Deep learning based sofc system operating condition recognition method

Abstract

The application discloses a kind of SOFC system working condition identification method based on deep learning, comprising: step 1, acquisition SOFC system different working condition state data;Step 2, data preprocessing;Step 3, construct deep neural network;Step 4, network training;Step 5, utilize genetic algorithm optimization network parameter;Step 6, obtain final network model;Step 7, SOFC system working condition identification.The application is by mining the regularity of different historical working condition state data with time variation, and learning to historical data, obtains the identification result with working condition type as output variable, reaches the effect of identifying SOFC system working condition type, overcomes the shortcoming of insufficient identification rate of traditional learning mode.LSTM long short-term memory network is used to learn the characteristics of working condition data, and genetic algorithm is used to optimize parameters, so as to improve the identification accuracy of SOFC system working condition identification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high-temperature fuel cell operating condition recognition, and particularly relates to a SOFC system operating condition recognition method based on deep learning. BACKGROUND

[0002] As a new type of clean energy with hydrogen as raw material, SOFC system has been widely used in many fields including transportation and energy storage. However, its weak load tracking ability and poor reliability limit its large-scale industrial development. During the operation of the SOFC system, water flooding and membrane drying often occur. Water flooding refers to the continuous accumulation of liquid water in the stack, which blocks the gas diffusion layer, catalyst layer and even gas flow channel, thereby affecting the smooth progress of the electrochemical reaction in the stack. Membrane drying refers to the lack of liquid water in the stack, which hinders the water production of the membrane electrode, reduces the electrical conductivity and increases the impedance of the membrane. In severe cases, membrane drying can also cause local high temperature in the stack, thereby reducing the load tracking ability of the SOFC system. Meanwhile, once these faults occur, the performance of the automatic system or the vehicle equipped with the SOFC system will be reduced, and in severe cases, it will lead to irreparable consequences.

[0003] For the operating condition recognition problem of SOFC system, the current mainstream methods include two categories based on model and non-model. The operating condition recognition method based on model needs to establish a corresponding model according to the physical process of the SOFC system, and realizes operating condition recognition and separation through residual error analysis between model simulation results and actual system output. The non-model-based operating condition recognition method refers to obtaining fault information of the SOFC system based on expert system, signal processing or a combination of the two methods and using visualization technology. Compared with the operating condition recognition method based on model, this method does not need to establish a model of the SOFC system, and the process of operating condition recognition and separation is replaced by expert system simulating human reasoning activities, but the fault types need to be predefined. With the rapid development of artificial intelligence, non-model operating condition recognition method has become more and more popular, and has been widely used in the field of operating condition recognition in engineering. Because of the complex relationship between the occurrence of SOFC system faults and the change of its state quantity with time, ordinary network model has been difficult to realize high-precision recognition model, so it is necessary to use deep neural network to construct the model. SUMMARY

[0004] The purpose of the present application is to provide a SOFC system operating condition recognition method based on deep learning which can improve the recognition rate.

[0005] The technical scheme adopted by the present application is as follows: a SOFC system operating condition recognition method based on deep learning, comprising the following steps:

[0006] Step 1, collect SOFC system working condition data X in different working conditions for a period of time in seconds according to the needs of SOFC system working condition identification task i , where i represents the working condition state data type, including: stack voltage, current, net output current change rate, minimum single piece voltage and stack inlet and outlet temperature; its output type is working condition type Y j , where j represents different working condition types, including: stack short circuit, stack gas leakage, stack overload, catalyst poisoning and carbon deposition;

[0007] Step 2, generate time series data Z in seconds based on the SOFC system working condition data collected in step 1, which has a dimension of i+1;

[0008] Step 3, normalize the SOFC system working condition data Z obtained after step 2 to map it to [0, 1]; the formula is: The normalized data is Z 1 ; where Z min and Z max represent the minimum and maximum values of such working condition state data, respectively;

[0009] Step 4, divide the normalized SOFC system working condition data obtained in step 3 into training set Z train and test set Z test according to the ratio of 7:3;

[0010] Step 5, based on the working condition time series data of SOFC system, a deep network model is constructed, which includes input layer, LSTM layer, full connection layer and classification layer; according to the collected working condition data, the batch size a, the maximum iteration number b, the initial learning rate c and the learning rate reduction factor d are set;

[0011] Step 6, input the normalized SOFC system working condition state data training set Z train into the input layer as input data and weighted sum into the LSTM layer;

[0012] Step 7, after the SOFC system working condition state data training set Z train is input into the LSTM layer, it is processed in turn through the dropper layer and the full connection layer, and the result function is used for output to obtain the LSTM data, the result function uses the dimension tanh function output, which is as follows:

[0013]

[0014]

[0015] The fully connected layer fuses the LSTM data and the data extracted from the SOFC system operating condition data at this moment; the processed data is then classified by a classification layer; its output is the operating condition type Y. j ={1,2,...,j};

[0016] Step 8: Test set Z of SOFC system operating condition status data test Input the data into the trained network and analyze the results;

[0017] Step 9: Construct the accuracy index for SOFC system operating condition identification: Where N right For the predicted result Y pre Compared with actual result Y j The same number, N SUM This represents the total number of samples.

[0018] Step 10: Optimize the LSTM parameters of the SOFC system operating condition identification network using a genetic algorithm;

[0019] Step 11: Collect SOFC system operating status data from a period of time prior to this moment and input it into the network to identify the current SOFC system operating status.

[0020] Furthermore, step 6 includes:

[0021] Step 6.1, Z train Obtain forgotten information f through the forgetting gate t :f t =σ(w f ×[h t-1 Z train The LSTM network forgets features irrelevant to the SOFC operating condition; where h t-1 w represents the network state at the previous moment. f Here is the forget gate weight matrix;

[0022] Step 6.2, Z train The input information is obtained after passing through the input gate: i t =σ(w i ×[h t-1 Z train ]), Where i t The input gate output value is a data characteristic related to the SOFC operating condition. This represents the current state value of the neuron.

[0023] Step 6.3: When passing through the output gate, its output value is O. t =σ(w o ×[ht-1 , Z train + b o ); The predicted value of the SOFC operating condition type for this batch is h t = O t × tanhC t ; where w o is the output gate weight, b o is the output gate bias vector, and σ is the output gate activation function.

[0024] Further, the step 10 includes:

[0025] Step 10.1: Use the current network parameter values as the parent, copy M identical individuals, where 30 < M < 80; set the termination evolution generation G of the genetic algorithm, where 200 < G < 400; the crossover probability P C , 0.5 < P C < 0.8; the mutation probability P m , 0.001 < P m < 0.05;

[0026] Step 10.2: Use the SOFC system operating condition identification accuracy T rate as the individual fitness function F(x); select the objective function J(X) = max{F(X)};

[0027] Step 10.3: Randomly cross the parent to generate a large number of subsets and add them to the new individual library, use the individual fitness function to screen the individual library, and select the M individuals with the largest F(X) as the parent again; repeat the above steps until the termination evolution generation G is reached;

[0028] Step 10.4: Select the individual with the largest F(X) value in this individual library, and the corresponding LSTM network parameters are the optimal SOFC system operating condition identification network parameters, and thus the final SOFC system operating condition identification model is obtained.

[0029] The beneficial effects of the present invention:

[0030] By mining the regularity of different historical operating condition state data over time and learning a large amount of historical data, the present invention obtains the identification result with the operating condition type as the output variable, achieving the effect of SOFC system operating condition identification. It overcomes the disadvantage that the traditional learning method has a poor training effect on the stack state data that has a high dependence on time, uses the LSTM long short-term memory network to learn the characteristics of the operating condition data, and uses the genetic algorithm to optimize the network parameters, thereby achieving a better identification effect. Description of the Drawings

[0031] Figure 1 It is the flow chart of the method of the present invention. Detailed Embodiments​

[0032] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. The specific examples described herein are only used to explain the technical scheme of the present application and are not limited to the present application.

[0033] As shown in Figure 1 , a deep learning-based SOFC system operating condition recognition method includes the following steps:

[0034] Step 1, for the needs of SOFC system operating condition recognition task, collect operating condition data X in different operating conditions of SOFC system in a period of time in seconds i , where i represents the operating condition data type, including: stack voltage, current, net output current change rate, minimum single piece voltage and stack inlet and outlet temperature. Its output type is operating condition type Y j . Where j represents different operating condition types, including: stack short circuit, stack gas leakage, stack overload, catalyst poisoning and carbon deposition;

[0035] Step 2, generate time series data Z in seconds for the SOFC system operating condition state data collected in step 1, which has a dimension of i+1;

[0036] Step 3, normalize the SOFC system operating condition data Z obtained after step 2 to map it to [0, 1]; the formula is: The normalized data is Z 1 ; where Z min and Z max represent the minimum and maximum values of such operating condition data, respectively;

[0037] Step 4, divide the normalized SOFC system operating condition data obtained in step 3 into training set Z train and test set Z test in the ratio of 7:3;

[0038] Step 5, construct a deep network model based on the operating condition time series data of the SOFC system, which includes an input layer, an LSTM layer, a fully connected layer and a classification layer; set the batch size a, the maximum number of iterations b, the initial learning rate c and the learning rate drop factor d according to the collected operating condition data;

[0039] Step 6, input the normalized SOFC system operating condition state data training set Z train as input data to the input layer and weighted sum to the LSTM layer;

[0040] Step 6.1, Z train gets the forgetting information f through the forgetting gatet : f t = σ(w f × [h t-1 , Z train ]) is forgotten by the LSTM network, which is irrelevant to the SOFC working condition. Wherein h t-1 represents the network state at the last time, w f is the forgetting gate weight matrix;

[0041] Step 6.2, Z train goes through the input gate to get the input information: i t = σ(w i × [h t-1 , Z train ]), Where i t is the input gate output value, which is mainly the data features related to the SOFC working condition. is the current state value of the neuron;

[0042] Step 6.3, when going through the output gate, its output value O t = σ(w o × [h t-1 , Z train ]+b o ); The SOFC working condition type prediction value of this batch is h t = O t × tanhC t ; Wherein w o is the output gate weight, b o is the output gate bias vector, and σ is the output gate activation function;

[0043] Step 7, the SOFC system working condition state data training set Z train goes through the LSTM layer, then the dropper layer and the full connection layer in turn, and the LSTM data is obtained by using the result function for output. The result function uses the dimension tanh function for output, which is as follows:

[0044]

[0045]

[0046] The full connection layer fuses and processes the LSTM data and the data of this moment extracted from the SOFC system working condition state data; The processed data is classified by the classification layer. Its output is the working condition type Y j = {1, 2, …, j}.

[0047] Step 8, the SOFC system working condition state data test set Ztest Input into the trained network, and statistics results;

[0048] Step 9, build SOFC system working condition recognition accuracy index: Where N right is the same number of predicted results Y pre and actual results Y j , N SUM is the total number of samples;

[0049] Step 10, using genetic algorithm to optimize SOFC system working condition recognition network LSTM parameters, the process is:

[0050] Step 10.1, with the current network parameter value as the father, copy M (30<M<80) same individuals; set the genetic algorithm termination evolution algebra G (200<G<400); crossover probability P C (0.5<P C <0.8); mutation probability P m (0.001<P m <0.05);

[0051] Step 10.2, SOFC system working condition recognition accuracy T rate as the individual fitness function F(x); select the objective function J(X) = max{F(X)};

[0052] Step 10.3, with the father to produce a large number of subsets and join in the new individual library, using individual fitness function to screen the individual library, take F(X) the largest M individuals as the father; repeat the above steps until the termination evolution algebra G;

[0053] Step 10.4, select the individual library F(X) value of the largest individual, its corresponding LSTM network parameters for the optimal SOFC system working condition recognition network parameters, thus obtaining the final SOFC system working condition recognition model;

[0054] Step 11, collect the SOFC system working condition state data in a period of time from this moment, input into the network, and the SOFC system working condition at this time can be identified. For example, in order to identify the working condition type of the SOFC system, the data of different working condition types are divided into training set and verification set according to the proportion of 7:3, the training set is input into the LSTM network model built to train, the network parameters are random, and the working condition data includes five parameters of stack short circuit, stack gas leakage, stack overload, catalyst poisoning and carbon deposition. The LSTM network includes: input layer, LSTM layer, classification layer and output layer. Then the recognition accuracy of the network model is obtained by using the verification set, which is used as the fitness function of the genetic algorithm individual, each group of parameters is used as an individual, the initial parameters are used as the parents for iteration, and the individual with the highest fitness is taken out as the finally determined parameters when the maximum iteration is reached. The network model at this time is the final SOFC system working condition identification model. The state data at a certain moment is input into the model to identify the working condition.

[0055] The above only expresses the preferred embodiments of the present application, which are described in detail, but cannot be understood as limiting the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications, improvements and substitutions can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A deep learning-based SOFC system operating condition recognition method, characterized in that, The method comprises the following steps: Step 1, according to the needs of SOFC system working condition recognition task, collect the working condition data X of SOFC system under different working conditions for a period of time in seconds i where i represents the working condition state data type, including: stack voltage, current, net output current change rate, minimum single chip voltage and stack inlet and outlet temperature; its output type is working condition type Y j where j represents different working condition types, including: stack short circuit, stack gas leakage, stack overload, catalyst poisoning and carbon deposition; Step 2, generating time series data Z in seconds with a dimension of i+1 from the working condition data of the SOFC system collected in step 1; Step 3, normalizing the SOFC system working condition data Z obtained in step 2 to map to [0, 1]; the formula is: The normalized data is Z 1 ; wherein Z min and Z max respectively represent the minimum value and the maximum value of the working condition data of this type. Step 4, divide the normalized SOFC system operating data obtained in step 3 into a training set Z and a test set Z according to a ratio of 7:3 train . test ; Step 5, constructing a deep network model based on the working condition time series data of the SOFC system, wherein the deep network model comprises an input layer, an LSTM layer, a full connection layer and a classification layer; setting the batch size a, the maximum iteration number b, the initial learning rate c and the learning rate reduction factor d according to the collected working condition data; Step 6, training set Z of normalized SOFC system operating condition state data train As input data input to the input layer and weighted summation into the LSTM layer; Step 7, SOFC system working condition state data training set Z train After the input LSTM layer, the dropper layer and the full connection layer are processed in sequence, and the result function is used for output to obtain the LSTM data, the result function adopts a dimension tanh function output, and the dimension tanh function is as follows: The full connection layer fuses the LSTM data and the data extracted from the SOFC system operating condition state data at this moment; the processed data is classified by the classification layer; and the output is the operating condition type Y j ={1,2, …, j} Step 8, test set Z of SOFC system working condition state data test Input into the trained network and count the results; Step 9, constructing SOFC system working condition recognition accuracy index: Where N right is the number of identical Y pre and actual results Y j , N SUM is the total number of samples; Step 10, optimizing the working condition identification network LSTM parameters of the SOFC system by using a genetic algorithm; Step 11, collecting the working condition state data of the SOFC system in a period of time before the current time, and inputting the data into the network to identify the working condition of the SOFC system at the current time.

2. The deep learning-based SOFC system operating condition recognition method according to claim 1, characterized by, The step 6 comprises: Step 6.1, Z train obtained through the forgetting gate t : f t = σ(w f × [h t-1 , Z train ]), forgetting the features irrelevant to the SOFC working condition by the LSTM network; wherein h t-1 represents the network state at the last time, and w f is the forgetting gate weight matrix; Step 6.2, Z train Get input information through input gate: i t = σ(w i x [h t-1 , Z train ]), where i t is the input gate output value, which is the data feature related to the SOFC working condition; is the current state value of the neuron; Step 6.3, when passing through the output gate, its output value O t = σ(w o × [h t-1 , Z train ]+b o ); the current SOFC working condition type prediction value is h t = O t × tanhC t ; wherein w o is the output gate weight, b o is the output gate bias vector, and σ is the output gate activation function.

3. The deep learning-based SOFC system operating condition recognition method according to claim 1, characterized by, The step 10 comprises: Step 10.1, copy M same individuals with current network parameter values as parent, 30 < M < 80; set the termination evolution generation G of genetic algorithm, 200 < G < 400; crossover probability P C , 0.5 < P C < 0.8 two mutation probability P m , 0.001 < P m < 0.05; Step 10.2, SOFC system operating condition recognition accuracy T rate For individual fitness function F(x); select objective function J(X) = max{F(X)}; Step 10.3, generating a large number of subsets by random crossover with the parents and adding the subsets to a new individual library, screening the individual library by using an individual fitness function, taking M individuals with the largest F(X) as the parents, and repeating the above steps until the termination evolution generation G is reached; Step 10.4, selecting an individual with the largest F(X) in the individual library, and the LSTM network parameters corresponding to the individual are the optimal SOFC system working condition identification network parameters, thereby obtaining the final SOFC system working condition identification model.

Images