Multi-objective combustion control method for coal-fired boiler based on VMD-multi-core LSSVM and GRU-DDPG

Abstract

This invention relates to the field of intelligent control technology for coal-fired boilers, specifically to a multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG. This invention employs an adaptive variational mode decomposition algorithm based on energy contribution thresholds to decompose the original combustion signal into several intrinsic mode functions, constructing a multi-core least squares support vector machine prediction model. The kernel weight parameters are optimized using a genetic algorithm, and the combustion control process is modeled as a continuous Markov decision process. A gated recurrent unit layer is integrated into the actuator and evaluator network to capture time-dependent features, constructing a weighted multi-objective reward function to form a control framework. The prediction model is then updated in real-time using an accumulation and change detection algorithm combined with a sliding window incremental learning mechanism, forming a complete closed-loop control. This achieves high-precision prediction and multi-objective collaborative optimization control of the combustion process under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for coal-fired boilers, specifically a multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG. Background Technology

[0002] Coal-fired power generation accounts for approximately 36% of global electricity production. Optimizing its combustion process is crucial for sustainable economic and environmental development. Modern coal-fired boilers face the dual pressures of increasingly stringent environmental regulations and competitive power generation costs. The inherent nonlinearity, strong coupling of variables, significant time delay, and complex process noise of the combustion process make it difficult for traditional control methods to effectively address these challenges. The core challenge lies in simultaneously optimizing multiple conflicting objectives: improving thermal efficiency to reduce fuel consumption while reducing nitrogen oxide emissions to meet environmental standards. Combustion signals have non-steady-state characteristics, and operating conditions continuously evolve with load changes, fuel quality fluctuations, and equipment aging, further exacerbating the complexity of multi-objective optimization. Traditional proportional-integral-derivative control systems are difficult to adapt to dynamic changes, with efficiency of only 88%-90% under variable load conditions and nitrogen oxide emissions exceeding 130 mg / m³. In existing technologies, although variational mode decomposition can adaptively decompose non-stationary signals, the selection of the number of modes and the interpretation of the physical meaning of the decomposed components still need to be deepened. Single-core support vector machines cannot simultaneously take into account the local and global features of the combustion process, which limits the ability to represent multi-scale characteristics. Deep deterministic policy gradient algorithms provide an effective framework for continuous action space optimization, but they are difficult to fully handle the time-dependent features in combustion dynamics, and the prediction model lacks online adaptive update capability, which can easily lead to a decrease in control performance due to model drift.

[0003] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention provides a multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG, which can effectively solve the problems of low prediction accuracy, poor signal processing effect, insufficient multi-objective optimization capability and poor real-time performance in existing technologies.

[0005] To achieve the above objectives, the present invention can be implemented through the following technical solutions: This invention provides a multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG, comprising the following steps: An adaptive variational mode decomposition algorithm based on energy contribution threshold is adopted and optimized by alternating direction multiplier method to decompose the original combustion signal into several intrinsic mode functions and determine the optimal number of modes as the target mode components for subsequent prediction. A multi-kernel least squares support vector machine prediction model integrating radial basis function kernel, polynomial kernel, and sigmoid kernel is constructed. The target modal components are used as the input data of the prediction model, and the weight parameters corresponding to the multi-kernel kernel function are optimized by genetic algorithm to complete the training and optimization of the multi-kernel least squares support vector machine prediction model. The combustion control process of a coal-fired boiler is modeled as a continuous Markov decision process. Gated recurrent unit layers are integrated into both the actuator network and the evaluator network to capture the temporal dependency characteristics of the combustion process. At the same time, a reward function adapted to the multi-objective control requirements is constructed to form the GRU-DDPG control framework. By combining the cumulative and change detection algorithm with the sliding window incremental learning mechanism, the trained multi-core least squares support vector machine prediction model is updated in real time. The updated prediction results from the multi-core least squares support vector machine prediction model are used as the state input of the GRU-DDPG controller, which then outputs the corresponding optimal combustion control command.

[0006] Furthermore, the specific process of the adaptive variational mode decomposition algorithm based on the energy contribution threshold is as follows: A VMD-constrained variational decomposition model is constructed, and Lagrange multipliers are introduced. At the same time, an extended Lagrange expression is constructed by combining a quadratic penalty term. The original combustion signal decomposition and solution is transformed into a constrained variational optimization problem. The alternating direction multiplier method is used to iteratively update each modal component and its corresponding center frequency in the frequency domain. The modal components are updated through frequency domain iterative formulas, and the center frequency is determined based on the centroid calculation of the modal power spectrum. Set an iteration convergence threshold. When the deviation of the modal components before and after the iteration meets the convergence condition, the iteration is terminated. The number of iterations for combustion signal decomposition does not exceed 235, and the average number of iterations is controlled within the range of 235±47 to achieve fast convergence.

[0007] Furthermore, the specific process for determining the optimal number of modes is as follows: The cumulative energy contribution criterion is used to adaptively select the number of modes. By calculating the energy proportion of each mode component, the component with a cumulative energy contribution exceeding 95% of the total energy of the original signal is retained, and the optimal number of decomposed modes is automatically determined.

[0008] Furthermore, the specific process of constructing a multi-kernel least squares support vector machine prediction model is as follows: We construct a combined kernel function that adapts to the characteristics of boiler combustion process by weighted linear combination of three complementary kernel functions: radial basis function kernel, polynomial kernel, and sigmoid kernel. With the goal of minimizing structural risk, we construct a multi-kernel least squares support vector machine optimization problem. We set a regularization parameter of 10.0 to balance model complexity and fitting accuracy. We introduce kernel function techniques to transform the optimization problem into a matrix solution form to complete the solution of the basic parameters of the prediction model.

[0009] Furthermore, the specific process of optimizing the weight parameters of the multi-kernel function using a genetic algorithm is as follows: A genetic algorithm combined with 5-fold cross-validation was used to optimize the weight parameters of the three kernel functions globally, with the goal of minimizing the prediction error. The optimization parameters of the genetic algorithm were set as follows: population size 50, crossover rate 0.8, and mutation rate 0.1. The algorithm converged within 50 generations, and the optimal kernel weight parameters were obtained through optimization.

[0010] Furthermore, the specific process of modeling the combustion control process of a coal-fired boiler as a continuous Markov decision process is as follows: Construct a quintuple model (S, A, P, R, γ), where S is the state space, A is the action space, P is the transition dynamics, R is the reward function, and γ is the discount factor. The state vectors in the state space specifically include: load, steam flow rate, steam pressure, steam temperature, coal feed rate A, coal feed rate B, coal feed rate C, coal feed rate D, primary air, secondary air, furnace temperature, flue gas temperature, efficiency, and nitrogen oxides. Each state vector is normalized to the historical operating boundary. ; The motion vector in the motion space is a continuous control increment, and the motion value is limited to ±10% of the nominal value; The transfer dynamics employ a hybrid model combining mechanistic and data analysis, and the boiler efficiency satisfies: ,in, The load baseline efficiency curve To learn the residuals, the formation of nitrogen oxides (NOx) follows an extended Zeldovitch mechanism.

[0011] Furthermore, the specific process for capturing the time-dependent features during combustion is as follows: Gated recurrent unit layers are integrated into both the executor network and the evaluator network of the deep deterministic policy gradient framework; The actuator network takes a state sequence of length 20 as input, which is processed by a gated recurrent unit layer and multiple fully connected layers, and then outputs the control action by an activation function. The gated recurrent unit layer processes the time sequence through a gating mechanism. The evaluator network first encodes the state sequence through a gated recurrent unit layer, then concatenates it with the action vector, and outputs a scalar Q value through a fully connected layer to evaluate the state-action value. The network parameters are initialized using Xavier and layer normalization is added. The target network updates its parameters through a soft update mechanism.

[0012] Furthermore, the expression for constructing the reward function is as follows:

[0013] in, Efficiency reward weighting, Weighting of emissions penalties As a stability reward weight, For safety constraint weights, , , and These are efficiency rewards, emissions penalties, stability rewards, and safety constraints.

[0014] Furthermore, the specific process of real-time updating the trained multi-core least squares support vector machine prediction model is as follows: The cumulative sum statistic is used to monitor the prediction residuals of the multi-kernel least squares support vector machine prediction model in real time. The formula for calculating the cumulative sum statistic is:

[0015] in, The prediction error for time t, To preset the mean of the expected error, For relaxation parameters, when When the threshold is exceeded, the prediction model is determined to have drifted significantly and an online update is triggered. Once the update is triggered, the sliding window incremental learning mechanism is immediately activated, and during the update process, only the support vectors corresponding to the newly added samples are updated using Lagrange multipliers.

[0016] The technical solution provided by this invention has the following advantages compared with the known prior art: 1. This invention employs an adaptive variational mode decomposition algorithm based on energy contribution thresholds. By automatically determining the optimal number of modes through the cumulative energy contribution criterion, the non-stationary combustion signal is decomposed into intrinsic mode functions with clear physical meaning. Based on this, a multi-kernel least squares support vector machine prediction model integrating radial basis function kernels, polynomial kernels, and sigmoid kernels is constructed. The kernel weight parameters are globally optimized through a genetic algorithm, giving full play to the complementary advantages of the three kernel functions in capturing local nonlinearities, fitting global trends, and representing operating conditions. This significantly improves the prediction accuracy of key combustion parameters and provides a reliable state estimation basis for subsequent control strategies.

[0017] 2. This invention models the combustion control process of a coal-fired boiler as a continuous Markov decision process. Gated recurrent unit layers are integrated into both the actuator and evaluator networks. Using a state sequence of length 20 as input, the temporal dependency characteristics of the combustion process are effectively captured. A weighted multi-objective reward function is constructed to comprehensively balance competing control objectives such as efficiency improvement, emission reduction, operational stability, and safety constraints. Under the premise of maintaining stable operation, the boiler thermal efficiency is improved by 3.13%, while nitrogen oxide emissions are reduced by 33.9 mg / m³, achieving synergistic optimization of efficiency and emissions.

[0018] 3. This invention employs a cumulative and change detection algorithm combined with a sliding window incremental learning mechanism. By monitoring the prediction residual in real time, it achieves rapid detection and triggering of model drift updates. The sliding window incremental learning only updates the support vectors corresponding to newly added samples using Lagrange multipliers, with a single model update time of less than 50 milliseconds. While ensuring prediction accuracy, it significantly reduces computational overhead and achieves long-term adaptive capability of the prediction model to fuel quality fluctuations, equipment performance aging, and changes in operating conditions, meeting the stringent requirements of real-time performance and reliability in industrial settings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is the overall architecture of the multi-objective combustion control process of the present invention; Figure 2 A comparison chart of prediction performance for multi-core LSSVM; Figure 3 This is a schematic diagram of the VMD decomposition results of the boiler combustion signal; Figure 4 This is a schematic diagram of the GRU-DDPG network structure; Figure 5 Train the convergence curve for the controller. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, the multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG includes the following steps: Step 1: Adaptive VMD Signal Decomposition: Addressing the characteristics of boiler combustion signals—strong non-stationarity, multi-frequency coupling, and complex noise—an adaptive variational mode decomposition (VMD) algorithm based on energy contribution thresholds is employed. Through Alternating Direction Multiplier Method (ADMM) optimization, the original combustion signal is decomposed into several intrinsic mode functions (IMFs), and the optimal number of modes is determined as the target mode components for subsequent prediction. The specific implementation process is as follows: A VMD-constrained variational model is constructed and Lagrange multipliers are introduced. The modal components and center frequency are iteratively updated in the frequency domain, specifically as follows: Let the input boiler combustion original signal be... The core objective of the VMD algorithm is to... Decomposed into Modal components and its corresponding center frequency The constrained variational optimization problem is constructed as follows:

[0023] Constraints:

[0024] in, Let be the Dirac distribution function. For convolution operations, For time Partial derivative operations; To achieve constrained optimization, a quadratic penalty term is introduced. With Lagrange multipliers Construct the extended Lagrange expression:

[0025] The Alternating Directional Multiplier Method (ADMM) is used to iteratively update each modal component and the center frequency in the frequency domain to complete the optimal decomposition solution; Specifically: Modal component update: 1st In the next iteration, the frequency domain expression of the modal components is:

[0026] Center frequency update: Based on the calculation of the centroid of the modal power spectrum, the expression is:

[0027] Set the iteration convergence threshold The iteration terminates when the following equation is satisfied:

[0028] In this invention, the number of decomposition iterations for all combustion signals does not exceed 235, and the average number of iterations is 235±47, which can quickly achieve convergence. The cumulative energy contribution criterion is used to achieve adaptive selection of the number of modes: by calculating the energy proportion of each IMF component, the mode components whose cumulative energy contribution exceeds 95% of the total signal energy are retained, and the optimal number of modes K is automatically determined.

[0029] Experiments have verified the optimal decomposition mode number of the boiler combustion signal. The specific physical meaning of each IMF component is as follows: IMF1 (0.01–0.1 Hz): Characterizes the variation and daily cycle fluctuation of unit load; IMF2 (0.1–0.5 Hz): Characterizes the system's control regulation and medium-term operational fluctuations; IMF3 (0.5–2 Hz): Characterizes combustion instability and pressure oscillations in the furnace; IMF4 (2–10 Hz): Characterizes measurement interference and flue gas turbulence fluctuations; IMF5 (>10Hz): Characterizes high-frequency environmental noise; like Figure 3 As shown, the energy distribution results indicate that IMF1-IMF3 contribute a total of 97.8% of the total signal energy, verifying the rationality and effectiveness of the adaptive mode selection threshold of the present invention. Finally, effective mode components are retained for subsequent prediction of target mode components.

[0030] Step 2: Multi-kernel Least Squares Support Vector Machine (LSSVM) Prediction Modeling: Construct a multi-kernel least squares support vector machine (LSSVM) prediction model that integrates radial basis function (RBF) kernels, multinomial kernels, and sigmoid kernels. Use the target modal components as input data for this prediction model, and optimize the weight parameters corresponding to the multi-kernel kernel function using a genetic algorithm. This completes the training and optimization of the multi-kernel least squares support vector machine (LSSVM) prediction model. The specific implementation process is as follows: By weighted linearly combining three kernel functions with complementary properties, a combined kernel function suitable for the combustion process is constructed, expressed as:

[0031] The kernel functions in the formula are defined as follows: RBF kernel: used to accurately capture local nonlinear characteristics of the combustion process, its expression is:

[0032] Polynomial kernel: Used to fit the global efficiency trend of the boiler to its load-related characteristics, the expression is:

[0033] The Sigmoid kernel is used to characterize the boiler's start-up, shutdown, and load change behavior. Its expression is:

[0034] With the goal of minimizing structural risk, a multi-core LSSVM optimization problem is constructed, expressed as:

[0035] Constraints:

[0036] in, This represents the implicit feature mapping corresponding to the combined kernel function. This is the regularization parameter, with a value of 10.0, used to balance model complexity and fitting accuracy. This represents the fitting error; By introducing kernel function techniques, the above optimization problem can be transformed into a matrix solution form:

[0037] in, Complete the parameter solution of the multi-core least squares support vector machine (LSSVM) prediction model; A genetic algorithm combined with 5-fold cross-validation is used to minimize the prediction error, adjusting the kernel weight parameters. , and Perform global optimization; Optimized parameter settings: population size 50, crossover rate 0.8, mutation rate 0.1, the algorithm converges within 50 generations; The optimal kernel weights obtained after optimization are: , , ; The target modal components obtained from VMD decomposition are used as input data for a multi-kernel least squares support vector machine (LSSVM) prediction model to achieve thermal efficiency and NO0. x High-precision prediction output of key boiler parameters such as emissions.

[0038] Step 3: GRU-DDPG Reinforcement Learning Control: The combustion control process of the coal-fired boiler is modeled as a continuous Markov decision process. Gated recurrent unit (GRU) layers are integrated into both the actuator and criterion networks to capture the temporal dependency features in the combustion process. Simultaneously, a reward function adapted to multi-objective control requirements is constructed, forming the GRU-DDPG control framework. The specific implementation process is as follows: The boiler combustion control problem is modeled as a continuous Markov decision process (MDP), and a five-tuple model (S, A, P, R, γ) is constructed, where S is the state space, A is the action space, P is the transition dynamics, R is the reward function, and γ is the discount factor. It should be noted that the state space is defined as: a state vector. This comprehensively characterizes the boiler's operating status, specifically including: load, steam flow rate, steam pressure, steam temperature, coal feed rate A, coal feed rate B, coal feed rate C, coal feed rate D, primary air, secondary air, furnace temperature, flue gas temperature, efficiency, and nitrogen oxides. Each state vector is normalized to historical operating boundaries. The normalization formula is: ; Action vectors For continuous control increment, the expression is: The action value is limited to ±10% of the nominal value. The calculation formula for the actual control input is: , where σ represents the scaling factor calibrated according to the physical actuator limits; The transfer dynamics employ a hybrid model combining mechanistic and data analysis, and the boiler efficiency satisfies: ,in, The load baseline efficiency curve To learn the residuals, the formation of nitrogen oxides (NOx) follows an extended Zeldowicz mechanism, expressed as: ; To address the inherent time-dependent characteristics of the combustion process, gated recurrent unit (GRU) layers are integrated into both the Actor and Critic networks of the Deep Deterministic Policy Gradient Framework (DDPG) to capture these time-dependent features. The Actor network takes a state sequence of length 20 as input, which is processed by a GRU layer with 128 hidden units and multiple fully connected layers. The Tanh activation then outputs a control action. The GRU layer processes the time sequence through a gating mechanism, specifically expressed as follows:

[0039] To update the door, To reset the door, In the candidate hidden state, This is the final hidden state; The Critic network first encodes the state sequence through a GRU layer, then concatenates it with the action vector, and outputs a scalar Q value through a fully connected layer to evaluate the state-action value. The network parameters are initialized with Xavier and layer normalization is added. The target network μ' and Q' are maintained through a soft update mechanism: θ'← τθ+(1-τ )θ', where the soft update coefficient τ=0.01; A weighted multi-objective reward function is used to achieve coordinated optimization of efficiency, emissions, stability, and safety constraints. The reward function expression is as follows:

[0040] Among them, efficiency reward weight The efficiency reward expression is: , Represents expected efficiency. Used to adjust the penalty coefficient; Emissions penalty weighting The expression for emissions penalties is:

[0041] in, , ; Stability reward weight The stability reward expression is: ,in, ; Safety constraint weights The safety constraint expression is: ; The furnace temperature is constrained to [1200, 1600]℃, and the steam pressure is constrained to [20, 28]MPa; The training process employs sequence experience replay technology, with the replay buffer storing sequence transition samples. : ; Each iteration samples 32 sequence samples to update the network. The loss function of the Critic network is... Among them, the target value ; The loss function for the Actor network is: ; The Adam optimizer is used to update parameters, and the Actor network learning rate is [missing information]. The learning rate of the Critic network is During training, Ornstein-Uhlenbeck noise with θ=0.15 and σ=0.2 is introduced to enhance the exploration capability. The noise gradually decays over 1000 training cycles and eventually converges to obtain the optimal control strategy.

[0042] Step 4: Online Adaptive Update: The trained multi-kernel least squares support vector machine prediction model is updated in real time using the cumulative sum (CUSUM) change detection algorithm combined with a sliding window incremental learning mechanism. The specific implementation process is as follows: The CUSUM statistic is used to monitor the prediction residuals of the multi-kernel LSSVM prediction model in real time. The formula for calculating the CUSUM statistic is:

[0043] in, The prediction error for time t, To preset the mean of the expected error, For relaxation parameters, when Exceeding the set threshold When the prediction model has significantly drifted, an online update is triggered. Once the update is triggered, the sliding window incremental learning mechanism starts immediately. The sliding window is defined as follows:

[0044] With a window length W = 100 samples, during the update process, only the support vectors corresponding to newly added samples are updated using Lagrange multipliers. The update formula is as follows:

[0045] While ensuring the prediction accuracy of the prediction model, the computational overhead is significantly reduced. The single update time of the prediction model is less than 50 milliseconds, which can fully meet the real-time control requirements of industrial sites. This online adaptive mechanism has a strong adaptability to changes in boiler operating modes. The false alarm rate of prediction model drift detection is less than 2%. When major process changes occur in boiler operating conditions, the average detection delay is only 3.2 sampling cycles, which can quickly track changes in combustion process characteristics.

[0046] Step 5, Closed-Loop Control: The prediction results output by the updated multi-core least squares support vector machine prediction model are used as the state input of the GRU-DDPG controller. The GRU-DDPG controller outputs the corresponding optimal combustion control command, and the response time of the overall control loop is less than 1 second. The specific implementation process is as follows: The updated multi-core LSSVM prediction model outputs key boiler parameters as the state input to the GRU-DDPG controller. Based on the current operating state and multi-objective optimization requirements, the GRU-DDPG reinforcement learning controller calculates and outputs the optimal combustion control command in real time. This optimal combustion control command is directly applied to the coal feeding mechanism, air distribution mechanism, and furnace pressure regulating device, forming a complete real-time closed-loop control process of signal decomposition, state prediction, optimization decision-making, execution control, and error feedback. During closed-loop operation, the system uses an online adaptive update module to synchronously and dynamically calibrate the prediction model and control strategy, continuously offsetting model drift caused by fuel quality fluctuations, equipment performance aging, and changes in operating conditions. This ensures long-term stable control accuracy, with a total control loop response time of less than 1 second. It can quickly track unit load changes and effectively suppress operating parameter disturbances, meeting the real-time, stability, and high-efficiency control application requirements of coal-fired boiler industrial sites.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG, characterized in that, Includes the following steps: An adaptive variational mode decomposition algorithm based on energy contribution threshold is adopted and optimized by alternating direction multiplier method to decompose the original combustion signal into several intrinsic mode functions and determine the optimal number of modes as the target mode components for subsequent prediction. A multi-kernel least squares support vector machine prediction model integrating radial basis function kernel, polynomial kernel, and sigmoid kernel is constructed. The target modal components are used as the input data of the prediction model, and the weight parameters corresponding to the multi-kernel kernel function are optimized by genetic algorithm to complete the training and optimization of the multi-kernel least squares support vector machine prediction model. The combustion control process of a coal-fired boiler is modeled as a continuous Markov decision process. Gated recurrent unit layers are integrated into both the actuator network and the evaluator network to capture the temporal dependency characteristics of the combustion process. At the same time, a reward function adapted to the multi-objective control requirements is constructed to form the GRU-DDPG control framework. By combining the cumulative and change detection algorithm with the sliding window incremental learning mechanism, the trained multi-core least squares support vector machine prediction model is updated in real time. The updated prediction results from the multi-core least squares support vector machine prediction model are used as the state input of the GRU-DDPG controller, which then outputs the corresponding optimal combustion control command.

2. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG as described in claim 1, characterized in that, The specific process of the adaptive variational mode decomposition algorithm based on energy contribution threshold is as follows: A VMD-constrained variational decomposition model is constructed, and Lagrange multipliers are introduced. At the same time, an extended Lagrange expression is constructed by combining a quadratic penalty term. The original combustion signal decomposition and solution is transformed into a constrained variational optimization problem. The alternating direction multiplier method is used to iteratively update each modal component and its corresponding center frequency in the frequency domain. The modal components are updated through frequency domain iterative formulas, and the center frequency is determined based on the centroid calculation of the modal power spectrum. Set an iteration convergence threshold. When the deviation of the modal components before and after the iteration meets the convergence condition, the iteration is terminated. The number of iterations for combustion signal decomposition does not exceed 235, and the average number of iterations is controlled within the range of 235±47 to achieve fast convergence.

3. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG as described in claim 1, characterized in that, The specific process for determining the optimal number of modes is as follows: The cumulative energy contribution criterion is used to adaptively select the number of modes. By calculating the energy proportion of each mode component, the component with a cumulative energy contribution exceeding 95% of the total energy of the original signal is retained, and the optimal number of decomposed modes is automatically determined.

4. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG as described in claim 1, characterized in that, The specific process of constructing a multi-kernel least squares support vector machine prediction model is as follows: We construct a combined kernel function that adapts to the characteristics of boiler combustion process by weighted linear combination of three complementary kernel functions: radial basis function kernel, polynomial kernel, and sigmoid kernel. With the goal of minimizing structural risk, we construct a multi-kernel least squares support vector machine optimization problem. We set a regularization parameter of 10.0 to balance model complexity and fitting accuracy. We introduce kernel function techniques to transform the optimization problem into a matrix solution form to complete the solution of the basic parameters of the prediction model.

5. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG as described in claim 1, characterized in that, The specific process of optimizing the weight parameters of a multi-kernel kernel function using a genetic algorithm is as follows: A genetic algorithm combined with 5-fold cross-validation was used to optimize the weight parameters of the three kernel functions globally, with the goal of minimizing the prediction error. The optimization parameters of the genetic algorithm were set as follows: population size 50, crossover rate 0.8, and mutation rate 0.

1. The algorithm converged within 50 generations, and the optimal kernel weight parameters were obtained through optimization.

6. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG as described in claim 1, characterized in that, The specific process of modeling the combustion control process of a coal-fired boiler as a continuous Markov decision process is as follows: Construct a quintuple model (S, A, P, R, γ), where S is the state space, A is the action space, P is the transition dynamics, R is the reward function, and γ is the discount factor. The state vectors in the state space specifically include: load, steam flow rate, steam pressure, steam temperature, coal feed rate A, coal feed rate B, coal feed rate C, coal feed rate D, primary air, secondary air, furnace temperature, flue gas temperature, efficiency, and nitrogen oxides. Each state vector is normalized to the historical operating boundary. ; The motion vector in the motion space is a continuous control increment, and the motion value is limited to ±10% of the nominal value; The transfer dynamics employ a hybrid model combining mechanistic and data analysis, and the boiler efficiency satisfies: ,in, The load baseline efficiency curve To learn the residuals, the formation of nitrogen oxides (NOx) follows an extended Zeldovitch mechanism.

7. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG according to claim 1, characterized in that, The specific process for capturing time-dependent features during combustion is as follows: Gated recurrent unit layers are integrated into both the executor network and the evaluator network of the deep deterministic policy gradient framework; The actuator network takes a state sequence of length 20 as input, which is processed by a gated recurrent unit layer and multiple fully connected layers, and then outputs the control action by an activation function. The gated recurrent unit layer processes the time sequence through a gating mechanism. The evaluator network first encodes the state sequence through a gated recurrent unit layer, then concatenates it with the action vector, and outputs a scalar Q value through a fully connected layer to evaluate the state-action value. The network parameters are initialized using Xavier and layer normalization is added. The target network updates its parameters through a soft update mechanism.

8. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG according to claim 1, characterized in that, The expression for constructing the reward function is: in, Efficiency reward weighting, Weighting of emissions penalties As a stability reward weight, For safety constraint weights, , , and These are efficiency rewards, emissions penalties, stability rewards, and safety constraints.

9. The multi-objective combustion control method for coal-fired boilers based on VMD-multi-core LSSVM and GRU-DDPG according to claim 1, characterized in that, The specific process of real-time updating the trained multi-core least squares support vector machine prediction model is as follows: The cumulative sum statistic is used to monitor the prediction residuals of the multi-kernel least squares support vector machine prediction model in real time. The formula for calculating the cumulative sum statistic is: in, The prediction error for time t, To preset the mean of the expected error, For relaxation parameters, when When the threshold is exceeded, the prediction model is determined to have drifted significantly and an online update is triggered. Once the update is triggered, the sliding window incremental learning mechanism is immediately activated, and during the update process, only the support vectors corresponding to the newly added samples are updated using Lagrange multipliers.

Images