A method for predicting the remaining life of a photovoltaic inverter
By constructing a CNN-LSTM hybrid network model and combining time series data of multiple failure parameters, the problem of photovoltaic inverter lifetime prediction, which fails to effectively consider multiple influencing factors in existing technologies, is solved, and higher accuracy of remaining lifetime prediction is achieved.
Patent Information
- Application Number
- CN202210337982.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2025-11-21
- Estimated Expiration
- 2042-03-31
AI Technical Summary
Existing methods for predicting the remaining lifetime of photovoltaic inverters fail to effectively consider multiple influencing factors, and deep learning algorithms with a single network structure struggle to achieve high-accuracy lifetime prediction.
A CNN-LSTM hybrid network model is adopted, which combines time series data of multiple failure parameters. Through z-score preprocessing and normalization, features are extracted using depthwise separable convolution, channel attention and spatial attention. The influence relationship between the most vulnerable components in the photovoltaic inverter and the inverter is established, and the remaining lifetime of multiple failure parameters is directly predicted.
The accuracy of photovoltaic inverter life prediction has been improved by taking into account the interaction between environmental factors and components, thus achieving a more accurate prediction of remaining life.
Smart Images
Figure CN114966256B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of inverter fault prediction and health management technology, specifically relating to a method for predicting the remaining life of a photovoltaic inverter. Background Technology
[0002] The lifespan of a photovoltaic inverter is determined by the shortest-lived component in the inverter. Electrolytic capacitors, MOSFETs / IGBTs, and fans are the most easily failed components in a photovoltaic inverter. Predicting the lifespan of the most easily failed components can help predict the remaining lifespan of the photovoltaic inverter.
[0003] Existing methods for predicting remaining lifetime typically predict the remaining lifetime of individual devices such as capacitors or MOSFETs / IGBTs, without establishing the influence relationship between the most vulnerable components and the photovoltaic inverter. Furthermore, the single-network structure deep learning algorithms used in inverter lifetime prediction methods struggle to achieve high accuracy. Summary of the Invention
[0004] The purpose of this invention is to provide a method for predicting the remaining lifetime of photovoltaic inverters that takes into account multiple influencing factors and can improve the accuracy of lifetime prediction.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A method for predicting the remaining lifetime of a photovoltaic inverter includes the following steps:
[0007] Step 1: Define multiple failure parameters for the photovoltaic inverter; collect time series data of the failure parameters under different operating conditions;
[0008] Step 2: Preprocess the time series data of the failure parameters to obtain preprocessed time series data of the failure parameters;
[0009] Step 3: Construct a CNN-LSTM network, train and validate the CNN-LSTM network using the preprocessed failure parameter time series data, and obtain the trained CNN-LSTM network;
[0010] Step 4: At the lifetime prediction time, acquire real-time data of the failure parameters of the photovoltaic inverter, and use the trained CNN-LSTM network to obtain the prediction data of the failure parameters, thereby predicting the remaining lifetime of the photovoltaic inverter and obtaining the prediction result.
[0011] In step 2, the preprocessing of the time series data of the failure parameters includes z-score preprocessing and normalization.
[0012] The failure parameters include multiple parameters among the following: temperature of each fan in the photovoltaic inverter, given speed of each fan in the photovoltaic inverter, actual speed of each fan in the photovoltaic inverter, operating current of each fan in the photovoltaic inverter, temperature of each switching transistor in the photovoltaic inverter, on-resistance of each switching transistor in the photovoltaic inverter, temperature of each capacitor in the photovoltaic inverter, and capacitance value of each capacitor in the photovoltaic inverter.
[0013] The formula for the z-score preprocessing method is: in This represents the time series data of the c-th failure parameter after preprocessing using the z-score method, x c μ represents the time series data of the c-th failure parameter before preprocessing using the z-score method. c Let σ represent the mean of the c-th failure parameter. c This represents the standard deviation of the c-th failure parameter.
[0014] In step 3, the CNN-LSTM network includes a CNN network with depthwise separable convolution, channel attention, and spatial attention, as well as an LSTM network. The input of the CNN network is the preprocessed failure parameter time series data, with each input channel corresponding to a failure device, which is the fan, switch, and capacitor in the photovoltaic inverter. The output of the CNN network is connected to the input of the LSTM network, and the output of the LSTM network is the predicted data of the failure parameters.
[0015] The CNN network includes a first depthwise separable convolutional layer, a second depthwise separable convolutional layer, a channel attention module, a spatial attention module, and a third depthwise separable convolutional layer connected in sequence; the LSTM network includes a first LSTM layer, a first DROP layer, a second LSTM layer, a second DROP layer, a first fully connected layer, and a second fully connected layer connected in sequence.
[0016] In step 3, when training the CNN-LSTM network, the Adam solver and the step learning rate descent strategy are used, and the loss value is set as the root mean square error of the failure parameter. The training of the CNN-LSTM network is completed when the loss value stops decreasing and the prediction accuracy stops increasing.
[0017] In step 4, the method for obtaining real-time data of the failure parameters of the photovoltaic inverter is as follows: a thermal simulation model of the photovoltaic inverter is established in advance. The thermal simulation model represents the correspondence between the temperature at the set temperature sampling points and the temperatures at each fan, each switch, and each capacitor in the photovoltaic inverter. Then, at the lifetime prediction time, the real-time temperature at the temperature sampling points is sampled, and the real-time temperature at each fan, each switch, and each capacitor in the photovoltaic inverter is obtained using the thermal simulation model. At the same time, the given speed of each fan, the actual speed of each fan, the operating current of each fan, the on-resistance of each switch, and the capacitance value of each capacitor in the photovoltaic inverter are sampled.
[0018] The digital signal processor (DSP) assigns a speed to the fan via PWM, thus the assigned speed of the fan is a known quantity. The DSP measures the actual speed of the fan through the capture mode of IO, and the DSP measures the operating current of the fan through a current sensor.
[0019] The method for sampling the on-resistance of each switch in the photovoltaic inverter is as follows: For any switch, control the switch to be turned on, inject a constant current into the switch using a measuring circuit, and at the same time isolate the devices connected to the switch through an isolation circuit. After measuring the voltage across the switch, the on-resistance of the switch is calculated based on Ohm's law.
[0020] The method for sampling the capacitance value of each capacitor in the photovoltaic inverter is as follows: For any capacitor, the capacitor is charged using a measuring circuit, and the devices connected to the capacitor are isolated using an isolation circuit. After measuring the voltage across the capacitor at two different times, the capacitance value is calculated based on the definition of a capacitor.
[0021] The measuring circuit includes a constant current source, at least one type of switch, and a voltmeter. The constant current source, the type of switch, and the switching transistor or the capacitor are connected in series to form a loop. The voltmeter is connected in parallel across the constant current source.
[0022] The isolation circuit includes an operational amplifier and two second-type switches. The non-inverting input of the operational amplifier is connected to one end of the isolated device via the first second-type switch, and the inverting input of the operational amplifier is connected to the output of the operational amplifier and connected to the other end of the isolated device via the second second-type switch.
[0023] In step 4, a failure state standard is established for each failure parameter, and the remaining lifespan of the photovoltaic inverter is predicted based on the predicted data of the failure parameters and the corresponding failure state standard.
[0024] Due to the application of the above technical solutions, this invention has the following advantages compared with the prior art: This invention establishes the influence relationship between the most easily failed components in the inverter and the photovoltaic inverter, uses the lifespan of the most easily failed components in the inverter, such as the fan, switching transistor, and electrolytic capacitor, to predict the inverter lifespan, and considers the influence of environmental factors such as temperature and humidity on the photovoltaic inverter, and proposes a direct remaining lifetime prediction method based on a CNN-LSTM network hybrid model with multiple failure parameters, thereby improving the accuracy of inverter lifespan prediction. Attached Figure Description
[0025] Appendix Figure 1 This refers to the measurement circuit and isolation circuit used in this invention.
[0026] Appendix Figure 2 This is a schematic diagram of the channel attention module and the spatial attention module involved in this invention.
[0027] Appendix Figure 3 This is a schematic diagram of the CNN-LSTM network involved in this invention. Detailed Implementation
[0028] The present invention will be further described below with reference to the embodiments shown in the accompanying drawings.
[0029] The lifespan of a photovoltaic (PV) inverter is determined by the shortest-lived components. Electrolytic capacitors, switching transistors (MOSFETs / IGBTs), and fans are the most prone to failure in PV inverters. Predicting the lifespan of these most vulnerable components allows for the prediction of the inverter's remaining lifespan. Meanwhile, the internal and external environment of the inverter is also one of the most important factors affecting its lifespan. Excessive temperature reduces component performance and lifespan. Generally, it is believed that for every 10°C increase in ambient temperature, the lifespan of electrolytic capacitors is halved, while for every 10°C decrease, the lifespan doubles. Temperature also has a significant impact on the lifespan of switching transistors (MOSFETs / IGBTs) and fans.
[0030] Based on this, a method for predicting the remaining lifetime of photovoltaic inverters is proposed, which includes the following steps:
[0031] Step 1: Define various failure parameters for the photovoltaic inverter. These failure parameters include the temperature T at each fan in the photovoltaic inverter. f (including T) f1 ,T f2 ,…,T fn1n1 represents the number of fans in the photovoltaic inverter (typically one fan is installed in one photovoltaic inverter), and the given speed V of each fan in the photovoltaic inverter is... i (including V) i1 V i2 ,…,V in1 The actual rotational speed V of each fan in the photovoltaic inverter o (including V) o1 V o2 ,…,V on1 The operating current I of each fan in the photovoltaic inverter f (including I) f1 ,I f2 ,…,I fn1 Temperature T at each switching transistor in the photovoltaic inverter m (including T) m1 ,T m2 ,…,T mn2 (n2 is the number of switching transistors in the photovoltaic inverter), and the on-resistance R of each switching transistor in the photovoltaic inverter. DS (including R) DS1 ,R DS2 ,…,R DSn2 The temperature T at each (electrolytic) capacitor in the photovoltaic inverter c (including T) c1 ,T c2 ,…,T cn3 n3 is the number of capacitors in the photovoltaic inverter, and the capacitance value C of each capacitor in the photovoltaic inverter (including C1, C2, ..., C3). n3 Multiple types of data are collected, including time-series data of failure parameters under different operating conditions.
[0032] Step 2: Preprocess the time series data of failure parameters to obtain preprocessed time series data of failure parameters.
[0033] Preprocessing of neural network input data can generally accelerate the convergence of the neural network and enhance its generalization ability. In step 2, the preprocessing of the time series data of failure parameters includes z-score preprocessing and normalization.
[0034] Since the amplitudes of the various failure parameters differ, each similar failure parameter is preprocessed using a z-score method, as shown in the following formula: in x represents the time series data of the c-th failure parameter after z-score preprocessing. c μ represents the time series data of the c-th failure parameter before z-score preprocessing. c Let σ represent the mean of the c-th failure parameter.c Let represent the standard deviation of the c-th failure parameter. Furthermore, the predicted values for each failure parameter need to be normalized to [0,1] before preprocessing. The preprocessed data is then divided into training data and validation data. In practice, the training data accounts for 80% and the validation data accounts for 20%.
[0035] Step 3: As attached Figure 3 As shown, a CNN-LSTM network is constructed, and the CNN-LSTM network is trained and validated using preprocessed failure parameter time series data to obtain the trained CNN-LSTM network.
[0036] In step 3, the CNN-LSTM network includes a CNN network with depthwise separable convolutions, channel attention, and spatial attention, as well as an LSTM network. The input to the CNN network is preprocessed time-series data of failure parameters, with each input channel corresponding to a failed device, such as a fan, switching transistor, or capacitor in a photovoltaic inverter. The CNN network consists of a first depthwise separable convolutional layer, a second depthwise separable convolutional layer, a channel attention module, a spatial attention module, and a third depthwise separable convolutional layer connected in sequence. The output of the CNN network is connected to the input of the LSTM network, and the output of the LSTM network is the predicted data of the failure parameters. The LSTM network consists of a first LSTM layer, a first drop layer, a second LSTM layer, a second drop layer, a first fully connected layer, and a second fully connected layer connected in sequence.
[0037] The input layer of a CNN network accepts a multi-channel time series X∈R consisting of multiple failure parameters. H×1×C Where H represents the length of the time series and C represents the type of failure parameter, the series is then passed through two depthwise separable convolutional layers to capture inter-channel and spatial features, followed by channel and spatial attention modules. Figure 2 Key features are extracted and refined through depthwise separable convolutional layers, and the output is fed into an LSTM network. After passing through two LSTM layers with 128 hidden units each and a Dropout layer with a probability of 0.2, and finally through two fully connected layers, the predicted value of each failure parameter is output.
[0038] When training the CNN-LSTM network using training data, the Adam solver was used with a learning rate of 0.001 and a learning rate descent strategy of step. The number of training samples per iteration was 32, and the number of training epochs was 200. The loss value was set as the root mean square error (RMSE) of the failure parameter. As training iterates, the loss value decreases, and the prediction accuracy increases. The training of the CNN-LSTM network is complete when the loss value stops decreasing and the prediction accuracy stops increasing. Then, the CNN-LSTM network is validated using validation data. After validation, a CNN-LSTM network (deep prediction model) suitable for predicting the remaining lifetime of photovoltaic inverters is obtained.
[0039] Step 4: At the lifetime prediction time, acquire real-time data of failure parameters of the photovoltaic inverter, and use the trained CNN-LSTM network to obtain predicted data of failure parameters, thereby predicting the remaining lifetime of the photovoltaic inverter and obtaining the prediction result.
[0040] In step 4, the method for obtaining real-time data of the failure parameters of the photovoltaic inverter is as follows: a thermal simulation model of the photovoltaic inverter is established in advance, and the thermal simulation model characterizes the temperature T at the set temperature sampling points. a The temperature T at each fan in the photovoltaic inverter is respectively related to the temperature T. f1 ,T f2 ,…,T fn1 Temperature T at each switching transistor m1 ,T m2 ,…,T mn2 Temperature T at each capacitor c1 ,T c2 ,…,T cn3 The correspondence between them.
[0041] The temperature T at the temperature sampling point is obtained at a certain moment. a At this moment, the temperature T at each fan is obtained through simulation. f1 ,T f2 ,…,T fn1 Temperature T at each switching transistor m1 ,T m2 ,…,T mn2 Temperature T at each capacitor c1 ,T c2 ,…,T cn3 By extensively simulating changes in external ambient temperature and inverter operating conditions, the corresponding relationships between different temperature values at sampling points and the temperature values at the electrolytic capacitor, switching transistor, and fan were obtained, i.e., T. c =f1(T a ), T m =f2(T a ), Tf =f3(T a ).
[0042] At the lifetime prediction point, the real-time temperature at the sampling point is sampled, and the real-time temperature at each fan, each switching transistor, and each capacitor in the photovoltaic inverter is obtained using a thermal simulation model. Simultaneously, the given speed, actual speed, operating current, on-resistance, and capacitance values of each fan, switching transistor, and capacitor in the photovoltaic inverter are sampled. Details are as follows:
[0043] When a fan ages or shows signs of wear, the digital signal processor (DSP) uses PWM to set the fan speed V. i Its actual rotational speed V o Less than the given speed V i Meanwhile, the operating current I f Increase. The digital signal processor provides the rotational speed V. i Given a known quantity, the actual rotational speed V o The fan operating current I is measured by a digital signal processor (DSP) through an I / O capture mode, and then by a current sensor. f The temperature T at the sampling point was obtained from the sampling. a According to formula T f =f3(T a The temperature T at the fan was obtained. f This is how the lifetime prediction time T is obtained. f V i V o I f .
[0044] An increase in the on-resistance of a switching transistor (MOSFET / IGBT) reduces its remaining lifetime. The method for sampling the on-resistance of each switching transistor in a photovoltaic inverter is as follows: For any given switching transistor, control its conduction, inject a constant current into the transistor using a measuring circuit, and simultaneously isolate the devices connected to the transistor using an isolation circuit. After measuring the voltage across the transistor, the on-resistance is calculated based on Ohm's law. The measuring circuit includes a constant current source, at least one type-1 switch, and a voltmeter. The constant current source, the type-1 switch, and the transistor or capacitor are connected in series to form a loop, and the voltmeter is connected in parallel across the constant current source. The isolation circuit includes an operational amplifier (op-amp) and two type-2 switches. The non-inverting input of the op-amp is connected to one end of the isolated device via the first type-2 switch, and the inverting input is connected to the output of the op-amp and then connected to the other end of the isolated device via the second type-2 switch. In inverter lifetime prediction mode, such as... Figure 1 As shown, when it is necessary to measure the on-resistance R of the switching transistor S1 DS1At this time, the digital signal processor controls the switching transistor S1 to turn on, closing the two first-type switches S7 and S8 of the measurement circuit, and the switch on-resistance R... S It is known that a constant current source injects a constant current I into the switching transistor S1. rt At this point, current may flow into devices connected to switching transistor S1, namely capacitor C1 or switch S2. Therefore, it is necessary to isolate the devices connected to switching transistor S1. Taking the isolation of switching transistor S2 as an example, the input and output voltages of the follower composed of operational amplifiers are applied across switching transistor S2, and the voltages across them are equal. Therefore, the current I flowing into switching transistor S2 is... S2 The current I flowing into the op-amp is 0, and because the input impedance of the op-amp is relatively high, the current I flowing into the op-amp is 0. a The voltage V across the switch S1 is approximately zero. rt Find the on-resistance R of switch S1. DS1 =V rt / I rt -2R S Similarly, the on-resistance R of other switching transistors was measured. DS2 ,…,R DSn2 Based on the temperature T at the sampling point a According to formula T m =f2(T a The temperature T at the switching transistor is obtained. m This is how the lifetime prediction time T is obtained. m R DS1 ,R DS2 ,…,R DSn2 .
[0045] The decay of the capacitance value of an electrolytic capacitor also reduces its remaining lifespan. The method for sampling the capacitance value of each capacitor in a photovoltaic inverter is as follows: For any capacitor, charge it using a measuring circuit while isolating the devices connected to the capacitor using an isolation circuit. Then, after measuring the voltage across the capacitor at two different times, calculate the capacitance value based on the definition of capacitance. Figure 1 The measurement circuit shown charges the capacitor while simultaneously isolating capacitors connected to the capacitor under test using an isolation circuit, as described earlier. If multiple capacitors are connected in parallel, the combined capacitance value is used as the measurement result. The charging current I from the constant current source is used... c And the sampled values of the voltage across the capacitor, V1 (corresponding to time T1) and V2 (corresponding to time T2), C1 = I c *(T1-T2) / (V1-V2), similarly, the capacitance values of other capacitors C2,…,C are measured. n3 Based on the temperature T at the sampling point a According to formula T c =f1(T a The temperature T at the capacitor is obtained. c This is how the lifetime prediction time T is obtained.c C1, C2, ..., C n3 .
[0046] After acquiring real-time data of the failure parameters of the photovoltaic inverter, it is input into the trained CNN-LSTM network to obtain predicted data of the failure parameters. A failure state standard is pre-established for each failure parameter, and the remaining lifetime of the photovoltaic inverter is predicted based on the predicted data of the failure parameters and the corresponding failure state standard.
[0047] To comprehensively consider the impact of multiple failure parameters on inverter lifespan, it is necessary to integrate characteristic parameters from different vulnerable components, including T. f V i V o I f T m R DS T c C. Data collected from the same faulty device is treated as a single channel. The input to the deep prediction model is then multi-channel time-series data, where each channel represents a sequence of faulty devices. The data within each channel exhibit temporal correlation, reflecting the characteristic value changes of a monitored parameter over the runtime. Simultaneously, due to mutual influence between different parameters, data from different channels also exhibit cross-channel correlation. Existing methods mostly use synchronous mapping of temporal and cross-channel correlations for lifetime prediction, which fails to effectively capture and model the dependencies between different failure parameters, thus limiting the prediction accuracy of lifetime prediction models. This approach uses separable convolution operations instead of standard convolution operations. First, convolution operations within each channel apply a single convolution kernel to each input channel, individually mapping the temporal correlation between each failure parameter sequence. Then, pointwise convolutions create feature combinations of different failure parameters after convolution, mapping the correlation between different failure parameters. Through these two independent steps, the aim is to effectively model the interrelationships between different failure parameters by decoupling temporal and cross-channel correlations.
[0048] Although inverter lifespan is affected by multiple failure parameters, the degree of influence varies among these parameters, and their impact also differs across time periods. Therefore, channel attention and spatial attention modules are introduced into lifespan prediction. The channel attention module enables the network to focus on which channel features have a stronger impact on overall lifespan prediction during the learning process; that is, the network can focus on failure parameters that play a key role in inverter lifespan prediction. The spatial attention module enables the network to focus on which stage of the sequence has a stronger impact on overall lifespan prediction during the learning process; that is, the network can focus on the key stages among the failure parameters that play a key role in inverter lifespan prediction. Features extracted by the CNN network are fed into the LSTM network to predict each failure parameter. Failure state criteria are established for individual failure parameters (actual fan speed and current, electrolytic capacitor capacitance, MOSFET / IGBT on-resistance). When a failure parameter reaches a certain stage of the failure state criteria, the inverter's lifespan is considered to have reached that stage.
[0049] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for predicting the remaining lifespan of a photovoltaic inverter, characterized in that: The method for predicting the remaining lifetime of a photovoltaic inverter includes the following steps: Step 1: Define multiple failure parameters for the photovoltaic inverter, including the temperature of each fan in the photovoltaic inverter, the given speed of each fan in the photovoltaic inverter, the actual speed of each fan in the photovoltaic inverter, the operating current of each fan in the photovoltaic inverter, the temperature of each switch in the photovoltaic inverter, the on-resistance of each switch in the photovoltaic inverter, the temperature of each capacitor in the photovoltaic inverter, and the capacitance value of each capacitor in the photovoltaic inverter; collect time series data of the failure parameters under different operating conditions; Step 2: Preprocess the time series data of the failure parameters to obtain preprocessed time series data of the failure parameters; Step 3: Construct a CNN-LSTM network, train and validate the CNN-LSTM network using the preprocessed failure parameter time series data, and obtain the trained CNN-LSTM network; Step 4: At the lifetime prediction time, acquire real-time data of the failure parameters of the photovoltaic inverter, and use the trained CNN-LSTM network to obtain the prediction data of the failure parameters, thereby predicting the remaining lifetime of the photovoltaic inverter and obtaining the prediction result. In step 3, the CNN-LSTM network includes a CNN network with depthwise separable convolution, channel attention, and spatial attention, as well as an LSTM network. The input of the CNN network is the preprocessed failure parameter time series data, with each input channel corresponding to a failure device, which is the fan, switching transistor, and capacitor in the photovoltaic inverter. The output of the CNN network is connected to the input of the LSTM network, and the output of the LSTM network is the predicted data of the failure parameters. The CNN network comprises a first depthwise separable convolutional layer, a second depthwise separable convolutional layer, a channel attention module, a spatial attention module, and a third depthwise separable convolutional layer connected in sequence; the LSTM network comprises a first LSTM layer, a first DROP layer, a second LSTM layer, a second DROP layer, a first fully connected layer, and a second fully connected layer connected in sequence; in step 4, the method for obtaining real-time data of the failure parameters of the photovoltaic inverter is as follows: a thermal simulation model of the photovoltaic inverter is pre-established, wherein the thermal simulation model characterizes the temperature at the set temperature sampling points and the temperature at each of the various parameters in the photovoltaic inverter. The relationship between the temperatures at each fan, each switching transistor, and each capacitor is established. At the lifetime prediction time, the real-time temperature at the temperature sampling point is sampled, and the real-time temperature at each fan, each switching transistor, and each capacitor in the photovoltaic inverter is obtained using the thermal simulation model. At the same time, the given speed of each fan, the actual speed of each fan, the operating current of each fan, the on-resistance of each switching transistor, and the capacitance value of each capacitor in the photovoltaic inverter are sampled.
2. The method for predicting the remaining life of a photovoltaic inverter according to claim 1, characterized in that: In step 2, the preprocessing of the time series data of the failure parameters includes z-score preprocessing and normalization.
3. The method for predicting the remaining life of a photovoltaic inverter according to claim 2, characterized in that: The formula for the z-score preprocessing method is: in This represents the time series data of the c-th failure parameter after preprocessing using the z-score method, x c μ represents the time series data of the c-th failure parameter before preprocessing using the z-score method. c Let σ represent the mean of the c-th failure parameter. c This represents the standard deviation of the c-th failure parameter.
4. The method for predicting the remaining life of a photovoltaic inverter according to claim 1, characterized in that: In step 3, when training the CNN-LSTM network, the Adam solver and the step learning rate descent strategy are used, and the loss value is set as the root mean square error of the failure parameter. The training of the CNN-LSTM network is completed when the loss value stops decreasing and the prediction accuracy stops increasing.
5. The method for predicting the remaining life of a photovoltaic inverter according to claim 1, characterized in that: The digital signal processor (DSP) assigns a speed to the fan via PWM, thus the assigned speed of the fan is a known quantity. The DSP measures the actual speed of the fan through the capture mode of IO, and the DSP measures the operating current of the fan through a current sensor. The method for sampling the on-resistance of each switch in the photovoltaic inverter is as follows: For any switch, control the switch to be turned on, inject a constant current into the switch using a measuring circuit, and at the same time isolate the devices connected to the switch through an isolation circuit. After measuring the voltage across the switch, the on-resistance of the switch is calculated based on Ohm's law. The method for sampling the capacitance value of each capacitor in the photovoltaic inverter is as follows: For any capacitor, the capacitor is charged using a measuring circuit, and the devices connected to the capacitor are isolated using an isolation circuit. After measuring the voltage across the capacitor at two different times, the capacitance value is calculated based on the definition of a capacitor.
6. The method for predicting the remaining life of a photovoltaic inverter according to claim 5, characterized in that: The measuring circuit includes a constant current source, at least one type of switch, and a voltmeter. The constant current source, the type of switch, and the switching transistor or the capacitor are connected in series to form a loop. The voltmeter is connected in parallel across the constant current source. The isolation circuit includes an operational amplifier and two second-type switches. The non-inverting input of the operational amplifier is connected to one end of the isolated device via the first second-type switch, and the inverting input of the operational amplifier is connected to the output of the operational amplifier and connected to the other end of the isolated device via the second second-type switch.
7. The method for predicting the remaining life of a photovoltaic inverter according to claim 1, characterized in that: In step 4, a failure state standard is established for each failure parameter, and the remaining lifespan of the photovoltaic inverter is predicted based on the predicted data of the failure parameters and the corresponding failure state standard.
Citation Information
Patent Citations
Railway train bearing residual life prediction method based on CAN-LSTM
CN113987834A