Intelligent decision-making method and system for coal power combustion based on deep learning

Abstract

This invention relates to the field of coal-fired power plant combustion technology, and more particularly to a deep learning-based intelligent decision-making method and system for coal-fired power plant combustion. The method collects combustion system parameters to form a sequence of operating condition data, extracts combustion features through multi-scale convolution and generates a weighted feature representation, then identifies the current combustion state and predicts its evolution through time-series processing, calculates the trajectory of combustion efficiency and emission concentration changes, and finally, under trajectory constraints, uses reinforcement learning to search for the optimal oxygen supply, fuel supply rate, and damper opening adjustment, generating control commands to regulate the combustion process, significantly improving combustion efficiency and ensuring that emission concentrations meet standards.

Description

Technical Field

[0001] This invention relates to the field of coal-fired power plant combustion technology, and in particular to a deep learning-based intelligent decision-making method and system for coal-fired power plant combustion. Background Technology

[0002] In the field of coal-fired power plant combustion control, traditional technologies generally rely on mechanistic model-based methods or fixed empirical rules. The conventional approach typically involves establishing simplified thermodynamic or kinetic equations, collecting data such as fuel parameters, airflow parameters, temperature, and pressure through sensors, and then controlling oxygen supply, fuel supply rate, and damper opening using manually set thresholds or PID control strategies. Some schemes introduce historical data statistical regression or fuzzy control to attempt to compensate for model uncertainties, but these are essentially still static or quasi-static response frameworks. While these methods can maintain basic operation under steady-state conditions, they struggle to cope with complex dynamic scenarios such as coal quality fluctuations and sudden load changes.

[0003] Existing conventional methods have a significant drawback: a lack of effective capture and utilization of the dynamic characteristics of the combustion process across multiple time scales. The fuel properties, airflow distribution, and thermodynamic state of the combustion system change coupled across different time scales. For example, the combustion reaction of fuel particles is completed on a second-by-second basis, while changes in furnace temperature distribution appear on a minute-by-minute basis. Traditional methods typically employ fixed sampling periods or single-scale feature extraction, failing to simultaneously identify combustion modes across different time spans, leading to delays or biases in the perception of evolving operating conditions. This limitation on time scales makes it difficult for control commands to match the actual state of the current combustion system in real time, especially during transient transitions, which can easily result in overshoot or instability.

[0004] Another drawback lies in the short-sightedness of control decisions. Existing technologies mostly rely on feedback corrections based on deviations at the current moment, lacking the ability to predict future changes in the state. For example, when combustion efficiency decreases, conventional methods immediately adjust the oxygen supply without considering the impact of this adjustment on emission concentrations or furnace temperature trajectories in the following minutes. This localized optimization strategy often sacrifices long-term performance, making it difficult to balance combustion efficiency and emission control, frequently resulting in over-adjustment or under-adjustment. Furthermore, the lack of quantitative assessment of the evolution direction of the combustion system state leads to poor coordination of control actions, failing to proactively explore potential areas for efficiency improvement while meeting emission limits. Summary of the Invention

[0005] This invention provides a deep learning-based intelligent decision-making method and system for coal-fired power plant combustion, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a deep learning-based intelligent decision-making method for coal-fired power plant combustion, comprising:

[0007] Collect fuel characteristic parameters, airflow dynamic parameters, thermodynamic state parameters and emission concentration parameters of the combustion system to form a sequence of operating condition data including timestamps;

[0008] By processing the working condition data sequence through convolution operations at different scales, combustion features at different time spans are extracted to generate multi-scale feature representations.

[0009] Calculate the correlation strength between the features of each time step in the multi-scale feature representation and the stability of the current combustion condition, and assign adaptive weights to the features of each time step according to the correlation strength to generate a weighted feature representation;

[0010] The weighted feature representation is processed temporally, and the current combustion state is identified and the direction of state change in future moments is predicted through the state transfer relationship between consecutive time steps, generating state evolution prediction information.

[0011] The trajectory of changes in combustion efficiency and emission concentration at multiple future time steps is calculated based on state evolution prediction information.

[0012] Under the constraint of the changing trajectory, the oxygen supply, fuel supply rate and damper opening are adjusted to maximize combustion efficiency and meet emission concentration limits through reinforcement learning strategy search. Control decision commands are generated and sent to the actuators to regulate the combustion process.

[0013] By processing the operating condition data sequence through convolution operations at different scales, combustion features at different time spans are extracted to generate multi-scale feature representations, including:

[0014] The operating condition data sequence is divided into multiple data windows according to the time dimension. The fluctuation variance of combustion efficiency in each data window is calculated. The receptive field range of the convolution kernel in the time dimension is determined according to the fluctuation variance. Convolution operation is performed on each data window along the time dimension to generate transient combustion features and evolutionary combustion features.

[0015] The error contribution values ​​of the transient combustion characteristics and the evolutionary combustion characteristics to the combustion efficiency prediction are calculated respectively, and the complementarity metric is calculated based on the difference between the error contribution values.

[0016] The transient fusion weight is calculated based on the ratio of the complementarity metric to the error contribution value of the transient combustion feature, and the evolutionary fusion weight is calculated based on the ratio of the complementarity metric to the error contribution value of the evolutionary combustion feature.

[0017] The transient combustion feature is multiplied by the transient fusion weight to obtain the weighted transient combustion feature, the evolved combustion feature is multiplied by the evolved fusion weight to obtain the weighted evolved combustion feature, and the weighted transient combustion feature and the weighted evolved combustion feature are concatenated to obtain the fused combustion feature.

[0018] The fusion combustion features are subjected to a nonlinear transformation to generate a multi-scale feature representation.

[0019] The correlation strength between the features at each time step in the multi-scale feature representation and the stability of the current combustion condition is calculated. Adaptive weights are assigned to the features at each time step based on the correlation strength, generating a weighted feature representation including:

[0020] Calculate the time interval between the time corresponding to each time step feature in the multi-scale feature representation and the current time, and calculate the time decay factor based on the time interval;

[0021] Calculate the rate of change of combustion efficiency and the rate of change of emission concentration at each time step corresponding to the characteristic, and calculate the intensity of operating condition fluctuation based on the rate of change of combustion efficiency and the rate of change of emission concentration;

[0022] Identify time steps in the operating condition fluctuation intensity that exceed a preset fluctuation threshold, calculate the time interval distribution characteristics between the time steps that exceed the preset fluctuation threshold and the current time, and calculate the fluctuation propagation delay coefficient based on the time interval distribution characteristics;

[0023] The correlation strength between the characteristics of each time step and the stability of the current combustion condition is calculated based on the time decay factor of each time step characteristic, the corresponding operating condition fluctuation intensity, and the fluctuation propagation delay coefficient.

[0024] The correlation strength is normalized to obtain the adaptive weights of the features at each time step. The features at each time step in the multi-scale feature representation are weighted according to the adaptive weights and then combined in chronological order to generate a weighted feature representation.

[0025] The weighted feature representation is processed temporally to identify the current combustion state and predict the direction of future state changes through the state transit relationship between consecutive time steps, generating state evolution prediction information including:

[0026] The weighted feature representation is processed temporally, the feature change gradient of the weighted feature representation between consecutive time steps is calculated, and the time step in which the absolute value of the feature change gradient exceeds the preset gradient threshold is identified as the combustion state transition time step.

[0027] Obtain the weighted feature representation corresponding to the combustion state transition time step, extract the combustion efficiency value and emission concentration value from the weighted feature representation corresponding to the combustion state transition time step, and identify the current combustion state according to the preset distribution interval to which the combustion efficiency value and emission concentration value belong;

[0028] Weighted features are extracted to represent the continuous time step feature sequence from the combustion state transition time step to the current time step. The state transfer relationship between continuous time steps is constructed through the continuous time step feature sequence. The state evolution trend function is obtained by fitting the state transfer relationship between continuous time steps.

[0029] The future time is used as a time parameter to input the state evolution trend function to calculate the predicted feature value of the future time. The difference between the predicted feature value of the future time and the weighted feature representation of the current time is calculated. The direction of state change in the future time is predicted based on the positive or negative sign of the difference.

[0030] The current combustion state and the direction of future state changes are combined to generate state evolution prediction information.

[0031] Constructing the state transfer relationship between consecutive time steps using feature sequences of consecutive time steps includes:

[0032] Calculate the feature difference between adjacent time step features in a continuous time step feature sequence, and calculate the state transition rate between adjacent time steps based on the feature difference;

[0033] Calculate the time span between the time corresponding to each time step feature in the continuous time step feature sequence and the time corresponding to the combustion state transition time step, and identify the time step in which the state transition rate fluctuates and reverses within the time span as the state evolution segment point.

[0034] The continuous time step feature sequence is divided into multiple evolution stage subsequences according to the state evolution segmentation point. The stage evolution rate corresponding to each evolution stage subsequence is calculated. The state transfer weight between each time step is calculated according to the time span and the stage evolution rate. The continuous time step feature sequence and the state transfer weight are combined to construct the state transfer relationship between continuous time steps.

[0035] Based on state evolution prediction information, the trajectory of changes in combustion efficiency and emission concentrations at multiple future time steps is calculated, including:

[0036] Extract the current combustion state and direction of state change from the state evolution prediction information, and obtain the baseline combustion efficiency and baseline emission concentration based on the current combustion state;

[0037] The weighted feature is calculated to represent the feature change between consecutive time steps. The evolution rate is calculated based on the correlation between the feature change and combustion efficiency and emission concentration. The direction of state change is converted into an evolution direction coefficient. The evolution rate is multiplied by the evolution direction coefficient to obtain the directional evolution rate.

[0038] The combustion efficiency and emission concentration for the future time steps are calculated by multiplying the directed evolution rate by the time step size of the future time steps and then superimposing the baseline combustion efficiency and baseline emission concentration.

[0039] The system identifies points where the difference between adjacent time steps in combustion efficiency and emission concentration exceeds a preset fluctuation threshold as segmentation points. Smoothing is then performed within the intervals defined by these segmentation points to generate the change trajectories of combustion efficiency and emission concentration for multiple future time steps.

[0040] Under the constraint of the changing trajectory, the control decision commands are generated by searching for oxygen supply, fuel supply rate, and damper opening to maximize combustion efficiency and meet emission concentration limits through reinforcement learning strategies.

[0041] The combustion efficiency constraint value and emission concentration constraint value for each time step are extracted from the trajectory of combustion efficiency change and emission concentration change over multiple future time steps.

[0042] Construct a coupling relationship surface between oxygen supply regulation and combustion efficiency constraints and emission concentration constraints. Identify the critical emission concentration value corresponding to the maximum combustion efficiency on the coupling relationship surface. Calculate the concentration gap between the critical emission concentration value and the emission concentration constraints. Based on the concentration gap, solve the compensation combination of fuel supply rate regulation and damper opening regulation to maximize combustion efficiency and meet emission concentration limits.

[0043] The oxygen supply rate adjustment, fuel supply rate adjustment, and damper opening adjustment obtained from the inverse solution are used as the initial search center point. A local search is performed in the neighborhood of the initial search center point using a reinforcement learning strategy to identify the oxygen supply rate adjustment, fuel supply rate adjustment, and damper opening adjustment that maximize combustion efficiency and meet emission concentration limits.

[0044] The identified oxygen supply adjustment, fuel supply rate adjustment, and damper opening adjustment are encapsulated into control decision commands.

[0045] A second aspect of this invention provides a deep learning-based intelligent decision-making system for coal-fired power plant combustion, comprising:

[0046] The data acquisition unit is used to collect fuel characteristic parameters, airflow dynamic parameters, thermodynamic state parameters and emission concentration parameters of the combustion system to form a sequence of operating condition data including timestamps.

[0047] Multi-scale feature units are used to process operating condition data sequences through convolution operations at different scales, extract combustion features at different time spans, and generate multi-scale feature representations.

[0048] An adaptive additive unit is used to calculate the correlation strength between the features of each time step in the multi-scale feature representation and the stability of the current combustion condition. Based on the correlation strength, an adaptive weight is assigned to each time step feature to generate a weighted feature representation.

[0049] The state prediction unit is used to perform time-series processing on the weighted feature representation, identify the current combustion state through the state transfer relationship between consecutive time steps, predict the direction of state change at future moments, and generate state evolution prediction information.

[0050] The trajectory calculation unit is used to calculate the trajectory of changes in combustion efficiency and emission concentration at multiple future time steps based on state evolution prediction information.

[0051] The control decision unit is used to search for oxygen supply, fuel supply rate, and damper opening adjustments to maximize combustion efficiency and meet emission concentration limits under the constraints of the changing trajectory through reinforcement learning strategies. It generates control decision commands and sends them to the actuators to regulate the combustion process.

[0052] A third aspect of the present invention provides an electronic device, comprising:

[0053] processor;

[0054] Memory used to store processor-executable instructions;

[0055] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0057] This invention calculates the correlation strength between features at each time step and combustion stability and assigns adaptive weights, making the weighted feature representation highlight current key influencing factors, effectively suppressing noise interference and redundant information, and providing high-quality input for subsequent state identification. Based on the time-series processing and state transmission relationship of the weighted feature representation, the current combustion state can be accurately identified and the future state evolution direction can be predicted. The generated combustion efficiency and emission concentration change trajectories reflect the dynamic trend of the combustion process in advance. This prediction mechanism enables the system to anticipate fluctuations in operating conditions, avoiding unstable combustion or emission exceedances due to response lag, and significantly improving the adaptability of the combustion system to complex dynamic operating conditions. Under the constraint of the change trajectory, the reinforcement learning strategy is used to simultaneously search for oxygen supply adjustment, fuel supply rate adjustment, and damper opening adjustment, not only pursuing maximum combustion efficiency but also strictly controlling emission concentration within limits, effectively balancing economic benefits and environmental protection requirements. Compared with traditional single-objective or empirical adjustment methods, it can avoid getting trapped in local optima, significantly reduce the generation of pollutants such as nitrogen oxides, and improve thermal efficiency. Once the generated control commands are sent directly to the actuators, the combustion process can be adjusted in real time in a closed loop, allowing it to continue operating at its optimal condition without human intervention. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating a deep learning-based intelligent decision-making method for coal-fired power plant combustion.

[0059] Figure 2 Flowchart for synergistic optimization control of combustion efficiency and emission concentration. Detailed Implementation

[0060] 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, and 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.

[0061] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0062] Figure 1 This is a flowchart illustrating the intelligent decision-making method for coal-fired power plant combustion based on deep learning, as described in an embodiment of the present invention.

[0063] Deep learning-based intelligent decision-making methods for coal-fired power plant combustion include:

[0064] Collect fuel characteristic parameters, airflow dynamic parameters, thermodynamic state parameters and emission concentration parameters of the combustion system to form a sequence of operating condition data including timestamps;

[0065] By processing the working condition data sequence through convolution operations at different scales, combustion features at different time spans are extracted to generate multi-scale feature representations.

[0066] Calculate the correlation strength between the features of each time step in the multi-scale feature representation and the stability of the current combustion condition, and assign adaptive weights to the features of each time step according to the correlation strength to generate a weighted feature representation;

[0067] The weighted feature representation is processed temporally, and the current combustion state is identified and the direction of state change in future moments is predicted through the state transfer relationship between consecutive time steps, generating state evolution prediction information.

[0068] The trajectory of changes in combustion efficiency and emission concentration at multiple future time steps is calculated based on state evolution prediction information.

[0069] Under the constraint of the changing trajectory, the oxygen supply, fuel supply rate and damper opening are adjusted to maximize combustion efficiency and meet emission concentration limits through reinforcement learning strategy search. Control decision commands are generated and sent to the actuators to regulate the combustion process.

[0070] By processing the operating condition data sequence through convolution operations at different scales, combustion features at different time spans are extracted to generate multi-scale feature representations, including:

[0071] The operating condition data sequence is divided into multiple data windows according to the time dimension. The fluctuation variance of combustion efficiency in each data window is calculated. The receptive field range of the convolution kernel in the time dimension is determined according to the fluctuation variance. Convolution operation is performed on each data window along the time dimension to generate transient combustion features and evolutionary combustion features.

[0072] The error contribution values ​​of the transient combustion characteristics and the evolutionary combustion characteristics to the combustion efficiency prediction are calculated respectively, and the complementarity metric is calculated based on the difference between the error contribution values.

[0073] The transient fusion weight is calculated based on the ratio of the complementarity metric to the error contribution value of the transient combustion feature, and the evolutionary fusion weight is calculated based on the ratio of the complementarity metric to the error contribution value of the evolutionary combustion feature.

[0074] The transient combustion feature is multiplied by the transient fusion weight to obtain the weighted transient combustion feature, the evolved combustion feature is multiplied by the evolved fusion weight to obtain the weighted evolved combustion feature, and the weighted transient combustion feature and the weighted evolved combustion feature are concatenated to obtain the fused combustion feature.

[0075] The fusion combustion features are subjected to a nonlinear transformation to generate a multi-scale feature representation.

[0076] When the operating condition data sequence is divided into multiple data windows along the time dimension, each window corresponds to a continuous time segment during the combustion process. The granularity of the window division determines the range of time span that subsequent convolution operations can perceive. For each data window, the variance of combustion efficiency fluctuation within that window is calculated. Specifically, let the first There are a total of windows The time step, the first The combustion efficiency at each time step is The average combustion efficiency of this window is The variance of the fluctuation is .when A larger value indicates that the combustion conditions within that window fluctuate drastically, requiring a smaller sensing field to capture local transient information; when A smaller receptive field indicates a relatively stable combustion process, making a larger receptive field suitable for capturing long-term evolutionary trends. Therefore, the receptive field range... and It exhibits an inverse mapping relationship, that is ,in and These are the preset upper and lower limits of the receptive field, respectively. and These represent the maximum and minimum values ​​of the variance across all windows. Through this adaptive receptive field mechanism, the convolutional kernel can flexibly adjust its time perception range according to the dynamic characteristics of actual combustion conditions, thereby focusing on short-term local features when conditions fluctuate drastically and taking into account long-term evolution patterns when conditions are stable.

[0077] When performing convolution operations along the time dimension for each data window, two types of convolution kernels with different scales are used for processing. Small-scale convolution kernels correspond to shorter receptive fields and can sensitively capture transient perturbation information during combustion, such as local temperature peaks caused by sudden changes in oxygen supply and short-term combustion rate fluctuations caused by changes in fuel particle size. Transient combustion features are obtained after small-scale convolution operations. Large-scale convolution kernels correspond to longer receptive fields, enabling them to perceive the gradual evolution of the combustion process over a longer time span, such as the slow drift of the furnace temperature field and the cumulative change trend of fly ash carbon content over time. Evolutionary combustion characteristics are obtained after large-scale convolution operations. The two types of features characterize the combustion process at different time scales, providing a complementary information basis for subsequent fusion.

[0078] In obtaining transient combustion characteristics and evolutionary combustion characteristics Next, the error contribution values ​​of the two types of features to the combustion efficiency prediction were calculated separately. Specifically, the following were performed: The value is input separately into the combustion efficiency prediction head, and the mean square error between the predicted result and the actual combustion efficiency is calculated and recorded as the transient error contribution value. ;Will Each input is fed into the same prediction head, and its corresponding mean square error is calculated and recorded as the evolution error contribution value. A larger error contribution value indicates a greater inadequacy of that type of feature when performing a prediction task alone, but it also implies a strong complementary potential between it and another type of feature. Complementarity measure. The calculation comprehensively considers the degree of difference between the error contribution values ​​of the two types of features, specifically expressed as follows: The greater the difference, the stronger the complementarity of the two types of features in terms of information content, and the more significant the gain after fusion; the smaller the difference, the more similar the information carried by the two types of features, and the limited complementary effect.

[0079] Based on complementarity metric The transient fusion weights are calculated by dividing each weight by its respective error contribution value. and evolutionary fusion weight The transient fusion weights are calculated as follows: The calculation method for evolutionary fusion weight is as follows: ,in To prevent extremely small positive numbers with a denominator of zero, the calculation logic is as follows: the larger the error contribution value of a certain type of feature, the more obvious its inadequacy in individual prediction; and the larger the ratio to the complementarity metric, the higher its fusion weight is assigned, prompting the fusion result to rely more on this type of feature to compensate for the information blind spots of the other type of feature. When the error contribution values ​​of the two types of features differ significantly, the feature with the larger error contribution value will receive a higher fusion weight, tilting the fusion result towards the direction with scarcer information, thus fully utilizing the complementarity of the two types of features. When the errors of the two types of features are similar, the fusion weights tend to be balanced, and the contributions of the two types of features to the fusion result are comparable.

[0080] transient combustion characteristics With transient fusion weights Element-wise multiplication yields the weighted transient combustion characteristics. ; Evolutionary combustion characteristics With evolutionary fusion weight Element-wise multiplication yields the weighted evolutionary combustion characteristics. Weighted transient combustion characteristics Weighted evolution combustion characteristics By splicing along the feature dimensions, the fused combustion features are obtained. ,in This represents a concatenation operation along the feature channel dimension. The concatenation operation preserves the complete information structure of each type of feature, avoiding the problem of feature information cancellation that may occur in direct addition fusion, while providing a richer input representation for subsequent nonlinear transformations.

[0081] Characteristics of fusion combustion When performing nonlinear transformations, a multi-layer fully connected network combined with an activation function is used. Specifically, the nonlinear transformation is implemented by... The model sequentially passes through a linear transformation layer and a nonlinear activation layer. The linear transformation layer reduces the dimensionality of the concatenated high-dimensional features, mapping the concatenated space of transient and evolutionary features to a unified low-dimensional representation space, eliminating feature redundancy introduced by concatenation. The nonlinear activation layer uses the ReLU function to perform element-wise nonlinear mapping on the linear transformation result, enhancing the model's ability to express the complex nonlinear dynamics of the combustion process. After the nonlinear transformation, a multi-scale feature representation is finally generated. This representation simultaneously encodes the transient perturbation characteristics and long-term evolution trends of the combustion process in a unified feature space, providing high-quality input features for subsequent adaptive weight allocation based on attention mechanisms and time-series state modeling, thereby supporting the accurate execution of combustion efficiency prediction and emission concentration prediction. In practical engineering applications, the pulverized coal combustion process in the furnace often involves millisecond-level local turbulence disturbances and minute-level overall operating condition drift. Single-scale feature extraction methods are difficult to cover both simultaneously. However, the above-mentioned multi-scale feature generation mechanism, through adaptive receptive field selection and complementary weight fusion, can effectively cover the combustion dynamic characteristics at different time scales, significantly improving the adaptability of the subsequent decision model to complex operating conditions.

[0082] The correlation strength between the features at each time step in the multi-scale feature representation and the stability of the current combustion condition is calculated. Adaptive weights are assigned to the features at each time step based on the correlation strength, generating a weighted feature representation including:

[0083] Calculate the time interval between the time corresponding to each time step feature in the multi-scale feature representation and the current time, and calculate the time decay factor based on the time interval;

[0084] Calculate the rate of change of combustion efficiency and the rate of change of emission concentration at each time step corresponding to the characteristic, and calculate the intensity of operating condition fluctuation based on the rate of change of combustion efficiency and the rate of change of emission concentration;

[0085] Identify time steps in the operating condition fluctuation intensity that exceed a preset fluctuation threshold, calculate the time interval distribution characteristics between the time steps that exceed the preset fluctuation threshold and the current time, and calculate the fluctuation propagation delay coefficient based on the time interval distribution characteristics;

[0086] The correlation strength between the characteristics of each time step and the stability of the current combustion condition is calculated based on the time decay factor of each time step characteristic, the corresponding operating condition fluctuation intensity, and the fluctuation propagation delay coefficient.

[0087] The correlation strength is normalized to obtain the adaptive weights of the features at each time step. The features at each time step in the multi-scale feature representation are weighted according to the adaptive weights and then combined in chronological order to generate a weighted feature representation.

[0088] After obtaining the multi-scale feature representations, it is necessary to evaluate the importance of the features at each time step so that the model can pay more attention to historical moments closely related to the stability of the current combustion conditions in subsequent processing. The core idea of ​​the evaluation is that historical features closer to the current time usually have a more direct impact on the current conditions, while drastic fluctuations that occur at historical moments will affect the current state with a delay through the thermodynamic inertia and flow field propagation effects of the combustion system. Therefore, it is necessary to consider three dimensions simultaneously: time distance, degree of fluctuation in operating conditions, and the delayed characteristics of fluctuation propagation, to comprehensively evaluate the correlation strength.

[0089] For each time step in the multi-scale feature representation, first calculate the time interval between the corresponding time step and the current time step. Let the th time step be... The time corresponding to each time step is The current time is The time interval is Based on this time interval, the time decay factor is calculated using an exponential decay method. Specifically ,in The decay rate hyperparameter controls the rate at which historical information is forgotten. When... When the value is large, the model tends to focus on the most recent time step; when When the value is small, the model retains more weight for features from more distant historical times. The specific value can be determined based on the thermodynamic response time constant of the combustion system, so that the decay curve matches the physical inertia of the system. Time decay factor The range of values ​​is The current time step has a value of 1, which decreases monotonically as the time interval increases.

[0090] When calculating the intensity of operating condition fluctuations, the rate of change in combustion efficiency and the rate of change in emission concentration are extracted for each time step. The rate of change in combustion efficiency reflects whether the combustion process is in a rapidly changing state at that moment, while the rate of change in emission concentration reflects the severity of the pollutant formation rate. Let the... The rate of change in combustion efficiency at each time step is The rate of change in emission concentration Then the intensity of the operating condition fluctuation Defined as a weighted combination of the two: ,in and These are the weighting coefficients for the rate of change in combustion efficiency and the rate of change in emission concentration, respectively. These can be set according to the relative importance placed on combustion efficiency and emission indicators in actual production. Using absolute values ​​ensures that the intensity of operating condition fluctuations remains non-negative; regardless of the direction of change, drastic changes should be assigned a higher fluctuation intensity value. Operating Condition Fluctuation Intensity The larger the value, the more unstable the combustion conditions at that time step are, and the historical characteristics at that time have a stronger reference value for understanding the stability of the current conditions.

[0091] Identify the intensity of operating condition fluctuations Exceeding the preset fluctuation threshold After settling the time steps, further analysis is needed to determine the propagation delay characteristics of these fluctuation events at the current moment. Because the propagation of the heat field, flue gas flow, and fuel mixing within the combustion furnace all involve physical delays, fluctuation events occurring in the past will not immediately manifest their full impact at the current moment; rather, there is a propagation delay. All events satisfying... The time steps are denoted as the set of fluctuation events. The time interval sequence between these fluctuation event moments and the current moment is calculated, and the mean of this time interval sequence is extracted. and standard deviation As a characteristic of time interval distribution, the wave propagation delay coefficient. The calculation is performed separately for each time step. The underlying idea is: if the time interval of a certain time step... It falls precisely within the typical range of wave propagation delay (i.e., close to) If the time step is such that its features are more likely to carry direct antecedent information about the stability of the current operating condition, then it should be assigned a higher correlation weight. Specifically, Calculated using Gaussian kernel form: ,in To prevent extremely small positive numbers with a denominator of zero from being compared with the defined... They have the same numerical magnitude but are declared separately here to maintain the independence of the parameter's meaning. exist It reaches its maximum value of 1 at time, and then... Deviation This reduces the time step, thus allowing the time step, which aligns with the historical pattern of wave propagation delay, to achieve a higher wave propagation delay coefficient.

[0092] Combining the quantitative results from the three dimensions, the first... The correlation strength between time step characteristics and the stability of current combustion conditions Defined as the product of the time decay factor, the intensity of the operating condition fluctuation, and the fluctuation propagation delay coefficient: This design constrains the correlation strength simultaneously by three factors: time distance, intensity of operating conditions, and matching propagation delay. The time decay factor ensures priority for recent characteristics, the intensity of operating condition fluctuations ensures that characteristics at moments of drastic change are valued, and the fluctuation propagation delay coefficient further filters out historical moments that are most likely to affect the current operating conditions in terms of physical propagation laws. The multiplication of these three factors ensures that only time steps that simultaneously satisfy the requirements of being close in time, having significant fluctuations, and matching delays can achieve the highest correlation strength, avoiding bias caused by single-dimensional evaluation.

[0093] Correlation strength over all time steps Normalization is performed to obtain the adaptive weights of the features at each time step. The normalization formula is The summation iterates through all time steps in the multi-scale feature representation. After normalization, the sum of the adaptive weights for all time steps is 1, and each weight value is non-negative, satisfying the basic properties of the probability distribution, which facilitates subsequent weighted operations with the feature vector.

[0094] Adaptive weights With the multi-scale feature vector at the corresponding time step Multiply to obtain the weighted eigenvector. All weighted feature vectors are arranged sequentially according to time, forming a weighted feature representation sequence. This sequence preserves the original time step order structure, ensuring that subsequent time series processing modules can correctly perceive the temporal relationship of features. Simultaneously, through adaptive weight modulation, historical moments with higher correlation to the stability of the current combustion condition have larger feature amplitudes in the sequence, thus obtaining stronger gradient signals during the time series modeling stage and guiding the model to focus on learning the state transition patterns of key historical moments.

[0095] The weighted feature representation is processed temporally to identify the current combustion state and predict the direction of future state changes through the state transit relationship between consecutive time steps, generating state evolution prediction information including:

[0096] The weighted feature representation is processed temporally, the feature change gradient of the weighted feature representation between consecutive time steps is calculated, and the time step in which the absolute value of the feature change gradient exceeds the preset gradient threshold is identified as the combustion state transition time step.

[0097] Obtain the weighted feature representation corresponding to the combustion state transition time step, extract the combustion efficiency value and emission concentration value from the weighted feature representation corresponding to the combustion state transition time step, and identify the current combustion state according to the preset distribution interval to which the combustion efficiency value and emission concentration value belong;

[0098] Weighted features are extracted to represent the continuous time step feature sequence from the combustion state transition time step to the current time step. The state transfer relationship between continuous time steps is constructed through the continuous time step feature sequence. The state evolution trend function is obtained by fitting the state transfer relationship between continuous time steps.

[0099] The future time is used as a time parameter to input the state evolution trend function to calculate the predicted feature value of the future time. The difference between the predicted feature value of the future time and the weighted feature representation of the current time is calculated. The direction of state change in the future time is predicted based on the positive or negative sign of the difference.

[0100] The current combustion state and the direction of future state changes are combined to generate state evolution prediction information.

[0101] When performing time-series processing on weighted feature representations, the first step is to calculate the feature change gradient between consecutive time steps to capture abrupt changes in the combustion process. Specifically, for the weighted feature vectors of two adjacent time steps, their difference is calculated as a local gradient estimate, and the absolute value of this gradient is compared with a preset gradient threshold. Compare them. When the absolute value of the feature change gradient at a certain time step exceeds... At this time, the time step is marked as the combustion state transition time step. Preset gradient threshold. The gradient is determined based on the statistical distribution of historical operating data, typically by adding a certain number of standard deviations to the mean of the historical gradient sequence. This ensures that only genuine changes in operating conditions are captured, rather than noise disturbances. For multi-dimensional weighted feature vectors, the absolute value of the gradient is taken as the L2 norm of the gradient in each dimension as a comprehensive measure, thereby avoiding interference from local fluctuations in a single dimension in identifying inflection points.

[0102] After identifying the combustion state transition time step, the weighted feature representation corresponding to that time step is extracted, and the combustion efficiency value is read from it. and emission concentration values Based on the design specifications and historical operating data of the combustion system, the range of values ​​for combustion efficiency and emission concentration is pre-defined into several distribution intervals. For example, combustion efficiency is divided into high-efficiency, medium-efficiency, and low-efficiency intervals, and emission concentration is divided into compliant, critical, and non-compliant intervals. and Each combustion state is matched with its corresponding distribution interval, and the category label is determined based on the combination of the intervals to which the two states belong. For example, when the combustion efficiency is in the high-efficiency interval and the emission concentration is in the compliant interval, it is identified as a "high-efficiency and compliant" combustion state; when the combustion efficiency is in the low-efficiency interval or the emission concentration is in the excessive interval, it is identified as an "abnormal" combustion state. This distribution interval-based state identification method can map continuous numerical features to discrete state categories, which facilitates subsequent combination with information on the direction of state changes.

[0103] After identifying the current combustion state, extract the continuous time step feature sequence from the combustion state transition time step to the current time step, and denote the total number of time steps contained in this sequence as . , No. The time corresponding to each time step is The corresponding weighted feature scalar representation (e.g., taking a representative component of the feature vector or its norm) is: ,in Using this time-feature to analyze the sequence A state transfer relationship is constructed between consecutive time steps, treating the eigenvalues ​​of adjacent time steps as input-output pairs for state transition, forming a state transfer sample set. This sample set is then fitted with a function, selecting a form that describes the evolution of the combustion state, such as a polynomial, exponential, or piecewise linear function. The fitting process employs the least squares criterion, aiming to minimize the sum of squared residuals between the predicted and observed values ​​at each time step. The function parameters are then solved to obtain the state evolution trend function. ,in This is a time parameter. The order or complexity of the fitting function depends on the sequence length. Adaptive selection: when the sequence is short, a low-order polynomial is used to prevent overfitting; when the sequence is long and shows a clear nonlinear trend, a high-order or nonlinear function form is used.

[0104] Obtain the state evolution trend function Then, the future moment Substituting the time parameter into the function, the predicted feature value for the future time is calculated. Future moments It can be set to one or more prediction step sizes after the current time, and the size of the prediction step size is determined according to the actual control requirements. Calculate the predicted eigenvalues ​​for future times. The corresponding scalar value of the weighted feature representation at the current time. The difference between Based on the difference The sign of the sign determines the direction of state change in the future: when When the predicted state change direction is "upward", it indicates that the combustion characteristic value will increase; when When the predicted state change direction is "decreasing", it indicates that the combustion characteristic value will decrease; when The absolute value is less than the preset direction judgment threshold. When the predicted state change direction is "stable," it indicates that the combustion characteristic values ​​remain basically stable within the prediction time domain. Direction judgment threshold. The resolution is determined by combining the sensor measurement accuracy and the control system response, in order to filter out minor trends that are too small to have a substantial impact on control decisions.

[0105] For scenarios requiring simultaneous prediction of the changing directions of two key indicators—combustion efficiency and emission concentration—the aforementioned state evolution trend function fitting and difference calculation process is performed independently on the combustion efficiency characteristic component and the emission concentration characteristic component, respectively, yielding two independent prediction results for the changing directions. The difference between the combustion efficiency characteristic components is denoted as... The difference between the characteristic components of the emission concentration is denoted as The positive and negative signs of the two indicators correspond to the future direction of change of their respective indicators. This method of predicting indicators independently can more precisely depict the possible inverse relationship between efficiency and emissions during combustion, such as the situation where combustion efficiency increases and emission concentrations rise simultaneously under certain operating conditions, thus providing more discriminative predictive information for subsequent reinforcement learning strategies.

[0106] The current combustion state category label is combined with the predicted direction of future state changes to generate state evolution prediction information. This prediction information is presented in a structured format, including the current combustion state category, the direction of combustion efficiency change, the direction of emission concentration change, and the corresponding prediction time. This structured information serves as input for subsequent calculations of combustion efficiency and emission concentration change trajectories. It also acts as part of the state input for the reinforcement learning strategy network, enabling the network to make forward-looking control adjustments based on trend direction information, in addition to understanding the current operating condition category. This allows for intervention measures to be taken before adverse changes in combustion conditions occur, improving the timeliness and effectiveness of overall control.

[0107] Constructing the state transfer relationship between consecutive time steps using feature sequences of consecutive time steps includes:

[0108] Calculate the feature difference between adjacent time step features in a continuous time step feature sequence, and calculate the state transition rate between adjacent time steps based on the feature difference;

[0109] Calculate the time span between the time corresponding to each time step feature in the continuous time step feature sequence and the time corresponding to the combustion state transition time step, and identify the time step in which the state transition rate fluctuates and reverses within the time span as the state evolution segment point.

[0110] The continuous time step feature sequence is divided into multiple evolution stage subsequences according to the state evolution segmentation point. The stage evolution rate corresponding to each evolution stage subsequence is calculated. The state transfer weight between each time step is calculated according to the time span and the stage evolution rate. The continuous time step feature sequence and the state transfer weight are combined to construct the state transfer relationship between continuous time steps.

[0111] After obtaining the weighted feature representation, it is necessary to further characterize the state transfer relationships between consecutive time steps in order to accurately capture the dynamic evolution law in the combustion process. Therefore, a quantitative analysis is performed on the feature changes between adjacent time steps in the consecutive time step feature sequence. Suppose the consecutive time step feature sequence contains... The time step, the first The weighted feature scalar value corresponding to each time step is Adjacent time steps and The feature difference between them is denoted as Then there is Based on this, combined with the actual time interval between adjacent time steps... (Right now ), calculate the state transition rate between adjacent time steps. The expression is State transition rate It reflects the rate at which the state of the combustion system changes between adjacent moments: when A large absolute value indicates that the combustion state has changed drastically during that period; when When the rate approaches zero, it indicates that the combustion system is in a relatively stable state. By calculating the state transition rate for each adjacent time step, a rate sequence describing the dynamic changes in the combustion state can be obtained, laying the foundation for subsequent segmented analysis.

[0112] After obtaining the state transition rate sequence, it is necessary to identify the key turning points in the combustion state evolution process, i.e., the state evolution segmentation points. Specifically, this involves calculating the corresponding time steps in the continuous time step feature sequence. The time corresponding to the transition step of the combustion state Time span between ,Right now The identification of the combustion state transition time step depends on the combustion efficiency value. and emission concentration values The extreme value locations in the historical sequence, and the preset gradient threshold. Constraints on the gradient of state changes are determined. Based on a defined time span, time steps in the rate sequence where "fluctuation reversal" occurs in the state transition rate are further detected; these are points where the sign of adjacent rate values ​​changes and the magnitude of the change exceeds a certain tolerance range. These time steps that satisfy the fluctuation reversal condition are marked as state evolution segment points, and the set of segment points is denoted as... ,in This represents the total number of segment points. The criterion for judging a fluctuation reversal is: if... and If the preset reversal amplitude threshold is exceeded, then the first... Each time step is used as a segmentation point. This mechanism can effectively distinguish between the steady evolution stage and the rapid transition stage in the combustion process, avoiding misjudging transient numerical noise as a real state transition.

[0113] Based on the identified state evolution segment points, the continuous time step feature sequence is divided into several evolutionary stage subsequences. Each evolutionary stage subsequence corresponds to a process segment of the combustion system with a relatively consistent evolutionary trend over a continuous time period. For the first... Let there be a sequence of evolutionary stages, with a time step index range of . The set of state transition rates for each adjacent time step within a subsequence is The stage evolution rate corresponding to this subsequence. The value is taken as the weighted average of all state transition rates within the subsequence. The weights are allocated based on the time distance between each time step and the center time of the subsequence; time steps closer to the center time are assigned higher weights to highlight the contribution of the core evolutionary trend of the subsequence. Stage evolution rate. It comprehensively reflects the overall rhythm of combustion state changes during this evolutionary stage and can effectively suppress the interference of local noise on trend judgment.

[0114] After obtaining the stage evolution rate of each evolution stage subsequence, the time span between each time step and the combustion state transition time step is combined. Calculate the state transfer weights between consecutive time steps. State transit weights are used to measure the weights between any two time steps in the sequence. and The strength of state information transfer between time steps is calculated by comprehensively considering the difference in the stage evolution rate of the two time steps and the ratio of their time spans to the transition time step. Specifically, if time step and If they belong to the same evolutionary stage, then their state transfer weights are... A higher correlation indicates a strong correlation between state information within the same stage; if the two belong to different evolutionary stages, then... It decreases as the difference in evolutionary rate increases, reflecting the attenuation effect of state information transfer between different evolutionary stages. Furthermore, the time span... The smaller the value, the closer the time step is to the combustion state inflection point, and the more significant its impact on subsequent state evolution. Therefore, when calculating the state transfer weights, additional weight gains are given to time steps with smaller time spans. Let the state transfer weight matrix be... , its first The element is The entire matrix fully describes the distribution of state transfer intensity between any two steps in a continuous time step feature sequence.

[0115] Combine the continuous time step feature sequence with the state transfer weight matrix By combining these elements, a complete state transfer relationship between consecutive time steps can be constructed. In practical applications, for the nth time step in a sequence... Weighted feature scalar values ​​at each time step The equivalent transitive representation of the state transitive relationship is obtained by weighting and aggregating the eigenvalues ​​of all other time steps according to their corresponding transitive weights, forming the enhanced state transitive representation for that time step. This enhanced representation retains the characteristic information of the current time step itself while incorporating the state transitive contributions from other time steps, thus enabling subsequent state evolution predictions to fully utilize the intrinsic correlations between different stages in the combustion process. The construction process of the state transitive relationship fully considers the nonlinear dynamic characteristics of the combustion system. Through piecewise rate analysis and inflection point identification, the complex evolution law of the combustion process is decomposed into several evolutionary stages with clear physical meanings, and expressed in the form of quantified transitive weights, providing a structured and interpretable input representation for the generation of subsequent state evolution prediction information.

[0116] Based on state evolution prediction information, the trajectory of changes in combustion efficiency and emission concentrations at multiple future time steps is calculated, including:

[0117] Extract the current combustion state and direction of state change from the state evolution prediction information, and obtain the baseline combustion efficiency and baseline emission concentration based on the current combustion state;

[0118] The weighted feature is calculated to represent the feature change between consecutive time steps. The evolution rate is calculated based on the correlation between the feature change and combustion efficiency and emission concentration. The direction of state change is converted into an evolution direction coefficient. The evolution rate is multiplied by the evolution direction coefficient to obtain the directional evolution rate.

[0119] The combustion efficiency and emission concentration for the future time steps are calculated by multiplying the directed evolution rate by the time step size of the future time steps and then superimposing the baseline combustion efficiency and baseline emission concentration.

[0120] The system identifies points where the difference between adjacent time steps in combustion efficiency and emission concentration exceeds a preset fluctuation threshold as segmentation points. Smoothing is then performed within the intervals defined by these segmentation points to generate the change trajectories of combustion efficiency and emission concentration for multiple future time steps.

[0121] Extracting the current combustion state and direction of change from the state evolution prediction information is the starting point for calculating the trajectory. The current combustion state encompasses the comprehensive operational characteristics of the combustion system at the current moment, reflected in the combustion efficiency level and emission concentration level mapped by the feature vector corresponding to the current moment in the weighted feature representation. Based on the current combustion state, the baseline combustion efficiency is obtained by querying the historical operating condition database or the current sensor sampling values. and baseline emission concentration The role of the baseline value is to provide a stable reference starting point for subsequent trajectory prediction, avoiding deviations of the predicted trajectory from the dimensional range of the actual physical quantities due to numerical drift of the feature vector in high-dimensional space. The direction of state change comes from the judgment result of the future combustion state trend in the state evolution prediction information, representing whether the combustion efficiency and emission concentration are in an increasing, decreasing, or stable trend category, respectively.

[0122] When calculating the weighted feature representation of the feature changes between consecutive time steps, the difference operation is performed on the weighted feature vectors of adjacent time steps to obtain the feature change vector between each pair of adjacent time steps. To establish a correlation between the high-dimensional feature changes and the two physical quantities of combustion efficiency and emission concentration, it is necessary to calculate the correlation degree between the feature changes and combustion efficiency and emission concentration separately. The correlation degree is calculated by extracting the subset of components related to combustion efficiency from the feature change vector and calculating its magnitude as the basis for the combustion efficiency evolution rate. Similarly, a subset of components related to emission concentration is extracted, and their modulus is calculated as a fundamental quantity for the rate of emission concentration evolution. The partitioning of quantum sets is based on the physical quantity annotation information corresponding to each dimension of the feature vector during the training phase, and is determined through the mapping relationship between feature channels and physical quantities.

[0123] Evolution direction coefficient and These correspond to the directions of state change for combustion efficiency and emission concentration, respectively. When the direction of state change is upward, the evolution direction coefficient is taken as... When the direction of state change is decreasing, the evolution direction coefficient is taken as... When the direction of state change is stationary, the evolution direction coefficient is taken as... Multiplying the fundamental evolution rate by the corresponding evolution direction coefficient yields the directional evolution rate. and ,Right now , The directional evolution rate carries information in both the magnitude and direction of change, and can be directly used for subsequent trajectory unfolding calculations.

[0124] After obtaining the directed evolution rate, trajectory prediction is performed sequentially for multiple future time steps. Let the index of the future time step be . ( ,in (To predict the total number of time steps in the time domain), the first The cumulative time step between each future time step and the current time is: . Future No. Predicted combustion efficiency at each time step and predicted emission concentrations They are respectively: , The above calculations, starting from a baseline value, are linearly extrapolated along the direction of the directional evolution rate to obtain a preliminary sequence of change trajectories. This linear extrapolation assumes that the evolution rate remains relatively stable within the prediction time domain, making it suitable for short-term prediction scenarios. For cases with longer prediction time domains, the basic evolution rate can be updated at each time step to reflect the nonlinear dynamic characteristics of the combustion process.

[0125] Preliminary combustion efficiency trajectory sequence and emission concentration trajectory sequence In the process, there may be locations where the difference between adjacent time steps is abnormally large. These locations correspond to abrupt changes in combustion conditions. If left unaddressed, these points can lead to unreasonable control actions by the subsequent reinforcement learning strategy under trajectory constraints. The method for identifying these segmentation points is as follows: for the combustion efficiency trajectory, calculate the difference between adjacent time steps. If the difference exceeds the preset fluctuation threshold Then the time step and The positions between these points are marked as combustion efficiency segment points; for the emission concentration trajectory, the time step difference between adjacent times is calculated. If the difference exceeds the preset fluctuation threshold The corresponding location will be marked as an emission concentration segment point. Preset fluctuation threshold. and The threshold is determined by statistical analysis of the normal fluctuation range of combustion efficiency and emission concentration in historical operating data. Typically, the 95th percentile of the distribution of adjacent differences in historical sequences is taken as the upper limit of the threshold.

[0126] Within each interval defined by the segmentation points, the trajectory sequence is smoothed to eliminate numerical discontinuities caused by abrupt changes. The smoothing process combines piecewise linear interpolation with local weighted averaging: several predicted values ​​from each time step are taken as anchor points on both sides of the segmentation point, and linear interpolation is used to regenerate the trajectory values ​​for the transition segment between these anchor points, ensuring the first-order difference of the trajectory is continuous at the segmentation points. For the trajectory values ​​in intervals outside the segmentation points, a sliding window weighted average is used for smoothing. The weight of each time step within the window decreases as the distance from the center time step increases, thereby suppressing the interference of local noise on the trajectory. The smoothed combustion efficiency change trajectory is shown below. And the trajectory of changes in emission concentrations Maintain physically reasonable continuity and monotonicity constraints throughout the entire prediction time domain.

[0127] The generated change trajectories serve as constraint boundaries for subsequent reinforcement learning policy search. The combustion efficiency trajectory requires that the actual combustion efficiency achieved by the control decision at each future time step is not lower than the corresponding predicted value, while the emission concentration trajectory requires that the actual emission concentration achieved by the control decision at each future time step does not exceed the smaller of the corresponding predicted value and the emission limit. By embedding the change trajectories into the reinforcement learning reward function and constraints, the policy search can strictly adhere to emission concentration compliance requirements while satisfying the combustion efficiency improvement target, thereby achieving fine-grained control of the combustion process.

[0128] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the synergistic optimization control operation of combustion efficiency and emission concentration in this embodiment.

[0129] Under the constraint of the changing trajectory, the control decision commands are generated by searching for oxygen supply, fuel supply rate, and damper opening to maximize combustion efficiency and meet emission concentration limits through reinforcement learning strategies.

[0130] The combustion efficiency constraint value and emission concentration constraint value for each time step are extracted from the trajectory of combustion efficiency change and emission concentration change over multiple future time steps.

[0131] Construct a coupling relationship surface between oxygen supply regulation and combustion efficiency constraints and emission concentration constraints. Identify the critical emission concentration value corresponding to the maximum combustion efficiency on the coupling relationship surface. Calculate the concentration gap between the critical emission concentration value and the emission concentration constraints. Based on the concentration gap, solve the compensation combination of fuel supply rate regulation and damper opening regulation to maximize combustion efficiency and meet emission concentration limits.

[0132] The oxygen supply rate adjustment, fuel supply rate adjustment, and damper opening adjustment obtained from the inverse solution are used as the initial search center point. A local search is performed in the neighborhood of the initial search center point using a reinforcement learning strategy to identify the oxygen supply rate adjustment, fuel supply rate adjustment, and damper opening adjustment that maximize combustion efficiency and meet emission concentration limits.

[0133] The identified oxygen supply adjustment, fuel supply rate adjustment, and damper opening adjustment are encapsulated into control decision commands.

[0134] The trajectory of combustion efficiency changes over multiple future time steps And the trajectory of changes in emission concentrations In the process, each future time step is extracted sequentially. ( The corresponding combustion efficiency constraint value and emission concentration limits These two sets of constraint values ​​reflect the expected levels of combustion efficiency and emission concentration at future time points under the current combustion state evolution trend. During the extraction process, the validity of the constraint values ​​at each time step is verified. If a constraint value at a certain time step exceeds the physically reasonable range (e.g., combustion efficiency exceeds 100% or emission concentration is negative), it is corrected by interpolation using the constraint values ​​of adjacent time steps to ensure the physical feasibility of the constraint value sequence. After extraction, the constraint value sequence is arranged in time step order to form a constraint input set for subsequent coupling relationship modeling.

[0135] After obtaining the combustion efficiency constraints and emission concentration constraints at each time step, a coupling surface is constructed between the oxygen supply adjustment and the combustion efficiency and emission concentration constraints. Specifically, let the oxygen supply adjustment be... Combustion efficiency constraint value Emission concentration limit value The coupling relationship surface describes the joint response characteristics of combustion efficiency and emission concentration under different oxygen supply adjustment values. This surface is obtained by multidimensional regression fitting of oxygen supply, combustion efficiency, and emission concentration from historical operating data, and a second-order polynomial surface model is used to approximate the nonlinear relationship among the three. On the coupling relationship surface, the maximum point is searched along the combustion efficiency axis to identify the oxygen supply adjustment value corresponding to the maximum combustion efficiency. and the critical value of emission concentration under this condition. Critical values ​​for emission concentrations This represents the natural response level of emission concentration when the oxygen supply is adjusted to the optimal level under the goal of maximizing combustion efficiency.

[0136] Calculate the critical value of emission concentration Compared with the current time step emission concentration constraint value Concentration gap between ,Right now .when When this occurs, it indicates that under the oxygen supply setting that maximizes combustion efficiency, the emission concentration exceeds the limit constrained by the trajectory, requiring compensation by adjusting the fuel supply rate and damper opening; when This indicates that the current oxygen supply setting can simultaneously meet the requirements of maximizing combustion efficiency and emission concentration limits, and no additional compensation is needed. Regarding... In this case, based on the size of the concentration gap, the solution is reversed to reduce the emission concentration. Required fuel supply rate adjustment And damper opening adjustment amount The compensation combination is as follows. In the reverse solution process, a partial response model of emission concentration to fuel supply rate and damper opening is established to estimate the marginal impact coefficients of unit fuel supply rate adjustment and unit damper opening adjustment on emission concentration, respectively. and By solving the system of linear equations To obtain the concentration gap compensation requirement Combination. When multiple solutions exist, the optimal compensation combination is selected based on minimizing combustion efficiency loss, and used as the initial estimate for the fuel supply rate adjustment and damper opening adjustment.

[0137] The oxygen supply adjustment amount obtained by inverse solution Fuel supply rate adjustment And damper opening adjustment amount The vector is combined into a three-dimensional control action vector. Using this vector as the initial search center point, a local search is performed within its neighborhood using a reinforcement learning strategy. The neighborhood range is determined based on the physical constraints of each regulation and the response capability of the actuator. The search radius for the oxygen supply regulation is set to... The preset proportional range, fuel supply rate adjustment amount, and damper opening adjustment amount are similarly set in the search radius. The reinforcement learning strategy adopts a proximal policy optimization method based on policy gradient, using maximizing combustion efficiency and meeting emission concentration limits as dual-objective reward signals. The reward function is designed as follows: when a candidate control action improves combustion efficiency and the emission concentration does not exceed the constraint limit, a positive reward is given; when the emission concentration exceeds the limit, a penalty term is given, and the penalty intensity is proportional to the amount of exceedance. During the local search process, the policy network takes the current state evolution prediction information and the initial search center point as input, and outputs the adjustment direction and step size of the control action. Through multiple iterations, it continuously approaches the optimal combination of control actions that maximizes combustion efficiency and meets emission concentration limits.

[0138] The termination conditions for local search iterations fall into two categories: one is reaching the preset maximum number of iterations. Secondly, the improvement in combustion efficiency corresponding to the optimal control action in two adjacent iterations is lower than the convergence threshold. This means the search is considered to have converged to a local optimum. During the iteration process, the feasibility of each candidate control action is verified to ensure that the adjustments to oxygen supply rate, fuel supply rate, and damper opening are all within the physical feasible range of the actuator. If a candidate action exceeds the feasible range, it is truncated to the boundary value before reward evaluation to avoid the policy network outputting unexecutable control commands.

[0139] After the search converges, the control action combination with the highest cumulative reward from all iterations is selected as the final identified optimal oxygen delivery adjustment. Fuel supply rate adjustment And damper opening adjustment amount The three components (oxygen supply, fuel supply pump, and damper actuator) are formatted and encapsulated according to the actuator interface specification to generate control decision instructions containing adjustment values, adjustment directions, execution sequences, and priority identifiers. These instructions also include the corresponding expected combustion efficiency improvement and emission concentration change, allowing the actuators to verify the effect after execution. The encapsulated control decision instructions are then sent to actuators such as the oxygen supply control valve, fuel supply pump, and damper actuator. These actuators then adjust the combustion process in real time according to the instructions, achieving a synergistic optimization goal of maximizing combustion efficiency and ensuring emission concentration compliance.

[0140] A second aspect of this invention provides a deep learning-based intelligent decision-making system for coal-fired power plant combustion, comprising:

[0141] The data acquisition unit is used to collect fuel characteristic parameters, airflow dynamic parameters, thermodynamic state parameters and emission concentration parameters of the combustion system to form a sequence of operating condition data including timestamps.

[0142] Multi-scale feature units are used to process operating condition data sequences through convolution operations at different scales, extract combustion features at different time spans, and generate multi-scale feature representations.

[0143] An adaptive additive unit is used to calculate the correlation strength between the features of each time step in the multi-scale feature representation and the stability of the current combustion condition. Based on the correlation strength, an adaptive weight is assigned to each time step feature to generate a weighted feature representation.

[0144] The state prediction unit is used to perform time-series processing on the weighted feature representation, identify the current combustion state through the state transfer relationship between consecutive time steps, predict the direction of state change at future moments, and generate state evolution prediction information.

[0145] The trajectory calculation unit is used to calculate the trajectory of changes in combustion efficiency and emission concentration at multiple future time steps based on state evolution prediction information.

[0146] The control decision unit is used to search for oxygen supply, fuel supply rate, and damper opening adjustments to maximize combustion efficiency and meet emission concentration limits under the constraints of the changing trajectory through reinforcement learning strategies. It generates control decision commands and sends them to the actuators to regulate the combustion process.

[0147] A third aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0152] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based intelligent decision-making method for coal-fired power plant combustion, characterized in that, include: Collect fuel characteristic parameters, airflow dynamic parameters, thermodynamic state parameters and emission concentration parameters of the combustion system to form a sequence of operating condition data including timestamps; By processing the working condition data sequence through convolution operations at different scales, combustion features at different time spans are extracted to generate multi-scale feature representations. Calculate the correlation strength between the features of each time step in the multi-scale feature representation and the stability of the current combustion condition, and assign adaptive weights to the features of each time step according to the correlation strength to generate a weighted feature representation; The weighted feature representation is processed temporally, and the current combustion state is identified and the direction of state change in future moments is predicted through the state transfer relationship between consecutive time steps, generating state evolution prediction information. The trajectory of changes in combustion efficiency and emission concentration at multiple future time steps is calculated based on state evolution prediction information. Under the constraint of the changing trajectory, the oxygen supply, fuel supply rate and damper opening are adjusted to maximize combustion efficiency and meet emission concentration limits through reinforcement learning strategy search. Control decision commands are generated and sent to the actuators to regulate the combustion process.

2. The method according to claim 1, characterized in that, By processing the operating condition data sequence through convolution operations at different scales, combustion features at different time spans are extracted to generate multi-scale feature representations, including: The operating condition data sequence is divided into multiple data windows according to the time dimension. The fluctuation variance of combustion efficiency in each data window is calculated. The receptive field range of the convolution kernel in the time dimension is determined according to the fluctuation variance. Convolution operation is performed on each data window along the time dimension to generate transient combustion features and evolutionary combustion features. The error contribution values ​​of the transient combustion characteristics and the evolutionary combustion characteristics to the combustion efficiency prediction are calculated respectively, and the complementarity metric is calculated based on the difference between the error contribution values. The transient fusion weight is calculated based on the ratio of the complementarity metric to the error contribution value of the transient combustion feature, and the evolutionary fusion weight is calculated based on the ratio of the complementarity metric to the error contribution value of the evolutionary combustion feature. The transient combustion feature is multiplied by the transient fusion weight to obtain the weighted transient combustion feature, the evolved combustion feature is multiplied by the evolved fusion weight to obtain the weighted evolved combustion feature, and the weighted transient combustion feature and the weighted evolved combustion feature are concatenated to obtain the fused combustion feature. The fusion combustion features are subjected to a nonlinear transformation to generate a multi-scale feature representation.

3. The method according to claim 1, characterized in that, The correlation strength between the features at each time step in the multi-scale feature representation and the stability of the current combustion condition is calculated. Adaptive weights are assigned to the features at each time step based on the correlation strength, generating a weighted feature representation including: Calculate the time interval between the time corresponding to each time step feature in the multi-scale feature representation and the current time, and calculate the time decay factor based on the time interval; Calculate the rate of change of combustion efficiency and the rate of change of emission concentration at each time step corresponding to the characteristic, and calculate the intensity of operating condition fluctuation based on the rate of change of combustion efficiency and the rate of change of emission concentration; Identify time steps in the operating condition fluctuation intensity that exceed a preset fluctuation threshold, calculate the time interval distribution characteristics between the time steps that exceed the preset fluctuation threshold and the current time, and calculate the fluctuation propagation delay coefficient based on the time interval distribution characteristics; The correlation strength between the characteristics of each time step and the stability of the current combustion condition is calculated based on the time decay factor of each time step characteristic, the corresponding operating condition fluctuation intensity, and the fluctuation propagation delay coefficient. The correlation strength is normalized to obtain the adaptive weights of the features at each time step. The features at each time step in the multi-scale feature representation are weighted according to the adaptive weights and then combined in chronological order to generate a weighted feature representation.

4. The method according to claim 1, characterized in that, The weighted feature representation is processed temporally to identify the current combustion state and predict the direction of future state changes through the state transit relationship between consecutive time steps, generating state evolution prediction information including: The weighted feature representation is processed temporally, the feature change gradient of the weighted feature representation between consecutive time steps is calculated, and the time step in which the absolute value of the feature change gradient exceeds the preset gradient threshold is identified as the combustion state transition time step. Obtain the weighted feature representation corresponding to the combustion state transition time step, extract the combustion efficiency value and emission concentration value from the weighted feature representation corresponding to the combustion state transition time step, and identify the current combustion state according to the preset distribution interval to which the combustion efficiency value and emission concentration value belong; Weighted features are extracted to represent the continuous time step feature sequence from the combustion state transition time step to the current time step. The state transfer relationship between continuous time steps is constructed through the continuous time step feature sequence. The state evolution trend function is obtained by fitting the state transfer relationship between continuous time steps. The future time is used as a time parameter to input the state evolution trend function to calculate the predicted feature value of the future time. The difference between the predicted feature value of the future time and the weighted feature representation of the current time is calculated. The direction of state change in the future time is predicted based on the positive or negative sign of the difference. The current combustion state and the direction of future state changes are combined to generate state evolution prediction information.

5. The method according to claim 4, characterized in that, Constructing the state transfer relationship between consecutive time steps using feature sequences of consecutive time steps includes: Calculate the feature difference between adjacent time step features in a continuous time step feature sequence, and calculate the state transition rate between adjacent time steps based on the feature difference; Calculate the time span between the time corresponding to each time step feature in the continuous time step feature sequence and the time corresponding to the combustion state transition time step, and identify the time step in which the state transition rate fluctuates and reverses within the time span as the state evolution segment point. The continuous time step feature sequence is divided into multiple evolution stage subsequences according to the state evolution segmentation point. The stage evolution rate corresponding to each evolution stage subsequence is calculated. The state transfer weight between each time step is calculated according to the time span and the stage evolution rate. The continuous time step feature sequence and the state transfer weight are combined to construct the state transfer relationship between continuous time steps.

6. The method according to claim 1, characterized in that, Based on state evolution prediction information, the trajectory of changes in combustion efficiency and emission concentrations at multiple future time steps is calculated, including: Extract the current combustion state and direction of state change from the state evolution prediction information, and obtain the baseline combustion efficiency and baseline emission concentration based on the current combustion state; The weighted feature is calculated to represent the feature change between consecutive time steps. The evolution rate is calculated based on the correlation between the feature change and combustion efficiency and emission concentration. The direction of state change is converted into an evolution direction coefficient. The evolution rate is multiplied by the evolution direction coefficient to obtain the directional evolution rate. The combustion efficiency and emission concentration for the future time steps are calculated by multiplying the directed evolution rate by the time step size of the future time steps and then superimposing the baseline combustion efficiency and baseline emission concentration. The system identifies points where the difference between adjacent time steps in combustion efficiency and emission concentration exceeds a preset fluctuation threshold as segmentation points. Smoothing is then performed within the intervals defined by these segmentation points to generate the change trajectories of combustion efficiency and emission concentration for multiple future time steps.

7. The method according to claim 1, characterized in that, Under the constraint of the changing trajectory, the control decision commands are generated by searching for oxygen supply, fuel supply rate, and damper opening to maximize combustion efficiency and meet emission concentration limits through reinforcement learning strategies. The combustion efficiency constraint value and emission concentration constraint value for each time step are extracted from the trajectory of combustion efficiency change and emission concentration change over multiple future time steps. Construct a coupling relationship surface between oxygen supply regulation and combustion efficiency constraints and emission concentration constraints. Identify the critical emission concentration value corresponding to the maximum combustion efficiency on the coupling relationship surface. Calculate the concentration gap between the critical emission concentration value and the emission concentration constraints. Based on the concentration gap, solve the compensation combination of fuel supply rate regulation and damper opening regulation to maximize combustion efficiency and meet emission concentration limits. The oxygen supply rate adjustment, fuel supply rate adjustment, and damper opening adjustment obtained from the inverse solution are used as the initial search center point. A local search is performed in the neighborhood of the initial search center point using a reinforcement learning strategy to identify the oxygen supply rate adjustment, fuel supply rate adjustment, and damper opening adjustment that maximize combustion efficiency and meet emission concentration limits. The identified oxygen supply adjustment, fuel supply rate adjustment, and damper opening adjustment are encapsulated into control decision commands.

8. A deep learning-based intelligent decision-making system for coal-fired power plant combustion, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to collect fuel characteristic parameters, airflow dynamic parameters, thermodynamic state parameters and emission concentration parameters of the combustion system to form a sequence of operating condition data including timestamps. Multi-scale feature units are used to process operating condition data sequences through convolution operations at different scales, extract combustion features at different time spans, and generate multi-scale feature representations. An adaptive additive unit is used to calculate the correlation strength between the features of each time step in the multi-scale feature representation and the stability of the current combustion condition. Based on the correlation strength, an adaptive weight is assigned to each time step feature to generate a weighted feature representation. The state prediction unit is used to perform time-series processing on the weighted feature representation, identify the current combustion state through the state transfer relationship between consecutive time steps, predict the direction of state change at future moments, and generate state evolution prediction information. The trajectory calculation unit is used to calculate the trajectory of changes in combustion efficiency and emission concentration at multiple future time steps based on state evolution prediction information. The control decision unit is used to search for oxygen supply, fuel supply rate, and damper opening adjustments to maximize combustion efficiency and meet emission concentration limits under the constraints of the changing trajectory through reinforcement learning strategies. It generates control decision commands and sends them to the actuators to regulate the combustion process.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images