Multi-parameter self-adaptive coordination system and method for high-alkali coal boiler

Abstract

The invention belongs to the technical field of thermal power generation process control, and provides a multi-parameter adaptive coordination system and method for a high-alkali coal boiler, and the method comprises the steps: extracting operation state features representing a current working condition based on the real-time operation data of the boiler; generating control action suggestions of at least two different sources in parallel based on the features, and synchronously performing security risk assessment; obtaining a dynamic boundary constraint based on the operation state characteristics, screening control suggestions in combination with an evaluation result, carrying out weighted fusion and boundary correction on the screened suggestions and a preset target weight, and outputting a final control instruction; issuing the instruction to an execution mechanism for coordination control, and collecting an actual system state; according to the method, multi-source control experience and real-time optimization are fused, multi-target coordination control within the dynamic safety boundary is achieved, and the operation economical efficiency, the environmental protection performance and the flexibility of the high-alkali coal boiler under the variable working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation process control technology, and in particular to a multi-parameter adaptive coordination system and method for high-alkali coal boilers. Background Technology

[0002] High-alkali coal, as a fuel with unique combustion characteristics, exhibits complex kinetic behavior during combustion in power plant boilers. The high alkali metal content (such as sodium and potassium) in the coal easily leads to severe slagging, fouling, and corrosion problems during combustion, significantly affecting combustion stability and pollutant formation characteristics. This makes the operation and control of high-alkali coal boilers exceptionally complex and sensitive.

[0003] In the actual operation of power plant boilers, especially in the context of participating in deep peak shaving of the power grid, boilers need to frequently and significantly adjust the load. This process involves the coordinated changes of multiple key parameters such as fuel quantity, primary and secondary air volume and their ratio, burnout air volume, and feedwater flow rate. Traditional control schemes mostly rely on static parameter ratio tables obtained from experiments under fixed operating conditions, or adopt single-loop control strategies based on classic PID. These methods have obvious limitations: First, static ratios cannot adapt to the inherent coal quality fluctuations of high-alkali coal (such as changes in alkali metal content and volatile matter), causing control parameters to deviate from optimal values, resulting in decreased combustion efficiency and excessive emissions of pollutants such as NOx; second, during dynamic load changes, fixed parameter coupling relationships are difficult to achieve rapid and stable load response, and are prone to safety risks such as combustion instability or even fire extinguishing due to asynchronous adjustments of various parameters; finally, existing methods lack the ability to dynamically balance multiple control objectives (economy, environmental protection, and flexibility), and cannot explore and track the optimal operating point in real time while meeting strict safety constraints.

[0004] Although some advanced process control methods (such as fuzzy control and model predictive control) have emerged in the existing technology and have been applied to boiler optimization, these methods are still insufficient in terms of the integration and utilization of multi-source control experience, dynamic perception of operational safety boundaries, and the controller's own ability to adapt to the slow time-varying characteristics of combustion system when facing the comprehensive challenges of strong nonlinearity, large delay, strong coupling of multiple variables and time-varying operating conditions in high-alkali coal combustion.

[0005] Therefore, developing a multi-parameter adaptive coordination method for high-alkali coal boilers that can intelligently integrate multi-source decision-making, perceive safety boundaries in real time, and continuously evolve itself is of great significance for ensuring boiler safety, improving operational economy, and achieving clean and flexible peak shaving. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a multi-parameter adaptive coordination system and method for high-alkali coal boilers. This system effectively solves the difficulties in multi-parameter coordination and the challenges in balancing safety assurance and multi-objective optimization during variable operating conditions of high-alkali coal boilers. It provides a systematic intelligent control solution for achieving safe, efficient, clean, and flexible operation of high-alkali coal boilers. This invention provides a multi-parameter adaptive coordination method for high-alkali coal boilers, the coordination method comprising the following steps: Based on real-time operating status data of high-alkali coal boilers, the operating status characteristics of the current operating conditions are extracted and characterized. Based on the aforementioned operational status characteristics, at least two different sources of control action suggestions are generated in parallel, and a safety risk assessment is simultaneously performed on each type of control action suggestion to obtain the assessment results. Based on the operating state characteristics, boundary constraints are obtained. At the same time, based on the evaluation results, various control action suggestions are screened. The screened control action suggestions, the operating state characteristics, and the preset target weights are combined for weighted fusion and boundary correction to output the final control command. The final control command is sent to the actuator corresponding to the high-alkali coal boiler for multi-parameter coordinated control, and the actual system status after the final control command is executed is collected. Based on the error between the actual system state and the predicted state corresponding to the final control command, the encoder parameters for extracting the operating state features are adjusted online.

[0007] Preferably, the extraction of operating status features based on real-time operating status data of the high-alkali coal boiler to characterize the current operating condition includes: Real-time acquisition of raw operating status data of high-alkali coal boilers and acquisition of coal quality data characterizing the current alkali metal properties of coal; fusion processing of the raw operating status data and the coal quality data to generate enhanced status data. Calculate the contribution weight of each feature dimension in the enhanced state data to the current combustion condition, and weight the enhanced state data according to the contribution weight to output the weighted state features; The weighted state features are mapped to the running state features by a pre-trained encoder.

[0008] Preferably, the step of generating control action suggestions from at least two different sources in parallel based on the operating state characteristics, and simultaneously conducting a safety risk assessment on each type of control action suggestion to obtain the assessment results, including: Using the operating state characteristics as input, the control strategy trained based on historical best control data is queried and invoked, a first control action suggestion is output, and a first safety risk value corresponding to the first control action suggestion is calculated. Using the aforementioned operating state characteristics as the initial state for rolling optimization, forward rolling calculation and optimization are performed within a preset finite time domain, with the criterion of satisfying the optimal multi-objective performance indicators. The second control action suggestion is obtained, and the second safety risk value corresponding to the second control action suggestion is calculated. The first security risk value and the second security risk value together constitute the evaluation result of various control action suggestions, and are then bound to their respective corresponding control action suggestions for output.

[0009] Preferably, the step of obtaining boundary constraints based on the operating state characteristics, simultaneously filtering various control action suggestions based on the evaluation results, and combining the filtered control action suggestions, the operating state characteristics, and preset target weights for weighted fusion and boundary correction to output the final control command includes: Based on the operating status characteristics, the category to which the current operating condition belongs is identified, and according to the identified operating condition category, the corresponding set of basic boundary constraints is matched and obtained from the pre-built multi-operating condition constraint knowledge base; Based on the aforementioned operating state characteristics and the aforementioned basic boundary constraint set, and combined with real-time device status information, constraint fusion is performed to output boundary constraints; Based on the safety risk values ​​in the assessment results, various control action suggestions are initially screened; based on the boundary constraints, the initially screened control action suggestions are verified for compliance, forming a set of control suggestions to be integrated. For each control action suggestion in the set of control suggestions to be fused, the dynamic fusion weight of each suggestion under the current working condition is calculated and weighted summed, using the corresponding operating state feature and the preset target weight as input, to generate the initial fusion control command; The parameter components of the initial fusion control command are bounded and normalized to output the final control command that satisfies all physical and safety constraints.

[0010] Preferably, the step of sending the final control command to the actuator corresponding to the high-alkali coal boiler for multi-parameter coordinated control and collecting the actual system status after executing the final control command includes: Based on the operating state characteristics, the final control command is decomposed into a multi-time-scale decomposition, which is a set of commands including fast-change loop control commands and slow-change loop control commands. Before the set of instructions is sent to each actuator, the control instructions are adaptively dead-zone compensated and corrected according to the real-time status of each actuator and the preset dead-zone characteristics, and the corrected control instructions are generated. The corrected control command is synchronously sent to the corresponding actuators. At the same time, a synchronous execution signal is sent to each actuator, so that the fuel supply mechanism, air supply mechanism, induced draft mechanism and water supply mechanism can perform coordinated actions according to the timing logic of the corrected control command to achieve multi-parameter coordinated control. During the operation of each actuator, process variables of the high-alkali coal boiler are collected in real time. These process variables include furnace temperature distribution, flue gas oxygen content, main steam parameters, and actual feedback position signals of each actuator, which constitute the execution process status data. When the execution process state data reaches a quasi-steady state, data within a preset effective observation window is extracted from the execution process state data and used as the actual system state.

[0011] Preferably, the step of adjusting the encoder parameters online based on the error between the actual system state and the predicted state corresponding to the final control command includes: Based on the actual system state and the predicted state made according to the operating state characteristics and the final control command, the model prediction error is calculated; and based on the model prediction error and combined with the historical distribution characteristics of the operating state characteristics, a comprehensive loss for encoder parameter optimization is constructed. With the goal of minimizing the overall loss, the parameters of the encoder are adjusted incrementally online; and the effect of the re-extracted operating state features after adjustment is verified. The verified actual system state, the final control command, and the corresponding optimized operating state characteristics will be stored as new samples.

[0012] This invention also provides a multi-parameter adaptive coordination system for high-alkali coal boilers, used to execute the aforementioned multi-parameter adaptive coordination method for high-alkali coal boilers, the coordination system comprising: The feature acquisition module is used to extract operating status features that characterize the current operating conditions based on the real-time operating status data of high-alkali coal boilers. The risk assessment module is used to generate control action suggestions from at least two different sources in parallel based on the operating status characteristics, and simultaneously conduct safety risk assessments on various control action suggestions to obtain assessment results; The control command output module is used to obtain boundary constraints based on the operating state characteristics, and at the same time, to filter various control action suggestions based on the evaluation results, and to perform weighted fusion and boundary correction by combining the filtered control action suggestions, the operating state characteristics and the preset target weights, and output the final control command. The instruction execution module is used to send the final control instruction to the corresponding actuator of the high-alkali coal boiler, perform multi-parameter coordinated control, and collect the actual system status after the final control instruction is executed. The parameter feedback module is used to adjust the parameters of the encoder that extracts the operating state features online based on the error between the actual system state and the predicted state corresponding to the final control command.

[0013] Compared with related technologies, the multi-parameter adaptive coordination system and method for high-alkali coal boilers provided by this invention have the following beneficial effects: This invention extracts operational status characteristics that can deeply characterize the current combustion conditions, and on this basis, it integrates empirical control strategies based on historical big data with feedforward optimization results based on model predictions in parallel. This enables the control system to no longer rely on fixed rules, but to make decisions intelligently by combining "experience" and "calculation" according to real-time operating conditions, which significantly enhances its adaptability to coal quality fluctuations and load changes.

[0014] This invention dynamically generates safety boundary constraints that match the current precise operating conditions and performs preliminary safety risk assessments and boundary compliance checks on multi-source control recommendations during the decision-making process, ensuring that all control actions remain within the dynamic safety domain. By combining the fusion and correction of multi-objective weights, it can dynamically optimize and track the optimal balance point of multiple objectives such as economy, environmental protection, and flexibility, while strictly ensuring the safe operation of the boiler.

[0015] This invention ensures high-precision coordinated operation of key parameters such as fuel, air, and water by employing multi-timescale decomposition, adaptive compensation for actuator dead zones, and synchronous coordinated control, thereby improving the quality of load response. Simultaneously, by collecting quasi-steady-state actual conditions and comparing them with predicted conditions, an online learning closed loop of "decision-execution-feedback" is constructed, enabling the feature extraction encoder to continuously self-optimize. This continuously improves the system's perception and prediction accuracy of boiler characteristic changes, achieving continuous and gradual improvement of the control strategy. Attached Figure Description

[0016] Figure 1 A flowchart of a multi-parameter adaptive coordination method for high-alkali coal boilers provided by the present invention; Figure 2 The present invention provides a modular structure diagram of a multi-parameter adaptive coordination system for high-alkali coal boilers. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the drawings, not all structures. Moreover, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0018] It should also be noted that, for ease of description, the accompanying drawings show only the parts relevant to the invention and not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the drawings. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0019] Example 1 To clearly understand the basis for implementing the subsequent method steps of this invention, the physical structure and inter-component relationships of the high-alkali coal boiler system to which this method is applied are first described. This system is a complex coupled system that includes subsystems such as fuel preparation and supply, forced draft and induced draft, boiler body and steam-water system.

[0020] The fuel preparation and supply system mainly includes components such as the coal feeder, coal mill, and burner. Its core function is to prepare high-alkali coal into pulverized coal and stably feed it into the furnace. The forced draft and induced draft system consists of a primary air fan, forced draft fan, induced draft fan, and related dampers, responsible for providing and organizing the air required for combustion (primary air, secondary air, and burnout air), while maintaining negative pressure in the furnace. The boiler body and steam-water system, encompassing the furnace, water-cooled walls, superheater, steam drum (for subcritical units) or start-up separator (for supercritical units), and feedwater pump, is the core of energy conversion, transforming the chemical energy of fuel into the thermal energy of superheated steam.

[0021] These components are tightly coupled through mass and energy flow, forming a complex interactive relationship. On the one hand, changes in fuel quantity must be coordinated with air supply to ensure combustion efficiency and control pollutant generation; on the other hand, changes in fuel quantity will cause changes in boiler heat absorption, which in turn requires corresponding adjustments to feedwater flow to maintain the balance of working fluid and energy in the steam-water system, ensuring stable main steam parameters and safe drum water level.

[0022] In addition, the supply air volume and induced air volume need to be dynamically coordinated to maintain a stable negative pressure in the furnace. The special properties of alkali metals in high-alkali coal, such as their tendency to induce slagging and corrosion, make the control of the temperature field distribution in the furnace, especially the local heat load, particularly critical. This further increases the difficulty and complexity of coordinating and controlling parameters such as fuel, air volume, and feedwater.

[0023] The multi-parameter adaptive coordination method provided by this invention aims to address the multiple challenges brought about by the characteristics of high-alkali coal and the demand for deep peak shaving. Through intelligent means, it achieves coordinated and optimized control of key actuators in systems such as fuel supply, air supply and exhaust, and steam-water circulation.

[0024] The following section will elaborate on the specific steps of this method. (Refer to the original text for more details.) Figure 1 As shown, the coordination method includes the following steps: S1: Extracting and characterizing the operating status features of the current operating condition based on real-time operating status data of high-alkali coal boilers.

[0025] Specifically, step S1 includes the following steps: S11: Real-time acquisition of raw operating status data of high-alkali coal boilers, acquisition of coal quality data characterizing the current alkali metal properties of coal, fusion processing of the raw operating status data and the coal quality data to generate enhanced status data.

[0026] In this embodiment, the original operating status data of the high-alkali coal boiler is collected in real time, specifically including: continuously acquiring process parameters such as furnace temperature field distribution (measured by a multi-layer thermocouple array), inlet and outlet working fluid temperature and pressure of each heating surface, flue gas oxygen content (measured by a zirconia oxygen meter), primary and secondary air volume and pressure, coal feeder speed feedback, steam drum water level, main steam flow rate and temperature, etc., through sensors installed at various key locations in the boiler.

[0027] Meanwhile, key coal quality characteristics data such as alkali metal content (mainly sodium and potassium oxide content), volatile matter, and calorific value of the current coal can be obtained through online coal quality monitoring systems or regular laboratory analysis.

[0028] The aforementioned real-time operating status data and coal quality data were timestamped and fused: First, the two types of data were normalized to eliminate dimensional differences; then, a sliding time window (e.g., a window length of 5 minutes) was used to apply a moving average filter to the real-time operating data to suppress noise; finally, the processed real-time operating data vector was concatenated with the coal quality feature vector to form an enhanced status data vector that comprehensively represents the current boiler operating status and fuel characteristics. This enhanced status data provides a more comprehensive information foundation for subsequent feature extraction.

[0029] S12: Calculate the contribution weight of each feature dimension in the enhanced state data to the current combustion condition, and weight the enhanced state data according to the contribution weight, and output the weighted state features.

[0030] In this embodiment, the enhanced state data obtained in step S11 is used to calculate the contribution weight of each feature dimension to the current combustion condition using a feature weighting method based on an attention mechanism.

[0031] In practice, a lightweight attention network is constructed. The input to this network is an augmented state data vector. Through a fully connected layer and a Softmax activation function, the output is an attention weight vector with the same dimension as the input features. The expression for the Softmax activation function is:

[0032] in For the input vector of the th One portion, It is the input vector The first in j The value of each component The total number of feature dimensions. and j It is an index variable used to point to a vector. A specific location within. `j` points to the feature whose weight is currently being calculated; `j` is used to iterate through all features in the vector (from the first to the last). indivual), This is the output value of the Softmax function. It represents the... Each feature has an attention weight under the current operating condition, with a value between 0 and 1, and the sum of the weights of all features is 1. It is a natural constant.

[0033] The weighting calculation process fully considers the characteristics of high-alkali coal combustion: for example, when coal quality data indicates that the current coal has a high alkali metal content, the attention mechanism will automatically assign higher weights to features related to slagging tendency (such as local furnace temperature and heating surface wall temperature); when the boiler is under variable load conditions, features related to dynamic response characteristics (such as steam pressure change rate and oxygen change rate) will receive greater attention.

[0034] After calculating the attention weights, they are multiplied element-wise with the original enhanced state data to achieve weighted feature reconstruction, outputting a weighted state feature vector. This process effectively highlights the key features most relevant to the current operating condition and suppresses the interference of secondary features.

[0035] S13: The weighted state features are mapped to the running state features by a pre-trained encoder.

[0036] In this embodiment, the weighted state feature vector obtained in step S12 is input into a pre-trained deep encoder to extract low-dimensional, dense, and semantically rich running state features.

[0037] The encoder typically adopts a stacked autoencoder or variational autoencoder structure, and its training process is as follows: In the offline stage, a large amount of historical operating data of the boiler under various typical operating conditions (covering different loads and different coal qualities) is collected, and a massive amount of weighted state feature samples are constructed according to the methods described in S11 and S12.

[0038] Using this as the training set, the encoder-decoder structure is trained using unsupervised or semi-supervised learning methods. The training objective is to minimize the reconstruction error. By adding a KL divergence regularization term between the latent variables of the encoder output and the Gaussian prior distribution to the loss function, the distribution of the running state features is made to conform to the expectation.

[0039] Through training, the encoder learns to nonlinearly map high-dimensional weighted state features to a low-dimensional latent space. In online applications, the weighted state feature vector is forward-propagated through this encoder, and the output of its bottleneck layer is the desired operating state feature. This feature not only significantly reduces the data dimensionality, which is beneficial for subsequent calculations, but more importantly, it deeply condenses and represents the essential characteristics of the boiler's current operating condition, providing a core basis for subsequent intelligent decision-making.

[0040] S2: Based on the operating state characteristics, generate control action suggestions from at least two different sources in parallel, and simultaneously conduct a safety risk assessment on each type of control action suggestion to obtain the assessment results.

[0041] Specifically, step S2 includes the following steps: S21: Using the operating state characteristics as input, query and call the control strategy trained based on historical best control data, output the first control action suggestion, and calculate the first safety risk value corresponding to the first control action suggestion.

[0042] In this embodiment, the operating status characteristics obtained in step S1 are used as input to query and call the control strategy trained based on historical best control data, output the first control action suggestion and simultaneously calculate its safety risk value.

[0043] In practical implementation, the first step is to construct a historical optimal control strategy library. During the offline phase, historical data segments demonstrating excellent control performance (e.g., high efficiency, low emissions, stable operation) under various operating conditions (covering various loads and coal quality combinations) during long-term boiler operation are collected. Corresponding "state-action" pairs are extracted from these segments to form a large-scale experience database. In online applications, a fast retrieval algorithm based on similarity matching (such as locality-sensitive hashing or k-nearest neighbor algorithm) is used to match the current operating state characteristics with the historical state characteristics in the strategy library, identifying the most similar historical operating conditions.

[0044] Then, the historical best control actions corresponding to these similar operating conditions are weighted and fused (the weights are determined by the similarity) to generate a first control action suggestion, which reflects the control experience that has been proven effective under similar operating conditions in historical operation.

[0045] Simultaneously, the current operating status characteristics are combined with the generated first control action suggestion and input into a pre-trained safety risk assessment model (e.g., a classifier based on support vector machines or deep neural networks). This model learns the characteristics of a large amount of historical normal operating data and abnormal / hazardous operating condition data (such as impending slagging, unstable combustion, etc.), and can output a quantified safety risk value, representing the probability that executing the suggested action may cause an unsafe situation, i.e., the first safety risk value.

[0046] S22: Using the aforementioned operating state characteristics as the initial state for rolling optimization, within a preset finite time domain, forward rolling calculation and optimization are performed based on the criterion of satisfying the optimal multi-objective performance indicators, to obtain a second control action suggestion and calculate the second safety risk value corresponding to the second control action suggestion.

[0047] In this embodiment, the operating state characteristics obtained in step S1 are used as the initial state of model predictive control. Within a preset finite time domain, forward rolling calculation and optimization are performed based on the criterion of satisfying the optimal multi-objective performance index to obtain the second control action suggestion, and its safety risk value is calculated simultaneously.

[0048] The specific implementation process is as follows: First, establish a simplified mechanism model or data-driven model that can predict the short-term dynamics of the boiler in the future (its input is the state and action, and the output is the state prediction at the next moment, such as efficiency, NOx concentration, main steam pressure, etc.).

[0049] In each control cycle, starting from the current operating state characteristics, an optimization problem with constraints is solved using an optimization algorithm (such as sequential quadratic programming or interior point method) within a given prediction time domain. The objective function of this optimization problem is a weighted sum of multiple performance indicators (such as average combustion efficiency, cumulative NOx emissions, and the sum of squares of load tracking errors) in the future time domain. The constraints include the range of changes in the manipulated variables, rate limits, and safety constraints derived from mechanistic knowledge (such as minimum air-coal ratio).

[0050] The first action in the optimal control action sequence obtained from the solution is the second control action suggestion, which represents the result of optimizing future performance based on the current model. In parallel with S21, the current operating state characteristics are combined with this second control action suggestion and input into the aforementioned safety risk assessment model to calculate the corresponding second safety risk value.

[0051] S23: Combine the first security risk value and the second security risk value to form the evaluation result of various control action suggestions, and bind and output them with their respective corresponding control action suggestions.

[0052] In this embodiment, the first safety risk value calculated in step S21 and the second safety risk value calculated in step S22 are combined to form the safety risk assessment result of the currently generated various control action suggestions.

[0053] Specifically, the process involves binding a first security risk value to a first control action suggestion, and a second security risk value to a second control action suggestion, forming a structured dataset as the output of this step. This assessment clearly indicates the potential security risk level of each control suggestion, providing crucial security information for suggestion screening and fusion decisions in subsequent steps. For example, a security risk threshold can be set; if the risk value of a suggestion exceeds this threshold, it will be marked as a high-risk suggestion in subsequent steps for the decision-making module to consider carefully or exclude directly.

[0054] S3: Obtain boundary constraints based on the operating state characteristics, and simultaneously filter various control action suggestions based on the evaluation results. Combine the filtered control action suggestions, the operating state characteristics, and the preset target weights to perform weighted fusion and boundary correction, and output the final control command.

[0055] Specifically, step S3 includes the following steps: S31: Based on the operating state characteristics, identify the category to which the current operating condition belongs, and according to the identified operating condition category, match and obtain the corresponding set of basic boundary constraints from the pre-built multi-operating condition constraint knowledge base.

[0056] In this embodiment, the first step in implementation is to construct the multi-condition constraint knowledge base: In the offline stage, through cluster analysis of historical boiler operation data (such as using K-means or Gaussian mixture model) and expert experience, the boiler's operating range is divided into several representative operating condition categories, such as "high load stable operation", "medium load coal quality change", and "low load deep peak shaving".

[0057] Each operating condition category is associated with a set of basic boundary constraints. These constraints are determined based on boiler design specifications, safe operation procedures, and historical safe operation data under that condition. They mainly include the operating range of key parameters (such as the upper and lower limits of coal feed and air volume), coupling relationship constraints between key parameters (such as the minimum air-coal ratio and the maximum flue gas temperature limit), and change rate constraints (such as the upper limit of load change rate).

[0058] In online applications, a lightweight classifier (such as Softmax regression or a shallow neural network) is used. It takes the current operating state features as input and outputs the probability of belonging to each preset operating condition category. The operating condition category with the highest probability is selected as the recognition result, and then the basic boundary constraint set corresponding to that category is retrieved from the knowledge base. This step provides an initial boundary framework that conforms to the current operating state for subsequent precise constraint fusion.

[0059] S32: Based on the operating state characteristics and the basic boundary constraint set, and combined with real-time device state information, constraint fusion is performed to output boundary constraints.

[0060] In this embodiment, based on the basic boundary constraint set obtained in step S31 and the operating state features obtained in step S1, and combined with real-time equipment status information, constraint fusion is performed to output dynamic boundary constraints that accurately match the current state. Real-time equipment status information includes, but is not limited to, the current available output of major auxiliary machines such as fans and pumps, and the health status of actuators (such as whether there is a jamming alarm).

[0061] During implementation, a constraint fusion model is constructed. This model can be a rule-based expert system or a pre-trained data-driven model. Its core function is to receive basic boundary constraints, operational status characteristics, and equipment status information, and dynamically adjust the basic boundaries. For example, when operational status characteristics indicate a local high-temperature risk in the current furnace temperature distribution (possibly related to the tendency of high-alkali coal to slag), the constraint fusion model will appropriately tighten the upper limit constraints on the airflow or fuel quantity in the relevant area. Similarly, when equipment status information shows that the output of a certain blower is limited, the model will adjust the feasible upper limit of the total airflow and the opening of the relevant dampers accordingly. Through this fusion, static, general basic boundary constraints are dynamically modified into personalized dynamic boundary constraints adapted to the specific equipment conditions and combustion state, providing a more precise safe operating space for control decisions.

[0062] S33: Based on the safety risk values ​​in the assessment results, perform preliminary screening of various control action suggestions; perform compliance verification on the preliminary screening of control action suggestions based on the boundary constraints to form a set of control suggestions to be integrated.

[0063] In this embodiment, based on the safety risk values ​​in the evaluation results obtained in step S2, various control action suggestions are initially screened; then, based on the dynamic boundary constraints generated in step S32, the control action suggestions after initial screening are verified for compliance, and finally a set of control suggestions to be integrated is formed.

[0064] The specific operation consists of two steps: Step 1, safety risk screening. Set one or more safety risk thresholds. Compare each control action suggestion obtained in S2 with its safety risk value, and eliminate suggestions whose safety risk values ​​exceed the preset high-risk threshold to prevent obviously unsafe actions from entering subsequent processes. Suggestions with moderate risk values ​​can be retained but marked with a warning label.

[0065] The second step is boundary compliance verification. Each parameter component in the control recommendations that passed the safety screening (such as the recommended coal feed rate, recommended primary air volume, etc.) is compared with the corresponding upper and lower limits of the dynamic boundary constraints. Any parameter component exceeding its corresponding dynamic boundary constraint will cause the recommendation to be marked as non-compliant. Ultimately, only those control action recommendations that both passed the safety risk screening (i.e., the risk value is below the threshold) and whose all parameter components fall completely within the dynamic boundary constraints are retained, forming the set of control recommendations to be integrated. This step ensures that the recommendations subsequently participating in the integration possess both low safety risk and operational feasibility under the current operating conditions.

[0066] S34: For each control action suggestion in the set of control suggestions to be fused, the dynamic fusion weight of each suggestion under the current working condition is calculated and weighted summed to generate an initial fusion control command, using the corresponding operating state feature and the preset target weight as input.

[0067] In this embodiment, for each control action suggestion in the set of control suggestions to be fused formed in step S33, the dynamic fusion weight of each suggestion under the current operating condition is calculated by taking its corresponding operating state characteristics and preset target weights (e.g., economic weight 0.5, environmental weight 0.3, and flexibility weight 0.2) as input, and then performing a weighted summation to generate an initial fusion control command.

[0068] Specifically, an attention-based fusion method is adopted. An offline-trained feedforward neural network is constructed as the attention network, whose inputs are the running state features, preset target weights, and the set of suggestions to be fused. This network is trained offline using historical best decision data and can automatically calculate the attention score (i.e., fusion weight) of each suggestion based on the features of the current running state and the preference of the target weights using the aforementioned Softmax function.

[0069] For example, when the operating status characteristics indicate that the current condition is under stringent environmental protection requirements, and the environmental protection weight is high in the target weights, control recommendations that have shown lower NOx emission potential in historical assessments or predictions will receive higher fusion weights. After calculating the fusion weights of each recommendation, these weights are multiplied by their corresponding control recommendation vectors and then summed to obtain the initial fused control command. This method allows the fusion process to adaptively prioritize control strategies that are more advantageous for achieving multiple objectives under the current state.

[0070] S35: Perform boundary normalization on the parameter components of the initial fusion control command, and output the final control command that satisfies all physical and safety constraints.

[0071] In this embodiment, the parameter components of the initial fusion control command generated in step S34 are bounded and normalized to output the final control command that satisfies all physical and safety constraints. The initial fusion command is mathematically weighted optimal, but may not perfectly conform to certain rigid physical laws or engineering constraints. Therefore, normalization and correction are necessary.

[0072] This step mainly addresses two types of issues: first, the coupling constraints between parameters. For example, even if the air volume and coal volume in the initial command are both within the boundaries, their ratio may be lower than the minimum air-to-coal ratio required to prevent slagging. In this case, it is necessary to increase the air volume or decrease the coal volume proportionally to satisfy this coupling constraint.

[0073] Secondly, there are constraints on the smoothness of dynamic processes. For example, control commands for high-inertia components such as water flow rate need to be smoothed to avoid abrupt changes and ensure that the rate of change matches the equipment's capacity. The regulation process can be implemented based on a set of predefined rules or a lightweight optimizer. Its goal is to minimize the adjustment range while preserving the initial command optimization intent as much as possible, so that all parameter components and their interrelationships strictly satisfy the dynamic boundary constraints obtained in step S32 as well as deeper physical laws.

[0074] The regulated instructions are the final control instructions that can be directly issued and executed.

[0075] S4: Send the final control command to the actuator corresponding to the high-alkali coal boiler, perform multi-parameter coordinated control, and collect the actual system status after executing the final control command.

[0076] Specifically, step S4 includes the following steps: S41: Based on the operating state characteristics, the final control command is decomposed into a multi-time-scale decomposition, which is divided into a set of commands including fast-changing loop control commands and slow-changing loop control commands.

[0077] In this embodiment, during specific implementation, it is first necessary to classify the boiler's subprocesses according to their dynamic response characteristics (i.e., time constants). Generally, loops with faster responses (time constants on the order of seconds), such as the combustion system and flue gas system, are classified as fast-changing loops, and their control commands include coal feed commands, secondary air damper opening commands for each layer, and burnout damper opening commands, etc.; while loops with slower responses (time constants on the order of minutes or even longer), such as the steam-water system, are classified as slow-changing loops, and their control commands mainly include feedwater flow commands, desuperheating water regulating valve commands, etc.

[0078] The decomposition process is accomplished by a rule-based or lightweight model classifier. This classifier takes operating state features as input and helps determine the dynamic characteristics of each parameter loop under the current operating condition, thereby accurately splitting the final control command vector into fast-changing loop control command subsets and slow-changing loop control command subsets. This step ensures that subsequent control commands with different response speeds can be issued to the corresponding actuators at appropriate cycles and rhythms, forming the foundation for coordinated control.

[0079] S42: Before sending the set of instructions to each actuator, adaptive dead-zone compensation correction is performed on the control instructions based on the real-time status of each actuator and the preset dead-zone characteristics, and the corrected control instructions are generated.

[0080] In this embodiment, the specific implementation process is as follows: a dead zone model is established for each key actuator (such as damper actuator, coal feeder frequency converter, and feedwater pump speed control device). This model describes the nonlinear relationship between the change in command and the actual change in displacement or speed of the mechanism, especially the characteristic of having an insensitive zone (dead zone).

[0081] In online applications, the actual feedback position of each actuator is monitored in real time and compared with the expected position of the currently received command. When the change in command is small and within the dead zone of the actuator, the compensation algorithm automatically calculates a compensated command that can overcome the dead zone and enable the actuator to produce effective action, based on a preset dead zone value. For example, if a damper actuator has a 2% dead zone, and the current command requires it to increase from 50% to 51%, the compensation algorithm will output a temporary command that causes the opening change to exceed 2% (such as increasing it to 52.5%) to ensure that the actuator actually moves, and then return to the target command. This correction process significantly improves the accuracy of control and avoids control lag or errors caused by actuator dead zones.

[0082] S43: The corrected control command is synchronously sent to the corresponding actuators. At the same time, a synchronous execution signal is sent to each actuator, so that the fuel supply mechanism, air supply mechanism, induced draft mechanism and water supply mechanism can perform coordinated actions according to the timing logic of the corrected control command to achieve multi-parameter coordinated control.

[0083] In this embodiment, the corrected control command generated in S42 is synchronously sent to the corresponding actuators, and by sending a synchronous execution signal, the actuators of the fuel, air supply, induced draft, and water supply systems perform coordinated actions according to the timing logic inherent in the command, thereby achieving true multi-parameter coordinated control.

[0084] Specifically, this is achieved through a coordination and control unit deployed in the control system. This unit first sends the corrected fast-change loop command and slow-change loop command to the corresponding lower-level controllers (such as the burner management system, fan control system, and water supply control system) through the control network.

[0085] The key point is that it's not simply issued simultaneously, but rather accompanied by a synchronization timing signal. For example, under a load increase command, the coordination control unit ensures that the command to increase fuel quantity and the command to increase air volume are precisely matched in time, usually by "increasing air first, then coal" or by increasing them synchronously in a certain proportion, to avoid incomplete combustion or drastic fluctuations in furnace pressure.

[0086] For slow-speed loops like feedwater, the start time and rate of change of commands are synchronized with the trend of heat load changes. All actuators begin to operate after receiving commands and synchronization signals, thereby ensuring that the boiler smoothly and safely transitions from one operating state to another target state.

[0087] S44: During the operation of each actuator, process variables of the high-alkali coal boiler are collected in real time. The process variables include furnace temperature distribution, flue gas oxygen content, main steam parameters, and actual feedback position signals of each actuator, which constitute the execution process status data.

[0088] In this embodiment, during the operation of each actuator in S43, key process variables of the high-alkali coal boiler and actual feedback signals of each actuator are collected in real time to form execution process status data.

[0089] Specifically, this is achieved through the data acquisition module of the boiler's distributed control system (DCS). The acquired process variables mainly include: furnace temperature field distribution reflecting combustion status (via multi-layer thermocouples), flue gas oxygen content and CO concentration, and exhaust gas temperature; main steam pressure and temperature, drum water level (for subcritical units) or water-cooled wall midpoint temperature / start-up separator water level (for supercritical units), and wall temperatures of each stage of superheater / reheater reflecting the status of the steam-water system; and parameters such as current and vibration reflecting the status of auxiliary equipment such as fans and pumps.

[0090] Simultaneously, the actuator feedback signals include the actual opening degree of each damper, the actual speed of the coal feeder, and the valve position feedback of the regulating valve. All these data are collected synchronously at a high frequency (e.g., once per second) and stamped with a unified timestamp, forming a complete and synchronous time-series dataset, i.e., execution process state data. This data serves as the direct basis for evaluating control effectiveness, determining whether the system has reached a new steady state, and for subsequent model adaptation.

[0091] S45: When the execution process state data reaches a quasi-steady state, extract the data within a preset effective observation window from the execution process state data as the actual system state.

[0092] In this embodiment, when the execution process status data collected in S44 reaches a quasi-steady state, data within a preset effective observation window is extracted from this data as the actual system state for subsequent feedback optimization. The quasi-steady state judgment criterion is: within a set time window (e.g., 2-3 minutes), the fluctuation amplitude of all key controlled parameters (such as main steam pressure, furnace negative pressure, drum water level (for subcritical units), water-cooled wall midpoint temperature, or slightly superheated steam enthalpy (for supercritical units)) is less than their respective set thresholds, and their changing trends have tended to stabilize. The system has a built-in quasi-steady state judgment logic module, which continuously monitors the real-time changes of these key parameters.

[0093] Once the system is determined to have entered a quasi-steady state, this module triggers a data extraction operation. The extracted data is the average value or a filtered representative value of all process variables and actuator feedback data within a preset effective observation window (e.g., 5 minutes) before and after entering the quasi-steady state. This extracted data eliminates interference from dynamic transient processes and stably reflects the new stable operating state reached by the boiler system after executing the final control command, i.e., the actual system state. This data will be used to compare with the previous predicted state and is key feedback information for driving the online adaptive adaptation of the controller parameters.

[0094] S5: Based on the error between the actual system state and the predicted state corresponding to the final control command, the encoder parameters for extracting the operating state features are adjusted online.

[0095] Specifically, step S5 includes the following steps: S51: Based on the actual system state and the predicted state made according to the operating state characteristics and the final control command, calculate the model prediction error; and based on the model prediction error and combined with the historical distribution characteristics of the operating state characteristics, construct a comprehensive loss for encoder parameter optimization.

[0096] In this embodiment, the actual system state obtained in step S4 is compared with the predicted state made based on the operating state characteristics in step S1 and the final control command in step S3. The model prediction error is calculated, and a comprehensive loss function for encoder parameter optimization is constructed.

[0097] In specific implementation, the prediction model (simplified mechanism model or data-driven model) used in step S22 is first invoked. The operating state characteristics and the final control command corresponding to the generation of the final control command are input, and the corresponding multi-step predicted state sequence is calculated. The steady-state value (or the predicted value of the last step) of the predicted state sequence is compared point by point with the key process variables (such as main steam pressure, furnace temperature, flue gas oxygen content, etc.) in the actual system state, and the absolute error or relative error of each variable is calculated to form a multi-dimensional error vector. Subsequently, a comprehensive loss function is constructed. This function not only includes the sum of squares of the above prediction errors (prediction accuracy loss term), but also introduces a regularization term based on the consistency of the distribution of operating state characteristics (distribution consistency loss term).

[0098] The distribution consistency loss term is obtained by calculating the KL divergence or other statistical distance between the current operating state features and the feature distribution under historical normal operating conditions (e.g., modeled using a Gaussian mixture model). Its purpose is to prevent the encoder from overfitting to single prediction errors during optimization, thus deviating from the normal feature distribution space and maintaining the physical rationality of the features. Finally, the comprehensive loss function is a weighted sum of the prediction accuracy loss term and the distribution consistency loss term, and the weighting coefficients can be adjusted according to different emphases on prediction accuracy and feature stability.

[0099] S52: With the goal of minimizing the overall loss, the parameters of the encoder are adjusted incrementally online; and the effect of the re-extracted operating status features after adjustment is verified.

[0100] In this embodiment, with the goal of minimizing the comprehensive loss constructed in step S51, the encoder parameters for extracting running state features are adjusted incrementally online, and the effect of feature extraction after adjustment is verified.

[0101] Specifically, an online gradient descent algorithm is used for parameter updates. First, the gradient of the comprehensive loss function with respect to all trainable parameters of the encoder is calculated. To stabilize the training process, stochastic gradient descent with momentum or an adaptive learning rate algorithm (such as Adam) is typically used. The learning rate is set relatively small (e.g., on the order of 1e-5) to ensure that the parameters are fine-tuned rather than drastically changed. Parameter updates are performed asynchronously in a background thread to avoid affecting the real-time control loop.

[0102] After the parameter update is completed, the effect needs to be verified immediately: using the updated encoder, the actual system state that triggered this adjustment (i.e., the data collected in S44) is re-encoded to generate new operating state characteristics. Then, based on these new characteristics and the previously issued final control commands, the predictive model is used again to predict the state and calculate the new prediction error.

[0103] The validation criterion is that the new prediction error should be significantly lower than the prediction error before adjustment. Simultaneously, it's checked whether the newly generated features still fall within the high-probability region of the historical normal feature distribution to ensure the features are not anomaly-free. If the validation passes, the parameter update is accepted; if the prediction error does not decrease or the feature distribution is abnormal, the update is rejected, and the encoder parameters are rolled back to their original state. This step ensures that the encoder optimization is robust and beneficial.

[0104] S53: Store the verified actual system state, the final control command, and the corresponding optimized operating state characteristics as new samples.

[0105] In this embodiment, the valid experience data verified in step S52 is stored as new samples to support subsequent periodic retraining of the encoder and associated model. Not all experiences are stored; they need to be filtered.

[0106] The stored sample is a complete data tuple, including: 1) The actual system state that triggers the adjustment (i.e., the raw data collected by S44); 2) The latest operating state characteristics generated by the optimized encoder in this state; 3) The final control command executed; 4) The actual system state reached after executing the instruction (for subsequent supervised learning).

[0107] The selection criteria for samples are: the adjustment is verified to be effective (i.e., S52 validation passes), and the reduction in overall loss or prediction error exceeds a certain minimum threshold, ensuring that only experiences that bring significant improvements are stored. During storage, the data undergoes anonymization and normalization. These new samples are stored in a fixed-size experience replay pool, following a first-in, first-out (FIFO) principle. After accumulating a certain number (e.g., several thousand samples), this data will serve as an incremental dataset, used for periodic, small-batch offline retraining of the encoder (and possibly the prediction model, policy network, etc.) during system idle periods, thereby achieving continuous accumulation of knowledge and gradual performance improvement for the entire system. This constitutes the core mechanism of the system's long-term adaptive capability.

[0108] Example 2 This invention also provides a multi-parameter adaptive coordination system for high-alkali coal boilers, used to execute the aforementioned multi-parameter adaptive coordination method for high-alkali coal boilers, with reference to... Figure 2 As shown, the coordination system includes: The feature acquisition module 100 is used to extract operating status features that characterize the current operating conditions based on the real-time operating status data of the high-alkali coal boiler.

[0109] The risk assessment module 200 is used to generate control action suggestions from at least two different sources in parallel based on the operating state characteristics, and simultaneously conduct safety risk assessments on various control action suggestions to obtain assessment results.

[0110] The control command output module 300 is used to obtain boundary constraints based on the operating state characteristics, and at the same time, to filter various control action suggestions based on the evaluation results, and to perform weighted fusion and boundary correction by combining the filtered control action suggestions, the operating state characteristics and the preset target weights, and output the final control command.

[0111] The instruction execution module 400 is used to send the final control instruction to the actuator corresponding to the high-alkali coal boiler, perform multi-parameter coordinated control, and collect the actual system status after executing the final control instruction.

[0112] The parameter feedback module 500 is used to adjust the parameters of the encoder that extracts the operating state features online based on the error between the actual system state and the predicted state corresponding to the final control command.

[0113] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0114] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0115] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A multi-parameter adaptive coordination method for high-alkali coal boilers, characterized in that, The coordination method includes the following steps: Based on real-time operating status data of high-alkali coal boilers, the operating status characteristics of the current operating conditions are extracted and characterized. Based on the aforementioned operational status characteristics, at least two different sources of control action suggestions are generated in parallel, and a safety risk assessment is simultaneously performed on each type of control action suggestion to obtain the assessment results. Based on the operating state characteristics, boundary constraints are obtained. At the same time, based on the evaluation results, various control action suggestions are screened. The screened control action suggestions, the operating state characteristics, and the preset target weights are combined for weighted fusion and boundary correction to output the final control command. The final control command is sent to the actuator corresponding to the high-alkali coal boiler for multi-parameter coordinated control, and the actual system status after the final control command is executed is collected. Based on the error between the actual system state and the predicted state corresponding to the final control command, the encoder parameters for extracting the operating state features are adjusted online.

2. The multi-parameter adaptive coordination method for high-alkali coal boilers according to claim 1, characterized in that, The extraction of operating status characteristics based on real-time operating status data of high-alkali coal boilers to characterize the current operating conditions includes: Real-time acquisition of raw operating status data of high-alkali coal boilers and acquisition of coal quality data characterizing the current alkali metal properties of coal; fusion processing of the raw operating status data and the coal quality data to generate enhanced status data. Calculate the contribution weight of each feature dimension in the enhanced state data to the current combustion condition, and weight the enhanced state data according to the contribution weight to output the weighted state features; The weighted state features are mapped to the running state features by a pre-trained encoder.

3. The multi-parameter adaptive coordination method for high-alkali coal boilers according to claim 2, characterized in that, The system generates control action suggestions from at least two different sources in parallel based on the operational state characteristics, and simultaneously performs a safety risk assessment on each type of control action suggestion to obtain the assessment results, including: Using the operating state characteristics as input, the control strategy trained based on historical best control data is queried and invoked, a first control action suggestion is output, and a first safety risk value corresponding to the first control action suggestion is calculated. Using the aforementioned operating state characteristics as the initial state for rolling optimization, forward rolling calculation and optimization are performed within a preset finite time domain, with the criterion of satisfying the optimal multi-objective performance indicators. The second control action suggestion is obtained, and the second safety risk value corresponding to the second control action suggestion is calculated. The first security risk value and the second security risk value together constitute the evaluation result of various control action suggestions, and are then bound to their respective corresponding control action suggestions for output.

4. The multi-parameter adaptive coordination method for high-alkali coal boilers according to claim 3, characterized in that, The process involves obtaining boundary constraints based on the operational state characteristics, simultaneously filtering various control action suggestions based on the evaluation results, and combining the filtered control action suggestions, the operational state characteristics, and preset target weights for weighted fusion and boundary correction to output the final control command, including: Based on the operating status characteristics, the category to which the current operating condition belongs is identified, and according to the identified operating condition category, the corresponding set of basic boundary constraints is matched and obtained from the pre-built multi-operating condition constraint knowledge base; Based on the aforementioned operating state characteristics and the aforementioned basic boundary constraint set, and combined with real-time device status information, constraint fusion is performed to output boundary constraints; Based on the safety risk values ​​in the assessment results, various control action suggestions are initially screened; based on the boundary constraints, the initially screened control action suggestions are verified for compliance, forming a set of control suggestions to be integrated. For each control action suggestion in the set of control suggestions to be fused, the dynamic fusion weight of each suggestion under the current working condition is calculated and weighted summed, using the corresponding operating state feature and the preset target weight as input, to generate the initial fusion control command; The parameter components of the initial fusion control command are bounded and normalized to output the final control command that satisfies all physical and safety constraints.

5. A multi-parameter adaptive coordination method for high-alkali coal boilers according to claim 4, characterized in that, The process of issuing the final control command to the actuator corresponding to the high-alkali coal boiler, performing multi-parameter coordinated control, and collecting the actual system status after executing the final control command includes: Based on the operating state characteristics, the final control command is decomposed into a multi-time-scale decomposition, which is a set of commands including fast-change loop control commands and slow-change loop control commands. Before the set of instructions is sent to each actuator, the control instructions are adaptively dead-zone compensated and corrected according to the real-time status of each actuator and the preset dead-zone characteristics, and the corrected control instructions are generated. The corrected control command is synchronously sent to the corresponding actuators. At the same time, a synchronous execution signal is sent to each actuator, so that the fuel supply mechanism, air supply mechanism, induced draft mechanism and water supply mechanism can perform coordinated actions according to the timing logic of the corrected control command to achieve multi-parameter coordinated control. During the operation of each actuator, process variables of the high-alkali coal boiler are collected in real time. These process variables include furnace temperature distribution, flue gas oxygen content, main steam parameters, and actual feedback position signals of each actuator, which constitute the execution process status data. When the execution process state data reaches a quasi-steady state, data within a preset effective observation window is extracted from the execution process state data and used as the actual system state.

6. A multi-parameter adaptive coordination method for high-alkali coal boilers according to claim 5, characterized in that, The online parameter adjustment of the encoder that extracts the operating state features based on the error between the actual system state and the predicted state corresponding to the final control command includes: Based on the actual system state and the predicted state made according to the operating state characteristics and the final control command, the model prediction error is calculated; and based on the model prediction error and combined with the historical distribution characteristics of the operating state characteristics, a comprehensive loss for encoder parameter optimization is constructed. With the goal of minimizing the overall loss, the parameters of the encoder are adjusted incrementally online; and the effect of the re-extracted operating state features after adjustment is verified. The verified actual system state, the final control command, and the corresponding optimized operating state characteristics will be stored as new samples.

7. A multi-parameter adaptive coordination system for high-alkali coal boilers, used to execute the multi-parameter adaptive coordination method for high-alkali coal boilers as described in any one of claims 1 to 6, characterized in that, The coordination system includes: The feature acquisition module is used to extract operating status features that characterize the current operating conditions based on the real-time operating status data of high-alkali coal boilers. The risk assessment module is used to generate control action suggestions from at least two different sources in parallel based on the operating status characteristics, and simultaneously conduct safety risk assessments on various control action suggestions to obtain assessment results; The control command output module is used to obtain boundary constraints based on the operating state characteristics, and at the same time, to filter various control action suggestions based on the evaluation results, and to perform weighted fusion and boundary correction by combining the filtered control action suggestions, the operating state characteristics and the preset target weights, and output the final control command. The instruction execution module is used to send the final control instruction to the corresponding actuator of the high-alkali coal boiler, perform multi-parameter coordinated control, and collect the actual system status after the final control instruction is executed. The parameter feedback module is used to adjust the parameters of the encoder that extracts the operating state features online based on the error between the actual system state and the predicted state corresponding to the final control command.

Images