A SOFC system performance prediction method based on LSTM-ARMA fusion algorithm

Abstract

The invention discloses a SOFC system performance prediction method based on LSTM-ARMA fusion algorithm, comprising the following steps: S1, preprocessing original data; S2, predicting the degradation trend of SOFC by using an ARMA model optimized by a robust Kalman filter; S3, predicting the output voltage of SOFC by using an LSTM algorithm, outputting a reconstruction voltage in combination with the degradation trend predicted by the ARMA model, calculating the loss function of the current node, and updating the weight parameters of each LSTM node according to the loss function; S4, repeating steps S2-S3, updating the ARMA model state, and completing the prediction of multiple nodes in the LSTM algorithm in sequence until a maximum iteration round is reached, or the loss function of the current node meets a preset reconstruction error threshold requirement, and parameter optimization is completed; S5, realizing the performance prediction of the SOFC system according to the LSTM-ARMA model after the parameter optimization is completed.

Description

Technical Field

[0001] The present invention relates to the technical field of solid oxide fuel cells, and in particular to a method for predicting SOFC system performance based on an LSTM-ARMA fusion algorithm. Background Art

[0002] In recent years, solid oxide fuel cells (SOFCs), as an electrochemical device that efficiently converts hydrogen into electrical energy, have attracted the attention of scholars at home and abroad. In order to achieve commercialization, it is urgent to optimize them in terms of cost, life, reliability, etc. Degradation performance prediction technology monitors SOFC system parameters and predicts battery performance degradation in advance, which is of great significance for preventing failures and indirectly reduces costs. With the development of machine learning and the maturity of data mining technology, the frontier direction of SOFC performance degradation prediction is mainly achieved by data-driven methods, learning system behavior from historical operating data and predicting system status. Existing technologies can be divided into two categories: non-neural network algorithms and neural network algorithms. Existing algorithms often have the following disadvantages:

[0003] ① The SOFC system has strong coupling between heat, force and electricity, which leads to mutual influence between system parameters and is easily affected by noise. However, most existing technologies are trained based on data generated by models, and their training sample structures are relatively simple, which may not be applicable to actual complex systems.

[0004] ②SOFC systems are often in long-term operation, and the data set used for performance prediction has a large time span, and most of the existing technologies may not be able to fit it well.

[0005] ③ The performance degradation of SOFC stack usually uses voltage as a reference indicator. The existing technology is too focused on fitting the actual voltage signal and ignores the trend information of voltage drop, which is not conducive to the assessment of the life limit of the stack. Summary of the invention

[0006] In order to solve the problems raised in the above background technology, the technical solution adopted by the present invention is:

[0007] A method for predicting SOFC system performance based on LSTM-ARMA fusion algorithm includes the following steps:

[0008] S1. Preprocess the original data;

[0009] S2, using the ARMA model optimized by robust Kalman filter to predict the degradation trend of SOFC;

[0010] S3, using LSTM algorithm to predict the output voltage of SOFC, combining the degradation trend predicted by ARMA model to output the reconstructed voltage, and calculating the loss function of the current node, and updating the weight parameters of each LSTM node according to the loss function;

[0011] S4, repeating steps S2-S3, updating the ARMA model state, and completing the prediction of multiple nodes in the LSTM algorithm in sequence until the maximum iteration round is reached, or the loss function of the current node meets the preset reconstruction error threshold requirement, and the parameter optimization is completed;

[0012] S5. Based on the LSTM-ARMA model after parameter optimization, the performance prediction of the SOFC system is realized.

[0013] Furthermore, in step S1, when the original data is preprocessed, an empirical analysis method or a PCA method is used to reduce the data dimension, while removing abnormal data and normalizing the data.

[0014] Further, in step S2, specifically:

[0015] Assuming that the order of SOFC system noise is ARMA(1,1), the ARMA model for predicting SOFC degradation trend is expressed as follows:

[0016] z(k)=a1z(k-1)+b1ε(k-1)+ε(k) (1)

[0017] Among them, z(k) is the current prediction point output, ε(k) is white noise, a1, b1 are constants;

[0018] The input parameters of ARMA(1,1) are used as state variables: X(k) = [a1(k), b1(k)] T ;

[0019] In the Kalman filtering process, equation (1) is rewritten as:

[0020]

[0021] in:

[0022]

[0023] is the observation matrix, v(k) is the virtual measurement noise caused by the white noise ε(k) and the measurement matrix error ΔH(k);

[0024] Since the performance degradation process of the SOFC system is a long-term process, the ARMA model is trained through a large-scale training data sample set so that the ARMA model parameters do not change with the new random noise samples, that is, Thus, the model parameters a1 and b1 are determined.

[0025] Furthermore, in step S3, when calculating the loss function L of the current i-th node, the cross entropy definition is adopted:

[0026]

[0027] Among them, h(i) represents the predicted output result of the current node, y i is the actual measured value.

[0028] Further, in step S4, the reconstruction error threshold is represented by ξ, and when the preset maximum iteration round is reached, or the loss function of the current node satisfies L<ξ, the parameter optimization is completed.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The SOFC system performance prediction method based on the LSTM-ARMA fusion algorithm provided by the present invention proposes an ARMA model optimized by LSTM combined with Kalman filter to solve the performance degradation problem of SOFC stacks; the "three-gate structure" of LSTM helps to capture the long-term and short-term dependencies of input time series data, prevents the abnormality of gradient information during information transmission, and can well complete the prediction of long-term voltage information in the SOFC degradation performance prediction process; and by combining the ARMA model optimized by Kalman filter to perform parameter optimization, the data interference caused by the complex coupling relationship between system parameters and the interference of system noise that may be generated during actual operation are eliminated to the greatest extent, thereby improving the system noise robustness.

[0031] The present invention solves the problem that the existing algorithm model is easily disturbed by system noise and environmental noise during the actual operation of the SOFC system, and can more reliably predict the degradation performance. In addition, compared with the existing algorithm that follows the actual voltage value, the present invention retains the system's performance degradation trend information while using LSTM to fit the voltage data, and can more accurately predict the life of the battery stack. On the premise of ensuring safe and efficient production, the utilization of the battery stack is maximized, thereby indirectly reducing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic diagram of the flow chart of the SOFC system performance prediction method based on the LSTM-ARMA fusion algorithm provided by the present invention. DETAILED DESCRIPTION

[0033] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following further describes how the present invention is implemented in conjunction with the accompanying drawings and specific implementation methods.

[0034] In a specific embodiment, referring to Figure 1 As shown, the present invention provides a SOFC system performance prediction method based on LSTM-ARMA fusion algorithm, comprising the following steps:

[0035] S1. Preprocess the raw data. When preprocessing the raw data, due to the complexity of the SOFC model and the large data set collected, empirical analysis or PCA method can be used to reduce the data dimension, which can greatly reduce the amount of calculation; at the same time, remove abnormal data and normalize the data to increase the reliability of the algorithm.

[0036] S2. The ARMA model (Auto-Regressive Moving Average Model) optimized by the robust Kalman filter is used to predict the degradation trend of SOFC.

[0037] Specifically: Assuming that the order of SOFC system noise is ARMA(1,1), the ARMA model for predicting SOFC degradation trend is expressed as follows:

[0038] z(k)=a1z(k-1)+b1ε(k-1)+ε(k) (1)

[0039] Among them, z(k) is the current prediction point output, ε(k) is white noise, a1, b1 are constants;

[0040] The input parameters of ARMA(1,1) are used as state variables: X(k) = [a1(k), b1(k)] T ;

[0041] In the Kalman filtering process, equation (1) is rewritten as:

[0042]

[0043] in:

[0044]

[0045] is the observation matrix, v(k) is the virtual measurement noise caused by the white noise ε(k) and the measurement matrix error ΔH(k);

[0046] Since the performance degradation process of the SOFC system is a long-term process (about 3% / kh), the scale of its training data sample set is usually large. By training the ARMA model with a large-scale training data sample set, the ARMA model can be fully learned and trained. The ARMA model parameters do not change with new random noise samples and have strong noise robustness, that is, Thus, the model parameters a1 and b1 are determined. At the same time, the system also reaches stability, that is, X(k) = X(k-1). At the same time, the state transfer matrix can be obtained by mathematical calculation. Considering that the system noise impact is 0 at this time, the iteration of the model parameters at the current moment is completed.

[0047] S3. Use LSTM (Long and Short-Term Memory Network) algorithm to predict the output voltage of SOFC, combine the degradation trend predicted by ARMA model to output the reconstructed voltage, calculate the loss function of the current node, and update the weight parameters of each LSTM node according to the loss function.

[0048] Recursive neural networks can be considered as repeated series of the same neural network, where the output of the previous neural network is used as the input of the next neural network to achieve chain extraction of information. By learning from historical information, future data can be predicted. As an improved version of the recursive neural network, LSTM uses three gates to adjust the traditional recursive neural network, so that it can retain more useful input information, with stronger scalability and more flexible structure and function.

[0049] Compared with RNN, the structure of LSTM input gate, forget gate, and output gate helps capture the long-term and short-term dependencies of input time series data and prevent the gradient reduction or explosion of information transmission. The key to LSTM's long-term memory is to keep the input information of each step in the storage unit. The hidden layer state of each output contains all the input information before the current moment, and then gradually learns the historical data, which can complete the time series prediction work well.

[0050] In addition, when calculating the loss function L of the current i-th node, the cross entropy definition is used:

[0051]

[0052] Among them, h(i) represents the predicted output result of the current node, y i is the actual measured value.

[0053] S4. Repeat steps S2-S3, update the ARMA model state, and complete the prediction of multiple nodes in the LSTM algorithm in sequence until the maximum iteration round is reached, or the loss function of the current node meets the preset reconstruction error threshold requirement, and the parameter optimization is completed.

[0054] The reconstruction error threshold is represented by ξ. When the preset maximum number of iterations is reached or the loss function of the current node satisfies L<ξ, the parameter optimization is completed.

[0055] S5. Based on the LSTM-ARMA model after parameter optimization, the performance prediction of the SOFC system is realized.

[0056] In summary, the SOFC system performance prediction method based on the LSTM-ARMA fusion algorithm provided by the present invention proposes an ARMA model optimized by LSTM combined with Kalman filter to address the performance degradation problem of SOFC stacks; the "three-gate structure" of LSTM helps to capture the long-term and short-term dependencies of input time series data, prevent the abnormality of gradient information during information transmission, and can well complete the prediction of long-term voltage information in the process of SOFC degradation performance prediction; and by combining the ARMA model optimized by Kalman filter to perform parameter optimization, the data interference caused by the complex coupling relationship between system parameters and the interference of system noise that may be generated during actual operation are eliminated to the greatest extent, thereby improving the system noise robustness.

[0057] The present invention solves the problem that the existing algorithm model is easily disturbed by system noise and environmental noise during the actual operation of the SOFC system, and can more reliably predict the degradation performance. In addition, compared with the existing algorithm that follows the actual voltage value, the present invention retains the system's performance degradation trend information while using LSTM to fit the voltage data, and can more accurately predict the life of the battery stack. On the premise of ensuring safe and efficient production, the utilization of the battery stack is maximized, thereby indirectly reducing costs.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A SOFC system performance prediction method based on LSTM-ARMA fusion algorithm, characterized in that: The steps include: S1. Preprocess the original data; S2, using the ARMA model optimized by robust Kalman filter to predict the degradation trend of SOFC; S3, using LSTM algorithm to predict the output voltage of SOFC, combining the degradation trend predicted by ARMA model to output the reconstructed voltage, and calculating the loss function of the current node, and updating the weight parameters of each LSTM node according to the loss function; S4, repeating steps S2-S3, updating the ARMA model state, and completing the prediction of multiple nodes in the LSTM algorithm in sequence until the maximum iteration round is reached, or the loss function of the current node meets the preset reconstruction error threshold requirement, and the parameter optimization is completed; Repeat steps S2 and S3, update the ARMA model status, and complete the prediction of multiple nodes in the LSTM algorithm in sequence until the maximum iteration round is reached or the loss function meets the preset threshold requirements. S5. Based on the LSTM-ARMA model after parameter optimization, the performance prediction of the SOFC system is realized.

2. The SOFC system performance prediction method based on the LSTM-ARMA fusion algorithm according to claim 1 is characterized in that: In step S1, when preprocessing the original data, an empirical analysis method or a PCA method is used to reduce the data dimension, remove abnormal data, and normalize the data.

3. The SOFC system performance prediction method based on LSTM-ARMA fusion algorithm according to claim 1 is characterized in that: In step S2, specifically: Assuming that the order of SOFC system noise is ARMA(1,1), the ARMA model for predicting SOFC degradation trend is expressed as follows: z(k)=a1z(k-1)+b1ε(k-1)+ε(k) (1) Among them, z(k) is the current prediction point output, ε(k) is white noise, a1, b1 are constants; The input parameters of ARMA(1,1) are used as state variables: X(k) = [a1(k), b1(k)] T ; In the Kalman filtering process, equation (1) is rewritten as: in: is the observation matrix, v(k) is the virtual measurement noise caused by the white noise ε(k) and the measurement matrix error ΔH(k); Since the performance degradation process of the SOFC system is a long-term process, the ARMA model is trained through a large-scale training data sample set so that the ARMA model parameters do not change with the new random noise samples, that is, Thus, the model parameters a1 and b1 are determined.

4. The SOFC system performance prediction method based on LSTM-ARMA fusion algorithm according to claim 3 is characterized in that: In step S3, when calculating the loss function L of the current i-th node, the cross entropy definition is used: Among them, h(i) represents the predicted output result of the current node, y i is the actual measured value.

5. The SOFC system performance prediction method based on LSTM-ARMA fusion algorithm according to claim 4 is characterized in that: In step S4, the reconstruction error threshold is represented by ξ. When the preset maximum number of iterations is reached or the loss function of the current node satisfies L<ξ, the parameter optimization is completed.

Images