A method for predicting and evaluating converter life

Abstract

The present invention belongs to the technical field of inverter life prediction, and discloses an inverter life prediction and evaluation method, step 1, data collection: construct a device to simulate the real process of the inverter, and collect the operation data of the inverter under different environments by continuously changing multiple working conditions; step 2, set health indicators; step 3, train a multi-step prediction model; step 4, global model prediction: input the real-time collected power loss data into the global multi-step prediction model to obtain the predicted values ​​of the future h power loss; step 5, switch to local model prediction: input the real-time collected power loss data of each component into the local multi-step prediction model to obtain the predicted values ​​of the future h power loss of M components. The present invention realizes a comprehensive evaluation of the performance of the inverter and an accurate prediction of the remaining life.

Description

Technical Field

[0001] The present invention belongs to the technical field of converter life prediction, and in particular relates to a converter life prediction and evaluation method. Background Art

[0002] As the core equipment of power conversion, the converter is mainly responsible for converting the power generated by renewable energy into the power required by the power grid. Over time, the performance of the converter gradually deteriorates, and even fails, which has a serious impact on the stable operation and economic benefits of the entire power system. Therefore, timely and accurate identification and replacement or repair of converters with degraded performance is crucial to ensure the stable operation of the power system. However, it takes tens of thousands of hours to obtain data under real conditions, and requires huge financial and time costs. In order to overcome this problem, accelerated life prediction provides a new method. Accelerated experiment is an efficient experimental method that simulates the long-term operation of the converter in a shorter time. With this method, data of the converter under different working conditions can be obtained in a short time, and these data can truly reflect the life change trend of the converter.

[0003] One of the existing methods for evaluating the life of converters is based on failure physics analysis. This method focuses on in-depth research and analysis of the failure mechanism of products or systems, aiming to reveal the root cause of failure. In the prediction of converter life, the method based on failure physics analysis can simulate and predict the failure of the converter during operation by establishing a physical model. This method helps to avoid building complex finite element physical models for components. Instead, it obtains the relevant parameters and working conditions of the components, calculates their temperature series, and uses a real-time rain flow algorithm (which can obtain and process load data from sensors, monitoring systems and other equipment in real time. By performing necessary preprocessing on these data, such as filtering and denoising, the accuracy and reliability of the data are improved) for statistics. Finally, the expected life of the converter can be obtained by combining the cumulative fatigue damage theory and the life model. For example, based on bipolar PWM linear modulation, a real-time power loss model of the power module of a low-voltage DC converter is constructed to accurately evaluate the expected life of the IGBT (Insulated Gate Bipolar Transistor) and diode submodules. Another method for predicting the life of a converter based on physical model mapping is to construct a physical model closely related to the actual operating conditions of the converter, map the operating parameters, working conditions and environmental factors of the converter into the model, and then predict the life of the converter through the analysis and calculation of the model. For example, through a detailed analysis of the temperature and injection power data of the MMC operating environment throughout the year, the average and effective current values ​​of the IGBT and Diode (diode) are calculated, and the average power loss in the fundamental frequency cycle is determined. Using the Foster network model, the average temperature rise of semiconductor devices and the average junction temperature of IGBTs (IGBTmodules, including IGBT and Diode) modules in the operating frequency cycle are obtained. Based on these data, the loss of IGBTs is corrected and the extreme value of the junction temperature is determined. Subsequently, the cycle and fluctuation information of the fundamental frequency and low-frequency junction temperature throughout the year are statistically analyzed, and the number of failure cycles of semiconductor devices is calculated using the Bayerer life model.

[0004] The degradation of heat exchanger performance may also be caused by the significant degradation of the performance of a large number of IGBT and diode components. Therefore, more in-depth professional analysis and judgment are required when deciding whether to replace the entire converter or only replace the internal components. Previous methods were mainly data-driven, with local modeling of IGBTs and diodes to directly evaluate their performance. In fact, redundant diodes and IGBTs are configured inside the converter, so when only a few components degrade in performance, the overall performance of the converter is often not significantly affected. In addition, the replacement of a few components is not only cumbersome, but also has no significant economic benefits. More importantly, if component replacement is based solely on local evaluation results, it is likely to lead to unnecessary replacement operations.

[0005] In general, the existing technology has the following disadvantages: 1) The process of obtaining the actual operating data of the inverter is complicated and resource-intensive, and the long data collection cycle can easily lead to time lags, affecting the timeliness of the data and the accuracy of the evaluation. 2) There is an in-depth understanding of the failure mechanism of the product or system, as well as professional knowledge and experience in the establishment and application of physical models, so the human resources and time costs may be high. 3) For the prediction of the life of the inverter, the prediction effect in the short term is relatively good, while the long-term prediction has greater uncertainty and the prediction effect is relatively poor. 4) For the performance evaluation of the component, only the local part of the component is modeled, and the overall performance of the inverter is not taken into account, resulting in poor economic benefits. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a method for predicting and evaluating the life of a converter, which comprehensively analyzes the converter through global and local modeling to more comprehensively evaluate its performance and accurately predict the remaining life of the converter. When the converter performance is detected to be degraded, the system will automatically switch to the local modeling mode, deeply evaluate the specific reasons for the performance degradation, and take corresponding solutions, thereby improving the accuracy and effectiveness of the prediction and ensuring the maximization of economic benefits.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A method for predicting and evaluating the life of a converter comprises the following steps:

[0009] Step 1: Data collection:

[0010] Build a device that simulates the real process of the converter, and collect the operating data of the converter under different environments by continuously changing various working conditions, including the effective current and average current of IGBT and diode, and switching frequency data;

[0011] Step 2: Set health indicators:

[0012] The sum of the IGBT and diode power losses P Z As the health indicator of the global multi-step prediction model, the loss rate P of IGBT and diode components is T and P D As a health indicator for local multi-step prediction models;

[0013] Step 3: Train the multi-step prediction model:

[0014] The multi-step prediction model is a deep learning model based on the encoder-decoder attention mechanism, including a global multi-step prediction model and a local multi-step prediction model;

[0015] During the model training phase, the collected loss power PZ ,P T ,P D The data is then used, and the five-fold cross-validation method is used for fine model tuning. When the RMSE / initial error of the validation set is reduced to below 15%, it can be determined that the global and local multi-step prediction models have reached the ideal accuracy standard and are determined to be the optimal models, thus completing the training process.

[0016] Step 4: Global model prediction

[0017] The real-time collected loss power P Z The data is input into the global multi-step prediction model trained in step 3 to obtain the predicted values ​​of h future power losses; then, it is determined whether there are β of the h predicted values ​​that continuously exceed the set health indicator critical value; if so, it is determined that the converter performance has degraded and the switching model is used to judge the performance of the M IGBT components;

[0018] Step 5: Switch to local model prediction:

[0019] The power loss P of each component will be collected in real time T and P D The data is input into the local multi-step prediction model trained in step 3 to obtain the predicted values ​​of the future h power losses of M components, where M is the total number of components; whether to issue an alarm is determined by judging the number of components among the M components whose predicted values ​​exceed the set value α.

[0020] Furthermore, in one fundamental cycle, the average conduction loss of the IGBT and the diode is:

[0021]

[0022] Where: Subscript x represents IGBT or diode; U x_0 is the turn-on threshold voltage of device x; r x_0 is the forward resistance of device x; I x_y is the effective current of device x, I x_p is the average current of device x; is the average conduction loss of device x in one fundamental cycle;

[0023] IGBT switching loss P sw and diode recovery loss P rec Respectively expressed as:

[0024]

[0025]

[0026] Where: f p is the switching frequency; Eon , E off , E rec They are the turn-on loss, turn-off loss and reverse recovery loss of the device at a certain DC voltage. is the average voltage of the capacitor of the corresponding module; U x_ref and I x_ref are the reference voltage and current of device x when measuring turn-on, turn-off or recovery loss respectively;

[0027] Combining the conduction loss, switching loss and recovery loss expressed by equations (2) to (3), the total loss of the IGBT and the diode is:

[0028]

[0029]

[0030] P Z =P T +P D (6)

[0031] Where: IGBT power loss is P T , the diode power loss is P D .

[0032] Furthermore, the multi-step prediction model is a Transformer model, which adopts an encoder-decoder structure. In the internal structure of the Transformer model, the loss power data is first processed by the embedding layer and converted into an information-rich embedding vector; then, these vectors are encoded and combined into a complete input sequence at the corresponding position, ensuring that the model can accurately identify the order and position of each time point in the time series; then, the data flows into the encoder part composed of multiple identical layers stacked together; in each layer, the self-attention mechanism enables the model to fully consider other time points in the time series when processing the current time point, thereby effectively capturing the global time dependency; at the same time, the feedforward neural network is responsible for further conversion and extraction of this information, strengthening the model's ability to understand time series data; and after each self-attention layer and feedforward neural network layer, residual connections and normalization layers are used;

[0033] After the encoder completes data processing, it passes it to the decoder part, which is also composed of multiple layers. The self-attention mechanism enables the model to focus on the current processing time point and deeply understand its relationship with other time points in the time series; the encoder-decoder attention mechanism enables the decoder to make full use of the encoder's output, that is, the processed global time information, to generate more accurate time predictions; finally, the output of the decoder is transformed by the linear layer and the Softmax function to form a probability distribution, that is, the model's predicted output for the time series.

[0034] Furthermore, in step 5, when the number exceeds the redundant amount, the system will issue a yellow alarm, indicating that these components need to be replaced; on the contrary, if the number of components exceeding the threshold does not exceed the redundant amount, the system will issue a red alarm, indicating that the overall performance of the converter can no longer be restored by replacing some components, and the entire converter needs to be replaced to ensure the stable and safe operation of the system.

[0035] Compared with the prior art, the present invention has the advantages of:

[0036] (1) The present invention uses the sum of the power losses of the IGBT and the diode as a health indicator. When multiple consecutive power loss values ​​are greater than the critical value of the health indicator, it indicates that the inverter life is reached.

[0037] (2) Based on the data, the present invention establishes a multi-step prediction model for the inverter life to achieve real-time, accurate and long-term prediction of the inverter life.

[0038] (3) Through switching between global and local models, a comprehensive analysis of the reasons for the degradation of converter performance is carried out.

[0039] (4) The present invention not only improves the prediction accuracy and efficiency, but also reduces the prediction cost and complexity, providing an efficient and reliable method for converter life prediction, which is of great significance to improving the reliability and economy of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0041] Figure 1 The present invention is a flow chart of the method.

[0042] Figure 2 This is the Transformer workflow diagram. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0044] Combination Figure 1 As shown, a method for predicting and evaluating the life of a converter includes the following steps:

[0045] Step 1: Data collection:

[0046] A device that simulates the real process of the converter is built, and the operating data of the converter in different environments is collected by continuously changing various working conditions, which greatly accelerates the experimental process. The device can simulate the operating conditions of the converter in various working environments, effectively improving the experimental efficiency and the accuracy of data collection.

[0047] The collected data include IGBT and diode effective current and average current, and switching frequency data.

[0048] Step 2: Set health indicators:

[0049] The sum of the IGBT and diode power losses P Z (hereinafter collectively referred to as loss power) as the health indicator of the global multi-step prediction model, and the loss rate P of each IGBT and diode element is taken into account. T and P D As a health indicator for local multi-step prediction models.

[0050] IGBT and diode are the core components of the converter. The sum of the power loss of the two is P Z (hereinafter collectively referred to as power loss) is a key indicator for evaluating the health status of the inverter and predicting its lifespan. Power loss reflects the energy conversion efficiency of the device during operation. An increase in power loss will cause the device temperature to rise, which may cause material aging and performance degradation. Therefore, the present invention uses power loss as a health indicator and judges the health status of the inverter based on its value. However, its power loss is also affected by the external environment, so it is unreasonable to judge its health status based on the power loss value at a single point in time. Therefore, the present invention uses the power loss value to exceed a certain set critical value for a long time. When the inverter performance is judged to have declined to the point where it cannot meet the working requirements, the inverter needs to be replaced or repaired in time to ensure the normal operation of the system.

[0051] In order to further determine whether the inverter should be repaired or replaced, the present invention also analyzes the power loss of the IGBT and the diode respectively, determines the number of components with degraded performance, and decides whether to replace the entire inverter or only replace specific components based on the overall performance of the components and the inverter, thereby ensuring that the economic benefits are maximized while ensuring the stable operation of the system.

[0052] In one fundamental cycle, the average conduction loss of IGBT and diode is:

[0053]

[0054] Where: Subscript x represents IGBT or diode; U x_0 is the turn-on threshold voltage of device x; r x_0is the forward resistance of device x; I x_y is the effective current of device x, I x_p is the average current of device x; is the average conduction loss of device x in one fundamental cycle.

[0055] The average current flowing through the IGBT is approximately linearly related to the IGBT turn-on loss and turn-off loss; the diode reverse recovery loss is less than the IGBT switching loss, and the diode turn-on loss is small and can be ignored compared with its reverse recovery loss. Therefore, the IGBT switching loss P sw and diode recovery loss P rec Respectively expressed as:

[0056]

[0057]

[0058] Where: f p is the switching frequency; E on , E off , E rec They are the turn-on loss, turn-off loss and reverse recovery loss of the device at a certain DC voltage. is the average voltage of the capacitor of the corresponding module; U x_ref and I x_ref They are the reference voltage and current of device x when measuring turn-on, turn-off or recovery loss respectively.

[0059] Combining the conduction loss, switching loss and recovery loss expressed by equations (2) to (3), the total loss of the IGBT and the diode is:

[0060]

[0061]

[0062] P Z =P T +P D (6)

[0063] Where: IGBT power loss is P T , the diode power loss is P D .

[0064] Step 3: Train the multi-step prediction model:

[0065] The multi-step prediction model is a deep learning model based on the encoder-decoder attention mechanism, including a global multi-step prediction model and a local multi-step prediction model.

[0066] In the model training stage, the loss power (P Z ,P T ,P D ) data, and use the five-fold cross-validation method to perform fine model tuning. When the RMSE / initial error of the validation set is reduced to below 15%, it can be determined that the global and local multi-step prediction models have reached the ideal accuracy standard and are determined to be the optimal models, thus completing the training process.

[0067] In this embodiment, the multi-step prediction model takes the Transformer model as an example. The model has demonstrated excellent performance in time series prediction with its unique self-attention mechanism. The core advantage of the Transformer model is that it can effectively capture long-distance dependencies in time series data, which is particularly important in multi-step time prediction tasks.

[0068] Combination Figure 2 As shown, the Transformer workflow is as follows:

[0069] In model training, encoding and processing information is crucial. For the Transformer model, positional encoding can ensure that the model accurately identifies the position of the loss power data in the sequence, improving processing efficiency and accuracy. The position encoding is a vector with the same dimension as the loss power data, which is added to the current loss power representation vector at each time step. It is usually calculated based on sine and cosine functions to maintain a certain distance relationship between different positions. Then, by combining data encoding (Input Emdendding) and position encoding, a complete representation of the input sequence can be formed. This representation can take into account both the loss power data and its position information, and can better capture the structure and relationship in the input sequence. The Transformer model adopts an encoder-decoder structure, and the encoder converts the input sequence composed of data encoding and position encoding into a series of time relationship representation vectors.

[0070] In the internal structure of the Transformer model, the loss power data is first processed by the embedding layer and converted into an information-rich embedding vector; then, these vectors are encoded at the corresponding positions and combined into a complete input sequence, ensuring that the model can accurately identify the order and position of each time point in the time series; then, the data flows into the encoder part which is stacked with multiple identical layers; in each layer, the self-attention mechanism enables the model to fully consider other time points in the time series when processing the current time point, thereby effectively capturing the global time dependency, and using residual connections and normalization layers after the attention layer; at the same time, the feedforward neural network is responsible for further conversion and extraction of this information, strengthening the model's ability to understand time series data, and using residual connections and normalization layers after the feedforward neural network layer to help the model better learn the representation of time series.

[0071] After the encoder completes data processing, it passes it to the decoder part, which is also composed of multiple layers. The self-attention mechanism (masked self-attention layer) enables the model to focus on the current processing time point and deeply understand its relationship with other time points in the time series; and the encoder-decoder attention mechanism (encoder-decoder attention layer) enables the decoder to make full use of the encoder's output, that is, the processed global time information, to generate more accurate time predictions; and then further transformed and extracted by the feedforward neural network layer, finally, the decoder output is transformed by the linear layer and the Softmax function to form a probability distribution, which represents the model's predicted probability for each possible value. These probability distributions (OutputProbabilities) are the model's predicted output for the time series. Similarly, residual connections and normalization layers are used after the masked self-attention layer, after the encoder-decoder attention layer, and after the feedforward neural network layer.

[0072] Step 4: Global model prediction:

[0073] The real-time collected loss power P Z The data is input into the global multi-step prediction model trained in step 3 to obtain the predicted values ​​of the future h power losses (the present invention uses an accelerated experiment to collect data once a minute. It is assumed that the prediction of the future 500 future values ​​can predict the future value after 500 minutes. In actual industrial application scenarios, this prediction time can be equivalently converted to 500 hours, providing sufficient operation time for the staff); then, it is determined whether there are β of the h predicted values ​​that continuously exceed the set health indicator critical value; if so, it is determined that the converter performance has degraded and the switching model is used to judge the performance of the M IGBT components.

[0074] Step 5: Switch to local model prediction:

[0075] The power loss P of each component will be collected in real time T and P D The data is input into the local multi-step prediction model trained in step 3 to obtain the predicted values ​​of the future h power losses of M components, where M is the total number of components. Whether to issue an alarm is determined by determining the number of components among the M components whose predicted values ​​exceed the set value α.

[0076] When the number exceeds the redundancy (for example, 20%M), the system will issue a yellow alarm, indicating that these components need to be replaced; on the contrary, if the number of components exceeding the threshold does not exceed the redundancy, the system will issue a red alarm, indicating that the overall performance of the inverter can no longer be restored by replacing some components, and the entire inverter needs to be replaced to ensure the stable and safe operation of the system.

[0077] In summary, the present invention is driven by accelerated experimental data and aims to accurately predict the life of the inverter. Its core advantage is that it uses accelerated experimental data to effectively avoid the high cost and data time lag required to obtain data in the actual operating environment. Compared with previous methods that rely on deep professional knowledge and experience to establish and apply physical models, the present invention uses deep learning technology to establish a multi-step prediction model, quickly revealing the inherent laws of the data, significantly simplifying the prediction process, and reducing dependence on professional knowledge and experience.

[0078] Compared with the previous data-driven methods, local modeling of components ignores the overall performance of the converter. Therefore, the present invention adopts a global multi-step prediction model of the converter according to the converter design mechanism to predict the overall performance of the converter. When the performance is found to have dropped to, it switches to the local model to determine the cause of the performance drop. The present invention not only improves the prediction accuracy and efficiency, but also reduces the prediction cost and complexity, providing an efficient and reliable method for predicting the life of the converter, which is of great significance to improving the reliability and economy of the power system.

[0079] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Any changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for predicting and evaluating the life of a converter, characterized in that: The following steps are involved: Step 1: Data collection: Build a device that simulates the real process of the converter, and collect the operating data of the converter under different environments by continuously changing various working conditions. The collected data include IGBT and diode effective current and average current, and switching frequency data; Step 2: Set health indicators: The sum of the IGBT and diode power losses P Z As the health indicator of the global multi-step prediction model, the loss rate P of IGBT and diode components is T and P D As a health indicator for local multi-step prediction models; Step 3: Train the multi-step prediction model: The multi-step prediction model is a deep learning model based on the encoder-decoder attention mechanism, including a global multi-step prediction model and a local multi-step prediction model; During the model training phase, the collected power loss P Z ,P T ,P D The data is then used, and the five-fold cross-validation method is used for fine model tuning. When the RMSE / initial error of the validation set is reduced to below 15%, it can be determined that the global and local multi-step prediction models have reached the ideal accuracy standard and are determined to be the optimal models, thus completing the training process. Step 4: Global model prediction The real-time collected loss power P Z The data is input into the global multi-step prediction model trained in step 3 to obtain the predicted values ​​of h future power losses; then, it is determined whether there are β of the h predicted values ​​that continuously exceed the set health indicator critical value; if so, it is determined that the converter performance has degraded and the switching model is used to judge the performance of the M IGBT components; Step 5: Switch to local model prediction: The power loss P of each component will be collected in real time T and P D The data is input into the local multi-step prediction model trained in step 3 to obtain the predicted values ​​of the future h power losses of M components, where M is the total number of components; whether to issue an alarm is determined by judging the number of components among the M components whose predicted values ​​exceed the set value α.

2. A method for predicting and evaluating the life of a converter according to claim 1, characterized in that: In one fundamental cycle, the average conduction loss of IGBT and diode is: Where: Subscript x represents IGBT or diode; U x_0 is the turn-on threshold voltage of device x; r x_0 is the forward resistance of device x; I x_y is the effective current of device x, I x_p is the average current of device x; is the average conduction loss of device x in one fundamental cycle; IGBT switching loss P sw and diode recovery loss P rec Respectively expressed as: Where: f p is the switching frequency; E on , E off , E rec They are the turn-on loss, turn-off loss and reverse recovery loss of the device at a certain DC voltage. is the average voltage of the capacitor of the corresponding module; U x_ref and I x_ref are the reference voltage and current of device x when measuring turn-on, turn-off or recovery loss respectively; Combining the conduction loss, switching loss and recovery loss expressed by equations (2) to (3), the total loss of the IGBT and the diode is: P Z =P T +P D (6) Where: IGBT power loss is P T , the diode power loss is P D .

3. A method for predicting and evaluating the life of a converter according to claim 1, characterized in that: The multi-step prediction model is a Transformer model, which adopts an encoder-decoder structure. In the internal structure of the Transformer model, the loss power data is first processed by the embedding layer and converted into an information-rich embedding vector; then, these vectors are encoded and combined into a complete input sequence at the corresponding position, ensuring that the model can accurately identify the order and position of each time point in the time series; then, the data flows into the encoder part composed of multiple identical layers stacked together; in each layer, the self-attention mechanism enables the model to fully consider other time points in the time series when processing the current time point, thereby effectively capturing the global time dependency; At the same time, the feedforward neural network is responsible for further conversion and extraction of this information, strengthening the model's ability to understand time series data; And after each self-attention layer and feed-forward neural network layer, residual connections and normalization layers are used; After the encoder completes data processing, it passes it to the decoder part, which is also composed of multiple layers. The self-attention mechanism enables the model to focus on the current time point and deeply understand its relationship with other time points in the time series; while the encoder-decoder attention mechanism enables the decoder to make full use of the encoder's output, that is, the processed global time information, to generate more accurate time predictions; Finally, the output of the decoder is transformed by the linear layer and the Softmax function to form a probability distribution, which is the model's predicted output for the time series.

4. A method for predicting and evaluating the life of a converter according to claim 1, characterized in that: In step 5, when the number exceeds the redundant amount, the system will issue a yellow alarm, indicating that these components need to be replaced; on the contrary, if the number of components exceeding the threshold does not exceed the redundant amount, the system will issue a red alarm, indicating that the overall performance of the converter can no longer be restored by replacing some components, and the entire converter needs to be replaced to ensure the stable and safe operation of the system.

Images