KR102440012B1 - Operating Attribute Value Prediction System including NOx generating Quantity at Power Station using AI and Method thereof - Google Patents

Abstract

The present invention provides a plurality of historical variable values, including fuel consumption, gas turbine output, NOx generation, denitrification reducing agent input, and exhaust heat recovery boiler temperature, which is the end temperature of the exhaust heat recovery boiler, that are time-series generated in the process of operating a power plant. a collection processing unit for collecting and processing raw data including; An artificial intelligence unit for predicting a plurality of predictive variable values corresponding to each of the past variable values for each unit time set after the present time following the past variable value by using artificial intelligence with the data stored in the collection and processing unit; characterized.
Accordingly, the power plant NOx generation using AI that can predict in advance various variable values including the generation amount of NOx, one of pollutants, using an artificial neural network based on past data generated in time series in the process of power generation. A driving attribute value prediction system and method may be provided.

Description

Operating Attribute Value Prediction System including NOx generating Quantity at Power Station using AI and Method thereof

The present invention relates to a system and method for predicting operating attribute values including the amount of NOx generated in a power plant using AI, and specifically, based on fuel consumption, which is an attribute value known in a power plant, the remaining attribute values such as NOx generation and ammonia reducing agent input Based on the predicted and predicted attribute value, the NOx generation amount, to minimize the NOx emission amount, the amount of ammonia reductant input predicted to be input is also calculated and the structure is improved so that the structure can be improved to preemptively respond. Operational attributes including the NOx generation amount of the power plant It relates to a value prediction system and method therefor.

In the course of industrial activities and human life, various types of garbage are emitted and hazardous substances are discharged into the atmosphere. Such air pollution pollutes the environment and threatens human life.

These environmental pollutants are also generated in power plants.

According to the Ministry of Trade, Industry and Energy's "Eighth Electricity Supply and Demand Plan" of the Ministry of Trade, Industry and Energy in Korea, as high-concentration fine dust has emerged as a social problem, domestic emissions of fine dust have been reduced by more than 30%. conversion is underway.

Combined cycle power generation generates power twice by turning a gas turbine and a steam turbine in one cycle to improve power generation efficiency. Recovery Steam Generator) has the advantage of greatly improving the efficiency of the power plant because it is used again to generate secondary power.

On the other hand, nitrogen oxides (NO, NO ₂ , N ₂ O, N ₂ O ₃ , N ₂ O ₄ , N ₂ O ₅ , There is a risk of environmental pollution due to the emission of 'NOx').

Nitrogen oxide, which is pointed out as a major source of fine dust, is one of the air pollutants emitted from gas combined cycle power plants. Gas combined cycle power generation is evaluated as a relatively safe and eco-friendly power source, but as these facilities increase, there is a concern that the generation of NOx, iron oxide, fine dust, etc. may increase.

Meanwhile, the government is strengthening the NOx emission limit of gas combined cycle power plant to 10ppm.

Currently, the amount of NOx generated from power plants is reduced by the oxidation catalyst method (SCR, Selective Catalystic Reduction), but there is a risk that the amount of NOx generated may exceed the emission standard due to uniform operation.

In the process of operating the combined cycle power plant, a lot of data is measured by sensors and generated in time series. such as the temperature or pressure of the location.

The conventional NOx reduction operation method is a method of operating the SCR based on the generated NOx concentration measurement value. Therefore, the conventional method has a problem in that it is difficult to respond to a case where a large amount of exhaust gas is instantaneously generated according to the operation mode of a gas turbine (GT), which is a core facility of a power plant.

In addition, in the “Basic Direction of the Korean Version of the Digital New Deal” announced in July 2020, it is proposed to establish a leading technology in the data SOC field and artificial intelligence field, and promote a policy of industrial structure change and job creation through this.

On the other hand, artificial intelligence, a computer program or a computer system including the human learning ability, reasoning ability, perceptual ability, argumentation ability, natural language comprehension ability, etc. artificially implemented is being grafted in various fields.

Therefore, it is desirable to predict the amount of NOx generated in industrial facilities such as a gas combined cycle power plant, and to preemptively prepare a countermeasure to reduce the predicted amount of NOx to minimize the amount of NOx that is finally generated and emitted. .

[Related technical literature]

Laid-open Patent Publication No. 10-2014-0130538 (published on November 10, 2014)

Unexamined Patent Publication No. 10-2019-0123837 (published on April 11, 2019)

Unexamined Patent Publication No. 10-2020-0014048 (published on February 10, 2020)

An object of the present invention is to predict in advance the value of various variables including the generation amount of NOx, one of pollutants, using an artificial neural network based on past data generated time-series in the process of power generation. An object of the present invention is to provide a system and method for predicting driving attribute values including an amount generated.

In addition, another object of the present invention is to predict the operating attribute value including the amount of NOx generated by the power plant using AI that can preemptively take measures to reduce pollutants including NOx predicted in advance according to the operation plan of the power plant To provide a system and a method therefor.

In addition, another object of the present invention is to improve the stability of the operation of the power plant and reduce the emission of pollutants, thereby predicting the operating attribute value including the amount of NOx generated by the power plant using AI for eco-friendly operation, and a method therefor will provide

An object of the present invention is to provide a plurality of historical variables including fuel consumption, gas turbine output, NOx generation, denitrification reducing agent input, and exhaust heat recovery boiler temperature, which is the end temperature of the exhaust heat recovery boiler, that are generated time-series in the process of operating a power plant. a collection processing unit for collecting and processing raw data including values; An artificial intelligence unit for predicting a plurality of predictive variable values corresponding to each of the past variable values for each unit time set after the present time following the past variable value by using artificial intelligence with the data stored in the collection and processing unit; Characterized by the driving attribute value prediction system.

In addition, at least one of the predictive variable values includes fuel consumption as a basic predictive variable value that is a pre-predicted value, and the power plant includes a combined cycle power plant using LNG as a fuel, and the rest based on the basic predictive variable value It is preferable to predict the predictor value.

In addition, the artificial intelligence unit preferably includes a multivariate prediction model in the artificial neural network.

On the other hand, an object of the present invention is to provide a plurality of various fuel consumption, gas turbine output, NOx generation, denitrification reducing agent input, and exhaust heat recovery boiler temperature, which is the end temperature of the exhaust heat recovery boiler, that are generated in time series in the process of operating a power plant. (1) process of collecting and processing raw data including past variable values in a collection and processing unit; (2) process of predicting, in the artificial intelligence unit, a plurality of predictive variable values corresponding to each of the past variable values for each unit time set after the present time following the past variable value by using artificial intelligence on the data stored in the collection and processing unit It is also achieved by the driving attribute value prediction method, characterized in that it includes.

In addition, the artificial intelligence unit preferably includes a multivariate prediction model in the artificial neural network.

In addition, it is preferable that the past variable value is a value measured by a sensor installed at a set position, and at least one of the predictive variable values is a basic predictive variable value that is a pre-predicted value and includes fuel consumption.

In addition, (3) step of extracting the value of the basic predictive variable after the current time and passing through the first linear layer; (4) selecting the predictor value that has passed through a Rectified Linear Unit (ReLU) that predicts the predictor value based on the historical variable value, the basic predictive variable value, and the historical variable value at the present time (4) It is preferable to further include a process;

In addition, in the process (4), it is preferable that the past variable value of the current time is initially selected, and then the predictor variable value that has passed through the ReLU is selected.

In addition, it is preferable to further include a (5) process in which the variable values passed through the steps (3) and (4) are combined in the contact layer and pass through the second linear layer.

In addition, it is preferable that the variable value passing through the process (5) and the variable value passing through the encoder include the process (6) passing through a Convolution Neural Network (CNN).

In addition, it is preferable that the value of the variable passed through the encoder in the process (6) is abbreviated by extracting the value of the past variable divided at a set time interval from the present time to the past.

In addition, the CNN includes the steps of (6-1) predicting variable values excluding the basic predictive variable values in the GRU (Gated Recurrent) layer using the variable values input in the step (6); It is preferable that the variable value passing the process (6-1) passes through the third linear layer, then through the ReLU layer, and then through the fourth linear layer to become the predictive variable value.

In addition, the process (3) and the process (4) are passed according to the value of the predictor according to the value of the basic variable corresponding to the set time after the value of the predictor predicted in the process (6-2), and the process (4) is performed. It is preferable to repeat the steps of predicting the predictor value through the steps (5) and (6).

Accordingly, according to the present invention, a power plant using AI that can predict in advance various variable values including the amount of NOx, one of pollutants, using an artificial neural network based on past data generated time-series in the process of power generation It is possible to provide a system and method for predicting driving attribute values including an amount of NOx generated.

In addition, it is possible to provide a system and method for predicting operating attribute values including the amount of NOx generated in a power plant using AI that can preemptively take a measure to reduce pollutants including NOx predicted in advance according to the operation plan of the power plant. can

In addition, it is possible to provide a system and method for predicting operating attribute values including the amount of NOx generated in a power plant using AI that can improve the stability of power plant operation and reduce the emission of pollutants for eco-friendly operation.

1 is a schematic diagram for explaining the relationship between combined cycle power generation and driving attribute value prediction system according to an embodiment of the present invention;
2 is a conceptual diagram of a driving attribute value prediction system according to an embodiment of the present invention;
3 is a conceptual diagram of the structure of ConvGRUNet, which is a time series prediction system;
Figure 4 is a flow chart showing a specific embodiment of Figure 3;
5 is a monitoring schematic diagram of the driving attribute value prediction system of the present invention;
6 is a graph comparing prediction performance of VAR and ConvGRUNet according to the present invention;
7 is a graph comparing predicted values and actual values of a plurality of driving attribute values according to the driving attribute value prediction system according to the present invention.

A system for predicting an operating attribute value including an amount of NOx generated by a power plant using AI according to an embodiment of the present invention and a method therefor (100, hereinafter referred to as a 'prediction system') will be described in detail with reference to FIGS. 1 to 7 below. It is explained as follows.

1 is a schematic diagram for explaining the relationship between combined cycle power generation and driving attribute value prediction system according to an embodiment of the present invention, FIG. 2 is a conceptual diagram of a driving attribute value prediction system according to an embodiment of the present invention, FIG. 3 is It is a conceptual diagram of the structure of ConvGRUNet as a time series prediction system, FIG. 4 is a flowchart showing a specific embodiment of FIG. 3 , FIG. 5 is a monitoring schematic diagram of the driving attribute value prediction system of the present invention, and FIG. 6 is VAR and ConvGRUNet according to the present invention It is a prediction performance comparison graph, and FIG. 7 is a graph comparing predicted values and actual values of a plurality of driving attribute values according to the driving attribute value prediction system according to the present invention.

Prior to describing the present invention in more detail, the present invention may have various changes and may have various forms, embodiments (態樣, aspect) (or implementation

ex) will be explained in detail in the text. However, this is not intended to limit the present invention to the specific disclosed form, it should be understood to include all modifications, equivalents and substitutes included in the spirit and scope of the present invention.

In each drawing, the same reference sign, particularly the number of the tens and one digit, or the reference numerals having the same tens, one, and alphabet, indicates a member having the same or similar function, and unless otherwise specified, each of the figures in the drawings The member indicated by the reference sign may be understood as a member conforming to these standards.

In addition, in each drawing, in consideration of the convenience of understanding, the size or thickness is exaggeratedly large (or thick) small (or thin) or simplified, but the protection scope of the present invention is limited by this. it shouldn't be

The terminology used herein is only used to describe a specific embodiment (aspect, aspect, aspect) (or embodiment), and is not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as comprises or consists of are intended to designate that the features, numbers, steps, operations, components, parts, or combinations thereof described in the specification exist, but one or more other features It is to be understood that it does not preclude the possibility of the presence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

The prediction system 100 according to the present invention is, as shown in FIGS. 1 to 5, fuel consumption, gas turbine output, NOx generation, denitrification reducing agent input, and heat recovery boiler time-series generated in the process of operating a power plant. a collecting processing unit 110 for collecting and processing raw data including various values of a plurality of past variables including the temperature of the exhaust heat recovery boiler which is the end temperature of the ; The artificial intelligence unit 130 for predicting a plurality of predictive variable values corresponding to each of the past variable values for each unit time set after the present time following the past variable values by using the data stored in the collecting and processing unit 110 using artificial intelligence. ); is preferred.

1, the power plant according to the present invention includes a combined cycle power plant. It is a power plant in which a steam turbine is installed along with a gas turbine (G/T). First, a gas turbine is operated to generate electricity with liquefied natural gas (LNG), and the high-temperature gas generated from the gas turbine is guided to a recovery device, and steam is generated from it to operate the steam turbine to generate electricity.

During the operation of such a power plant, the state of each device, operating conditions, power generation, and pollutant generation are monitored by sensors installed at each location, and these monitored data are sent raw to a device called a data collection system (DCS). Data (refer to 'DSC (Raw Data)' in FIG. 2) is stored.

These raw data include the amount of NOx emitted to the atmosphere, the amount of NOx generated before the emission, and the amount of decontamination agent used in the SCR device to reduce the generated NOx, G/T output, G/T fuel consumption, The outside temperature is also included.

In order to apply and predict various data generated during the operation of the power plant, time is divided into the present time, the past before the present time, and the future after the present time, and these past times and future times are set as needed. It can be divided into units and applied.

In addition, among various data generated during the driving process, there are variable values that can be predicted in advance among the driving dependent values (hereinafter referred to as 'variable values') according to the driving conditions. This is the input amount of LNG fuel input - this fuel input amount can be calculated based on the power generation amount set in advance according to the operation plan of the power plant) is an example, and the fuel input amount that can be predicted in advance is referred to as the basic predictive variable value below.

The prediction system according to the present invention is a predictive system 100 according to the present invention, which is a prediction system 100 according to the present invention, which will be described later based on the variable values that occur, occur, or will occur in the combined cycle power generation as shown in FIG. can be predicted and the predicted variable values can be managed. In the present invention, in particular, in order to predict the amount of NOx generated in the process of power generation and input the predicted reducing agent input amount to reduce NOx in response to the predicted amount of NOx, to minimize the amount of NOx finally discharged to the atmosphere, preemptively can be controlled and managed.

The prediction system 100 according to the present invention includes a time series data-based deep learning prediction system, and collects, stores, manages, and utilizes time series data as shown in FIGS. 2 to 3 . That is, the present invention is related to multivariate time series prediction that predicts the pollution situation by using a deep learning model after collecting time series data from a gas turbine of an LNG power plant.

In the prediction system 100, as shown in FIG. 2, multivariate time-series data is processed through a collection-store-utilization step, and in the collection step, ETL (Extract/Transform/Load), MQ Various techniques are used, including preprocessing techniques such as (Message Queue) and CEP (Complex Event Processing), and in the storage stage, they can be stored in the Enterprise Data Warehouse or Cloud.

Time series data that has gone through this data engineering stage is learned and predicted through statistical analysis, rule-based machine learning, and the recent deep neural network technology.

First, in order to learn a deep learning-based predictive model, it is necessary to acquire a large amount of time series data from the power plant site, and this time series data generates multi-channel time series data through various sensors installed in LNG Gas Turbine and HRSG (Heat Recovery Boiler). and data is accumulated in DCS. The raw data of DCS is collected through the data engineering stage and pre-processed in a form suitable for machine learning through processing, transformation, and labeling processing.

An example showing the NOx generation amount and the associated variance collected from the power plant is shown in <Table 1>.

<Table 2> shows the relationship between variables in the data through correlation analysis between the collected multivariate data, and an example of understanding this relationship.

As the capacity of the deep learning model increases, the generalization ability of the model improves, but the parameters also increase exponentially and a large amount of training data is required. In addition, it may cost a lot of money and time to use only the data collected from the power plant for learning, and since it is collected in a situation where the surrounding environmental conditions are limited, the prediction accuracy may decrease in a situation where the conditions change. Here, the data augmentation technique is effective as a method for solving this data collection problem. Although many data augmentation techniques have been studied in the computer vision field, there are few studies related to augmentation on 1D time series data, and it is desirable to reinterpret and effectively apply the augmentation techniques in the vision field to 1D data.

In multivariate properties, not only correlations between properties, but also short-term time-series data dependence and long-term time-series data dependence can all exist, so it is important to understand the regularity based on these correlations. Therefore, it is desirable to use AttentionNet, Temporal ConvNet, and RNN/LSTM/GRU time series pattern analysis algorithms as a base technology to improve it to fit the power plant time series data.

Existing time series data prediction techniques (ARIMA, VAR, SVR, GP, etc.) have a problem in that it is difficult to predict long-term temporal patterns or high-dimensional multivariate time series prediction due to high computational costs.

The prediction system 100 according to the present invention introduces a deep neural network-based model such as DeepGLO, DeepAR, AttentionNet, ConvGRUNet, etc., and through this, various variables are put as input values, and various variables are put as input values, unlike the existing prediction models that use one variable as a predictive value. It includes a multivariate prediction model that can predict

In other words, the existing time series prediction models only show trends and predict results with different scales, but DeepGLO and ConvGRUNet can predict both the trend and scale of the results well by combining the autoregressive models.

3 shows a conceptual diagram of a structure of a deep neural network-based ConvGRUNet applied to the prediction system 100 of FIG. 2 . It will be described in detail with reference to FIGS. 2 and 3 as follows.

G/T operation generation data and multivariate data (multiple sensor channels) generated from NOx pollution prevention facilities are streamed and collected through DCS (Data Collection System). The collected raw data is processed and stored through the ETL and data engineering process through the collection processing unit 110 , and is input to the artificial intelligence unit 130 including a deep neural network for time series prediction.

In the first part of the deep neural network, cross-correlation and short-time correlation between variables (channels) are modeled through a 2D convolutional layer. After that, long-time correlation is extracted through the LSTM/GRU layer, and the result goes out to the output node through the FCN.

6 is a graph comparing the performance of the currently widely used multivariate prediction model VAR (Vector Autoregression) and the ConvGRUNet model of the prediction system 100 of the present invention. ConvGRUNet predicts not only the signal change tendency but also the scale well. show

It is preferable to use PyTorch, CUDA, and CuDNN under Ubuntu Linux OS for the development environment of the prediction system 100 according to the present invention.

PyTorch is a Python-based machine learning and deep learning library maintained by Facebook's artificial intelligence research team. CUDA is a GPGPU that enables parallel processing algorithms performed by the graphics processing unit (GPU) to be written using industry standard languages including the C programming language. CUDNN is a high-performance Deep Neural Networks library provided by Nvidia and is used to improve the computation speed of deep learning.

Here, various sensor signals from G/T itself are collected in a device called DCS. DCS is a kind of data collection server, and it is desirable to use data received by connecting to DCS for all subsequent processing.

In addition, the collection processing unit 110 processes the data so that it can be used in deep learning because raw data from DCS is unorganized data and is difficult to use for deep learning learning or prediction as it is, and time series data obtained from DCS Processes multiple measured values (some sensor data is missing at a certain time), removes high-frequency or noise signals, normalizes values within a certain range (usually between 0~1 or -1~1), ETL (Extract, Load) , Transform) process and save the data log.

The artificial intelligence unit 130 includes a high-performance server computer that performs learning and prediction tasks for MTSR (Multivariate Time Series Regression), MTSC (multivariate time series classification), and the like.

In addition, as shown in FIG. 5, the monitoring server in the prediction system 100 outputs the results and predicted values from the collection processing unit 110 and the artificial intelligence unit 130 on the screen, and when an abnormal value is predicted, an alert is issued to the manager. This is the module that generates

And, in the field of deep learning, there is a lib called a framework, and the network structure is constructed using Tensorflow (Google), Keras (Google), MxNet (Microsoft), Pytorch (Facebook), etc.

In addition, for data processing, python libraries such as Pandas and Tableau are used together, and it is preferable to use libraries such as CUDA and cuDNN (Nvidia, USA) for high-speed parallel learning. Also, the OS is based on Ubuntu Linux, but windows10 is also possible. In addition, hardware (HardWare) uses several high-speed parallel processing GPUs (Graphic Processing Units) in the artificial intelligence unit 130 for high-speed parallel learning. Most preferably, chips made by Nvidia in the United States are used, and the desktop CPU has 16 cores. It is preferable to use a high-performance CPU or higher.

An embodiment of the prediction system 100 according to the present invention will be described in detail with reference to FIG. 5 as follows.

Here, ① is sensor measurement values such as Nox emission amount, gas turbine output, denitrification agent amount, heat recovery boiler temperature, various temperatures, and pressure values, and 15 are included as examples in this embodiment.

② is the basic predictor value that is the source of all measured values in ①, and is the (fuel) consumption of LNG as a fuel. That is, various measurement values (attribute values) of ① are generated by the combustion of fuel.

③ is all the measured values except ② at the moment of current time t, the length of the measured values at current time t is 1, and there are 15 sensor (property) values, so they are displayed as (1,15).

④ is a time series of 15 properties that should be predicted in future time. If the sensor measurement time step (set unit time) is 1 minute, and 15 attribute values of 5/5 of the future are to be predicted, the size of the predicted value becomes (5,15).

⑤ is a value known as a value indicating how much fuel you plan to inject in the future. a, b, c,… is 1 minute, 2 minutes, 3 minutes after the current time t… is the amount of fuel to be injected.

On the other hand, since all data of the past time is known based on the current time t, if we consider the past 32 minutes (assuming that the set unit time is 1 minute), ① is (32,15) and ② is (32,1). becomes

The prediction system 100 first predicts data 1 minute after the current time t. Extract a data (size=(1,1)) from ⑤, and it enters the first linear layer (neural network) of ⑦. The size that passes through ⑦ becomes (1,16). We increased the attribute size 1 to 16 using neural network connectivity.

On the other hand, (1,15) data of ③ enters the selector ⑧. The role of ⑧ is to select the signal from ③ and the signal from ⑭. At first, signal ③ is selected, but in the next step, the signal from ⑭ is selected. ⑧ If either (1,15) signal is selected from the inside, it is converted to (1,32) size through the linear layer inside, The two properties of (1, 16) that have been passed through ⑦ are combined to make the size (1,48). It goes through the second linear layer again and becomes (1,15) size and enters the GRU in ⑪.

⑪ is a part including the GRU layer, which is a type of recurrent neural network (RNN) net for time series prediction, and the GRU plays a role of predicting the attribute value of the next time using the current time input.

⑫ GRU has two inputs, the (1,15) size passing through the second linear layer and the (1,512) size passing through the encoder of ⑩.

Here, the size (1,512) is a vector abbreviated by extracting the features of the past time series using the (32,15) component of ① and the (32,1) size, which is the past time component before t time of ②.

⑫ predicts the size of the attribute value (1,64) at the next instant using two vectors of magnitude (1,15) and (1,512), which passes through the 3rd linear layer and the activation layer (ReLU layer) of ⑭ Enter the 4th linear layer again.

The output of ⑭ is of magnitude (1,15), which is fed back to predict the next moment in ⑧. ⑮ is a module that accumulates and stores the currently predicted future value (1,15) as an output.

At time b (t+2) after the output of ⑭ predicted at time a (t+1), which is after time t, which is the current time, is generated, the following process is performed.

In ⑧, instead of (1,15) in ③, select (1,15) of ⑭ and output (1,32) through the linear layer inside ⑧. Then, b, which is the value at the other instant (t+2), is extracted from ⑤ and combined with the data output from ⑧ through ⑦ and combined in ⑨.

Through repetition of this process, after 1 minute (t+1), after 2 minutes (t+2), after 3 minutes (t+3), ... , after n minutes (t+n), it is possible to predict as many times as we want in the future.

That is, using the pre-planned LNG fuel consumption value, the remaining 15 attribute values that users are curious about (this includes the pollution prediction value of interest, Nox emission) can be predicted simply and conveniently.

7 is a graph comparing the sensor data extracted from the LNG gas combined cycle gas turbine and the predicted value in relation to the prediction system 100 of the present invention.

The horizontal axis of FIG. 7 is time unit (minutes), and the vertical axis is normalized to a value between 0 and 1 for the generation amount of each attribute. And, in the graph, the solid line is a graph showing actual measurements generated during actual operation in time series, and the dotted line is a graph showing the deep learning prediction values in time series.

Here, ‘G/T fuel consumption’ is the amount of LNG fuel that goes into the gas turbine. All other attribute (sensor) values can be obtained by burning the fuel. ‘G/T load’ is the output of the gas turbine. The amount of electricity produced is usually in MW. 'HRSG #3' is the temperature value at the end (#3) of the heat recovery boiler (HRSG). Combined power generation is the primary power generation by rotating the G/T, and operating the heat recovery boiler with the remaining high-temperature and high-pressure steam. It is a structure that generates secondary power, and the reason for using this value is that the HRSG supplies the heat required for vaporization of ammonia, a pollution inhibitor.

The ‘NOx outlet’ indicates the amount of NOx, which is a pollutant, that is generated when LNG is combusted, which is nitrogen oxide (NOx).

‘Nox eliminator’ is a substance that is added to suppress NOx emission by the amount of ammonia reducing agent input.

On the other hand, FIG. 6 is a diagram schematically illustrating the contents of monitoring and transmitting the prediction system 100 according to the present invention to the user's mobile by wire or wireless or when an abnormality occurs in the set data. it will be shown

Of course, the prediction system 100 according to the present invention may include all of security, monitoring, mobile, and alarm as shown in FIG. 6 .

Accordingly, according to the present invention, a power plant using AI that can predict in advance various variable values including the amount of NOx, one of pollutants, using an artificial neural network based on past data generated time-series in the process of power generation It is possible to provide a system and method for predicting driving attribute values including an amount of NOx generated.

In addition, it is possible to provide a system and method for predicting operating attribute values including the amount of NOx generated in a power plant using AI that can preemptively take a measure to reduce pollutants including NOx predicted in advance according to the operation plan of the power plant. can

In addition, it is possible to provide a system and method for predicting operating attribute values including the amount of NOx generated in a power plant using AI that can improve the stability of power plant operation and reduce the emission of pollutants for eco-friendly operation.

Here, although several embodiments of the present invention have been illustrated and described, those skilled in the art to which the present invention pertains will appreciate that modifications can be made to the present embodiments without departing from the principles or spirit of the present invention. will be. The scope of the invention will be defined by the appended claims and their equivalents.

100: prediction system 110: collection and storage unit
130: artificial intelligence department

Claims (13)

Raw data including a plurality of historical variable values, including fuel consumption, gas turbine output, NOx generation, denitrification reducing agent input, and exhaust heat recovery boiler temperature, which is the end temperature of the exhaust heat recovery boiler, that are generated in time-series in the process of operating a power plant a collection processing unit that collects and processes (Raw Data);
An artificial intelligence unit for predicting a plurality of predictive variable values corresponding to each of the past variable values for each unit time set after the present time following the past variable values by using the data stored in the collection and processing unit using artificial intelligence;
It passes through the first linear layer by extracting the basic predictor value after the current time,
The method further includes selecting the predictor value that has passed through a Rectified Linear Unit (ReLU) that predicts the predictor value based on the historical variable value, the basic predictive variable value, and the historical variable value of the current time. Driving attribute value prediction system, characterized in that. According to claim 1,
At least one of the predictive variable values includes fuel consumption as a basic predictive variable value that is a pre-predicted value,
The power plant includes a combined cycle power plant using LNG as a fuel,
A driving attribute value prediction system, characterized in that predicting the remaining predictive variable values based on the basic predictive variable values. According to claim 1,
The artificial intelligence unit driving attribute value prediction system, characterized in that it includes a multivariate prediction model in the artificial neural network. Raw data including a plurality of historical variable values, including fuel consumption, gas turbine output, NOx generation, denitrification reducing agent input, and exhaust heat recovery boiler temperature, which is the end temperature of the exhaust heat recovery boiler, that are generated in time-series in the process of operating a power plant (1) process of collecting and processing (Raw Data) in the collection and processing department;
(2) process of predicting, in the artificial intelligence unit, a plurality of predictive variable values corresponding to each of the past variable values for each unit time set after the present time following the past variable value by using artificial intelligence on the data stored in the collection and processing unit class;
(3) extracting the value of the basic predictor after the current time and passing it through the first linear layer;
(4) selecting the predictor value that has passed through a Rectified Linear Unit (ReLU) that predicts the predictor value based on the historical variable value, the basic predictive variable value, and the historical variable value at the present time (4) Process; driving attribute value prediction method, characterized in that it further comprises. delete delete delete 5. The method of claim 4,
In step (4), the value of the past variable of the present time is initially selected, and thereafter, the value of the predictor that has passed the ReLU is selected. 5. The method of claim 4,
The method of predicting a driving attribute value further comprising: (5) step of passing through a second linear layer in which the variable values passing the steps (3) and (4) are merged in the contact layer. 10. The method of claim 9,
The method for predicting driving attribute values, characterized in that the variable values that have passed the step (5) and the variable values that have passed through the encoder pass through a convolutional neural network (CNN) (6). 11. The method of claim 10,
The method of predicting a driving attribute value, characterized in that the variable value passed through the encoder in the step (6) is abbreviated by extracting the past variable value divided at a set time interval from the present time to the past. 11. The method of claim 10,
(6-1) of predicting, in the GRU (Gated Recurrent) layer, variable values excluding the basic predictive variable values using the variable values input in the step (6);
The method of predicting driving attribute value, characterized in that the variable value that has passed the step (6-1) passes through the third linear layer, passes through the ReLU layer, and passes through the fourth linear layer to become the predictive variable value. delete

Images