Pre-trained online Gaussian process device and method for predicting hydrogen fuel cell voltage

Abstract

This invention discloses a pre-trained online Gaussian process device and method for predicting hydrogen fuel cell voltage, applicable to marine propulsion systems. Addressing the problems of traditional models relying on real-ship data, weak generalization ability, and inability to adapt to battery aging, the device includes a sensor module, a memory, and a processor. The sensor module collects pressure, current, temperature, and voltage data; the memory stores the model and program; and the processor performs data processing and prediction calculations. The method pre-trains a Gaussian process regression model using experimental data, updates it online according to preset rules using real-time ship data, and outputs the prediction results after data preprocessing, model construction, updating, and prediction. This invention achieves high-precision prediction in zero-sample / few-sample stages, dynamically adapts to complex operating conditions and battery aging, operates efficiently without redundancy, and ensures reliable operation of marine hydrogen fuel cell propulsion systems.

Description

Technical Field

[0001] This invention relates to the field of hydrogen fuel cell technology, specifically to a hydrogen fuel cell voltage prediction device and method based on a pre-trained online Gaussian process suitable for marine power systems. This device and method are used to accurately predict the output voltage of shipborne hydrogen fuel cells under complex operating conditions and during performance aging, thereby ensuring the reliable operation of the hydrogen fuel cell power system. Background Technology

[0002] The operating status of shipborne hydrogen fuel cells directly determines the navigation safety and energy efficiency of ships. As a core parameter reflecting the performance of fuel cell stacks, accurate voltage prediction is of great significance for real-time monitoring, fault warning and optimized control of the power system.

[0003] However, the operating environment of shipborne hydrogen fuel cells has significant unique characteristics: on the one hand, ships face complex and variable operating conditions during navigation, such as load fluctuations, ambient temperature changes, and unstable fuel supply pressure, which differ greatly from the static experimental conditions in the laboratory; on the other hand, after long-term operation, fuel cells experience performance degradation phenomena such as membrane electrode aging and catalyst activity decay, leading to continuous changes in output voltage under the same operating conditions. Traditional data-driven voltage prediction models (such as neural networks and support vector machines) have obvious shortcomings: firstly, model training heavily relies on large-scale real-ship data, resulting in extremely low prediction accuracy in the zero-sample or few-sample stage, making it difficult to meet the monitoring needs of the initial operation phase of a ship; secondly, the model has weak generalization ability and cannot adapt to the differences between experimental and real-ship operating conditions, easily producing large prediction biases; thirdly, the model parameters are fixed and cannot dynamically track the aging process of the fuel cell, leading to a continuous decline in prediction accuracy after long-term operation; fourthly, some models lack effective sample selection mechanisms, easily introducing redundant data during online updates, resulting in reduced model computational efficiency and increased prediction latency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a pre-trained online Gaussian process device and method for predicting hydrogen fuel cell voltage. This solves the technical problems of traditional data-driven models, which rely on large-scale real-ship data, have low prediction accuracy in zero-sample / few-sample stages, weak generalization ability, and cannot adapt to complex real-ship operating conditions and battery aging. Furthermore, online updates are prone to redundancy and are inefficient.

[0005] A pre-trained online Gaussian process device for predicting hydrogen fuel cell voltage includes: a sensor module, a memory, and a processor; The sensor module is used to collect the anode hydrogen pressure of the hydrogen fuel cell stack. Cathode oxygen pressure fuel cell current fuel cell stack temperature and voltage data; The memory is used to store the pre-trained Gaussian process regression model, online update rules, and data preprocessing program; The processor is communicatively connected to the sensor module and the memory, and is used to call the program in the memory to perform data preprocessing, pre-trained model construction, online model updating and voltage prediction operations on the data collected by the sensor module, and output the hydrogen fuel cell voltage prediction result.

[0006] Furthermore, the sensor module includes at least two pressure sensors, one current sensor, one temperature sensor, and one voltage sensor, with a total number of sensors not less than five.

[0007] Furthermore, the two pressure sensors are used to collect the hydrogen pressure at the anode. and cathode oxygen pressure The current sensor is used to collect the stack current. The temperature sensor is used to collect the temperature of the fuel cell stack. The voltage sensor is used to collect voltage. .

[0008] Furthermore, the processor uses an ARM or x86 architecture chip, and the memory uses a solid-state drive, flash memory, or random access memory.

[0009] A method for predicting the voltage of a hydrogen fuel cell based on a pre-trained online Gaussian process includes the following steps: S1: Data Acquisition Steps: Acquire the anode hydrogen pressure of the hydrogen fuel cell stack via the sensor module. Cathode oxygen pressure fuel cell current fuel cell stack temperature and voltage data; S2: Data preprocessing steps: Preprocessing the collected data... , , , , The data were subjected to outlier removal, missing value imputation, and normalization. S3: Pre-trained model construction steps: Construct an input-output set from the preprocessed experimental data, construct a pre-trained model based on Gaussian process regression, and calculate the predicted mean and variance; S4: Online model update steps: Based on the pre-trained model, combined with real-time collected shipboard hydrogen fuel cell data, optimize the model according to the preset update rules; S5: Voltage prediction step: Based on the updated online Gaussian process regression model, input the preprocessed real-time data and output the voltage prediction result.

[0010] Furthermore, in step S1, the sensor module's acquisition frequency is 0.5Hz~2Hz.

[0011] Furthermore, in step S2, outlier removal employs 3... In principle, calculate the mean of each variable. and variance Eliminate those that meet the requirements Abnormal data; missing values ​​are filled using the average of the preceding and following values. The first and last missing values ​​are filled with the most recent value, and when a value is missing near the next missing value, the value closest to the missing value is used to replace the nearest value.

[0012] Furthermore, in step S2, normalization uses 0-1 normalization, and the normalization formula is as follows: ,in For the maximum value of the corresponding variable, This represents the minimum value of the corresponding variable.

[0013] Furthermore, in step S3, the kernel function of the Gaussian process regression is: ,in For signal variance, For kernel width parameter, , These are sample vectors from different times in the input set.

[0014] Furthermore, in step S4, the preset update rule includes: S41: Determine if the real-time data acquisition is located within... Data within the specified interval and data outside the interval are considered erroneous data and will not be included in the model update. S42: Predict the voltage based on the current model. If the prediction result is not within the confidence interval... ( Within the standard normal distribution quantile, add this real-time data to the training set; S43: If the predicted result is within the confidence interval and the actual value is also within the confidence interval, do not update the model; S44: If the predicted result is within the confidence interval but the actual value is not, add the real-time data corresponding to the actual voltage value to the training set, and use the Euclidean distance formula. Calculate the similarity between the new input vector and the training set samples, and remove the samples with the highest similarity.

[0015] The technical effects of this invention are as follows: 1. High prediction accuracy in zero-sample / few-sample stages: By pre-training a Gaussian process regression model with laboratory experimental data and combining it with online optimization using real-time data from actual ships, the dependence of traditional models on large-scale actual ship data is resolved, and the prediction error can be controlled within ±2%, meeting the monitoring needs of the initial operation stage of ships.

[0016] 2. Dynamically adapt to complex operating conditions and battery aging: The online update rules accurately identify complex operating conditions such as ship load fluctuations and environmental changes, as well as performance degradation phenomena such as membrane electrode aging and catalyst activity decay, through confidence interval judgment and sample screening mechanisms. The model parameters are optimized in real time to ensure stable prediction accuracy over long-term operation.

[0017] 3. High efficiency and no redundancy: The sensor module, processor, and memory are modularly coordinated. The parallel computing capability of the processor is combined with high-speed storage, and the single prediction response time does not exceed 50ms. The sample selection mechanism avoids training set redundancy, reduces the amount of model computation, and adapts to the low latency requirements of real-time monitoring of ship power systems.

[0018] 4. Convenient installation and application, strong compatibility: The device adopts standardized installation interface and bus protocol, which can be seamlessly connected with the existing ship power control system; the acquisition frequency, confidence level and other parameters can be flexibly adjusted according to different ship types and power levels of hydrogen fuel cells, with wide adaptability.

[0019] 5. Reliable data processing: Through The combined preprocessing principles of outlier removal, mean filling of preceding and following values, and 0-1 normalization improve data quality, reduce noise interference, provide high-quality input for the model, and further ensure prediction accuracy. Attached Figure Description

[0020] The invention will be further described below with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the structure of the present invention; Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 and 2 As shown, the pre-trained online Gaussian process device for voltage prediction of shipborne hydrogen fuel cells includes: a sensor module, a memory, and a processor; The sensor module is used to collect the anode hydrogen pressure of the hydrogen fuel cell stack. Cathode oxygen pressure fuel cell current fuel cell stack temperature and voltage The data provides input parameters for voltage prediction; the memory, which is the storage medium of this invention, is used to permanently store related programs and temporarily store data, thus realizing the core function of the storage medium. The memory is used to store the pre-trained Gaussian process regression model, online update rules, and data preprocessing program. The separation design of program solidification and temporary data caching ensures data security and access efficiency. The processor is communicatively connected to the sensor module and the memory. By calling the program in the memory, it connects the "data reception-preprocessing-model calculation-result output" process, performs data preprocessing, pre-trained model construction, online model update and voltage prediction operations on the data collected by the sensor module, and outputs the hydrogen fuel cell voltage prediction result and the corresponding confidence interval data, realizing an integrated closed loop from data acquisition to prediction output, which is suitable for the real-time monitoring requirements of the shipborne hydrogen fuel cell power system.

[0023] Furthermore, the sensor module includes at least two gas pressure sensors (e.g., MPX5700DP), one Hall current sensor (e.g., ACS758LCB-050B), one platinum resistance temperature sensor (e.g., PT1000A), and one differential voltage sensor (e.g., ADS1256), for a total of no less than five. By using the sensors to capture the key parameters of pressure, current, temperature, and voltage required for voltage prediction, the module avoids functional overlap or parameter omissions, ensuring the comprehensiveness and accuracy of the collected data. This lays a solid foundation of high-quality data for subsequent model training and prediction, and helps improve the accuracy of voltage prediction in the zero-sample / few-sample stage.

[0024] Furthermore, the two gas pressure sensors are used to collect the hydrogen pressure at the anode. and cathode oxygen pressure The Hall current sensor is used to collect the current of the fuel cell stack. The platinum resistance temperature sensor is used to collect the temperature of the fuel cell stack. The differential voltage sensor is used to collect voltage. ;in The hydrogen pressure at the anode of the fuel cell stack. The oxygen pressure at the cathode of the fuel cell stack. This is the operating current of the fuel cell stack. This refers to the core temperature of the fuel cell stack. This refers to the output voltage of the fuel cell stack; it follows the mapping relationship between voltage and key parameters. This ensures accurate correspondence between the input parameters of each sensor and the prediction model, avoids prediction deviations caused by misaligned parameter acquisition, and guarantees the consistency and accuracy of the prediction logic.

[0025] Furthermore, the processor adopts an ARM or x86 architecture chip (e.g., Cortex-A9 or AM3359). The processor establishes a bidirectional communication connection with the sensor module via the SPI bus and a bidirectional communication connection with the memory via the I2C bus, and is configured with an RS485 serial port and an Ethernet interface. The memory adopts a solid-state drive, flash memory, or random access memory, specifically including independent program storage areas and data storage areas. The program storage area is a NAND Flash chip with a capacity of not less than 16GB, supporting ISP online programming updates; the data storage area is a DDR3 SDRAM chip with a capacity of not less than 2GB. The processor's architecture and floating-point arithmetic unit are adapted to the matrix operations of Gaussian process regression and the real-time requirements of online updates. The ADC converter realizes accurate conversion of analog signals to digital data. The memory partitioning design takes into account both program storage stability and high-speed data read and write. The bus interface ensures efficient data transmission between modules, jointly supporting the long-term stable operation of the device in the complex environment of a ship, avoiding prediction delays or data anomalies caused by insufficient hardware performance.

[0026] A method for predicting the voltage of a hydrogen fuel cell based on a pre-trained online Gaussian process includes the following steps: S1: Data Acquisition Steps: Acquire the anode hydrogen pressure of the hydrogen fuel cell stack via the sensor module. Cathode oxygen pressure fuel cell current fuel cell stack temperature and voltage Data, voltage The mapping relationship between the variable and the given variable satisfies the formula ; S2: Data preprocessing steps: Preprocessing the collected data... , , , , The data were subjected to outlier removal, missing value imputation, and normalization. S3: Pre-trained model construction steps: Constructing input-output sets from pre-processed experimental data. , A pre-trained model is built based on Gaussian process regression, and the predicted mean and variance are calculated. S4: Online model update steps: Based on the pre-trained model, combined with real-time collected shipboard hydrogen fuel cell data, optimize the model according to the preset update rules; S5: Voltage prediction step: Based on the updated online Gaussian process regression model, input the preprocessed real-time data and output the voltage prediction result; Through a closed-loop design encompassing "data acquisition - preprocessing - pre-trained model construction - online model update - voltage prediction," the model is first initialized and trained using a large amount of experimental data from the laboratory. Then, the model parameters are dynamically optimized using real-time data from the actual ship, achieving an organic combination of "offline pre-training + online adaptive learning." This overcomes the problem of traditional data-driven models' high dependence on large-scale data from actual ships, enabling high-precision voltage prediction even in zero-sample / few-sample stages. Simultaneously, it dynamically adapts to the complex and ever-changing operating conditions of actual ships and the performance aging characteristics of hydrogen fuel cells, ensuring the stability of prediction accuracy during long-term operation.

[0027] Furthermore, in step S1, the sensor module's acquisition frequency is 0.5Hz~2Hz, which is an adjustable frequency. The sampling time deviation of each sensor does not exceed 1ms to ensure data time synchronization. The number of experimental data acquisitions is not less than 1000 sets of valid data. The range of acquisition frequency values ​​balances data timeliness and hardware computing load, capturing key information about changes in ship operating conditions while avoiding data redundancy caused by excessive acquisition. The sampling time deviation is controlled within 1ms to ensure time synchronization of multiple parameters and avoid distortion of correlation caused by parameter time misalignment. Not less than 1000 sets of experimental data provide a sufficient fitting basis for the pre-trained model. The adjustable acquisition frequency adapts to the operating condition monitoring needs of different ships.

[0028] Furthermore, in step S2, outlier removal employs... In principle, calculate the mean of each variable. and variance Eliminate those that meet the requirements Abnormal data, among which The original data for the variables is used. Missing values ​​are filled using the mean of the preceding and following values. The first and last missing values ​​are filled with the most recent value. When a value is missing near the next missing value, the value closest to the missing value is used to replace the nearest value. The value for filling missing data points in the middle is the arithmetic mean of the data at the previous time step and the data at the next time step. If the data at the next time step is still missing, the mean is calculated by replacing the data at the next time step with the most recent valid data at the next time step that is closest to the missing value. Interference data that deviates from the normal distribution is identified and removed based on statistical laws to avoid outliers misleading model training. At the same time, the continuity of data time is used to fill in missing information and restore the original appearance of the data to the greatest extent. Through combined processing, the quality of the original data is significantly improved, noise interference is reduced, and a reliable and complete data foundation is provided for model training.

[0029] Furthermore, in step S2, normalization uses 0-1 normalization, and the normalization formula is as follows: ,in For the maximum value of the corresponding variable, This represents the minimum value of the corresponding variable. The data are normalized and standardized; all experimental data and real-time data from the actual ship are calculated independently. and The predicted value needs to be obtained through the inverse normalization formula. Restore to actual voltage value Normalization eliminates the dimensional differences between different parameters, ensuring that all parameters are within the same numerical range. This avoids dimensional interference that could cause model parameters to be biased towards parameters with a larger value range, thus ensuring the fairness and stability of model training. Inverse normalization, through reverse calculation, restores the standardized predicted values ​​to the actual voltage values ​​that conform to the actual application scenario, ensuring the practical significance of the prediction results. Together, these two methods improve the accuracy and practicality of the prediction results.

[0030] Furthermore, in step S3, the kernel function of the Gaussian process regression is: ,in For signal variance, For kernel width parameter, , respectively input sets The Middle , Sample vector at time step; predicted mean The calculation formula is: Prediction variance The calculation formula is: ,in The kernel function matrix has a dimension of number of experimental data sets × number of experimental data sets, and its i-th dimension is... Line number Column elements are , The kernel function calculates the vector of results for predicting the independent variable and the input set X. To predict the kernel function results of the independent variable and itself, It is the identity matrix. The noise variance is determined using the maximum likelihood estimation method. , and The initial solution yields , , The kernel function constructs a nonlinear mapping relationship between input parameters and voltage by measuring the similarity between samples, adapting to the correlation characteristics of complex operating conditions. The prediction mean formula calculates the core value of voltage prediction, and the prediction variance formula calculates the prediction confidence, providing a reference for prediction reliability. The maximum likelihood estimation method optimizes the model parameters to minimize the fitting error of experimental data. The nonlinear characteristics of the kernel function improve the model's adaptability to complex operating conditions on real ships. Parameter optimization ensures that the pre-trained model has high-quality initial fitting accuracy, providing a reliable basic model for subsequent online updates.

[0031] Furthermore, in step S4, the preset update rule includes: S41: Determine if the real-time data acquisition is located within... Data within and outside the specified interval is considered erroneous data and is not included in model updates; abnormal data affected by external interference are filtered based on the current model's prediction distribution to avoid erroneous data contaminating the training set. S42: Predict the voltage based on the current model. If the prediction result is not within the confidence interval... ( The value is 1.96 (equivalent to the 95% confidence level, ranging from 1.64 to 2.58) of the standard normal distribution quantile. This real-time data is added to the training set. New working condition data not covered by the model are identified, and the model's adaptability is expanded by supplementing the training set. S43: If the predicted result is within the confidence interval and the actual value is also within the confidence interval, do not update the model; determine that the model meets the requirements for the prediction accuracy of the current working condition, and avoid meaningless redundant updates. S44: If the predicted result is within the confidence interval but the actual value is not, add the real-time data corresponding to the actual voltage value to the training set, and use the Euclidean distance formula. Calculate the new input vector With training set sample vectors The similarity was determined by removing the items with the highest similarity. The sample with the smallest value; identify conflicting data caused by battery aging, and optimize the training set to avoid model parameter drift.

[0032] Example: I. Installation Method: Two gas pressure sensors are installed on the anode hydrogen outlet and cathode oxygen outlet pipes of the hydrogen fuel cell stack respectively through sealed flanges. The sensor probes are in direct contact with the gas, and the sealing rating reaches IP67. The Hall current sensor is connected in series on the copper busbar of the main power supply circuit of the stack, with an installation gap of no more than 2mm. The platinum resistance temperature sensor is embedded in the bipolar plate groove of the stack core, with a contact area with the bipolar plate of no less than 1cm². The differential voltage sensor is connected in parallel to the positive and negative output terminals of the stack through shielded wires, with a wire length of no more than 1m. The processor and memory are integrated on a control board and installed in the stack control cabinet. The control board is connected to each sensor through shielded cables with a cable length of no more than 5m.

[0033] II. The above-mentioned device was applied to a 100kW hydrogen fuel cell stack for a ship. The specific operation process is as follows: S1. Data Acquisition: The sensor module is set to an acquisition frequency of 1Hz, continuously acquiring 1000 sets of laboratory experimental data and 500 sets of real-time data during actual ship navigation; the experimental data covers the operating conditions of hydrogen pressure 0.2~0.5MPa, oxygen pressure 0.1~0.3MPa, current 50~200A, and temperature 50~80℃; the actual ship data includes the operating conditions of typical navigation states such as ship departure, cruising, acceleration, and deceleration; S2. Data Preprocessing: The 1500 sets of collected data were processed. 32 sets of outlier data were removed using the 3σ principle, 18 sets of missing data were filled, and the remaining 1468 sets of valid data were normalized to 0-1. Among these, voltage data... , The normalized data range is [0,1]; S3. Pre-trained model construction: Based on 1000 sets of experimental data, an input-output set is constructed, and the kernel function parameters are obtained through maximum likelihood estimation. , , A pre-trained model was generated, which had a fitting error of 1.2% on the experimental data. S4. Online Model Update: During actual ship operation, the pre-trained model was updated based on 500 sets of real-time data. Of these, 28 sets of data were added to the training set because their predicted values ​​exceeded the confidence interval; 15 sets of data were added to the training set and their corresponding similar samples were removed because their true values ​​exceeded the confidence interval; the remaining 457 sets of data did not trigger a model update. The updated model parameters were adjusted as follows: , , ; S5. Voltage Prediction: The updated model is used to predict the voltage of the real-time data of the ship. The prediction results are output after inverse normalization. 95% of the predicted values ​​have a confidence interval width of no more than 1.5V, the average prediction error is 1.1%, and the maximum error is no more than 2.0%, which meets the accuracy requirements for voltage monitoring of shipborne hydrogen fuel cells.

[0034] III. The method of this invention was compared and verified with traditional BP neural network models and support vector machine (SVM) models. The verification indicators included prediction error in the zero-sample stage, prediction error under actual ship operating conditions, and prediction error decay rate over long-term operation (1000 hours). The results are shown in the table below:

[0035] As shown in the table above, the prediction error of the method of the present invention is significantly lower than that of the traditional model in the zero-sample stage and under actual ship operating conditions. Moreover, the error decay rate is small during long-term operation, and it can stably adapt to battery aging and changes in operating conditions, thus having obvious technical advantages.

[0036] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A pre-trained online Gaussian process device for predicting hydrogen fuel cell voltage, characterized in that, include: Sensor module, memory, and processor; The sensor module is used to collect the anode hydrogen pressure of the hydrogen fuel cell stack. Cathode oxygen pressure fuel cell current fuel cell stack temperature and voltage data; The memory is used to store the pre-trained Gaussian process regression model, online update rules, and data preprocessing program; The processor is communicatively connected to the sensor module and the memory, and is used to call the program in the memory to perform data preprocessing, pre-trained model construction, online model updating and voltage prediction operations on the data collected by the sensor module, and output the hydrogen fuel cell voltage prediction result.

2. The pre-trained online Gaussian process device for hydrogen fuel cell voltage prediction according to claim 1, characterized in that, The sensor module includes at least two pressure sensors, one current sensor, one temperature sensor, and one voltage sensor, with a total number of no less than five sensors.

3. The pre-trained online Gaussian process device for hydrogen fuel cell voltage prediction according to claim 2, characterized in that, The two pressure sensors are used to collect the hydrogen pressure at the anode. and cathode oxygen pressure The current sensor is used to collect the stack current. The temperature sensor is used to collect the temperature of the fuel cell stack. The voltage sensor is used to collect voltage. .

4. The pre-trained online Gaussian process device for hydrogen fuel cell voltage prediction according to claim 1, characterized in that, The processor uses an ARM or x86 architecture chip, and the memory uses a solid-state drive, flash memory, or random access memory.

5. A method for using a pre-trained online Gaussian process device for hydrogen fuel cell voltage prediction according to any one of claims 1 to 4, characterized in that, Includes the following steps: S1: Data Acquisition Steps: Acquire the anode hydrogen pressure of the hydrogen fuel cell stack via the sensor module. Cathode oxygen pressure fuel cell current fuel cell stack temperature and voltage data; S2: Data preprocessing steps: Preprocessing the collected data... , , , , The data were subjected to outlier removal, missing value imputation, and normalization. S3: Pre-trained model construction steps: Construct an input-output set from the preprocessed experimental data, construct a pre-trained model based on Gaussian process regression, and calculate the predicted mean and variance; S4: Online model update steps: Based on the pre-trained model, combined with real-time collected shipboard hydrogen fuel cell data, optimize the model according to the preset update rules; S5: Voltage prediction step: Based on the updated online Gaussian process regression model, input the preprocessed real-time data and output the voltage prediction result.

6. The hydrogen fuel cell voltage prediction method based on a pre-trained online Gaussian process according to claim 5, characterized in that, In step S1, the sensor module's acquisition frequency is 0.5Hz~2Hz.

7. The hydrogen fuel cell voltage prediction method based on a pre-trained online Gaussian process according to claim 5, characterized in that, In step S2, outlier removal uses 3 In principle, calculate the mean of each variable. and variance Eliminate those that meet the requirements Abnormal data; Missing values ​​are filled using the average of the preceding and following values. The first and last missing values ​​are filled with the most recent value, and when a value is missing near the next missing value, the value closest to the missing value is used to replace the nearest value.

8. A method for predicting the voltage of a hydrogen fuel cell based on a pre-trained online Gaussian process according to claim 5 or 7, characterized in that, In step S2, normalization is performed using 0-1 normalization, and the normalization formula is as follows: ,in The maximum value of the corresponding variable. This represents the minimum value of the corresponding variable.

9. The hydrogen fuel cell voltage prediction method based on a pre-trained online Gaussian process according to claim 5, characterized in that, In step S3, the kernel function for Gaussian process regression is: ,in For signal variance, For kernel width parameter, , These are sample vectors from different times in the input set.

10. The hydrogen fuel cell voltage prediction method based on a pre-trained online Gaussian process according to claim 5, characterized in that, In step S4, the preset update rules include: S41: Determine if the real-time data acquisition is located within... Data within the specified interval and data outside the interval are considered erroneous data and will not be included in the model update. S42: Predict the voltage based on the current model. If the prediction result is not within the confidence interval... ( Within the standard normal distribution quantile, add this real-time data to the training set; S43: If the predicted result is within the confidence interval and the actual value is also within the confidence interval, do not update the model; S44: If the predicted result is within the confidence interval but the actual value is not, add the real-time data corresponding to the actual voltage value to the training set, and use the Euclidean distance formula. Calculate the similarity between the new input vector and the training set samples, and remove the samples with the highest similarity.

Images