Boiler combustion control system and method based on hierarchical collaborative optimization and online correction

Abstract

The invention discloses a boiler combustion control system and method based on hierarchical collaborative optimization and online correction. Offline hierarchical collaborative optimization comprises the following steps: S1, data preparation and working condition clustering; s2, training a hierarchical expert predictive control model; s3, carrying out model integration; and the online self-adaptive correction comprises the following content: S4, real-time circulation control. According to the method, a complex boiler combustion control problem is decomposed into a plurality of local optimization sub-problems through offline layered collaborative optimization, and the limitation that a traditional single model is prone to falling into local optimum and the full-working-condition generalization ability is insufficient is effectively avoided. Furthermore, through an on-line self-adaptive correction mechanism, the system can continuously learn a residual error between an expert prediction control model and an actual optimal value, and performs real-time fine adjustment and compensation on a control signal, thereby effectively coping with dynamic changes such as boiler operation load, fuel composition or equipment aging and the like. And efficient and stable control of the boiler under full-working-condition and long-period operation is ensured.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a boiler combustion control system and method based on hierarchical collaborative optimization and online correction. Background Technology

[0002] The combustion process of industrial gas-fired boilers exhibits significant characteristics of large inertia, large time delay, and strong coupling of multiple variables, accompanied by strong nonlinearity and time-varying uncertainty. Traditional proportional-integral-derivative (PID) controllers, with their fixed parameter structure, struggle to adapt to fluctuations in boiler operating load, changes in fuel composition, and dynamic shifts in environmental conditions, leading to reduced energy efficiency and excessive pollutant emissions. Existing intelligent control schemes attempt to incorporate fuzzy logic, neural networks, or swarm optimization algorithms, but these have significant limitations: fuzzy logic relies excessively on manual empirical rules, neural network training is time-consuming and cannot meet real-time control requirements, and swarm optimization algorithms, when dealing with boiler efficiency surfaces, are prone to getting trapped in suboptimal solutions under specific operating conditions due to the presence of multiple local extrema that dynamically drift with operating conditions, resulting in insufficient generalization ability of the model across the entire operating range. Current technical solutions generally lack systematic modeling of the boiler's inherent physical characteristics, failing to effectively address the fundamental differences in optimal control strategies across different load ranges, and lacking an adaptive mechanism for the slow drift of model parameters caused by equipment aging. This technical defect causes the control model to gradually deviate from the optimal state during long-term operation, making it impossible to simultaneously achieve the dual goals of energy efficiency improvement and emission control, thus restricting the efficient and stable operation of the boiler system throughout its entire life cycle. Summary of the Invention

[0003] In view of this, the present invention provides a boiler combustion control method based on hierarchical collaborative optimization and online correction, which has the advantages of effectively adapting to load fluctuations and operating condition changes during boiler operation, improving control accuracy and stability, thereby improving energy utilization efficiency and reducing pollutant emissions.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A boiler combustion control method based on hierarchical collaborative optimization and online correction includes two stages: offline hierarchical collaborative optimization and online adaptive correction. Offline hierarchical collaborative optimization includes the following: S1, data preparation and operating condition clustering: Based on historical operating data, an unsupervised learning algorithm is used to automatically divide the continuous operating conditions of the boiler into multiple discrete typical operating condition intervals with significantly different characteristics; S2, hierarchical expert predictive control model training: For each operating condition interval, a hybrid leapfrog-simulated annealing algorithm is used independently and in parallel to optimize the hyperparameters of the broadband learning system, thereby training an expert predictive control model for each operating condition interval; S3, model integration: The expert predictive control models of all operating condition intervals are integrated with an operating condition identifier to form a unified hierarchical collaborative optimization control model. Online adaptive correction includes the following: S4, real-time cyclic control: The control system runs an online sequential learning module in parallel. The online sequential learning module continuously learns the residual between the predicted value and the actual optimal value of the current expert predictive control model. Based on this residual, the online sequential learning module fine-tunes and compensates the basic control signal to achieve online adaptive correction of the expert predictive control model.

[0006] Preferably, the data preparation and operating condition clustering specifically include the following: S11, collecting historical data of the boiler within at least one complete operating cycle; S12, using the K-Means clustering algorithm, with steam pressure and load change rate as the main features, automatically dividing the dataset into K operating condition clusters.

[0007] Preferably, the broadband learning system is defined as follows: an independent broadband learning system model is defined for each operating condition cluster.

[0008] Preferably, the hybrid frog-leap-simulated annealing algorithm specifically includes the following: Optimization objective: For each case cluster, the goal of the hybrid frog-leap-simulated annealing algorithm is to find a set of optimal broadband learning system hyperparameters to minimize a composite fitness function; Collaborative optimization mechanism: The hybrid frog-leap-simulated annealing algorithm divides the frog swarm into multiple meme groups, each group performs a depth search in its own local space, and periodically exchanges information between groups to achieve global information sharing; On this basis, a simulated annealing operator is introduced to perturb the optimal frog in each meme group, using its probabilistic leaping ability to help the algorithm effectively escape the local optimum within the case interval.

[0009] Preferably, the mathematical expression for the composite fitness function is: ;in, The mean square error of the predictions made by the expert predictive control model. The nitrogen oxide emission predictions are from the nitrogen oxide surrogate model. The integral of the rate of change of the control signal. , , All are weighting factors.

[0010] Preferably, the nitrogen oxide surrogate model is a multilayer perceptron neural network, and its construction steps are as follows: T1, extract data from the historical dataset... Generate closely related feature parameters as input to the nitrogen oxide surrogate model; T2, use the actual measured... The concentration value is used as the output label of the nitrogen oxide surrogate model; T3, the multilayer perceptron neural network is trained using historical datasets until its prediction error converges.

[0011] Preferably, the real-time cyclic control specifically includes the following: S41, the control system receives real-time operating parameters of the boiler; S42, the current operating condition range is determined by the operating condition identifier and the corresponding expert predictive control model is activated; S43, the activated expert predictive control model calculates and generates a basic optimal control signal, which includes the blower frequency setpoint and the gas valve opening setpoint; S44, the residual between the basic optimal control signal and the actual optimal control signal is continuously learned by the online sequential learning module to generate a compensation signal; S45, the compensation signal is superimposed with the basic optimal control signal to form the final control signal and sent to the actuator.

[0012] Preferably, the online adaptive correction also includes the following: S5, Online learning and correction: The control system calculates the actual operating thermal efficiency of the boiler over a period of time, i.e. the actual optimal control value. At the same time, the control system uses a performance curve model with a higher-order load as input and theoretical maximum efficiency as output to deduce the residual between the actual optimal control value and the model output value. This residual is used as a new training sample to update the online sequential learning module.

[0013] Preferably, the calculation of the actual operating thermal efficiency of the boiler specifically includes the following: the control system calculates the actual operating thermal efficiency of the boiler over a past period using a preset time scale through the boiler thermal efficiency evaluation module; the boiler thermal efficiency evaluation module calculates the actual flue gas heat loss and heat dissipation loss of the boiler based on the heat loss method for calculating the thermal efficiency of industrial boilers, so as to obtain the actual operating thermal efficiency of the boiler.

[0014] This invention also proposes a boiler combustion control system based on hierarchical collaborative optimization and online correction, comprising: a control module, a sensor module, and an execution module; both the sensor module and the execution module are signal-connected to the control module; the control module is configured to execute any of the boiler combustion control methods based on hierarchical collaborative optimization and online correction in the above embodiments; the control module integrates an operating condition identifier, a hierarchical collaborative optimization control model, and an online correction module; the operating condition identifier determines the current operating condition range based on the real-time operating data of the boiler; the hierarchical collaborative optimization control model includes multiple expert predictive control models based on a broadband learning system, which are obtained in the offline hierarchical collaborative optimization stage through a hybrid leapfrog-simulated annealing algorithm; the online correction module is a residual compensator based on an online sequential learning module, used to perform online correction on the basic optimal control signal output by the expert predictive control model; the execution module includes a blower frequency converter and a gas valve execution servo motor, the blower frequency converter adjusts the blower frequency according to the final control signal, and the gas valve execution servo motor adjusts the gas valve opening according to the final control signal.

[0015] Preferably, the boiler combustion control system based on hierarchical collaborative optimization and online correction further includes a monitoring module, which is signal-connected to the control module. The monitoring module includes a cloud server and a user interface. The control module packages the collected and calculated key data according to a predefined format and transmits it to the cloud server via 4G, 5G, or a wired network through the MQTT lightweight IoT protocol. The application logic on the cloud server parses the data and persistently stores it in a time-series database. Users can log in and access this data through a client application on a network-connected device to achieve visualized monitoring and management of the boiler status.

[0016] The beneficial effects of this invention are as follows: Compared with the prior art, this application decomposes the complex boiler combustion control problem into multiple local optimization sub-problems through offline hierarchical collaborative optimization, effectively avoiding the limitations of traditional single models that are prone to getting trapped in local optima and have insufficient generalization ability under all operating conditions. Furthermore, through an online adaptive correction mechanism, the system can continuously learn the residual between the expert predictive control model and the actual optimal value, and perform real-time fine-tuning and compensation of the control signal, thereby effectively coping with dynamic changes such as boiler operating load, fuel composition, or equipment aging, and ensuring efficient and stable control of the boiler under all operating conditions and long-term operation.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture of the control system of the present invention;

[0019] Figure 2 This is a flowchart of the offline hierarchical collaborative optimization and online adaptive correction control algorithm of the present invention;

[0020] Figure 3 This is a system architecture diagram of the monitoring module of the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] The following is for reference. Figures 1 to 3 This invention describes a boiler combustion control method based on hierarchical collaborative optimization and online correction in an embodiment of the invention.

[0024] This application discloses a boiler combustion control method based on hierarchical collaborative optimization and online correction, which includes two stages: offline hierarchical collaborative optimization and online adaptive correction.

[0025] Offline hierarchical collaborative optimization includes the following:

[0026] S1. Data Preparation and Operating Condition Clustering: Based on historical operating data, an unsupervised learning algorithm is used to automatically divide the continuous operating conditions of the boiler into multiple discrete typical operating condition intervals with significantly different characteristics.

[0027] S2. Training of hierarchical expert predictive control model: For each operating condition interval, the hybrid leapfrog-simulated annealing algorithm is used independently and in parallel to optimize the hyperparameters of the broadband learning system, thereby training an expert predictive control model for each operating condition interval.

[0028] S3. Model Integration: Integrate the expert predictive control models for all operating conditions with a single operating condition identifier to form a unified hierarchical collaborative optimization control model.

[0029] Online adaptive correction includes the following:

[0030] S4. Real-time cyclic control: The control system runs an online sequential learning module in parallel. The online sequential learning module continuously learns the residual between the predicted value and the actual optimal value of the current expert predictive control model. Based on the residual, the online sequential learning module fine-tunes and compensates the basic control signal to achieve online adaptive correction of the expert predictive control model.

[0031] Specifically, in the data preparation and operating condition clustering stage, historical operating data is collected, such as obtaining long-term series operating parameters of the boiler, including steam flow, fuel consumption, flue gas temperature, and oxygen content. This data is stored and preprocessed, for example, removing outliers and imputing missing data. Subsequently, unsupervised learning algorithms, such as hierarchical clustering, DBSCAN algorithm, or Gaussian mixture model, are used to process this data, automatically dividing the boiler's continuous operating conditions into multiple discrete typical operating condition intervals with significantly different characteristics. These intervals can be divided based on load level or combustion efficiency characteristics.

[0032] Furthermore, in the training phase of the hierarchical expert predictive control model, an expert predictive control model is trained independently for each defined operating condition interval. This model aims to learn the dynamic characteristics of the boiler under that specific operating condition and predict future operating states. As a model architecture, the hyperparameters of the broadband learning system can be optimized using various optimization algorithms, such as particle swarm optimization, genetic algorithms, or differential evolution algorithms. These algorithms iteratively search and adjust the internal parameters of the broadband learning system to achieve the expected predictive performance within the specified operating condition interval. The training process can be performed in parallel to improve overall efficiency.

[0033] Subsequently, in the model integration phase, all trained expert predictive control models are integrated with a condition identifier to form a unified hierarchical collaborative optimization control model. The condition identifier can be a rule-based judgment module, such as switching operating conditions based on real-time load thresholds, or a classification algorithm-based model, such as a support vector machine or decision tree, used to determine the current operating condition range of the boiler based on real-time sensor data. Once the condition identifier determines the current operating condition, the corresponding expert predictive control model is activated and responsible for generating control commands for that condition. This integration method enables the system to dynamically switch and apply the most suitable control strategy based on the actual operating state of the boiler.

[0034] In the real-time cyclic control during the online adaptive correction phase, the control system continuously receives data from sensors during the boiler's real-time operation, and the activated expert predictive control model generates basic control signals. Simultaneously, an online sequential learning module runs in parallel with the main control flow. This module can be a neural network based on incremental learning, such as an online extreme learning machine. Its task is to continuously compare the differences—the residuals—between the predicted values ​​output by the expert predictive control model and the optimal values ​​reflected by the boiler's actual operating data. Based on these residuals, the online sequential learning module calculates a compensation signal, which is then superimposed on the basic control signal to form the final control command. This mechanism enables the control model to adapt in real-time to minute changes in the boiler's operating environment or slow drifts in equipment performance.

[0035] This method decomposes the complex boiler combustion control problem into multiple local optimization sub-problems through offline hierarchical collaborative optimization, effectively avoiding the limitations of traditional single models that are prone to getting trapped in local optima and lack generalization ability under all operating conditions. Furthermore, through an online adaptive correction mechanism, the system can continuously learn the residuals between the expert-predicted control model and the actual optimal value, and perform real-time fine-tuning and compensation of the control signal. This effectively addresses dynamic changes such as boiler operating load, fuel composition, or equipment aging, ensuring efficient and stable control of the boiler under all operating conditions and long-term operation.

[0036] In some embodiments, data preparation and operational condition clustering specifically include the following:

[0037] S11. Collect historical data for at least one complete operating cycle of the boiler.

[0038] S12. Using the K-Means clustering algorithm, with steam pressure and load change rate as the main features, the dataset is automatically divided into K operating condition clusters.

[0039] Specifically, S11 aims to acquire comprehensive and representative boiler operating data. Data from a complete operating cycle can cover various operating conditions of the boiler, including startup, low-load stable combustion, medium-load efficient operation, high-load response, and shutdown, thereby ensuring the accuracy and generalization ability of subsequent operating condition clustering and model training. In implementation, the data acquisition and storage module built into the boiler control system can automatically record various operating parameters, including steam pressure, load command, fuel flow, air supply volume, flue gas temperature, and oxygen content, at preset time intervals (e.g., per second or per minute), and store them in a local database or cloud data platform. Alternatively, externally deployed sensor networks and data loggers can be used to perform high-frequency sampling and storage of key boiler operating parameters, and unified management through a data integration platform to ensure data integrity and traceability.

[0040] In S12, the aim is to discretize continuous boiler operation data into typical operating condition intervals with similar characteristics using unsupervised learning methods. The K-Means clustering algorithm is an efficient and widely used unsupervised learning algorithm. Its core idea is to iteratively calculate cluster centers and reallocate data points to the nearest clusters until the cluster assignments no longer change or a preset number of iterations is reached. When selecting clustering features, steam pressure directly reflects the boiler's load demand and energy output level, serving as a core indicator for distinguishing different operating load intervals; the load change rate characterizes the dynamic characteristics and response speed requirements of boiler operation, effectively distinguishing between steady-state operation and dynamic conditions such as rapid load changes. By using these key parameters as primary features, it can be ensured that the clustering results accurately reflect the actual operating state of the boiler. The K value can be evaluated using various methods, such as the Elbow Method or the Silhouette Coefficient, to select the number of clusters that best reflect the inherent structure of the data.

[0041] Through the above technical solutions, this application clarifies the specific implementation details of data preparation and operating condition clustering, effectively solving the problems of inaccurate or inefficient operating condition classification. By collecting historical data from at least one complete operating cycle of the boiler in S11, the comprehensiveness and representativeness of the data are ensured, avoiding clustering bias caused by insufficient or biased data, and providing a solid data foundation for subsequent operating condition classification. Furthermore, S12 employs the K-Means clustering algorithm, using steam pressure and load change rate as the main features, which can efficiently and accurately divide the continuous operating conditions of the boiler into K typical operating condition clusters with significantly different characteristics. The introduction of the K-Means algorithm provides a standardized clustering method, while steam pressure and load change rate, as core operating parameters, directly reflect the boiler's load state and dynamic characteristics, enabling the clustering results to accurately capture the essential differences between different operating conditions. This precise division of operating conditions provides a reliable foundation for the subsequent independent training of expert predictive control models for each operating condition range. It ensures that each expert model can perform in-depth optimization for the specific operating condition range it is responsible for, thereby significantly improving the overall performance and control accuracy of the hierarchical collaborative optimization control model, and ultimately achieving more efficient and stable operation of the boiler across the entire operating range.

[0042] In some embodiments, a broadband learning system is defined as follows: an independent broadband learning system model is defined for each cluster of operating conditions. The broadband learning system model is a novel neural network architecture characterized by its ability to train quickly and learn incrementally without requiring deep stacking. Specifically, the broadband learning system model typically includes feature mapping layers and enhancement layers, and its output layer weights can be directly calculated analytically, rather than through traditional iterative optimization. This characteristic makes the broadband learning system model perform well in handling large-scale data and scenarios requiring rapid response. Furthermore, the broadband learning system model supports online incremental learning; that is, when new data is received, the entire model does not need to be retrained, only some parameters need to be updated, providing a technical foundation for subsequent online adaptive correction.

[0043] Defining an independent broadband learning system model for each operating condition cluster means that after clustering the boiler operating conditions, a dedicated broadband learning system model is built and trained separately for each identified operating condition cluster. This means that if there are K operating condition clusters, K independent broadband learning system models will be trained, each model specifically responsible for its corresponding operating condition cluster. In one implementation, each independent broadband learning system model is trained and optimized using only the historical operating data of its corresponding operating condition cluster, ensuring that the model can deeply learn the unique operating patterns and control strategies of that operating condition cluster. In another implementation, each independent broadband learning system model can adopt different hyperparameter configurations or even fine-tune its internal structure according to the characteristics of its corresponding operating condition cluster to maximize performance under that specific operating condition cluster.

[0044] Through the above technical solution, this application effectively solves the problem that a single model is difficult to adapt to the complex characteristics of the boiler under all operating conditions by defining an independent broadband learning system model for each operating condition cluster. Specifically, in the offline hierarchical collaborative optimization stage, after the continuous operating conditions of the boiler are divided into multiple discrete operating condition clusters using the K-Means clustering algorithm, a broadband learning system model is independently trained for each operating condition cluster. This allows each model to focus on learning and mastering the unique operating rules and optimal control strategies of its corresponding operating condition cluster. This "cluster-specific" strategy avoids the deficiency of generalized models in different operating condition ranges, significantly improving the prediction accuracy and control effect of the model in specific load ranges. Furthermore, since the broadband learning system model itself has the characteristics of fast training and incremental learning, it is possible and efficient to train multiple independent models in parallel for multiple operating condition clusters.

[0045] In some embodiments, the hybrid frog-leap-simulated annealing algorithm specifically includes the following:

[0046] Optimization objective: For each cluster of operating conditions, the goal of the hybrid frog-leap-simulated annealing algorithm is to find a set of optimal hyperparameters for the broadband learning system to minimize a composite fitness function.

[0047] Collaborative optimization mechanism: The hybrid frog-leap-simulated annealing algorithm divides the frog swarm into multiple meme groups. Each group performs a depth search in its own local space and periodically exchanges information between groups to achieve global information sharing. On this basis, a simulated annealing operator is introduced to perturb the best frog in each meme group. By utilizing its probabilistic leaping ability, it helps the algorithm effectively escape the local optimum in the working range.

[0048] Specifically, the optimization objective aims to provide a clear direction for the hybrid frog-leap-simulated annealing algorithm. The "optimal hyperparameters of the broadband learning system" refer to the parameter combination that enables the broadband learning system model to achieve optimal performance under specific conditions. Examples include the number of nodes in the feature mapping layer, the number of nodes in the enhancement layer, and the regularization coefficient. These hyperparameters directly affect the model's learning ability, generalization ability, and computational efficiency. The "composite fitness function" is a comprehensive evaluation metric that integrates multiple interrelated performance objectives (such as prediction accuracy, emission targets, and control stability) through a weighted or multi-objective function, quantifying the merits of the current hyperparameter combination. By minimizing this composite fitness function, the algorithm can simultaneously consider multiple optimization objectives, thereby obtaining a more comprehensive and practically suitable hyperparameter configuration.

[0049] In the collaborative optimization mechanism, the hybrid frog-leap-simulated annealing algorithm achieves structured management of the search space by logically dividing the entire "frog swarm," i.e., the set of candidate solutions, into multiple independent "meme groups." Each meme group can be regarded as an independent subpopulation, responsible for exploring different regions of the search space. This division helps the algorithm maintain diversity while conducting parallel searches of multiple potential optimal regions, thereby improving the global search efficiency. For example, the frog swarm can be divided by random assignment, randomly distributing the initially generated candidate solutions into a preset number of meme groups; or, clustering can be performed based on the fitness values ​​of candidate solutions or their positional similarity in the search space, ensuring that each meme group has certain local characteristics or diversity in the initial stage.

[0050] Based on the meme group division, each meme group is endowed with the ability to perform "deep search" within its own local search space. This means that each meme group concentrates its resources on fine-grained iterative optimization of its internal candidate solutions in order to find the optimal solution within that local region. This deep search is typically implemented using the local search strategy of the Simulated Frog Leaping Algorithm (SFLA), where frogs within a group update their positions based on information from the best frog within the group and the globally best frog, gradually converging to a local optimum. For example, local deep search can employ a gradient descent-based local optimization method, calculating gradients within a small range on the fitness function within the meme group to guide frogs towards better solutions; alternatively, a random walk strategy can be used, randomly adjusting the positions of frogs within the meme group with a certain step size and direction, accepting better solutions while also accepting worse solutions with a certain probability, thus enhancing local exploration capabilities.

[0051] To avoid each meme group getting trapped in its own local optima due to independent searches, the hybrid frog-leap-simulated annealing algorithm introduces a "periodic information exchange" mechanism. This means that after a certain number of local depth search iterations, different meme groups will exchange information, such as sharing information about their respective optimal solutions, or copying excellent individuals from one meme group to other meme groups. This global information sharing mechanism ensures that the algorithm can integrate the discoveries from different local search regions, promote the spread of knowledge, and thus guide the entire frog swarm to converge toward the global optimum. For example, information exchange can be manifested as comparing the best frog in each meme group (local optimum) with the best frog in the entire swarm (global optimum) and updating the position of the corresponding frog based on the comparison result; or, an individual migration strategy between meme groups can be adopted, such as migrating some excellent frogs from one meme group to other meme groups, while removing poorer frogs from other meme groups, to promote population diversity and information flow.

[0052] Building upon this foundation, to further enhance the algorithm's ability to escape local optima, the hybrid frog-leap-simulated annealing algorithm introduces a "simulated annealing operator" to perturb the "optimal frog" within each meme group. The simulated annealing operator is a heuristic search method based on the physical annealing process, allowing the algorithm to accept solutions worse than the current one with a certain probability during the search process. By applying this perturbation to the locally optimal frog, the algorithm can temporarily deviate from the current optimal path, exploring a wider search space within its neighborhood, thus creating opportunities to discover better global optima. For example, the simulated annealing operator can gradually reduce the probability of accepting worse solutions exponentially based on a preset annealing temperature curve, maintaining high exploratory activity in the early stages of the search and tending towards convergence in the later stages; alternatively, an adaptive annealing strategy can be adopted, dynamically adjusting the annealing temperature according to the algorithm's search progress and population diversity to better balance exploration and exploitation.

[0053] One of the core advantages of simulated annealing is its "probabilistic leapfrogging capability." This capability allows the algorithm to avoid simply stagnating when faced with local optima, but rather to "jump out" of the current local optimum with a certain probability and enter a new search space. This "jump" is not blindly random, but is influenced by the annealing temperature and fitness difference, making the algorithm more likely to accept worse solutions for broad-area exploration in the early stages of the search, and more inclined to accept better solutions for fine-tuning in the later stages. Through this mechanism, the hybrid frog-leapfrog-simulated annealing algorithm can overcome the weakness of traditional optimization algorithms that are prone to getting trapped in local optima, significantly improving its global search capability in complex multimodal optimization problems. For example, the probabilistic leapfrogging capability can be achieved through the Metropolis criterion: if the new solution is better than the current solution, the new solution is accepted; if the new solution is worse than the current solution, the new solution is accepted with a probability related to the temperature and fitness difference.

[0054] Through the above technical solution, this application effectively solves the problem of hyperparameter optimization easily getting trapped in local optima during the training process of hierarchical expert predictive control models by introducing a hybrid frog-leapfrog-simulated annealing algorithm and designing its unique collaborative optimization mechanism. This is because the boiler efficiency surface has multiple local optima. The algorithm first sets a clear optimization objective: to find a set of optimal broadband learning system hyperparameters to minimize a composite fitness function. This provides a clear optimization direction for the algorithm, ensuring focus on key performance indicators of the model.

[0055] In the collaborative optimization mechanism, the algorithm divides the frog swarm into multiple meme groups, allowing each meme group to conduct a depth search in its own local space. This enables in-depth exploration of potential optimal regions, given the complex structure of the boiler efficiency surface. Simultaneously, by periodically exchanging information between groups, global information sharing is achieved, effectively overcoming the limitations of local search, promoting the fusion of knowledge across different regions, and avoiding duplication or omissions caused by independent searches by each meme group.

[0056] Building upon this, a simulated annealing operator is introduced to perturb the optimal frog within each meme group. Utilizing its probabilistic jump capability, the algorithm can accept temporarily inferior solutions with a certain probability, thus effectively escaping the local optimum trap. This mechanism is particularly well-suited to the multi-peak characteristics faced by hyperparameter optimization in boiler combustion control, significantly enhancing the algorithm's robustness and global search capability under varying operating conditions.

[0057] Through the aforementioned technical solutions, the hybrid leapfrog-simulated annealing algorithm can more comprehensively and reliably find the optimal hyperparameter combination of the broadband learning system, thereby training an expert predictive control model with superior performance and stronger generalization ability. This not only improves the accuracy and stability of boiler combustion control within each operating condition range, but also lays a solid foundation for subsequent model integration and online correction, ultimately contributing to energy conservation, emission reduction, and pollutant reduction of the boiler across the entire operating range.

[0058] In some embodiments, the mathematical expression for the composite fitness function is: ;in, The mean square error of the predictions made by the expert predictive control model. The nitrogen oxide emission predictions are from the nitrogen oxide surrogate model. The integral of the rate of change of the control signal. , , All are weighting factors.

[0059] This composite fitness function aims to quantify the overall performance of a given combination of hyperparameters in a broadband learning system. Its mathematical expression integrates multiple interrelated but potentially conflicting optimization objectives into a single evaluation metric through a weighted summation.

[0060] Mean square error (MSE) is a commonly used metric for measuring the accuracy of a prediction model. It calculates the average of the squares of the differences between the predicted and actual values. Minimizing MSE ensures that the expert predictive control model can accurately predict the boiler's operating status or control output. Besides MSE, other statistical metrics such as mean absolute error (MAE), root mean square error (RMSE), or coefficient of determination (R²) can be used to evaluate the model's prediction accuracy to meet the error sensitivity requirements in different scenarios.

[0061] Nitrogen oxides (NOx) are one of the main pollutants produced during boiler combustion, and their emissions directly affect environmental protection requirements. Here, NOx represents the predicted NOx emission value using a pre-built NOx surrogate model. This surrogate model can quickly estimate NOx emission levels based on boiler operating parameters (such as combustion temperature, oxygen content, and fuel type), thus serving as a basis for evaluating environmental performance during the optimization process.

[0062] The integral term of the control signal's rate of change measures the smoothness and stability of the control signal. Excessively rapid changes in the control signal can lead to frequent actuator movements, increased wear, and even system oscillations. By penalizing the rate of change of the control signal, optimization algorithms can be prompted to find more stable and gradual control strategies, thereby extending equipment lifespan and improving system operational stability.

[0063] Weighting factors , , These weights are used to adjust the relative importance of each objective term in the composite fitness function. By setting these weights appropriately, model accuracy, pollutant emissions, and control stability can be balanced according to actual operational needs (e.g., prioritizing energy conservation in some operating conditions and environmental protection in others). These weight factors can be preset based on expert experience or dynamically optimized through offline experiments, multi-objective optimization algorithms, or online adaptive adjustment mechanisms to adapt to changes in the boiler operating environment.

[0064] The above technical solution integrates the three key performance indicators of model prediction accuracy, pollutant emission, and control signal stability into a unified composite fitness function, enabling the hybrid frog-leap-simulated annealing algorithm to comprehensively and balancedly consider multiple optimization objectives when optimizing the hyperparameters of a broadband learning system.

[0065] Specifically, by minimizing the MSE term, the accurate prediction of boiler operating status by the expert predictive control model is ensured, laying the foundation for precise control; by minimizing the NOx term, environmental protection requirements are directly incorporated into the optimization objective, effectively guiding the control strategy towards low nitrogen oxide emissions; by minimizing the integral term of the control signal change rate, drastic fluctuations in the control signal are effectively suppressed, thereby reducing wear on the actuators and improving the system's operational stability and equipment lifespan.

[0066] In addition, by flexibly adjusting the weighting factors , , This application can dynamically adjust the emphasis of various optimization objectives according to the priority of different operating conditions or operating strategies, and realize the effective balance and synergistic optimization among multiple objectives. Thus, while ensuring prediction accuracy, it also takes into account environmental benefits and system stability, and significantly improves the overall performance of boiler combustion control.

[0067] In some embodiments, the nitrogen oxide surrogate model is a multilayer perceptron neural network, and its construction steps are as follows:

[0068] T1. Extracting data from historical datasets... The closely related feature parameters are generated as input to the nitrogen oxide surrogate model.

[0069] T2, the actual measurement The concentration value is used as the output label of the nitrogen oxide surrogate model.

[0070] T3. Train the multilayer perceptron neural network using historical datasets until its prediction error converges.

[0071] The nitrogen oxide surrogate model is a mathematical model used to predict nitrogen oxide (NOx) emissions during boiler combustion. Its core function is to quickly and accurately assess the impact of different control strategies on NOx emissions during the optimization process. The multilayer perceptron neural network is a feedforward artificial neural network consisting of at least three layers: an input layer, a hidden layer, and an output layer. Its key feature is its ability to handle complex nonlinear relationships through nonlinear activation functions and to learn the mapping relationship between inputs and outputs through backpropagation training.

[0072] Step T1 aims to provide high-quality input data for the NOx surrogate model, ensuring that the model can capture key factors influencing NOx emissions. Historical datasets typically contain various sensor data and control parameters from boiler operation. Feature parameters closely related to NOx formation refer to physical quantities that directly or indirectly affect NOx formation conditions such as combustion temperature, oxygen concentration, and residence time. By extracting these key features, the input dimensionality of the model can be effectively reduced, noise interference can be minimized, and the training efficiency and prediction accuracy of the model can be improved. For example, parameters such as boiler load, air supply, fuel gas volume, furnace temperature, flue gas oxygen content, and flue gas outlet temperature can be selected as input features based on domain expert knowledge and experience. These parameters directly reflect the combustion state and flue gas conditions and have a significant impact on NOx formation. In addition, data-driven methods can be used, such as feature selection techniques like correlation analysis, mutual information, or principal component analysis, to automatically filter out feature parameters with strong correlation to NOx emissions from a large amount of historical operating data, ensuring the effectiveness and independence of the input features.

[0073] Step T2 clarifies the learning objective of the nitrogen oxide surrogate model: predicting actual NOx emission concentrations. The actual measured NOx concentration values ​​are real emission data acquired in real-time or periodically by equipment such as flue gas analyzers during boiler operation. These values ​​represent the actual NOx emission level of the boiler under specific operating conditions. Using these real values ​​as the model's output labels allows the multilayer perceptron neural network to learn the mapping relationship between input features and actual emissions, thereby achieving accurate prediction of NOx emissions.

[0074] Step T3 describes the training process of the multilayer perceptron neural network, aiming to bring its predictive ability to a stable and acceptable level. The training process typically involves dividing the historical dataset into training, validation, and test sets, and iteratively adjusting the weights and biases of the neural network using backpropagation and an optimizer to minimize the error between the predicted values ​​and the actual labels. "Prediction error convergence" means that the model's performance metrics on the training and validation sets have stabilized and no longer significantly decreased, indicating that the model has sufficiently learned the patterns in the data and avoided overfitting or underfitting.

[0075] By explicitly defining the NOx surrogate model as a multilayer perceptron neural network and standardizing its construction steps, this application effectively addresses the potential limitations of surrogate models in terms of accuracy, efficiency, and generalization ability. The multilayer perceptron neural network, with its powerful nonlinear fitting capability, can accurately capture the nonlinear relationship between the complex NOx formation mechanism and operating parameters during boiler combustion, thereby providing more reliable NOx emission predictions in the composite fitness function.

[0076] Specifically, step T1 extracts feature parameters closely related to NOx generation as input, effectively focusing on the core factors affecting emissions, reducing interference from irrelevant noise, and improving the quality of input data and model training efficiency. Step T2 uses the actual measured NOx concentration value as the output label, ensuring that the model directly learns the real emission pattern, making the prediction results more realistic and applicable. Step T3 trains the multilayer perceptron neural network using historical datasets until the prediction error converges, ensuring that the model can stably and efficiently generate high-precision emission predictions during the offline optimization phase.

[0077] This explicit and standardized approach to constructing the NOx surrogate model enables the hybrid leapfrog-simulated annealing algorithm to more accurately assess the impact of different control strategies on NOx emissions when optimizing the hyperparameters of the broadband learning system. This allows for effective optimization of environmental performance within the composite fitness function. This not only improves the overall performance of the hierarchical expert predictive control model, giving it superior environmental performance across the entire operating range, but also avoids optimization direction deviations caused by inaccurate surrogate models by providing reliable NOx predictions. Ultimately, this contributes to achieving efficient and low-emission operation of the boiler combustion process.

[0078] In some embodiments, real-time loop control specifically includes the following:

[0079] S41. The control system receives real-time operating parameters from the boiler. These parameters are the basis for precise control, ensuring that control decisions are based on the latest actual operating status of the boiler.

[0080] S42. The operating condition identifier determines the current operating condition range and activates the corresponding expert predictive control model. The function of the operating condition identifier is to quickly and accurately determine the current operating state of the boiler, thereby matching the most suitable pre-trained control model for that state.

[0081] S43. The activated expert predictive control model calculates and generates a basic optimal control signal, which includes the blower frequency setpoint and the gas valve opening setpoint. This expert predictive control model is optimized and trained for a specific operating condition range, and can provide preliminary and accurate control commands. Specifically, this expert predictive control model is a pre-trained neural network model, such as a broadband learning system (BLS), which is optimized and trained offline for a specific operating condition range and can quickly output the corresponding blower frequency setpoint and gas valve opening setpoint based on real-time input.

[0082] S44. The online sequential learning module continuously learns the residual between the basic optimal control signal and the actual optimal control signal to generate a compensation signal. This online sequential learning module is responsible for capturing and compensating for model prediction errors caused by factors such as equipment aging, fuel composition fluctuations, or unmodeled dynamics in real time. For example, the online sequential learning module uses the Online Sequential Extreme Learning Machine (OS-ELM) algorithm, which can learn new data samples incrementally and update its internal parameters in real time, thereby capturing the dynamic residual between the basic optimal control signal and the actual optimal control signal.

[0083] S45. The compensation signal is superimposed on the basic optimal control signal to form the final control signal, which is then sent to the actuator. This superposition operation integrates the basic control strategy and real-time correction information, ensuring the accuracy and coordination of the final control signal.

[0084] Through the above technical solution, this application effectively solves the problems of low efficiency, response delay, and unstable control signals in real-time control loops. Specifically, by acquiring boiler operating parameters in real time through step S41, the timeliness of control decisions and the accuracy of the data basis are ensured. In step S42, the operating condition identifier can quickly determine the current operating condition range and activate the corresponding expert predictive control model, significantly reducing the delay of model switching and improving the timeliness and pertinence of control response. In step S43, the activated expert predictive control model generates a preliminary and accurate basic optimal control signal for the specific operating condition, laying a reliable foundation for subsequent compensation. In step S44, the online sequential learning module continuously learns the residual between the basic optimal control signal and the actual optimal control signal and generates a compensation signal, enabling the system to capture and compensate for model prediction errors caused by equipment aging, fuel fluctuations, or unmodeled dynamics in real time, significantly enhancing the adaptability and robustness of the control system. Finally, step S45 superimposes the compensation signal with the basic optimal control signal to form a precise and coordinated final control signal, which is then sent to the actuator. This mechanism maintains the stability of the basic control strategy while achieving rapid response to dynamic changes through fine-tuning, ensuring the stability of the control signal execution, thereby improving the overall control effect of the boiler combustion process and realizing efficient, low-emission, and stable operation of the boiler under all operating conditions.

[0085] In some embodiments, online adaptive correction also includes the following:

[0086] S5. Online Learning and Correction: The control system calculates the actual operating thermal efficiency of the boiler over a period of time, i.e. the actual optimal control value. At the same time, the control system uses a performance curve model with a higher-order load as input and the theoretical maximum efficiency as output to deduce the residual between the actual optimal control value and the model output value. This residual is used as a new training sample to update the online sequential learning module.

[0087] Specifically, the control system calculates the boiler's actual operating thermal efficiency over a past period and uses it as the actual optimal control value. Here, "actual operating thermal efficiency" refers to the efficiency with which the boiler converts the chemical energy of fuel into usable thermal energy under actual operating conditions; it is a key indicator for measuring the boiler's operational economy. Using it as the actual optimal control value aims to provide a performance benchmark based on real physical processes to evaluate the actual effectiveness of the control system.

[0088] Simultaneously, the control system uses a performance curve model, which takes a higher-order load as input and outputs the theoretical maximum efficiency, to deduce the residual between the actual optimal control value and the model output value. This "performance curve model" is a model used to describe the theoretical optimal operating efficiency of the boiler under different loads. The model takes a "higher-order load" as input, which can be understood as including not only instantaneous load but also potentially more comprehensive load information such as load change rate and historical load trends, to more accurately reflect the boiler's operating state. The model's output is the "theoretical maximum efficiency" under that load condition, i.e., the efficiency achievable under ideal or optimal control strategies. This performance curve model can be constructed in various forms. For example, it can be a mathematical model fitted using statistical methods such as polynomial regression and nonlinear regression based on a large amount of historical operating data, used to describe the relationship between load and theoretical efficiency; or, the model can be a simulation model based on the boiler's physical mechanisms, predicting the theoretical maximum efficiency by simulating the thermodynamic processes under different loads. By comparing the actual optimal control value (i.e., the actual operating thermal efficiency) with the theoretical maximum efficiency output by this performance curve model, the residual between the two can be deduced. This residual quantifies the deviation between actual operation and theoretical optimal operation, and may reflect problems such as the control system not being fully optimized, equipment performance degradation, or changes in environmental factors.

[0089] Furthermore, this residual is used as a new training sample to update the online sequential learning module. Here, the "residual" is considered an error signal or learning target, representing the gap between the current control strategy and the actual optimal performance. Using it as a new training sample means the system will utilize this error information to improve its learning ability. The "online sequential learning module" is a machine learning module capable of continuously learning and adapting to changes. When it receives a new training sample (i.e., the residual), the module adjusts its parameters or structure according to its internal update mechanism. For example, the online sequential learning module is an online sequential extreme learning machine (OS-ELM), which uses this residual as a target value to adjust its output layer weights through an online update algorithm to reduce future prediction errors.

[0090] Through the above technical solution, the control system can periodically calculate the actual operating thermal efficiency of the boiler and use it as the actual optimal control value, which provides a reliable performance benchmark based on real physical processes for the online sequential learning module.

[0091] Given that factors such as equipment aging, fuel quality fluctuations, or changes in environmental conditions can cause the boiler's actual optimal operating point to drift, relying solely on the residual between model predictions and the basic control signal for correction may not accurately capture these long-term changes. By introducing a performance curve model with a higher-order load as input and theoretical maximum efficiency as output, the system can establish a dynamic theoretical optimal efficiency benchmark, which more comprehensively reflects the ideal performance of the boiler under different operating conditions. Comparing the actual optimal control value with the output of this performance curve model, the residual between the two is derived. This residual not only includes short-term deviations in model predictions but, more importantly, effectively captures the drift of the actual optimal operating point caused by long-term factors. Using this more representative residual as a new training sample to update the online sequential learning module allows the module to continuously learn and adapt to long-term changes in boiler operating characteristics, thereby significantly improving the accuracy and robustness of online adaptive correction. This ensures that even with long-term boiler operation and gradual degradation of equipment performance, the control system can continuously operate the boiler close to its actual optimal efficiency, maintaining efficient and low-emission operation, effectively avoiding the problem of decreased control accuracy caused by model drift.

[0092] In some embodiments, the calculation of the actual operating thermal efficiency of the boiler specifically includes the following:

[0093] The control system calculates the actual operating thermal efficiency of the boiler over a past period using a preset time scale through the boiler thermal efficiency evaluation module. The boiler thermal efficiency evaluation module calculates the actual flue gas heat loss and heat dissipation loss of the boiler based on the heat loss method for calculating the thermal efficiency of industrial boilers, so as to obtain the actual operating thermal efficiency of the boiler.

[0094] Specifically, the "preset time scale" refers to the time interval at which thermal efficiency calculations are performed periodically and pre-defined in the control system. This time scale can be set to a fixed time interval, such as performing calculations every 1 minute, 5 minutes, or 10 minutes, to ensure the periodicity and consistency of the thermal efficiency calculations. Furthermore, this time scale can also be dynamically adjusted according to changes in boiler operating conditions. For example, when the load change rate exceeds a certain threshold, the calculation cycle is shortened; when the load is stable, the calculation cycle is extended, thereby providing a stable and reliable data foundation for online correction and avoiding data fluctuations caused by irregular calculation frequencies.

[0095] The "Boiler Thermal Efficiency Assessment Module" is a dedicated software or hardware unit for performing boiler thermal efficiency calculations. This module can be a standalone software subroutine or function block integrated into the central processing unit of the control system, responsible for centrally processing data acquisition, algorithm execution, and result output related to thermal efficiency calculations. Alternatively, this module can be a dedicated embedded hardware module with built-in thermal efficiency calculation algorithms and logic, interacting with the main control system via a communication interface to ensure the professionalism and accuracy of the calculations.

[0096] In the calculation process, the "actual operating thermal efficiency of the boiler over a past period" refers to the thermal efficiency over a continuous time window prior to the current calculation moment. Using data over a period of time, rather than instantaneous data, aims to smooth out instantaneous fluctuations and provide a more representative and stable thermal efficiency value. This can be achieved by averaging multiple instantaneous thermal efficiency values ​​collected within that time period, or by calculating based on the cumulative input and output energy within that time period, thereby enhancing the robustness of online corrections.

[0097] The "heat loss method for calculating the thermal efficiency of industrial boilers" is a standardized approach that indirectly calculates thermal efficiency by measuring various heat losses during boiler operation. This method typically includes the measurement and calculation of various losses, such as flue gas heat loss, heat dissipation loss, heat loss from incomplete chemical combustion, and heat loss from incomplete mechanical combustion. Alternatively, a simplified heat loss method can be used, focusing primarily on flue gas heat loss and heat dissipation loss. This simplifies the calculation process while maintaining a certain level of accuracy, ensuring the scientific validity, accuracy, and reliability of the thermal efficiency calculation and avoiding biases caused by subjective estimations or simplified models.

[0098] Among them, "actual flue gas heat loss" refers to the heat loss carried away by the boiler flue gas, which is one of the largest heat loss items of the boiler. It is usually calculated by measuring parameters such as flue gas temperature, flue gas oxygen content, ambient temperature, and fuel composition, combined with the flue gas volume calculation formula.

[0099] Through the above technical solutions, the control system calculates thermal efficiency on a preset time scale, ensuring the periodicity and consistency of the calculation and avoiding data fluctuations caused by inappropriate time intervals, thus providing a stable basis for residual calculation. Specialized calculations are performed through the boiler thermal efficiency evaluation module, improving the reliability and professionalism of the calculations and preventing errors introduced by non-specialized methods. Calculating thermal efficiency over a past period provides more stable values ​​based on historical data, reducing the impact of instantaneous fluctuations on the results and enhancing the representativeness of the thermal efficiency values. The boiler thermal efficiency evaluation module uses the heat loss method based on industrial boiler thermal efficiency calculations, employing standardized methods to calculate actual flue gas heat loss and heat dissipation loss, ensuring the scientific validity and accuracy of the calculation method and avoiding deviations caused by subjective estimations or simplified models. Finally, the actual operating thermal efficiency of the boiler is obtained, which is used to deduce the residual between the actual optimal control value and the model output value, thereby accurately updating the online sequential learning module. This effectively compensates for model drift, improves the robustness of long-term system operation, and thus improves the accuracy of online correction and system stability, achieving more stable and efficient boiler operation.

[0100] This invention also proposes a boiler combustion control system based on hierarchical collaborative optimization and online correction, comprising: a control module, a sensor module, and an execution module; both the sensor module and the execution module are signal-connected to the control module; the control module is configured to execute any of the boiler combustion control methods based on hierarchical collaborative optimization and online correction in the above embodiments; the control module integrates an operating condition identifier, a hierarchical collaborative optimization control model, and an online correction module; the operating condition identifier determines the current operating condition range based on the real-time operating data of the boiler; the hierarchical collaborative optimization control model includes multiple expert predictive control models based on a broadband learning system, which are obtained in the offline hierarchical collaborative optimization stage through a hybrid leapfrog-simulated annealing algorithm; the online correction module is a residual compensator based on an online sequential learning module, used to perform online correction on the basic optimal control signal output by the expert predictive control model; the execution module includes a blower frequency converter and a gas valve execution servo motor, the blower frequency converter adjusts the blower frequency according to the final control signal, and the gas valve execution servo motor adjusts the gas valve opening according to the final control signal.

[0101] The operating condition identifier determines the current operating condition range based on the boiler's real-time operating data, ensuring that the system can dynamically match the most suitable control strategy according to the actual operating state. The hierarchical collaborative optimization control model includes multiple expert predictive control models based on a broadband learning system. These models are obtained in the offline hierarchical collaborative optimization stage through a hybrid leapfrog-simulated annealing algorithm, performing in-depth optimization for specific operating condition ranges. This effectively avoids the defect of a single global model easily getting trapped in local optima on complex efficiency surfaces. The online correction module, as a residual compensator based on the online sequential learning module, is used to correct the basic optimal control signal output by the expert predictive control model online, continuously learning and compensating for model drift caused by factors such as equipment aging and changes in fuel composition.

[0102] The execution module includes a blower frequency converter and a gas valve execution servo motor. The blower frequency converter adjusts the blower frequency according to the final control signal, and the gas valve execution servo motor adjusts the gas valve opening according to the final control signal to achieve precise control of the combustion process.

[0103] The core innovation of this embodiment lies in effectively solving the problem of insufficient control accuracy caused by dynamic load changes, equipment aging, and unmodeled dynamics during long-term boiler operation under all operating conditions. Specifically, the operating condition identifier determines the current operating condition range in real time, activates the corresponding expert predictive control model to generate basic control signals, while the online correction module works in parallel to fine-tune and compensate for the basic control signals, ensuring efficient and low-emission operation under different load conditions. This architecture not only guarantees local optimal control in each operating condition range but also maintains long-term operational stability and accuracy through an online adaptive mechanism.

[0104] Through the above technical solution, the system can provide precise control strategies for the dynamic characteristics of boiler operation, significantly improving energy utilization efficiency, reducing pollutant emissions, and extending the efficient operation cycle of the control system. Compared with the basic solution, this application has the dual advantages of segmented optimization and continuous self-adaptation, which can effectively cope with the complex control challenges of boilers across the entire operating range and ensure that the control model maintains high accuracy and robustness during equipment aging.

[0105] In some embodiments, the boiler combustion control system based on hierarchical collaborative optimization and online correction further includes a monitoring module, which is signal-connected to the control module. The monitoring module includes a cloud server and a user interface. The control module packages the collected and calculated key data in a predefined format and transmits it to the cloud server via 4G, 5G, or wired networks through the MQTT lightweight IoT protocol. The application logic on the cloud server parses the data and persistently stores it in a time-series database. Users can log in and access this data through a client application on a network-connected device to achieve visualized monitoring and management of the boiler status.

[0106] Other configurations and operations of the boiler combustion control system and method based on hierarchical collaborative optimization and online correction according to embodiments of the present invention are known to those skilled in the art and will not be described in detail here.

[0107] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0108] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A boiler combustion control method based on hierarchical collaborative optimization and online correction, characterized in that, It includes two stages: offline hierarchical collaborative optimization and online adaptive correction. The offline hierarchical collaborative optimization includes the following contents: S1, data preparation and working condition clustering: based on historical operation data, an unsupervised learning algorithm is used to automatically divide the continuous operation working conditions of the boiler into multiple discrete, typical working condition intervals with significantly different characteristics; S2, hierarchical expert predictive control model training: for each working condition interval, independently and in parallel, a hybrid frog leap-simulated annealing algorithm is used to optimize the hyperparameters of the wideband learning system, thereby training an expert predictive control model for each working condition interval; S3, model integration: integrate the expert predictive control models of all working condition intervals with a working condition identifier to form a unified hierarchical collaborative optimization control model; The online adaptive correction includes the following contents: S4, real-time cycle control: the control system runs an online sequential learning module in parallel, which continuously learns the residual error between the predicted value and the actual optimal value of the current expert predictive control model, and adjusts and compensates the basic control signal according to the residual error, realizing online adaptive correction of the expert predictive control model.

2. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 1, wherein, Data preparation and working condition clustering specifically includes the following contents: S11, collect historical data of the boiler for at least one complete operation cycle; S12, use the K-Means clustering algorithm to automatically divide the data set into K working condition clusters based on steam pressure and load change rate as the main features.

3. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 2, wherein, Wideband learning system definition: define an independent wideband learning system model for each working condition cluster.

4. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 3, wherein, The hybrid frog leap-simulated annealing algorithm specifically includes the following contents: Optimization target: for each working condition cluster, the goal of the hybrid frog leap-simulated annealing algorithm is to find a set of optimal wideband learning system hyperparameters to minimize a composite fitness function; Collaborative optimization mechanism: the hybrid frog leap-simulated annealing algorithm divides the frog population into multiple meme groups, each group performs deep search in its local space, and periodically exchanges information between groups to achieve global information sharing; On this basis, the simulated annealing operator is introduced to disturb the optimal frog in each meme group, and its probability jump ability is used to help the algorithm effectively escape from the local optimal solution in the working condition interval.

5. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 4, wherein, The mathematical expression of the compound fitness function is: ; wherein, is the mean square error of the expert predictive control model prediction, is the nitrogen oxide emission prediction value of the nitrogen oxide proxy model, is the integral of the control signal change rate, , , are all weight factors.

6. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 5, wherein, The nitrogen oxide proxy model is a multi-layer perceptron neural network, and its construction steps are as follows: T1, extracting from the historical data set the features related to Generating closely related feature parameters as inputs of the nitrogen oxide proxy model; T2, the actual measured concentration value as an output label for the nitrogen oxides proxy model; T3, use the historical data set to train the multi-layer perceptron neural network until its prediction error converges.

7. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 1, wherein, Real-time cycle control specifically includes the following contents: S41, the control system receives real-time operation parameters of the boiler; S42, determine the current working condition interval and activate the corresponding expert predictive control model through the working condition identifier; S43, calculate and generate the basic optimal control signal from the activated expert predictive control model, which includes the frequency set value of the air supply fan and the opening set value of the gas valve; S44, continuously learn the residual error between the basic optimal control signal and the actual optimal control signal through the online sequential learning module to generate a compensation signal; S45, superimpose the compensation signal and the basic optimal control signal to form the final control signal and send it to the actuator.

8. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 1, wherein, The online adaptive correction further comprises the following: S5, online learning and correction: the control system calculates the actual operating thermal efficiency of the boiler in the past period, that is, the actual optimal control value, and at the same time, the control system reverses the residual error between the actual optimal control value and the model output value through a performance curve model with higher order load as input and theoretical maximum efficiency as output, and the residual error is used as a new training sample to update the online sequential learning module.

9. The boiler combustion control method based on hierarchical co-optimization with online correction of claim 8, wherein, The calculation of the actual operating thermal efficiency of the boiler specifically comprises the following: The control system calculates the actual operating thermal efficiency of the boiler in the past period through the boiler thermal efficiency evaluation module at a preset time scale; The boiler thermal efficiency evaluation module calculates the actual flue gas heat loss and heat loss of the boiler based on the heat loss method of industrial boiler thermal efficiency calculation to obtain the actual operating thermal efficiency of the boiler.

10. A boiler combustion control system based on hierarchical co-optimization with online correction, characterized in that, It comprises: a control module, a sensor module and an execution module; The sensor module and the execution module are both in signal connection with the control module; The control module is configured to execute the boiler combustion control method based on hierarchical collaborative optimization and online correction according to any one of claims 1-9; The control module is integrated with a working condition identifier, a hierarchical collaborative optimization control model and an online correction module; The working condition identifier determines the current working condition interval based on real-time operation data of the boiler; The hierarchical collaborative optimization control model comprises a plurality of expert predictive control models based on a wideband learning system, which are obtained through a hybrid frog leap-simulated annealing algorithm in an offline hierarchical collaborative optimization stage; The online correction module is a residual error compensator based on an online sequential learning module, used for online correction of the basic optimal control signal output by the expert predictive control model; The execution module comprises a blower frequency converter and a gas valve execution servo motor, the blower frequency converter adjusts the blower frequency according to the final control signal, and the gas valve execution servo motor adjusts the gas valve opening degree according to the final control signal.

11. The boiler combustion control system based on hierarchical co-optimization with online correction of claim 10, wherein, It also comprises a monitoring module in signal connection with the control module; The monitoring module comprises a cloud server and a user interface, the control module packages the key data collected and calculated according to a predefined format, transmits it to the cloud server through the MQTT lightweight Internet of Things protocol via 4G, 5G or wired network, and the application logic on the cloud server analyzes the data and stores it in a time series database; Users can log in and access these data through the client application on the network-connected device, realize the visual monitoring and management of the boiler state.

Images