CN117349620A

CN117349620A - Application method and system for predicting safety state of nuclear power plant equipment

Info

Publication number: CN117349620A
Application number: CN202311352276.1A
Authority: CN
Inventors: 秦绪涛; 张钧鸣; 宋忠洋; 许列琦; 朱旭东; 王晨成; 杨强; 朱云飞; 樊金龙; 刘铭洋
Original assignee: Jiangsu Nuclear Power Corp
Current assignee: Jiangsu Nuclear Power Corp
Priority date: 2023-10-18
Filing date: 2023-10-18
Publication date: 2024-01-05

Abstract

The invention provides an application method for predicting the safety state of nuclear power plant equipment, which comprises the following steps: and (5) preprocessing data. Sampling the data to obtain the power of a transmitter; aligning the sampling frequency, and supplementing the missing data forward by using the later data; respectively calculating the power difference of each transmitter in the same pump; calculating the maximum value, the minimum value, the average value and the sum of the daily power data of each transmitter in the data set to obtain the power statistical data of the current day; and (3) calculating the maximum value, the minimum value, the intermediate value and the sum of the diff of each transmitter every day in the data set to obtain the difference statistical data of the current day. The invention also provides a system for predicting the safety state of the nuclear power plant equipment, which comprises a data preprocessing module, an equipment safety state prediction model training module and an equipment safety state prediction module. The invention greatly reduces the difference of the prediction amplitude, and early warning is carried out on the early degradation of the equipment performance in real time, thereby reducing the analysis time of business personnel.

Description

Application method and system for predicting safety state of nuclear power plant equipment

Technical Field

The invention relates to the technical field of prediction of the safety state of nuclear power plant equipment, in particular to an application method and system for prediction of the safety state of the nuclear power plant equipment.

Background

The power system has more unstable factors, wherein the power equipment is used as a basic element of the power system, the fault is an important potential safety hazard affecting the operation of the power grid, and the loss caused by the fault risk of the power equipment can be effectively reduced by predicting the fault risk of the power equipment. In order to ensure the development of safe operation and maintenance economy and society of the power network, the improvement of the power equipment fault diagnosis strategy is one of the key problems to be solved in the current urgent need. In the detection of the power system, the existing equipment is utilized to monitor the system operation data and the sensor data in real time, a fault prediction and health state evaluation model based on multi-source heterogeneous data can be established for the equipment, early degradation early warning of the equipment performance is realized, the operation and maintenance labor cost is reduced, and the equipment operation safety is improved.

Currently, physical model and data-driven based research methods for fault prediction are used. The physical model comprises a physical failure model and a filter model, and the data driving can be divided into reliability analysis, a statistical method and an intelligent neural network. Conventional failure analysis predictions have been popular in recent years using machine learning models, such as Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs), for factor-dominant approaches, and have shown some competitiveness in the knowledge learning task. Machine learning is more flexible for high-dimensional training data and is able to capture more feature information from historical power plant operational data sets than simple statistical models. This is advantageous for obtaining better power plant operating conditions and stability analysis results.

Disclosure of Invention

The invention aims to provide an application method and system for predicting the safety state of nuclear power plant equipment, which solve the problems that the equipment has large power change under different environments and can not judge the new operation condition of the equipment and adapt to the environment change, thereby improving the stability of the equipment and improving the efficiency and performance of the equipment.

In order to achieve the above object, the present invention provides the following technical solutions:

an application method for predicting the safety state of nuclear power plant equipment, comprising the following steps:

step 1: preprocessing data to obtain a device safety state prediction data set;

step 2: training a device safety state prediction model on the device safety state prediction data set;

step 3: predicting the device security status.

The step 1 comprises the following steps:

step 1.1: sampling the data to obtain the power of a transmitter;

step 1.2: aligning the sampling frequency, and supplementing the missing data forward by using the later data;

step 1.3: respectively calculating the power difference of each transmitter in the same pump;

step 1.4: calculating the maximum value, the minimum value, the average value and the sum of the daily power data of each transmitter in the data set to obtain the power statistical data of the current day;

step 1.5: and calculating the maximum value, the minimum value, the intermediate value and the sum of the power difference values of each transmitter in the same pump of each transmitter in the data set to obtain the difference statistical data of the current day.

In step 1.1, the sampling frequency is aligned to within 10 seconds.

In step 1.1, the sampling frequency is aligned to 1 second.

The step 2 comprises the following steps:

step 2.1: digitizing the device status in the dataset the next day of the day;

step 2.2: the power statistics data, the power difference statistics data, the pump and the transmitter are taken as characteristics, the equipment state on the next day is taken as a label, and the equipment state is input into a model for training, so as to obtain and store an equipment safety state prediction model.

In step 2.1, the data is modeled using the LGBMClassifier tool in the sklearn toolkit.

Step 2.2 comprises:

step 2.2.1: reading a device security state prediction data set and adding a dataset to the device security state prediction data set;

step 2.2.2: dividing a device safety state prediction data set by using a train_test_split function, wherein the device safety state prediction training data set is defined as train_data set, and the device safety state prediction verification data set is defined as val_data set;

step 2.2.3: performing model training of the LGBMClassifier model by using the equipment safety state prediction training data set, and storing the equipment safety state prediction model; verifying the equipment safety state prediction model by using the equipment safety state prediction verification data set, so as to select an optimal equipment safety state prediction model; the device safety state prediction model is trained and verified by using default parameters of the LGBMClassiier model to obtain a reference device safety state prediction model.

Step 2.2.3 comprises:

step 2.2.3.1: adjusting and optimizing the parameters of max_depth, num_leave, min_data_in_leaf, min_split_gain, subsamples and colsample_byte in the classification model LGBMClassier;

step 2.2.3.2: the method comprises the steps of adjusting and optimizing min_split_gain, subsamples and colsample_byte parameters in a classification model LGBMClassiier;

step 2.2.3.3: parameter adjustment optimization is carried out on the lambda_l1 and lambda_l2 parameters in the classification model LGBMClassiier;

step 2.2.3.4: parameter optimization is carried out on the classification model LGBMClassiier, and the learning_rate parameter is adjusted.

The step 3 comprises the following steps:

step 3.1: acquiring real-time power data of each transmitter of the equipment on the same day from an equipment sensor monitoring system;

step 3.2: performing data processing on the power data of each transmitter of the equipment in the mode of the step 1 to obtain the power data of the equipment;

step 3.3: predicting different transmitters and predicting equipment pumps by using the equipment safety state prediction model in the step 2;

step 3.4: and feeding the prediction result information back to the equipment state page for display.

Furthermore, the present invention provides a system for predicting a safety state of a nuclear power plant apparatus, comprising:

the data preprocessing module is used for extracting characteristics of monitoring data of different equipment sensors;

the equipment safety state prediction model training module is used for training the equipment safety state prediction model for the equipment data set by using the classification model;

and the equipment safety state prediction module is used for predicting the equipment safety state of the real-time data of the sensors in different areas of the equipment and giving out whether the equipment is normal or not.

Compared with the prior art, the application method and the system for predicting the safety state of the nuclear power plant equipment have the following beneficial effects:

according to the invention, through the provided historical equipment time sequence data and through extracting the characteristic values, the design of a variable rule is searched, a power equipment fault prediction model is realized, and equipment fault prediction is realized. Based on the modeling results, the potentially malfunctioning device is predicted and information is pushed to the engineer or operator responsible for the device.

According to the invention, the device fault prediction model is constructed by extracting and analyzing the extracted characteristics of the data of each time period in the device sensor monitoring system and adopting a machine learning algorithm, and whether the tomorrow device is abnormal or not is predicted based on the characteristics of the same day.

The method uses multi-source heterogeneous data to carry out modeling, so that the difference of the prediction amplitude is greatly reduced; early warning is carried out on early degradation of the equipment performance in real time; dynamic prediction equipment trend state conditions are adopted, so that analysis time of business personnel is shortened; in the same pump, monitoring and judging the power of each transmitter device; and comparing with the historical conditions, predicting and analyzing the abnormal conditions of the equipment, and giving out similar equipment fault conditions.

Drawings

In order to more clearly illustrate the technical solution of the present invention, the drawings that are used in the technical description will be briefly described below.

FIG. 1 is a flow chart of the fault prediction model construction provided by the present invention.

Detailed Description

Further details are provided below with reference to the specific embodiments.

As shown in fig. 1, the present invention provides an application method for predicting a safety state of a nuclear power plant device, including:

step 1: and (5) preprocessing data.

Step 1.1: sampling the data to obtain the power of a transmitter;

step 1.2: aligning the sampling frequency to 1 second, and supplementing the missing data with the following data;

step 1.5: and (3) calculating the maximum value, the minimum value, the intermediate value and the sum of the diff of each transmitter every day in the data set to obtain the difference statistical data of the current day.

Illustratively, the power data of each transmitter of the device pump is sampled to obtain the power of the transmitter. In this embodiment four pumps are included, each pump containing 3 transmitters.

The sampling frequency is aligned to within 10 seconds, where missing date data is forward-padded with data from the date closest to the missing date. For example: assuming that the data on days 7-02 and 10 are missing in this embodiment, the missing data is complemented with the data on day 7-01, and the missing data is complemented with the data on day 7-01, so that the missing data is complemented with the data on day 7-01, which will cause partial data duplication, but will not cause interference to the equipment safety state prediction due to the duplicated data.

In view of the multiple transmitter condition of the same pump, the difference of the power of each transmitter in the same pump needs to be calculated. For example, assuming that in the pump No. 1 of 7 months 01, the power of the transmitters No. 1, 2, and 3 is 5338.17, 5186.52, and 5340.64, respectively, the difference diff_1 and diff_2 between the transmitters No. 1 and 2 is 151.65.

And calculating the maximum value, the minimum value, the average value, the sum and the intermediate value of the daily power data of each transmitter in the data set to obtain the power statistical data of the current day. For example, assuming that the power of the transmitters is 5338.17, 5186.52, and 5340.64 in pump 1 of 7 months 01 day 1, the maximum power transmitter_1_max, the minimum power transmitter_1_min, the average power transmitter_1_mean, and the sum power transmitter_1_sum are 5340.64, 5186.52, 5288.44, 15865.33, and 5338.17, respectively, of the transmitter 1 of pump 1.

The maximum, minimum, median, sum, average of the differences in power of each transmitter in the same pump of each transmitter in the dataset are calculated. For example, assume that in the 7 month 01 day 1 pump of this embodiment, the difference diff_1 between the No. 1 transmitter and the No. 2 transmitter is 151.65, 154.12,2.47. The difference between the transmitter No. 1 and the transmitter No. 2 in pump No. 1 has a maximum power diff_1_max of 154.12, a minimum power diff_1_min of 2.47, an average power diff_1_mean of 102.74, a total power diff_1_sum of 308.24, and an intermediate power diff_1_mean of 151.65.

In order to better distinguish equipment fault conditions, the equipment state in the data set on the next day of the day is subjected to numerical treatment, the occurrence of faults is recorded as 1, and the non-occurrence of faults is recorded as 0.

And taking the power statistics data and the power difference statistics data of each transmitter every day, the pump, the transmitter and the identification state of the last day as characteristics, and taking the equipment state of the next day as a label to obtain an equipment safety state prediction data set.

Step 2: and training a device safety state prediction model.

Step 2.1: the device status on the next day of the day in the dataset was digitized. Modeling the data by using an LGBMClassifier tool in a sklearn toolkit;

Illustratively, the device security state prediction model is trained on the device security state prediction dataset using LGBMClassifier function files in the sklearn toolkit.

The LGBMClassifier is a LGBMClassifier function file in sklearn toolkit, is a classification model in machine learning, and is iteratively trained by using a weak classifier (decision tree) to obtain an optimal model for multi-classification prediction. LGBM (Light Gradient Boosting Machine) is a framework for realizing GBDT algorithm, supports high-efficiency parallel training, and has the advantages of faster training speed, lower memory consumption, better accuracy, support of distributed type and capability of rapidly processing mass data.

Training the equipment safety state prediction model, wherein the specific step flow comprises the following steps:

step 2.2.2: dividing the device safety state prediction data set by using a train_test_split function, namely dividing the dataset into a device safety state prediction training data set (80%) and a device safety state prediction verification data set (20%), wherein the device safety state prediction training data set is defined as train_dataset, and the device safety state prediction verification data set is defined as val_dataset;

step 2.2.3: performing model training of the LGBMClassifier model by using the equipment safety state prediction training data set, and storing the equipment safety state prediction model; and verifying the equipment safety state prediction model by using the equipment safety state prediction verification data set, so as to select the optimal equipment safety state prediction model. The device safety state prediction model is trained and verified by using default parameters of the LGBMClassiier model to obtain a reference device safety state prediction model.

Wherein, step 2.2.3 specifically further includes:

step 2.2.3.1: adjusting and optimizing the parameters of max_depth, num_leave, min_data_in_leaf, in_split_gain, subsamples and colsample_byte in the classification model LGBMClassier;

step 2.2.3.4: parameter optimization is performed on the classification model LGBMClassiier, and a learning_rate parameter is adjusted, wherein a suitable candidate value of the learning_rate parameter is [0.01,0.015,0.025,0.05,0.075,0.1,0.2,0.3].

The boosting_type parameter under the LGBMClassifier tool introduces the type of the lifting tree, and a common gradient lifting method includes gbdt, dart, goss, rf. Different types of gradual enhancement lifting methods may be attempted to be run.

(1) gbdt: this is a traditional gradient-lifting decision tree, and is also an algorithm behind the excellent libraries based on XGBoost and pGBRT. The gbdt has high precision, high efficiency and good stability, and is widely applied at present. However, it has the major disadvantage that finding the best split point in each tree node is time consuming and consumes memory. Other lifting methods below attempt to solve this problem.

(2) dart: namely Dropouts meet Multiple Additive Regression Trees, dart solves the overfitted Regression Trees using dropout techniques (derived from deep neural networks), improving model regularization. The problem of over-specialization exists with gbdt, which means that trees added in later iterations tend to affect predictions for only a few instances, while the contribution to the remaining instances is negligible. Adding dropout makes it more difficult to specialize those few examples in later iterations, thereby improving performance. The principle is that the generated decision tree is randomly discarded, and then the optimized promotion tree is iterated from the rest decision tree set. It is characterized by slow training: training is slower because random dropouts do not use buffers for storing the prediction results. Random resulting in instability: early stops may not be stable enough because of randomness. Wherein dart differs from gbdt: when calculating the gradient to be fitted to the next tree, only a part is randomly selected from the already generated trees. Note that the addition of dart to a tree requires prior normalization.

(3) goss: the most important reason for this method named lightgbm, based on single-sided sampling of gradients, is that it uses the Goss-based method. The basic idea of goss is that firstly, training set data are ordered according to gradients, a proportion is preset to divide the gradient size, and data samples with large gradients in all samples are reserved; setting a sampling proportion, and proportionally extracting the sample from the sample with small gradient. To compensate for the effects on sample distribution, the goss algorithm multiplies the smaller gradient data set by a factor for amplification when calculating the information gain. Thus, the algorithm may focus more on "undertrained" sample data in calculating the information gain. goss significantly reduces computational effort by estimating gain for smaller sample data sets. Moreover, the goss algorithm does not excessively reduce the training accuracy. Standard gbdt is reliable but not fast enough on large data sets. Goss therefore proposes a gradient-based sampling method to avoid searching the entire search space. In essence, for each data instance, when the gradient is small, this means that there is no concern that the data is well trained, and when the gradient is large, it should be retrained. Examples of data have large and small fades. Therefore, goss saves all data in one large gradient and randomly samples data in one small gradient. This results in a smaller search space and faster convergence of goss.

(4) rf: random forests. If the enhancement is set to rf, the lightgbm algorithm appears as a random forest rather than an enhancement tree. From the document, to use rf, it is necessary to use bagging_fraction and feature_fraction smaller than 1.

And inputting the train_dataset into a model for training, obtaining a device safety state prediction model and storing the model. For example, in this embodiment, the input data is:

where transmissitter represents the transmitter device, diff represents the difference value and pump represents the pump device. The field details are as follows:

transmitter_1_max: the maximum value of the power of the transmitter No. 1;

diff_1_max: the maximum difference of the power of the transmitter No. 1;

transmitter_1_min: transmitter number 1 minimum power;

diff_1_min: the minimum difference of the power of the transmitter No. 1;

transmitter_1_mean: a transmitter number 1 power average;

diff_1_mean: the average difference of the power of the transmitter No. 1;

transmitter_1_sum: a transmitter power sum value number 1;

diff_1_sum: a transmitter power difference sum value number 1;

transmitter_1_media: a transmitter power intermediate value number 1;

diff_1_media: a median value of the transmitter power difference No. 1;

pump_1: pump power No. 1;

transmitter_1: transmitter power No. 1;

last_label: the device pump day before status;

performing predictive test on the stored model by using val_dataset, and evaluating a predictive result; the model meeting the standard is stored under a specified folder.

Step 3: device security state prediction.

Step 3.1: and acquiring real-time power data of each transmitter of the device on the same day from the device sensor monitoring system.

Step 3.2: and (3) performing data processing on the power data of each transmitter of the equipment by using the mode of the step (1) to obtain the power data of the equipment.

Step 3.3: and (3) predicting different transmitters and predicting the equipment pump by using the equipment safety state prediction model in the step (2).

In addition, the invention also provides a system for predicting the safety state of the nuclear power plant equipment, which comprises the following steps:

and the data preprocessing module is used for extracting characteristics of monitoring data of different equipment sensors.

And (3) sampling sensor data of each transmitter of the equipment pump, obtaining the power of the transmitter, and using data closest to the missing date to forward fill missing date data.

In view of the multiple transmitter condition of the same pump, the difference of the power of each transmitter in the same pump needs to be calculated. And calculating the maximum value, the minimum value, the average value, the sum and the intermediate value of the daily power data of each transmitter in the data set to obtain the power statistical data of the current day. The maximum value, the minimum value, the intermediate value, the sum and the average value of the daily diff of each transmitter in the data set are calculated, and in order to better distinguish equipment fault conditions, the equipment state in the data set on the next day of the day is subjected to numerical treatment, the occurrence of faults is recorded as 1, and the non-occurrence of faults is recorded as 0. And processing the data of each transmitter of different pumps on the same day and the maximum value, the minimum value, the intermediate value, the sum and the average value of the difference values of the power of each transmitter in the same pump, and integrating the data of the next year to obtain a device data set.

And the equipment safety state prediction model training module is used for training the equipment safety state prediction model for the equipment data set by using the classification model.

Device security state prediction model training is performed on the device data using the LGBMClassifier model in the sklearn toolkit.

Reading data in the equipment safety state prediction data set, and dividing a training set and a verification set;

parameter tuning optimization is carried out on parameters in the LGBMClassification model by using training data and verification data;

and the equipment safety state prediction module predicts the equipment safety state of the real-time data of the sensors in different areas of the equipment and gives out whether the equipment is normal or not.

Acquiring real-time power data of each transmitter of the equipment on the same day from an equipment sensor monitoring system;

performing characteristic processing on the real-time power data of each transmitter of the equipment to obtain equipment power data;

calling a device safety state prediction model, predicting the device safety state of different transmitters and predicting the device safety state of a device pump;

and returning the equipment safety state prediction result information to the background.

The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any changes or substitutions easily contemplated by those skilled in the art within the scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An application method for predicting the safety state of nuclear power plant equipment is characterized by comprising the following steps:

step 1: preprocessing data to obtain a device safety state prediction data set;

step 3: predicting the device security status.

2. The application method for predicting the safety state of a nuclear power plant equipment according to claim 1, wherein step 1 comprises:

step 1.1: sampling the data to obtain the power of a transmitter;

step 1.5: and calculating the maximum value, the minimum value, the intermediate value and the sum of the daily diff of each transmitter in the data set to obtain the difference statistical data of the current day.

3. The method for predicting the safety state of a nuclear power plant according to claim 2, wherein in step 1.1, the sampling frequency is aligned to within 10 seconds.

4. An application method for predicting a safety state of a nuclear power plant equipment according to claim 3, wherein the sampling frequency is aligned to 1 second.

5. The application method for predicting the safety state of a nuclear power plant equipment according to claim 1, wherein step 2 comprises:

step 2.1: digitizing the device status in the dataset the next day of the day;

6. The method according to claim 5, wherein in step 2.1, the LGBMClassifier tool in sklearn kit is used to model the data.

7. The method for predicting the safe state of a nuclear power plant according to claim 5, wherein step 2.2 comprises:

8. The method for predicting the safe state of a nuclear power plant according to claim 7, wherein the step 2.2.3 comprises:

9. The application method for predicting the safety state of a nuclear power plant equipment according to claim 1, wherein step 3 comprises:

10. A system for predicting a safety state of a nuclear power plant installation, comprising: