Nuclear power key index prediction device based on optimal deep learning

Abstract

The invention discloses a nuclear power key index prediction device based on optimal deep learning. The device comprises a sodium iodide spectrometer database, a data acquisition and preprocessing module, a one-dimensional CNN-LSTM cascade model module, a multi-target adaptive particle swarm optimization algorithm module and a nuclear power key index prediction and online correction module. According to the method, online prediction and online optimization of parameter optimization are realized, nuclear power key indexes in the operation process of nuclear power facilities can be predicted robustly and quickly, and prediction characteristics or rules do not need to be designed manually; the defects that an existing nuclear power key index prediction device is not high in precision and prone to being influenced by human factors, a prediction model is prone to mismatch, and therefore the prediction effect becomes poor and even a system breaks down are overcome, and the prediction efficiency and accuracy of long-term online operation of nuclear power key index prediction in the radiation environment are improved.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power key indicator prediction, deep learning and optimization algorithm, and in particular relates to a nuclear power key indicator prediction device based on optimal deep learning. Background Art

[0002] In the field of nuclear power, nuclear safety is the top priority of nuclear power operation. my country's radiation environment monitoring automatic stations have always adopted manual prediction and analysis methods, which are affected by expert human factors. At the same time, manual methods are very inaccurate and require a long prediction time. Therefore, high-precision prediction of key nuclear power indicators is an urgent problem and challenge that needs to be solved.

[0003] In recent years, deep learning methods have rarely been used in the prediction of key indicators of nuclear power in China. The use of deep learning algorithm models combined with advanced optimization algorithms can improve the accuracy, robustness and adaptability of predictions, thereby improving the safety and stability of nuclear power systems. The existing nuclear power key indicator prediction devices have the characteristics of low prediction accuracy, susceptibility to human factors, and easy mismatch of prediction models, which leads to poor prediction results or even system crashes, making it difficult to operate effectively online for a long time. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies in the prior art and provide a nuclear power key indicator prediction device based on optimal deep learning.

[0005] The object of the present invention is achieved through the following technical solutions: a nuclear power key indicator prediction device based on optimal deep learning, the device comprising a sodium iodide spectrometer database, a data acquisition and preprocessing module, a one-dimensional CNN-LSTM cascade model module, a multi-objective adaptive particle swarm optimization algorithm module and a nuclear power key indicator prediction and online correction module;

[0006] The data acquisition and preprocessing module is used to perform data preprocessing on the data x of the sodium iodide spectrometer input from the sodium iodide spectrometer database, obtain the preprocessed data x′ and upload it to the one-dimensional CNN-LSTM cascade model module;

[0007] The one-dimensional CNN-LSTM cascade model module is used to classify the preprocessed data x′ through the one-dimensional CNN-LSTM cascade model to obtain a classification vector y;

[0008] The multi-objective adaptive particle swarm optimization algorithm module is used to optimize the training times, learning rate, and LSTM time window size of the one-dimensional CNN-LSTM cascade model using an improved particle swarm optimization algorithm;

[0009] The nuclear power key indicator prediction and online correction module is used to update the one-dimensional CNN-LSTM cascade model online, regularly input the current data stream of the sodium iodide spectrometer into the training set, and use the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model.

[0010] Furthermore, the method of preprocessing the data x input from the sodium iodide spectrometer database to obtain preprocessed data x′ specifically includes the following sub-steps:

[0011] (a.1) Clean the data x, including deleting duplicate information and abnormal information, and obtain the cleaned data

[0012] (a.2) Then the cleaned data Perform data type conversion and clean the data The character features in are converted into numerical features to obtain data

[0013] (a.3) Then the data Normalize the data to obtain the preprocessed data x′. The calculation formula is as follows:

[0014]

[0015] in, For data The eigenvalue at the i-th position in ; For data The smallest eigenvalue in ; For data The largest eigenvalue in i ′ is the eigenvalue at the i-th position in the preprocessed data x′.

[0016] Furthermore, the one-dimensional CNN-LSTM cascade model module includes a one-dimensional CNN-LSTM cascade model, and the specific structure of the one-dimensional CNN-LSTM cascade model is:

[0017] The first layer is a one-dimensional convolution layer: the convolution kernel size is 16×1, the stride is 2, and the number of convolution kernels is 64;

[0018] The second layer, the maximum pooling layer: the pooling window size is 2;

[0019] The third layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64; the second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64;

[0020] The fourth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 2, and the number of convolution kernels is 64; the second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64;

[0021] The fifth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128. The second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128.

[0022] The sixth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 2, and the number of convolution kernels is 128. The second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128.

[0023] The seventh layer, the fully connected layer, has 256 neurons, followed by a batch standard layer and a nonlinear transformation layer;

[0024] The eighth layer, LSTM layer: the number of units is 10, and the activation function uses the tanh function;

[0025] The ninth layer, LSTM layer: the number of units is 10, and the activation function uses the tanh function;

[0026] The tenth layer, the fully connected layer, has 128 neurons, followed by a batch standard layer and a nonlinear transformation layer;

[0027] Eleventh layer, classification layer: softmax classifier, classifies the results;

[0028] Each one-dimensional convolutional layer is followed by a nonlinear transformation layer and a batch normalization layer.

[0029] Furthermore, the multi-objective adaptive particle swarm optimization algorithm module uses an improved particle swarm optimization algorithm to optimize the number of training times, learning rate, and LSTM time window size of the one-dimensional CNN-LSTM cascade model, specifically:

[0030] (a.1) Initialize the particle swarm optimization algorithm parameters: population size P, maximum number of iterations t max ;

[0031] (a.2) Randomly initialize the particle swarm position r i and speed v i ;

[0032] (a.3) Calculate the fitness value F and calculate the historical optimal Lbest;

[0033] (a.4) According to the Pareto dominance principle, the current non-inferior solution is stored in the Archive set;

[0034] (a.5) Divide the target space into small areas with grids, and use the number of particles contained in each area as the density information of the particles; the more particles a particle contains in the grid, the greater its density value, and vice versa, the smaller it is, so as to calculate the crowdedness of the Archive set;

[0035] (a.6) Select the particle with the lowest density in the Archive set as the global optimal Gbest;

[0036] (a.7) Calculate the evolution factor f, which is the length of the sum of the connection vectors between the optimal particle and other particles divided by the sum of the distances between the optimal particle and other particles;

[0037] (a.8) Update the particle's velocity and position:

[0038]

[0039] ω is the inertia factor, c 1 、c 2 is the acceleration coefficient, generally 2; rand 1 、rand 2 is a random number between [0,1]; is the updated particle velocity, is the particle velocity before updating; r i t+1 is the updated particle position, r i t is the particle position before updating;

[0040] (a.9) Repeat steps (a.3) to (a.8) until the iteration stopping condition is met.

[0041] Furthermore, the nuclear power key indicator prediction and online correction module updates the one-dimensional CNN-LSTM cascade model online, regularly inputs the current data stream of the sodium iodide spectrometer into the training set, and uses the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model, specifically:

[0042] The nuclear power key indicator prediction and online correction module adopts an online correction strategy to perform real-time correction on the one-dimensional CNN-LSTM cascade model, by inputting abnormal data as new training data into the training set, and using the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model in real time;

[0043] When the one-dimensional CNN-LSTM cascade model is working, the prediction analysis value MI at time t is obtained in real time a (t) and the predicted value MI p (t), calculate the analytical value MI a and the predicted value MIp The deviation e o (t):

[0044] e o (t)=|MI a (t)-MI p (t)|;

[0045] If e o (t) is greater than a positive error tolerance ε o :

[0046] e o (t)>ε o ;

[0047] Then the predicted analysis value MI at time t a (t) and parameter value x(t) as new training data {x(t), MI a (t)} is added to the training data set, and the prediction model parameters are retrained to perform online correction of the prediction model; otherwise, the model parameters are considered to be accurate and no online correction is required.

[0048] The beneficial effects of the present invention are as follows: a nuclear power key indicator prediction device based on optimal deep learning provided by the present invention predicts sodium iodide spectrometer data, mines more distinguishable subtle features in multiple scales, and has higher prediction accuracy; a one-dimensional CNN-LSTM cascade model is used to realize the prediction of industrial control nuclear power key indicators, and a new improved particle swarm optimization algorithm is used to select optimization parameters, which partially overcomes the disadvantage that the traditional parameter adjustment method is easily affected by human factors, and can target rare abnormal sodium iodide spectrometer data, so that the results have higher robustness; and an online correction method is used to adjust the model parameters online, and the mismatched samples are collected and trained through the online correction module, which protects the operation of nuclear power through real-time online monitoring, thereby improving the prediction efficiency and accuracy of the long-term online operation of the nuclear power key indicator prediction device. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a structural diagram of a nuclear power key indicator prediction device based on optimal deep learning;

[0050] In the figure, 1-sodium iodide spectrometer database; 2-data acquisition and preprocessing module; 3-one-dimensional CNN-LSTM cascade model module 3; 4-multi-objective adaptive particle swarm optimization algorithm module; 5-nuclear power key indicator prediction and online correction module. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical scheme and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0052] The technical concept of the present invention is as follows: the present invention is based on an online correction nuclear power key indicator prediction device, that is, it plays the role of protecting nuclear power operation through real-time online monitoring. The nuclear power key indicator prediction device based on optimal deep learning uses the original sodium iodide spectrometer data as input, mines more subtle features that can distinguish different differences in multiple scales, and uses the optimal deep learning algorithm to establish a nuclear power key indicator prediction model with higher prediction accuracy; the multi-objective adaptive particle swarm algorithm used in this patent can target rare abnormal sodium iodide spectrometer data, so that the results have higher robustness. The nuclear power key indicator prediction and online correction module collects and trains mismatched samples, and has the advantages of adaptability, generalization and agility, etc., which solves the shortcomings of the traditional nuclear power key indicator prediction device with low prediction accuracy, and only targets the known sodium iodide spectrometer data anomaly types but cannot predict the new sodium iodide spectrometer data anomaly types that appear during the operation of nuclear power facilities. A nuclear power key indicator prediction device with strong prediction ability, high accuracy, strong robustness, generalization and agility is invented. The beneficial effects of the present invention are mainly manifested in.

[0053] Example 1

[0054] like Figure 1 As shown, a nuclear power key indicator prediction device based on optimal deep learning includes a sodium iodide spectrometer database 1, a data acquisition and preprocessing module 2, a one-dimensional CNN-LSTM cascade model module 3, a multi-objective adaptive particle swarm optimization algorithm module 4 and a nuclear power key indicator prediction and online correction module 5.

[0055] The data acquisition and preprocessing module 2 is used to preprocess the data x of the sodium iodide spectrometer input from the sodium iodide spectrometer database 1, obtain the preprocessed data x′ and upload it to the one-dimensional CNN-LSTM cascade model module 3.

[0056] The method of preprocessing the data x input from the sodium iodide spectrometer database to obtain preprocessed data x′ specifically includes the following sub-steps:

[0057] (a.1) Clean the data x, including deleting duplicate information and abnormal information, and obtain the cleaned data

[0058] (a.2) Then the cleaned data Perform data type conversion and clean the data The character features in are converted into numerical features to obtain data

[0059] (a.3) Then the data Normalize the data to obtain the preprocessed data x′. The calculation formula is as follows:

[0060]

[0061] in, For data The eigenvalue at the i-th position in ; For data The smallest eigenvalue in ; For data The largest eigenvalue in i ′ is the eigenvalue at the i-th position in the preprocessed data x′.

[0062] The one-dimensional CNN-LSTM cascade model module 3 is used to classify the preprocessed data x′ through the one-dimensional CNN-LSTM cascade model to obtain a classification vector y.

[0063] The one-dimensional CNN-LSTM cascade model module 3 includes a one-dimensional CNN-LSTM cascade model, which mines more distinguishable subtle features in multiple scales and has higher prediction accuracy. The specific structure of the one-dimensional CNN-LSTM cascade model is:

[0064] The first layer is a one-dimensional convolution layer: the convolution kernel size is 16×1, the stride is 2, and the number of convolution kernels is 64;

[0065] The second layer, the maximum pooling layer: the pooling window size is 2;

[0066] The third layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64; the second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64;

[0067] The fourth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 2, and the number of convolution kernels is 64; the second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64;

[0068] The fifth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128. The second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128.

[0069] The sixth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 2, and the number of convolution kernels is 128. The second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128.

[0070] The seventh layer, the fully connected layer, has 256 neurons, followed by a batch standard layer and a nonlinear transformation layer;

[0071] The eighth layer, LSTM layer: the number of units is 10, and the activation function uses the tanh function;

[0072] The ninth layer, LSTM layer: the number of units is 10, and the activation function uses the tanh function;

[0073] The tenth layer, the fully connected layer, has 128 neurons, followed by a batch standard layer and a nonlinear transformation layer;

[0074] Eleventh layer, classification layer: softmax classifier, classifies the results;

[0075] Each one-dimensional convolutional layer is followed by a nonlinear transformation layer and a batch normalization layer.

[0076] The core algorithm of LSTM controls the state of each unit through three control gates, thereby controlling the long and short-term memory of the unit. The specific meanings are as follows:

[0077] 2.2.1) Forget gate. The forget gate calculates and generates the memory weight fgate t Control the state C of the previous moment t-1 How many C are saved to the current moment? t The unit state is calculated as follows:

[0078] fgate t =σ(W F [h t-1 ,p t ]+bf);

[0079] Among them, W F is the weight matrix of the forget gate, h t-1 is the hidden state of the previous moment, p t is the network input value at the current time t, bf is the forget gate bias, and σ is the Sigmoid function.

[0080] 2.2.2) Input gate. The input gate is responsible for generating the input weight i t And the current input unit status And control the input unit status How many inputs are there to the current cell state C? t , input weight igatet and input unit status The calculation formula is as follows:

[0081] igate t =σ(W I [h t-1 ,p t ]+bi);

[0082]

[0083] Among them, W I is the input gate weight matrix, bi is the input gate bias; W C is the input state weight matrix, bc is the input state bias, and tanh is the hyperbolic tangent function.

[0084] 2.2.3) Generate the current cell state C t The current state is determined by the forget gate fgate t , the unit state at the previous moment C t-1 , input gate igate t and the current input unit status Jointly determined, the calculation formula is:

[0085]

[0086] 2.2.4) Output gate. The output gate is responsible for generating the output weight ogate t Control the current cell state C t How many are the hidden layer outputs h at the current moment? t The calculation formula is:

[0087] ogate t =σ(W O [h t-1 ,p t ]+bo);

[0088] h t-1 =ogate t ×tanh(C t );

[0089] Among them, W O is the output gate weight matrix, h t-1 is the hidden layer state at the previous moment, p t is the network input value at the current time t, bo is the output state bias, and tanh is the hyperbolic tangent function.

[0090] The nonlinear transformation layer performs nonlinear transformation on the features through the activation function, and the calculation formula is as follows:

[0091]

[0092] Among them, z is the input feature.

[0093] The batch normalization layer makes the distribution more consistent with the actual distribution of the data, ensuring the nonlinear expression ability of the model.

[0094] The one-dimensional residual layer: The one-dimensional residual layer is composed of two one-dimensional convolutional layers, each of which is followed by a nonlinear transformation layer and a batch normalization layer. There is a residual connection at both ends of the two connected one-dimensional convolutional layers. Note that when the input and output feature dimensions of the two convolutional layers as a whole change, the residual connection also needs to make the same dimensional change to the features, so that the output of the residual connection is added to the output of the two one-dimensional convolutional layers as a whole.

[0095] The multi-objective adaptive particle swarm optimization algorithm module 4 is used to optimize the hyperparameters such as the number of training times, learning rate, and LSTM time window size of the one-dimensional CNN-LSTM cascade model using an improved particle swarm optimization algorithm, which partially overcomes the disadvantage that the traditional parameter adjustment method is easily affected by human factors, and can target rare abnormal sodium iodide spectrometer data to make the results more robust, specifically:

[0096] (a.1) Initialize the particle swarm optimization algorithm parameters: population size P, maximum number of iterations t max ;

[0097] (a.2) Randomly initialize the particle swarm position r i and speed v i ;

[0098] (a.3) Calculate the fitness value F and calculate the historical optimal Lbest;

[0099] (a.4) According to the Pareto dominance principle, the current non-inferior solution is stored in the Archive set;

[0100] (a.5) Use a grid to divide the target space into small areas, and use the number of particles in each area as the density information of the particles. The more particles a particle contains in the grid, the greater its density value, and vice versa. This is used to calculate the crowdedness of the Archive set;

[0101] (a.6) Select the particle with the lowest density in the Archive set as the global optimal Gbest;

[0102] (a.7) Calculate the evolution factor f, which is the length of the sum of the connection vectors between the optimal particle and other particles divided by the sum of the distances between the optimal particle and other particles;

[0103] (a.8) Update the particle's velocity and position:

[0104]

[0105] ω is the inertia factor, c 1 、c 2 is the acceleration coefficient, generally 2; rand 1 、rand 2 is a random number between [0,1]; is the updated particle velocity, is the particle velocity before updating; r i t+1 is the updated particle position, r i t is the particle position before updating.

[0106] (a.9) Repeat steps (a.3) to (a.8) until the iteration stopping condition is met.

[0107] The nuclear power key indicator prediction and online correction module 5 is used to update the one-dimensional CNN-LSTM cascade model online, regularly input the current data stream of the sodium iodide spectrometer into the training set, use the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model, use the online correction method to adjust the model parameters online, collect and train the mismatched samples through the online correction module, and protect the operation of nuclear power through real-time online monitoring, thereby improving the prediction efficiency and accuracy of the long-term online operation of the nuclear power key indicator prediction device. The one-dimensional CNN-LSTM cascade model can adapt to its changes and maintain good prediction performance, specifically:

[0108] The nuclear power key indicator prediction and online correction module 5 performs online update on the one-dimensional CNN-LSTM cascade model, regularly inputs the current data stream of the sodium iodide spectrometer into the training set, and uses the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model, specifically:

[0109] The nuclear power key indicator prediction and online correction module 5 adopts an online correction strategy to perform real-time correction on the one-dimensional CNN-LSTM cascade model, by inputting abnormal data as new training data into the training set, and using the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model in real time;

[0110] When the one-dimensional CNN-LSTM cascade model is working, the prediction analysis value MI at time t is obtained in real time a (t) and the predicted value MI p (t), calculate the analytical value MI a and the predicted value MI p The deviation e o (t):

[0111] e o (t)=|MI a (t)-MI p (t)|;

[0112] If e o (t) is greater than a positive error tolerance ε o :

[0113] e o (t)>ε o ;

[0114] Then the predicted analysis value MI at time t a (t) and parameter value x(t) as new training data {x(t), MI a (t)} is added to the training data set, and the prediction model parameters are retrained to perform online correction of the prediction model; otherwise, the model parameters are considered to be accurate and no online correction is required.

[0115] The proposed multi-objective adaptive particle swarm optimization module 4 is used to obtain the optimal weight of the influence of the new sample data x(t) on the training model results in the training set, including the number of training times, learning rate, and LSTM time window size hyperparameters of the CNN-LSTM prediction module, so as to quickly correct the mismatched nuclear power key indicator prediction model and obtain better nuclear power key indicator prediction effect.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A nuclear power key indicator prediction device based on optimal deep learning, characterized in that: The device includes a sodium iodide spectrometer database, a data acquisition and preprocessing module, a one-dimensional CNN-LSTM cascade model module, a multi-objective adaptive particle swarm optimization algorithm module, and a nuclear power key indicator prediction and online correction module; The data acquisition and preprocessing module is used to perform data preprocessing on the data x of the sodium iodide spectrometer input from the sodium iodide spectrometer database, obtain the preprocessed data x′ and upload it to the one-dimensional CNN-LSTM cascade model module; The one-dimensional CNN-LSTM cascade model module is used to classify the preprocessed data x′ through the one-dimensional CNN-LSTM cascade model to obtain a classification vector y; The multi-objective adaptive particle swarm optimization algorithm module is used to optimize the training times, learning rate, and LSTM time window size of the one-dimensional CNN-LSTM cascade model using an improved particle swarm optimization algorithm; The nuclear power key indicator prediction and online correction module is used to update the one-dimensional CNN-LSTM cascade model online, regularly input the current data stream of the sodium iodide spectrometer into the training set, and use the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model.

2. The device for predicting key nuclear power indicators based on optimal deep learning according to claim 1, characterized in that: The method of preprocessing the data x input from the sodium iodide spectrometer database to obtain preprocessed data x′ specifically includes the following sub-steps: (a.1) Clean the data x, including deleting duplicate information and abnormal information, and obtain the cleaned data (a.2) Then the cleaned data Perform data type conversion and clean the data The character features in are converted into numerical features to obtain data (a.3) Then the data Normalize the data to obtain the preprocessed data x′. The calculation formula is as follows: in, For data The eigenvalue at the i-th position in ; For data The smallest eigenvalue in ; For data The largest eigenvalue in i ′ is the eigenvalue at the i-th position in the preprocessed data x′.

3. The device for predicting key nuclear power indicators based on optimal deep learning according to claim 1, characterized in that: The one-dimensional CNN-LSTM cascade model module includes a one-dimensional CNN-LSTM cascade model, and the specific structure of the one-dimensional CNN-LSTM cascade model is: The first layer is a one-dimensional convolution layer: the convolution kernel size is 16×1, the step size is 2, and the number of convolution kernels is 64; The second layer, the maximum pooling layer: the pooling window size is 2; The third layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64; the second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64; The fourth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 2, and the number of convolution kernels is 64; the second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 64; The fifth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128. The second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128. The sixth layer is a one-dimensional residual layer: the convolution kernel of the first one-dimensional convolution layer is 16×1, the stride is 2, and the number of convolution kernels is 128. The second one-dimensional convolution layer is 16×1, the stride is 1, and the number of convolution kernels is 128. The seventh layer, the fully connected layer, has 256 neurons, followed by a batch standard layer and a nonlinear transformation layer; The eighth layer, LSTM layer: the number of units is 10, and the activation function uses the tanh function; The ninth layer, LSTM layer: the number of units is 10, and the activation function uses the tanh function; The tenth layer, the fully connected layer, has 128 neurons, followed by a batch standard layer and a nonlinear transformation layer; Eleventh layer, classification layer: softmax classifier, classifies the results; Each one-dimensional convolutional layer is followed by a nonlinear transformation layer and a batch normalization layer.

4. The device for predicting key nuclear power indicators based on optimal deep learning according to claim 1, characterized in that: The multi-objective adaptive particle swarm optimization algorithm module uses an improved particle swarm optimization algorithm to optimize the training times, learning rate, and LSTM time window size of the one-dimensional CNN-LSTM cascade model, specifically: (a.1) Initialize the particle swarm optimization algorithm parameters: population size P, maximum number of iterations t max ; (a.2) Randomly initialize the particle swarm position r i and speed v i ; (a.3) Calculate the fitness value F and calculate the historical optimal Lbest; (a.4) According to the Pareto dominance principle, the current non-inferior solution is stored in the Archive set; (a.5) Divide the target space into small areas with grids, and use the number of particles contained in each area as the density information of the particles; the more particles a particle contains in the grid, the greater its density value, and vice versa, the smaller it is, so as to calculate the crowdedness of the Archive set; (a.6) Select the particle with the lowest density in Archive as the global optimal Gbest; (a.7) Calculate the evolution factor f, which is the length of the sum of the connection vectors between the optimal particle and other particles divided by the sum of the distances between the optimal particle and other particles; (a.8) Update the particle's velocity and position: ω is the inertia factor, c1 and c2 are acceleration coefficients, usually 2; rand1 and rand2 are random numbers between [0,1]; is the updated particle velocity, is the particle velocity before updating; r i t+1 is the updated particle position, r i t is the particle position before updating; (a.9) Repeat steps (a.3) to (a.8) until the iteration stopping condition is met.

5. The device for predicting key nuclear power indicators based on optimal deep learning according to claim 1, characterized in that: The nuclear power key indicator prediction and online correction module updates the one-dimensional CNN-LSTM cascade model online, regularly inputs the current data stream of the sodium iodide spectrometer into the training set, and uses the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model, specifically: The nuclear power key indicator prediction and online correction module adopts an online correction strategy to perform real-time correction on the one-dimensional CNN-LSTM cascade model, by inputting abnormal data as new training data into the training set, and using the new training set to modify the parameters in the one-dimensional CNN-LSTM cascade model in real time; When the one-dimensional CNN-LSTM cascade model works, the prediction analysis value MI at time t is obtained in real time a (t) and the predicted value MI p (t), calculate the analytical value MI a and the predicted value MI p The deviation e o (t): e o (t)=|MI a (t)-MI p (t)|; If e o (t) is greater than a positive error tolerance ε o : e o (t)>e o ; Then the predicted analysis value MIa(t) and parameter value x(t) at time t are used as new training data {x(t), MI a (t)} is added to the training data set, and the prediction model parameters are retrained to perform online correction of the prediction model; otherwise, the model parameters are considered to be accurate and no online correction is required.

Images