Power station condenser backpressure prediction method suitable for few samples
Through deep neural network and transfer learning technology, a pre-trained network is established using data from the same type of unit, and the newly built units are fine-tuned, solving the problem of low back pressure prediction accuracy of the newly built units, improving the adaptability and accuracy of the model, and reducing the The consumption of computing resources.
Patent Information
- Application Number
- CN202411110569.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-22
AI Technical Summary
Due to the small amount of historical data, it is difficult to achieve high-precision condenser backpressure prediction. The existing model generalization capabilities are insufficient and cannot adapt to the differences between different units.
Deep neural network (DNN) combined with transfer learning, a pre-trained network is established using complete data from the same type of unit, and fine-tuning is carried out through a small amount of data from the target domain unit to build a backpressure prediction model suitable for the target domain unit.
It effectively reduces the impact of differences between units on back pressure prediction, improves the adaptability and prediction accuracy of the model, and reduces the consumption of computing resources.
Smart Images

Figure CN119003989A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of back pressure prediction of coal-fired power generation units, and in particular to a power plant condenser back pressure prediction method suitable for a small number of samples. Background Art
[0002] Under the guidance of the "dual carbon" goal, the main energy source of my country's power system has been transformed from traditional fossil energy to renewable energy. Non-fossil energy power generation installed capacity will exceed thermal power generation capacity for the first time in 2023, and the proportion of total installed capacity will exceed 50% for the first time in 2023, but coal-fired power generation still accounts for nearly 60% of the power generation. With the continuous increase in the proportion of new energy installed capacity, traditional coal-fired power plants need to further improve their operating level to ensure that the units have sufficient flexibility to cope with the strong randomness and volatility of new energy power generation such as wind and light, and give full play to their basic support and regulation role, so as to ensure the stable operation of the power grid.
[0003] As an important auxiliary equipment of coal-fired power generation units, the condenser mainly provides a stable low-temperature environment for the unit, thereby condensing the exhaust steam from the low-pressure cylinder and maintaining the vacuum at the exhaust port of the turbine. As the most important parameter of the cold end of the steam turbine, the condenser back pressure directly affects the economy and safety of the unit operation. Generally speaking, the lower the condenser back pressure, the better the performance. The condenser performance will directly affect the overall performance of the generator set in terms of power output and thermal efficiency. Therefore, ensuring the normal operation of the condenser has an extremely important impact on the smooth and efficient operation of the unit.
[0004] At present, the back pressure prediction method for condensers mainly adopts the data-driven method. This method is based on historical data to directly obtain the correlation between operating parameters. It does not need to understand the complex characteristic relationship, and the prediction effect is closely related to the completeness of the data sample. The new unit has a short operation time and a small amount of accumulated historical data. The data-driven method lacks sufficient training samples. The established model cannot learn the characteristics under unknown conditions and cannot achieve high-precision back pressure prediction. The same type of units have been in operation for a long time, and the database has a large amount of data and complete types. However, when the model trained directly based on the unit data is applied to the new unit, the differences between units will lead to insufficient model generalization ability. How to effectively avoid the above problems and ensure that the model prediction accuracy meets the actual operation needs has become one of the problems that need to be solved urgently. Summary of the invention
[0005] Purpose of the invention: In order to overcome the shortcomings of the prior art, a method for predicting the back pressure of a power plant condenser suitable for a small number of samples is provided, a deep neural network (DNN) is used to predict the back pressure of the condenser, the idea of transfer learning is adopted, and the network is fine-tuned using the target domain unit data to obtain a prediction model suitable for the target domain unit, thereby improving the adaptability of back pressure prediction in actual production applications.
[0006] Technical solution: To achieve the above purpose, the present invention provides a power plant condenser back pressure prediction method suitable for a small number of samples, comprising the following steps:
[0007] S1: Collect historical operation data of two units of the same type, one of which is the target domain unit and the other is the source domain unit;
[0008] S2: pre-processing the collected operation data;
[0009] S3: Use the pre-processed source domain unit operation data as source domain data to train the built deep neural network to obtain a pre-trained network suitable for the source domain unit;
[0010] S4: input a set amount of target domain unit operation data to retrain the pre-trained network structure to obtain a target network;
[0011] S5: Input the remaining target domain unit operation data into the trained target network as test samples to obtain the predicted value of condenser back pressure.
[0012] Furthermore, the preprocessing in step S2 includes:
[0013] A1: Select input parameters related to condenser back pressure through mechanism analysis and correlation analysis;
[0014] A2: Perform data cleaning, missing value processing and normalization on the historical operating data of the two units; 1 As the target domain data, the preprocessed source domain unit dataset D 2 As the source domain data; select D 1 Specified quantity data D 1tr As the input of backpressure prediction, the remaining data is recorded as D 1te .
[0015] Furthermore, the method of correlation analysis in step A1 is:
[0016] Grey correlation analysis compares the changing trend of data series and quantitatively calculates the correlation with the reference series, thereby judging the correlation between variables and the main influencing factors;
[0017] Among them, the correlation degree r of the sequence to be compared i The definition is as follows:
[0018]
[0019] Δ oi (k)=|x o (k)-x i (k)|
[0020] Among them, Δ oi (k) is the difference between the comparison sequence and the reference sequence, where the maximum difference and the minimum difference are denoted by Δ max and Δ min , x o (k) is the reference sequence, x i (k) is the sequence to be compared, i = 0, 1, 2, ..., n; k = 1, 2, ..., m, n is the number of samples, m is the number of indicators to be compared, ρ is the discrimination coefficient, ρ∈(0,1);
[0021] Based on the Pearson coefficient of two variables, determine whether there is a correlation between the variables;
[0022] Among them, Pearson coefficient is defined as follows:
[0023]
[0024] in, The parameters x are a ,x b The average value, n is the number of data samples, the absolute value The larger the parameter x is, the a ,x b The stronger the correlation between them.
[0025] Furthermore, the data cleaning and missing value processing in step A2 are specifically as follows: removing data samples without values for a long time and data samples of shutdown parts, and removing abnormal data points; wherein the characteristic thresholds are: back pressure greater than 30kPa; unit load greater than 670MW and less than 100MW;
[0026] Normalization processing: Convert the input parameter samples of the two units into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows:
[0027]
[0028] Among them, z is the normalized value, x is the original value of the sample data, μ is the mean of the sample data, and σ is the standard deviation of the sample data.
[0029] Furthermore, the pre-trained network in step S3 includes a five-layer network structure, namely an input layer, three hidden layers and an output layer; the input layer is used for data input; each hidden layer includes multiple neurons, which are used to perform feature learning on the input data samples and predict the condenser back pressure to obtain the corresponding predicted value; the output layer is used for outputting the results.
[0030] Furthermore, each layer of the hidden layer neurons in the pre-trained network has 10 neurons, that is, the hidden layer output is a 10-dimensional vector, and weighted summation and activation are performed on each layer of neurons. The weighted summation is a linear combination of all neuron outputs in the previous layer, and an activation function is applied to the linear combination result to provide nonlinear changes for the neural network; the hidden layer output, weighted summation, application of activation function and the adopted ReLU activation function are expressed as follows:
[0031]
[0032]
[0033] A [l] =g(Z [l] )
[0034]
[0035] Among them, A [l] is the output of the lth hidden layer, Z [l] is the weighted sum of the lth hidden layer, W [l] is the weight between the lth hidden layer neuron and the l-1th hidden layer neuron, b [l] is the constant bias in the linear combination, g ReLU (x) is the activation function.
[0036] Furthermore, the training process of the pre-trained network in step S3 is:
[0037] The training samples are input into the neural network for forward propagation. After the loss function is calculated by comparing with the true value, the partial derivatives of the loss function with respect to the weight and bias are calculated through back propagation. Then, the principle of gradient descent is used to iteratively update each parameter using partial derivatives to minimize the network loss function. The calculation formulas for forward propagation, loss function, partial derivatives, and iterative updates are as follows:
[0038] A [l] =g [l] (W [l] A [l-1] +b [l] ),l=1,2,…,L
[0039]
[0040] W [l] =W [l] -lrdW [l] ,b [l] =b [l] -lrdb [l]
[0041] Where L is the number of hidden layers, g[l] is the activation function from the l-1th hidden layer to the lth hidden layer, E loss is the loss function of the network, m is the number of training samples, and j are the network output value and true value of the jth training sample, dW [l] and db [l] are the partial derivatives of the loss function with respect to the weights and biases of the lth hidden layer, respectively, and lr is the learning rate that determines the update step size of the weights and biases in the gradient direction.
[0042] Furthermore, the learning rate attenuation strategy used in the training process of the pre-trained network is the Adam algorithm, and the formulas for the optimization process, the biased estimation correction of the gradient, and the parameter update method are as follows:
[0043]
[0044]
[0045] Among them, v t and t is a biased estimate of the first and second moments of the gradient, β 1 is a hyperparameter, is the loss function E about the model parameters θ loss The gradient of (θ), and The biased estimates of the first and second moments of the gradient are corrected respectively, and ∈ is a small constant to prevent division by zero.
[0046] Furthermore, in the training process of the target network in step S4: based on the pre-trained network structure and parameters, by fixing part of the model structure, using the optimized weights and biases as initialization values, supplementing a small amount of target domain data D 1tr Fine-tune the hidden layer structure and parameters except the last layer, initialize the last layer parameters, and then use the target domain data D 1tr The training obtains the new last layer weights and biases. The amount of training data and the number of training steps required for fine-tuning will be greatly reduced compared to directly obtaining a large amount of data for training, saving the computing resources required for model training.
[0047] The present invention introduces a transfer learning method to solve the problem of model accuracy of newly built units. After establishing the model using complete data samples in the database of units of the same type, the prediction model is fine-tuned through the model migration method, thereby enhancing the prediction accuracy of the model and providing a guarantee for the flexible adjustability of the unit.
[0048] Beneficial effect: Compared with the prior art, the present invention constructs a pre-trained prediction model for the source domain unit based on a deep neural network; at the same time, with the help of the idea of transfer learning, some structures and parameters of the pre-trained model are fixed, the deeper hidden layer weights and biases of the model network are reinitialized, and a small amount of data from the target domain unit is used to fine-tune the model, so as to obtain a target domain condenser back pressure prediction model suitable for different units of the same type, thereby achieving a lower prediction error, effectively reducing the impact of differences between motors on back pressure prediction, and improving the adaptability of condenser back pressure prediction in actual production. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 is a flow chart of the method of the present invention;
[0050] Figure 2 It is a DNN network structure diagram in the present invention;
[0051] Figure 3 It is a diagram of the training process of the source network in the present invention;
[0052] Figure 4 A diagram showing the fine-tuning process of the target network in the present invention;
[0053] Figure 5 A process diagram of network migration in the present invention;
[0054] Figure 6 It is the target domain unit data distribution diagram in the present invention;
[0055] Figure 7 This is a comparison chart of the synthetic sample output before and after network migration in the present invention;
[0056] Figure 8 This is a comparison chart of evaluation indicators before and after network migration in the present invention. DETAILED DESCRIPTION
[0057] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0058] like Figure 1 As shown, the present invention provides a method for predicting the back pressure of a power plant condenser suitable for a small number of samples, comprising the following steps:
[0059] S1: Use sensors to collect the only spring operation-related historical data sample set of a newly built unit (target domain unit) and the complete historical data sample set of the same type of unit (source domain unit) throughout the year.
[0060] S2: Preprocess the collected operation data, including:
[0061] A1: Select input parameters related to condenser back pressure through mechanism analysis and correlation analysis, including unit load, exhaust temperature, fan speed and ambient temperature;
[0062] The method of correlation analysis is:
[0063] Grey correlation analysis compares the changing trend of data series and quantitatively calculates the correlation with the reference series, thereby judging the correlation between variables and the main influencing factors;
[0064] Among them, the correlation degree r of the sequence to be compared i The definition is as follows:
[0065]
[0066] Δ oi (k)=|x o (k)-x i (k)|
[0067] Among them, Δ oi (k) is the difference between the comparison sequence and the reference sequence, where the maximum difference and the minimum difference are denoted by Δ max and Δ min , x o (k) is the reference sequence, x o (k) is the sequence to be compared, i = 0, 1, 2, ..., n; k = 1, 2, ..., m, n is the number of samples, m is the number of indicators to be compared, ρ is the discrimination coefficient, ρ∈(0,1);
[0068] Based on the Pearson coefficient of two variables, determine whether there is a correlation between the variables;
[0069] Among them, Pearson coefficient is defined as follows:
[0070]
[0071] in, The parameters x are a ,x b The average value, n is the number of data samples, the absolute value The larger the parameter x is, the a ,x b The stronger the correlation between them;
[0072] A2: Perform data cleaning, missing value processing and normalization on the historical operation data of the two units; 1As the target domain data, the preprocessed source domain unit dataset D 2 As the source domain data; select D 1 Specified quantity data D 1tr As the input of backpressure prediction, the remaining data is recorded as D 1te ;
[0073] Data cleaning and missing value processing are as follows: remove data samples with no values for a long time and data samples of shutdown parts, remove abnormal data points; remove abnormal samples with back pressure greater than 30kPa and unit load greater than 670MW and less than 100MW;
[0074] Normalization processing: Convert the input parameter samples in the two data sets into distributions with a mean of 0 and a standard deviation of 1. The calculation formula is as follows:
[0075]
[0076] Among them, z is the normalized value, x is the original value of the sample data, μ is the mean of the sample data, and σ is the standard deviation of the sample data.
[0077] S3: Use the preprocessed source domain unit operation data as the source domain data to train the built deep neural network, obtain a pre-trained network suitable for the source domain unit, and save the pre-trained network structure parameters;
[0078] like Figure 2 As shown, the pre-trained network includes a five-layer network structure, namely an input layer, three hidden layers and an output layer; the input layer is used for data input; each hidden layer includes multiple neurons, which are used to learn the features of the input data samples and predict the condenser back pressure to obtain the corresponding predicted value; the output layer is used to output the results;
[0079] Each layer of the hidden layer neurons in the pre-trained network has 10 neurons, that is, the hidden layer output is a 10-dimensional vector. Weighted summation and activation are performed on each layer of neurons. The weighted summation is a linear combination of all neuron outputs in the previous layer. The activation function is applied to the linear combination result to provide nonlinear changes for the neural network. The hidden layer output, weighted summation, activation function application and the adopted ReLU activation function are expressed as follows:
[0080]
[0081] Z [l] =W [l] A [l-1] +b [l]
[0082]
[0083] A[l] =g(Z [l] )
[0084]
[0085] Among them, A [l] is the output of the lth hidden layer, Z [l] is the weighted sum of the lth hidden layer, W [l] is the weight between the lth hidden layer neuron and the l-1th hidden layer neuron, b [l] is the constant bias in the linear combination, g ReLU (x) is the activation function;
[0086] Reference Figure 3 , the training process of the pre-trained network is:
[0087] The training samples are input into the neural network for forward propagation. After the loss function is calculated by comparing with the true value, the partial derivatives of the loss function with respect to the weight and bias are calculated through back propagation. Then, the principle of gradient descent is used to iteratively update each parameter using partial derivatives to minimize the network loss function. The calculation formulas for forward propagation, loss function, partial derivatives, and iterative updates are as follows:
[0088] A [l] =g [l] (W [l] A [l-1] +b [l] ),l=1,2,…,L
[0089]
[0090] W [l] =W [l] -lrdW [l] ,b [l] =b [l] -lrdb [l]
[0091] Where L is the number of hidden layers, g [l] is the activation function from the l-1th hidden layer to the lth hidden layer, E loss is the loss function of the network, m is the number of training samples, and j are the network output value and true value of the jth training sample, dW [l] and db [l] are the partial derivatives of the loss function with respect to the weights and biases of the lth hidden layer, respectively, and lr is the learning rate that determines the update step size of the weights and biases in the gradient direction.
[0092] The learning rate attenuation strategy used in the training process of the pre-trained network is the Adam algorithm. The formulas for the optimization process, the correction of the biased estimation of the gradient, and the parameter update method are as follows:
[0093]
[0094]
[0095] Among them, v t and t is a biased estimate of the first and second moments of the gradient, β 1 is a hyperparameter, is the loss function E about the model parameters θ loss The gradient of (θ), and The biased estimates of the first and second moments of the gradient are corrected respectively, and ∈ is a small constant to prevent division by zero.
[0096] S4: Input 10% of the target domain unit operation data D 1tr Retrain the pre-trained network structure to obtain the target network;
[0097] like Figure 4 As shown in Figure 1, during the training process of the target network: read the structure and parameters obtained by the pre-trained network training, freeze the weights and biases of the first two hidden layers, fine-tune the last layer of the pre-trained network, initialize the weights and biases of this layer of the network and use the target domain data D 1tr Retrain, obtain new weights and biases, and get the target network; the amount of training data and the number of training steps required for fine-tuning will be greatly reduced compared to directly obtaining a large amount of data for training, saving the computing resources required for model training.
[0098] S5: The remaining data D of the target domain unit operation data 1te The trained target network is input as a test sample, and the predicted value of the condenser back pressure is obtained as output.
[0099] like Figure 5 As shown, the present invention transfers the pre-trained network model and inputs D 1tr Get the target network model, for the data set D 1te Perform the test and obtain the back pressure prediction value.
[0100] In order to verify the effectiveness of the method of the present invention, the above scheme is applied as an example in this embodiment, as follows:
[0101] This embodiment takes a 660MW unit air-cooled condenser system as an example. Historical operation data is collected from the target domain unit and another unit of the same type, namely the source domain unit. Specifically, Figure 6As shown in FIG. , it is the target domain unit data distribution diagram in the present invention. 1tr The initial target domain network model is trained, and then the pre-trained network model and the target domain network model are established according to the above steps. Synthetic samples are used to evaluate the interpretability of the three models respectively. Except for the load and ambient temperature, the input parameters of the synthetic samples remain unchanged. The load of the synthetic samples is set to 200MW to 600MW, with a step size of 100MW, and an additional group is added at the highest load of 660MW. The circulating water inlet temperature range of the synthetic samples is 20℃ to 36℃, with a step size of 1℃. Specifically, Figure 7 As shown, it is a comparison diagram of the output of the three network models of the present invention for synthetic samples. Figure 7 (a) shows the output result of the initial target domain network model. In the low temperature range with high sample coverage, the back pressure prediction effect is better. However, when the temperature is high, the prediction effect is seriously deviated due to incomplete samples in the data set. Figure 7 (b) shows the output result of the pre-trained network model. Since the samples in the source domain data set are complete, the prediction effect of the overall synthetic sample conforms to the actual operating rules. However, due to the differences between the units and insufficient generalization ability, the prediction accuracy of the back pressure of the target domain units is reduced. Figure 7 (c) shows the output result of the target domain network model. The target domain network model fine-tuned using part of the target domain data retains the parameter relationship of the initial pre-training model while enhancing the generalization ability, making the prediction of the back pressure of the target domain unit more accurate.
[0102] Specifically, Figure 8 The figure shows the comparison of evaluation indicators before and after network migration in the present invention. 2 ) Comparison of the three models on the target domain data D 1te As can be seen from the figure, the R 2 It is obviously higher than the other two models, thus verifying the effectiveness of the present invention.
[0103] The power plant condenser back pressure prediction method based on transfer learning proposed in the present invention can effectively retain the potential physical laws of the source network model, and ensure the generalization ability of the model through target domain data, and effectively realize the power plant condenser back pressure prediction under few samples.
Claims
1. A method for predicting the back pressure of a power plant condenser suitable for a small number of samples, characterized in that: The steps include: S1: Collect historical operation data of two units of the same type, one of which is the target domain unit and the other is the source domain unit; S2: pre-processing the collected operation data; S3: Use the pre-processed source domain unit operation data as source domain data to train the built deep neural network to obtain a pre-trained network suitable for the source domain unit; S4: input a set amount of target domain unit operation data to retrain the pre-trained network structure to obtain a target network; S5: Input the remaining target domain unit operation data into the trained target network as test samples to obtain the predicted value of condenser back pressure.
2. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 1, characterized in that: The pre-processing in step S2 includes: A1: Select input parameters related to condenser back pressure through mechanism analysis and correlation analysis; A2: Perform data preprocessing operations such as data cleaning, missing value processing and normalization on the historical operation data of the two units; use the preprocessed target domain unit data set D1 as the target domain data, and use the preprocessed source domain unit data set D2 as the source domain data; select the specified amount of data D in D1 1tr As the input of backpressure prediction, the remaining data is recorded as D 1te .
3. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 2, characterized in that: The method of correlation analysis in step A1 is: Grey correlation analysis compares the changing trend of data series and quantitatively calculates the correlation with the reference series, thereby judging the correlation between variables and the main influencing factors; Among them, the correlation degree r of the sequence to be compared i The definition is as follows: Δ oi (k)=|x o (k)-x i (k)| Among them, Δ oi (k) is the difference between the comparison sequence and the reference sequence, where the maximum difference and the minimum difference are denoted by Δ max and Δ min , x o (k) is the reference sequence, x i (k) is the sequence to be compared, i = 0, 1, 2, ..., n; k = 1, 2, ..., m, n is the number of samples, m is the number of indicators to be compared, ρ is the discrimination coefficient, ρ∈(0,1); Based on the Pearson coefficient of two variables, determine whether there is a correlation between the variables; Among them, Pearson coefficient is defined as follows: in, The parameters x are a ,x b The average value of , n is the number of data samples.
4. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 2, characterized in that: The data cleaning and missing value processing in step A2 specifically include: removing data samples without values for a long time and data samples of downtime, and removing abnormal data points; Normalization processing: Convert the input parameter samples of the two units into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: Among them, z is the normalized value, x is the original value of the sample data, μ is the mean of the sample data, and σ is the standard deviation of the sample data.
5. The method for predicting the back pressure of a power plant condenser suitable for a small number of samples according to claim 1, characterized in that: The pre-trained network in step S3 includes a five-layer network structure, namely an input layer, three hidden layers and an output layer; the input layer is used for data input; each hidden layer includes multiple neurons, which are used to perform feature learning on the input data samples and predict the condenser back pressure to obtain the corresponding predicted value; the output layer is used for outputting the results.
6. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 5, characterized in that: Each layer of neurons in the hidden layer of the pre-trained network has 10 neurons, that is, the output of the hidden layer is a 10-dimensional vector, and two steps of weighted summation and activation are performed on each layer of neurons. The weighted summation is to linearly combine the outputs of all neurons in the previous layer, and the activation function is applied to the linear combination result to provide nonlinear changes for the neural network; the hidden layer output, weighted summation, application of activation function and the adopted ReLU activation function are expressed as follows: A [l] =g(Z [l] ) Among them, A [l] is the output of the lth hidden layer, Z [l] is the weighted sum of the lth hidden layer, W [l] is the weight between the lth hidden layer neuron and the l-1th hidden layer neuron, b [l] is the constant bias in the linear combination, g ReLU (x) is the activation function.
7. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 6, characterized in that: The training process of the pre-trained network in step S3 is: The training samples are input into the neural network for forward propagation. After the loss function is calculated by comparing with the true value, the partial derivatives of the loss function with respect to the weight and bias are calculated through back propagation. Then, the principle of gradient descent is used to iteratively update each parameter using partial derivatives to minimize the network loss function. The calculation formulas for forward propagation, loss function, partial derivatives, and iterative updates are as follows: A [l] =g [l] (W [l] A [l-1] +b [l] ),l=1,2,…,L W [l] =W [l] -lrdW [l] ,b [l] =b [l] -lrdb [l] Where L is the number of hidden layers, g [l] is the activation function from the l-1th hidden layer to the lth hidden layer, E loss is the loss function of the network, m is the number of training samples, and j are the network output value and true value of the jth training sample, dW [l] and db [l] are the partial derivatives of the loss function with respect to the weights and biases of the lth hidden layer, respectively, and lr is the learning rate that determines the update step size of the weights and biases in the gradient direction.
8. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 7, characterized in that: The learning rate attenuation strategy used in the training process of the pre-trained network is the Adam algorithm. The formulas for the optimization process, the biased estimation correction of the gradient, and the parameter update method are as follows: Among them, v t and t is a biased estimate of the first and second moments of the gradient, β1 is a hyperparameter, is the loss function E about the model parameters θ loss The gradient of (θ), and The biased estimates of the first and second moments of the gradient are corrected respectively, and ∈ is a small constant to prevent division by zero.
9. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 6, characterized in that: In the training process of the target network in step S4: based on the pre-trained network structure and parameters, by fixing part of the model structure, using the optimized weights and biases as initialization values, a small amount of target domain data D is supplemented. 1tr Make fine adjustments.
10. A power plant condenser back pressure prediction method applicable to a small number of samples according to claim 9, characterized in that: In step S4, a small amount of target domain data D is supplemented 1tr The fine-tuning method is to freeze the hidden layer structure and parameters except the last layer, initialize the last layer parameters, and then use the target domain data D 1tr Train to get the new final layer weights and biases.
Citation Information
Patent Citations
Method and device for monitoring ash pollution condition of heat dissipation fins
CN111428747A
Indirect air cooling tower shutter regulation and control method giving consideration to freezing prevention and economical efficiency
CN117029557A
Gas turbine fault prediction and diagnosis method based on deep transfer learning
CN117588312A
Cited By
Circulating water energy-saving scheduling method and system based on model fusion
CN121836287A
