KR20200133311A - Early warning device and method of power plant using switching technology of multiple-predictive models - Google Patents

Abstract

The present invention relates to an early warning device of a power plant using a switching technology of a multi-prediction model and to a method thereof. The early warning device comprises: a multi-prediction model module including a plurality of prediction models for receiving information on an operation state of devices of a power plant and outputting a predicted value and the reliability of the predicted value for each of a plurality of sections; a reliability analysis module quantifying and analyzing the reliability output from the multi-prediction model, switching one prediction model among the plurality of prediction models with a current prediction model based on the reliability output by the plurality of prediction models included in the multi-prediction, and outputting a final predicted value; a predicted value flattening module flattening discontinuities of the predicted value output from the plurality of prediction models, comparing the final predicted value output for each of the plurality of sections with an actual measured value, and outputting the residual; and a determination module analyzing the residual and determining whether the operation state of the devices of the power plant has a defect.

Description

Early warning device and method of power plant using switching technology of multiple-predictive models}

The present invention relates to an early warning apparatus and method of a power plant using a switching technology of a multi-prediction model, and more specifically, an early warning apparatus and a method for improving detection performance and reliability for early warning using a multi-prediction model. It's about how.

The early warning device is a system used to prevent device failure in advance, and has been widely introduced and used in military, aviation, and power generation fields.

The principle of the early warning device is that the predictive model makes a predicted value for the driving state by using the past normal operation data, and by comparing the predicted value with the measured value to calculate the residual, it will alert when the residual is out of the normal operation range. Is to occur.

Conventionally, various mathematical methods are used to calculate predicted values for the driving state of devices, and representative methods include a kernel regression method using a statistical method, a Gaussian process regression method, a neural network method, and a Kalman filter.

Early warning devices can have a practical effect only when the probability of false alarms is minimized. Otherwise, the operator may miss important alarms while analyzing a large number of alarms, and the operator's concentration may be degraded, making it impossible to spot the alarms that are caused by real machine problems.

In this situation, conventionally, there have been attempts to improve detection performance and reliability for early warning by using a multiple prediction model, but in this case, it is necessary to define an algorithm for calculating a final predicted value from a plurality of predicted values. Therefore, there is a need for research on a process of outputting a plurality of predicted values while using a multiple prediction model and flattening the reliability of the plurality of predicted values.

Korean Patent Registration No. 10-1827108

The present invention is to solve the problems of the prior art as described above, and it is an object of the present invention to perform an early warning by detecting abnormal signs during starting and stopping of a power plant.

In order to achieve the above object, an early warning device of a power plant using the switching technology of a multi-prediction model according to the present invention receives information on the operating state of the devices of the power plant and outputs a predicted value and a reliability of the predicted value for each of a plurality of sections. A multiple prediction model module having a plurality of prediction models; Quantifying and analyzing the reliability output from the multiple prediction model, and switching one prediction model among the plurality of prediction models with the current prediction model based on the reliability output from the plurality of prediction models included in the multiple prediction model And a reliability analysis module that outputs a final predicted value; A predicted value flattening module for flattening discontinuities of the predicted values output from the plurality of predictive models, comparing the final predicted values output for each of the plurality of sections with actual measured values, and outputting a residual; And a determination module that analyzes the residual and determines whether or not there is a defect in the operation state of the devices of the power plant.

Here, the reliability analysis module is the first reliability when the first reliability output from the first prediction model and the second reliability output from the second prediction model are the highest in succession for each of the plurality of sections. And comparing the second reliability and determining that the difference between the first reliability and the second reliability is within a preset value, the first prediction model or the second prediction model may be used continuously.

In addition, the first prediction model or the second prediction model may be used continuously in five of the plurality of sections.

In addition, the reliability analysis module may compare the reliability of the current section and the next section in real time when switching from the one section to another section among the plurality of sections, and determine section switching when the reliability of the next section is high.

In addition, when the reliability of at least two prediction models is similar in a certain interval, the reliability analysis module may determine the final prediction value by applying a weight to each prediction value according to the reliability to calculate an average.

In addition, the predicted value flattening module may smooth the discontinuity of the final predicted value using one of a moving average filter or a low-pass filter.

On the other hand, in the early warning method of a power plant using the switching technology of a multi-prediction model, a multi-prediction model module equipped with a plurality of predictive models receives information about the operation status of the equipment of the power plant, and the predicted value and the predicted value are Outputting the reliability; The reliability analysis module quantifies and analyzes the reliability output from the multiple prediction model, and determines one prediction model among the plurality of prediction models based on the reliability output from the plurality of prediction models included in the multiple prediction model. Switching to the prediction model and outputting a final prediction value; A predicted value flattening module flattening discontinuities of the predicted values output from the plurality of predictive models, comparing the final predicted values output for each of the plurality of sections with an actual measured value, and outputting a residual; And determining, by the determination module, whether or not there is a defect in the operation state of the devices of the power plant by analyzing the residual.

The early warning apparatus and method of a power plant using the switching technology of a multi-prediction model according to the present invention has the following effects.

First, it is possible to overcome the vulnerability that occurs in a single prediction model. The early warning apparatus according to the present invention uses multiple prediction models rather than a single prediction model. Accordingly, a case in which the predicted value may be different from the actual value due to defects in each predictive model can be compensated for, thereby overcoming the vulnerability.

Second, it increases reliability. In the present invention, by applying a multi-prediction model, early warning is performed on the operating states of devices of a power plant. Accordingly, it is possible to enhance the capability of detecting a defect than a single prediction model, and increase the reliability of an early warning.

1 is a block diagram of an early warning device of a power plant using a switching technology of a multiple prediction model according to an embodiment of the present invention.
FIG. 2 is a diagram schematically illustrating a function execution between components of an early warning device of a power plant using a switching technology of a multiple prediction model according to an embodiment of the present invention.
3 is a flowchart of an early warning method of a power plant using a switching technique of a multiple prediction model.
FIG. 4 is an exemplary graph between time and predicted value of power plant output when a plurality of predictive models provided in a multi-prediction model switches in a plurality of sections in an early warning device of a power plant using a switching technology of a multi-prediction model It is a drawing shown as.
FIG. 5 is a diagram exemplarily showing that a predicted value flattening module serializes predicted values for power plant outputs for a plurality of sections in the graph between the time shown in FIG. 4 and predicted values for power plant outputs.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings. However, some configurations irrelevant to the gist of the invention will be omitted or compressed, but the omitted configuration is not necessarily a configuration unnecessary in the present invention, and will be combined and used by a person having ordinary knowledge in the technical field to which the present invention belongs. I can.

1 is a block diagram of an early warning device of a power plant using a switching technology of a multiple prediction model according to an embodiment of the present invention.

As shown in Fig. 1, the early warning device of a power plant using the switching technology of a multiple prediction model according to an embodiment of the present invention includes a signal measurement unit 100, a multiple prediction model module 200, and a reliability analysis module 300. ), a predicted value flattening module 400, and a determination module 500.

The signal measuring unit 100 is configured to receive a monitoring signal related to the operation state of the equipment of the power plant, calculate an actual measured value for the operation state of the equipment of the power plant, and transmit the calculated measured value to the determination module 500. In addition, the signal measuring unit 100 receives a monitoring signal related to the operation state of the equipment of the power plant and transmits it to the multiple prediction model module 200. Here, the monitoring signal for the operating state of the equipment of the power plant relates to pressure, temperature, flow data, and the like. The signal measurement unit 100 is attached directly to or around the power plant devices to measure the operating conditions of the devices and transmit a monitoring signal. The transmission interval can be transmitted at a time interval specified by the user. For example, it may be transmitted at 1 second intervals or for each of a plurality of driving sections.

The multiple prediction model module 200 is a component that receives a monitoring signal measured by the signal measurement unit 100 and generates a predicted value for the driving state of the devices for each of a plurality of sections. Here, a plurality of sections refers to a time-series division of the operating states of the equipment of the power plant. The multiple prediction model module 200 may be provided with a plurality of prediction models from 1 to n. For example, the first prediction model 210, the second prediction model 220, ... , N-th prediction models may be provided, and different algorithms may be applied to the plurality of prediction models to generate prediction values. In addition, each prediction model outputs not only the predicted value of the current monitoring signal, but also the reliability of the predicted value. Therefore, by monitoring the operation status of the equipment of the power plant in each of the plurality of sections, the measured value and the predicted value are compared and monitored based on the monitoring signal whether the equipment of the power plant operates normally in the plurality of sections, and the residual, which is the difference between the measured value and the predicted value, is preliminarily determined If it is calculated outside the set range, it can trigger an early warning.

The reliability analysis module 300 receives the reliability of the predicted value from a plurality of prediction models, quantifies and analyzes the received reliability to determine the predicted value of the power plant devices in one section from among a plurality of sections for the operation state of the power plant devices. To do this, one prediction model is determined from among a plurality of prediction models, and a predicted value of the determined prediction model is output. The reliability analysis module 300 may switch a prediction model applied in a previous section to a new prediction model based on reliability in a plurality of sections. Here, switching refers to changing and using a prediction model with high reliability among a plurality of prediction models for each of a plurality of sections. For example, among a plurality of sections, a first section is a first prediction model 210, a second prediction model 220, which is a plurality of prediction models. , Among the n-th prediction models, the second prediction model 220 having the highest reliability can be used, and the second section is the first prediction model 210, the second prediction model 220, ... , Among the n-th prediction models, the first prediction model 210 having the highest reliability may be used.

The predicted value flattening module 400 is configured to flatten discontinuities of predicted values output from the predicted model switched for each of a plurality of sections, compare the final predicted values output for each of the plurality of sections with actual measured values, and output a residual. Here, as the final predicted value, the predicted value of the predicted model with the highest ranking may be used, the arithmetic average of the upper two predicted values may be used, or a weighted average according to the ranking may be used.

The determination module 500 is a component that compares and analyzes a residual output from the predicted value flattening module 400 and an actual measured value to determine whether or not there is a defect in the operation state of the power plant devices.

Hereinafter, a functionally connected relationship between components of the early warning device of a power plant using the switching technology of the multiple prediction model module 200 will be described.

First, let's take a brief look at that early warning is performed in the operating state of the power plant.

FIG. 2 is a diagram schematically illustrating a function performance between components of an early warning device of a power plant using the switching technology of the multiple prediction model module 200 according to an embodiment of the present invention.

As shown in Figure 2, the early warning device of a power plant using the switching technology of the multi-prediction model module 200 according to the present invention measures a signal to perform an early warning when there is a problem with the devices of the power plant in the operating state. The unit 100 measures a monitoring signal related to the output of the operating state of the power plant devices for each of a plurality of sections.

The monitoring signal measured by the signal measurement unit 100 is received by the multi-prediction model module 200, and the first prediction model 210, the second prediction model 220 and the multi-prediction model module 200 are provided. … , The reliability analysis module 300 outputs a prediction value and a reliability level based on a monitoring signal for each of a plurality of sections received by a plurality of prediction models, such as an n-th prediction model, and transmits the output to the prediction value flattening module 400.

In addition, the predicted value flattening module 400 flattens the predicted values output from different predictive models, selects a final predicted value, and compares and analyzes the predicted value with the measured value, and outputs a residual.

Thereafter, the residual is transmitted to the determination module 500, and the determination module 500 analyzes the residual to determine whether or not there are defects in the devices of the power plant, so that an early warning may be performed to the user.

Hereinafter, an early warning method of a power plant using a switching technology of a multiple prediction model will be described with reference to the drawings.

3 is a flowchart of an early warning method of a power plant using a switching technique of a multiple prediction model.

As shown in Fig. 3, the power plant early warning method to which the multi-prediction model according to an embodiment of the present invention is applied initially receives information on the operation status of the equipment of the power plant, The prediction model of outputs the predicted value and the reliability of the predicted value. <S30>

In this case, the signal measuring unit 100 may be attached directly to or around the devices of the power plant to receive a monitoring signal for the devices of the monitoring power plant.

The received monitoring signals are analog signals, and accordingly, the signal measurement unit 100 processes the data of the monitoring signal and undergoes an analog-to-digital conversion step, thereby converting them into digital signals. The signal measuring unit 100 may transmit the measured values of the monitoring signals converted into digital signals to the predicted value flattening module 400.

In addition, the monitoring signal measured by the signal measurement unit 100 may be transmitted to the multiple prediction model module 200 configured with a plurality of prediction models. Accordingly, among the devices of the power plant measured by the signal measuring unit 100, the multi-prediction model module 200 configured of a plurality of prediction models may output a predicted value of a corresponding device. Depending on the use of such a plurality of prediction models, the first prediction model 210, the second prediction model 220, ... , Up to the nth prediction model can be used. Here, the first prediction model 210, the second prediction model 220, ... , The n-th prediction model may output device prediction values by applying different algorithms.

For example, the first prediction model 210 may use a data model using an Auto Associative Kernel Regression (AAKR) technique, and the second prediction model 220 may use a physical model using a Kalman filter. have. Accordingly, the first prediction model 210 may output a prediction value using the AAKR technique, and the second prediction model 220 may output a prediction value using a Kalman filter.

Specifically, the AAKR technique generates a pattern model by learning past data, determines similarity by comparing the current measured value and the pattern model, and generates a predicted value by weighting the pattern model based on the similarity. Therefore, the AAKR technique is an interpolation type prediction model that can be predicted within the learned pattern model, and can be predicted only when the measured value exists in the learned pattern model. In determining the similarity, it is possible to check whether the measured value exists in the learned pattern region, and the reliability of the predicted value can be calculated by synthesizing the similarities.

When a prediction model to which a plurality of AAKR techniques are applied is applied, each AAKR model may output a prediction value and reliability. Here, the reliability means the similarity between the measured value and the pattern model.

The Kalman filter is a statistical model based on a physical model, and predicted values and covariances can be output. Such a Kalman filter can be predicted even in an untrained region, and the uncertainty (reliability) of a predicted value can be determined through covariance.

After the prediction value is output, the first prediction model 210, the second prediction model 220, ... , In the n-th prediction model, reliability for each predicted value may be output. At this time, the size and meaning of the reliability, and the calculation method may be different for each algorithm.

In addition, the reliability is the first prediction model 210, the second prediction model 220, ... , It is an index that roughly shows how accurate the predicted value output for the device in the n-th prediction model is.

In addition, the first prediction model 210, the second prediction model 220, ... , The n-th prediction models are the first prediction model 210, the second prediction model 220,… as all AAKR techniques or all Kalman filters are used. , And the n-th prediction models may all use the same algorithm. Even if the same algorithm is used, multiple prediction models can be used for each operating condition of the device.

For example, the first prediction model 210, the second prediction model 220, the third prediction model (not shown) and the fourth prediction model for the four condenser temperatures for each season in consideration of seawater temperature change. (Not shown), and the first prediction model 210, the second prediction model 220, the third prediction model (not shown), and the fourth prediction model (not shown) are applied only with the AAKR technique to obtain a predicted value. Is to print.

Here, the condenser is a device that cools water vapor and returns it to water, and is a kind of condenser. Seawater is the ultimate heat sink of a power plant, and almost all of the heat generated by a nuclear power plant is discharged into seawater. Therefore, when the temperature of the seawater changes, the heat exchange process between the power plant and seawater changes, and as a result, the operation pattern of almost all facilities changes. One of the facilities affected by the seawater temperature is a condenser.

While operating an early warning system at a nuclear power plant, there is a case of a large number of false alarms occurring in the summer of 2016 due to abnormal high temperatures in seawater (2 degrees of temperature rise compared to normal learning data). Accordingly, in the existing single prediction model, the predicted range of the prediction model is increased by additionally learning data on the changed seawater temperature, or the alert sensitivity is reduced through a change that increases the allowable value of the determination module 500. .

However, such a method is not desirable because it reduces the overall reliability and detection capability of the early warning device. Therefore, by applying the multiple prediction model module 200 based on the four AAKR techniques for each seawater temperature, each operation pattern of nuclear power plant variables is learned according to the seawater temperature.

That is, a prediction model for spring, a prediction model for summer, a prediction model for autumn, and a prediction model for winter are generated. Then, when the measured value is input, it is determined which season-related prediction model is most similar, and the predicted value of the most similar prediction model is used.

As a result, since the prediction range is narrow and a plurality of precise prediction models are used to predict, it is to generate an effect that can improve the reliability of the early warning device than before.

In contrast, in the early warning device of a power plant using the switching technology of a multiple prediction model according to the present invention, the multiple prediction model module 200 equipped with a plurality of prediction models provides a prediction model to which the AAKR technique is applied and a prediction model to which a Kalman filter is applied. They can be used interchangeably, and the prediction model to which the AAKR technique is applied outputs similarity with reliability, and the prediction model to which the Kalman filter is applied outputs the covariance, so that the predicted value and reliability can be output.

After that, the reliability analysis module 300 compares and analyzes the reliability outputs of the plurality of prediction models provided in the multiple prediction model module 200 in one of the plurality of sections of the operating state of the power plant, and a plurality of prediction models are performed. By selecting one of them, the final predicted value of the corresponding prediction model for one section is output. <S31> This will be described with reference to the drawings.

FIG. 4 is a graph between time and predicted values for power plant output when a plurality of predictive models provided in the multi-prediction model module 200 switch each of a plurality of sections in an early warning device of a power plant using a switching technology of a multi-prediction model. It is a diagram exemplarily shown through.

As shown in Fig. 4, a plurality of sections relating to the operating state are formed between the operating states of the power plant devices. Accordingly, the multiple prediction model module 200 provided with a plurality of prediction models includes a first prediction model 210, a second prediction model 220, ... , The n-th prediction model is a first prediction model 210, a second prediction model 220,… for each of a plurality of sections from section 1 to section n. , The n-th prediction model is used to output the predicted values and reliability for all of one section.

Thereafter, in the reliability analysis module 300, the first prediction model 210, the second prediction model 220,… for one section. , By comparing and analyzing the reliability output from the n-th prediction model, a prediction model with the highest reliability may be selected in a corresponding section. This means that the reliability analysis module 300 selects a prediction model for each section among a plurality of sections in which the operating states of the devices of the power plant are classified, and the first prediction model 210, the second prediction model 220, ... , The predicted value of the predictive model with the highest reliability among the predicted values output from the n-th predictive model is output as the final predicted value.

In this case, when the reliability is higher than a preset threshold value, the reliability analysis module 300 determines that a normal signal for classifying the corresponding section is received, so that the conversion of the section is determined. I can.

In addition, as another embodiment of determining section switching, when the reliability of the next section is high by comparing the reliability of the current section and the next section in real time, section switching may be determined.

In addition, it is also possible to determine the section change through the state of the representative signal.

*Specifically, the representative signals are the main signals of the power plant such as the speed of the reactor coolant pump (RCP), the temperature of the reactor coolant system (RCS), the temperature of the reactor coolant, the power of the reactor, the speed of the turbine, and the power of the generator. In addition to the result of the reliability analysis module 300, it can be used as an auxiliary.

For some monitoring signals related to CCW (Component Cooling Water), if the RCP speed is 1100RPM or higher and the RCS low temperature tube temperature is 294℃, the operation can be the same as normal operation.

On the other hand, some monitoring signals of the CVCS (chemical and volume control system) system can start normal operation from the point when the reactor output exceeds 0.1%. These monitoring signals can be used by creating two groups of predictive models, such as start and stop and normal operation, and convert them for each section.

In addition, in the case of some monitoring signals of the main steam system or condenser system, the operation data may change continuously depending on the power plant output. These monitoring signals can be selectively used by having a number of prediction models for each output section.

Therefore, the section is not uniformly classified by applying the same criteria, and a user or an expert directly classifies the section in consideration of the characteristics of driving data for each monitoring signal. The standard for dividing the section is the value of the representative monitoring signal, and the approximate standard is, for example, the RCP speed above 1100RPM, the RCS low temperature tube temperature above 294℃, the reactor output 0.1%, 20% or 80%, turbine It can be divided into standards with a speed of 20RPM or higher and generator output of 20% and 80%.

For example, even if the reliability of the second prediction model 220 is highest based on the final prediction value output by the reliability analysis module 300, it is more suitable to use the first prediction model 210 based on the state of the representative signal. If the reliability analysis module 300 determines that it is, the final prediction value output by the reliability analysis module 300 may be ignored and the prediction value of the first prediction model 210 may be used.

Here, the multiple prediction model module 200 uses a plurality of prediction models, but the types of the plurality of prediction models may use the same prediction models.

Meanwhile, when a final prediction value corresponding to a plurality of sections is determined, by setting a switching condition of a prediction model, it is possible to prevent a plurality of prediction models from being frequently switched in a plurality of sections. For example, when the reliability corresponding to sections 1 to 8 is output from the multiple prediction model module 200 equipped with a plurality of prediction models, the first prediction model 210 and the second prediction model 220 output. It may be the most reliable. At this time, the reliability of the first prediction model 210 is the highest in the first section, the reliability of the second prediction model 220 is the highest in the second section, and the reliability of the first prediction model 210 again is the highest in the third section. , In section 4, the reliability of the second prediction model 220 is again the highest, which means that the first prediction model 210 and the second prediction model 220 alternately up to section 8 to output the highest reliability. At this time, there is a difference in reliability between the first prediction model 210 and the second prediction model 220 in terms of high and low in sections 1 to 8, but the prediction model with the highest reliability is the prediction model of the next ranking reliability. The reliability may be slightly higher. In other words, although the reliability of the first prediction model 210 in the first section is higher than that of the second prediction model 220, the difference may be within 5% of the reliability.

In this way, if the first prediction model 210 and the second prediction model 220 are alternately changed based on the highest reliability in the first to eighth sections, the reliability analysis module increases as the process increases due to continuous change. (300) And the load borne by the early warning device is bound to increase.

Therefore, in the present invention, when a prediction model is used for each of a plurality of sections, when the difference between the highest reliability and the next highest reliability for each adjacent section is not large, the prediction model used in the previous section is used as it is, so that the reliability analysis module 300 And to reduce the load borne by the early warning device.

For example, in an arbitrary section among a plurality of sections, the first prediction model 210 has the highest reliability and the second prediction model 220 has the highest reliability, and the second prediction model ( 220) has the highest reliability and the second prediction model 220 has the highest reliability, and the first prediction model 210 has the highest reliability and the second prediction model ( 220) is assumed to be highly reliable. That is, the reliability of the first prediction model 210 and the second prediction model 220 is continuously output in a certain section.

At this time, the difference in reliability output by each of the first and second prediction models 210 and 220 may be insignificant in each section described above, which means that the reliability of the first prediction model 210 is the second prediction model. It can be higher than the reliability of (220) within 5%. Therefore, in this case, after using the first prediction model 210 in an arbitrary section for each of the plurality of sections, the first prediction model 210 is not changed to the second prediction model 220 in the section following the arbitrary section. The prediction model 210 is used as it is. Therefore, only when the reliability is 5% or more, the first prediction model 210 is changed to the second prediction model 220 and the final predicted value is output. Here, 5% may be a value preset by the user in the reliability analysis module 300, and the preset value may be applied not only to 5% but also various values.

In addition, comparing the difference in reliability in an arbitrary interval may be determined by synthesizing samples of reliability before an arbitrary interval. For example, the first prediction model 210 has the highest reliability in the entire section immediately before an arbitrary section, and the first prediction model 210 has the highest reliability in a random section. , It is assumed that the reliability of the first prediction model 210 is next highest.

In this case, if the reliability analysis module 300 used the first prediction model 210 because the reliability of the first prediction model 210 is highest in five periods before a certain period, the reliability analysis module 300 Also, the first prediction model 210 may be used. If the reliability of the second prediction model 220 is highest in five sections including an arbitrary section in the future, the reliability analysis module 300 may use the second prediction model 220 in the next section of the arbitrary section. Here, the number of sections to be used can be applied in various ways as well as five sections. Accordingly, by continuously using the first prediction model 210 in a plurality of sections including an arbitrary section, it is possible to reduce the load on the switching of the prediction model in each section.

As a result, it is possible to reduce the load borne by the reliability analysis module 300 and the early warning device by preventing frequent switching of the prediction model.

In addition, when the reliability of the first prediction model 210 and the second prediction model 220 is almost the same and it is difficult to determine only the reliability, an external signal may be additionally used. For example, when a prediction model for each season is used, the reliability analysis module 300 may prevent frequent switching by additionally using seawater temperature data for switching.

In addition, high and low reliability output from a plurality of prediction models may be utilized in calculating a final predicted value. For example, when the reliability of one prediction model is overwhelmingly high, the final prediction value of the prediction model can be used, but when the reliability of two or more prediction models is similar, the weighted average of the prediction values according to the reliability is used. Can be printed.

For example, a case in which the reliability of the first prediction model 210, the second prediction model 220, and the third prediction model (not shown) is similar in an arbitrary section will be described. Here, the reliability of the first prediction model 210 is a, the reliability of the second prediction model 220 is b, and the reliability of the third prediction model is c. Here, the reliability is high in the order of a>b>c, and the difference between a and c is within 1%. Therefore, different weights are assigned to a, b, and c in the high order of a, b, and c, and the average of the weighted a, b, and c is calculated and used as the final predicted value for an arbitrary section.

Next, the predicted value flattening module 400 flattens discontinuities of predicted values output from different predictive models for each of the plurality of sections, compares the final predicted values output for each of the plurality of sections with the measured values, and outputs a residual. S32> This will be described with reference to FIG. 5.

FIG. 5 is a diagram exemplarily showing that the predicted value flattening module 400 serializes predicted values for the power plant output for a plurality of sections in the graph between the time shown in FIG. 4 and the final predicted value for the power plant output.

The final predicted value output by the multi-prediction model module 200 may be provided with a plurality of prediction models and may be output with different algorithms for each of a plurality of sections. Therefore, in order to compare the final predicted value with the actual measured value and output the residual, it is necessary to convert the final predicted value with one reference. Accordingly, as shown in FIG. 5, the final predicted value can be transformed through the operation of flattening the discontinuities.

A moving average filter or a low pass filter may be used for the smoothing of the predicted values.

The moving average filter is a filter that outputs an average of an arbitrary number of continuously input values, and the low-pass filter is a filter that cuts off a high-frequency portion and passes a low-frequency portion of the frequency components of the input values. Both of these filters smoothly connect sudden changes occurring in the input value.

For example, the moving average filter,

Can be represented by Here, N may represent the number of samples.

In addition, in the case of the first-order digital low-pass filter of the low-pass filter,

는 샘플링 타임, r은 시정수를 나타낼 수 있다.Can be represented by here,

Denotes a sampling time and r denotes a time constant.

In addition, the residual is calculated by using the difference between the actual measured value of the device and the final predicted value. That is, the residual can be calculated and output through |actual value-final predicted value|.

Finally, the determination module 500 analyzes the residual to determine whether there is a defect in the starting and stopping states of the power plant. <S33>

The determination module 500 may determine whether a monitoring signal for a corresponding device is included in a preset normal operation range based on the residual output from the predicted value flattening module 400.

Here, the normal driving range may be determined differently according to the user's setting. Therefore, if the residual value is out of the preset normal operation range, the monitoring signal for the device determines that the device is in an abnormal operation state and generates an early warning, and if the residual value does not exceed the preset normal operation range, an alarm Does not occur.

At this time, the user can adjust the normal operation range for generating an alarm.

For example, the normal operating range of the feed water pump is 45 (minimum) to 55 (maximum), and the deviation between the minimum and maximum values is 10.

Therefore, if the residual value is programmed to generate an alarm when it exceeds 30% (ㅁ3) of the deviation (10) of the normal operating range, the residual value is 60 and the calculated predicted value is 52. Becomes 8 and exceeds 30% of the deviation and an alarm is generated.

Conversely, if the measured value is 43 and the predicted value is 45, the residual is 2, and the alarm does not occur.

When an alarm occurs, the determination module 500 may transmit alarm status information to an integrated center device (not shown) that monitors all power plants so that a user can check it.

As described above, in the present invention, a plurality of predicted values and reliability are output based on a monitoring signal related to start and stop of a power plant using a multiple prediction model, and a residual value is output by comparing a predicted value with a high ranking according to the reliability with the measured value. An early warning is generated by judging whether the generated residual is within the normal operation range.

Preferred embodiments of the present invention described above are disclosed for the purpose of illustration, and those skilled in the art with ordinary knowledge of the present invention will be able to make various modifications, changes and additions within the spirit and scope of the present invention. And additions will be seen as falling within the scope of the claims of the present invention.

100: signal measuring unit
200: multiple prediction model module
210: first prediction model
220: second prediction model
300: reliability analysis module
400: predicted value flattening module
500: decision module

Claims (1)

In the early warning device of a power plant using a switching technology of a multiple prediction model,
A multi-prediction model module including a plurality of prediction models for receiving information on operating states of devices of a power plant and outputting a predicted value and a reliability of the predicted value for each of a plurality of sections;
By quantifying and analyzing the reliability output from the multiple prediction model, switching one prediction model among the plurality of prediction models with the current prediction model based on the reliability output from the plurality of prediction models provided in the multiple prediction model And a reliability analysis module that outputs a final predicted value;
A predicted value flattening module for flattening discontinuities of the predicted values output from the plurality of predictive models, comparing the final predicted values output for each of the plurality of sections with actual measured values, and outputting a residual; And
And a determination module that analyzes the residual to determine whether or not there is a defect in the operation state of the devices of the power plant,
The reliability analysis module is an early warning device for a power plant using a switching technology of a multi-prediction model for determining the switching of the section based on a state of a representative signal when switching from one section to another section among the plurality of sections.

Images