Boiler combustion optimization control system for thermal power plant

Abstract

The invention relates to a boiler combustion optimization control system of a thermal power plant, relates to the technical field of thermal power generation control, and aims to solve the problems of inaccurate multi-source data perception, slow control response and insufficient multi-target collaborative optimization capability under fire coal quality fluctuation and load change. The system comprises a data perception and fusion module, a working condition self-adaptive identification module, a multi-target dynamic optimization decision module and a distributed execution control module, and through multi-source data acquisition and feature extraction, depth time sequence mode identification, multi-target optimization model solving and distributed coordination control, the multi-target dynamic optimization decision module and the distributed execution control module are subjected to multi-source data fusion. Dynamic balance of boiler heat efficiency improvement and nitrogen oxide emission reduction is achieved, and adaptability and control precision of the system under complex working conditions are enhanced.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation control technology, and in particular to a combustion optimization control system for boilers in thermal power plants. Background Technology

[0002] Thermal power generation is one of the main methods of global energy supply, and the combustion efficiency and pollutant emission control of its core equipment, the boiler, are directly related to the economic efficiency and environmental protection of the power plant. Boiler combustion optimization control aims to achieve high efficiency and cleanliness in the combustion process by adjusting key parameters such as the fuel-air ratio, furnace temperature, and air volume.

[0003] Among them, the combustion optimization control system for boilers in thermal power plants is a key technology for improving boiler operating performance. This system collects real-time data during boiler operation and dynamically adjusts the combustion conditions based on a preset control model or algorithm to achieve the goal of reducing coal consumption and reducing the generation of pollutants such as nitrogen oxides.

[0004] Existing technologies typically employ static models based on historical data or simple PID control strategies, which struggle to accurately adapt to complex operating conditions such as fluctuations in coal quality and load changes. Multi-source data collected by sensors contains noise and latency, leading to inaccurate system perception of the combustion state. Furthermore, traditional control algorithms exhibit slow response and limited optimization effectiveness when handling nonlinear, strongly coupled combustion processes. In addition, existing systems lack the ability to collaboratively optimize multiple objectives (such as efficiency and environmental protection) during combustion, easily resulting in sacrificing boiler efficiency while reducing nitrogen oxide emissions, or control instability during rapid load changes. Therefore, there is an urgent need for an intelligent boiler combustion control system capable of adapting to complex operating conditions and achieving multi-objective dynamic optimization. Summary of the Invention

[0005] The present invention aims to solve at least one of the problems existing in the prior art and to provide a combustion optimization control system for boilers in thermal power plants.

[0006] This invention provides a combustion optimization control system for boilers in thermal power plants, comprising: The data perception and fusion module is used to collect multi-source real-time data during boiler operation, and to perform noise reduction, alignment and feature extraction on the data to generate a highly reliable combustion state feature vector. The adaptive operating condition identification module, based on the combustion state feature vector, identifies the operating condition category of the current combustion process and its dynamic transfer trend in real time through deep temporal pattern recognition and operating condition clustering analysis. The multi-objective dynamic optimization decision module calls the corresponding multi-objective optimization model according to the identified operating condition category, and solves the optimal set of combustion control parameters with the optimization objectives of maximizing boiler thermal efficiency and minimizing nitrogen oxide emissions. The distributed execution control module receives the optimal combustion control parameter set and decomposes it into coordinated control commands for each actuator, thereby achieving precise linkage adjustment of air-coal ratio, secondary air damper opening and burnout air volume. The data perception and fusion module includes a multi-source data acquisition unit, a time-series alignment and filtering unit, and a feature extraction unit. The multi-source data acquisition unit is used to simultaneously acquire real-time data from boiler furnace temperature sensor, flue gas oxygen sensor, nitrogen oxide concentration detector, primary and secondary air flow meters, coal feeder speed sensor and load command signal; The timing alignment and filtering unit uses a sliding window mechanism to align the timestamps of multi-source data and applies an adaptive filtering algorithm based on wavelet thresholds to suppress sensor noise. The feature extraction unit extracts feature indicators that characterize combustion stability, feature indicators that characterize combustion efficiency, and feature indicators that characterize pollutant generation intensity from the aligned and filtered data stream, and combines them to form a combustion state feature vector.

[0007] Optionally, the adaptive identification module for operating conditions includes a deep temporal network unit and a dynamic clustering analysis unit; The deep temporal network unit takes the combustion state feature vector as input, learns the dynamic temporal dependency of the combustion process through the gated recurrent unit network, and outputs the implicit state vector of the current operating condition. The dynamic clustering analysis unit uses an incremental clustering algorithm to divide and update historical operating conditions online based on the implicit state vector of the operating conditions, and calculates the matching degree between the current operating condition and each cluster center. The operating condition represented by the cluster with the highest matching degree is determined as the current combustion operating condition category.

[0008] Optionally, the adaptive working condition identification module further includes a working condition transition prediction submodule; The working condition transition prediction submodule uses the working condition hidden state vector output by the gated recurrent unit network to predict the probability of working condition transition within the next three sampling periods through a fully connected neural network. When the predicted probability exceeds 0.85, it sends an early warning signal to the multi-objective dynamic optimization decision module to trigger the early switching of the optimization model.

[0009] Optionally, the multi-objective dynamic optimization decision module includes a multi-objective optimization model library and a real-time solver; The multi-objective optimization model library pre-stores optimization models corresponding to different operating conditions. Each optimization model uses the boiler thermal efficiency function and the nitrogen oxide emission function as objective functions, and uses the air-coal ratio constraint, furnace temperature safety constraint and fan output limit as constraint conditions. After receiving the operating condition category signal, the real-time solver loads the corresponding multi-objective optimization model and uses a constrained non-dominated sorting genetic algorithm to solve it online, outputting a set of Pareto optimal solutions. Then, according to the preset preference weights, it selects the final set of optimal combustion control parameters from the Pareto optimal solution set.

[0010] Optionally, the preference weights in the multi-objective dynamic optimization decision-making module can be dynamically configured according to the power plant operation strategy; When the operating strategy focuses on environmental protection requirements, the weight of the nitrogen oxide emission target is set to 0.7, and the weight of the thermal efficiency target is set to 0.3. When the operating strategy focuses on economic efficiency, the weight of the thermal efficiency target is set to 0.7, and the weight of the nitrogen oxide emission target is set to 0.3.

[0011] Optionally, the distributed execution control module includes an instruction decomposition unit and an execution coordination unit; The instruction decomposition unit parses the optimal combustion control parameter set into specific setting values ​​for the blower frequency converter, induced draft fan damper, secondary air damper actuator and coal feeder controller. The execution coordination unit adopts a model predictive control method, with the goal of minimizing the fluctuation of the boiler main steam pressure. It dynamically adjusts the timing and amplitude of the control commands of each actuator to ensure that the system maintains combustion stability during load changes.

[0012] Optionally, the model predictive control method establishes a dynamic model of the boiler main steam pressure, with the model input being the control commands of each actuator and the output being the predicted value of the main steam pressure; In each control cycle, the execution coordination unit takes minimizing the main steam pressure fluctuation over the next 10 cycles as the optimization objective and continuously calculates the optimal adjustment sequence of control commands for each actuator. The optimization problem is solved using quadratic programming, taking into account the action rate constraints of the actuators. The maximum rate of change of the blower frequency converter is 0.5 Hz per second, the maximum rate of change of the induced draft fan damper is 2% per second, the maximum rate of change of the secondary damper is 5 degrees per second, and the maximum rate of change of the coal feeder speed is 1% per second.

[0013] Optionally, the data sensing and fusion module also integrates sensor health status monitoring function; The sensor health status monitoring function analyzes the variance and autocorrelation index of the output data of each sensor to determine whether the sensor is drifting or malfunctioning, and automatically uses virtual sensor data based on a soft measurement model to replace it when an anomaly is detected.

[0014] Optionally, the sensor health status monitoring function determines the sensor status by calculating the variance and autocorrelation index of the output data of each sensor; For furnace temperature sensors, the variance should be less than 15 degrees Celsius and the autocorrelation coefficient should be greater than 0.8 during normal operation. For flue gas oxygen sensors, the normal variance should be less than 0.3%, and the autocorrelation coefficient should be greater than 0.7. When any indicator exceeds the threshold range for three consecutive cycles, the system determines that the sensor is abnormal and automatically switches to the soft measurement model.

[0015] Optionally, during operation, the data perception and fusion module generates a set of combustion state feature vectors every second, the operating condition adaptive identification module updates the operating condition category judgment every 5 seconds, the multi-objective dynamic optimization decision module performs optimization calculations every 30 seconds, and the distributed execution control module outputs a control command every second. Data interaction between modules adopts a publish-subscribe model, with the data perception and fusion module acting as the feature vector publisher and the working condition adaptive identification module acting as the subscriber. The adaptive working condition identification module acts as the working condition category publisher, while the multi-objective dynamic optimization decision module acts as the subscriber. The multi-objective dynamic optimization decision-making module acts as the publisher of optimization parameters, while the distributed execution control module acts as the subscriber.

[0016] Compared with existing technologies, the present invention provides a combustion optimization control system for boilers in thermal power plants. By constructing a closed-loop system architecture encompassing data perception and fusion, adaptive operating condition identification, multi-objective dynamic optimization decision-making, and distributed execution control, it achieves multi-dimensional and accurate perception and intelligent decision-making of the boiler combustion process. It can automatically identify operating condition shifts caused by fluctuations in coal quality and load changes, and switch corresponding optimization strategies in real time, significantly improving adaptability under complex operating conditions. The operating condition identification method, combining deep temporal networks and incremental clustering, overcomes the shortcomings of traditional static models in describing dynamic processes, enhancing the accuracy of state perception. Through the collaboration of a multi-objective optimization model and a real-time solver, a dynamic trade-off between efficiency and environmental protection goals can be achieved, avoiding operational imbalances caused by single-objective optimization. The distributed execution control module, through a model prediction coordination mechanism, ensures coordinated responses of multiple actuators during load changes, effectively improving the system's control accuracy and stability. Overall, the present invention solves the core contradictions of control lag, conflicting optimization goals, and poor operating condition adaptability in existing technologies, providing reliable technical support for the efficient and clean operation of boilers in thermal power plants. Attached Figure Description

[0017] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0018] Figure 1 This is a schematic diagram of the architecture of the boiler combustion optimization control system for thermal power plants proposed in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0020] Example 1: Please refer to the appendix for the overall architecture of the boiler combustion optimization control system in thermal power plants. Figure 1 The system consists of a data sensing and fusion module, a working condition adaptive identification module, a multi-objective dynamic optimization decision-making module, and a distributed execution control module connected sequentially. Each module interacts with the others via industrial Ethernet, using Modbus TCP / IP as the communication protocol. The data frame format includes a timestamp, source module identifier, target module identifier, and payload field. The sampling period for the boiler combustion optimization control system in the thermal power plant is fixed at 1 second, and all real-time data includes time stamp information. Clock synchronization among the modules is achieved through a network time protocol.

[0021] The data sensing and fusion module is the core unit for multi-source data acquisition and preprocessing. It includes a multi-source data acquisition unit, a time-series alignment and filtering unit, and a feature extraction unit. It is used to acquire real-time multi-source data during boiler operation and to perform noise reduction, alignment, and feature extraction on the data to generate a highly reliable combustion state feature vector. Specifically, the multi-source data acquisition unit synchronously acquires real-time data from boiler furnace temperature sensors, flue gas oxygen sensors, nitrogen oxide concentration detectors, primary and secondary air flow meters, coal feeder speed sensors, and load command signals. This is achieved through a parallel acquisition interface that synchronously receives real-time measurements from these sensors. The furnace temperature sensor uses a type K thermocouple with a measurement range of 0 to 1300 degrees Celsius and an output signal of 4 to 20 mA analog quantity. The flue gas oxygen sensor is based on the zirconium oxide principle and has a measurement accuracy of 0.1%. The nitrogen oxide concentration detector uses ultraviolet differential absorption spectroscopy technology with a detection limit of 1 mg / m³. The primary and secondary air flow meters are vortex flow meters with an accuracy class of 0.5%. The coal feeder speed sensor outputs a pulse frequency signal, generating 60 pulses per revolution. The load command signal comes from the power plant's distributed control system and is expressed as a percentage of the current unit load rate. All sensor signals are converted into digital quantities by an isolated analog-to-digital converter with a conversion resolution of 16 bits and transmitted to the acquisition unit buffer via a dedicated data bus.

[0022] The timing alignment and filtering unit performs time consistency processing on the raw data stream output by the multi-source data acquisition unit. This includes aligning the timestamps of the multi-source data using a sliding window mechanism and applying an adaptive filtering algorithm based on wavelet thresholding to suppress sensor noise. The timing alignment and filtering unit maintains a sliding window of length 30, with data within the window arranged in ascending order of acquisition timestamps. For data points with a time deviation exceeding 100 milliseconds, Lagrange interpolation is used for time axis alignment. The noise suppression stage employs an adaptive filtering algorithm based on wavelet thresholding. The specific process includes: performing a 5-level wavelet decomposition on each sensor data sequence using the sym8 wavelet basis function; calculating the noise variance estimate based on the wavelet coefficients at each level; and filtering the detail coefficients using a soft threshold function, with the threshold set to [value missing]. Where σ represents the noise standard deviation and N is the data sequence length. An adaptive filtering algorithm based on wavelet thresholding is applied: a 5-level decomposition is performed using the sym8 wavelet basis, and the noise standard deviation σ is estimated by the median of the detail coefficients at each level. ), threshold according to The algorithm calculates and dynamically adjusts the scaling factor (ranging from 0.8 to 1.2) based on the signal energy gradient to achieve adaptive behavior. The wavelet coefficients, after thresholding, are reconstructed into a denoised data sequence via inverse transform. This algorithm effectively suppresses impulse interference and high-frequency noise, improving the signal-to-noise ratio by at least 12 dB.

[0023] The feature extraction unit extracts three types of combustion state feature indicators from the filtered data stream. Specifically, it extracts feature indicators representing combustion stability, combustion efficiency, and pollutant generation intensity from the aligned and filtered data stream, and combines them to form a combustion state feature vector. Features representing combustion stability include the standard deviation of furnace temperature fluctuation, the mean absolute value of the flue gas oxygen change rate, and the coefficient of variation of the coal feeder speed. The calculation period for the standard deviation of furnace temperature fluctuation is 10 seconds, and the standard deviation of 10 consecutive sampled values ​​is taken as the instantaneous stability indicator; the flue gas oxygen change rate is calculated using first-order difference, and the absolute value is then used to calculate a 30-second moving average; the coefficient of variation of the coal feeder speed is the ratio of the standard deviation to the mean of the speed, and the calculation window length is 15 seconds. Features representing combustion efficiency include the estimated value of flue gas heat loss and the estimated value of unburned carbon loss. Flue gas heat loss is calculated based on the difference between the flue gas temperature and the reference temperature and the flue gas oxygen concentration; unburned carbon loss is indirectly estimated using a fly ash carbon content regression model. Features representing pollutant generation intensity include the nitrogen oxide generation rate and the carbon monoxide concentration gradient. The nitrogen oxide formation rate was obtained by multiplying the concentration change by the flue gas flow rate, and the carbon monoxide concentration gradient was obtained by linearly fitting the slope of five adjacent sampling points. All characteristic indicators were normalized and combined into a combustion state feature vector with a dimension of 8. The vector elements ranged from 0 to 1, and the normalization parameters were dynamically updated based on the extreme values ​​of historical operating data.

[0024] The adaptive operating condition identification module receives the combustion state feature vector output by the data perception and fusion module and performs operating condition pattern recognition. The adaptive operating condition identification module includes a deep temporal network unit and a dynamic clustering analysis unit. The deep temporal network unit takes the combustion state feature vector as input, learns the dynamic temporal dependencies of the combustion process through a gated recurrent unit network, and outputs the implicit state vector of the current operating condition. Based on the implicit state vector, the dynamic clustering analysis unit uses an incremental clustering algorithm to classify and update historical operating conditions online, calculates the matching degree between the current operating condition and each cluster center, and determines the operating condition represented by the cluster with the highest matching degree as the current combustion operating condition category.

[0025] Specifically, the deep temporal network unit adopts a gated recurrent unit (GRU) architecture, with an input layer, two hidden layers, and an output layer. The input layer has 8 neurons, consistent with the dimension of the feature vector; the first hidden layer contains 64 GRUs, and the second hidden layer contains 32 GRUs; the output layer maps the final hidden state to a 16-dimensional working condition hidden state vector through a fully connected layer. The update and reset gates of the GRUs use the sigmoid function as their activation function, and the candidate hidden state activation function uses the hyperbolic tangent function. The network is trained using three years of historical data, with the Adam algorithm selected for optimization. The initial learning rate is set to 0.001, and the batch size is 128. An early stopping strategy is employed during training, terminating training when the validation set loss does not decrease for 10 consecutive epochs. During forward inference, the feature vector sequence of 20 consecutive time steps is input into the network. Before network training, the feature vectors are normalized by Z-score, and the hidden layer weights are initialized using Xavier. During real-time inference, zeros are padded at the front of sequences with fewer than 20 time steps, and for very long sequences, a sliding window is used to truncate the most recent 20 sampling points. The output of the last time step is taken as the hidden state vector of the current working condition.

[0026] The dynamic clustering analysis unit performs online working condition classification on the implicit state vector of the working conditions output by the deep temporal network unit. The dynamic clustering analysis unit adopts an incremental clustering algorithm. The initial cluster centers are obtained by preprocessing historical data using the K-means algorithm. The number of clusters is determined based on the silhouette coefficient, typically ranging from 5 to 8. When new data arrives, its Euclidean distance to all cluster centers is calculated. If the minimum distance is less than the dynamic threshold, the data is assigned to the corresponding cluster and the cluster center is updated; otherwise, a new cluster is created. The dynamic threshold is set at 1.5 times the cluster radius, which is obtained by calculating the average distance from the data points within a cluster to the center point. The current working condition category is determined based on the matching degree, defined as the difference between 1 and the normalized distance, which is the ratio of the actual distance to the threshold. Every 5 minutes, the system recalculates the statistical characteristics of all clusters, including intra-cluster compactness and inter-cluster separation. When the intra-cluster compactness falls below 0.7, cluster reconstruction is triggered.

[0027] The multi-objective dynamic optimization decision-making module calls the corresponding optimization model based on the operating condition category. This module includes a multi-objective optimization model library and a real-time solver. The multi-objective optimization model library stores mathematical models of optimization problems corresponding to different operating conditions, i.e., pre-stored optimization models corresponding to different operating condition categories. Each model contains two objective functions: a boiler thermal efficiency function and a nitrogen oxide emission function. The boiler thermal efficiency function is defined as the ratio of the input fuel calorific value to the output steam effective calorific value, and its relationship with the operating variables is established through the boiler heat balance equation. The nitrogen oxide emission function uses an empirical model, relating the nonlinear relationship between nitrogen oxide generation concentration and furnace temperature and excess air coefficient. Constraints include air-fuel ratio constraints, furnace temperature safety constraints, and fan output limits. The air-to-coal ratio constraint requires that the ratio of primary air volume to coal feed be maintained between 1.8 and 2.2; the furnace temperature safety constraint limits the temperature at each measuring point in the furnace to not exceed 1250 degrees Celsius; the fan output limit stipulates that the frequency of the blower inverter is in the range of 30 Hz to 50 Hz, and the opening of the induced draft fan damper is in the range of 40% to 100%.

[0028] Upon receiving the operating condition category signal, the real-time solver loads the corresponding multi-objective optimization model and uses a constrained non-dominated sorting genetic algorithm for online solution, outputting a set of Pareto optimal solutions. Then, based on preset preference weights, it selects the final optimal set of combustion control parameters from this Pareto optimal solution set. Specifically, the real-time solver uses a constrained non-dominated sorting genetic algorithm for multi-objective optimization. The algorithm's population size is set to 100, the crossover probability is 0.9, the mutation probability is 0.1, and the maximum number of generations is 200. Constraint handling employs a penalty function method, where the fitness of individuals violating constraints is weighted according to the degree of violation. The solution process first randomly generates an initial population that satisfies the constraints, then generates offspring populations through simulated binary crossover and polynomial mutation. Non-dominated sorting divides the population into multiple Pareto fronts, and crowding calculation is used to maintain solution set diversity. After the algorithm terminates, it outputs a Pareto optimal solution set, where each solution contains a set of air-coal ratio, secondary damper opening, and burnout air volume setpoints. The final decision selects the optimal solution from the solution set based on preset preference weights. The weight configuration is received from the power plant operators through the human-machine interface. The weight values ​​are limited to the range of 0 to 1, and the sum of the two weights must be equal to 1.

[0029] The distributed execution control module converts the optimal combustion control parameter set into actuator commands. The module includes an instruction decomposition unit and an execution coordination unit. The instruction decomposition unit parses the optimal combustion control parameter set into specific setpoints for the forced draft fan inverter, induced draft fan damper, secondary damper actuator, and coal feeder controller. Based on equipment characteristics, the instruction decomposition unit maps the optimized setpoints to specific control signals. The forced draft fan inverter setpoint is a frequency value, the correspondence of which is obtained through linear interpolation of the fan performance curve; the induced draft fan damper setpoint is an opening percentage, directly converted into a 4-20 mA analog signal; the secondary damper actuator setpoint is an angle value, ranging from 0 to 90 degrees, driven by a pulse width modulation signal; the coal feeder controller setpoint is a speed percentage, converted into a 0-10 V voltage signal. All control signals are amplitude-limited before output to prevent actuators from operating beyond their limits.

[0030] The execution coordination unit employs model predictive control (MMCC) to ensure coordinated action of multiple actuators. Specifically, it uses MMCC with the goal of minimizing boiler main steam pressure fluctuations, dynamically adjusting the timing and amplitude of control commands for each actuator to maintain combustion stability during load changes. The execution coordination unit establishes a dynamic model of the boiler main steam pressure, with the control commands from each actuator as input and the predicted main steam pressure as output. The model structure is a multi-input single-output transfer function matrix, identified through step response testing. In each control cycle, the execution coordination unit continuously calculates the optimal adjustment sequence of control commands for each actuator, aiming to minimize main steam pressure fluctuations over the next 10 cycles. The optimization problem is solved using quadratic programming, considering actuator action rate constraints: the maximum rate of change for the forced draft fan inverter is 0.5 Hz, for the induced draft fan damper it is 2 percent per second, for the secondary damper it is 5 degrees per second, and for the coal feeder speed it is 1 percent per second. Only the first control variable of the optimized sequence is implemented; prediction optimization is performed again in the next cycle, forming a closed-loop control.

[0031] The adaptive operating condition identification module also includes an operating condition transition prediction submodule. This submodule, based on the operating condition hidden state vector output by the gated recurrent unit network, predicts the probability of an operating condition transition within the next three sampling periods (i.e., the operating condition transition probability) using a fully connected neural network. The neural network structure consists of an input layer, two hidden layers, and an output layer. The input layer has 16 neurons, consistent with the dimension of the hidden state vector; the first hidden layer contains 32 neurons, and the second hidden layer contains 16 neurons; the output layer outputs four probability values ​​through a softmax activation function, corresponding to the probabilities of an operating condition transition occurring in the 1st, 2nd, and 3rd periods, respectively. The network training data comes from historical operating condition transition records, and the ratio of positive to negative samples is balanced using oversampling techniques. When any predicted probability exceeds 0.85, the operating condition transition prediction submodule sends an early warning signal to the multi-objective dynamic optimization decision module, triggering the optimization model preloading mechanism to reduce control latency during operating condition switching.

[0032] The data perception and fusion module integrates sensor health status monitoring. By analyzing the variance and autocorrelation indices of each sensor's output data, it determines whether the sensors are drifting or malfunctioning, and automatically switches to virtual sensor data based on a soft sensor model when an anomaly is detected. The sensor health status monitoring function determines the sensor status by calculating the variance and autocorrelation indices of each sensor's output data. The variance index calculates the standard deviation of the data within a sliding window with a window length of 60 seconds; the autocorrelation index calculates the autocorrelation coefficient lagging 1 to 5 sampling points. For the furnace temperature sensor, the variance should be less than 15 degrees Celsius and the autocorrelation coefficient should be greater than 0.8 during normal operation; for the flue gas oxygen sensor, the normal variance should be less than 0.3 percent and the autocorrelation coefficient should be greater than 0.7. When any indicator exceeds the threshold range for three consecutive cycles, the system determines that the sensor is abnormal and automatically switches to the soft sensor model. The soft sensor model is based on partial least squares regression, with input variables being data from other normal sensors and output being an estimated value of the faulty sensor. The soft measurement model employs partial least squares regression: the input variables are data from seven normal sensors, including furnace temperature, flue gas oxygen content, and primary air volume. Latent variables are solved using the NIPALS algorithm. The objective function is to minimize the estimation error of faulty sensors. The number of latent variables is determined through 10-fold cross-validation, with RMSE < 5% as the convergence criterion. The model is updated every 24 hours, and the model parameters are retrained using data from the most recent 30 days.

[0033] During operation, the boiler combustion optimization control system in a thermal power plant generates a set of combustion state feature vectors every second using the data perception and fusion module. The adaptive operating condition identification module updates the operating condition category judgment every 5 seconds, the multi-objective dynamic optimization decision module performs optimization calculations every 30 seconds, and the distributed execution control module outputs a control command every second. Data interaction between modules adopts a publish-subscribe model: the data perception and fusion module acts as the feature vector publisher, and the adaptive operating condition identification module acts as the subscriber; the adaptive operating condition identification module acts as the operating condition category publisher, and the multi-objective dynamic optimization decision module acts as the subscriber; the multi-objective dynamic optimization decision module acts as the optimization parameter publisher, and the distributed execution control module acts as the subscriber. RabbitMQ is used as the message middleware, and persistent message storage ensures no data loss.

[0034] The boiler combustion optimization control system for thermal power plants is deployed on an industrial server cluster. The data sensing and fusion module runs on a real-time operating system to ensure the timeliness of data acquisition and processing; the adaptive operating condition identification module and the multi-objective dynamic optimization decision-making module run on high-performance computing nodes, equipped with GPU acceleration cards to improve the inference speed of deep learning models; and the distributed execution control module runs on redundant controllers, employing a hot backup mechanism to ensure control continuity. The boiler combustion optimization control system is integrated with the power plant's distributed control system via the OPC UA protocol, receiving load commands and equipment status in real time, while simultaneously uploading optimization control commands and system operating status. The human-machine interface adopts a web architecture, supports multi-terminal access, and displays real-time combustion status curves, operating condition category identifiers, optimization target weight setting interfaces, and a system health status dashboard.

[0035] Example 2: Building upon Example 1, this example enhances the feature extraction capabilities of the data perception and fusion module to address the specific operating conditions of drastic fluctuations in coal quality. The feature extraction unit adds a coal quality characteristic identification subunit. This subunit indirectly estimates the volatile matter and calorific value of the coal by analyzing the current fluctuation spectrum of the coal feeder and the primary air pressure variation pattern. The coal feeder current signal undergoes Fast Fourier Transform (FFT) to extract the energy proportion in the 3Hz to 10Hz frequency band as a coal hardness indicator. The primary air pressure variation pattern is obtained through Empirical Mode Decomposition (EMD) to obtain intrinsic mode functions (EMFs), and the energy entropy of the first three EMFs is calculated as a coal uniformity indicator. These two indicators are combined with the 8-dimensional feature vector from Example 1 to form a 10-dimensional enhanced combustion state feature vector.

[0036] The deep temporal network unit of the adaptive working condition identification module adjusts the network structure to enhance feature vectors. The number of neurons in the input layer is increased to 10, the number of gated recurrent units in the first hidden layer is increased to 72, and the number of gated recurrent units in the second hidden layer remains at 32 dimensions. Network retraining uses historical data including coal quality mutation scenarios, and the training strategy introduces a focus loss function, increasing the penalty weight for a minority of working conditions to improve the model's sensitivity to identifying new working conditions caused by coal quality changes. The incremental learning mechanism of the dynamic clustering analysis unit adds a category merging function. When the distance between two cluster centers is less than a threshold and the number of samples within each cluster is less than 100, they are automatically merged into a single cluster to prevent fragmentation of working condition categories.

[0037] The model library of the multi-objective dynamic optimization decision-making module has been expanded to include an adaptive optimization model based on coal quality. This model adds a coal quality compensation term to the standard objective function, introduces a coal calorific value correction coefficient into the boiler thermal efficiency function, and incorporates a volatile matter influencing factor into the nitrogen oxide emission function. A primary air pressure fluctuation limit has been added to the constraints, requiring that the primary air pressure change not exceed ±200 Pascals within 10 seconds to prevent coal quality changes from causing mill blockage. The real-time solver optimization algorithm parameters are adaptively adjusted; when the detected coal quality index change rate exceeds 5%, the population size is increased to 150, and the crossover probability is reduced to 0.85, enhancing the algorithm's global search capability.

[0038] The execution coordination unit of the distributed execution control module strengthens its anti-interference strategy. The optimization objective of model predictive control adds a coal-air ratio tracking term with a weight of 0.2 to reduce the impact of coal quality fluctuations on the coal-air balance. Feedforward compensation is added to the actuator control commands to predict coal quality trends based on the feeder current spectrum characteristics and adjust the blower speed setpoint in advance. The feedforward compensation is calculated through fuzzy inference, with input variables being coal hardness and uniformity indices, and output variable being the blower speed compensation percentage, limited to within ±3%.

[0039] The boiler combustion optimization control system for thermal power plants incorporates an operation mode switching mechanism. It defaults to the standard mode of Example 1. When a change in coal hardness exceeding 15% is detected for three consecutive cycles, it automatically switches to the enhanced mode of this example. During mode switching, the data sensing and fusion module gradually transitions to the enhanced feature vector, with a transition period of five sampling cycles. A linear interpolation method is used to mix the two feature vectors to avoid abrupt changes in control commands. During enhanced mode operation, the system assesses the degree of coal quality fluctuation every 10 minutes. Once the coal quality index stabilizes, it automatically switches back to the standard mode.

[0040] This embodiment significantly improves the adaptability of the boiler combustion optimization control system in thermal power plants to fluctuations in coal quality by enhancing feature extraction and employing specialized optimization strategies. Actual test data shows that even with a 20% fluctuation in coal calorific value, the system can still maintain a boiler thermal efficiency of over 93% and a nitrogen oxide emission concentration below 100 mg / m³, representing a 12% performance improvement compared to the standard mode in Example 1.

[0041] Those skilled in the art will understand that the above embodiments are specific implementations of the present invention, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A combustion optimization control system for boilers in thermal power plants, characterized in that, include: The data perception and fusion module is used to collect multi-source real-time data during boiler operation, and to perform noise reduction, alignment and feature extraction on the data to generate a highly reliable combustion state feature vector. The adaptive operating condition identification module, based on the combustion state feature vector, identifies the operating condition category of the current combustion process and its dynamic transfer trend in real time through deep temporal pattern recognition and operating condition clustering analysis. The multi-objective dynamic optimization decision module calls the corresponding multi-objective optimization model according to the identified operating condition category, and solves the optimal set of combustion control parameters with the optimization objectives of maximizing boiler thermal efficiency and minimizing nitrogen oxide emissions. The distributed execution control module receives the optimal combustion control parameter set and decomposes it into coordinated control commands for each actuator, thereby achieving precise linkage adjustment of air-coal ratio, secondary air damper opening and burnout air volume. The data perception and fusion module includes a multi-source data acquisition unit, a time-series alignment and filtering unit, and a feature extraction unit. The multi-source data acquisition unit is used to simultaneously acquire real-time data from boiler furnace temperature sensor, flue gas oxygen sensor, nitrogen oxide concentration detector, primary and secondary air flow meters, coal feeder speed sensor and load command signal; The timing alignment and filtering unit uses a sliding window mechanism to align the timestamps of multi-source data and applies an adaptive filtering algorithm based on wavelet thresholds to suppress sensor noise. The feature extraction unit extracts feature indicators that characterize combustion stability, feature indicators that characterize combustion efficiency, and feature indicators that characterize pollutant generation intensity from the aligned and filtered data stream, and combines them to form a combustion state feature vector.

2. The combustion optimization control system for a thermal power plant boiler according to claim 1, characterized in that, The adaptive identification module for operating conditions includes a deep temporal network unit and a dynamic clustering analysis unit; The deep temporal network unit takes the combustion state feature vector as input, learns the dynamic temporal dependency of the combustion process through the gated recurrent unit network, and outputs the implicit state vector of the current operating condition. The dynamic clustering analysis unit uses an incremental clustering algorithm to divide and update historical operating conditions online based on the implicit state vector of the operating conditions, and calculates the matching degree between the current operating condition and each cluster center. The operating condition represented by the cluster with the highest matching degree is determined as the current combustion operating condition category.

3. The combustion optimization control system for a thermal power plant boiler according to claim 2, characterized in that, The adaptive working condition identification module also includes a working condition transition prediction submodule; The working condition transition prediction submodule uses the working condition hidden state vector output by the gated recurrent unit network to predict the probability of working condition transition within the next three sampling periods through a fully connected neural network. When the predicted probability exceeds 0.85, it sends an early warning signal to the multi-objective dynamic optimization decision module to trigger the early switching of the optimization model.

4. A combustion optimization control system for a thermal power plant boiler according to any one of claims 1 to 3, characterized in that, The multi-objective dynamic optimization decision module includes a multi-objective optimization model library and a real-time solver; The multi-objective optimization model library pre-stores optimization models corresponding to different operating conditions. Each optimization model uses the boiler thermal efficiency function and the nitrogen oxide emission function as objective functions, and uses the air-coal ratio constraint, furnace temperature safety constraint and fan output limit as constraint conditions. After receiving the operating condition category signal, the real-time solver loads the corresponding multi-objective optimization model and uses a constrained non-dominated sorting genetic algorithm to solve it online, outputting a set of Pareto optimal solutions. Then, according to the preset preference weights, it selects the final set of optimal combustion control parameters from the Pareto optimal solution set.

5. The combustion optimization control system for a thermal power plant boiler according to claim 4, characterized in that, The preference weights in the multi-objective dynamic optimization decision-making module can be dynamically configured according to the power plant operation strategy; When the operating strategy focuses on environmental protection requirements, the weight of the nitrogen oxide emission target is set to 0.7, and the weight of the thermal efficiency target is set to 0.

3. When the operating strategy focuses on economic efficiency, the weight of the thermal efficiency target is set to 0.7, and the weight of the nitrogen oxide emission target is set to 0.

3.

6. A combustion optimization control system for a thermal power plant boiler according to any one of claims 1 to 3, characterized in that, The distributed execution control module includes an instruction decomposition unit and an execution coordination unit; The instruction decomposition unit parses the optimal combustion control parameter set into specific setting values ​​for the blower frequency converter, induced draft fan damper, secondary air damper actuator and coal feeder controller. The execution coordination unit adopts a model predictive control method, with the goal of minimizing the fluctuation of the boiler main steam pressure. It dynamically adjusts the timing and amplitude of the control commands of each actuator to ensure that the system maintains combustion stability during load changes.

7. The combustion optimization control system for a thermal power plant boiler according to claim 6, characterized in that, The model predictive control method establishes a dynamic model of the boiler main steam pressure. The model input is the control commands of each actuator, and the output is the predicted value of the main steam pressure. In each control cycle, the execution coordination unit takes minimizing the main steam pressure fluctuation over the next 10 cycles as the optimization objective and continuously calculates the optimal adjustment sequence of control commands for each actuator. The optimization problem is solved using quadratic programming, taking into account the action rate constraints of the actuators. The maximum rate of change of the blower frequency converter is 0.5 Hz per second, the maximum rate of change of the induced draft fan damper is 2% per second, the maximum rate of change of the secondary damper is 5 degrees per second, and the maximum rate of change of the coal feeder speed is 1% per second.

8. A combustion optimization control system for a thermal power plant boiler according to any one of claims 1 to 3, characterized in that, The data sensing and fusion module also integrates sensor health status monitoring function; The sensor health status monitoring function analyzes the variance and autocorrelation index of the output data of each sensor to determine whether the sensor is drifting or malfunctioning, and automatically uses virtual sensor data based on a soft measurement model to replace it when an anomaly is detected.

9. A combustion optimization control system for a thermal power plant boiler according to claim 8, characterized in that, The sensor health status monitoring function determines the sensor status by calculating the variance and autocorrelation index of the output data of each sensor. For furnace temperature sensors, the variance should be less than 15 degrees Celsius and the autocorrelation coefficient should be greater than 0.8 during normal operation. For flue gas oxygen sensors, the normal variance should be less than 0.3%, and the autocorrelation coefficient should be greater than 0.

7. When any indicator exceeds the threshold range for three consecutive cycles, the system determines that the sensor is abnormal and automatically switches to the soft measurement model.

10. A combustion optimization control system for a thermal power plant boiler according to any one of claims 1 to 3, characterized in that, During operation, the boiler combustion optimization control system of a thermal power plant generates a set of combustion state feature vectors every second by the data perception and fusion module, updates the operating condition category judgment every 5 seconds by the operating condition adaptive identification module, performs an optimization calculation every 30 seconds by the multi-objective dynamic optimization decision module, and outputs a control command every second by the distributed execution control module. Data interaction between modules adopts a publish-subscribe model, with the data perception and fusion module acting as the feature vector publisher and the working condition adaptive identification module acting as the subscriber. The adaptive working condition identification module acts as the working condition category publisher, while the multi-objective dynamic optimization decision module acts as the subscriber. The multi-objective dynamic optimization decision-making module acts as the publisher of optimization parameters, while the distributed execution control module acts as the subscriber.

Images