CN110689171A - A method for predicting the state of health of steam turbines based on E-LSTM - Google Patents

Abstract

The invention provides a method for predicting the state of health of a steam turbine based on E-LSTM. Collect steam turbine operation data from sensors and perform preprocessing; feed the preprocessed data into the LSTM network for multiple iterative training; input multiple trained model parameters into the genetic algorithm as the initial population, and run the genetic algorithm. Algorithm, select the optimal model parameters; use more steam turbine operation data to verify the generalization performance of the optimal model; according to the optimal model parameters, predict the test data set and evaluate the model error. The invention can improve the accuracy of model prediction and avoid over-fitting, and can realize multiple linear regression prediction, so that the prediction model has a better fitting effect on real data, can greatly reduce the error of human monitoring, and improve the efficiency of fault diagnosis. Be proactive about the occurrence of failures. It can be widely used in the state management of various thermal and nuclear power plants and even steam turbines of ships.

Description

A method for predicting the state of health of steam turbines based on E-LSTM

technical field

The invention relates to a health state prediction method, in particular to a health state prediction method of a steam turbine generator in a nuclear power and thermal power plant.

Background technique

According to statistics, China's annual thermal and nuclear power generation accounts for nearly 80% of the total power generation, and the steam turbine generator is one of the core equipment in the thermal power and nuclear power generation systems. Ensuring the safe and stable operation of the turbo-generator has always been one of the most important links in the power supply system. However, in the era of Industry 4.0, the traditional sensor + manual monitoring method faces many problems such as high cost and low efficiency, and an intelligent and efficient power supply system state prediction scheme is urgently needed.

From the current research results, it can be seen that the traditional way of understanding the health status of steam turbines by observing sensor data is quite subjective and one-sided, and the interpretation of data depends entirely on human experience. In the past few decades, people have built up a rule-based expert system by accumulating a lot of experience on steam turbine operation. However, the expert system has the following obvious shortcomings: (1) The relationship between the rules is not transparent. The logical relationship between a large number of rules may be opaque and lack hierarchical knowledge representation. (2) Inefficient search strategy. The inference engine searches all the rules in each cycle. When there are many rules, the system will run very slowly, and large-scale rule-based expert systems are not suitable for real-time applications. (3) No learning ability. General rule-based expert systems do not have the ability to learn from experience, and it is difficult to deal with special or emergency situations.

For a steam turbine generator set, if it is regularly maintained, the economic benefits will be low. If it is repaired after a failure, it often misses the opportunity to prevent the further expansion of the failure loss, which is more than the gain. The technology based on the knowledge and experience of domain experts in the past has been unable to meet the requirements of safe and economical operation of the unit. The development of artificial intelligence technology such as neural network and its rapid penetration into the engineering field have brought new vitality to the fault state prediction technology, making the modern diagnosis technology enter a new stage. The implementation of artificial intelligence algorithms does not require users to have very rich prior knowledge, and can directly mine fault features from data, and then perform fault classification and state prediction. The model obtained based on artificial intelligence algorithm has the characteristics of small size and strong transferability, which is suitable for industrial fault diagnosis and has become an important research topic in the field of fault diagnosis technology today.

Summarizing the existing research results, it is found that the current steam turbine health status monitoring system has the following problems that need to be solved:

(1) Human monitoring is costly and inefficient, and human error cannot be avoided.

(2) Manual judgment of faults based on sensor data is subjective, and the judgment results depend on human experience. In addition, it is difficult to fully explore the internal relationship between various parameters manually, so that the fault information cannot be fully interpreted. The expert system is relatively rigid, lacks real-time performance, and does not have the ability to learn the newly discovered fault characteristics, so it is difficult to cope with the complex and changeable production environment.

(3) The existing fault monitoring method is "later aware" of the impending occurrence of the fault, and there is not enough time to respond when the fault is discovered. However, over-maintenance and early replacement to avoid failures have low economic benefits.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an E-LSTM-based steam turbine health state prediction method with high prediction accuracy, small error and high diagnostic efficiency.

The object of the present invention is achieved in this way:

Step 1. Collect steam turbine operation data from sensors and preprocess;

Step 2: Feed the preprocessed data into the LSTM network for multiple iterative training;

Step 3: Input the trained multiple model parameters into the genetic algorithm as the initial population, run the genetic algorithm, and select the optimal model parameters;

Step 4. Use more steam turbine operation data to verify the generalization performance of the optimal model;

Step 5: Predict the test data set according to the optimal model parameters, and evaluate the model error.

The present invention can also include:

1. Preprocess the sampled data, denote the sequence as Y after normalization, Y=|y ₀ , y ₁ , y ₂ ,..., y _r , y _r-1 |, and use Y as the training data input to initialize The LSTM network completes parameter learning, uses the actual value as the input for the next step in each step of prediction at time t ∈ (0, T) in the training phase, and updates the neuron state, looping through the remaining predictions,

Let h _t = y _t , the steps of the prediction method are as follows:

In the formula, h is the value of the output gate of the previous layer, y is the input value of the current node, f is the weight output by the forgetting gate through the sigmoid activation function, and C is the output value of the forgetting gate and the input gate after confirming the update and forgetting.

2. The training is divided into the following three training methods:

①For the initial training, all neural networks are trained, and the network is trained by setting the variable learning rate to optimize the cross loss function through the Adam gradient descent algorithm and the SGD gradient descent algorithm;

②When a new category needs to be added as training data, under the premise of ①training results, set a small learning rate for the main structure of the LSTM network for learning, then freeze all neural network layers except the fully connected layer, retrain and finally the fully connected layer;

③ When it is necessary to control new measurement points, use the training result of ① as a pre-training model, activate all neural networks, and set a variable learning rate to optimize the cross loss function to train the network.

3. The method of selecting the model parameters with the best effect is: train multiple models according to the data sampled at different time intervals, use the parameters of these models as the initial population, carry out iterative optimization of the genetic algorithm, and select the parameter sequence of the optimal offspring is the optimal model.

4. The evaluation model error is specifically: substituting the optimal model parameters into the LSTM, inputting the test set, and calculating the error between the predicted value and the actual value; the error calculation has the following two methods:

Mean Squared Error:

Root Mean Square Error:

where N is the number of datasets, N is the number of datasets, Y _i is the real dataset, Y _i ^* is the predicted dataset,

According to the error calculation results, check whether the accuracy of the model meets the requirements, if not, continue training and seek optimization.

Aiming at the health state monitoring problem of steam turbines, the present invention proposes a method for predicting the state of health of steam turbines based on long short-term memory neural network (LSTM). Prediction method of steam turbine health. Multiple models are obtained by training the neural network multiple times, and genetic algorithm optimization is performed on the model parameters obtained by multiple training, and the model with excellent prediction effect and the best generalization ability is selected to improve the accuracy of model prediction and avoid overfitting. combine. Through the LSTM neural network, the various parameters of the steam turbine system (pressure, vibration, temperature, speed, etc.) are used to fully explore the internal relationship between the parameters and realize multiple linear regression prediction. Then, by using the genetic algorithm, the parameters of multiple trained LSTM models are selected, so that the prediction model has a better fitting effect on the real data. Using the optimal model to predict the health status of the turbo-generator can greatly reduce the error of human monitoring, improve the efficiency of fault diagnosis, and achieve "prescientific awareness" of the occurrence of faults. It can be widely used in the state management of various thermal and nuclear power plants and even steam turbines of ships.

In order to overcome the defects in the prior art, the present invention proposes a steam turbine health state prediction model E-LSTM on the basis of previous research, that is, by using a long short-term memory neural network (LSTM) training model combined with an evolutionary algorithm ( Evolutionary algorithms) for model selection to improve prediction accuracy and avoid overfitting.

Description of drawings

Fig. 1 Functional structure diagram of steam turbine state prediction system.

Figure 2. Model training and selection flow chart.

Figure 3 is a flow chart of the implementation of steam turbine health state prediction.

Figure 4 E-LSTM structure diagram.

Figure 5. The error between the predicted value and the actual value of the optimal model.

Detailed ways

The present invention will be described in more detail with examples below.

The present invention is a method for predicting the health degree of a steam turbine based on E-LSTM, the structure of which is shown in Figure 1, including collecting the operation data of the steam turbine to obtain the distribution characteristics of its health status;

In order to overcome the defects in the prior art, the present invention proposes a steam turbine health state prediction model E-LSTM on the basis of previous research, that is, by using a long short-term memory neural network (LSTM) training model combined with an evolutionary algorithm ( Evolutionary algorithms) to carry out model selection, in order to achieve the purpose of improving prediction accuracy and reducing overfitting, the present invention adopts the following steps to realize the state prediction of steam turbine:

Step 01. Collect turbine operation data from sensors and preprocess the data.

Step 02. Feed the processed data into the LSTM network for multiple iterative training.

Step 03. Input the trained multiple models into the genetic algorithm as the initial population, run the genetic algorithm, and select the model with the best effect.

Step 04. Use more steam turbine operation data to verify the generalization performance of the optimal model.

Step 05. According to the optimal model, make predictions on the test data set and evaluate the model error.

The step 01 is specifically as follows:

Step 0101. Arrange sensors at each monitoring point of the steam turbine, perform verification and sensor data fusion on multiple types of sensor data, and obtain effective and reliable data that truly reflects the operation status of the steam turbine.

Step 0102. Sampling the data, and the sampling time interval is 5 minutes, 10 minutes, 15 minutes, 30 minutes, and 60 minutes.

Step 0103. Take 70% of the sampled data as the training set and 30% as the test set according to each time interval.

The step 02 is specifically:

According to the data preprocessing described above, the sequence is normalized and expressed as Y, Y=|y ₀ , y ₁ , y ₂ , ..., y _r , y _r-1 |. Input Y as training data to the initialized LSTM network to complete parameter learning. In order to predict the value of L time steps, the traditional LSTM network makes each prediction based on the predicted value of the previous step. This is improved to use the actual value as the input for the next step in each prediction step at time t ∈ (0, T) in the training phase, and update the neuron state, reducing the gradient propagation of the error, looping over the remaining predictions.

Let h _t = y _t , the steps of the improved prediction method are as follows:

In the formula, h is the value of the output gate of the previous layer, y is the input value of the current node, f is the weight output by the forgetting gate through the sigmoid activation function, and C is the output value of the forgetting gate and the input gate after confirming the update and forgetting.

Combined with the actual problems of steam turbine fault diagnosis, it can be divided into the following three training methods:

①For the initial training, all neural networks are trained, and the network is trained by setting the variable learning rate to optimize the cross loss function through the Adam gradient descent algorithm and the SGD gradient descent algorithm;

②When a new category needs to be added as training data, under the premise of ①training results, set a small learning rate for the main structure of the LSTM network for learning, then freeze all neural network layers except the fully connected layer, retrain and finally the fully connected layer;

③ When it is necessary to control new measurement points, use the training result of ① as a pre-training model to activate all neural networks, and set a variable learning rate to optimize the cross loss function to train the network;

The step 03 is specifically:

The parameters to be optimized for the LSTM neural network prediction model include: the number of hidden layers of the LSTM neural network, the time window step size, the number of training times, and the forgetting rate Dropout. The genetic algorithm optimizes the model of the LSTM neural network in the parameter search space, with the minimum prediction error and the strongest generalization ability as the objective function, to optimize the parameter combination to form a composite E-LSTM, including the following steps:

Step 0301: Step S21, population initialization and decoding;

Step 0302, using the mean square error of the LSTM neural network as a fitness function;

Step 0303, performing a selection crossover mutation operation on the individual of the solution;

Step 0304, if the target value of the fitness function reaches the optimal value, proceed to the next step; otherwise, return to step 0303;

Step 0305, obtain the fitness function target value and optimal parameter;

Step 0306, calculate the predicted mean square error based on the best parameter;

Step 0307, judging the termination condition, if the number of iterations of the population is satisfied, stop the calculation, and at this time, the global optimal parameter combination of the LSTM network; otherwise, return to step 0306;

Further, the step 04 is specifically:

Take the data sampled at different time intervals in step 01 and input it into the optimal model in step 03 to obtain the error between the predicted value of the model and the actual value. If the error is greater than the threshold allowed by the system, go to step 02. The error is calculated as follows:

Mean Squared Error:

In the formula, N is the number of data sets, which is the real data set and the predicted data set.

Finally, the step 05 is specifically:

Use the optimal model to predict the health of the steam turbine on the predicted data set, and calculate the error between the predicted data and the actual data. The error calculation uses two indicators of mean square error and root mean square error to restore the predicted data for output. In the prediction, The smaller the value of mean square error and root mean square error, the higher the prediction accuracy, where:

Mean Squared Error:

Root Mean Square Error:

where N is the number of datasets, Y _i is the real dataset, and Y _i ^* is the predicted dataset.

Claims (5)

1. a steam turbine health state prediction method based on E-LSTM is characterized in that: Step 1. Collect steam turbine operation data from sensors and preprocess; Step 2: Feed the preprocessed data into the LSTM network for multiple iterative training; Step 3: Input the trained multiple model parameters into the genetic algorithm as the initial population, run the genetic algorithm, and select the optimal model parameters; Step 4. Use more steam turbine operation data to verify the generalization performance of the optimal model; Step 5: Predict the test data set according to the optimal model parameters, and evaluate the model error. 2. The method for predicting the state of health of a steam turbine based on E-LSTM according to claim 1, wherein the preprocessing operation is performed on the sampled data, and the sequence is standardized and expressed as Y, Y=|y ₀ , y ₁ ,y ₂ ,…,y _r ,y _r-1 |, input Y as training data to the initialized LSTM network, complete parameter learning, and use the actual value as the next input, and update the neuron state, looping over the remaining predictions, Let h _t = y _t , the steps of the prediction method are as follows: Input: Y={y _o ,y ₁ ,...,y _T-1 ,y _T }, Output predicted value={y′ _T+1 ,y′ _T+2 ,…,y′ _T+L }, for t=0, t≤T, t++: C _t ←input _gate ←C _t-1 ,h _t-1 ,y _t ,f _t f _t ←forget _gate ←h _t-1 ,y _t y′ _t ←output _gate ←h _t-1 ,y _t Loss(y _t ,y′ _t ) end; for l=1, l＜=L, l++: C _T+1 ←input _gate ←C _T+L-1 ,h _T+1 ,y′ _T+l ,f _T+ f _T+l ←forget _gate ←h _T+L-1 ,y′ _T+l y′ _T+l ←output _gate ←C _T+l ,h _T+l-1 ,y′ _T+l end; In the formula, h is the value of the output gate of the previous layer, y is the input value of the current node, f is the weight output by the forgetting gate through the sigmoid activation function, and C is the output value of the forgetting gate and the input gate after confirming the update and forgetting. 3. the steam turbine health state prediction method based on E-LSTM according to claim 2, is characterized in that described training is divided into following 3 kinds of training methods: ①For the initial training, all neural networks are trained, and the network is trained by setting the variable learning rate to optimize the cross loss function through the Adam gradient descent algorithm and the SGD gradient descent algorithm; ②When a new category needs to be added as training data, under the premise of ①training results, set a small learning rate for the main structure of the LSTM network for learning, then freeze all neural network layers except the fully connected layer, retrain and finally the fully connected layer; ③ When it is necessary to control new measurement points, use the training result of ① as a pre-training model, activate all neural networks, and set a variable learning rate to optimize the cross loss function to train the network. 4. the steam turbine state-of-health prediction method based on E-LSTM according to claim 3 is characterized in that the method for selecting the optimal model parameter of effect is: a plurality of models are trained according to the data sampled at different time intervals, and these models are The parameters are used as the initial population, the genetic algorithm is iteratively optimized, and the parameter sequence of the optimal offspring is selected as the optimal model. 5. the steam turbine state-of-health prediction method based on E-LSTM according to claim 4, is characterized in that: described evaluation model error is specifically: substitute optimal model parameter in LSTM, input test set, calculate predicted value and The error between the true values; there are two ways to calculate the error: Mean Squared Error:

Root Mean Square Error: where N is the number of datasets, N is the number of datasets, Y _i is the real dataset, Y _i * is the predicted dataset, According to the error calculation results, check whether the accuracy of the model meets the requirements, if not, continue training and seek optimization.

Images