Machine learning based boiler combustion control method and system

Abstract

This invention discloses a boiler combustion control method and system based on machine learning, belonging to the interdisciplinary field of artificial intelligence and industrial control. It includes: acquiring continuous combustion cycle operation and emission time-series data of the boiler; constructing a multi-dimensional dataset through alignment and standardization; extracting spatiotemporal coupling features to generate a combustion state evolution feature tensor; inputting it into a combustion process evolution prediction model to obtain a predicted distribution of future combustion states; generating a Pareto optimal control command set through multi-objective optimization; and selecting combustion control commands from the set and sending them to the boiler actuators, considering safety constraints and actuator adjustment margins. By employing a spatiotemporal coupling feature extraction network and an adaptive evolutionary combustion process evolution prediction model, combined with a multi-objective evolutionary algorithm and a safety constraint screening mechanism, multi-objective collaborative optimization and safe, stable control of the boiler combustion process can be achieved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and industrial control, and in particular to a boiler combustion control method and system based on machine learning. Background Technology

[0002] As a core piece of equipment in industrial production and energy conversion, the boiler's combustion process is a complex physicochemical process involving gas-solid multiphase flow, chemical reactions, and heat transfer. The boiler combustion control system needs to adjust key parameters such as coal feed and air supply in real time based on load demand, fuel characteristics, and environmental standards to maintain stable and efficient combustion within the furnace. With the intelligent transformation of industry, machine learning technology has been gradually introduced into the field of boiler control. By analyzing massive amounts of operational data, mathematical models of the combustion process are established, thereby guiding the formulation of control strategies. Machine learning models can handle high-dimensional nonlinear data, providing a new approach to solving the problem that traditional mechanistic models struggle to accurately describe complex combustion processes.

[0003] Existing technologies for boiler combustion control typically employ static data-driven models based on multilayer perceptrons or support vector machines. These methods collect steady-state operating data under specific conditions, extract features, and train predictive models to establish a mapping relationship between input parameters and combustion efficiency or pollutant emissions. During the control phase, the trained static model is combined with traditional single-objective optimization algorithms, such as gradient descent, to find the optimal combination of control parameters for a single indicator. The calculated parameters are then directly sent to the boiler actuators for adjustment.

[0004] Existing technologies struggle to effectively capture the dynamic evolution of parameters over time and their spatial coupling when processing continuous boiler operation data, leading to decreased model prediction accuracy under varying operating conditions. Static models cannot adapt to characteristic drift caused by wear and tear on internal boiler equipment or changes in fuel properties, and lack online adaptive update mechanisms. In the control command generation stage, optimization is focused on a single objective, making it difficult to simultaneously balance conflicting indicators such as combustion efficiency, pollutant emissions, and load response rate. The generated control commands do not adequately consider the dynamic adjustment limitations of actuators and current state deviations, and direct issuance may result in equipment overshoot or triggering safety protection actions. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a boiler combustion control method and system based on machine learning. It employs a spatiotemporally coupled feature extraction network and an adaptive evolutionary combustion process prediction model, combined with a multi-objective evolutionary algorithm and a safety constraint screening mechanism, to achieve multi-objective collaborative optimization and safe and stable control of the boiler combustion process.

[0006] The above objectives can be achieved through the following approach: A machine learning-based boiler combustion control method includes: acquiring the boiler's operating parameter sequence and flue gas emission monitoring sequence over multiple continuous combustion adjustment cycles; constructing a multi-dimensional time-series operation dataset through time-series alignment and standardization; extracting spatiotemporal coupling features from the multi-dimensional time-series operation dataset to generate a combustion state evolution feature tensor; inputting the combustion state evolution feature tensor into a combustion process evolution prediction model trained on historical combustion operation data to obtain the predicted combustion state distribution for future times; using a multi-objective optimization function to solve the predicted combustion state distribution for future times to generate a Pareto optimal control instruction set; selecting combustion control instructions from the Pareto optimal control instruction set based on current operating condition safety constraints and actuator adjustment margin, and sending them to the boiler actuator.

[0007] Optionally, the boiler's operating parameter sequences and flue gas emission monitoring sequences are obtained over multiple continuous combustion regulation cycles. After time-series alignment and standardization, a multi-dimensional time-series operating dataset is constructed, which also includes: Collect time-series data on furnace temperature, oxygen content, coal feed rate, primary air volume, secondary air volume, and nitrogen oxide and carbon monoxide concentrations in flue gas. The furnace temperature, oxygen content, coal feed rate, primary air volume, and secondary air volume constitute an operating parameter sequence, and the nitrogen oxide concentration and carbon monoxide concentration constitute a flue gas emission monitoring sequence. The operating parameter sequence and the flue gas emission monitoring sequence are aligned by timestamps and standardized using a sliding window to ensure that each parameter dimension has the same scale, thus obtaining the multi-dimensional time-series operating dataset.

[0008] Optionally, performing spatiotemporal coupling feature extraction on the multi-dimensional time-series runtime dataset to generate a combustion state evolution feature tensor further includes: The multi-dimensional time-series running dataset is input into a feature extraction network composed of a concatenated convolutional neural network and a bidirectional long short-term memory network; By sliding the convolutional layers of the convolutional neural network along the time dimension, local parameter coupling features within each window of the multi-dimensional time-series running dataset are extracted. The local parameter coupling features are input into a bidirectional long short-term memory network in a temporal sequence, and the forward and reverse temporal dependency information is extracted and fused to form the combustion state evolution feature tensor.

[0009] Optionally, inputting the combustion state evolution feature tensor into a combustion process evolution prediction model trained with historical combustion operation data to obtain the predicted distribution of combustion state at future times further includes: The historical combustion operation data is acquired, and a historical multidimensional time series sample including the operation parameter sequence and emission monitoring sequence is constructed in chronological order, and the actual combustion state at future moments is marked. The operating parameter sequence and emission monitoring sequence in the historical multidimensional time series samples are feature extracted and then input into the time series self-attention transformation network, which outputs the predicted combustion state for future moments. Calculate the multi-task loss function between the predicted combustion state and the actual combustion state, wherein the multi-task loss function includes both a combustion efficiency deviation term and a pollutant emission concentration deviation term. Gradient backpropagation is performed based on the multi-task loss function to update the parameters of the temporal self-attention transformation network until the loss converges, thereby obtaining the combustion process evolution prediction model; Real-time acquisition of the boiler's newly generated operating parameter sequence and flue gas emission monitoring sequence after control execution, and updating the multi-dimensional time-series operating dataset; The spatiotemporal coupling feature extraction is repeated on the updated multi-dimensional time-series running dataset to generate an online combustion state evolution feature tensor, which is then input into the combustion process evolution prediction model to obtain the online combustion state prediction distribution. Calculate the deviation between the online combustion state prediction distribution and the actual combustion state monitored at the corresponding time, and construct an adaptive loss function for the model; The weights of the temporal self-attention transformation network are incrementally updated according to the model adaptive loss function, so that the combustion process evolution prediction model can continuously evolve during operation.

[0010] Optionally, the multi-objective optimization function is used to solve the predicted combustion state distribution at the future time point to generate a Pareto optimal control instruction set, which further includes: Construct a first objective function with the goal of maximizing combustion efficiency, and a second and third objective functions with the goals of minimizing nitrogen oxide emissions and carbon monoxide emissions, to form the multi-objective optimization function; The boiler's adjustable variables are used as decision variables, including coal feed rate, primary air volume, and secondary air volume; Using the predicted distribution of combustion state at future moments as state input, and combining the range of change of the decision variables, the predicted values ​​of combustion efficiency, nitrogen oxide emissions, and carbon monoxide emissions corresponding to each combination of decision variables are calculated. A multi-objective evolutionary algorithm is used to search for optimizations in the decision variables to obtain a non-dominated solution set in which the first objective function, the second objective function, and the third objective function are not mutually dominant. The Pareto optimal control instruction set is composed of the non-dominated solution set. Obtain the load dispatch command of the boiler at the current moment, and convert the target load change rate required by the command into the load response rate objective function; The load response rate objective function is weighted and combined with the highest combustion efficiency objective function and the lowest pollutant emission objective function to form a comprehensive fitness function; During the multi-objective evolutionary algorithm search process, the comprehensive fitness function guides the population evolution direction to obtain a non-dominated solution set that simultaneously considers load following rate, combustion efficiency and emission performance, and the Pareto optimal control command set is updated by the non-dominated solution set.

[0011] Optionally, selecting combustion control commands from the Pareto optimal control command set based on current operating condition safety constraints and actuator adjustment margins, and sending them to the boiler actuator, further includes: Obtain the upper limit of furnace pressure and upper limit of bed temperature in the current working condition safety constraints, and the limit of the fan damper opening rate change and the limit of the coal feeder speed adjustment rate in the actuator adjustment margin; Remove instructions from the Pareto optimal control instruction set that are expected to cause the furnace pressure to exceed the furnace pressure limit or the bed temperature to exceed the bed temperature limit after execution, and obtain a feasible instruction subset; Within the feasible instruction subset, candidate instructions are further selected that simultaneously satisfy the limit of the fan damper opening rate and the limit of the coal feeder speed adjustment rate. The combustion control command is selected from the candidate commands that has the smallest deviation from the current fan damper opening and coal feeder speed.

[0012] Based on the same inventive concept, this invention also provides a machine learning-based boiler combustion control system, comprising: a time-series data acquisition module for acquiring the boiler's operating parameter sequences and flue gas emission monitoring sequences over multiple continuous combustion adjustment cycles, and constructing a multi-dimensional time-series operation dataset through time-series alignment and standardization; a feature extraction module for performing spatiotemporal coupling feature extraction on the multi-dimensional time-series operation dataset to generate a combustion state evolution feature tensor; an evolution prediction module, embedded with a combustion process evolution prediction model trained using historical combustion operation data, for inputting the combustion state evolution feature tensor to obtain a predicted combustion state distribution for future times; a multi-objective optimization module for using a multi-objective optimization function to solve for the predicted combustion state distribution for future times, generating a Pareto optimal control instruction set; and an instruction selection module for selecting combustion control instructions from the Pareto optimal control instruction set based on current operating condition safety constraints and actuator adjustment margin, and sending them to the boiler actuator.

[0013] Compared with the prior art, the present invention has the following advantages: This invention constructs a feature extraction network that includes a multi-layered gridded feature extractor and a bidirectional sequence memory unit. It performs parameter correlation extraction within a local window on the multi-parameter sequence of boiler operation data, such as furnace temperature, coal feed rate, and air volume. At the same time, it encodes the forward and reverse temporal dependencies over a long time span. The generated combustion state evolution feature tensor provides an input representation for subsequent prediction models.

[0014] This invention utilizes a sequence mapping network with a self-attention mechanism to allocate time weights to construct a combustion process evolution prediction model. This enables the model to differentiate the weights of information from different historical moments when predicting future combustion states. Simultaneously, a multi-task loss function containing combustion efficiency deviation and pollutant emission concentration deviation terms is used for training, and an adaptive loss function is introduced to incrementally update the network weights, ensuring that the output distribution of the prediction model continuously approximates changes in boiler operating conditions.

[0015] This invention employs a parallel search strategy based on population iterative optimization, using coal feed rate, primary air volume, and secondary air volume as decision variables. It combines a comprehensive fitness function to perform multi-objective optimization on the predicted distribution of future combustion states, generating a Pareto optimal control instruction set. This instruction set contains multiple non-dominant control schemes for subsequent selection based on actual constraints.

[0016] This invention acquires current operating condition safety constraints such as furnace pressure upper limit and bed temperature upper limit, as well as actuator adjustment margins such as fan damper opening rate limit and coal feeder speed adjustment rate limit during the command issuance stage. By calculating the dimensionless deviation distance between the candidate command and the current fan damper opening and coal feeder speed state, the command with the smallest deviation distance is selected as the combustion control command and sent to the boiler actuator, so that the finally issued command simultaneously meets the safety boundary and the dynamic adjustment capability of the actuator. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a machine learning-based boiler combustion control method according to an embodiment of the present invention.

[0018] Figure 2 This is a flowchart of the online adaptive update process of the combustion process evolution prediction model according to an embodiment of the present invention.

[0019] Figure 3 This is a framework diagram of a machine learning-based boiler combustion control system according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 One embodiment of this invention proposes a machine learning-based boiler combustion control method. This method collects time-series data on furnace temperature, oxygen content, coal feed rate, air volume, and nitrogen oxide and carbon monoxide concentrations in flue gas using sensors. After time-series alignment and standardization, a multi-dimensional time-series operational dataset is constructed. This dataset is then input into a feature extraction network composed of a multi-layered gridded feature extractor and bidirectional sequence memory units to extract local parameter coupling features and global time-series dependency information, fusing them to generate a combustion state evolution feature tensor. A combustion process evolution prediction model is constructed using a self-attention sequence mapping network, trained with a multi-task loss function, and incrementally updated online. A parallel search strategy combined with a comprehensive fitness function is employed for multi-objective optimization, generating a Pareto optimal control instruction set that considers load response, combustion efficiency, and emission performance. During the instruction issuance stage, safety constraints such as furnace pressure upper limits and bed temperature upper limits, as well as adjustment margins for actuators such as fan dampers and coal feeders, are introduced. Instructions with the smallest state deviation are selected and sent to the boiler actuators, forming a complete control link from data acquisition, feature extraction, state prediction, multi-objective optimization to safety selection.

[0022] The method described in this embodiment specifically includes: S1. Obtain the boiler's operating parameter sequence and flue gas emission monitoring sequence in multiple continuous combustion adjustment cycles, and construct a multi-dimensional time-series operating dataset after time-series alignment and standardization. In another embodiment of the present invention, step S1 further includes: Collect time-series data on furnace temperature, oxygen content, coal feed rate, primary air volume, secondary air volume, and nitrogen oxide and carbon monoxide concentrations in flue gas. The furnace temperature, oxygen content, coal feed rate, primary air volume, and secondary air volume constitute an operating parameter sequence, and the nitrogen oxide concentration and carbon monoxide concentration constitute a flue gas emission monitoring sequence. The operating parameter sequence and the flue gas emission monitoring sequence are aligned by timestamps and standardized using a sliding window to ensure that each parameter dimension has the same scale, thus obtaining the multi-dimensional time-series operating dataset.

[0023] Specifically, sensors deployed at key nodes of the boiler collect time-series data on furnace temperature, oxygen content, coal feed rate, primary air volume, secondary air volume, and nitrogen oxide and carbon monoxide concentrations in the flue gas. Furnace temperature represents the thermodynamic state of the combustion zone inside the boiler. Oxygen content reflects the remaining amount of combustion-supporting gas during combustion. Coal feed rate refers to the mass flow rate of fuel input into the boiler. Primary and secondary air volumes represent the air flow rates required for fuel delivery and supplementary combustion, respectively. Nitrogen oxide and carbon monoxide concentrations reflect the pollutant content in the flue gas after combustion. The furnace temperature, oxygen content, coal feed rate, primary air volume, and secondary air volume are arranged in chronological order to form an operating parameter sequence. The nitrogen oxide and carbon monoxide concentrations are arranged in chronological order to form a flue gas emission monitoring sequence. These sequences cover multiple continuous combustion adjustment cycles, which refer to the time interval between the boiler receiving a control command and the stabilization of the combustion state, preparing to receive the next command. Due to differences in sampling frequencies among the sensors, the operating parameter sequence and the flue gas emission monitoring sequence are time-series aligned by timestamps. Time alignment refers to interpolating or truncating data from different sampling frequencies to the same time node. A sliding window is used to standardize the aligned data. A sliding window is a method of truncating data segments of a fixed length on the time series; the window length is set to 60 sampling points based on the autocorrelation function decay characteristics of the boiler combustion process. Standardization eliminates the differences in physical dimensions of the parameters and maps them to a unified numerical range, ensuring that each parameter dimension has the same scale, ultimately resulting in a multi-dimensional time-series operating dataset. The multi-dimensional time-series operating dataset is a matrix that integrates operating parameter and emission monitoring data and has undergone dimensionless processing. The standardization process is calculated using a specific formula, as follows: , in The value is a dimensionless value after standardization. The original parameter measurement value at the current time point within the sliding window, the dimension of which depends on the specific parameter such as degrees Celsius or milligrams per cubic meter; This is the arithmetic mean of the measured values ​​of this parameter within the sliding window, with the same dimensions as the original parameter. This is the standard deviation of the measured values ​​of this parameter within the sliding window, with the same dimensions as the original parameter. To prevent extremely small constants with a denominator of zero, their dimensions are the same as the original parameters, and they are set to 0.001 based on 500 sets of measured data from industrial sensors. By subtracting from this formula and dividing by the standard deviation of the same dimension, the coefficients are completely eliminated, ensuring that the output values ​​are dimensionless.

[0024] For example, the furnace temperature time-series data is collected within the current continuous combustion regulation cycle, and the raw parameter measurement values ​​at the current time point are obtained. =950 degrees Celsius. Within a set sliding window containing furnace temperature data from the past 60 sampling points, the arithmetic mean of the measured values ​​of this parameter within the sliding window is calculated. =940 degrees Celsius. The standard deviation of the measured values ​​of this parameter within the sliding window was calculated. =20 degrees Celsius. Set a minimum constant. =0.001 degrees Celsius. Substituting the above value into the standardization formula, the dimensionless value after standardization is obtained through calculation. 0.4999. This calculation process transforms the furnace temperature, which has physical dimensions, into a dimensionless scale. Together with other parameters that have undergone the same processing, such as oxygen, coal feed, primary air volume, secondary air volume, nitrogen oxide concentration, and carbon monoxide concentration, it forms a multi-dimensional time-series operational dataset, ensuring that each parameter dimension has the same scale.

[0025] S2. Perform spatiotemporal coupling feature extraction on the multi-dimensional time-series running dataset to generate a combustion state evolution feature tensor; In another embodiment of the present invention, step S2 further includes: The multi-dimensional time-series running dataset is input into a feature extraction network composed of a concatenated convolutional neural network and a bidirectional long short-term memory network; By sliding the convolutional layers of the convolutional neural network along the time dimension, local parameter coupling features within each window of the multi-dimensional time-series running dataset are extracted. The local parameter coupling features are input into a bidirectional long short-term memory network in a temporal sequence, and the forward and reverse temporal dependency information is extracted and fused to form the combustion state evolution feature tensor.

[0026] Specifically, the multi-dimensional time-series runtime dataset generated in step S1 is obtained and input into a feature extraction network composed of a multi-layer gridded feature extractor and a bidirectional sequence memory unit. The multi-layer gridded feature extractor is used to capture the correlation between different parameters at adjacent time points. The bidirectional sequence memory unit is used to remember and transmit the evolution patterns of the past and future in the time series. The feature extraction network refers to a computational model that integrates these two structures to process complex time-series data. The gridded computation layer of the multi-layer gridded feature extractor slides along the time dimension to extract the local parameter coupling features within each window of the multi-dimensional time-series runtime dataset. The local parameter coupling features refer to the mathematical expression of the mutual influence between parameters such as furnace temperature, oxygen content, and coal feed rate within a short time window. This extraction process is calculated using a specific formula, as follows: , in Time node The local parameter coupling characteristics at a given point are dimensionless. The length of the sliding window is set to 5 based on 300 sets of industrial boiler operation test data; This is the weight matrix of the grid computing layer, which is dimensionless. For the time nodes in the multi-dimensional time series running dataset obtained in step S1 The dimensionless value at the location; The bias vector is dimensionless. Local parameter coupling features are input into a bidirectional sequential memory unit according to time sequence, while forward and reverse time-series dependency information is extracted. Forward time-series dependency information refers to the parameter evolution pattern from the past to the present, and reverse time-series dependency information refers to the parameter evolution pattern from the future to the present. The fused output forms a combustion state evolution feature tensor. The combustion state evolution feature tensor is a multi-dimensional data matrix that integrates local parameter correlations and global time evolution patterns. The fusion process is calculated using a specific formula, as follows: , in Time node The combustion state evolution characteristic tensor at the location is dimensionless; It represents positive time-dependent information and is dimensionless. It is inverse temporal dependency information and is dimensionless; The positive weighting coefficient is dimensionless and is set to 0.6 based on 300 sets of industrial boiler operation test data. The inverse weighting coefficient is dimensionless and set to 0.4 based on 300 sets of industrial boiler operation test data.

[0027] For example, a multi-dimensional time-series running dataset is input into the feature extraction network. At time nodes... At this point, the length of the sliding window The value is set to 5. Assume the calculated local parameter coupling characteristics... Sequence input bidirectional sequence memory unit. At time node. Bidirectional sequential memory units calculate positive temporal dependency information. =0.75, inverse timing dependency information =0.65. Set the positive weighting coefficient. =0.6, set the reverse weight coefficient. =0.4. Substituting the above value into the fusion formula, the time node is calculated. Combustion state evolution characteristic tensor at the location =0.71. This calculation process effectively fuses local parameter coupling features with temporal dependency information to generate a combustion state evolution feature tensor for subsequent prediction.

[0028] S3. Input the combustion state evolution feature tensor into the combustion process evolution prediction model trained by historical combustion operation data to obtain the predicted distribution of combustion state at future moments; In another embodiment of the present invention, step S3 further includes: The historical combustion operation data is acquired, and a historical multidimensional time series sample including the operation parameter sequence and emission monitoring sequence is constructed in chronological order, and the actual combustion state at future moments is marked. The operating parameter sequence and emission monitoring sequence in the historical multidimensional time series samples are feature extracted and then input into the time series self-attention transformation network, which outputs the predicted combustion state for future moments. Calculate the multi-task loss function between the predicted combustion state and the actual combustion state, wherein the multi-task loss function includes both a combustion efficiency deviation term and a pollutant emission concentration deviation term. Gradient backpropagation is performed based on the multi-task loss function to update the parameters of the temporal self-attention transformation network until the loss converges, thereby obtaining the combustion process evolution prediction model.

[0029] In another embodiment of the present invention, step S3 further includes: Real-time acquisition of the boiler's newly generated operating parameter sequence and flue gas emission monitoring sequence after control execution, and updating the multi-dimensional time-series operating dataset; The spatiotemporal coupling feature extraction is repeated on the updated multi-dimensional time-series running dataset to generate an online combustion state evolution feature tensor, which is then input into the combustion process evolution prediction model to obtain the online combustion state prediction distribution. Calculate the deviation between the online combustion state prediction distribution and the actual combustion state monitored at the corresponding time, and construct an adaptive loss function for the model; The weights of the temporal self-attention transformation network are incrementally updated according to the model adaptive loss function, so that the combustion process evolution prediction model can continuously evolve during operation.

[0030] Specifically, historical combustion operation data accumulated during the boiler's past operating phases is acquired. Historical combustion operation data refers to sensor measurements recorded over a long period under various operating conditions. A multi-dimensional historical time-series sample, including operating parameter sequences and emission monitoring sequences, is constructed in chronological order. This multi-dimensional historical time-series sample refers to standardized historical data segments used to train the prediction model. The actual combustion state at future times is labeled onto these samples. The actual combustion state refers to the combustion efficiency and pollutant concentration values ​​that actually occur at future time points in the historical records, which are converted to dimensionless values ​​after the same standardization process as in step S1. The operating parameter sequences and emission monitoring sequences from the historical multi-dimensional time-series samples are then processed using the same feature extraction process as in step S2 and input into a sequence mapping network that assigns time weights through a self-attention mechanism. This network automatically assigns weights based on the degree of influence of data at different time points on the future state. The network outputs the predicted combustion state at future times. The predicted combustion state refers to the dimensionless predicted values ​​of combustion efficiency and pollutant concentration at future time points calculated by the model. A multi-task loss function is calculated between the predicted combustion state and the actual combustion state. The multi-task loss function is a mathematical index that measures the overall error of the model across multiple prediction targets. The multi-task loss function includes both a combustion efficiency deviation term and a pollutant emission concentration deviation term. The combustion efficiency deviation term refers to the mean square error between the predicted and actual combustion efficiency. The pollutant emission concentration deviation term refers to the mean square error between the predicted and actual pollutant concentrations. The formula for calculating the multi-task loss function is as follows: , in This is a multi-task loss function, dimensionless. This is a dimensionless term representing the combustion efficiency deviation. This is the pollutant emission concentration deviation term, which is dimensionless. The combustion efficiency weighting coefficient is dimensionless and set to 0.5 based on 500 sets of industrial boiler operation test data. The pollutant emission weighting coefficient is dimensionless and set to 0.5 based on 500 sets of industrial boiler operation test data. Gradient backpropagation is performed using a multi-task loss function to update the parameters of the sequence mapping network, which allocates time weights through a self-attention mechanism, until the loss converges, thus obtaining the combustion process evolution prediction model. The combustion process evolution prediction model refers to a computational architecture that has been trained on historical data and can accurately predict future combustion states.

[0031] The combustion state evolution feature tensor generated in step S2 is input into the combustion process evolution prediction model to obtain the predicted combustion state distribution for future time steps. The predicted combustion state distribution refers to the probability or numerical range of the combustion state within multiple future time steps output by the model. Real-time acquisition of the newly generated operating parameter sequence and flue gas emission monitoring sequence after boiler control execution updates the multi-dimensional time-series operating dataset. Repeated spatiotemporal coupling feature extraction is performed on the updated multi-dimensional time-series operating dataset to generate an online combustion state evolution feature tensor. The online combustion state evolution feature tensor refers to the feature matrix extracted based on the latest real-time data. This is input into the combustion process evolution prediction model to obtain the online combustion state prediction distribution. The online combustion state prediction distribution refers to the future combustion state predicted based on rolling real-time data. The deviation between the online combustion state prediction distribution and the actual monitored combustion state at the corresponding time is calculated to construct the model adaptive loss function. The model adaptive loss function is an error evaluation index used for online fine-tuning of model parameters. The calculation formula is as follows: , in The adaptive loss function for the model is dimensionless. The total number of time steps is dimensionless and set to 10 based on the industrial boiler control cycle. For the first Dimensionless predicted values ​​in the online combustion state prediction distribution at each time step; Representing the The dimensionless true value of the combustion state actually monitored at each time step; the weights of the sequence mapping network, which allocates time weights through a self-attention mechanism, are incrementally updated according to the model's adaptive loss function, so that the combustion process evolution prediction model can continuously evolve during operation.

[0032] For example, during the model training phase, the combustion efficiency deviation term is calculated. =0.04, Pollutant emission concentration deviation item =0.06. Set the combustion efficiency weighting coefficient. =0.5, setting the pollutant emission weighting coefficient =0.5. Substituting into the multi-task loss function formula, the multi-task loss function is obtained through calculation. =0.05. During the online update phase, the total number of prediction time steps is set. =10. Assume that at the first time step, the dimensionless predicted value in the online combustion state prediction distribution is... =0.80, the dimensionless true value of the actual monitored combustion state. =0.82. The adaptive loss function of the model is obtained through calculation. = 0.0004. This value is used for incremental weight updates to ensure the model adapts to the current actual working conditions.

[0033] S4. Use a multi-objective optimization function to solve the predicted distribution of combustion state at the future time and generate a Pareto optimal control instruction set. In another embodiment of the present invention, step S4 further includes: Construct a first objective function with the goal of maximizing combustion efficiency, and a second and third objective functions with the goals of minimizing nitrogen oxide emissions and carbon monoxide emissions, to form the multi-objective optimization function; The boiler's adjustable variables are used as decision variables, including coal feed rate, primary air volume, and secondary air volume; Using the predicted distribution of combustion state at future moments as state input, and combining the range of change of the decision variables, the predicted values ​​of combustion efficiency, nitrogen oxide emissions, and carbon monoxide emissions corresponding to each combination of decision variables are calculated. The decision variables are optimized using a multi-objective evolutionary algorithm to obtain a set of non-dominated solutions that make the first objective function, the second objective function, and the third objective function non-dominated. The Pareto optimal control instruction set is composed of the set of non-dominated solutions.

[0034] In another embodiment of the present invention, step S4 further includes: Obtain the load dispatch command of the boiler at the current moment, and convert the target load change rate required by the command into the load response rate objective function; The load response rate objective function is weighted and combined with the highest combustion efficiency objective function and the lowest pollutant emission objective function to form a comprehensive fitness function; During the multi-objective evolutionary algorithm search process, the comprehensive fitness function guides the population evolution direction to obtain a non-dominated solution set that simultaneously considers load following rate, combustion efficiency and emission performance, and the Pareto optimal control command set is updated by the non-dominated solution set.

[0035] Specifically, the predicted combustion state distribution for future moments, output in step S3, is obtained as the state input. Adjustable variables of the boiler are used as decision variables. Decision variables refer to physical quantities that can be actively adjusted by the system during the control process. Adjustable variables include coal feed rate, primary air volume, and secondary air volume. Combining the range of change of the decision variables, the predicted values ​​of combustion efficiency, nitrogen oxide emissions, and carbon monoxide emissions corresponding to each combination of decision variables are calculated. A first objective function is constructed with the goal of maximizing combustion efficiency, and a second and third objective function are constructed with the goals of minimizing nitrogen oxide and carbon monoxide emissions, forming a multi-objective optimization function. A multi-objective optimization function refers to a mathematical model that simultaneously evaluates multiple conflicting performance indicators. The load dispatch command for the boiler at the current moment is obtained. The load dispatch command refers to the power demand change requirement issued by the power grid or the superior control system. The target load change rate required by the command is converted into a load response rate objective function. The target load change rate refers to the required rate of change of boiler power over time. The load response rate objective function is weighted and combined with the objective functions for maximizing combustion efficiency and minimizing pollutant emissions to form a comprehensive fitness function. The comprehensive fitness function refers to an evaluation criterion that integrates multiple evaluation metrics into a single value to guide the search direction. The formula for calculating the comprehensive fitness function is as follows: , in The fitness function is dimensionless. The predicted combustion efficiency output by the first objective function is dimensionless. The predicted nitrogen oxide emissions output by the second objective function are dimensionless. The predicted carbon monoxide emissions output by the third objective function are dimensionless. The numerical value output by the load response rate objective function is dimensionless. The combustion efficiency weighting coefficient is dimensionless and set to 0.4 based on industrial boiler operating specifications. The nitrogen oxide emission weighting coefficient is dimensionless and set to 0.25 based on environmental emission standards. The carbon monoxide emission weighting coefficient is dimensionless and set to 0.15 based on environmental emission standards. The load response weighting coefficient is dimensionless and set to 0.2 based on grid dispatch requirements. A parallel search strategy based on population iterative optimization is used to optimize the decision variables. This parallel search strategy simulates biological evolution by simultaneously evaluating a large number of candidate solutions in a multidimensional space. During the search, a comprehensive fitness function guides the population's evolutionary direction, obtaining a set of non-dominated solutions where the first, second, and third objective functions are mutually non-dominated. A non-dominated solution set refers to the set of decision variable combinations that cannot be simultaneously surpassed by other solutions across all objectives. This yields a set of non-dominated solutions that simultaneously considers load following rate, combustion efficiency, and emission performance, forming a Pareto optimal control instruction set. A Pareto optimal control instruction set is a set containing multiple optimal regulation schemes that balance different objectives.

[0036] For example, under a specific combination of decision variables, the predicted combustion efficiency value output by the first objective function is calculated. =0.92. The predicted value of nitrogen oxide emissions from the output of the second objective function is calculated. =0.18. The predicted carbon monoxide emissions from the third objective function were calculated. =0.12. The calculated value of the load response rate objective function output is 0.12. =0.85. Set the combustion efficiency weighting coefficient. =0.4, setting the nitrogen oxide emission weighting factor. =0.25, setting the carbon monoxide emission weighting coefficient. =0.15, set the load response weighting factor. =0.2. Substituting the above value into the formula for the comprehensive fitness function, the comprehensive fitness function is obtained through calculation. =0.4×0.92 0.25×0.18 0.15×0.12 0.2 × 0.85 = 0.475. The parallel search strategy based on population iterative optimization uses this comprehensive fitness function value to evaluate the merits of the current combination of decision variables. After multiple iterations, it selects a set of non-dominated solutions that are not mutually dominant, and finally generates a Pareto optimal control instruction set.

[0037] like Figure 2As shown, after the boiler combustion control command is executed, the latest boiler operating parameter sequence and flue gas emission monitoring sequence are collected in real time to update the multi-dimensional time-series operating dataset. Subsequently, spatiotemporal coupling feature extraction is repeatedly performed on the updated data to generate an online combustion state evolution feature tensor, which is then input into the combustion process evolution prediction model to obtain the online combustion state prediction distribution. An adaptive loss function is constructed based on the deviation between the prediction results and the actual monitored combustion state, and incremental learning is used to update the parameters of the combustion process evolution prediction model, enabling the model to continuously adapt to changes in boiler operating conditions and improve the accuracy of subsequent combustion state predictions. After the model update is completed, the updated prediction model is used in the next control cycle to achieve online closed-loop adaptive optimization in the boiler combustion control process.

[0038] S5. From the Pareto optimal control instruction set, select the combustion control instruction based on the current operating condition safety constraints and the actuator adjustment margin, and send it to the boiler actuator.

[0039] In another embodiment of the present invention, step S5 further includes: Obtain the upper limit of furnace pressure and upper limit of bed temperature in the current working condition safety constraints, and the limit of the fan damper opening rate change and the limit of the coal feeder speed adjustment rate in the actuator adjustment margin; Remove instructions from the Pareto optimal control instruction set that are expected to cause the furnace pressure to exceed the furnace pressure limit or the bed temperature to exceed the bed temperature limit after execution, and obtain a feasible instruction subset; Within the feasible instruction subset, candidate instructions are further selected that simultaneously satisfy the limit of the fan damper opening rate and the limit of the coal feeder speed adjustment rate. The combustion control command is selected from the candidate commands that has the smallest deviation from the current fan damper opening and coal feeder speed.

[0040] Specifically, obtain the Pareto optimal control instruction set generated in step S4. From the Pareto optimal control instruction set, select combustion control instructions based on the current operating condition safety constraints and the actuator adjustment margin. The current operating condition safety constraints refer to the physical parameter boundaries set to ensure the safe operation of the boiler equipment. The actuator adjustment margin refers to the maximum adjustment range that the boiler's physical actuators can safely respond to per unit time. Obtain the furnace pressure upper limit and bed temperature upper limit from the current operating condition safety constraints. The furnace pressure upper limit refers to the maximum allowable positive or negative pressure extreme value inside the boiler; exceeding this value may cause damage to the furnace walls. The bed temperature upper limit refers to the highest allowable temperature in the combustion zone; exceeding this value may cause coking or equipment burnout. Obtain the fan damper opening change rate limit and coal feeder speed adjustment rate limit from the actuator adjustment margin. The fan damper opening change rate limit refers to the maximum percentage change in the fan damper opening per second. The coal feeder speed adjustment rate limit refers to the maximum percentage change in the coal feeder motor speed per second. From the Pareto optimal control instruction set, instructions that are expected to cause the furnace pressure to exceed the upper limit or the bed temperature to exceed the upper limit after execution are removed, resulting in a feasible instruction subset. The feasible instruction subset refers to the set of control instructions that satisfy all static safety boundaries. Within this subset, candidate instructions are further selected that simultaneously satisfy the limits on the rate of change of the blower damper opening and the rate of adjustment of the coal feeder speed. Candidate instructions are those that satisfy both static safety boundaries and dynamic adjustment capability limits. From the candidate instructions, the instruction with the smallest deviation from the current blower damper opening and coal feeder speed is selected as the combustion control instruction. The combustion control instruction refers to the adjustment signal ultimately sent to the physical equipment. The selection process for the instruction with the smallest deviation is calculated using a specific formula, as follows: , in The dimensionless deviation distance between the candidate instruction and the current state; The difference between the fan damper opening required by the candidate command and the current fan damper opening is expressed as a percentage. The maximum physical range of the fan damper opening is expressed as a percentage. The difference between the feeder speed required by the candidate command and the current feeder speed, expressed in revolutions per minute; The maximum physical range of the coal feeder's rotational speed is expressed in revolutions per minute. The fan adjustment penalty coefficient is dimensionless and set to 0.6 based on the equipment wear characteristics. The feeder adjustment penalty coefficient is dimensionless and set to 0.4 based on equipment wear characteristics. The coefficient is dimensionless by dividing by the maximum physical range, ensuring that deviations of different physical quantities can be calculated on the same dimensionless scale. The deviation distance of all candidate commands is calculated, and a selection is made. The smallest candidate command is sent to the boiler actuator as a combustion control command.

[0041] For example, a candidate instruction is obtained. The difference between the required fan damper opening and the current fan damper opening is... =5% . Maximum physical range of fan damper opening. =100% The difference between the required feeder speed and the current feeder speed in this candidate instruction. =30 revolutions per minute. Maximum physical range of coal feeder speed. =1000 revolutions per minute. Set the fan adjustment penalty coefficient. =0.6, set the feeder adjustment penalty coefficient. =0.4. Substituting the above value into the state deviation distance formula, the dimensionless deviation distance between the candidate command and the current state is calculated. ≈0.0431. The same calculation was performed on all candidate commands, and the comparison revealed that this candidate command had the smallest deviation distance. Therefore, it was selected as the combustion control command and sent to the boiler actuator for execution.

[0042] Based on the same inventive concept, such as Figure 3 As shown, the present invention also provides a boiler combustion control system based on machine learning, the system comprising: The time-series data acquisition module is used to acquire the boiler's operating parameter sequence and flue gas emission monitoring sequence in multiple continuous combustion adjustment cycles. After time-series alignment and standardization, a multi-dimensional time-series operation dataset is constructed. The feature extraction module is used to perform spatiotemporal coupled feature extraction on the multi-dimensional time-series running dataset to generate a combustion state evolution feature tensor. The evolution prediction module has an embedded combustion process evolution prediction model trained with historical combustion operation data. It is used to input the combustion state evolution feature tensor to obtain the predicted distribution of combustion state at future times. The multi-objective optimization module is used to solve the multi-objective optimization problem on the predicted distribution of combustion state at the future time using a multi-objective optimization function, and generate a Pareto optimal control instruction set. The instruction selection module is used to select combustion control instructions from the Pareto optimal control instruction set, based on the current operating condition safety constraints and the actuator adjustment margin, and send them to the boiler actuator.

[0043] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the disclosure of the specification and practices. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A boiler combustion control method based on machine learning, characterized in that, The method includes: The boiler's operating parameter sequence and flue gas emission monitoring sequence were obtained in multiple continuous combustion adjustment cycles. After time-series alignment and standardization, a multi-dimensional time-series operating dataset was constructed. Spatiotemporal coupling feature extraction is performed on the multi-dimensional time-series running dataset to generate a combustion state evolution feature tensor; The combustion state evolution feature tensor is input into the combustion process evolution prediction model trained with historical combustion operation data to obtain the predicted distribution of combustion state at future times; The combustion state prediction distribution at the future time is solved by multi-objective optimization function to generate a Pareto optimal control command set; Combustion control commands are selected from the Pareto optimal control command set based on the current operating condition safety constraints and the actuator adjustment margin, and then sent to the boiler actuator.

2. The boiler combustion control method based on machine learning according to claim 1, characterized in that, The obtained boiler operating parameter sequences and flue gas emission monitoring sequences over multiple continuous combustion adjustment cycles are time-series aligned and standardized to construct a multi-dimensional time-series operating dataset, including: Collect time-series data on furnace temperature, oxygen content, coal feed rate, primary air volume, secondary air volume, and nitrogen oxide and carbon monoxide concentrations in flue gas. The furnace temperature, oxygen content, coal feed rate, primary air volume, and secondary air volume constitute an operating parameter sequence, and the nitrogen oxide concentration and carbon monoxide concentration constitute a flue gas emission monitoring sequence. The operating parameter sequence and the flue gas emission monitoring sequence are aligned by timestamps and standardized using a sliding window to ensure that each parameter dimension has the same scale, thus obtaining the multi-dimensional time-series operating dataset.

3. The boiler combustion control method based on machine learning according to claim 1, characterized in that, The step of extracting spatiotemporally coupled features from the multi-dimensional time-series running dataset to generate a combustion state evolution feature tensor includes: The multi-dimensional time-series running dataset is input into a feature extraction network composed of a concatenated convolutional neural network and a bidirectional long short-term memory network; By sliding the convolutional layers of the convolutional neural network along the time dimension, local parameter coupling features within each window in the multi-dimensional time-series running dataset are extracted. The local parameter coupling features are input into a bidirectional long short-term memory network in a temporal sequence, and the forward and reverse temporal dependency information is extracted and fused to form the combustion state evolution feature tensor.

4. The boiler combustion control method based on machine learning according to claim 1, characterized in that, The combustion process evolution prediction model is a temporal self-attention transform network, and it is obtained through the following training process: The historical combustion operation data is acquired, and a historical multidimensional time series sample including the operation parameter sequence and emission monitoring sequence is constructed in chronological order, and the actual combustion state at future moments is marked. The operating parameter sequence and emission monitoring sequence in the historical multidimensional time series samples are feature extracted and then input into the time series self-attention transformation network, which outputs the predicted combustion state for future moments. Calculate the multi-task loss function between the predicted combustion state and the actual combustion state, wherein the multi-task loss function includes both a combustion efficiency deviation term and a pollutant emission concentration deviation term. Gradient backpropagation is performed based on the multi-task loss function to update the parameters of the temporal self-attention transformation network until the loss converges, thereby obtaining the combustion process evolution prediction model.

5. The boiler combustion control method based on machine learning according to claim 1, characterized in that, The step of using a multi-objective optimization function to solve the predicted combustion state distribution at future times and generating a Pareto optimal control command set includes: Construct a first objective function with the goal of maximizing combustion efficiency, and a second and third objective functions with the goals of minimizing nitrogen oxide emissions and carbon monoxide emissions, to form the multi-objective optimization function; The boiler's adjustable variables are used as decision variables, including coal feed rate, primary air volume, and secondary air volume; Using the predicted distribution of combustion state at future moments as state input, and combining the range of change of the decision variables, the predicted values ​​of combustion efficiency, nitrogen oxide emissions, and carbon monoxide emissions corresponding to each combination of decision variables are calculated. The decision variables are optimized using a multi-objective evolutionary algorithm to obtain a set of non-dominated solutions that make the first objective function, the second objective function, and the third objective function non-dominated. The Pareto optimal control instruction set is composed of the set of non-dominated solutions.

6. The boiler combustion control method based on machine learning according to claim 1, characterized in that, The step of selecting combustion control commands based on current operating condition safety constraints and actuator adjustment margins includes: Obtain the upper limit of furnace pressure and upper limit of bed temperature in the current working condition safety constraints, and the limit of the fan damper opening rate change and the limit of the coal feeder speed adjustment rate in the actuator adjustment margin; Remove instructions from the Pareto optimal control instruction set that are expected to cause the furnace pressure to exceed the furnace pressure limit or the bed temperature to exceed the bed temperature limit after execution, and obtain a feasible instruction subset; Within the feasible instruction subset, candidate instructions are further selected that simultaneously satisfy the limit of the fan damper opening rate and the limit of the coal feeder speed adjustment rate. The combustion control command is selected from the candidate commands that has the smallest deviation from the current fan damper opening and coal feeder speed.

7. The boiler combustion control method based on machine learning according to claim 4, characterized in that, It also includes online adaptive updates to the combustion process evolution prediction model: Real-time acquisition of the boiler's newly generated operating parameter sequence and flue gas emission monitoring sequence after control execution, and updating the multi-dimensional time-series operating dataset; The spatiotemporal coupling feature extraction is repeated on the updated multi-dimensional time-series running dataset to generate an online combustion state evolution feature tensor, which is then input into the combustion process evolution prediction model to obtain the online combustion state prediction distribution. Calculate the deviation between the online combustion state prediction distribution and the actual combustion state monitored at the corresponding time, and construct an adaptive loss function for the model; The weights of the temporal self-attention transformation network are incrementally updated according to the model adaptive loss function, so that the combustion process evolution prediction model can continuously evolve during operation.

8. The boiler combustion control method based on machine learning according to claim 5, characterized in that, The multi-objective optimization function also includes a load response rate objective function generated based on power grid dispatch instructions, and the multi-objective optimization solution further includes: Obtain the load dispatch command of the boiler at the current moment, and convert the target load change rate required by the command into the load response rate objective function; The load response rate objective function is weighted and combined with the highest combustion efficiency objective function and the lowest pollutant emission objective function to form a comprehensive fitness function; During the multi-objective evolutionary algorithm search process, the comprehensive fitness function guides the population evolution direction to obtain a non-dominated solution set that simultaneously considers load following rate, combustion efficiency and emission performance, and the Pareto optimal control command set is updated by the non-dominated solution set.

9. A boiler combustion control system based on machine learning, characterized in that, include: The time-series data acquisition module is used to acquire the boiler's operating parameter sequence and flue gas emission monitoring sequence in multiple continuous combustion adjustment cycles. After time-series alignment and standardization, a multi-dimensional time-series operation dataset is constructed. The feature extraction module is used to perform spatiotemporal coupled feature extraction on the multi-dimensional time-series running dataset to generate a combustion state evolution feature tensor. The evolution prediction module has an embedded combustion process evolution prediction model trained with historical combustion operation data. It is used to input the combustion state evolution feature tensor to obtain the predicted distribution of combustion state at future times. The multi-objective optimization module is used to solve the multi-objective optimization problem on the predicted distribution of combustion state at the future time using a multi-objective optimization function, and generate a Pareto optimal control instruction set. The instruction selection module is used to select combustion control instructions from the Pareto optimal control instruction set, based on the current operating condition safety constraints and the actuator adjustment margin, and send them to the boiler actuator.

Images