establishes the number of cycles before failure. With junction temperature fluctuation and average junction temperature The relationship between them is usually in the form of .

[0046] For example, the physical degradation state of an IGBT It can be defined as the cumulative value of damage. In each power cycle... The damage it causes can be seen as By monitoring junction temperature fluctuations in real time and average junction temperature By substituting the values ​​into the Coffin-Manson model, the following can be calculated for this working condition. Then, using Miner's linear cumulative damage rule, the damage from the start of the run to the current time will be calculated. The damage caused by all power cycles is added together, i.e. ,in Is to The total number of loops experienced up to this point. This cumulative damage value. This refers to the physical degradation state of the IGBT. Theoretically, when At that time, the device fails. For example, in The time was calculated This indicates that the fatigue life of the IGBT has been reduced by 62%.

[0047] For example, the stress index in the Coffin-Manson model for the Al-Si solder layer commonly used in IGBT modules is... Typical values ​​for this index range from 2.5 to 4.0 (dimensionless). This index reflects the material's resistance to thermomechanical fatigue damage. These parameters are also highly dependent on the specific packaging materials (such as solder and bonding wire materials) and the device geometry. Those skilled in the art can obtain or calibrate accurate model parameters for specific application scenarios by consulting relevant materials science literature, semiconductor device reliability manuals, or by performing active power cycling tests on the device.

[0048] Through S200, this application transforms the internal aging state, which is difficult to measure directly but is crucial for fault prediction, into two specific and calculable quantitative indicators. and .

[0049] S300: Based on the physical degradation state, generate a dynamic mechanism prior mask.

[0050] Reference Figure 4 In one embodiment, the core task of this step is to transform the low-dimensional physical degradation state vector obtained in S200 (e.g., a two-dimensional vector) This is transformed into a high-dimensional control signal that can interact with the internal state of the constrained long short-term memory network model, namely the dynamic mechanism prior mask (DMPM). The "dynamic" nature of this mask is reflected in its continuous updating with the real-time changes in the physical degradation state, while the "priority" is reflected in the fact that its numerical distribution is completely determined by the physical mechanism model.

[0051] Specifically, the generation process includes:

[0052] First, the multiple physical degradation states characterizing different key components obtained from the S200 assessment are combined into a health state vector. .

[0053] For example, at time If only capacitors and IGBTs are considered, then the health state vector is: This is... A vector of dimension, where This refers to the number of critical components being monitored. In this example... .

[0054] Then, the health state vector is transformed into the dynamic mechanism prior mask through a nonlinear mapping function. The purpose of this mapping process is to expand (or broadcast) the low-dimensional health state information to a dimension consistent with the hidden state dimension of the constrained long short-term memory network model. In a matching high-dimensional space. This is done so that the mask can be applied directly, element-wise, to the hidden state vector or the activation vector of the gated unit.

[0055] Optionally, the nonlinear mapping function can be a Sigmoid function or a Softmax function. These functions are chosen because they can map inputs of any real range to a probability distribution of [0, 1] or sum to 1, which is well-suited as a "weight" or "attention" mask.

[0056] The mapping process can be specifically implemented as follows: .

[0057] in, It is the final generated dynamic mechanism prior mask, which is a A dimensional vector. This represents the Sigmoid activation function, i.e. . It is a dimension The mapping weight matrix, It is a dimension The bias vector. This linear transformation Its function is to... Dimensional health information, linearly combined and expanded 3D space.

[0058] The mapping weight matrix The value of can be preset based on expert knowledge, or it can be learned end-to-end via backpropagation as part of the overall hybrid model. As it is learned, the model automatically identifies which neuron patterns in the hidden layers should be enhanced or suppressed for which physical degradation states.

[0059] For example, suppose the hidden layer dimension of a constrained Long Short-Term Memory network model is... , .but It is The matrix. In At time 1, the health state vector is calculated. After linear transformation and sigmoid activation, a 4-dimensional mask vector may be obtained. .

[0060] Each element of this mask vector corresponds to a neuron in the hidden layer. For example, the higher values ​​of the second element (0.88) and the third element (0.91) might mean that when the IGBT is at a higher aging level (0.62), the model should pay particular attention to specific temporal patterns captured by the second and third hidden neurons (e.g., patterns associated with drastic junction temperature fluctuations or increased switching losses). The lower value of the first element (0.15) indicates that the weight of the first neuron pattern associated with capacitor aging is suppressed at the current moment. In this way, the DMPM acts like a dynamic "dimmer," adjusting the brightness of the information flow pathways within the neural network in real time based on the physical health of the device.

[0061] To ensure that the dynamic mechanism prior mask does not unnecessarily suppress the normal information flow of the constrained long short-term memory network model when processing new or fully healthy inverters, this method further includes a mask initialization or bias adjustment strategy. In one embodiment, when the physical degradation state indicators (i.e., health state vectors) of all critical components are... When all elements of the mapping weight matrix are 0, the mapping weight matrix is... and bias vector The parameters are set to enable the dynamic mechanism prior mask output by a nonlinear mapping function (such as Sigmoid). All element values ​​are close to or equal to 1. This is equivalent to an "all-pass" mask, ensuring that during the initial healthy phase of the device, the gating units of the constrained Long Short-Term Memory network model are not decayed and can learn and infer with their original performance. In another embodiment, this can be achieved by adjusting the bias vector during training. Apply a positive bias regularization term to encourage the model to default to outputting a higher mask value when there is no explicit degradation signal.

[0062] In a preferred embodiment, the mapping weight matrix and bias vector These are not pre-fixed parameters, but rather an integral part of the entire constrained long short-term memory network model, participating in the end-to-end training process. Specifically, the mapping weight matrix... and bias vector The parameter values ​​are learned and updated through gradient-based optimization algorithms (such as the Adam optimizer or stochastic gradient descent (SGD)). During model training, a loss function, such as cross-entropy loss, is first calculated based on the difference between the model's final predicted output and the true label. Then, using backpropagation, the gradient of this loss value is propagated back to each trainable parameter of the network, starting from the output layer and layer by layer, according to the chain rule of calculus. This gradient flow updates not only the weights within the Long Short-Term Memory network (such as...) (etc.), which will also be similarly fed back to the mapping weight matrix in the dynamic mechanism prior mask generation step. and bias vector In this way, the model can automatically learn how to optimally map the physical degradation state into an internal control signal that maximizes the final prediction accuracy, achieving adaptive and deep coupling between mechanistic knowledge and the data-driven learning process.

[0063] S400: Input the real-time running data and the prior mask of the dynamic mechanism into a constrained long short-term memory network model to perform temporal feature learning on the real-time running data under the constraints of the mask, and obtain a temporal feature learning result.

[0064] In one embodiment, this step describes how the dynamic mechanism prior mask generated by S300 specifically acts on a specially designed deep learning model—a constrained long short-term memory (C-LSTM) model—to achieve a deep fusion of mechanism and data.

[0065] The constrained long short-term memory network model can have a macroscopic architecture that is a multi-layered stacked network.

[0066] For example, the model can consist of an input layer, two stacked C-LSTM layers, and an output layer. The input layer receives the preprocessed data from S100. 3D sequence samples. The first C-LSTM layer is responsible for extracting low-order temporal features from the original sequence, and its output is then fed into a second C-LSTM layer to extract higher-order, more abstract temporal features. The number of hidden units in each C-LSTM layer... It can be set to, for example, 128 or 256. Finally, the output layer (usually a fully connected layer with a Softmax activation function) maps the high-order feature vector of the last time step to a probability distribution representing different failure modes.

[0067] The key to this step lies in constraining the microscopic computational mechanisms within the Long Short-Term Memory (LSTM) network model. For example... Figure 3 As shown

[0068] This demonstrates the internal structure of a modified Constrained Long Short-Term Memory (LSTM) network unit. A standard LSTM unit contains three key gating units: the forget gate... Input gate and output gate These gating units control the flow of information in memory by generating values ​​between 0 and 1 using the Sigmoid activation function. The flow within.

[0069] Forgotten Gate Determines the memory of the previous moment How much should be retained?

[0070] Input gate The candidate information at the current moment is determined. How much should be written into memory cells.

[0071] The formula for updating memory units is: ,in This represents element-wise multiplication.

[0072] In this invention, the "constraint" is implemented by introducing a dynamic mechanism prior mask during the memory unit update process. .

[0073] Specifically, at each time step of the constrained long short-term memory network model In the calculation, the a priori mask of the dynamic mechanism is used. Element-wise multiplication is performed with the activation values ​​of one or more gating units within the model.

[0074] Optionally, the modulated gating unit may be an input gate and / or a forget gate.

[0075] For example, when both the input gate and the forget gate are constrained, the update formula for the memory cell is modified as follows:

[0076]

[0077] The significance of this modification lies in the fact that prior physical knowledge (embodied in...) (China) is now directly involved in decision-making regarding information flow.

[0078] if A high value (close to 1) for a particular element means that the physical model considers the feature pattern associated with that neuron to be very important at that moment. Therefore, this element will "amplify" the input gate at the corresponding location. And the Gate of Oblivion This makes the network more prone to forgetting old, potentially irrelevant information (because...). (Being amplified), while more actively absorbing new and important information (because) (Magnified)

[0079] Conversely, if If the value of a certain element is very low (close to 0), it will "suppress" the gating signal at the corresponding position, so that the information flow is basically blocked in that path.

[0080] Through this fine-tuning of time steps and neurons, the mechanistic model is no longer simply an additional input feature, but plays the role of the commander of the entire learning process. It guides the attention of the constrained Long Short-Term Memory network model, enabling it to focus on the weak signals most relevant to the laws of physical degradation from massive amounts of data, while ignoring irrelevant noise interference.

[0081] After multiple time-step iterations and possible multi-layered network structures, the Constrained Long Short-Term Memory (LSTM) network model ultimately outputs a temporal feature learning result. This result is typically the network's performance at the last time step. Hidden state vector Or some aggregation of the hidden states at all time steps (such as average pooling). This vector It is a high-dimensional, compact representation that contains deep temporal features related to fault prediction throughout the input sequence, filtered and enhanced by physical prior knowledge.

[0082] S500: Based on the learning results of the time-series features, predict one or more potential fault modes of the photovoltaic-storage inverter.

[0083] In one embodiment, step S500 is executed. This step is the final link in the entire process and is performed by the fault prediction and decision-making module 150. Its task is to transform the abstract feature vector extracted in step S400 into user-friendly and clear fault warning information.

[0084] Specifically, the process includes:

[0085] First, the temporal feature learning results obtained from S400 (e.g., the hidden state vector of the last time step) are processed. The vector is linearly transformed through one or more fully connected layers and finally passed through a Softmax activation function. The Softmax function transforms a real-valued vector into a probability distribution vector. .

[0086] in, It is the total number of failure modes that need to be predicted (plus a "healthy" state), for example These could correspond to "health," "capacitor failure," "IGBT failure," and "fan failure," respectively. Each element in the vector... Represents the current moment Under the input sequence, the inverter is in the first... The probability of each state, and the sum of all probabilities is 1.

[0087] Then, the probability distribution vector The probability of occurrence of each fault mode is compared with its corresponding preset fault warning threshold.

[0088] It is necessary to set different warning thresholds for different failure modes because different failures have different severity and development speeds.

[0089] For example, for capacitor aging faults that develop relatively slowly, the warning threshold is... It can be set relatively high, for example, 0.85. However, for IGBT short-circuit faults that could lead to catastrophic consequences, the warning threshold... In this case, a more sensitive threshold may be needed, such as 0.70. These thresholds can be determined based on historical data, expert experience, or risk assessment.

[0090] Finally, if the probability of any of the occurrences is greater than the corresponding preset fault warning threshold, the system outputs the corresponding potential fault mode warning.

[0091] For example, at time The probability vector output by the model is These correspond to "health," "capacitor failure," "IGBT failure," and "fan failure," respectively. At this point, the probability of capacitor failure is... Greater than its warning threshold Therefore, the system will trigger a warning message, which may read: "Warning: Potential capacitor failure risk detected. Failure probability: 88%. Repair recommended."

[0092] Through the above steps, the method proposed in this invention can effectively integrate the prior knowledge of the mechanism model and the data mining capability of the deep learning model to achieve early and accurate prediction of potential faults in various key components of the photovoltaic-storage inverter.

[0093] In one specific embodiment, the large number of labeled fault datasets required for training the constrained long short-term memory network model can be generated by constructing a digital twin model of the photovoltaic-storage inverter. Specifically, a high-fidelity simulation model can be built in a professional power electronics simulation environment (e.g., but not limited to Simulink / PLECS) based on the circuit topology, control strategy, and key component parameters of the target inverter. Subsequently, by systematically modifying the parameters of key components in this digital twin model, their physical behavior under different fault modes can be simulated. For example, capacitor aging faults can be simulated by gradually increasing the equivalent series resistance (ESR) value of the electrolytic capacitor model; solder layer fatigue can be simulated by introducing additional thermal resistance into the IGBT thermal model. By running a large number of such simulation tasks injected with different degrees and types of faults, a large-scale dataset with accurate fault labels covering the entire life cycle of the inverter, including healthy states and multiple fault modes, can be generated. The dataset generated by the simulation can be used for model pre-training. Subsequently, the model can be fine-tuned using limited normal operation data or a small amount of fault data collected from actual equipment, thereby solving the data sparsity problem while ensuring the model's adaptability to actual working conditions.

[0094] Reference Figure 5The present invention also provides a fault prediction device, which can be used to execute the fault prediction method described in any of the foregoing embodiments. In one specific embodiment, the device may include: a processor, a memory, and an input / output interface connected to the processor and the memory.

[0095] The memory is used to store computer program instructions, which, when executed by the processor, can implement all or part of the steps S100 to S500 in the aforementioned method embodiments. For example, the memory may contain the code of functional units such as a physical model calculation module, a DMPM generation module, and a constrained long short-term memory network module.

[0096] The input / output interface is used to communicate with external devices. For example, the interface can be connected to a sensor to receive multi-channel real-time operating data from the photovoltaic-storage inverter; the interface can also be connected to a warning output device (such as a display screen or alarm) to output a warning signal when a potential fault is detected.

[0097] The processor, memory, and input / output interfaces can be connected and communicate with each other via a bus or other means.

[0098] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0099] In several embodiments of the present invention, the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0102] If the aforementioned integrated units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-mode fault prediction method for photovoltaic-storage inverters that integrates mechanistic models and deep learning, characterized in that, include: Acquire multi-channel real-time operating data of the photovoltaic-storage inverter; Based on the physical mechanism model of the key components in the photovoltaic-storage inverter, a physical degradation state of the key components is evaluated and obtained. Based on the physical degradation state, a dynamic mechanism prior mask is generated; The real-time running data and the prior mask of the dynamic mechanism are input into a constrained long short-term memory network model to perform temporal feature learning on the real-time running data under the constraints of the mask, and a temporal feature learning result is obtained. Based on the learning results of the time-series features, one or more potential fault modes of the photovoltaic-storage inverter are predicted.

2. The method according to claim 1, characterized in that, After acquiring the multi-channel real-time operating data of the photovoltaic-storage inverter, the process further includes: The real-time running data is preprocessed, and the preprocessing includes at least one of data cleaning, data normalization, and time window sliding.

3. The method according to claim 1, characterized in that, The key components include at least one of an electrolytic capacitor and an insulated-gate bipolar transistor. The physical mechanism model includes at least one of the Arrhenius model for evaluating the aging state of the electrolytic capacitor and the Coffin-Manson model for evaluating the cumulative thermal cycling damage of the insulated gate bipolar transistor.

4. The method according to claim 1, characterized in that, The generation of a dynamic mechanism prior mask based on the physical degradation state includes: Multiple physical degradation states representing different key components are combined into a health state vector; The health state vector is transformed into the dynamic mechanism prior mask through a nonlinear mapping function; wherein the dimension of the dynamic mechanism prior mask matches the dimension of a hidden state of the constrained long short-term memory network model.

5. The method according to claim 1, characterized in that, The step of inputting the real-time running data and the prior mask of the dynamic mechanism into a constrained long short-term memory network model to learn the temporal features of the real-time running data under the constraints of the mask includes: In the calculation of each time step of the constrained long short-term memory network model, the dynamic mechanism prior mask is multiplied element-wise with the activation values ​​of one or more gating units inside the constrained long short-term memory network model to dynamically adjust the model's selective memorization and forgetting of input information.

6. The method according to claim 5, characterized in that, The one or more gating units include at least one of an input gate and a forget gate.

7. The method according to claim 1, characterized in that, The prediction of one or more potential fault modes of the photovoltaic-storage inverter based on the time-series feature learning results includes: The probability of occurrence of different fault modes represented by the time-series feature learning results is compared with their respective preset fault warning thresholds. If the probability of any of the occurrences is greater than the corresponding preset fault warning threshold, then the corresponding potential fault mode warning is output.

8. The method according to claim 1, characterized in that, The multi-channel real-time operating data includes at least one of the following: inverter output voltage, output current, DC bus voltage, core temperature of key components, and switching frequency.

9. The method according to claim 4, characterized in that, The nonlinear mapping function is either the Sigmoid function or the Softmax function.

10. The method according to claim 5, characterized in that, The constrained long short-term memory network model is a multi-layer stacked constrained long short-term memory network model.

Images