JP7250205B1 - Learning model generation method, learning model generation device, combustion prediction method, combustion prediction device, and program - Google Patents

Abstract

Kind Code: A1 To improve the prediction accuracy of an objective variable related to combustion in a furnace. A computer obtains, as teaching data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information, and the computer obtains the teaching data. Provided is a learning model generation method for generating a learning model in which the amount of input and state information are input and the predicted value of the objective variable is output. [Selection drawing] Fig. 1

Description

The present invention relates to a learning model generation method, a learning model generation device, a combustion prediction method, a combustion prediction device, and a program.

The operation and maintenance of industrial combustion furnaces has long been the responsibility of field workers. However, individual adjustment by workers causes problems in skill inheritance in addition to variations in accuracy. In addition, adjustment by workers is basically subjective. Therefore, it is not known whether the current adjustment is the optimum solution.

JP 2019-183698 A JP 2017-089928 A JP 2019-190802 A

The conditions inside an industrial combustion furnace change from moment to moment. For example, the amount and type of waste (hereinafter referred to as "state variables") charged into the furnace may fluctuate depending on the emission source. In addition, the relationship between the fuel gas and air (hereinafter referred to as "control variables") and the state variables is complex, and the gas discharged from the combustion furnace as a result of combustion (hereinafter referred to as "exhaust gas"). It is difficult to accurately predict the concentration of , the amount of flame radiation, the surface temperature of the object to be heated, etc. (hereinafter referred to as "objective variables").

SUMMARY OF THE INVENTION It is an object of the present invention to improve the accuracy of prediction of target variables related to furnace combustion.

In the invention according to claim 1, a computer acquires, as teaching data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information. A learning model generation method in which a computer generates a learning model using the training data, inputting the input amount and the state information, and outputting the predicted value of the objective variable, wherein the training data is used for the furnace the period during which only the fuel is burned in the furnace, the period during which only the waste is burned in the furnace, and the period during which the fuel and the waste are burned together in the furnace, and the computer obtains by the said period a learning model generation method for generating the learning model for each period using the teacher data obtained by the learning model.
In the invention according to claim 2, the learning model corresponding to a period in which only the waste is burned or a period in which the fuel and the waste are burned together is acquired for each difference in the composition of the waste. 2. The method of generating a learning model according to claim 1, wherein the learning model is generated for each difference in composition using the teacher data that has been obtained.
The invention according to claim 3 is characterized in that the waste is at least one of waste liquid from other equipment, waste oil from other equipment, process exhaust gas from other equipment, and incinerated matter. 2, the learning model generating method.
In the invention according to claim 4, the computer acquires the input amount of fuel as the control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information as teacher data. A learning model generation method in which a computer generates a learning model using the teacher data, the input amount and the state information, and the predicted value of the objective variable as an output, wherein the teacher data is acquired and the The method of generating the learning model reflects at least one of the number of nozzles used for charging the waste into the furnace and the difference in the position of the nozzle used for charging.
The invention according to claim 5 is an acquisition unit that acquires, as teacher data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information. , a learning model generation device having a learning unit that generates a learning model that uses the teacher data, inputs the input amount and the state information, and outputs a predicted value of the objective variable, wherein the acquisition unit includes the Acquiring teacher data for each period in which only the fuel is burned in the furnace, a period in which only waste is burned in the furnace, and a period in which the fuel and the waste are burned together in the furnace, The learning model generation device, wherein the learning unit generates the learning model for each period using the teacher data acquired for each period.
The invention according to claim 6 is an acquisition unit that acquires, as teacher data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information. a learning unit that generates a learning model that uses the teacher data, inputs the amount of input and the state information, and outputs a predicted value of the objective variable, wherein the acquisition of the teacher data; The learning model generating device reflects at least one of the number of nozzles used for charging the waste into the furnace and the difference in the position of the nozzle used for charging, in the generation of the learning model.
In the invention according to claim 7, the computer acquires the input amount of fuel as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information as teaching data. and a function of generating a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable, wherein the function of acquiring However, the training data is acquired for each period during which only the fuel is burned in the furnace, only waste is burned in the furnace, and the fuel and waste are burned together in the furnace. and the function of generating the learning model is a program for generating the learning model for each period using teacher data acquired for each period.
According to the eighth aspect of the invention, the computer acquires, as teaching data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information. and a function of generating a learning model using the teacher data, inputting the input amount and the state information, and outputting the predicted value of the objective variable, wherein Acquisition and generation of the learning model is a program that reflects at least one of the number of nozzles used to inject waste into the furnace and/or differences in nozzle positions used in injecting.
According to the ninth aspect of the invention, the computer learns the relationship between the amount of fuel input as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information. A combustion prediction method for inputting a measured value of the input amount and a measured value of the state information to a model, and outputting a predicted value of the corresponding objective variable, wherein the learning model is the fuel only in the furnace. is burned, the period when only the waste is burned in the furnace, and the period when the fuel and the waste are burned together in the furnace. A combustion prediction method that uses a different learning model.
In the invention according to claim 10, the computer learns the relationship between the input amount of fuel as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information. A combustion prediction method for inputting a measured value of the input amount and a measured value of the state information to a model, and outputting a predicted value of the corresponding objective variable, wherein the learning model is obtained by obtaining teacher data and obtaining the The combustion prediction method uses a learning model that reflects at least one of the number of nozzles used for charging waste into the furnace and the difference in the position of the nozzle used for charging when generating the learning model.
The invention according to claim 11 is a learning model that learns the relationship between the input amount of fuel as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information, a prediction unit that, when the input amount and the state information are given as inputs, provides the measured value of the input amount and the measured value of the state information to the learning model, and outputs the predicted value of the objective variable. The combustion prediction device, wherein the learning model includes a period in which only the fuel is burned in the furnace, a period in which only waste is burned in the furnace, and a period in which the fuel and the waste are burned together in the furnace. This is a combustion prediction device that uses a learning model for each period generated using teacher data acquired for each period.
The invention according to claim 12 is a learning model that learns the relationship between the input amount of fuel as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information, a prediction unit that, when the input amount and the state information are given as inputs, provides the measured value of the input amount and the measured value of the state information to the learning model, and outputs the predicted value of the objective variable. The combustion prediction device, wherein at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for inputting waste into the furnace is used as the learning model when acquiring teacher data and generating the learning model. A combustion predictor that uses a reflected learning model.
According to a thirteenth aspect of the invention, the computer learns the relationship between the amount of fuel input as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information. A program for realizing a function of inputting a measured value of the input amount and a measured value of the state information to a model and outputting a predicted value of the objective variable, wherein the learning model is the furnace. Generated using training data obtained by the period in which only fuel is burned, the period in which only waste is burned in the furnace, and the period in which the fuel and the waste are burned together in the furnace It is a program that uses a learning model for each period.
In the invention according to claim 14, the computer learns the relationship between the amount of fuel input as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information. A program for realizing a function of outputting a predicted value of the objective variable by giving a measured value of the input amount and a measured value of the state information as inputs to the model, and acquiring teacher data as the learning model. and a program that uses a learning model that reflects at least one of the number of nozzles used to charge the waste into the furnace and the position of the nozzle used for charging when generating the learning model.
In the fifteenth aspect of the present invention, the computer measures an input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and an objective variable corresponding to the input amount and the state information. A learning model generation method in which values are acquired as teaching data, and a computer generates a learning model using the teaching data, inputting the input amount and the state information, and outputting the predicted value of the objective variable. .
According to a sixteenth aspect of the invention, the state information includes the input amount of waste containing the causative substance of the specific gas discharged from the furnace, the concentration of the causative substance in the waste, and the viscosity of the waste. The learning model generation method according to claim 15 , comprising at least one of
According to a seventeenth aspect of the invention , the teaching data includes the input amount of the fuel gas , the input amount of the waste, and the measured value measured after a predetermined time from the input of the fuel gas and the waste. 16. The learning model generation method according to claim 15 , which is a combination of
The invention according to claim 18 is the learning model generation according to claim 15 , wherein the computer corrects the learning model using an error between the predicted value predicted using the learning model and the actual measured value. The method.
The invention according to claim 19 , wherein the state information includes at least one of process exhaust gas, waste oil, waste liquid, material to be incinerated, oxygen concentration in the furnace, temperature in the furnace, and temperature of the furnace wall . is the learning model generation method described in .
In the invention according to claim 20 , the objective variable is CO, CO ₂ , NO _x , SO _x , dust, dioxin, unburned fuel gas, greenhouse gas, amount of flame radiation, surface temperature of the object to be heated. 16. The learning model generation method according to claim 15 , comprising at least one of
In the invention according to claim 21 , the input amount of the fuel gas used for combustion of the object as the control variable, the state information of the furnace, and the measurement value of the objective variable corresponding to the input amount and the state information. A learning model generation device, comprising: an acquisition unit for acquiring training data; and a learning unit for generating a learning model using the training data, inputting the input amount and the state information, and outputting a predicted value of the objective variable. be.
In the invention according to claim 22 , the computer stores, as control variables, an input amount of fuel gas used for combustion of an object , state information of the furnace, and measurement of objective variables corresponding to the input amount and the state information. values as teacher data, and a function of generating a learning model using the teacher data, inputting the input amount and the state information, and outputting the predicted value of the objective variable. It's a program.
In the twenty-third aspect of the present invention, the computer measures an input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and objective variables corresponding to the input amount and the state information. This is a combustion prediction method in which the measured value of the input amount and the measured value of the state information are given as inputs to a learning model that has learned the relationship with the value, and the predicted value of the corresponding objective variable is output.
According to the twenty-fourth aspect of the present invention, the amount of fuel gas used for combustion of an object as a control variable, state information of the furnace, and the measured value of the objective variable corresponding to the amount of input and the state information. When a learning model that has learned the relationship and the input amount and the state information are given as inputs, the measured value of the input amount and the measured value of the state information are given to the learning model, and the predicted value of the objective variable is calculated. and a prediction unit that outputs a combustion prediction device.
According to the twenty-fifth aspect of the invention, the computer stores, as control variables, an input amount of fuel gas used for combustion of an object , furnace state information, and measurement of objective variables corresponding to the input amount and the state information. It is a program for realizing the function of outputting the predicted value of the objective variable by giving the measured value of the input amount and the measured value of the state information as inputs to the learning model that has learned the relationship with the value.

According to the first aspect of the invention, by generating a learning model for each combustion environment, it is possible to further improve the prediction accuracy of the objective variable.
According to the second aspect of the invention, the prediction accuracy of the objective variable can be further improved by generating a learning model for each waste composition.
According to the third aspect of the invention, the prediction accuracy of the target variable can be further improved by generating a learning model for each waste composition.
According to the fourth aspect of the invention, the prediction accuracy of the objective variable can be further improved by generating the learning model including the number of nozzles and the difference in the positions of the nozzles used for injection.
According to the fifth aspect of the invention, by generating a learning model for each combustion environment, it is possible to further improve the prediction accuracy of the objective variable.
According to the sixth aspect of the invention, by generating a learning model including the number of nozzles and the difference in the positions of the nozzles used for injection, it is possible to further improve the prediction accuracy of the objective variable.
According to the seventh aspect of the invention, by generating a learning model for each combustion environment, it is possible to further improve the prediction accuracy of the objective variable.
According to the eighth aspect of the invention, by generating a learning model including the number of nozzles and the difference in the positions of the nozzles used for injection, it is possible to further improve the prediction accuracy of the objective variable.
According to the ninth aspect of the invention, by generating a learning model for each combustion environment, it is possible to further improve the prediction accuracy of the objective variable.
According to the tenth aspect of the invention, the prediction accuracy of the objective variable can be further improved by generating the learning model including the number of nozzles and the difference in the positions of the nozzles used for injection.
According to the eleventh aspect of the invention, by generating a learning model for each combustion environment, it is possible to further improve the prediction accuracy of the objective variable.
According to the twelfth aspect of the invention, by generating a learning model including the number of nozzles and the difference in the positions of the nozzles used for injection, it is possible to further improve the prediction accuracy of the objective variable.
According to the thirteenth aspect of the invention, by generating a learning model for each combustion environment, it is possible to further improve the prediction accuracy of the objective variable.
According to the fourteenth aspect of the invention, by generating a learning model including the number of nozzles and the difference in the positions of the nozzles used for injection, it is possible to further improve the prediction accuracy of the objective variable.
According to the fifteenth aspect of the invention, it is possible to improve the prediction accuracy of the target variable relating to the combustion of the furnace.
According to the sixteenth aspect of the invention, it is possible to improve the prediction accuracy of the specific gas discharged from the furnace.
According to the seventeenth aspect of the invention , it is possible to further improve the prediction accuracy of the objective variable.
According to the eighteenth aspect of the invention, the accuracy of the learning model can be efficiently improved.
According to the nineteenth aspect of the invention, it is possible to improve the prediction accuracy of the objective variable relating to the combustion of the furnace.
According to the twentieth aspect of the invention, it is possible to improve the prediction accuracy of the objective variable relating to the combustion of the furnace.
According to the twenty- first aspect of the invention, it is possible to improve the prediction accuracy of the objective variable relating to the combustion of the furnace.
According to the twenty-second aspect of the invention, it is possible to improve the prediction accuracy of the objective variable relating to the combustion of the furnace.
According to the twenty-third aspect of the present invention, it is possible to improve the prediction accuracy of the objective variable relating to combustion in the furnace.
According to the twenty-fourth aspect of the invention, it is possible to improve the prediction accuracy of the objective variable relating to the combustion of the furnace.
According to the twenty-fifth aspect of the invention, it is possible to improve the prediction accuracy of the objective variable relating to the combustion of the furnace.

1 is a diagram illustrating a conceptual configuration of a combustion furnace system assumed in Embodiment 1; FIG. It is a figure explaining the structural example of a prediction model production|generation apparatus. 4 is a diagram illustrating an example of variables used for generating a prediction model in Embodiment 1; FIG. 4 is a flowchart for explaining a procedure for generating a prediction model that uses the concentration of exhaust gas as a prediction value; FIG. 4 is a diagram for explaining a function of predicting the concentration of exhaust gas using a prediction model; FIG. 4 is a diagram illustrating an example of variables used for generating a prediction model in generation processing 1; FIG. 10 is a diagram illustrating another example of variables used for generating a prediction model in generation processing 2; 2 is a diagram illustrating combustion patterns assumed in one combustion furnace; FIG. 13 is a diagram for explaining prediction model generation processing according to Embodiment 3. FIG. FIG. 4 is a diagram for explaining a function of classifying teacher data; FIG. 10 is a diagram illustrating a learning function of a dedicated prediction model corresponding to each pattern using classified teacher data; FIG. 11 is a diagram for explaining prediction processing of predicted values in Embodiment 3; FIG. 4 is a diagram illustrating a case where one nozzle is provided for each combustible material on the ceiling of the combustion furnace; FIG. 4 is a diagram illustrating a case where two nozzles are provided for each combustible material on the ceiling of the combustion furnace; FIG. 10 is a diagram illustrating another example in which a total of two nozzles are provided for each combustible material distributed in the ceiling portion and the wall portion of the combustion furnace; FIG. 12 is a diagram for explaining prediction model generation processing according to Embodiment 4. FIG. FIG. 4 is a diagram for explaining a function of classifying teacher data; FIG. 10 is a diagram illustrating a learning function of a dedicated prediction model corresponding to each pattern using classified teacher data; FIG. 12 is a diagram for explaining prediction processing of predicted values in Embodiment 4; FIG. FIG. 20 is a diagram for explaining prediction model generation processing according to Embodiment 5. FIG. 4 is a chart for explaining an example in which measured values such as an input amount of combustible material input at a reference time (current time) and measured values of concentration of exhaust gas after a certain period of time from the reference time are used as teacher data. FIG. 4 is a chart for explaining an example in which a measured value of exhaust gas concentration at a reference time (current time) and a measured value of an input amount of combustible material, etc., input a certain time before the reference time are used as teacher data. It is a figure explaining the correction method of the prediction model by a gradient boosting decision tree. FIG. 13 is a flowchart for explaining a procedure for generating a prediction model according to Embodiment 6. FIG. FIG. 10 is a diagram illustrating a method of correcting a learning model using teacher data generated by interpolation; It is a figure explaining the generation method of teacher data. (A) shows a method of increasing the number of teacher data samples using linear interpolation, and (B) shows a method of increasing the number of teacher data samples using a polynomial or regression model. FIG. 12 is a diagram for explaining a conceptual configuration of a combustion furnace system assumed in an eighth embodiment;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings.
<Embodiment 1>
<System configuration>
In this embodiment, a combustion furnace for burning waste will be described. In particular, a combustion furnace in which the input amount of waste and the combination and ratio of the types of waste to be input change from moment to moment according to the convenience of the emission source will be described.

In this type of combustion furnace, the state inside the furnace changes from moment to moment. For this reason, it is much more difficult to predict the exhaust gas and the like discharged from the combustion furnace than in the case of the combustion furnace in which the amount of waste input is controlled to be constant. Therefore, in the present embodiment, a mechanism for generating a prediction model for accurately predicting the concentration of waste generated by combustion at each time point by machine learning will be described.

FIG. 1 is a diagram illustrating a conceptual configuration of a combustion furnace system 1 assumed in Embodiment 1. FIG.
In the case of FIG. 1, the outline indicated by the dashed line represents the outer edge of the building or site where the combustion furnace 10 and the predictive model generation device 20 are installed. Although FIG. 1 shows only one combustion furnace 10, it is not intended to limit the number of combustion furnaces 10 in the building or site to one.
In the combustion furnace 10 shown in FIG. 1, waste (e.g., process exhaust gas, waste oil, waste liquid) and fuel gas (city gas, air) are introduced as materials to be burned, and exhaust gas is discharged as a result of the combustion reaction. be.

City gas and air are fuels for burning waste. City gas is, for example, natural gas whose main component is methane. Instead of city gas, for example, liquefied petroleum gas (so-called LPG) containing propane or butane as a main component may be introduced. City gas and LPG are representative examples of hydrocarbon fuels. In addition, ammonia gas, a mixed gas of ammonia and city gas, a mixed gas of ammonia and hydrogen, or the like may be fed into the combustion furnace 10 as the fuel gas.
The city gas flow rate Q _N [Nm ³ /h] and the air flow rate Q _A [Nm ³ /h] are examples of control variables.

Waste refers to substances that are by-products of manufacturing processes and chemical reaction processes and that are subject to disposal. Waste may be discharged as a gas, as a liquid, or as a solid.
A representative example of waste as a gas is process exhaust gas. Typical examples of liquid waste include waste oil and waste liquid.

Waste oil refers to a liquid containing oil.
Wastewater is the liquid recovered from each process. In the present embodiment, liquid that does not contain oil is referred to as waste liquid. The waste liquid includes contaminated water after being used for cooling and cleaning, as well as organic liquids and the like. It also includes mixtures of liquids discharged from multiple processes separately.
In this embodiment, the liquid combusted in the combustion furnace 10 may be generically called "waste liquid". Liquid waste in this sense also includes oil-containing liquids.
Solid waste includes dust, shavings, scraps, and the like.

In addition, the flow rate of process exhaust gas Qp [Nm ³ /h], the flow rate of waste oil _QH [Nm ³ /h], the flow rate of waste liquid _QD [Nm ³ /h], the oxygen concentration in the furnace X _O [%], the furnace The internal temperature T [°C] and the furnace wall temperature T [°C] are examples of variables representing the state of the furnace (that is, “state variables”). The state variable here is an example of "state information" in the claims.
Incidentally, the weight of waste as a state variable is represented by [ton/h].
Each flow rate is measured in real time by a flow meter, oxygen concentration by an oxygen concentration meter, temperature by a temperature sensor, and weight by a weight scale.

Exhaust gas varies depending on the substance burned in the combustion furnace 10, but includes, for example, CO, _CO2 , NOx, _SOx , dust, _dioxin , unburned fuel gas, greenhouse gas, hydrocarbon gas, and the like. In the present embodiment, a specific gas component contained in exhaust gas is referred to as "specific gas".
In addition, the information representing the results of the combustion reaction includes the amount of flame radiation [W/m ² ] and the surface temperature T [°C] of the object to be heated. For example, the concentration of exhaust gas, the amount of flame radiation, and the surface temperature of the object to be heated are examples of objective variables. In addition, the objective variable includes the heating efficiency [%].

Concentration is measured in real time by a densitometer, flame radiation and heating efficiency are calculated, and surface temperature is measured by a temperature sensor. The heating efficiency is calculated using, for example, the amount of heat input to the combustion furnace 10 and the surface temperature of the object to be heated.
However, as far as the combustion furnace 10 assumed in the present embodiment is concerned, there is no object to be heated. Therefore, the surface temperature of the object to be heated is measured when the combustion furnace 10 is used as a heating furnace or a melting furnace. Objects to be heated include, for example, glass, iron, aluminum oxide, and ash.

<Device configuration>
FIG. 2 is a diagram illustrating a configuration example of the prediction model generation device 20. As shown in FIG.
The prediction model generation device 20 includes a processor 201 that controls the operation of the entire device, a ROM 202 that stores BIOS and the like, a RAM 203 that is used as a work area for the processor 201, an auxiliary storage device 204 that records various data, It has a communication interface 205 that communicates with sensors installed in various parts of the combustion furnace 10 (see FIG. 1). Note that the processor 201 and other processing units are interconnected through a bus or other signal line 206 .

The processor 201, ROM 202, and RAM 203 function as a so-called computer. That is, the predictive model generation device 20 is configured by, for example, a server, a desktop computer, or a notebook computer.
The processor 201 implements various functions through execution of programs.

In this embodiment, as one of the programs, each measured value of the control variable, the state information, and the objective variable is used as teacher data, and the objective variable is predicted when the control variable and the status information are given as inputs. Consider a program that generates a predictive model that outputs values.
As one of the other programs, a program is assumed in which measured values of control variables and state information acquired from sensors provided in each part of the combustion furnace 10 are input to a prediction model, and predicted values of objective variables are output. .

The auxiliary storage device 204 is, for example, a hard disk device or a semiconductor memory. The generated prediction model 204A is stored in the auxiliary storage device 204 in the present embodiment. The prediction model 204A here is an example of a so-called learning model. Therefore, the prediction model generation device 20 that generates the prediction model 204A is an example of a "learning model generation device."

<Prediction model generation processing>
The process of generating the prediction model 204A (see FIG. 2) for predicting the exhaust gas concentration will be described with reference to FIGS. 3 and 4. FIG.
FIG. 3 is a diagram illustrating an example of variables used to generate prediction model 204A in Embodiment 1. In FIG. In FIG. 3, parts corresponding to those in FIG. 1 are shown with reference numerals corresponding thereto. As shown in FIG. 3, it is not necessary to use all of the variables shown in FIG. 1 to generate prediction model 204A.

The flow rate of city gas introduced into the combustion furnace 10 is adjusted by the opening of the valve 101 .
The flow rate of air introduced into the combustion furnace 10 is adjusted by the degree of opening of the valve 102 .
Waste liquid is introduced into the combustion furnace 10 from the waste liquid tank 103 . In the case of FIG. 3, no valve for adjusting the flow rate is provided between the waste liquid tank 103 and the combustion furnace 10 . Therefore, the flow rate of the waste liquid introduced into the combustion furnace 10 from the waste liquid tank 103 always fluctuates. However, it does not prevent the installation of the valve.
In FIG. 3, the waste liquid is introduced into the combustion furnace 10 from the waste liquid tank 103, but the waste liquid generated in each step of the process may be directly introduced into the combustion furnace 10.

Waste liquid and process exhaust gas are introduced into the combustion furnace 10 from a production line 104 or the like. It should be noted that both the waste liquid and the process exhaust gas do not need to be introduced into the combustion furnace 10, and only one of them may be introduced into the combustion furnace 10. The flow rate of waste liquid and process exhaust gas here also fluctuates all the time.
In the combustion furnace 10, city gas, air, waste liquid, and process exhaust gas introduced into the furnace are burned, and exhaust gas is discharged. The combustion reaction in the combustion furnace 10 is affected by the temperature inside the furnace and the temperature of the furnace wall. The temperature inside the furnace and the temperature of the furnace wall also change from moment to moment.
The waste liquid tank 103 and the production line 104 are examples of "other equipment" in the scope of claims.

FIG. 4 is a flow chart for explaining the procedure for generating a prediction model using the concentration of exhaust gas as a prediction value. The symbol S shown in the figure means a step. The generation procedure shown in FIG. 4 is an example of a learning model generation method.
The generation procedure shown in FIG. 4 is realized through program execution by the processor 201 (see FIG. 2).

First, the processor 201 acquires teacher data (step 1). The training data here consist of the input amount of fuel (for example, city gas and air) as control variables, furnace status information (for example, the flow rate of waste liquid, the flow rate of process exhaust gas, the temperature inside the furnace, and the temperature of the furnace wall). , is given as a measure of the target variable (e.g. concentration of exhaust gas) corresponding to input and status information.

Next, processor 201 learns teacher data (step 2). Specifically, when the processor 201 inputs the control variables and the state information of the furnace to the input layer, it advances the learning of the parameters of the intermediate layer so that the predicted value of the objective variable is output from the output layer.
After learning is performed for one teacher data, the processor 201 determines whether or not there is unprocessed teacher data (step 3).

A positive result is obtained in step 3 if there is unprocessed training data. In this case, processor 201 returns to step 1 and one of the unprocessed teacher data is selected for training.
If there is no unprocessed teacher data, a negative result is obtained in step 3. In this case, processor 201 terminates the process of generating prediction model 204A (see FIG. 2). As a result, the city gas flow rate, air flow rate, waste liquid flow rate, amount of combustible material, process exhaust gas flow rate, furnace temperature, and furnace wall temperature are input, and the predicted concentration of exhaust gas is output. A predictive model 204A is generated.

<Output of predicted value>
FIG. 5 is a diagram illustrating the exhaust gas concentration prediction function using the prediction model 204A (see FIG. 2). This prediction function is also realized through program execution by the processor 201 . The function of the prediction unit 210 shown in FIG. 5 is an example of the "combustion prediction device" in the claims.
An exhaust gas concentration prediction unit 210 realized through execution of a program by the processor 201 inputs the current values of the input amount of fuel gas measured in the combustion furnace 10 and state information (such as the input amount of waste liquid).

Prediction unit 210 in the present embodiment provides the input current value to the prediction model, and outputs a predicted value of the exhaust gas concentration. The prediction operation by this prediction unit 210 is an example of a combustion prediction method.
Note that the prediction model 204A is stored in the auxiliary storage device 204. FIG.

<Summary of Embodiment 1>
By using the prediction model 204A described in the present embodiment, even if the amount and type of waste (process exhaust gas, waste oil, waste liquid) introduced into the combustion furnace 10 constantly fluctuates due to the emission source, the combustion furnace It is possible to predict the concentration of exhaust gas emitted from 10 (objective variable) in real time with high accuracy.
The learning of the prediction model 204A includes the input amount of fuel (city gas and air), the input amount of furnace state information (process exhaust gas, waste liquid), and the concentration of exhaust gas emitted when each value is input ( Objective variable) is used as teacher data. When the input amount of fuel (city gas and air) and the input amount of furnace state information (process exhaust gas, waste liquid) are input to this prediction model 204A, the concentration of exhaust gas discharged from the combustion furnace 10 is output as a predicted value. be done.

In the present embodiment, the prediction model 204A is used to predict the concentration of the exhaust gas discharged from the combustion furnace 10, but the amount of flame radiation in the combustion furnace 10 may be predicted, The surface temperature of the object to be heated in the combustion furnace 10 may be predicted. Also, the combustion efficiency of the combustion furnace 10 may be predicted.
Moreover, the physical quantity to be predicted is not limited to one, and any two or all three of the concentration of the exhaust gas, the amount of flame radiation, and the surface temperature of the object to be heated may be predicted at once. In that case, the prediction model 204A is trained to output these physical quantities as prediction values.

In addition, in the description of FIG. 3, the state variables are assumed to be the input amount of waste, the temperature in the furnace, and the temperature of the furnace wall, but the state variables may include the oxygen concentration in the furnace.
By the way, the input amount of waste does not have to be all three of process exhaust gas, waste liquid and waste oil, and at least one of them may be focused on. In other words, the combustion furnace 10 may assume input of any one or more of process exhaust gas, waste liquid, and waste oil.
In addition, in the present embodiment, the input amounts of city gas and air are assumed as control variables. is.

<Embodiment 2>
<Learning focusing on the content of causative substances>
Here, in order to explain the technique according to the second embodiment, a case of predicting the concentration of CO gas discharged from the combustion furnace 10 will be explained.
Mechanisms 1 and 2 below are conceivable for the mechanism by which CO gas is generated from the combustion furnace 10 .

・ Mechanism 1:
The flame is cooled by the atomized waste liquid introduced into the combustion furnace 10, and the oxidation reaction that produces _CO2 is prematurely frozen and discharged as CO.
・Mechanism 2:
As a result of a large amount of waste liquid containing a large amount of carbon being put in, part of the carbon is not completely burned and is discharged as CO. The C content here is the causative substance of CO.

This embodiment focuses on the mechanism 2. FIG.
Even if the amount of the entire waste liquid is the same, the concentration of CO discharged from the combustion furnace 10 differs between when a waste liquid containing a large amount of C is burned and when a waste liquid containing little C is burned.
Therefore, in the present embodiment, in addition to the input amount of waste liquid and the like, information on the concentration and viscosity of the C component, which is the causative substance of the exhaust gas, is used for learning of the prediction model 204A. Incidentally, the viscosity of a waste liquid with a high content of C is high, and the viscosity of a waste liquid with a low content of C is low. Therefore, if the viscosity is known, the content of C in the waste liquid can be specified.

<Prediction model generation processing 1>
FIG. 6 is a diagram illustrating an example of variables used to generate the prediction model 204A (see FIG. 2) in generation processing 1. In FIG. In FIG. 6, parts corresponding to those in FIG. 3 are shown with reference numerals corresponding thereto. In the case of FIG. 6, the training data includes not only the input amount of the waste liquid, etc., but also the concentration of C contained in the waste liquid, etc., and the measured values of the viscosity of the waste liquid, etc. FIG. The concentration and viscosity measurements here are real-time values.

In the generation process 1, the difference in the content of C contained in the waste liquid or the like that is input at each point is reflected in the learning of the prediction model.
Specifically, teacher data is prepared for each difference in content of C, which is a causative substance of CO gas, and a prediction model for each content of C is learned using the prepared teacher data. By preparing a prediction model for each C component content in this way, it is expected that the prediction accuracy of the concentration of CO gas will be improved as compared with a prediction model that focuses only on the input amount of waste liquid or the like.

It should be noted that the content of C charged into the combustion furnace 10 may change with time or be constant regardless of time.
The change in the content of C content over time occurs, for example, when the concentration of C content in the waste liquid changes with time for each discharge source. In this way, when the C content changes over time, it is better to switch the prediction model according to the C content at each time than to use a prediction model learned only from the input amount. The accuracy of predicting the concentration of CO gas emitted at the time is improved.

Even if the concentration of C content is constant regardless of time, the combustion furnace 10 of Company A that burns the waste liquid with the C content of C1, etc., and B that burns the waste liquid with the C content of C2, etc. The concentration of CO gas is different between the combustion furnace 10 of the company and the same amount of input waste liquid.
Therefore, the prediction model learned with the content C1 is used for the combustion furnace 10 of Company A, and the prediction model learned with the content C2 is used for the combustion furnace 10 of Company B, so that learning is performed only with the input amount. The accuracy of predicting the concentration of CO gas emitted at each point in time is improved as compared with the case of using the predictive model.

Thus, in the generation process 1, the input amount of fuel (city gas and air), the input amount of furnace state information (process exhaust gas, waste liquid), and the cause of the exhaust gas are used as teacher data for learning the prediction model 204A. The concentration of the substance C in the waste liquid and the concentration (objective variable) of the exhaust gas discharged at the time of inputting each value are used.
The teacher data may use the viscosity of the waste liquid instead of the concentration of the C component in the waste liquid, or may use both the concentration of the C component in the waste liquid and the viscosity of the waste liquid.
Also, the concentration of C in the process exhaust gas may be included in the training data.

<Prediction model generation processing 2>
FIG. 7 is a diagram illustrating another example of variables used to generate the prediction model 204A (see FIG. 2) in the generation process 2. In FIG. In FIG. 7, parts corresponding to those in FIG. 6 are shown with reference numerals corresponding thereto.
In the case of FIG. 7, instead of measuring the concentration and viscosity of C contained in the waste liquid at each point in time, the cumulative amount of C (time integral value) is included in the teaching data. The cumulative amount of C is calculated, for example, as a time integral value of the amount of C (=input amount×concentration) at each time point.

In the generation process 2, when the cumulative amount of C contained in the combustion furnace 10 is large, a temporary increase or decrease in the content of C contained in the waste liquid or the like does not appear in the change in the concentration of CO gas. Assume that when the cumulative amount of C charged into the combustion furnace 10 is small, a temporary increase or decrease in the content of C contained in the waste liquid or the like is immediately reflected in the change in the concentration of CO gas.
In such a case, as shown in FIG. 7, it is required to include the cumulative amount of C in the teacher data.

Thus, in the generation process 2, the input amount of fuel (city gas and air), the input amount of furnace state information (process exhaust gas, waste liquid), and the cause of the exhaust gas are used as teacher data for learning the prediction model 204A. The cumulative amount of C in the waste liquid, which is a substance, and the concentration (objective variable) of the exhaust gas discharged at the time of inputting each value are used.
When the viscosity of the waste liquid is given as information representing the content of C, the cumulative amount of C is calculated by the time integral value of the amount of C (=α×input amount×viscosity) at each time point. α is a coefficient.
In addition, the teacher data may include the cumulative amount of C derived from the process exhaust gas fed into the combustion furnace 10 .

<Summary of Embodiment 2>
Using the prediction model 204A described in the present embodiment, the difference in the concentration of the causative substance of the specific gas (that is, CO) contained in the waste (process exhaust gas, waste oil, waste liquid) thrown into the combustion furnace 10 It is possible to predict the concentration (objective variable) of the exhaust gas discharged from the combustion furnace 10 in real time with high accuracy regardless of the time change of the concentration, etc.
Incidentally, both the concentration and the viscosity of the causative substance contained in the waste liquid or the like may be used for the learning of the prediction model 204A.

6 and 7, CO is assumed to be the specific gas contained in the exhaust gas, and C is assumed to be the causative substance. Nitride can be considered as Further, when SOx is assumed as the specific gas, sulfide may be considered as the causative substance. When dioxin is assumed as the specific gas, chlorine compounds and aromatic compounds may be considered as causative substances.

Incidentally, a method similar to that shown in FIGS. 9 to 11, which will be described later, may be applied to generate the prediction model 204A for each content of the causative substance contained in the waste liquid or the like.
Also, a method similar to that shown in FIG. 12, which will be described later, may be applied to predict the concentration of the exhaust gas. In other words, the prediction model 204A used for predicting the exhaust gas concentration may be switched according to the current content of the causative substance contained in the waste liquid or the like.

<Embodiment 3>
<Learning focusing on differences in combinations of waste content>
Here, a case will be described where the combination of the contents of the waste thrown into the combustion furnace 10 changes with the lapse of time.
In the embodiment described above, it is assumed that the combination of the contents of the wastes thrown into the combustion furnace 10 is constant.
However, the above-mentioned waste oil, waste liquid, and process exhaust gas are not always fed at the same time.

FIG. 8 is a diagram illustrating a combustion pattern assumed in one combustion furnace 10. As shown in FIG. In FIG. 8, the materials to be burned in the combustion furnace 10 are assumed to be fuel and waste. Also, waste liquid, waste oil, incinerated materials, and process exhaust gas are assumed as waste.
However, the waste shown in FIG. 8 is an example, and it is not necessary to burn all of these in one combustion furnace 10 .

・Pattern #1:
Period/Pattern #2 in which only fuel (city gas and air) is supplied:
Period/Pattern #3 in which only the waste liquid is introduced:
Period/Pattern #4 in which only waste oil is supplied:
Period/Pattern #5 in which only incinerators are put in:
Period/Pattern #6 to #15 in which only process exhaust gas is introduced:
Period/Pattern #16 to #25 in which only any two of fuel, waste liquid, waste oil, incinerated matter, and process exhaust gas are input:
Period/pattern #26 to #30 in which only three of fuel, waste liquid, waste oil, incinerator, and process exhaust gas are input:
Period/Pattern #31 in which only four of fuel, waste liquid, waste oil, incinerated matter, and process exhaust gas are input:
The period during which all of the fuel, waste liquid, waste oil, incinerators, and process exhaust gas are input

In the present embodiment, a prediction model 204A (see FIG. 2) is generated for each of these patterns (for each period), and the concentration of the exhaust gas is predicted according to the combination of the combustibles put into the combustion furnace 10. The prediction accuracy is improved by switching the prediction model 204A to be used.

<Prediction model generation process according to combination of combustible materials>
Generating processing of the prediction model 204A in Embodiment 3 will be described below with reference to FIGS. 9 to 11. FIG.
FIG. 9 is a diagram illustrating the process of generating prediction model 204A according to the third embodiment.
The processor 201 classifies the teacher data according to the combination of combustible materials (step 11). The processor 201 classifies the teacher data according to the combination of combustible materials introduced into the combustion furnace 10 at each point in time, for example, based on information from sensors provided for each flow path assumed as combustible materials.

FIG. 10 is a diagram for explaining the function of classifying teacher data. The teacher data acquisition unit 211 realized through program execution by the processor 201 has a classification function. In Embodiment 3, the classification function is enabled, and the measured values and the like at each point of time are classified into teacher data for each combustible material combination.
For example, if the current combination is pattern #1, the acquisition unit 211 obtains, for example, the temperature in the furnace, the temperature of the furnace wall, the oxygen concentration in the furnace, the flow rate of city gas, the flow rate of air, and the concentration of exhaust gas in pattern #1. classify as training data for 1. As a result, the acquisition unit 211 generates 31 types of teacher data.

Returning to the description of FIG. After classifying the teacher data, the processor 201 uses the classified teacher data to generate a prediction model corresponding to each classification (step 12).
FIG. 11 is a diagram for explaining the learning function of a dedicated prediction model corresponding to each pattern using teacher data after classification. The learning unit 212 shown in FIG. 11 generates a dedicated prediction model corresponding to each combination by learning a prediction model using teacher data for each combination of combustible materials. In this embodiment, 31 types of prediction models are generated.

<Output of predicted value>
12A and 12B are diagrams for explaining prediction processing of predicted values according to the third embodiment.
An exhaust gas concentration prediction unit 213 realized through execution of a program by the processor 201 acquires current values of each unit measured in the combustion furnace 10 and current combination pattern information.
The current combination pattern is determined by the processor 201, for example, by analyzing the current values of each part measured in the combustion furnace 10. FIG.

The prediction unit 213 in the present embodiment inputs the current values of each part measured in the combustion furnace 10 to a dedicated prediction model corresponding to the current combination pattern, and outputs the predicted value of the exhaust gas concentration. For example, if the current combination pattern is pattern #2, the prediction unit 213 selects a dedicated prediction model corresponding to pattern #2 and uses it to predict the exhaust gas concentration.
The prediction models #1 to #31 are stored in the auxiliary storage device 204. FIG.

<Summary of Embodiment 3>
According to this embodiment, when the combination of the contents of the combustible materials put into the combustion furnace 10 changes with the passage of time, the exhaust gas concentration can improve the prediction accuracy of
In other words, by preparing a dedicated prediction model that has been learned according to the combination of combustible materials, the prediction accuracy is higher than when predicting exhaust gas concentration using a highly versatile prediction model. be able to.

<Embodiment 4>
<Learning focusing on the difference in the number of nozzles used for inputting combustible materials and their mounting positions>
The piping structure of the combustion furnace 10 (see FIG. 1) may differ depending on the difference in the combustible materials, site restrictions, and the like. In this embodiment, attention is paid to the difference in the number of nozzles used for throwing in the waste and the combination of the positions.

The combination of the number and positions of nozzles used for throwing in waste may depend on the structure or may change during combustion. When the number and position of nozzles change during combustion, for example, if the amount of fuel injected increases, the number of nozzles used for injection increases, but if the amount injected decreases, the number of nozzles used for injection decreases. be.

In any case, the difference in the number and position of nozzles used for inputting fuel gas, waste, etc. affects the combustion of waste, etc., and even if the amount of input to the combustion furnace 10 is the same, the concentration of exhaust gas, etc. changes. there's a possibility that.
For example, even if the input amount is the same when the waste liquid is input using one nozzle and when the waste liquid is input using two nozzles, the concentration of the exhaust gas discharged from the combustion furnace 10 may differ. .
In addition, even if the input amount is the same when each nozzle is attached to the ceiling surface of the combustion furnace 10 and when each nozzle is attached to the side surface of the combustion furnace 10, the concentration of the exhaust gas discharged from the combustion furnace 10 is different. there is a possibility.

<Example of nozzle number and arrangement pattern>
FIG. 13 is a diagram illustrating a case where one nozzle is provided for each combustible material on the ceiling of the combustion furnace 10. As shown in FIG. In FIG. 13, the parts corresponding to those in FIG. 7 are indicated by the reference numerals. Note that the wastes in FIG. 13 are waste liquid and waste oil.
In the case of FIG. 13, one nozzle for city gas, one nozzle for air, one nozzle for waste liquid, and one nozzle for waste oil, four nozzles in total are attached to the ceiling of the combustion furnace 10. provided in

FIG. 14 is a diagram illustrating a case where two nozzles are provided for each combustible material in the ceiling portion of the combustion furnace 10. As shown in FIG. In FIG. 14, the parts corresponding to those in FIG. 7 are indicated by the reference numerals. Wastes in FIG. 14 are also waste liquid and waste oil.
In the case of FIG. 14, two nozzles for city gas, two nozzles for air, two nozzles for waste liquid, and two nozzles for waste oil, eight in total, are installed on the ceiling of the combustion furnace 10. provided in

FIG. 15 is a diagram for explaining another example in which a total of two nozzles are provided for each combustible material dispersed in the ceiling and wall portions of the combustion furnace 10 . In FIG. 15, parts corresponding to those in FIG. 7 are shown with reference numerals corresponding thereto. Note that the wastes in FIG. 15 are also waste liquid and waste oil.
In the case of FIG. 15, four nozzles, one for city gas, one for air, one for waste liquid, and one for waste oil, are installed on the ceiling of combustion furnace 10 . A total of four nozzles, one for city gas, one for air, one for waste liquid, and one for waste oil, are provided on the side of the combustion furnace 10 is provided.

In addition, the number of nozzles for each combustible material may be three or more, and the mounting position of the nozzles may be on the bottom side of the combustion furnace 10 . Also, the nozzles attached to the side may be attached near the top surface or near the bottom surface. When a plurality of nozzles are attached to one side surface, the nozzles may be arranged vertically, horizontally, or diagonally. Further, when a plurality of nozzles are attached to two side surfaces, the nozzles may be attached to two opposing side surfaces or adjacent side surfaces.

Further, in the case of FIGS. 13 to 15, the case where the same number of nozzles are attached to each combustible material put into the combustion furnace 10 has been described, but the number of nozzles attached to each combustible material may be different. . For example, two nozzles may be prepared for inputting waste liquid, and one nozzle may be prepared for inputting waste oil.
In this way, the combination of the number of nozzles for ejecting the combustible material to the combustion furnace 10 and the mounting position can be classified into many patterns, considering the difference in the internal volume and the shape of the internal space of the combustion furnace 10. .

<Prediction model generation processing according to the combination of the number of nozzles and the mounting position>
The prediction model generation process according to the fourth embodiment will be described below with reference to FIGS. 16 to 19. FIG.
FIG. 16 is a diagram illustrating generation processing of prediction model 204A in the fourth embodiment. In FIG. 16, parts corresponding to those in FIG. 9 are shown with reference numerals corresponding thereto.

The processor 201 classifies the teacher data according to the combination of the number of nozzles used for charging the combustible material and the mounting position (step 21). The processor 201 determines where to classify measurements obtained from the combustion furnace 10 based on the structure of the combustion furnace 10 .
Note that when the number and mounting positions of nozzles used for charging the combustible material are switched according to the operator's operation, a plurality of teaching data are generated for one combustion furnace 10 .

FIG. 17 is a diagram for explaining the function of classifying teacher data. The teacher data acquisition unit 214 realized through program execution by the processor 201 has a classification function. The acquisition unit 214 in FIG. 17 focuses on the number and mounting position of the nozzles used for charging the combustible material, and classifies the measured values and the like at each point into teacher data for each combination of the number of nozzles and mounting position.

Returning to the description of FIG. After classifying the teacher data, the processor 201 uses the classified teacher data to generate a prediction model corresponding to each classification (step 12).
FIG. 18 is a diagram for explaining the learning function of a dedicated prediction model corresponding to each pattern using teacher data after classification. The learning unit 215 shown in FIG. 18 generates a dedicated prediction model corresponding to each combination by learning the prediction model using teacher data for each combination of the number of nozzles and the mounting position.

<Prediction processing of predicted value>
19A and 19B are diagrams for explaining prediction processing of predicted values according to the fourth embodiment.
The exhaust gas concentration prediction unit 216, which is realized through the execution of the program by the processor 201, obtains the current values of each unit measured in the combustion furnace 10 and the current combination pattern information. However, the combination pattern here is different from that of the third embodiment in that it is determined by the number of nozzles used for charging the combustibles and their mounting positions.

The prediction unit 216 in the present embodiment inputs the current values of each part measured in the combustion furnace 10 to a dedicated prediction model corresponding to the current combination pattern, and outputs the predicted value of the exhaust gas concentration. For example, if the current combination pattern is pattern #2, the prediction unit 216 selects a dedicated prediction model corresponding to pattern #2 and uses it to predict the exhaust gas concentration.
The prediction models #1 to #N are stored in the auxiliary storage device 204. FIG.

<Summary of Embodiment 4>
According to the present embodiment, since a dedicated prediction model is learned for each combination of the number of nozzles used for charging the combustible material and the mounting position, compared to the case where the difference in the number of nozzles and the mounting position is not considered. , the accuracy of predicting the concentration of the exhaust gas can be improved.
In other words, by preparing a dedicated prediction model that has learned according to the combination of the number of nozzles and their mounting positions, the prediction accuracy is higher than when predicting the exhaust gas concentration using a highly versatile prediction model. can be enhanced.

<Embodiment 5>
<Learning that focuses on the time lag between changes in the input amount and changes in the exhaust gas concentration>
Depending on the combustion furnace 10 , the time lag between changes in the input amount of fuel gas, wastes, and the like appearing as changes in the concentration of the exhaust gas discharged from the combustion furnace 10 may not be negligible. For the combustion furnace 10 with a large time lag, there is a possibility that a highly accurate prediction result cannot be obtained even if the prediction model is learned without considering the time lag.

<Prediction model generation processing considering time lag>
The prediction model generation process according to the fifth embodiment will be described below with reference to FIGS. 20 to 22. FIG.
FIG. 20 is a diagram illustrating generation processing of the prediction model 204A (see FIG. 2) according to the fifth embodiment.
The processor 201 acquires the time difference until the change in the input amount of the combustible material affects the exhaust gas concentration (step 31).

Next, the processor 201 generates teacher data by combining the measured values of the input amount and the exhaust gas at times different by the time difference obtained in step 31 (step 32).
As a method for determining the teacher data, a case where the input amount measurement value is used as the reference time (current time) and a case where the exhaust gas measurement value is used as the reference time (current time) are conceivable.
FIG. 21 is a diagram for explaining an example in which measured values such as the amount of combustible material put in at the reference time (current time) and measured values of the concentration of exhaust gas after a certain period of time from the reference time are used as teacher data. be. In FIG. 21, it is assumed that the time difference (time lag) between the change in the input amount of the combustible material and the change in exhaust gas concentration is 3 minutes. Of course, 3 minutes is just an example.

The chart shown in FIG. 21 shows the input amount of fuel gas (city gas and air) at each point, and the state information (furnace temperature, furnace oxygen concentration, flow rate of waste liquid, flow rate of process exhaust gas, amount of waste material, flow rate of waste oil). ) and the measured value of the exhaust gas concentration (CO concentration).
In the case of FIG. 21, "2022/4/10:01" is the current time. In this case, the exhaust gas concentration associated with the input amount of fuel gas and the state information at the current time is 50 [ppm] at "2022/4/10:04" three minutes after the current time. That is, in FIG. 21, combinations of measured values that satisfy the relationship of colored numerical values are acquired as teacher data and used for learning the prediction model.

FIG. 22 is a chart for explaining an example in which the measured value of the exhaust gas concentration at the reference time (current time) and the measured value of the input amount of combustible material that was input a certain time before the reference time are used as teacher data. be. In the case of FIG. 22 as well, it is assumed that the time difference (time lag) between the change in the input amount of the combustible material and the change in the exhaust gas concentration is 3 minutes.
However, in the case of FIG. 22, the time when the measured value of the concentration of the exhaust gas is obtained is set as the reference time (current time), and the input amount of the fuel gas (city gas and air) and the state information (furnace temperature , furnace concentration, waste liquid flow rate, process exhaust gas flow rate, waste amount, waste oil flow rate), and exhaust gas concentration (CO concentration).

Specifically, with respect to the exhaust gas concentration value (that is, 50 [ppm]) at "2022/4/10:04", the input amount of fuel gas at "2022/4/10:01" three minutes ago Associate the exhaust gas concentration associated with the state information.
That is, in FIG. 22, combinations of measured values that satisfy the relationship of the colored numerical values are acquired as teacher data and used for learning the prediction model.
21 and 22, the combination of data recorded as teacher data is the same, but the time at which any measured value is obtained is different from the reference time.

<Prediction processing of predicted value>
In the case of this embodiment, one prediction model is generated.
Therefore, the process of predicting the exhaust gas concentration is the same as in the first embodiment.

<Summary of Embodiment 5>
According to the present embodiment, it is possible to improve the prediction accuracy of the exhaust gas concentration compared to the case where the time lag until the result of the introduction of the combustible material appears in the exhaust gas concentration is not considered.
As for the length of the time lag used to generate the teacher data, for example, the time difference at which the matching rate between the measured value and the predicted value becomes high may be specified from experimental results.

<Embodiment 6>
<Modification of prediction model using error between predicted value and actual value>
Many samples are required to improve the learning accuracy of the prediction model 204A (see FIG. 2). However, collecting a large number of samples is time consuming.
Therefore, in the present embodiment, a Gradient Boosting Decision Tree (GBDT) is applied as a method of increasing the learning accuracy of the prediction model 204A using limited samples.

FIG. 23 is a diagram illustrating a method of correcting prediction model 204A using a gradient boosting decision tree. In FIG. 23, the correction of the prediction model 204A progresses in the order of numbers 1, 2, 3, .
As shown in FIG. 23, the difference between the objective variable (concentration of exhaust gas) and the predicted value by the prediction model 204A created so far is learned, and prediction models are added sequentially until the difference becomes small.
FIG. 24 is a flowchart for explaining the procedure for generating prediction model 204A (see FIG. 2) according to the sixth embodiment. In FIG. 24, parts corresponding to those in FIG. 4 are shown with reference numerals.

First, processor 201 acquires teacher data (step 1), and learns the acquired teacher data (step 2).
After learning is performed for one teacher data, the processor 201 determines whether or not there is unprocessed teacher data (step 3).
A positive result is obtained in step 3 if there is unprocessed training data. In this case, processor 201 returns to step 1 .

If there is no unprocessed teacher data, a negative result is obtained in step 3. In this case, the processor 201 calculates the difference between the predicted value of the prediction model and the correct value (step 41). The correct value here is the concentration value of the exhaust gas measured for the input value corresponding to the predicted value.
Processor 201 then determines whether the difference is less than or equal to a threshold (step 42).

If the difference is greater than the threshold, step 42 yields a negative result. In this case, processor 201 modifies the predicted value of the prediction model by multiplying the difference by the learning rate (step 43). After modifying the prediction model, processor 201 returns to step 41 .
If the difference is less than or equal to the threshold, a positive result is obtained at step 42 . In this case, processor 201 terminates the process of generating prediction model 204A. As a result, a prediction model 204A with high prediction accuracy is generated even with a small number of samples.

<Summary of Embodiment 6>
As described above, the difference between the predicted values and the actual measurements can be used to modify the prediction model 204A to efficiently improve the accuracy of the prediction model 204A.
In the present embodiment, the method for correcting prediction model 204A using a gradient boosting decision tree was exemplified, but the specific correction method is intended to be limited to the method for correcting prediction model 204A using a gradient boosting decision tree. do not.

<Embodiment 7>
In Embodiment 6, a highly accurate learning model is obtained even with a small number of samples of teacher data by correcting prediction model 204A using a gradient boosting decision tree. Increase the number of samples of training data in advance.
FIG. 25 is a diagram explaining a method of correcting a learning model using teacher data generated by interpolation calculation. In FIG. 25, the parts corresponding to those in FIG. 4 are shown with reference numerals.

First, when the processor 201 acquires the measured values (step 51), the acquired measured values are used to generate teacher data (step 52).
FIG. 26 is a diagram illustrating a method of generating teacher data. (A) shows a method of increasing the number of teacher data samples using linear interpolation, and (B) shows a method of increasing the number of teacher data samples using a polynomial or regression model.
In FIG. 26, for convenience of explanation, the vertical axis represents the concentration of the specific gas, and the horizontal axis represents the input amount of city gas. In the case of FIG. 26, the purpose is to increase the number of samples for the concentration of the specific gas and the input amount of city gas.

In FIG. 26A, the midpoint between two measured values is calculated as a complementary value by linear interpolation, but two or more complementary values may be calculated between two measured values.
In FIG. 26(B), a function passing through the measured values is defined, and the midpoint between the two measured values is calculated as the interpolated value. may
After completing the generation of complementary values to be used as teacher data, the same processing (steps 1 to 3) as in FIG. 4 is executed.

<Summary of Embodiment 7>
As described above, it is possible to efficiently improve the accuracy of the prediction model 204A by learning the prediction model 204A using the teacher data including the complementary value calculated using the interpolation operation.

<Embodiment 8>
<Example of preparing a server for generating a prediction model>
FIG. 27 is a diagram for explaining the conceptual configuration of the combustion furnace system 1A assumed in the seventh embodiment. In FIG. 27, parts corresponding to those in FIG.
In the case of the combustion furnace system 1A shown in FIG. 27, the information processing server 30 existing on the network N executes the process of generating the prediction model. The network N here is assumed to be, for example, a LAN (Local Area Network), the Internet, or a mobile communication system such as 4G or 5G.

The information processing server 30 acquires teacher data, generates a prediction model using the acquired teacher data, predicts the exhaust gas concentration using the prediction model, and the like.
The information processing server 30 is a server type computer. However, the information processing server 30 does not need to be composed of a single computer, and may be composed of a plurality of computers connected via a network N and cooperating to execute prediction model generation processing and the like. .

Acquisition processing of teaching data is executed by an information terminal (server, desktop computer, notebook computer, etc.) in the building or site where the combustion furnace 10 is installed, and the acquired teaching data is processed. It may be uploaded to the server 30 .
Further, the predicted value of the exhaust gas concentration predicted by the information processing server 30 may be fed back to an information terminal in the building or site where the combustion furnace 10 is provided.

<Other embodiments>
(1) Although the embodiments of the present invention have been described above, the technical scope of the present invention is not limited to the scope described in the above-described embodiments. It is clear from the scope of claims that the technical scope of the present invention includes various modifications and improvements to the above-described embodiment.

(2) In the above-described embodiments, an industrial furnace having a manufacturing process and a chemical reaction process is assumed as a waste generation source, but the waste generation source is not limited to an industrial furnace.

(3) In the above-described embodiment, for convenience of explanation, the prediction model for predicting the exhaust gas concentration was explained as the objective variable. or the surface temperature of the object to be heated. Also, the combustion efficiency P of the combustion furnace 10 may be predicted as an objective variable. The combustion efficiency P can be calculated from the amount of heat taken away by the object to be heated/the amount of heat input.

(4) In the above-described embodiment, the flow rate of the process exhaust gas, the flow rate of the waste oil, the flow rate of the waste liquid, etc. were exemplified as the state information. ratio of city gas and air), and the opening degree of an exhaust damper provided in the exhaust path of the combustion furnace 10 may be included.

(5) In the above embodiment, the combustion furnace was described, but an incinerator or a heating furnace may be used. An incinerator is a furnace that uses fuel gas to incinerate waste and other materials. A heating furnace is a furnace that heats an object to be heated using a fuel gas.

DESCRIPTION OF SYMBOLS 1, 1A... Combustion furnace system 10... Combustion furnace 20... Prediction model generation apparatus 30... Information processing server 101, 102... Valve 103... Waste liquid tank 104... Production line etc. 201... Processor 202... ROM , 203...RAM, 204...auxiliary storage device, 204A...prediction model, 205...communication interface, 210, 213, 216...prediction unit, 211, 214...acquisition unit, 212, 215...learning unit

Claims (25)

The computer acquires, as teaching data, the input amount of fuel as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information,
A computer generates a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable.
A learning model generation method,
Acquiring the training data for each period in which only the fuel is burned in the furnace, a period in which only waste is burned in the furnace, and a period in which the fuel and the waste are burned together in the furnace,
A computer generates the learning model for each period using the teacher data acquired for each period ;
Learning model generation method. The learning model corresponding to the period in which only the waste is burned or the period in which the fuel and the waste are burned together uses the training data obtained for each difference in the composition of the waste. , generated for each difference in the composition,
The learning model generation method according to claim 1 . The waste is at least one of waste liquid from other equipment, waste oil from other equipment, process exhaust gas from other equipment, and incinerated matter.
The learning model generation method according to claim 1 or 2 . The computer acquires, as teaching data, the input amount of fuel as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information,
A computer generates a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable.
A learning model generation method,
Acquisition of the teacher data and generation of the learning model reflect at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for input .
Learning model generation method. an acquisition unit that acquires, as teacher data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information;
A learning unit that uses the teacher data, inputs the input amount and the state information, and generates a learning model that outputs the predicted value of the objective variable,
A learning model generation device,
The acquisition unit acquires the training data for a period in which only the fuel is burned in the furnace, a period in which only waste is burned in the furnace, and the fuel and the waste are burned together in the furnace. obtained by period,
The learning unit generates the learning model for each period using the teacher data acquired for each period.
Learning model generator. an acquisition unit that acquires, as teacher data, a fuel input amount as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information;
A learning unit that uses the teacher data, inputs the input amount and the state information, and generates a learning model that outputs the predicted value of the objective variable,
A learning model generation device,
Acquisition of the teacher data and generation of the learning model reflect at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for input.
Learning model generator. to the computer,
A function of acquiring, as teacher data, a fuel input amount as a control variable, reactor state information, and a measured value of an objective variable corresponding to the input amount and the state information;
A function of generating a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable;
It is a program to realize
The function of acquiring the training data includes a period in which only the fuel is burned in the furnace, a period in which only waste is burned in the furnace, and a period in which the fuel and the waste are burned together in the furnace. obtained by period,
The function to generate the learning model uses the teacher data acquired for each period to generate the learning model for each period.
program. to the computer,
A function of acquiring, as teacher data, a fuel input amount as a control variable, reactor state information, and a measured value of an objective variable corresponding to the input amount and the state information;
A function of generating a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable;
It is a program to realize
Acquisition of the teacher data and generation of the learning model reflect at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for input.
program. the computer
For a learning model that has learned the relationship between the amount of fuel input as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information, the measured value of the input amount and the relevant Given a measured value of state information as an input, and outputting a predicted value of the corresponding objective variable,
A combustion prediction method,
As the learning model,
Using the training data acquired by the period in which only the fuel is burned in the furnace, the period in which only waste is burned in the furnace, and the period in which the fuel and the waste are burned together in the furnace using a learning model for each period generated by
Combustion prediction method. the computer
For a learning model that has learned the relationship between the amount of fuel input as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information, the measured value of the input amount and the relevant Given a measured value of state information as an input, and outputting a predicted value of the corresponding objective variable,
A combustion prediction method,
As the learning model,
When acquiring teacher data and generating the learning model, using a learning model that reflects at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for input.
Combustion prediction method. a learning model that learns the relationship between the input amount of fuel as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information;
a prediction unit that, when the input amount and the state information are given as inputs, supplies the measured value of the input amount and the measured value of the state information to the learning model, and outputs a predicted value of the objective variable;
A combustion prediction device having
As the learning model,
Using the training data acquired by the period in which only the fuel is burned in the furnace, the period in which only waste is burned in the furnace, and the period in which the fuel and the waste are burned together in the furnace using a learning model for each period generated by
Combustion predictor. a learning model that learns the relationship between the input amount of fuel as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information;
a prediction unit that, when the input amount and the state information are given as inputs, supplies the measured value of the input amount and the measured value of the state information to the learning model, and outputs a predicted value of the objective variable;
A combustion prediction device having
As the learning model,
When acquiring teacher data and generating the learning model, using a learning model that reflects at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for input.
Combustion predictor. to the computer,
For a learning model that has learned the relationship between the amount of fuel input as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information, the measured value of the input amount and the relevant A function that gives the measured value of state information as an input and outputs the predicted value of the objective variable
It is a program to realize
As the learning model,
Using the training data acquired by the period in which only the fuel is burned in the furnace, the period in which only waste is burned in the furnace, and the period in which the fuel and the waste are burned together in the furnace using a learning model for each period generated by
program. to the computer,
For a learning model that has learned the relationship between the amount of fuel input as a control variable, the state information of the reactor, and the measured value of the objective variable corresponding to the input amount and the state information, the measured value of the input amount and the relevant A function that gives the measured value of state information as an input and outputs the predicted value of the objective variable
It is a program to realize
As the learning model,
When acquiring teacher data and generating the learning model, using a learning model that reflects at least one of the number of nozzles used for inputting waste into the furnace and the difference in the position of the nozzle used for input.
program. The computer acquires, as teaching data, the input amount of fuel gas used for combustion of the object as a control variable, the state information of the furnace, and the measured value of the objective variable corresponding to the input amount and the state information,
A computer generates a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable.
Learning model generation method. The state information includes at least one of an input amount of waste containing a causative substance of the specific gas discharged from the furnace, a concentration of the causative substance in the waste, and a viscosity of the waste,
The learning model generation method according to claim 15 . The training data is a combination of the input amount of the fuel gas , the input amount of the waste, and the measured value measured after a predetermined time from the input of the fuel gas and the waste.
The learning model generation method according to claim 15 . The computer corrects the learning model using the error between the predicted value predicted using the learning model and the actual measured value;
The learning model generation method according to claim 15 . The state information includes at least one of process exhaust gas, waste oil, waste liquid, material to be incinerated, oxygen concentration in the furnace, temperature in the furnace, temperature of the furnace wall,
The learning model generation method according to claim 15 . The target variable includes at least one of CO, CO ₂ , NO _x , SO _x , dust, dioxin, unburned fuel gas, greenhouse gas, amount of flame radiation, and surface temperature of the object to be heated.
The learning model generation method according to claim 15 . an acquisition unit that acquires, as teacher data, an input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information;
A learning unit that uses the teacher data, inputs the input amount and the state information, and generates a learning model that outputs the predicted value of the objective variable,
Learning model generator. to the computer,
A function of acquiring, as teacher data, an input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and a measured value of an objective variable corresponding to the input amount and the state information;
A function of generating a learning model using the teacher data, inputting the amount of input and the state information, and outputting the predicted value of the objective variable;
program to make it happen. the computer
For a learning model that has learned the relationship between the input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and the measured value of the objective variable corresponding to the input amount and the state information, Giving as input the measured value of the input amount and the measured value of the state information, outputting the predicted value of the corresponding objective variable;
Combustion prediction method. A learning model that learns the relationship between the input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and the relationship between the input amount and the measured value of the objective variable corresponding to the state information;
a prediction unit that, when the input amount and the state information are given as inputs, supplies the measured value of the input amount and the measured value of the state information to the learning model, and outputs a predicted value of the objective variable;
Combustion prediction device with to the computer,
For a learning model that has learned the relationship between the input amount of fuel gas used for combustion of an object as a control variable, furnace state information, and the measured value of the objective variable corresponding to the input amount and the state information, A program for realizing the function of outputting the predicted value of the objective variable by giving the measured value of the input amount and the measured value of the state information as input.

Images