JP2022134564A - Feature amount extraction method, performance prediction model generation method, plant operation support method, feature amount extraction device, performance prediction model creation device, and plant operation support device - Google Patents

Abstract

To shorten a time required for optimizing operation parameters by improving prediction accuracy, and facilitate automation of extracting one or more feature amounts.SOLUTION: A feature amount extraction method for extracting one or more feature amounts for creating a performance prediction model showing a relation between operation parameters of a fuel-burning plant and process values of the plant, comprises the steps of: obtaining image data of a combustion area of the plant; converting the image data into color image data segmented into a plurality of color regions according to color information by clustering processing; and extracting the one or more feature amounts from the color image data.SELECTED DRAWING: Figure 2

Description

The present disclosure relates to a feature quantity extraction method, a performance prediction model creation method, a plant operation support method, a feature quantity extraction device, a performance prediction model creation device, and a plant operation support device.

Patent Literature 1 discloses a method of creating a performance prediction model that can accurately predict process values even when plant specifications and fuel properties change. The performance prediction model shows the relationship between input parameters (input) and process values (output) of a plant that burns fuel. In this patent document 1, plant operation data and feature amounts extracted from image data of the combustion area of the plant taken from above are treated as input parameters of a performance prediction model.

WO2019/225457

By the way, it is desirable that the feature quantity is information that improves the prediction accuracy of the performance prediction model, but no specific method for extracting the feature quantity from the image data of the combustion region of the plant is disclosed. In addition, depending on the method of extracting the feature amount from the image data of the combustion region of the plant, manual adjustment work such as preprocessing and threshold setting is required, and automation may not be easy.

The present disclosure has been made in view of the above-described problems, and provides a feature quantity extraction method and a feature quantity extraction device capable of improving prediction accuracy and facilitating automation of feature quantity extraction. for the purpose.

In order to solve the above problems, for example, the configurations described in the claims are adopted. The present application includes a plurality of means for solving the above problems, and one example thereof will be given. A feature quantity extraction method according to the present disclosure is a feature quantity extraction method for extracting a feature quantity for creating a performance prediction model showing the relationship between an operating parameter of a plant that burns fuel and a process value of the plant. a step of obtaining image data of a combustion region of the plant; a step of converting the image data into color image data segmented into a plurality of color regions according to color information by a clustering process; and extracting features.

A feature quantity extraction device according to the present disclosure is a feature quantity extraction device that extracts a feature quantity for creating a performance prediction model that indicates the relationship between an operating parameter of a plant that burns fuel and a process value of the plant. an acquisition unit for acquiring image data of the combustion area of the plant; a conversion unit for converting the image data into color image data segmented into a plurality of color areas according to color information by clustering processing; and the color image data. and an extraction unit that extracts the feature amount from.

According to the feature quantity extraction method and feature quantity extraction device of the present disclosure, it is possible to improve prediction accuracy and facilitate automation of feature quantity extraction.

It is a schematic diagram of a boiler plant. 4 is a flow chart showing a feature quantity extraction method according to an embodiment; FIG. 4 is a diagram showing color image data according to one embodiment; FIG. 4 is a diagram showing color image data according to one embodiment; It is a figure which shows the color image data which concerns on another one embodiment. 4 is a flow chart showing a method of creating a performance prediction model according to one embodiment; 8 is a flow chart showing a method of creating a performance prediction model according to another embodiment; It is a flowchart which shows the operation assistance method of the plant which concerns on one Embodiment. 1 is a schematic functional block diagram of a feature extraction device according to an embodiment; FIG. 1 is a schematic functional block diagram of a performance prediction model creation device according to an embodiment; FIG. BRIEF DESCRIPTION OF THE DRAWINGS It is a figure which shows the structure of the operation assistance apparatus of the plant which concerns on one Embodiment.

Hereinafter, the feature quantity extraction method, the performance prediction model creation method, the plant operation support method, the feature quantity extraction device, the performance prediction model creation device, and the plant operation support device according to the embodiments of the present disclosure will be described based on the drawings. do. Such an embodiment shows one aspect of the present disclosure, does not limit the present disclosure, and can be arbitrarily changed within the scope of the technical idea of the present disclosure.

A feature quantity extraction method for extracting feature quantities for creating a performance prediction model showing the relationship between the operating parameters (inputs) of a fuel-burning plant and the process values (outputs) of the plant will be described.

<Feature extraction method>
First, the configuration of a boiler plant 100 is shown as an example of a plant to which the performance prediction model according to the present disclosure is applied. FIG. 1 is a schematic diagram of a boiler plant 100. As shown in FIG. In addition, the plant is not limited to the boiler plant 100, and can be applied to a wide range of plants that produce industrial products or materials. For example, plants that burn fuel include steam supply plants and steelmaking plants in addition to power plants.

A boiler 102 included in the boiler plant 100 is configured to burn solid fuel. The boiler 102 is a coal-fired boiler capable of generating steam from feed water using heat generated by combustion of pulverized coal (solid fuel) obtained by pulverizing coal Fc. The solid fuel fed into the boiler 102 is not limited to pulverized coal, and may be solid fuel other than pulverized coal, such as biomass, as long as it is combustible in the boiler 102 .

As shown in FIG. 1 , the boiler 102 includes a furnace 104 , a combustion device (combustion burner 106 ), a primary air supply 108 , a secondary air supply 115 and a flue 110 .

The furnace 104 has, for example, a hollow shape of a square cylinder and extends along the vertical direction. The furnace 104 burns pulverized coal with air (primary air) supplied from a primary air supply device 108, and a primary combustion region R1 with air (secondary air) newly supplied from a secondary air supply device 115. It is a so-called two-stage combustion type furnace that burns unburned pulverized coal (unburned fuel) flowing from the secondary combustion region R2.

A primary combustion region R<b>1 in the furnace 104 of the boiler plant 100 is imaged by the imaging device 105 . At least one imaging device 105 is provided in the furnace 104 at a position capable of imaging the primary combustion region R1 so that the fuel supply port 109, which will be described later, is also imaged. The imaging device 105 is provided, for example, on the lower side in the vertical direction of the furnace wall that constitutes the furnace 104 . The imaging device 105 is configured to be capable of capturing at least one of moving images and still images, and generates image data of the primary combustion region R1. Such an imaging device 105 is, for example, a camera such as a digital camera, a video camera, or an infrared camera capable of detecting specific wavelengths. The format of the image data may be TIFF or BMP for still images, and non-compressed data such as AVI for moving images. Alternatively, the format of the image data may be compression format data represented by JPEG or PNG for still images, and MPEG or MP4 for moving images.

A combustion device is provided on the lower side of the furnace wall of the furnace 104 and is, for example, a combustion burner 106 . Combustion burner 106 is connected to pulverizer 114 (coal pulverizer/mill) via pulverized coal supply pipe 112 . The pulverizer 114 pulverizes coal Fc transported by a transport system (not shown) to a predetermined fine powder size to generate pulverized coal. The pulverizer 114 supplies a mixture of pulverized coal and air (primary air) supplied from the primary air supply device 108 to the combustion burner 106 via the pulverized coal supply pipe 112 . The combustion burner 106 injects the air-fuel mixture supplied from the pulverizer 114 into the furnace 104 to form a flame Fr and burn the pulverized coal.

A primary air supply device 108 supplies primary air used for combustion of pulverized coal to the furnace 104 . Such a primary air supply device 108 includes a primary air supply pipe 116 and a primary blower 118 provided on the primary air supply pipe 116 . The air generated by the primary blower 118 flows through the primary air supply pipe 116, the pulverizer 114, the pulverized coal supply pipe 112, and the combustion burner 106 in this order, and is supplied to the furnace 104 as primary air. A primary air (air-fuel mixture) injection port (fuel supply port 109) formed in the furnace wall is located below a secondary air supply port 113 formed in the furnace wall.

The secondary air supply device 115 supplies the furnace 104 with secondary air used for burning the pulverized coal (unburned fuel) that has not burned out in the primary combustion region R1. The secondary air supply device 115 includes a secondary air supply pipe 117 and a secondary blower 119 provided on the secondary air supply pipe 117 . The air generated by the secondary blower 119 flows through the secondary air supply pipe 117 and is supplied to the furnace 104 as secondary air. The secondary air supply device 115 may be configured to supply the air generated by the secondary blower 119 to the combustion burner 106 .

The flue 110 communicates with the upper portion of the furnace 104, through which the combustion gas Cg generated by burning the pulverized coal flows. A gas analyzer 111 is provided in the flue 110 to measure the concentration of gas components such as oxygen and carbon monoxide contained in the combustion gas Cg. The installation position of the gas analyzer 111 is not particularly limited, and may be the outlet of the flue 110 or the inlet of the flue 110 . In some embodiments, gas analyzer 111 is provided in furnace 104 . Further, the flue 110 is provided with a heat exchanger 120 for generating steam from the thermal energy of the combustion gas Cg. The installation position of the heat exchanger 120 is not limited to the form illustrated in FIG. Although not shown, in some embodiments, the boiler plant 100 further includes a steam turbine and a power generator, and the steam generated by the heat exchanger 120 rotates the steam turbine to rotate the rotating shaft of the steam turbine. It is configured to generate power by rotationally driving a generator connected to.

In the form illustrated in FIG. 1 , the boiler plant 100 further includes a denitration device 150 , a dust collection device 152 and a stack 154 . The denitrification device 150 is connected to the flue 110 of the boiler 102 and purifies the combustion gas Cg discharged from the flue 110 of the boiler 102 . The dust collector 152 is connected to the denitration device 150 and collects fly ash contained in the combustion gas Cg purified by the denitration device 150 . The chimney 154 is connected to the dust collector 152 and discharges the combustion gas Cg from which the fly ash has been collected by the dust collector 152 to the outside of the boiler plant 100 such as the atmosphere.

(Configuration of feature quantity extraction method)
FIG. 2 is a flow chart showing a feature amount extraction method according to one embodiment. In the feature quantity extraction method illustrated in FIG. 2, the image data of the primary combustion region R1 in the boiler plant 100 that burns pulverized coal is used to predict the concentration (process value) of the gas component contained in the combustion gas Cg. Extract features to create prediction models. Note that the process value is not limited to the concentration of gaseous components. For example, there is also the unburned content concentration in the ash collected by the dust collector 152 .

As shown in FIG. 2, the feature extraction method includes an acquisition step S2, a conversion step S4, and an extraction step S6.

First, in acquisition step S2, the image data of the primary combustion region R1 of the boiler plant 100 is acquired from the imaging device 105 provided in the furnace 104. FIG. Note that the image data may be acquired from the imaging device 105 in real time, or may be acquired from a storage device such as a database by storing previously acquired image data in a storage device.

Next, in a conversion step S4, the image data obtained from the imaging device 105 is converted into color image data segmented into a plurality of color regions according to color information by clustering processing.

“Clustering processing” uses an unsupervised clustering method such as k-means, and converts image data into color image data divided into clusters (plurality of color regions). Note that the number of clusters may be a value automatically determined in the clustering process, or may be a value set in advance.

A description will be given of "dividing into a plurality of color areas according to color information". The color information is, for example, brightness, and each of the plurality of brightness regions A (plurality of color regions) is set by clustering processing so that the brightness does not overlap with each other. Then, the luminance is calculated for each pixel of the image data acquired from the imaging device 105, and the luminance area A including this pixel is determined. Note that the present disclosure does not limit the color information to luminance, and may be color information other than luminance, such as RGB color components and saturation.

3 and 4 are diagrams showing luminance image data 130 (color image data) according to one embodiment. In the form illustrated in FIGS. 3 and 4, the number of clusters is 4, and the brightness image data 130 is divided into four brightness areas A. In FIG. Each of the four luminance areas A is converted to black and white (gray scale) gradation values, and the luminance area A with lower luminance is darker. The brightness region A having the highest brightness among the four brightness regions A includes a region where pulverized coal supplied to the primary combustion region R1 is burned, and such a brightness region A is defined as a high brightness region A1 (A). In other words, the high-brightness area A1 corresponds to the portion of the image data where the flame Fr is imaged. The remaining luminance area A includes an area where the pulverized coal supplied to the primary combustion area R1 is not burned, and such a luminance area A is defined as a low luminance area A2 (A).

Next, in an extraction step S6, a first feature amount is extracted from the luminance image data 130. FIG. In other words, the first feature amount is extracted based on the image data of the primary combustion region R1. The first feature amount is extracted from at least one of the high brightness area A1 and the low brightness area A2. The first feature amount is, for example, the area of the high brightness area A1, the area of the low brightness area A2, the ignition distance X1, the width X2 of the flame, the length X3 of the flame, the ejection angle θ of the flame, and the like.

In some embodiments, the first feature quantity includes the area of the high brightness area A1. The area of the high-luminance area A1 is the total number of pixels divided into the high-luminance area A1. The area of the high luminance region A1 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the area of the high-brightness region A1 in the first feature amount.

In some embodiments, the first feature amount includes the amount of change in the area of the high-brightness area A1. In the present disclosure, "variation" refers to spatial and/or temporal variation. The same applies to the "amount of change" described below. The amount of change in the area of the high-brightness region A1 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the area of the high-brightness region A1 in the first feature amount.

In some embodiments, the first feature quantity includes the area of the low luminance area A2. The area of the low-luminance area A2 is the total number of pixels divided into the low-luminance area A2. The area of the low luminance region A2 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the area of the low luminance region A2 in the first feature amount.

In some embodiments, the first feature quantity includes the amount of change in the area of the low luminance area A2. The amount of change in the area of the low luminance region A2 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the area of the low-luminance region A2 in the first feature amount.

A specific example of extracting the ignition distance X1 from the luminance image data 130 will be described with reference to FIG. A specific example of extracting the width X2 of the flame Fr, the length X3 of the flame Fr, and the ejection angle θ of the flame Fr from the brightness image data 130 will be described with reference to FIG. As shown in FIGS. 3 and 4, the position of the fuel supply port 109 in the brightness image data 130 is calculated in advance. 3 and 4, fuel is injected downward (inside the furnace 104) from the fuel supply port 109. As shown in FIG.

3 and 4, the direction crossing the opening of the fuel supply port 109 is defined as the length direction D1, and the direction orthogonal to the length direction D1 is defined as the width direction D2. A virtual straight line passing through the center O of the fuel supply port 109 in the width direction D2 along the length direction D1 is defined as a first virtual line L1 (indicated by a dashed line in FIGS. 3 and 4).

In some embodiments, the first feature includes firing distance X1. As shown in FIG. 3, for example, the ignition distance X1 is the distance between the fuel supply port 109 for supplying pulverized coal to the furnace 104 and the boundary line L3 between the high luminance area A1 and the low luminance area A2. be. More specifically, the ignition distance X1 is the distance between the center O and the position P1 where the first imaginary line L1 and the boundary line L3 intersect. The ignition distance X1 has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, by including the ignition distance X1 in the first feature amount, it is possible to improve the prediction accuracy of the performance prediction model. In another embodiment, the ignition distance X1 may be extracted from a first imaginary line shifted from the center O in the width direction D2. In another embodiment, the ignition distance X1 is the ignition distance extracted from the first virtual line L1 and the ignition extracted from the first virtual line shifted in the width direction D2 from the center O of the first virtual line L1. It may be an average value with the distance.

In some embodiments, the first feature quantity includes the amount of change in firing distance X1. The amount of change in the ignition distance X1 has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the ignition distance X1 in the first feature amount.

In some embodiments, the first feature includes width X2 of flame Fr. As shown in FIG. 4, for example, the width X2 of the flame Fr is the size in the width direction of the high-brightness area A1. More specifically, it is the average value of the size in the width direction of the high luminance area A1. The width X2 of the flame Fr has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the width X2 of the flame Fr in the first feature amount. The width X2 of the flame Fr is not limited to the average value of the widthwise size of the high-luminance area A1, and may be the maximum value of the widthwise size of the high-luminance area A1.

In some embodiments, the first feature includes a variation in width X2 of flame Fr. The amount of change in the width X2 of the flame Fr has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the width X2 of the flame Fr in the first feature amount.

In some embodiments, the first feature includes the length X3 of the flame Fr. As shown in FIG. 4, for example, the length X3 of the flame Fr is the length direction D1 of the high brightness area A1. More specifically, it is the average value of the distance from one end to the other end in the length direction D1 of the high-brightness area A1. The length X3 of the flame Fr has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, by including the length X3 of the flame Fr in the first feature amount, it is possible to improve the prediction accuracy of the performance prediction model. The length X3 of the flame Fr may be the maximum value of the distance from the fuel supply port 109 to the other end in the length direction D1 of the high luminance area A1.

In some embodiments, the first feature includes a variation in the length X3 of the flame Fr. The amount of change in the length X3 of the flame Fr has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the length X3 of the flame Fr in the first feature amount.

In some embodiments, the first feature includes an ejection angle θ of the flame Fr. As shown in FIG. 4, for example, the ejection angle θ of the flame Fr is defined by the first virtual line L1 and the second virtual line L2. The second imaginary line L2 is an imaginary straight line passing through the center O and the position P2, assuming that the farthest tip from the fuel supply port 109 of the high brightness area A1 is the position P2 in the length direction D1. The center O corresponds to an intersection where the fuel supply port 109 and the first imaginary line L1 intersect. The ejection angle θ of the flame Fr has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, by including the ejection angle θ of the flame Fr in the first feature amount, it is possible to improve the prediction accuracy of the performance prediction model.

In some embodiments, the first feature quantity includes a change in the ejection angle θ of the flame Fr. The amount of change in the ejection angle θ of the flame Fr has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the ejection angle θ of the flame Fr in the first feature amount.

(Action and effect of feature quantity extraction method)
According to the embodiment illustrated in FIG. 2, the image data in which the primary combustion region R1 of the boiler plant 100 is imaged is converted into luminance image data 130 divided into a plurality of luminance regions A according to the luminance by clustering processing. A first feature amount is extracted from the luminance image data 130 . Therefore, it is possible to extract various feature quantities (for example, ignition distance, etc.), and improve the prediction accuracy of the performance prediction model. Therefore, by improving the prediction accuracy, the time required for optimizing the operating parameters can be shortened.

According to the form illustrated in FIG. 2, since clustering processing is applied when image data is converted into luminance image data 130, there is no need to prepare teacher data. In addition, the clustering process does not require manual adjustment work such as preprocessing and threshold setting. Therefore, it is possible to facilitate the automation of feature quantity extraction.

According to the form illustrated in FIG. 2, the image data of the primary combustion area R1 of the boiler plant 100 is divided into a high luminance area A1 and a low luminance area A2 according to luminance by clustering processing. Then, the first feature amount is extracted from at least one of the high brightness area A1 and the low brightness area A2. Therefore, since the first feature amount includes information on the primary combustion region R1, it is possible to improve the prediction accuracy of the performance prediction model.

In the form illustrated in FIG. 2, the image data of the primary combustion region R1 is utilized to extract the feature quantity for creating a performance prediction model for predicting the concentration of gaseous components contained in the combustion gas Cg. The present disclosure is not limited to this form. In some embodiments, in addition to the image data of the primary combustion region R1, the image data of the secondary combustion region R2 is also utilized to create a performance prediction model that predicts the concentration (process value) of gaseous components contained in the combustion gas Cg. Extract features to create.

<Feature Extraction Method According to Another Embodiment>
A feature amount extraction method according to another embodiment different from the form illustrated in FIG. 2 will be described. Another embodiment differs from the form illustrated in FIG. 2 in that the second feature amount is further extracted based on the image data of the secondary combustion region R2, but otherwise the form illustrated in FIG. is the same as In the description of another embodiment, only the parts that differ from the form illustrated in FIG. 2 will be described.

In the feature amount extraction method according to another embodiment, image data of the secondary combustion region R2 of the boiler plant 100 is acquired in addition to image data of the primary combustion region R1 of the boiler plant 100 in the acquisition step S2. The image data of the secondary combustion region R2 may be captured by the imaging device 105, or may be captured by an imaging device other than the imaging device 105. In the form illustrated in FIG. 1, the secondary combustion region R2 is imaged by an imaging device 122 provided in the upper portion of the furnace 104. In the embodiment shown in FIG. This imaging device 122 is located above the secondary combustion area. The image data of the secondary combustion region R2 is captured at the same time as the image data of the primary combustion region R1.

Next, in the feature amount extraction method according to another embodiment, in the conversion step S4, the image data of the secondary combustion region R2 is clustered into a plurality of brightness regions B according to the brightness. Convert to data 140 . The second luminance image data 140 is divided into a plurality of luminance regions B according to luminance different from the luminance with which the luminance image data 130 is divided into a plurality of luminance regions A. FIG. In some embodiments, the second luminance image data 140 is partitioned into a plurality of luminance regions B according to the same luminance that partitioned the luminance image data 130 into a plurality of luminance regions A.

FIG. 5 is a diagram showing second luminance image data 140 (color image data) according to another embodiment. In the form illustrated in FIG. 5, the number of clusters is 4, and the second luminance image data 140 is divided into four luminance regions B. In FIG. Each of the four luminance regions B is converted to a black and white (gray scale) grayscale value, and the lower the luminance region B, the darker it becomes. The brightness region B, which has the highest brightness among the four brightness regions B, includes a region where unburned pulverized coal flowing from the primary combustion region R1 of the boiler plant 100 to the secondary combustion region R2 is burned. Let B be a second high-brightness region B1(B). The remaining brightness region B includes a region in which unburned pulverized coal that has flowed from the primary combustion region R1 to the secondary combustion region R2 of the boiler plant 100 is not burned, and such a brightness region B is referred to as a second low brightness region. B2(B).

Next, in the feature amount extraction method according to another embodiment, the second feature amount is extracted from the second luminance image data 140 in extraction step S6. In other words, the second feature amount is extracted based on the image data of the secondary combustion region R2. The second feature amount is extracted from at least one of the second high brightness area B1 and the second low brightness area B2. The second feature amount is, for example, the area of the second high-luminance region B1 or the area of the second low-luminance region B2.

In some embodiments, the second feature quantity includes the area of the second high-brightness region B1. The area of the second high-brightness area B1 is the total number of pixels of the second brightness image data 140 divided into the second high-brightness areas B1. The area of the second high-brightness region B1 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg and the concentration of unburned matter in the ash. Therefore, the prediction accuracy of the performance prediction model can be improved by including the second high-brightness region B1 in the second feature amount.

In some embodiments, the second feature quantity includes the amount of change in the area of the second high-brightness region B1. The amount of change in the area of the second high-brightness region B1 has a high correlation with the concentration of gas components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the area of the second high-brightness region B1 in the second feature amount.

In some embodiments, the second feature quantity includes the area of the second low-brightness region B2. The area of the second low-luminance region B2 is the total number of pixels of the second luminance image data 140 divided into the second low-luminance regions B2. The area of the second low luminance region B2 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the area of the second low-luminance region B2 in the second feature amount.

In some embodiments, the second feature quantity includes the amount of change in the area of the second low-luminance region B2. The amount of change in the area of the second low-brightness region B2 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the area of the second low-luminance region B2 in the second feature amount.

In some embodiments, the second feature quantity includes the positional relationship between the second high-brightness region B1 and the second low-brightness region B2. The positional relationship between the second high-brightness region B1 and the second low-brightness region B2 is, for example, the position of the center of gravity derived from the shape of the second high-brightness region B1 and the position of the center of gravity derived from the shape of the second low-brightness region B2. is a vector quantity calculated from The positional relationship between the second high-luminance region B1 and the second low-luminance region B2 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the positional relationship between the second high-brightness region B1 and the second low-brightness region B2 in the second feature amount.

In some embodiments, the second feature quantity includes the amount of change in the positional relationship between the second high-brightness area B1 and the second low-brightness area B2. The amount of change in the positional relationship between the second high-luminance region B1 and the second low-luminance region B2 has a high correlation with the concentration of gaseous components contained in the combustion gas Cg. Therefore, the prediction accuracy of the performance prediction model can be improved by including the amount of change in the positional relationship between the second high-brightness region B1 and the second low-brightness region B2 in the second feature amount.

Actions and effects of a feature amount extraction method according to another embodiment will be described. The second feature quantity extracted from the second luminance image data 140 obtained by converting the image data of the secondary combustion region R2 of the boiler plant 100 may have a high correlation with the concentration of gaseous components contained in the combustion gas Cg.

According to another embodiment, the image data of the secondary combustion region R2 of the boiler plant 100 is divided into a second high-brightness region B1 and a second low-brightness region B2 by clustering processing according to brightness. It is converted into luminance image data 140 . Then, the second feature amount is extracted from at least one of the second high-brightness region B1 and the second low-brightness region B2. Therefore, since the second feature amount includes information on the secondary combustion region R2, it is possible to improve the prediction accuracy of the performance prediction model.

<Performance prediction model creation method>
A performance prediction model creation method for creating a performance prediction model using the feature quantity extracted by the feature quantity extraction method described above will be described.

FIG. 6 is a flow chart showing a performance prediction model creation method according to one embodiment. In the performance prediction model creation method illustrated in FIG. 6, a performance prediction model for predicting combustion performance such as the concentration of gas components contained in the combustion gas Cg and the concentration of unburned ash in the ash is created from the operating parameters of the boiler plant 100. The operating parameters are, for example, the amount of air supplied to the furnace 104, the amount of pulverized coal supplied to the furnace 104, and the like. The operating parameters are measured in real time by a measuring device (not shown). The gas component concentration contained in the combustion gas Cg, which is a process value, is measured in real time by the gas analyzer 111 (see FIG. 1).

(Configuration of performance prediction model creation method)
As shown in FIG. 6, the performance prediction model creation method includes a feature quantity extraction step S12, an analysis feature quantity calculation step S14, an optimum physical model parameter set determination step S16, and a performance prediction model creation step S18.

The feature amount extraction step S12 includes the acquisition step S2, conversion step S4, and extraction step S6 described above. In this feature amount extraction step S12, the feature amount (first feature amount) is extracted from the image data of the primary combustion region R1 of the boiler plant 100 captured at the first timing by the method described above. In addition to or instead of the first feature amount extracted from the image data of the primary combustion area R1, the feature amount is a second feature extracted based on the image data of the secondary combustion area R2. quantity may be included.

Next, in the analysis feature quantity calculation step S14, based on the operating parameters and the physical model parameter set at the first timing of the boiler plant 100, the physical model parameter set is changed to perform sensitivity analysis, and at the first timing An analysis feature quantity corresponding to the feature quantity is calculated. In other words, the analysis feature amount is calculated by repeating the sensitivity analysis in which the physical model parameter set is changed without changing the operating parameter. The sensitivity analysis in the analysis feature amount calculation step S14 is an analysis for obtaining the sensitivity of the feature amount to the physical model parameter set. For example, a response surface model is created and the analysis feature amount is calculated from this response surface model. A physical model (combustion analysis model, etc.) including a chemical reaction model, a turbulence model, etc. is required to perform the sensitivity analysis. A physical model parameter is a parameter set in this physical model.

Next, in the optimum physical model parameter set determination step S16, the optimum physical model parameter set is determined based on the difference between the feature amount and the analysis feature amount at the first timing. For example, the optimum physical model parameter set is determined such that the difference between the feature quantity and the analysis feature quantity at the first timing is minimized.

Next, in the performance prediction model creation step S18, based on the optimum physical model parameter set and the operating parameters of the boiler plant 100, sensitivity analysis is performed by changing the operating parameters to determine the relationship between the operating parameters and the process values. Create a performance prediction model to show. The sensitivity analysis in the performance prediction model creation step S18 is an analysis for obtaining the sensitivity of process values to operating parameters, and for example, creates a response surface model (performance prediction model).

(Action and effect of performance prediction model creation method)
According to such a performance prediction model creation method, the optimum physical model parameter set is determined based on the difference between the feature amount and the analysis feature amount at the first timing (for example, so that the difference is minimized). Then, a performance prediction model showing the relationship between the operating parameters and the process values is created by changing the operating parameters and performing sensitivity analysis based on the optimum physical model parameter set and the plant operating parameters. For this reason, the performance prediction model according to the present disclosure is created as a model (so-called physical model) that analyzes the relationship between the operating parameter and the process value, and is a model that shows the relationship between the actually obtainable feature amount and the process value ( The prediction accuracy of the performance prediction model can be improved compared to the so-called statistical model).

A method of creating a performance prediction model according to another embodiment will be described. A performance prediction model creation method according to another embodiment creates a performance prediction model using the feature quantity extracted by the feature quantity extraction method described above. FIG. 7 is a flow chart showing a performance prediction model creation method according to another embodiment.

As shown in FIG. 7, the performance prediction model creation method according to another embodiment includes the feature quantity extraction step S12, the operation parameter acquisition step S30, the process value acquisition step S32, and the second performance prediction model. and a creation step S34. Since the feature quantity extraction step S12 has already been described, the description thereof will be omitted.

In the operating parameter acquisition step S30, the operating parameters of the boiler plant 100 at a plurality of timings are acquired from the measuring device. In the process value acquisition step S32, process values of the boiler plant 100 at a plurality of timings are acquired from the gas analyzer 111, respectively. In addition, in the form illustrated in FIG. 7, the feature amount extraction step S12, the operating parameter acquisition step S30, and the process value acquisition step S32 are executed in this order, but the present disclosure is not limited to this form. For example, the feature quantity extraction step S12, the operating parameter acquisition step S30, and the process value acquisition step S32 may be performed simultaneously.

In the second performance prediction model creation step S34, based on learning data including each feature amount at a plurality of timings, each operating parameter at a plurality of timings, and each process value at a plurality of timings, the operating parameters and process Create a performance prediction model that shows the relationship between Specifically, regression analysis is performed on multiple pieces of learning data (feature values + operating parameters + process values) with different timings, and a regression formula (performance prediction model) for calculating process values from operating parameters is created. create.

According to the performance prediction model creation method according to another embodiment illustrated in FIG. 7, the performance prediction model is created based on learning data including each feature amount, operating parameter, and process at a plurality of timings. (since learning data at multiple timings are machine-learned), the prediction accuracy of the performance prediction model can be improved.

<Plant operation support method>
An operation support method for a plant (boiler plant 100) using a performance prediction model created by the performance prediction model creation method illustrated in FIG. 6 will be described. In some embodiments, the plant operation support method uses a performance prediction model created by the performance prediction model creation method illustrated in FIG.

FIG. 8 is a flowchart showing a plant operation support method according to one embodiment. As shown in FIG. 8, the plant operation support method includes an operation parameter setting step S22, a process value output step S24, a judgment step S26, and a resetting step S28.

First, in the operating parameter setting step S22, the operating parameters measured in real time by the measuring device are set (obtained). In other words, it prepares the operating parameters to be input to the performance prediction model.

Next, in the process value output step S24, the operating parameters set in the operating parameter setting step S22 are input to the performance prediction model, and the concentrations of gas components (process values of the boiler plant 100) are output.

Next, in determination step S26, it is determined whether or not the concentration of the gas component output in process value output step S24 is included in a predetermined condition. The predetermined condition is, for example, the exhaust gas standard of the boiler plant 100 . The predetermined condition may further include an index related to reliability such as the metal temperature of the heat exchanger, or another index related to efficiency such as heat absorption amount. If the concentration of the gas component is included in the predetermined condition (S26: Yes), the plant operation support method according to the present disclosure ends. If the concentration of the gas component is not included in the predetermined condition (S26: No), the process proceeds to resetting step S28.

In resetting step S28, the operating parameters are reset. Then, in the process value output step S24, the operating parameters reset in the reset step S28 are input to the performance prediction model, and the concentrations of the gas components are output again. The resetting step S28 is repeated until the concentration of the gas component output by the performance prediction model is included in the predetermined condition.

According to such a plant operation support method, since the performance prediction model according to the present disclosure is used, it is possible to quickly calculate an operation parameter in which the concentration of the gaseous component is included in the predetermined condition. Further, according to such a plant operation support method, before actually adjusting the operation parameters of the boiler plant 100, the operation parameters are repeatedly calculated so that the concentration of the gas component is included in the predetermined conditions. Therefore, the time for actually determining the optimum operating parameters of the boiler plant 100 can be shortened.

<Feature value extraction device>
The feature quantity extraction device 1 according to the present disclosure is for creating a performance prediction model that shows the relationship between the feature quantity and the process value (gas component concentration) of a plant (boiler plant 100) that burns fuel (pulverized coal). Extract features.

The feature extraction device 1 is a computer such as an electronic control device, and includes a processor such as a CPU or GPU, a memory such as a ROM or a RAM, and an I/O interface (not shown). The feature quantity extraction device 1 implements each functional unit included in the feature quantity extraction device 1 by the processor operating (computing, etc.) according to the instructions of the program loaded in the memory. Each functional unit of the feature quantity extraction device 1 according to one embodiment will be described with reference to FIG. 9 .

FIG. 9 is a schematic functional block diagram of the feature quantity extraction device 1 according to one embodiment. As shown in FIG. 9 , the feature quantity extraction device 1 includes an acquisition unit 2 , a conversion unit 4 and an extraction unit 6 .

The acquisition unit 2 acquires image data of the primary combustion region R1 of the boiler plant 100 from the imaging device 105 provided in the furnace 104 .

The conversion unit 4 converts the image data acquired by the acquisition unit 2 from the imaging device 105 into luminance image data 130 divided into a plurality of luminance regions A according to luminance by clustering processing.

The extraction unit 6 converts the brightness image data 130 converted by the conversion unit 4 into first feature amounts (area of high brightness region A1, area of low brightness region A2, ignition distance X1, width of flame X2, length of flame X3, flame ejection angle θ, etc.).

According to the feature quantity extraction device 1, the conversion unit 4 divides the image data of the primary combustion region R1 of the boiler plant 100 into a plurality of luminance regions A according to the luminance by clustering processing. Convert to data 130 . The extraction unit 6 extracts the first feature amount from the brightness image data 130 . Therefore, it is possible to extract various feature quantities (for example, ignition distance, etc.), and improve the prediction accuracy of the performance prediction model.

Further, according to the feature quantity extraction device 1, since clustering processing is applied when image data is converted into luminance image data 130, there is no need to prepare teacher data. In addition, the clustering process does not require manual adjustment work such as preprocessing and threshold setting. Therefore, it is possible to facilitate the automation of feature quantity extraction.

<Performance prediction model creation device>
The performance prediction model creation device 50 according to the present disclosure uses the performance prediction model created by the feature quantity extraction device 1 described above to use the operating parameters and processes of the plant (boiler plant 100) that burns solid fuel (pulverized coal). Create a performance prediction model that shows the relationship with values (concentrations of gaseous components).

The performance prediction model creation device 50 is a computer such as an electronic control device, and includes a processor such as a CPU or GPU, a memory such as a ROM or a RAM, and an I/O interface (not shown). The performance prediction model generation device 50 realizes each functional unit included in the performance prediction model generation device 50 by the processor operating (computing, etc.) according to the instructions of the program loaded in the memory. Each functional unit of the performance prediction model creation device 50 according to one embodiment will be described with reference to FIG. 10 .

FIG. 10 is a schematic functional block diagram of the performance prediction model creation device 50 according to one embodiment. As shown in FIG. 10, the performance prediction model creation device 50 includes the feature quantity extraction device 1 described above, an analysis feature quantity calculation unit 54, an optimum physical model parameter set determination unit 56, a performance prediction model creation unit 58, Prepare. Note that the performance prediction model creation device 50 can acquire each of the operating parameters (measured values of the measuring device) and the process values (measured values of the gas analyzer 111).

The feature quantity extraction device 1 extracts the feature quantity (first feature quantity) from the image data of the primary combustion region R1 of the boiler plant 100 captured at the first timing by the method described above. Incidentally, in the form illustrated in FIG. 10 , the performance prediction model creation device 50 includes the feature quantity extraction device 1 . In other words, the performance prediction model creation device 50 includes the functional units of the feature quantity extraction device 1 . In another embodiment, the performance prediction model creation device 50 is provided separately from the feature quantity extraction device 1 and acquires the feature quantity from the feature quantity extraction device 1 .

The analysis feature quantity calculation unit 54 performs sensitivity analysis by changing the physical model parameter set based on the operating parameters and the physical model parameter set at the first timing of the boiler plant 100, and calculates the feature quantity at the first timing. Calculate the corresponding analysis feature amount.

The optimum physical model parameter set determination unit 56 determines the optimum physical model parameter set based on the difference between the feature amount and the analysis feature amount at the first timing.

Based on the optimal physical model parameter set and the operating parameters of the plant, the performance prediction model creation unit 58 performs sensitivity analysis while changing the operating parameters to create a performance prediction model showing the relationship between the operating parameters and the process values. create.

According to such a performance prediction model creation device 50, the optimum physical model parameter set is determined based on the difference between the feature amount and the analysis feature amount at the first timing (for example, so that the difference is minimized). Then, a performance prediction model showing the relationship between the operating parameters and the process values is created by changing the operating parameters and performing sensitivity analysis based on the optimum physical model parameter set and the plant operating parameters. For this reason, the performance prediction model according to the present disclosure is created as a model (so-called physical model) that analyzes the relationship between the operating parameter and the process value, and is a model that shows the relationship between the actually obtainable feature amount and the process value ( Compared to the so-called statistical model), it is possible to make highly accurate predictions even under extrapolation conditions when plant operating conditions and fuel properties change.

Although not shown, the performance prediction model creation device 50 may further include a database that stores feature quantities, operating parameters, and process values. According to such a configuration, by accessing the database 60, it is possible to acquire each of the feature quantity, the operating parameter, and the process value. According to such a configuration, the operator of the boiler plant 100 obtains the feature values, the operating parameters, and the process values stored in the database 60, thereby obtaining the feature values for suitably operating the boiler plant 100. I can grasp it. For example, it is possible to suppress the judgment of the operator that the combustion state of the boiler plant 100 is better when the ignition distance X1 is shorter.

<Plant operation support device>
The plant (boiler plant 100) operation support device according to the present disclosure uses the performance prediction model created by the performance prediction model creation device 50 to support the operation of the plant.

A plant operation support device (hereinafter referred to as an operation support device 70) is a computer such as an electronic control device, and includes a processor such as a CPU or GPU, a memory such as a ROM or a RAM, and an I/O interface (not shown). The driving assistance device 70 implements each functional unit included in the driving assistance device 70 by the processor operating (computing, etc.) according to the instructions of the program loaded in the memory. Each functional unit of the driving support device 70 according to one embodiment will be described with reference to FIG. 11 .

FIG. 11 is a diagram showing the configuration of a plant operation support device 70 according to one embodiment. As shown in FIG. 11 , the driving support device 70 includes a driving parameter setting section 72 , a process value output section 74 , a determination section 76 and a resetting section 78 . The driving support device 70 can acquire the driving parameters (measured values of the measuring device) and the process values (measured values of the gas analyzer 111).

In the form illustrated in FIG. 11 , the driving support device 70 is provided separately from the performance prediction model creation device 50 and acquires the performance prediction model from the performance prediction model creation device 50 . In some embodiments, the driving support device 70 includes the performance prediction model creation device 50 described above.

The operating parameter setting unit 72 sets (acquires) operating parameters measured by the measuring device in real time. In other words, it prepares the operating parameters to be input to the performance prediction model.

The process value output unit 74 inputs the operating parameters set by the operating parameter setting unit 72 to the performance prediction model, and outputs the concentrations of gaseous components (process values).

The determination unit 76 determines whether the concentration of the gas component output by the process value output unit 74 is included in the predetermined condition.

When the determination unit 76 determines that the concentration of the gas component output by the process value output unit 74 is not included in the predetermined condition, the reset unit 78 resets the operating parameter. Then, the process value output unit 74 inputs the operating parameters reset by the resetting unit 78 to the performance prediction model, and outputs the concentration of the gas component again. The resetting unit 78 resets the operating parameters until the determination unit 76 determines that the concentration of the gas component output by the process value output unit 74 is included in the predetermined condition.

According to the driving support device 70, since the performance prediction model created by the performance prediction model creation device 50 is used, it is possible to quickly set the operating parameters in which the concentration of the gas component is included in the predetermined condition. .

The contents described in each of the above embodiments are understood as follows, for example.

[1] The feature quantity extraction method according to the present disclosure includes:
A feature quantity extraction method for extracting a feature quantity for creating a performance prediction model showing the relationship between an operating parameter of a fuel-burning plant (100) and a process value of the plant,
a step (S2) of acquiring image data of the combustion areas (R1, R2) of the plant;
a step (S4) of converting the image data into color image data (130, 140) divided into a plurality of color areas (A, B) according to color information by clustering processing;
and a step (S6) of extracting the feature amount from the color image data.

According to the method described in [1] above, image data is converted into color image data segmented into a plurality of color regions according to color information by clustering processing, and feature amounts are extracted from the color image data. Various feature quantities (for example, ignition distance, which will be described later) are extracted. Therefore, it is possible to improve the prediction accuracy of the performance prediction model.

According to the method described in [1] above, since clustering processing is applied when image data is converted into color image data, there is no need to prepare teacher data. In addition, the clustering process does not require manual adjustment work such as preprocessing and threshold setting. Therefore, it is possible to facilitate the automation of feature quantity extraction.

[2] In some embodiments, in the method described in [1] above,
The image data includes image data of a primary combustion zone (R1) of the plant;
The plurality of color regions include a first color region (A1) including a region where the fuel supplied to the primary combustion region burns, and a region where the fuel supplied to the primary combustion region is unburned. a second color area (A2) included,
The feature quantity includes a first feature quantity extracted from at least one of the first color region and the second color region.

The feature quantity extracted from the color image data obtained by converting the image data of the primary combustion area of the plant has a high correlation with the process value. According to the method described in [2] above, the image data of the primary combustion area of the plant is converted into color image data divided into the first color area and the second color area according to the color information by the clustering process. be. Then, the first feature amount is extracted from at least one of the first color area and the second color area. Therefore, the first feature amount includes information on the primary combustion region, and the prediction accuracy of the performance prediction model can be improved.

[3] In some embodiments, in the method described in [2] above,
The first feature amount is the distance between a fuel supply port (109) for supplying the fuel to the plant and a boundary line (L3) between the first color area and the second color area. Includes ignition distance (X1).

The ignition distance has a high correlation with the process value. According to the method described in [3] above, since the first feature amount includes the ignition distance, it is possible to improve the prediction accuracy of the performance prediction model.

[4] In some embodiments, in the method described in [2] or [3] above,
The first feature amount is the amount of change in the ignition distance, which is the distance between the fuel supply port for supplying the fuel to the plant and the boundary line between the first color area and the second color area. include.

The amount of change in ignition distance has a high correlation with the process value. According to the method described in [4] above, since the first feature amount includes the amount of change in the ignition distance, it is possible to improve the prediction accuracy of the performance prediction model.

[5] In some embodiments, in the method of any one of [2] to [4] above,
The first feature quantity includes at least one of the area of the first color region and the area of the second color region.

Each of the area of the first color region and the area of the second color region has a high correlation with the process value. According to the method described in [5] above, since the first feature amount includes at least one of the area of the first color region and the area of the second color region, the prediction accuracy of the performance prediction model can be improved. .

[6] In some embodiments, in the method of any one of [2] to [5] above,
The first feature amount includes at least one of an area change amount of the first color region and an area change amount of the second color region.

Each of the amount of change in the area of the first color region and the amount of change in the area of the second color region has a high correlation with the process value. According to the method described in [6] above, since the first feature amount includes at least one of the amount of change in the area of the first color region and the amount of change in the area of the second color region, the prediction accuracy of the performance prediction model can be improved.

[7] In some embodiments, in the method of any one of [2] to [6] above,
The first feature amount includes at least one of a flame width (X2), a flame length (X3), and a flame ejection angle (θ) generated by combustion of the fuel,
Assuming that the direction crossing the opening of the fuel supply port for supplying the fuel to the plant is the length direction (D1) and the direction perpendicular to the length direction is the width direction (D2),
the width of the flame is the size in the width direction of the first color region;
The length of the flame is the size of the first color region in the longitudinal direction,
The flame ejection angle is defined by a first imaginary line (L1) passing through the fuel supply port in the longitudinal direction, an intersection (P1) where the fuel supply port and the first imaginary line intersect, and the second and a second imaginary line (L2) passing through the end of the one-color area on the side opposite to the fuel supply port side in the length direction.

Each of the flame width, flame length, and flame ejection angle also has a high correlation with the process value. According to the method described in [7] above, the first feature amount includes at least one of the width of the flame, the length of the flame, and the ejection angle of the flame, so that the prediction accuracy of the performance prediction model is improved. can be done.

[8] In some embodiments, in the method of any one of [2] to [7] above,
The first feature amount includes at least one of a variation in the width of the flame generated by burning the fuel, a variation in the length of the flame, and a variation in the ejection angle of the flame,
Assuming that the direction intersecting with the opening of the fuel supply port for supplying the fuel to the plant is the length direction, and the direction perpendicular to the length direction is the width direction,
the width of the flame is the size in the width direction of the first color region;
The length of the flame is the size of the first color region in the longitudinal direction,
The flame ejection angle is defined by a first virtual line passing through the fuel supply port in the length direction, an intersection point where the fuel supply port and the first virtual line intersect, and the length of the first color region. and a second imaginary line passing through the tip on the side opposite to the fuel supply port side in the vertical direction.

Each of the amount of change in flame width, the amount of change in flame length, and the amount of change in flame ejection angle also has a high correlation with the process value. According to the method described in [8] above, the first feature amount includes at least one of the amount of change in the width of the flame, the amount of change in the length of the flame, and the amount of change in the ejection angle of the flame. The prediction accuracy of the prediction model can be improved.

[9] In some embodiments, in the method of any one of [1] to [8] above,
The image data includes image data of a secondary combustion area of the plant,
The plurality of color regions include a third color region (B1) including a region where unburned fuel that has flowed from the primary combustion region of the plant to the secondary combustion region burns, and a fourth color region (B2) including a region where the unburned fuel that has flowed to the secondary combustion region is unburned;
The feature amount includes a second feature amount extracted from at least one of the third color area and the fourth color area.

The feature quantity extracted from the color image data obtained by converting the image data of the secondary combustion area of the plant has a high correlation with the process value. According to the method described in [9] above, the image data of the secondary combustion area of the plant is converted into color image data divided into the third color area and the fourth color area according to the color information by the clustering process. be done. Then, the second feature amount is extracted from at least one of the third color area and the fourth color area. Therefore, the second feature amount includes information on the secondary combustion region, and the prediction accuracy of the performance prediction model can be improved.

[10] In some embodiments, in the method of [9] above,
The second feature quantity includes the area of the third color region.

The area of the third color region has a high correlation with the process value. According to the method described in [10] above, since the second feature amount includes the area of the third color region, it is possible to improve the prediction accuracy of the performance prediction model.

[11] In some embodiments, in the method of [9] or [10] above,
The second feature amount includes the amount of change in the area of the third color area.

The amount of change in the area of the third color region has a high correlation with the process value. According to the method described in [11] above, since the second feature amount includes the amount of change in the area of the third color region, it is possible to improve the prediction accuracy of the performance prediction model.

[12] In some embodiments, in the method of any one of [9] to [11] above,
The second feature amount includes a positional relationship between the third color area and the fourth color area.

The positional relationship between the third color area and the fourth color area has a high correlation with the process value. According to the method described in [12] above, since the second feature amount includes the positional relationship between the third color area and the fourth color area, it is possible to improve the prediction accuracy of the performance prediction model.

[13] In some embodiments, in the method of any one of [9] to [12] above,
The second feature amount includes the amount of change in the positional relationship between the third color area and the fourth color area.

The amount of change in the positional relationship between the third color area and the fourth color area has a high correlation with the process value. According to the method described in [13] above, since the second feature amount includes the amount of change in the positional relationship between the third color area and the fourth color area, it is possible to improve the prediction accuracy of the performance prediction model.

[14] The performance prediction model creation method according to the present disclosure includes:
A performance prediction model creation method for creating a performance prediction model showing the relationship between the operating parameters of a plant that burns fuel and the process values of the plant,
a step (S12) of extracting the feature quantity at a first timing using the feature quantity extraction method according to any one of [1] to [13] above;
Based on the operating parameter and the physical model parameter set at the first timing of the plant, sensitivity analysis is performed by changing the physical model parameter set, and an analysis feature corresponding to the feature amount at the first timing. a step of calculating the amount (S14);
a step of determining an optimum physical model parameter set based on the difference between the feature amount and the analysis feature amount at the first timing (S16);
Based on the optimal physical model parameter set and the operating parameters of the plant, sensitivity analysis is performed by varying the operating parameters to create the performance prediction model showing the relationship between the operating parameters and the process values. and a step (S18).

According to the method described in [14] above, the optimum physical model parameter set is determined based on the difference between the feature amount and the analysis feature amount at the first timing. Then, a performance prediction model showing the relationship between the operating parameters and the process values is created by changing the operating parameters and performing sensitivity analysis based on the optimum physical model parameter set and the plant operating parameters. For this reason, the performance prediction model according to the present disclosure is created as a model (so-called physical model) that analyzes the relationship between the operating parameter and the process value, and is a model that shows the relationship between the actually obtainable feature amount and the process value ( The prediction accuracy of the performance prediction model can be improved compared to the so-called statistical model).

[15] The performance prediction model creation method according to the present disclosure includes:
A performance prediction model creation method for creating a performance prediction model showing the relationship between the operating parameters of a plant that burns fuel and the process values of the plant,
a step (S12) of extracting the feature amounts at a plurality of timings using the feature amount extraction method according to any one of [1] to [13];
a step (S30) of obtaining the operating parameters of the plant at the plurality of timings;
obtaining the process values of the plant at the plurality of timings (S32);
Based on learning data including each of the feature values at the plurality of timings, each of the operation parameters at the plurality of timings, and each of the process values at the plurality of timings, the operation parameter and the process value and a step (S34) of creating the performance prediction model showing the relationship.

According to the method described in [15] above, the performance prediction model is created based on learning data including each feature quantity, operating parameter, and process at a plurality of timings (learning data at a plurality of timings machine learning), it is possible to improve the prediction accuracy of the performance prediction model.

[16] The plant operation support method according to the present disclosure includes:
A plant operation support method using the performance prediction model created by the performance prediction model creation method according to [14] or [15] above,
a step of setting the operating parameters (S22);
a step of inputting the operating parameters set in the step of setting the operating parameters into the performance prediction model and outputting the process values (S24);
a step of determining whether the process value output in the step of outputting the process value is included in a predetermined condition (S26);
and a step (S28) of resetting the operating parameter when it is determined that the process value is not included in the predetermined condition.

According to the method described in [16] above, since the performance prediction model created by the performance prediction model creation method described in [14] or [15] is used, the process value is included in the predetermined conditions. Parameters can be calculated quickly.

[17] The feature quantity extraction device (1) according to the present disclosure is
A feature quantity extraction device for extracting a feature quantity for creating a performance prediction model showing the relationship between an operating parameter of a plant that burns fuel and a process value of the plant,
an acquisition unit (2) for acquiring image data of the combustion area of the plant;
a conversion unit (4) for converting the image data into color image data segmented into a plurality of color regions according to color information by clustering processing;
an extraction unit (6) for extracting the feature quantity from the color image data;

According to the configuration described in [17] above, image data is converted into color image data segmented into a plurality of color regions according to color information by clustering processing, and feature amounts are extracted from the color image data. Various feature quantities (for example, ignition distance, which will be described later) can be extracted, and the prediction accuracy of the performance prediction model can be improved.

According to the configuration described in [17] above, since clustering processing is applied when image data is converted into color image data, there is no need to prepare teacher data. In addition, the clustering process does not require manual adjustment work such as preprocessing and threshold setting. Therefore, it is possible to facilitate the automation of feature quantity extraction.

[18] The performance prediction model creation device (50) according to the present disclosure is
A performance prediction model creation device for creating a performance prediction model showing the relationship between the operating parameters of a plant that burns fuel and the process values of the plant,
The feature quantity extraction device according to [17] above, wherein the feature quantity extraction device extracts the feature quantity at a first timing;
Based on the operating parameter and the physical model parameter set at the first timing of the plant, sensitivity analysis is performed by changing the physical model parameter set, and an analysis feature corresponding to the feature amount at the first timing. an analysis feature quantity calculation unit (54) for calculating the quantity;
an optimum physical model parameter set determination unit (56) for determining an optimum physical model parameter set based on the difference between the feature amount and the analysis feature amount at the first timing;
Based on the optimal physical model parameter set and the operating parameters of the plant, sensitivity analysis is performed while varying the operating parameters to create the performance prediction model showing the relationship between the operating parameters and the process values. and a performance prediction model creation unit (58).

According to the configuration described in [18] above, the optimum physical model parameter set is determined based on the difference between the feature amount and the analysis feature amount at the first timing. Then, a performance prediction model showing the relationship between the operating parameters and the process values is created by changing the operating parameters and performing sensitivity analysis based on the optimum physical model parameter set and the plant operating parameters. For this reason, the performance prediction model according to the present disclosure is created as a model (so-called physical model) that analyzes the relationship between the operating parameter and the process value, and is a model that shows the relationship between the actually obtainable feature amount and the process value ( The prediction accuracy of the performance prediction model can be improved compared to the so-called statistical model).

[19] The plant operation support device (70) according to the present disclosure includes:
A plant operation support device using the performance prediction model created by the performance prediction model creation device according to [18] above,
an operating parameter setting unit (72) for setting the operating parameters;
a process value output unit (74) that inputs the operating parameters set by the operating parameter setting unit to the performance prediction model and outputs the process values;
a determination unit (76) that determines whether the process value output by the process value output unit is included in a predetermined condition;
a resetting unit (78) for resetting the operating parameter when the determination unit determines that the process value output by the process value output unit is not included in the predetermined condition.

According to the configuration described in [19] above, since the performance prediction model created by the performance prediction model creation device described in [18] above is used, the process value quickly determines the operating parameter included in the predetermined condition. can be set.

1 feature quantity extraction device 2 acquisition unit 4 conversion unit 6 extraction unit 50 performance prediction model creation device 54 analysis feature quantity calculation unit 56 optimum physical model parameter set determination unit 58 performance prediction model creation unit 70 driving support device 72 driving parameter setting unit 74 Process value output unit 76 Judgment unit 78 Reset unit 100 Boiler plant 109 Fuel supply port 130 Luminance image data (color image data)
140 second luminance image data (color image data)
A luminance region (luminance region segmented from image data of primary combustion region)
A1 high brightness area (first color area)
A2 Low luminance area (second color area)
B Luminance region (luminance region segmented from image data of secondary combustion region)
B1 Second high luminance area (third color area)
B2 Second low luminance area (fourth color area)
Cg Combustion gas D1 Length direction D2 Width direction Fr Flame L1 First virtual line L2 Second virtual line L3 Boundary line P1 Position (intersection of first virtual line L1 and boundary line L3)
R1 Primary combustion region R2 Secondary combustion region S2 Acquisition step S4 Conversion step S6 Extraction step S12 Feature value extraction step S14 Analysis feature value calculation step S16 Optimal physical model parameter set determination step S18 Performance prediction model creation step S22 Operation parameter setting step S24 Process Value output step S26 Judgment step S28 Reset step S30 Operation parameter acquisition step S32 Process value acquisition step S34 Second performance prediction model creation step X1 Ignition distance X2 Flame width X3 Flame length θ Flame ejection angle

Claims (19)

A feature quantity extraction method for extracting a feature quantity for creating a performance prediction model showing the relationship between an operating parameter of a plant that burns fuel and a process value of the plant,
obtaining image data of a combustion area of the plant;
a step of converting the image data into color image data segmented into a plurality of color regions according to color information by a clustering process;
extracting the feature quantity from the color image data;
Feature extraction method. The image data includes image data of a primary combustion area of the plant;
The plurality of color regions include a first color region including a region where the fuel supplied to the primary combustion region burns and a second color region including a region where the fuel supplied to the primary combustion region is unburned. a two-color region;
The feature amount includes a first feature amount extracted from at least one of the first color region and the second color region,
The feature amount extraction method according to claim 1. The first feature amount includes an ignition distance, which is the distance between a fuel supply port for supplying the fuel to the plant and a boundary line between the first color area and the second color area,
The feature amount extraction method according to claim 2. The first feature amount is the amount of change in the ignition distance, which is the distance between the fuel supply port for supplying the fuel to the plant and the boundary line between the first color area and the second color area. include,
The feature extraction method according to claim 2 or 3. The first feature amount includes at least one of the area of the first color region and the area of the second color region,
The feature quantity extraction method according to any one of claims 2 to 4. The first feature amount includes at least one of an amount of change in the area of the first color region and an amount of change in the area of the second color region,
The feature quantity extraction method according to any one of claims 2 to 5. The first feature amount includes at least one of a width of a flame generated by combustion of the fuel, a length of the flame, and an ejection angle of the flame,
Assuming that the direction intersecting with the opening of the fuel supply port for supplying the fuel to the plant is the length direction, and the direction perpendicular to the length direction is the width direction,
the width of the flame is the size in the width direction of the first color region;
The length of the flame is the size of the first color region in the longitudinal direction,
The flame ejection angle is defined by a first virtual line passing through the fuel supply port in the length direction, an intersection point where the fuel supply port and the first virtual line intersect, and the length of the first color region. and a second virtual line passing through the tip on the side opposite to the fuel supply port side in the vertical direction,
The feature quantity extraction method according to any one of claims 2 to 6. The first feature amount includes at least one of a variation in the width of the flame generated by burning the fuel, a variation in the length of the flame, and a variation in the ejection angle of the flame,
Assuming that the direction intersecting with the opening of the fuel supply port for supplying the fuel to the plant is the length direction, and the direction perpendicular to the length direction is the width direction,
the width of the flame is the size in the width direction of the first color region;
The length of the flame is the size of the first color region in the longitudinal direction,
The flame ejection angle is defined by a first virtual line passing through the fuel supply port in the length direction, an intersection point where the fuel supply port and the first virtual line intersect, and the length of the first color region. and a second virtual line passing through the tip on the side opposite to the fuel supply port side in the vertical direction,
The feature quantity extraction method according to any one of claims 2 to 7. The image data includes image data of a secondary combustion area of the plant,
The plurality of color regions include a third color region including a region where unburned fuel that has flowed from the primary combustion region of the plant to the secondary combustion region is burned, and a fourth color region including a region where the unburned fuel that has flowed into the region is unburned;
The feature quantity includes a second feature quantity extracted from at least one of the third color region and the fourth color region,
The feature quantity extraction method according to any one of claims 1 to 8. The second feature amount includes the area of the third color region,
The feature quantity extraction method according to claim 9 . The second feature amount includes an amount of change in the area of the third color region,
The feature extraction method according to claim 9 or 10. The second feature amount includes a positional relationship between the third color area and the fourth color area,
The feature quantity extraction method according to any one of claims 9 to 11. The second feature amount includes an amount of change in positional relationship between the third color area and the fourth color area,
The feature quantity extraction method according to any one of claims 9 to 12. A performance prediction model creation method for creating a performance prediction model showing the relationship between the operating parameters of a plant that burns fuel and the process values of the plant,
A step of extracting the feature quantity at a first timing using the feature quantity extraction method according to any one of claims 1 to 13;
Based on the operating parameter and the physical model parameter set at the first timing of the plant, sensitivity analysis is performed by changing the physical model parameter set, and an analysis feature corresponding to the feature amount at the first timing. calculating the amount;
determining an optimum physical model parameter set based on the difference between the feature amount and the analysis feature amount at the first timing;
Based on the optimal physical model parameter set and the operating parameters of the plant, sensitivity analysis is performed while varying the operating parameters to create the performance prediction model showing the relationship between the operating parameters and the process values. and
Performance prediction model creation method. A performance prediction model creation method for creating a performance prediction model showing the relationship between the operating parameters of a plant that burns fuel and the process values of the plant,
A step of extracting the feature amount at a plurality of timings using the feature amount extraction method according to any one of claims 1 to 13;
obtaining each of the operating parameters of the plant at the plurality of timings;
obtaining each of the process values at the plurality of timings of the plant;
Based on learning data including each of the feature values at the plurality of timings, each of the operation parameters at the plurality of timings, and each of the process values at the plurality of timings, the operation parameter and the process value creating the performance prediction model showing the relationship;
Performance prediction model creation method. A plant operation support method using the performance prediction model created by the performance prediction model creation method according to claim 14 or 15,
setting the operating parameters;
a step of inputting the operating parameter set in the step of setting the operating parameter into the performance prediction model and outputting the process value;
a step of determining whether the process value output in the step of outputting the process value is included in a predetermined condition;
and resetting the operating parameter when it is determined that the process value is not included in the predetermined condition. A feature quantity extraction device for extracting a feature quantity for creating a performance prediction model showing the relationship between an operating parameter of a plant that burns fuel and a process value of the plant,
an acquisition unit that acquires image data of the combustion area of the plant;
a conversion unit that converts the image data into color image data segmented into a plurality of color regions according to color information by clustering processing;
an extraction unit that extracts the feature amount from the color image data;
Feature extractor. A performance prediction model creation device for creating a performance prediction model showing the relationship between the operating parameters of a plant that burns fuel and the process values of the plant,
18. The feature quantity extraction device according to claim 17, wherein the feature quantity extraction device extracts the feature quantity at a first timing;
Based on the operating parameter and the physical model parameter set at the first timing of the plant, sensitivity analysis is performed by changing the physical model parameter set, and an analysis feature corresponding to the feature amount at the first timing. an analysis feature quantity calculation unit that calculates the quantity;
an optimum physical model parameter set determination unit that determines an optimum physical model parameter set based on the difference between the feature amount and the analysis feature amount at the first timing;
Based on the optimal physical model parameter set and the operating parameters of the plant, sensitivity analysis is performed while varying the operating parameters to create the performance prediction model showing the relationship between the operating parameters and the process values. and a performance prediction model creation unit for
Performance prediction model creation device. A plant operation support device that uses the performance prediction model created by the performance prediction model creation device according to claim 18,
an operating parameter setting unit configured to set the operating parameters;
a process value output unit that inputs the operating parameter set by the operating parameter setting unit to the performance prediction model and outputs the process value;
a determination unit that determines whether the process value output by the process value output unit is included in a predetermined condition;
a resetting unit that resets the operating parameter when the determination unit determines that the process value output by the process value output unit is not included in the predetermined condition;
Plant operation support equipment.

Images