Control of selective catalytic reduction using machine learning

Abstract

To provide control of selective catalytic reduction using machine learning.SOLUTION: A method includes acquiring input data (102) including one or a plurality of input values, by a computing system including a machine learning model (104). The method includes generating output data (106) indicating an amount (106) of a reactant (110), by the machine learning model (104) on the basis of the input data (102). The method includes providing a signal allowing a selective catalytic reduction system (112) to provide the amount (106) of the reactant (110), by the computing system.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This disclosure relates generally to selective catalytic reduction of emissions, and more particularly to systems and methods for controlling selective catalytic reduction systems using machine learning. [Background technology]

[0002] Selective catalytic reduction is a process that can convert one or more input compounds (e.g., gases such as NO, NO2, etc.) into one or more exhaust compounds (e.g., common atmospheric gases such as pure nitrogen, water vapor, etc.) using a reactant (e.g., ammonia, urea, etc.). In some cases, the catalytic reduction process can include providing a reactant (e.g., a liquid reactant such as ammonia, urea, diesel exhaust fluid, etc.) to a catalytic reduction system (e.g., via a catalyst bed or catalyst chamber). In some cases, optimizing the selective catalytic reduction process can include optimizing the amount of reactant provided. For example, providing too much reactant can cause reactant slip (e.g., ammonia slip, etc.), where the reactant can be unintentionally released from the catalyst chamber without causing conversion of the input compound. As another example, providing too little reactant can cause a portion of the input compound to be released from the catalyst chamber without being converted to an output compound. Summary of the Invention

[0003] Aspects and advantages of the systems and methods according to the present disclosure will be set forth in part in the description that follows, or will be obvious from the description, or may be learned by practice of the techniques.

[0004] According to one embodiment, a method is provided. The method includes obtaining, by a computing system including a machine learning model, input data including one or more input values. The method includes generating, by the machine learning model based on the input data, output data indicative of an amount of a reactant. The method includes providing, by the computing system, a signal to cause the amount of the reactant to be provided to a selective catalytic reduction system.

[0005] According to another embodiment, a method is provided. The method includes obtaining, by a self-adjusting reactant flow control system, emissions data indicative of one or more emissions. The method includes adjusting, by the self-adjusting reactant flow control system, a first amount of a reactant supplied to a selective catalytic reduction system based on the emissions data. The method includes monitoring the adjustment by a computing system comprising one or more computing devices. The method includes determining that the first amount of the reactant supplied to the selective catalytic reduction system has stabilized. The method includes training a machine learning model by the computing system using one or more example data including data indicative of the stabilized first amount.

[0006] According to another embodiment, a computing system is provided. The computing system includes one or more processors. The computing system includes one or more non-transitory computer-readable media that collectively store a machine learning model and instructions that, when executed by the one or more processors, cause the computing system to perform operations. The operations include obtaining input data including one or more input values. The operations include generating, by the machine learning model based on the input data, output data indicative of an amount of a reactant. The operations include providing a signal to cause the amount of the reactant to be provided to a selective catalytic reduction system.

[0007] These and other features, aspects, and advantages of the present systems and methods will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present technology and, together with the description, serve to explain the principles of the technology.

[0008] A full and enabling disclosure of the present systems and methods, including the best mode of making and using the same, directed to one of ordinary skill in the art, is set forth in this specification, which makes reference to the accompanying figures. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a block diagram of a system for selective catalytic reduction using machine learning, according to an embodiment of the present disclosure. [Figure 2] FIG. 1 is a block diagram of a system for training a machine learning system for controlling a selective catalytic reduction system, according to an embodiment of the present disclosure. [Figure 3] FIG. 1 is a block diagram of a machine learning system for controlling a selective catalytic reduction system, according to an embodiment of the present disclosure. [Figure 4] FIG. 1 is a block diagram of a system for controlling a selective catalytic reduction system using machine learning while continuously training the machine learning control system during operation, in accordance with an embodiment of the present disclosure. [Figure 5] FIG. 1 is a flowchart diagram of an exemplary method for controlling a selective catalytic reduction system using machine learning, according to an embodiment of the present disclosure. [Figure 6] FIG. 1 is a flowchart diagram of an exemplary method for training a machine learning system, according to an embodiment of the present disclosure. [Figure 7] FIG. 1 is a block diagram of an exemplary computing system for performing various operations according to embodiments of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0010] Reference will now be made in detail to the present system and method embodiments, one or more examples of which are illustrated in the drawings. Each example is provided by way of explanation of the present technology, not as a limitation thereof. Indeed, it will be apparent to those skilled in the art that modifications and variations can be made in the present technology without departing from the scope or spirit of the claimed technology. For example, features illustrated or described as part of one embodiment can be used in another embodiment to yield still a further embodiment. Accordingly, the present disclosure is intended to cover such modifications and variations as come within the scope of the appended claims and their equivalents.

[0011] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Additionally, unless otherwise specified, all embodiments described herein should be considered exemplary.

[0012] The detailed description uses numerical and letter designations to refer to features in the drawings. Like or similar designations in the drawings and description are used to refer to like or similar parts of the invention. As used herein, the terms "first," "second," and "third" may be used interchangeably to distinguish one component from another and are not intended to denote the location or importance of the individual components.

[0013] The term "fluid" can be a gas or a liquid. The term "fluid communication" means that a fluid is capable of making a connection between designated areas.

[0014] Approximate terms such as "approximately," "about," "approximately," and "substantially" are not intended to be limited to the exact value stated. In at least some cases, approximating language can correspond to the precision of an instrument for measuring a value or the precision of a method or machine for constructing or manufacturing a component and / or system. In at least some cases, approximating language can correspond to the precision of an instrument for measuring a value or the precision of a method or machine for constructing or manufacturing a component and / or system. For example, approximating language can refer to within a margin of 1, 2, 4, 5, 10, 15, or 20% for a particular value, a range of values, and / or any of the endpoints defining the range of values. When used in the context of angles or directions, such terms include a range of plus or minus 10 degrees of the stated angle or direction. For example, "approximately vertical" includes directions within 10 degrees of any direction, e.g., clockwise or counterclockwise, from vertical.

[0015] As used herein, the terms "comprises," "comprising," "includes," "including," "has," "having," or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, article, or apparatus that includes a list of features is not necessarily limited to only those features and may include other features not expressly listed or inherent to such process, method, article, or apparatus. Furthermore, unless expressly stated to the contrary, "or" refers to an inclusive or, not an exclusive or. For example, condition A or B is satisfied by any one of the following: A is true (or present) and B is false (or absent), A is false (or absent) and B is true (or present), and both A and B are true (or present).

[0016] Herein and throughout the specification and claims, range limitations are combinable and interchangeable, and unless the context and language dictate otherwise, such ranges are identified and include all subranges subsumed therein. For example, all ranges disclosed herein are inclusive of the endpoints, and the endpoints are independently combinable with each other.

[0017] overview The present disclosure relates generally to systems and methods for machine learning control of selective catalytic reduction systems. More particularly, the present disclosure relates to systems and methods for training machine learning models to predict optimal or near-optimal amounts of reactants to be delivered to a particular system (e.g., a particular machine, plant, device, turbine, catalyst chamber, etc.) based on the historical performance of the particular system.

[0018] For example, the system may include a self-tuning catalyst flow control component, such as a proportional-integral-derivative (PID) controller, that can automatically adjust reactant flow based on sensor data (e.g., emissions data collected from an emissions stack). Upon one or more changes to the system (e.g., a change in turbine load or other operating variables in a gas turbine system), the self-tuning control system may iteratively (e.g., continuously) adjust reactant flow until the reactant flow rates converge to an optimal flow rate value. Upon convergence, the optimal flow rate value may be stored as training data examples, along with additional input data that may be relevant to predicting the optimal reactant flow rate (e.g., emissions data, reactant slip data, ambient conditions, flue gas temperature, system load such as turbine load, duct burner fuel flow rate, etc.). The training examples may be used to perform machine learning predictions of optimal flow rates for a particular system based on example data collected from the particular system.

[0019] In some cases, exemplary machine learning algorithms for predicting optimal flow rates may include nonparametric or semiparametric machine learning. For example, in some cases, collected data examples may be stored in a data structure that correlates various input data values ​​with corresponding optimal flow rate values. Based on input data corresponding to the current state of the catalytic reduction system, the machine learning system may obtain multiple data examples that indicate past optimal flow rate values ​​associated with similar system states. Based on the obtained data examples, the machine learning system may predict optimal reactant flow rate values ​​for the current state of the system. Various methods (e.g., statistical methods, machine learning methods, etc.), such as parametric, nonparametric, or semiparametric regression, linear, polynomial, or other interpolation, may be used to combine the data examples to generate a prediction. As a simplified illustrative example, a machine learning system that predicts optimal flow rates based on one input variable may receive the input value, obtain two or more data examples associated with similar input values, and perform linear regression or linear interpolation between the data examples to generate an output. Similar methods can be adapted to predict optimal flow based on input data including multiple input variables (e.g., multiple input variables related to a gas turbine system, such as turbine load, flue gas temperature, duct burner fuel flow rate, ambient conditions such as temperature, emissions sensor data such as NOx data or reactant slip data, start-up or shutdown condition data, or other relevant data).

[0020] In some cases, the machine learning flow control system can continuously self-calibrate using a selective forgetting process. For example, the machine learning flow control system can continue to collect additional training data while the machine learning flow control system is operating and can periodically remove old data from the training data structure. For example, the machine learning model can predict an optimal reactant flow rate, and the predicted reactant flow rate can be provided over a waiting period (e.g., related to the waiting period for exhaust data), after which the self-tuning flow control system can determine a true optimal flow rate based on the exhaust data. If the true optimal flow differs significantly from the predicted flow (e.g., according to an error threshold, etc.), new training data examples can be added based on the true optimal flow. In response to the new training data examples being added to the training data structure, one or more older data examples can be determined to be stale (e.g., based on the age of the data example, etc.) and can be removed from the training data structure. In this way, for example, the machine learning flow control system can continuously recalibrate as a system (e.g., a selective catalytic reduction system, a turbine system, etc.) ages, undergoes maintenance, or otherwise experiences changes that may alter its operating behavior.

[0021] An example application of some of the provided systems and methods may include industrial turbine applications. For example, an industrial turbine system may include a gas turbine, a heat recovery steam generator with a selective catalytic reduction system, an exhaust stack, and stack exhaust (e.g., NO xand one or more sensors for monitoring the system's operation (e.g., turbine load, duct burner fuel flow rate, reactant flow rate, etc.). In some cases, calibrating such systems via traditional methods (e.g., using a PID controller, etc.) can be difficult due to long latency times (e.g., minutes, etc.) between when a change in system operation (e.g., a change in turbine load, duct burner fuel flow rate, reactant flow rate, etc.) is made or detected (e.g., ambient conditions, etc.) and when the corresponding effect of the change on emissions can be detected (e.g., by sensors associated with the exhaust stack, etc.). Advantageously, some provided systems and methods can predict optimal flow rates sooner than some alternative methods, thereby providing improved catalytic reduction and reducing undesirable emissions (e.g., NO x Reduces the amount of ammonia slip, etc.

[0022] In some cases, the exemplary machine learning flow rate prediction can be combined with other methods for controlling a selective catalytic reduction system (e.g., in a turbine-based industrial system, etc.). For example, in some cases, the selective catalytic reduction system can operate in a certain manner during continuous operation of the turbine system and behave differently during a shutdown or startup period (e.g., one hour before shutdown, one hour after startup, etc.). In some cases, the machine learning flow rate control can include one or more adjustments for such startup or shutdown periods. For example, in some cases, a reactant flow rate multiplier can be used to adjust the reactant flow rate during startup and shutdown. For example, the reactant flow rate can be stopped (e.g., reduced to zero) or significantly reduced before shutdown (e.g., one hour before shutdown) to prevent reactant slippage during startup. As another example, the reactant flow rate multiplier may be increased (e.g., from zero to one) during the startup time window. In some cases, controlling the reactant flow rate during startup or shutdown may include using a machine learning system to predict an expected reactant amount; determining, by a computing system, a reactant multiplier based on the time to shutdown or startup; and providing, to the selective catalytic reduction system, an amount of reactant equal to the predicted reactant amount multiplied by the reactant multiplier.

[0023] Systems and methods according to exemplary aspects of the present disclosure may provide improved emissions-related outcomes of selective catalytic reduction systems (e.g., reduced ammonia slip, NO x The machine learning model may provide various technical effects and benefits, such as improved performance (e.g., reduced emissions, improved regulatory compliance, reduced compliance costs, etc.), and improved capabilities of computing systems that run the machine learning model (e.g., reduced memory footprint, reduced training data requirements, reduced computational costs, etc.).

[0024] For example, some alternative methods of reactant flow rate control (e.g., PID-based methods) can suffer from long latency (e.g., several minutes) between when a system change (e.g., a PID-based adjustment to a reactant flow rate) is made and when the effect of the change can be detected (e.g., via an exhaust sensor). In some cases, the system may require several cycles of change and delayed feedback (e.g., cycles of overcorrection, undercorrection, etc.) before ultimately converging to an optimal solution. In systems where each feedback cycle may take several minutes (e.g., 4 minutes), these repeated cycles of delayed feedback may cause the control system to provide suboptimal amounts of reactant for an extended period of time before converging to the optimal amount. Advantageously, systems and methods according to some aspects of the present disclosure can predict optimal reactant amounts more quickly than some alternative control systems.

[0025] As another example, some industrial systems (e.g., turbine systems, etc.) and selective catalytic reduction systems may change their operational behavior over time for a variety of reasons, such as equipment aging, equipment maintenance or replacement, system modifications, system operator behavior, etc. Furthermore, individual industrial systems and selective catalytic reduction systems may behave differently from other similar systems. Some alternative methods may not be able to account for differences between systems or within a particular system over time. Advantageously, systems and methods according to some aspects of the present disclosure can predict optimal flow values ​​for a particular system and can adapt to changes in that system's operational behavior over time. In this way, for example, prediction accuracy can be improved, providing more optimal reactant flow rates for a particular system over time. The more optimal reactant flow rates can then be calculated based on emissions (e.g., ammonia slip, NO ). x emissions), improved regulatory compliance, and improved compliance costs.

[0026] Systems and methods according to some aspects of the present disclosure may also provide improved capabilities for computing systems that execute machine learning models. For example, in some cases, the provided systems and methods may predict optimal reactant flow rates using less training data compared to some alternative methods. For example, in some cases, a training data structure including only 32 data examples may be sufficient to model complex nonlinear relationships between input variables (e.g., turbine load) and optimal flow rates. Predicting optimal flows based on less training data may provide various technical effects and benefits, such as reduced memory footprint, reduced data collection costs, and shorter data collection periods. For example, shorter data collection periods may enable machine learning control systems to be installed or calibrated more quickly compared to some alternative methods, thereby resulting in faster improvements in emissions and regulatory compliance.

[0027] Furthermore, systems and methods according to certain exemplary aspects of the present disclosure may provide machine learning inference at reduced computational costs (e.g., memory usage, processor usage, power costs, etc.). For example, some alternative machine learning methods may use neural networks. In some cases, performing inference with a neural network may require a large number (e.g., thousands, millions, billions, etc.) of floating-point operations associated with a large number of neural network parameters. In contrast, exemplary systems and methods according to certain aspects of the present disclosure may use a small number of operations to perform inference based on a small number of acquired data examples. In this manner, for example, the computational cost of machine learning inference may be reduced, and the functionality of the computing system itself may be improved.

[0028] As another example, some alternative methods use machine learning methods to generate predicted emissions (e.g., NO) as output based on reactant flow rates as input. xAlternatively, alternative methods may be used to predict optimal reactant flow rates (e.g., emission predictions). However, determining optimal reactant flow rates using such alternative methods may require multiple machine learning inference calculations. For example, determining optimal reactant flow rates may require selecting multiple candidate flow rates, predicting multiple emissions associated with the multiple candidate flow rates, and determining the amount of reactant to be supplied to the selective catalytic reduction system based on a comparison between the multiple emission predictions. In contrast, exemplary systems and methods according to certain aspects of the present disclosure may directly predict the optimal flow rates using only one machine learning inference calculation, thereby reducing the computational cost (e.g., processor usage, power costs, etc.) of selecting optimal reactant flow rates. In this manner, for example, the computational cost of machine learning inference may be reduced, and the functionality of the computing system itself may be improved compared to some alternative methods.

[0029] Exemplary System Referring now to the drawings, Figure 1 shows a block diagram of a system for selective catalytic reduction using machine learning according to an embodiment of the present disclosure. A machine learning system 104 can receive input data 102 and can output one or more reactant quantity predictions 106 based on the input data 102. Based on the reactant quantity predictions 106, a reactant flow control system 108 can provide an amount of reactant 110 to a selective catalytic reduction system 112 of an industrial system 114. The selective catalytic reduction system 112 can process the effluent of the industrial system 114 to produce a catalytic reduction effluent 116, which is exhausted via an exhaust system 118.

[0030] The input data 102 may include, for example, any data relevant to estimating optimal reactant flow rates for the selective catalytic reduction system 112. For example, the input data 102 relevant to a turbine-based industrial system 114 may include turbine load, flue gas temperature, duct burner fuel flow rate, ambient conditions (e.g., temperature, etc.), emissions sensor data (e.g., NO xThe input data 102 may include data indicative of turbine load, reactant slip data, etc., start-up or shutdown status data or other turbine status data, or any other relevant data. The input data 102 associated with other systems may include similar types of data (e.g., load, temperature, fuel flow, emissions data, etc.) or different types of data. The input data 102 may include snapshot data from a single point in time or may include time-series data from multiple points in time. The input data 102 may be one type of data or multiple types of data. For example, in some cases, the input data 102 may include numeric (e.g., floating point, etc.) data indicative of one or more numeric values ​​(e.g., turbine load, flue gas temperature, fuel flow, ambient temperature, emissions, etc.).

[0031] The machine learning system 104 can include various machine learning architectures, and can include one or more types of machine learning architectures. The machine learning system 104 can use parametric learning techniques (e.g., neural networks, etc.) or non-parametric learning techniques (e.g., nearest neighbors, etc.), sequence-based techniques (e.g., attention-based techniques, convolutional techniques, recursive or long short-term memory techniques, etc.) to predict optimal reactant flow rates based on time-series data, or non-sequence techniques to predict reactant flow rates based on input data 102 from a single time. In some cases, the machine learning system 104 can include a learning-based machine learning system, where relevant data (e.g., historical data correlating past input data 102 with past optimal reactant flow rates) can be learned by the machine learning system during inference. Exemplary details of the exemplary machine learning system 104 are further described below with respect to FIG. 3.

[0032] The reactant quantity prediction 106 may be, for example, a machine learning prediction of an optimal amount of reactant 110 to be provided to the selective catalytic reduction system 112. In some cases, the reactant quantity prediction 106 may include a final reactant flow rate value, and the reactant flow control system 108 may provide an amount of reactant 110 equal to or approximately equal to the reactant quantity prediction 106. In some cases, the reactant quantity prediction 106 may include an intermediate value, which may be further processed to determine the final amount of reactant 110 to be provided. For example, in some cases, the selective catalytic reduction system 112 may behave differently during a startup or shutdown process (e.g., the last hour before shutdown, the first hour after startup, etc.) compared to other times. In such cases, the reactant quantity prediction 106 may include an intermediate value indicative of an optimal reactant flow rate for non-startup / shutdown conditions, and the reactant quantity prediction 106 may be further processed to determine the amount of reactant 110 to be provided during the startup or shutdown process. For example, the reactant quantity prediction 106 can be adjusted according to a reactant flow rate adjustment value associated with the start-up or shutdown process, and the reactant quantity prediction 106 can be modified based on the adjustment value (e.g., by adding, subtracting, multiplying, dividing, or performing another operation in response to the adjustment value). In some cases, the reactant quantity prediction 106 can be multiplied by a reactant flow rate multiplier associated with the start-up or shutdown process to provide an amount of reactant 110 equal to the result of the multiplication. However, this is not required. For example, in some cases, start-up / shutdown condition data can be provided as input data 102, and the machine learning system 104 can output a final reactant quantity prediction 106 that takes the start-up / shutdown condition into account.

[0033] The reactant flow control system 108 can include, for example, any system for directly or indirectly controlling the flow of a fluid. As non-limiting illustrative examples, the reactant flow control system 108 can include one or more distributed control system (DCS) devices, one or more programmable logic controller (PLC) devices or human-machine interface (HMI) devices, one or more supervisory control and data acquisition (SCADA) devices, proportional-integral-derivative (PID) control devices, or other control devices. In some cases, the reactant flow control system 108 can include one or more computing devices or computing systems (e.g., a system such as that described below with respect to FIG. 7). In some cases, the reactant flow control system 108 can include one or more hardware devices (e.g., valves, actuators, etc.) for controlling the flow of a fluid. In some cases, the one or more hardware devices can be controlled by another component of the reactant flow control system (e.g., a DCS device or another computing device, etc.). In some cases, reactant flow control system 108 may include a computing device (e.g., a DCS device, a PLC device, an HMI device, etc.) configured to provide a signal (e.g., an internal signal of the computing device, an external signal, etc.) to cause an amount of reactant 110 (e.g., an amount equal to reactant quantity prediction 106, etc.) to be provided to selective catalytic reduction system 112. The signal may include, for example, an internal signal to a component of the computing device (e.g., from a processor to a hardware component of the computing device, etc.), an external signal to another device (e.g., a hardware device such as a valve, actuator, etc., a computing device such as a DCS controller, a PLC controller, etc.), or other signal. The signal may be provided via any suitable connection, such as an internal connection (e.g., a bus, an interconnect, a transmission line, etc.) or an external connection (e.g., a network, a wire, a cable, a port, an input / output device, etc.) of the computing device.

[0034] Reactant 110 may include, for example, any reactant (e.g., reactant fluid) that can be supplied to selective catalytic reduction system 112. In some cases, reactant 110 may include nitrogen oxides (NO x ) into nitrogen (N2) and water (HO). Exemplary reactants for converting nitrogen oxides include, for example, ammonia (e.g., anhydrous ammonia, aqueous ammonia, etc.), urea, solutions thereof (e.g., diesel exhaust fluid, etc.), or exhaust (e.g., NO x The reaction system may include any suitable reactants for treating the reaction mixture (e.g., effluent).

[0035] The selective catalytic reduction system 112 may, for example, combine reactants 110 with one or more types of emissions (e.g., NO x The selective catalytic reduction system 112 may include any system for using a catalyst to reduce or convert reactants 110 (e.g., effluents). In some cases, the selective catalytic reduction system 112 may include a catalyst bed or a catalyst chamber. In some cases, the reactants 110 may be deposited on or in a catalyst bed or catalyst chamber, and a flue gas or exhaust gas stream may pass through the bed or chamber to react with the reactants 110.

[0036] Industrial system 114 may include, for example, any system that may include or be associated with selective catalytic reduction system 112. Exemplary industrial systems may include, for example, any system that may include a combustion component (e.g., a gas turbine, a boiler, an incinerator, an internal combustion engine such as a diesel engine, etc.).

[0037] The catalytic reduction effluent 116 may include, for example, any output (e.g., a fluid such as a gas) processed by the selective catalytic reduction system 112 before being removed (e.g., emitted, vented, captured, etc.) from the industrial system 114. An exemplary catalytic reduction effluent 116 may include, for example, a proportion of nitrogen (N) and water (H0), which may have been converted from other input gases (e.g., NO, NO, etc.).

[0038] Exhaust system 118 may include, for example, any system for emitting catalytically reduced emissions 116 from industrial system 114 (eg, an exhaust stack in an industrial plant, a vehicle exhaust system, etc.).

[0039] 2 is a block diagram of a system for training a machine learning system for controlling a selective catalytic reduction system according to an embodiment of the present disclosure. A self-adjusting reactant flow control system 220 can receive sensor data 222 from one or more sensors 224 and can adjust the flow rate of reactant 110 in response to the sensor data 222. Once the self-adjusting reactant flow control system 220 converges on an optimal flow rate for a particular set of input conditions, reactant quantity data 226 indicative of the optimal flow rate can be provided to the machine learning system 104 along with input data 102 indicative of the input conditions. The machine learning system 104 can then be trained based on the received data such that the machine learning system 104 can learn to predict the optimal flow rate based on the input data.

[0040] The self-regulating flow control system 220 may include any system for adjusting the flow rate of the reactant 110 in response to data, including sensor data 222. In some cases, the self-regulating flow control system 220 may be, include, or be included in the reactant flow control system 108. In some cases, the self-regulating flow control system 220 may include a proportional-integral-derivative (PID) controller or similar controller for adjusting the flow rate of the reactant 110 in response to the sensor data 222. For example, the PID controller may adjust the flow rate of the reactant 110 in response to one or more variables of the sensor data 222 (e.g., NO xThe sensor data 222 may be configured with one or more desired set points describing target values ​​(e.g., about 2 parts per million, about 5 parts per million, etc.) of the reactants 110 (emissions data, such as emissions, ammonia slip, etc.) and may continuously adjust the flow of reactants 110 based on the difference between the current value of the variable in the sensor data 222 and the target value. For example, the PID controller may select adjusted reactant flow values ​​based on an equation such as:

number

[0041] The sensor data 222 may include, for example, any data that may be relevant in adjusting the amount of reactant 110 provided to the selective catalytic reduction system. For example, in some cases, the sensor data 222 may include emissions data (e.g., NO xThe sensor data 222 may include other data such as temperature data, pressure data, leak sensor data, exhaust flow data, or other related sensor data 222. The sensor data 222 may include one or more types of data. For example, in some cases, the sensor data 222 may include numeric or binary data (e.g., floating point numeric data, etc.) indicative of one or more numerical characteristics being sensed (e.g., emissions concentration in parts per million, etc.).

[0042] The sensor 224 may, for example, detect one or more aspects of the environment (e.g., NO x The sensor 224 may include any device or system configured to sense one or more aspects of the exhaust system 118 (e.g., carbon dioxide, nitrogen, etc.) and provide sensor data 222 (e.g., emissions data, etc.) describing the one or more aspects of the exhaust system 118 (e.g., near the top of the exhaust stack of an industrial plant, etc.).

[0043] The reactant quantity data 226 may include data indicative of the amount of reactant 110 actually provided by the self-regulating reactant flow control system 220 under the circumstances (e.g., after an adjustment period). For example, the self-regulating reactant flow control system 220 may iteratively adjust the flow rate of the reactant 110 based on the sensor data 222 (e.g., according to a PID equation, etc.) until it converges to a steady-state flow rate of the reactant 110 (e.g., according to one or more exhaust setpoints, etc.). In some cases, the steady-state flow rate may include a flow rate that has not changed (e.g., not at all) for a particular period of time (e.g., several seconds, etc.). In some cases, the steady-state flow rate may include a flow rate that has not changed significantly over a particular period of time (e.g., a predetermined time window, a dynamically determined time window, etc.). For example, in some cases, a change threshold may be defined, and the flow rate may be determined to be at a steady state if the change in flow rate (e.g., net change, distance from maximum peak to trough, etc.) over a particular period of time is less than the threshold. After reaching the steady-state flow rate, the self-adjusting reactant flow control system 220 can provide reactant quantity data 226 indicative of the steady-state flow rate to the machine learning system 104 for training or other learning. In this way, for example, the machine learning system 104 can learn the true steady-state value selected by the self-adjusting flow control system 220, which may correspond to an optimal or preferred flow rate in some cases based on the sensor data 222. The reactant quantity data can include one or more types of data. In some cases, the reactant quantity data 226 can include numeric (e.g., floating point, etc.) or binary data indicative of the numeric flow rate of the reactant 110.

[0044] Based on the reactant quantity data 226, the machine learning system 104 can be trained to generate reactant quantity predictions 106 that predict future reactant quantity data 226. Training the machine learning system 104 based on the reactant quantity data 226 can include any process that uses the reactant quantity data 226 to modify one or more future reactant quantity predictions 106. For example, in some cases, training a parameterized machine learning system 104 (e.g., a neural network, etc.) can include any process for updating one or more parameters of the parameterized machine learning system 104 based on the reactant quantity data 226. As another example, training an acquisition-based machine learning system 104 (e.g., a neighborhood-based acquisition system, etc.) can include storing the reactant quantity data 226 in a data structure from which the acquisition-based machine learning system 104 acquires the data at inference time. For example, training the machine learning system 104 may include correlating the reactant quantity data 226 with corresponding input data 102 describing one or more circumstances under which the reactant quantity data 226 was generated (e.g., turbine load, ambient conditions, flue gas temperature, duct burner fuel flow rate, exhaust sensor data, start-up or shutdown conditions, etc.). Training the machine learning system 104 may further include storing the correlation data in a data structure that the machine learning system 104 uses at inference time to generate reactant quantity predictions 106. In this manner, for example, the reactant quantity data 226 may be used to modify one or more future reactant quantity predictions 106, thereby training the machine learning system 104. Further details of an exemplary method for training the exemplary machine learning system 104 based on the reactant quantity data 226 are provided further below with respect to FIG. 3 .

[0045] FIG. 3 is a block diagram of an exemplary machine learning system for controlling a selective catalytic reduction system, according to an embodiment of the present disclosure. The machine learning system 304 can receive input data 102 and can retrieve one or more stored example data 328 based on the input data 102. For example, in some cases, a nearest neighbor acquisition 330 system can retrieve stored example data 328 associated with one or more operating conditions similar to the operating conditions associated with the input data 102. Based on the retrieved neighbor data 332, the machine learning system 304 can perform neighborhood-based inference 334 to determine a reactant quantity prediction 106. In some cases, the reactant quantity prediction 106 can be provided to the reactant flow control system 108 shown in FIG. 1. In some cases, the machine learning model 304 can be trained based on reactant quantity data 226 as shown in FIG. 2. For example, the machine learning model 304 can be provided with the reactant quantity data 226 as shown in FIG. 2, and a comparison system 338 can be used to compare the reactant quantity data 226 to the reactant quantity prediction 106. Based on the comparison, the machine learning system 304 can determine whether to add the reactant amount data 226 to the stored example data 328. For example, if the reactant amount data 226 is “surprising” to the machine learning system 304 (e.g., different from the reactant amount predictions 106 expected by the machine learning system 304), a surprising example 340 can be added to the stored example data 328. Furthermore, in some cases, the machine learning system can perform selective forgetting, where one or more stored example data 328 can be deleted under various circumstances (e.g., in response to a surprising example 340 being added, etc.).

[0046] In some cases, machine learning system 304 may be identical to, include, be included in, or share one or more characteristics of machine learning system 104. In some cases, machine learning system 304 may include some or all of the components shown in Figure 3. However, Figure 3 shows one example configuration of machine learning system 304, and other configurations may be used as well. For example, individual components shown may be omitted, rearranged, or added without departing from the scope of the present disclosure.

[0047] The example stored data 328 can include one or more data structures (e.g., data collections such as databases, arrays, lists, stacks, data objects, files, folders, or other file-based data structures) containing multiple example data. Each example stored data 328 can include, for example, data correlating one or more input values ​​(e.g., input data 102, etc.) with data indicative of reactant quantities (e.g., optimal or steady-state reactant quantities, reactant quantity data 226, etc.). The example stored data 328 can include one or more types of data, such as numeric data (e.g., integer, floating point, quantized numeric data, etc.), binary data, raw sensor data, or any other data format.

[0048] In some cases, the data structure for storing the stored data examples 328 may include a fixed-size or variable-size data structure for storing a fixed or variable number of data examples. For example, a data structure correlating one input variable with reactant amounts may include a small, fixed number (e.g., 8, 16, 32, 48, 64, etc.) of data examples. As another example, a data structure correlating a plurality of n input variables may include a small, fixed number (e.g., 16n, 32n, 16 * 2 n , 8 * 4 n , etc.)

[0049] In some cases, a fixed-size data structure can be used in combination with the selective forgetting process. For example, in some cases, the data structure can have a fixed maximum number of data examples. Once a surprising example 340 is identified, the computing system can determine whether the number of stored data examples 328 already equals the maximum number. If so, the computing system can delete the previously stored data example 328 before adding the new surprising example 340. In some cases, the selection of which stored data example 328 to delete can depend on one or more of the age of each stored data example 328, the proximity or similarity of the stored data example 328 to the surprising example 340 (e.g., according to a vector distance measure, etc.), whether each stored data example 328 was part of the obtained neighborhood data 332, or other suitable factors. For example, the stored data instance 328 selected for deletion may include the oldest entire stored data instance 328 (e.g., throughout the entire data structure of stored data instances 328), the oldest retrieved neighbor 332, the closest retrieved neighbor 332, or other suitable stored data instance 328.

[0050] In some cases, the size of a fixed-size data structure can be selected in combination with a selective forgetting process to optimize the performance of the machine learning system 304. In some cases, the optimal data structure size of the stored data examples 328 may depend on the complexity of the relationships between the input variables and reactant amounts, the measurement range of the input variables, and other relevant factors. In some cases, the appropriate scaling of the data structure size may depend on the level of correlation or independence among the n input variables.

[0051] In some cases, variable-sized data structures can be used in combination with a selective forgetting process. The decision whether to delete stored data examples 328 may depend, for example, on the age of each stored data example 328, the proximity of each stored data example 328 to newly added surprising examples 340, or other suitable factors.

[0052] In some cases, a stored data example 328 can be removed based on similarity to its immediate neighbors. For example, if the stored data example 328 includes reactant quantity data 226 that can be accurately predicted by the machine learning system 304 without retaining the stored data example 328, the stored data example 328 can be deleted (e.g., from a fixed-size data structure, from a variable-size data structure, etc.). As an illustrative example, if the neighborhood-based inference 334 (e.g., as described further below) operates by performing linear interpolation between nearby data points, the stored data example 328 can be safely deleted if it is sufficiently close to a line interpolating between two nearby stored data examples 328. In some cases, determining whether to delete the stored data example 328 can include determining reactant quantity predictions 106 associated with the input data 102 of the stored data example 328, comparing the reactant quantity predictions 106 to the reactant quantity data 226 of the stored data example 328, and determining whether to delete the stored data example 328 based on the comparison. In some cases, comparing the reactant amounts 106, 226 may include determining an absolute value of the difference between the reactant amounts 106, 226. In some cases, determining whether to delete the stored data example 328 may include comparing the absolute value of the difference to a difference threshold, and deleting the stored data example 328 if the difference is less than the difference threshold. In some cases, if two or more stored data examples 328 can be safely deleted in this manner, the oldest data example of the two or more stored data examples 328 may be deleted first.

[0053] Nearest neighbor acquisition 330 may include any acquisition of one or more (e.g., multiple) stored data examples 328 based on input data 102. In some cases, nearest neighbor acquisition 330 may include acquisition of multiple stored data examples 328 based on similarity between input data 102 and one or more corresponding data fields of stored data examples 328. As a simplified illustrative example, input data 102 may include data indicative of turbine load, and stored data examples 328 may include data correlating turbine load to reactant quantity. In such an example, nearest neighbor acquisition 330 may acquire stored data examples 328 associated with turbine loads similar (e.g., numerically close, etc.) to the turbine load of input data 102. In some cases, nearest neighbor acquisition 330 may acquire one or more nearest neighbors in each of two or more directions (e.g., nearest neighbors with higher turbine loads and nearest neighbors with lower turbine loads, etc.). In some cases, nearest neighbor acquisition 330 can acquire close neighbors regardless of direction (eg, two nearest neighbors, k nearest neighbors, etc.).

[0054] In some cases, the input data 102 may include multiple input variables (e.g., turbine load, flue gas temperature, duct burner fuel flow rate, ambient conditions, exhaust sensor data, start-up or shutdown condition data, etc.), and the nearest neighbor acquisition 330 may determine similarity based on some combination of the multiple input variables. For example, in some cases, the input data 102 may include multiple numeric input variables or input variables that can be converted to numeric values ​​(e.g., Boolean variables, etc.). In such cases, the input data 102 may be treated as a vector, and the nearest neighbor acquisition 330 may be based on a vector-based distance metric (e.g., cosine distance, Euclidean distance, Manhattan distance, etc.) between the input data 102 and each stored data instance 328.

[0055] In some cases, one or more input variables of the input data 102 may be normalized (e.g., to a common scale, such as a 0 to 1 scale) before calculating the distance metric. In some cases, normalization may include identifying a minimum and a maximum value for each input variable of the input data 102 and rescaling the variables so that the minimum value corresponds to a common minimum value (e.g., 0) and the maximum value corresponds to a common maximum value (e.g., 1) that may be common to all input variables of the input data 102. As an illustrative example, a temperature range may include a minimum temperature of 550 degrees Fahrenheit and a maximum temperature of 700 degrees Fahrenheit. In such a case, normalizing the temperature range to a 0 to 1 scale may include subtracting 550 degrees (the identified minimum value minus 0) from the temperature values ​​of the input data 102, and then dividing by 150 (the difference between the maximum and minimum values ​​divided by 1). The minimum and maximum values ​​may include true minimum or maximum values ​​(e.g., theoretically or physically defined minimum or maximum values ​​such as 0%, 100%, 0 Kelvin, etc.) or values ​​selected for use as minimum or maximum values ​​(e.g., values ​​corresponding to possible actual minimum or maximum values). If the minimum or maximum value is not a true minimum or maximum, input data 102 having a value below the minimum or above the maximum value may be treated as equal to the minimum or maximum value for purposes of normalization.

[0056] The retrieved neighborhood data 332 may include one or more (e.g., multiple) stored data examples 328 retrieved by nearest neighbor acquisition 330. In some cases, the number of retrieved stored data examples 328 may be a predetermined or fixed number (e.g., k nearest neighbors). In some cases, the number of stored data examples 328 to retrieve may be dynamically determined. For example, in some cases, the number of retrieved stored data examples 328 may be determined based on a similarity threshold (e.g., a vector distance threshold, etc.). In some cases, each stored data example 328 having a similarity above the similarity threshold (e.g., a vector distance below the vector distance threshold, etc.) may be retrieved. Other methods are also possible.

[0057] The neighborhood-based inference 334 may include any method for combining the obtained neighborhood data 332 to generate the reactant quantity prediction 106. For example, various statistical and machine learning methods, such as parametric, nonparametric, or semiparametric regression, linear, polynomial, or other interpolation, may be used to combine the data examples to generate the prediction. As a simplified illustrative example, a machine learning system 304 that generates a reactant quantity prediction 106 based on one input variable may receive the input value, obtain two stored data examples 328 associated with similar input values, and perform linear interpolation between the two stored data examples 328 to generate the reactant quantity prediction 106. As another example, a machine learning system 304 that predicts an optimal flow based on two or more input variables may receive the input data 102, obtain a plurality of stored data examples 328 associated with similar input values ​​(e.g., at least one more than the number of input variables, etc.), and perform linear regression based on the plurality of stored data examples 328 to generate the reactant quantity prediction 106. Other methods of combining the obtained neighborhood data 332 to generate the reactant quantity prediction 106 are possible (eg, averaging, weighted averaging such as distance-weighted averaging, non-linear methods, etc.).

[0058] The comparison system 338 can include any system for comparing the reactant amount predictions 106 with the reactant amount data 226 and determining, based on the comparison, whether to store a new stored data example 328. In some cases, the comparison can include determining a measure of similarity between the reactant amount predictions 106 and the reactant amount data 226 (e.g., the absolute value of the difference between the two reactant amounts 106, 226, etc.). In some cases, determining whether to store the data can include comparing the difference between the two reactant amounts 106, 226 to a difference threshold and storing a surprising example 340 if the absolute value of the difference is greater than the difference threshold. Other methods are also possible.

[0059] Surprising example 340 may include, for example, any data example that comparison system 338 selects for storage. Surprising example 340 may be identical to stored data example 328, may include stored data example 328, may be contained by stored data example 328, or may share one or more characteristics of stored data example 328.

[0060] 4 is a block diagram of a system for controlling a selective catalytic reduction system using machine learning while continuously training the machine learning control system during operation, according to an embodiment of the present disclosure. The machine learning system 104 can provide reactant quantity predictions 106 to a self-adjusting reactant flow control system 220, which can provide an amount of reactant 110 to the selective catalytic reduction system 112 based on the reactant quantity predictions 106 (e.g., as shown in FIG. 1 ). The sensor 224 can then provide sensor data 222 to the self-adjusting reactant flow control system 220, which can adjust the flow rate of the reactant 110 based on the sensor data 222. After convergence to a steady-state flow rate, the self-adjusting flow control system 220 can provide reactant quantity data 226 to the machine learning system 104 (e.g., as shown in FIG. 2 ). The machine learning system 104 can then train based on the new reactant quantity data 226 (e.g., as shown in FIG. 3 ). In this way, for example, the machine learning system 104 can continuously learn during operation of the industrial system, even while the output of the machine learning system 104 is being used to control the flow of reactants 110.

[0061] In some cases, the amount of reactant 110 supplied to the selective catalytic reduction system 112 may be held constant throughout a wait period before the self-adjusting reactant flow control system 220 begins adjusting the amount of reactant 110 (e.g., based on sensor data 222, etc.). For example, in some cases, operational changes in the industrial system 114 (e.g., industrial turbine system 414) may occur based on sensor data 222 (e.g., NO xFor example, a change in an operating variable (e.g., turbine load, flue gas temperature, duct burner fuel flow rate, ambient conditions, start-up or shutdown conditions, etc.) can cause a time-delayed change in emissions at the top of the exhaust stack 418. In some cases, this time-delayed change may not occur until several minutes (e.g., four minutes) after the operating variable is changed. In such cases, the amount of reactant 110 based on the reactant quantity prediction 106 can be provided for the first few minutes after the operating variable change, and then the amount of reactant 110 can be automatically adjusted by the self-tuning reactant flow control system 220 (e.g., based on the sensor data 222) after a waiting period has expired (e.g., after sufficient time has passed for the effects of the change in the operating variable to be seen in the sensor data 222). After the wait period ends, the self-adjusting reactant flow control system 220 can continue to adjust the flow rate of the reactant 110 until it converges on a steady-state flow rate (e.g., as described with respect to FIG. 2) or until another operational change is made, thereby triggering another wait period that can hold the amount of reactant 110 stable according to the new reactant quantity prediction 106.

[0062] 4, the selective catalytic reduction system 112 may be a component of an industrial turbine system 414 that includes a gas turbine 442, a heat recovery steam generator 444, an exhaust stack 418, etc. However, other industrial systems 114 are possible. The industrial turbine system 414 may be, comprise, or be provided by the industrial system 114. The exhaust stack 418 (e.g., a chimney, exhaust stack, etc.) may be, comprise, or be provided by the exhaust system 118.

[0063] The gas turbine 442 and the heat recovery steam generator 444 may comprise any type of turbine or steam generation system. For example, the gas turbine 442 may include a compressor section, a combustion section, a turbine section, and an exhaust section. The compressor section gradually increases the pressure of a working fluid entering the gas turbine engine and supplies the compressed working fluid to the combustion section. The compressed working fluid and fuel (e.g., natural gas) are mixed in the combustion section and combusted in a combustion chamber to generate high-pressure, high-temperature combustion gases. The combustion gases flow from the combustion section to the turbine section, where they expand to generate work. For example, the expansion of the combustion gases in the turbine section may rotate a rotor shaft connected to, for example, a generator, generating electricity. The combustion gases then exit the gas turbine via the exhaust section. The heat recovery steam generator 444 may include, for example, any system for generating steam from a heat source (e.g., exhaust from the gas turbine 442). In some cases, the heat recovery steam generator 444 may include a selective catalytic reduction system 112 (eg, at a location selected to optimize the temperature at which the catalytic reaction occurs).

[0064] Exemplary Methods 5 is a flowchart diagram of an exemplary method for controlling selective catalytic reduction using machine learning, according to an exemplary embodiment of the present disclosure. While FIG. 5 depicts steps performed in a particular order for purposes of illustration and explanation, the methods of the present disclosure are not limited to the particularly depicted order or arrangement. Various steps of the exemplary method 500 may be omitted, rearranged, combined, and / or adapted in various ways without departing from the scope of the present disclosure.

[0065] At 502, the example method 500 may include obtaining, by a computing system including the machine learning model, input data (e.g., input data 102) including one or more input values. In some cases, the example method 500 may include using one or more systems or performing one or more activities described with respect to Figures 1, 3, or 4.

[0066] At 504, the example method 500 may include, based on the input data, obtaining, by the computing system, two or more example data (e.g., stored example data 328, obtained neighborhood data 332, etc.) indicative of previous reactant amounts (e.g., reactant amount data 226, etc.). In some cases, the example method 500 may include, at 504, using one or more systems or performing one or more activities described with respect to Figures 1, 3, or 4.

[0067] At 506, the example method 500 may include generating output data (e.g., reactant amount prediction 106, etc.) indicative of an amount of a reactant (e.g., reactant 110, etc.) by the machine learning model based on the two or more example data. In some cases, the example method 500 may include using one or more systems or performing one or more activities described with respect to Figures 1, 3, or 4.

[0068] At 508, the example method 500 may include providing, by a computing system, a signal to cause the amount of reactant to be provided to a selective catalytic reduction system (e.g., selective catalytic reduction system 112). In some cases, the example method 500 may include, at 508, using one or more systems or performing one or more activities described with respect to FIG. 1, FIG. 3, or FIG. 4.

[0069] 6 is a flowchart diagram of an exemplary method for training a machine learning system (e.g., 104, 304, etc.) according to an exemplary embodiment of the present disclosure. While FIG. 6 depicts steps performed in a particular order for purposes of illustration and explanation, the methods of the present disclosure are not limited to the particularly depicted order or arrangement. Various steps of the exemplary method 600 may be omitted, rearranged, combined, and / or adapted in various ways without departing from the scope of the present disclosure.

[0070] At 602, the example method 600 may include obtaining, by a self-regulating reactant flow control system, emissions data indicative of one or more emissions. In some cases, the example method 600 may include using one or more systems or performing one or more activities described with respect to FIG. 2 or FIG. 4.

[0071] At 604, the example method 600 may include adjusting, by a self-adjusting reactant flow control system, a first amount of reactant supplied to the selective catalytic reduction system based on the emissions data. In some cases, the example method 600 may include using one or more systems or performing one or more activities described with respect to FIG. 2 or FIG. 4 at 604.

[0072] At 606, the example method 600 may include monitoring the adjustment by a computing system comprising one or more computing devices (e.g., computing system 702, etc.) In some cases, the example method 600 may include using one or more systems or performing one or more activities described with respect to Figure 2 or Figure 4 at 606.

[0073] At 608, the example method 600 may include determining that the first amount of the reactant supplied to the selective catalytic reduction system has stabilized. In some cases, the example method 600 may include using one or more systems or performing one or more activities described with respect to FIG. 2 or FIG. 4 at 608.

[0074] At 610, the example method 600 may include training a machine learning system (e.g., machine learning system 104, 304) by a computing system using one or more example data (e.g., stored example data 328) including data indicative of the stabilized first quantity (e.g., reactant quantity data 226). In some cases, the example method 600 may include using one or more systems or performing one or more activities described with respect to Figures 2, 3, or 4 at 610.

[0075] Exemplary Computing Systems and Devices 7 is a block diagram of an exemplary computing system. Computing system 702 can include one or more computing devices 704, each of which can include a processor device 706, a memory device 712, a storage device 714, or an input / output device 716. Computing device 704 can include one or more machine learning models 718 (e.g., machine learning system 104, machine learning system 304), or portions thereof, which can be located, for example, in storage device 714 or memory device 712. Computing system 702 can be connected via network 720 to one or more other systems, such as one or more industrial input systems 722 (e.g., systems for acquiring or providing input data 102, sensors 224 for providing sensor data 222 input, etc.), computing systems 724 (e.g., client computing systems, third-party computing systems, computing systems for controlling or monitoring industrial processes, etc.), or industrial control systems 726 (e.g., reactant flow control system 108, etc.).

[0076] The computing system 702 may include any number of computing devices 704, such as one computing device 704 or many computing devices 704. The computing devices 704 may include any type of computing device, such as a DCS controller, a PLC controller, a PID controller, a server, a workstation, a desktop, a laptop, a virtual machine, a mobile device, or other computing device.

[0077] Processor 706 may include, for example, one or more central processing units (CPUs) 708 and one or more application specific integrated circuits (ASICs) 710, such as an ASIC for performing floating-point operations (e.g., a GPU, etc.), an ASIC for performing matrix multiplication, an ASIC for performing machine learning or artificial intelligence tasks, or other ASICs. CPU 708 may comprise, for example, any hardware configured to operate as a CPU (e.g., a microprocessor, microcontroller, soft-core processor, etc.).

[0078] The memory device 712 may include, for example, one or more non-transitory computer-readable storage media for temporarily storing data (e.g., to facilitate faster data access from the storage device 714). The temporary storage medium may include, for example, one or more memory devices such as high-bandwidth memory, random access memory (e.g., RAM, DRAM, SDRAM, DDR SRAM, etc.), virtual memory, cache memory, etc. The memory device 712 may include, for example, volatile memory, non-volatile memory, and semi-volatile memory.

[0079] The storage device 714 may include, for example, one or more non-transitory computer-readable storage media for persistent storage of data (e.g., including when the computing device 704 is powered down). A persistent storage device may include a non-volatile storage device such as read-only memory (e.g., ROM, PROM, EPROM, EEPROM, etc.), flash memory (e.g., NAND flash memory, etc.), magnetic storage device (e.g., hard disk drive, floppy disk, etc.), optical storage device (e.g., CD, DVD, Blu-Ray, etc.), or other non-volatile memory. In some cases, the storage device 714 may include volatile or semi-volatile memory coupled to a continuous power source (e.g., power grid power, backup battery, etc.) provided to semi-volatile memory to preserve data when the computing device 704 is shut down.

[0080] The input / output devices 716 may include, for example, any device or component that receives input from or provides output to a device, system, person, or other entity other than the computing device 704. The input / output devices 716 may include, for example, a network connection, a network card or adapter for communicating over a network connection, human input / output devices such as a keyboard, a mouse, a display monitor, speakers, a camera, a microphone, or other input / output devices. In some cases, the input / output devices 716 may include one or more sensors, such as sensors 224, associated with the industrial system 114. In some cases, the input / output devices 716 may include one or more devices for communicating with such sensors.

[0081] The machine learning model 718 may include, for example, example stored data 328 of the data or acquisition-based machine learning system 304 that stores one or more parameters of the machine learning model, computer-readable instructions (e.g., source code, object code, etc.) that, when executed by one or more processors 706, cause the processor(s) to perform one or more operations of the machine learning model, or any other data or components related to the machine learning model (e.g., machine learning system 104, etc.).

[0082] Network 720 may be or include, for example, the Internet or any other network configured to transfer computer-readable data between computing devices (e.g., a local area network, a wide area network, a peer-to-peer network, etc.). Network 720 may include, for example, wired connections, wireless connections, or a combination of both wired and wireless connections. In some cases, network 720 may be associated with one or more communication protocols for communicating over network 720, such as Transmission Control Protocol (TCP), Internet Protocol (IP), Hypertext Transfer Protocol (HTTP), User Datagram Protocol (UDP), Border Gateway Protocol (BGP), Address Resolution Protocol (ARP), etc. In some cases, network 720 may be associated with one or more security protocols for secure communication over network 720, such as Secure Sockets Layer (SSL) protocol or Transport Layer Security (TLS) protocol.

[0083] Industrial input system 722 may include any type of computing device, such as a workstation, a server, a laptop, a desktop, a mobile device, a virtual device, or other computing system. Industrial input system 722 may include, for example, a client computing system, a server computing system, or a third-party computing system. In some cases, industrial input system 722 may include any component or have any of the characteristics described above with respect to computing device 704. In some cases, industrial input system 722 may include a single computing device or multiple computing devices.

[0084] Computing system 724 may include any type of computing device, such as a workstation, a server, a laptop, a desktop, a mobile device, a virtual device, or other computing system. Computing system 724 may include, for example, a client computing system, a server computing system, or a third-party computing system. In some cases, computing system 724 may include any component or have any of the characteristics described above with respect to computing device 704. In some cases, computing system 724 may include a single computing device or multiple computing devices.

[0085] The industrial control system 726 may include any type of computing device, such as a PLC controller, a DCS controller, a PID controller, a workstation, a server, a laptop, a desktop, a mobile device, a virtual device, or other computing system. The industrial control system 726 may include, for example, a client computing system, a server computing system, or a third-party computing system. In some cases, the industrial control system 726 may include any component or have any of the characteristics described above with respect to the computing device 704. In some cases, the industrial control system 726 may include a single computing device or multiple computing devices.

[0086] 7 illustrates one exemplary configuration of a computing system that can be used to implement the present disclosure. Other computing system configurations can be used as well. For example, individual components shown can be omitted, rearranged, or added without departing from the scope of the present disclosure.

[0087] Examples are used herein to disclose the invention, including the best mode, and to enable any person skilled in the art to practice the invention, including making and using any devices or systems, and performing any methods incorporated therein. The patentable scope of the invention is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are considered to be within the scope of the claims if they include structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements that do not differ insubstantial way from the literal language of the claims.

[0088] Further aspects of the present invention are provided by the subject matter of the following clauses.

[0089] 1. A method for selective catalytic reduction, comprising: obtaining, by a computing system including a machine learning model, input data comprising one or more input values; generating, by the machine learning model, output data indicative of an amount of a reactant based on the input data; and providing, by the computing system, a signal to cause the amount of the reactant to be provided to a selective catalytic reduction system.

[0090] The method of one or more of these clauses, wherein generating the output data includes obtaining, by a computing system, two or more data examples from a data structure indicating previous reactant amounts; and generating, by a machine learning model, the output data based on the two or more data examples.

[0091] The method of one or more of these clauses, wherein the two or more data examples are obtained based on a similarity metric between the input data and each of the two or more data examples.

[0092] The method of one or more of these clauses, wherein generating the output data based on the two or more data examples includes at least one of interpolation and regression.

[0093] The method of one or more of these clauses, wherein generating the output data based on the two or more data examples includes at least one of linear interpolation and linear regression.

[0094] The method of any one or more of these clauses, wherein the input data includes data indicative of a turbine load.

[0095] The method of one or more of these clauses, wherein the output data indicative of the amount includes a first reactant quantity value, further including determining, by the computing system, a reactant quantity adjustment value based on a start-up or shutdown process of the industrial system including the selective catalytic reduction system; and adjusting, by the computing system, the first reactant quantity value based on the reactant quantity adjustment value to produce the amount of reactant.

[0096] The method of one or more of these clauses, further including providing, by a computing system, one or more signals to cause amounts of reactants to be provided to the selective catalytic reduction system throughout a first period associated with a latency of emissions data associated with the selective catalytic reduction system, and adjusting, after the first period, by a self-regulating reactant flow control system, the reactant flows provided to the selective catalytic reduction system.

[0097] The method of any one or more of these clauses, wherein the self-regulating reactant flow control system comprises a proportional-integral-derivative controller.

[0098] The method of one or more of these clauses, further including: determining, by a computing system, adjusted reactant amounts based on the adjusting; and storing, by the computing system, data indicative of the adjusted reactant amounts in a training data structure associated with the machine learning model.

[0099] The method of one or more of these clauses, further including determining, by the computing system based on a comparison between the adjusted reactant amounts and the amounts of the reactants, whether to store data indicative of the adjusted reactant amounts in a training data structure associated with the machine learning model, wherein the storing is performed in response to a determination that the data should be stored.

[0100] The method of one or more of these clauses, wherein determining the adjusted reactant amounts includes monitoring, by a computing system, the adjusting; determining that the reactant flow rates have stabilized; and determining, by the computing system, the adjusted reactant amounts based on the stabilized reactant flow rates.

[0101] 1. A method for training a machine learning model to output a reactant amount, the method comprising: obtaining, by a self-adjusting reactant flow control system, emissions data indicative of one or more emissions; adjusting, by the self-adjusting reactant flow control system, a first amount of a reactant supplied to a selective catalytic reduction system based on the emissions data; monitoring, by a computing system comprising one or more computing devices, the adjusting; determining that the first amount of the reactant supplied to the selective catalytic reduction system has stabilized; and training, by the computing system, a machine learning model using one or more example data including data indicative of the stabilized first amount.

[0102] The method of one or more of these clauses, wherein the machine learning model is configured to, in response to receiving one or more inference inputs, retrieve data from a data structure that includes one or more data examples based on the one or more inference inputs, and generate an output based on the retrieved data.

[0103] The method of one or more of these clauses, further including: providing, by a computing system, input data related to a system state of an industrial system including the selective catalytic reduction system to a machine learning model; generating, by the machine learning model, output data indicative of a second amount of the reactant based on the input data; and determining, based on a comparison between the first amount and the second amount, whether to store the data indicative of the first amount in a data structure including one or more example data.

[0104] The method of one or more of these clauses, further comprising: deleting, by the computing system, previous data examples from the data structure including the one or more data examples, the previous data examples including data indicative of previous reactant amounts.

[0105] The method of one or more of these clauses, further including determining whether the industrial system including the self-regulating reactant flow control system is in one or more predetermined conditions indicative of a lack of emissions data quality, and in response to determining that the industrial system is not in any of the one or more predetermined conditions, training a machine learning model using the one or more data examples.

[0106] The method of any one or more of these clauses, wherein the one or more predetermined conditions include at least one of a start-up process of the industrial system, a shutdown process of the industrial system, a calibration mode of a component of the industrial system, and a data signal having a current or voltage outside an expected current or voltage range.

[0107] The method of any one or more of these clauses, wherein the self-regulating reactant flow control system comprises a proportional-integral-derivative controller.

[0108] A computing system comprising one or more processors and one or more non-transitory computer-readable media, the one or more non-transitory computer-readable media collectively storing a machine learning model and instructions that, when executed by the one or more processors, cause the computing system to perform one or more of the methods of these clauses. [Explanation of symbols]

[0109] 102 Input Data 104 Machine Learning Systems 106 Reactant amount prediction, reactant amount 108 Reactant Flow Control System 110 Reactants 112 Selective Catalytic Reduction System 114 Industrial Systems 116 Catalytic Reduction Emissions 118 Exhaust System 220 Self-regulating reactant flow control system, Self-regulating flow control system 222 Sensor Data 224 Sensors 226 Reactant amount data, reactant amount 304 Machine Learning Systems, Machine Learning Models 328 Example of stored data 330 Get nearest neighbor 332 Obtained neighborhood data, obtained data 334 Neighborhood-Based Reasoning 338 Comparison System 340 Amazing Examples 414 Industrial Turbine Systems 418 exhaust stack 442 Gas Turbine 444 Heat Recovery Steam Generator 500 Exemplary Methods 600 Exemplary Method 702 Computing Systems 704 Computing Devices 706 Processors, processor devices 708 CPU, Central Processing Unit (CPU) 710 Application Specific Integrated Circuits (ASIC) 712 Memory Devices 714 Storage Devices 716 Input / Output Devices 718 Machine Learning Models 720 Network 722 Industrial Input Systems 724 Computing Systems 726 Industrial Control Systems

Claims

1. A method for selective catalytic reduction comprising: obtaining, by a computing system (702) including a machine learning model (718), input data (102) including one or more input values; generating, by the machine learning model (718), output data indicative of the amount (106) of a reactant (110) based on the input data (102); providing, by the computing system (702), a signal to cause the amount (106) of the reactant (110) to be provided to a selective catalytic reduction system (112); A method comprising:

2. generating the output data, retrieving, by said computing system (702), from a data structure, two or more data instances (328) indicative of previous reactant amounts; generating, by the machine learning model (718), the output data based on the two or more data examples (328); The method of claim 1 , comprising:

3. the two or more data examples (328) are obtained based on a similarity metric between the input data (102) and each of the two or more data examples (328); and / or generating the output data based on the plurality of data examples (328); Interpolation, preferably linear interpolation, and Regression, preferably linear regression The method of claim 2 , comprising at least one of:

4. The method of any one of claims 1 to 3, wherein the input data (102) includes data indicative of turbine load.

5. the output data indicative of the quantity (106) includes a first reactant quantity value; determining, by the computing system (702), reactant amount adjustment values ​​based on a start-up or shut-down process of an industrial system (114) including the selective catalytic reduction system (112); adjusting the first reactant quantity value by the computing system (702) based on the reactant quantity adjustment value to produce the amount (106) of the reactant (110); The method of claim 1 , further comprising:

6. providing, by the computing system (702), one or more signals to cause the selective catalytic reduction system (112) to provide the amount (106) of the reactant (110) throughout a first time period associated with a latency of emissions data associated with the selective catalytic reduction system (112); adjusting, after the first period of time, the reactant flow rate provided to the selective catalytic reduction system (112) by a self-regulating reactant flow control system (220); The method of claim 1 , further comprising:

7. determining, by the computing system (702), adjusted reactant amounts based on the adjusting; and storing, by the computing system (702), data indicative of the adjusted reactant amounts in a training data structure associated with the machine learning model (718); The method of claim 6 further comprising:

8. determining, by the computing system, whether to store the data indicative of the adjusted reactant amounts in the training data structure associated with the machine learning model (718) based on a comparison between the adjusted reactant amounts and the amounts (106) of the reactants (110); said storing being performed in response to determining that the data should be stored; and The method of claim 7 further comprising:

9. determining the adjusted reactant amounts; monitoring, by the computing system (702), the adjusting; and determining that the reactant flow rates have stabilized; determining, by the computing system (702), the adjusted reactant amounts based on the stabilized reactant flow rates; 9. The method of claim 7 or 8, comprising:

10. 1. A method for training a machine learning model (718) to output reactant quantities (106), comprising: acquiring, by a self-regulating reactant flow control system (220), emissions data indicative of one or more emissions; adjusting, by the self-adjusting reactant flow control system (220), a first amount of reactant supplied to a selective catalytic reduction system (112) based on the emissions data; monitoring the adjusting by a computing system (702) comprising one or more computing devices (704); determining that the first amount of the reactant supplied to the selective catalytic reduction system (112) has stabilized; training, by the computing system (702), the machine learning model (718) using one or more data examples (328) including data indicative of the stable first quantity; A method comprising:

11. The machine learning model (718) in response to receiving one or more inference inputs, obtaining data from a data structure containing the one or more data examples (328) based on the one or more inference inputs; generating an output based on the acquired data (332); The method of claim 10, wherein the method is configured to:

12. providing, by the computing system (702), to the machine learning model (718), input data (102) related to a system state of an industrial system (114) including the selective catalytic reduction system (112); generating, by the machine learning model (718), output data indicative of a second amount of the reactant based on the input data (102); determining whether to store the data indicative of the first amount in a data structure containing the one or more data instances (328) based on a comparison between the first amount and the second amount; 12. The method of claim 10 or 11, further comprising:

13. removing, by said computing system (702), from the data structure containing said one or more data examples (328), previous data examples including data indicative of previous reactant amounts.

13. The method of any one of claims 10 to 12, further comprising:

14. determining whether the industrial system (114) including the self-regulating reactant flow control system (220) is experiencing one or more predetermined conditions indicative of a lack of emissions data quality; training the machine learning model (718) using the one or more data examples (328) in response to determining that the industrial system (114) does not meet any of the one or more predetermined conditions; Further includes The one or more predetermined conditions are: a start-up process for said industrial system (114); a shutdown process for the industrial system (114); a calibration mode for components of the industrial system (114); and Data signals with currents or voltages outside the expected current or voltage ranges at least one of:

14. The method according to any one of claims 11 to 13.

15. one or more processors (706); one or more non-transitory computer-readable media; 1. A computing system (702) comprising: Machine learning models (718) and instructions that, when executed by the one or more processors (706), cause the computing system (702) to perform the method of any one of claims 1 to 9; Collectively remembering A computing system (702).

Images