System and method for online identification of coal heat value and optimization of air-coal ratio based on the same

Abstract

The application provides an online identification and wind-coal ratio optimization system and method based on the calorific value of coal entering a furnace, and relates to the field of combustion optimization control of coal-fired generating units. The method comprises the following steps: collecting and preprocessing real-time DCS data, constructing a mechanism model channel and a BP neural network channel, calculating a rough identification calorific value and decomposing error components, and embedding physical consistency constraints in a loss function; a two-way correction mechanism is established, error components are corrected by using dynamic parameters, and multiple source uncertainty factors are fused to output a calorific value feature information package; based on efficiency prediction uncertainty, an optimization period and a constraint boundary are adaptively adjusted, and an optimal wind-coal ratio parameter set is solved. The application realizes deep coupling of a mechanism model and a data model, and integration and cooperation of calorific value identification and optimization decision, and significantly improves control precision and system robustness.

Description

Technical Field

[0001] This application relates to the field of combustion optimization control of coal-fired power generating units, and in particular to a system and method for online identification of the calorific value of coal fed into the furnace and optimization of the air-coal ratio. Background Technology

[0002] Currently, the core of combustion optimization control in coal-fired power plant boilers lies in the accurate online sensing of the calorific value of the coal fed into the furnace and the dynamic optimization matching of the air-coal ratio. In existing technologies, the mainstream solutions typically adopt a serial combination architecture of "mechanism model + BP neural network error compensation". This architecture uses the heat balance equation of the pulverizing system to construct a mechanism model to calculate a rough calorific value, and also uses a BP neural network to predict and correct the errors of the mechanism model.

[0003] However, in practical applications, due to the lack of a deep interaction mechanism based on physical characteristics between the mechanistic model and the data model, when the type of coal fed into the furnace changes or the operating conditions are in a sparse region of the training samples, the output directions of the two models may conflict, leading to non-physical oscillations in the final calorific value identification result. Furthermore, the existing calorific value identification and air-coal ratio optimization layers only have a serial data transmission relationship. Uncertainty information such as confidence level and error sources in the calorific value identification process is lost during transmission, causing the air-coal ratio optimization module to be unable to adjust its decision-making strategy based on the reliability of the calorific value. Therefore, there is an urgent need for a method and system that can achieve deep coupling between the mechanistic model and the data model, and realize integrated collaborative decision-making for calorific value identification and air-coal ratio optimization. Summary of the Invention

[0004] To address the issues of insufficient interaction between the mechanism model and data model, and the disconnect between calorific value identification and optimization decision-making in existing technologies, which leads to a decrease in control accuracy, this application constructs an integrated optimization architecture based on online identification of the calorific value of the coal fed into the furnace and optimization of the air-coal ratio. This architecture features a dual-channel deep coupling mechanism and a feature-driven approach, thereby improving the accuracy of calorific value identification and enabling adaptive adjustment of the air-coal ratio optimization decision-making.

[0005] To achieve the above objectives, this application adopts the following technical solution: The first aspect involves an online identification method for the calorific value of coal fed into the furnace and an optimal air-coal ratio method. This includes: collecting real-time data from the DCS system of the coal-fired power generation unit, performing wavelet denoising and time-series alignment, and outputting feature vectors. A mechanistic model channel and a BP neural network channel are constructed. The mechanistic model channel calculates the coarsely identified calorific value and outputs dynamic parameters. The BP neural network channel uses the feature vectors as input to obtain the decomposed dynamic characteristic error components and coal quality characteristic error components, embedding physical consistency constraints into the loss function of the BP neural network channel. A bidirectional correction mechanism is established. Within each control cycle, the dynamic parameters are used as input features to the BP neural network channel to correct the error components online using dynamic characteristics, and multi-source uncertainties are integrated to output a calorific value feature information package. The risk aversion coefficient is calculated based on the calorific value feature information package to obtain the efficiency prediction uncertainty. Based on the efficiency prediction uncertainty, an optimization objective is set, and the optimization cycle and search space constraint boundaries are adaptively adjusted. A multi-objective optimization algorithm is used to solve for the optimal air-coal ratio parameter set and output execution instructions.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves wavelet denoising and time-series alignment to output a feature vector. Specifically, this includes: constructing a multi-scale wavelet denoising unit; decomposing the original signal based on the Daubechies-4 wavelet basis; extracting low-frequency components for reconstruction; and outputting a purified signal sequence. A cross-correlation function is used to calculate the optimal delay time between the coal feeding command and the combustion response. The purified signal sequence is then time-aligned based on this optimal delay time to compensate for the transmission lag from coal powder preparation to the combustion response, outputting a time-synchronized feature vector. This feature vector includes at least: coal feed rate, mill current, mill inlet and outlet differential pressure, mill inlet air temperature, mill outlet air temperature, primary air volume, secondary air volume, flue gas oxygen content, exhaust gas temperature, and unit load.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the calculation process of the mechanistic model channel specifically includes: constructing a static mechanistic model based on the heat balance equation of the pulverizing system; performing inverse calculation of the heat balance equation on real-time data from the DCS system to obtain the coarse calorific value of the coal fed into the furnace; and calculating the dynamic parameters of the first-order inertial plus pure time delay model within the pulverizing system. The dynamic parameters include the inertial time constant and the pure time delay. Within the current control cycle, the dynamic parameters are used as input features of the BP neural network channel, and the dynamic characteristic error components from the previous control cycle are received. Based on the dynamic characteristic error components from the previous control cycle, the inertial time constant is corrected online using a recursive formula, where the correction step size is determined by the product of the feedback gain coefficient and the dynamic characteristic error components processed by the low-pass filter.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the specific form of the loss function for physical consistency constraints includes: constructing a composite loss function for the BP neural network channels. This composite loss function consists of a data fitting loss term, a physical consistency penalty term, and an output smoothing regularization term, weighted and summed using preset weight coefficients. The data fitting loss term is calculated based on the mean square error between the fused calorific value output by the BP neural network channels and the true coal quality analysis value. The physical consistency penalty term is calculated based on the degree to which the sum of the coarsely identified calorific value output by the mechanistic model channels and the error correction amount output by the BP neural network channels deviates from the preset physical calorific value boundary. It adopts a hinged loss function form; when the fused calorific value is within the preset physical calorific value boundary, the penalty term is zero; when the fused calorific value exceeds the preset physical calorific value boundary, the penalty term is proportional to the excess amount. The output smoothing regularization term is calculated based on the L2 norm of the time gradient of the error correction amount output by the BP neural network channels. During the forward and backward propagation iterations of the BP neural network channels, the composite loss function is calculated synchronously, and the preset weight coefficients are updated with the goal of minimizing this composite loss function.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the decomposition process of the dynamic characteristic error component and the coal quality characteristic error component specifically includes: the output layer of the BP neural network channel is a dual-output layer containing a first output node and a second output node. The first output node outputs the dynamic characteristic error component, and the second output node outputs the coal quality characteristic error component. The dynamic characteristic error component is used to characterize the deviation between the first-order inertial plus pure lag model parameters of the mechanism model channel and the dynamic characteristics of the actual pulverizing system. The coal quality characteristic error component is used to characterize the steady-state error of the mechanism model caused by unconventional coal quality changes in the coal fed into the furnace. During the training phase of the BP neural network channel, the difference between the true value of coal quality analysis and the coarsely identified calorific value output by the mechanism model channel is used as the total error target, and the error backpropagation algorithm is adopted. The weight parameters corresponding to the first output node and the second output node are updated. During the inference phase of the BP neural network channel, the feature vector is input into the BP neural network channel, and the dynamic characteristic error component and the coal quality characteristic error component are output synchronously, and the sum of the two is output as the total error correction amount.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the bidirectional correction mechanism specifically includes: establishing a forward propagation channel, using the dynamic parameters of the first-order inertial plus pure time delay model calculated by the mechanistic model channel as input features of the BP neural network channel; the dynamic parameters include the inertial time constant and the pure time delay. Establishing a reverse correction channel, receiving the dynamic characteristic error components output by the BP neural network channel, filtering them through a first-order low-pass filter, multiplying them by the feedback gain coefficient to obtain the correction step size, and updating the inertial time constant of the mechanistic model channel online. Before the next control cycle, the updated inertial time constant is input again into the BP neural network channel through the forward propagation channel, forming a closed bidirectional interactive loop, used to achieve coordinated adaptive adjustment of dynamic parameters and neural network input features under coal type switching or pulverizing system characteristic changes.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves fusing multiple sources of uncertainty, specifically including: These multiple sources of uncertainty include input spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty. For the input spatial distance uncertainty, the spatial distance uncertainty is calculated based on the Mahalanobis distance between the current feature vector and the training sample set of the BP neural network channel, used to characterize the sparsity of the current operating condition in the training sample space. For the model parameter identification uncertainty, the model parameter identification uncertainty is calculated based on the variance estimation of the output layer weights of the BP neural network channel, used to characterize the confidence level of the model parameters themselves. For the sensor noise uncertainty, the sensor noise uncertainty is calculated based on the real-time signal-to-noise ratio estimation of the DCS system's real-time data, used to characterize the reliability of the measurement data. The spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty are weighted and fused to obtain a comprehensive uncertainty, where the weighting coefficients are adaptively adjusted according to the unit load change rate to enhance uncertainty perception under dynamic operating conditions.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the calorific value characteristic information package specifically includes the final fused calorific value, comprehensive uncertainty, physical consistency deviation, dynamic characteristic intensity, and coal quality characteristic change intensity. The final fused calorific value is obtained by adding the coarsely identified calorific value output from the mechanistic model channel to the total error correction amount output from the BP neural network channel. The comprehensive uncertainty is obtained by weighted fusion of input spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty. The physical consistency deviation is calculated based on the sum of the coarsely identified calorific value and the total error correction amount, deviating from the preset physical calorific value boundary, and is used to characterize whether the current fused calorific value is close to or exceeds the physically reasonable range. The dynamic characteristic intensity is taken as the absolute value of the dynamic characteristic error component output from the BP neural network channel, and is used to characterize the intensity of the dynamic response under the current operating condition. The coal quality characteristic change intensity is taken as the absolute value of the coal quality characteristic error component output from the BP neural network channel, and is used to characterize the degree of change in the coal quality entering the furnace.

[0013] In conjunction with the first aspect mentioned above, one possible implementation further includes: after executing the instruction, collecting the actual boiler efficiency and calculating the efficiency prediction deviation. Based on the efficiency prediction deviation and calorific value characteristic information package, selectively triggering BP neural network channel updates. Specifically, when the physical consistency deviation exceeds a first preset threshold and the duration exceeds a first preset time window, incremental retraining of the BP neural network channel is triggered. When the moving average of the dynamic feature intensity continuously exceeds a second preset time window and is non-zero, and the efficiency prediction deviation exceeds a second preset threshold, recursive least squares online correction of the inertial time constant of the mechanistic model channel is triggered. When the triggering conditions are not met, the current parameters of the BP neural network channel and the mechanistic model channel remain unchanged.

[0014] Secondly, a system for online identification of the calorific value of coal fed into the furnace and optimization of the air-coal ratio is provided, including a communication unit and a processing unit. The communication unit is connected to the DCS system of the coal-fired power generation unit to collect real-time data from the DCS system and transmit the real-time data to the processing unit.

[0015] The processing unit is used to construct a mechanistic model channel and a BP neural network channel. The mechanistic model channel calculates the coarsely identified calorific value and outputs dynamic parameters. The BP neural network channel takes the feature vector as input to obtain the decomposed dynamic characteristic error components and coal quality characteristic error components, and embeds physical consistency constraints into the loss function of the BP neural network channel. A two-way correction mechanism is established. In each control cycle, the dynamic parameters are used as input features of the BP neural network channel to correct the error components online with dynamic characteristics, and multi-source uncertainties are fused to output a calorific value feature information package. The risk aversion coefficient is calculated based on the calorific value feature information package to obtain the efficiency prediction uncertainty. Based on the efficiency prediction uncertainty, an optimization objective is set, and the optimization cycle and search space constraint boundary are adaptively adjusted. A multi-objective optimization algorithm is used to solve for the optimal air-coal ratio parameter set and outputs execution instructions.

[0016] This application provides a system and method for online identification of calorific value of incoming coal and optimization of air-coal ratio. By constructing a deep dual-channel coupling architecture of a mechanistic model channel and a BP neural network channel, and embedding physical consistency constraints into the loss function of the BP neural network, the output of the neural network can be constrained by physical boundaries. This effectively avoids the problems of conflicting output directions and non-physical oscillations between the two models in traditional schemes, improving the accuracy and stability of calorific value identification. Simultaneously, by establishing a bidirectional correction mechanism, the dynamic characteristic error component of the BP neural network output is used to inversely correct the inertial time constant of the mechanistic model, achieving adaptive calibration of model parameters to equipment state drift and solving the problem of decreased model accuracy due to equipment aging. Furthermore, the calorific value feature information package generated by integrating multi-source uncertainties can be transmitted to the optimization decision layer as a connecting link, realizing integrated optimization decision based on risk perception. This allows the system to adaptively adjust the optimization cycle, search space, and constraint boundaries according to the reliability of the calorific value, avoiding erroneous decisions when the calorific value is unreliable, thus improving the system's safety and robustness. In addition, through a closed-loop feedback and feature-driven parameter self-calibration mechanism, online updating and evolution of the model are achieved, ensuring the long-term accuracy of the system.

[0017] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the method for online identification of calorific value of coal fed into the furnace and optimization of air-coal ratio provided in this application embodiment; Figure 2 The flowchart of data preprocessing and time-series alignment in the online identification of calorific value of coal and optimization of air-coal ratio provided in the embodiments of this application; Figure 3 The dual-channel deep coupling architecture diagram in the online identification of calorific value of coal fed into the furnace and optimization of air-coal ratio provided in the embodiments of this application; Figure 4 The bidirectional correction mechanism interaction diagram in the online identification of calorific value of coal fed into the furnace and the optimization of air-coal ratio provided in the embodiments of this application; Figure 5 A schematic diagram of error component decomposition in the online identification of calorific value of coal fed into the furnace and optimization of air-coal ratio provided in the embodiments of this application; Figure 6 The flowchart for calculating the composite loss function in the online identification of calorific value of coal fed into the furnace and the optimization of the air-coal ratio provided in the embodiments of this application is as follows: Figure 7 In the online identification of calorific value of coal fed into the furnace and optimization of air-coal ratio provided in the embodiments of this application, a multi-source uncertainty factor fusion and calorific value feature information package generation diagram is provided. Figure 8 The integrated optimization decision-making flowchart is provided in the online identification of calorific value of coal fed into the furnace and the optimization method of air-coal ratio in the embodiments of this application. Figure 9 The flowchart of closed-loop feedback and model self-calibration in the online identification of calorific value of coal fed into the furnace and optimization of air-coal ratio provided in the embodiments of this application; Figure 10 The overall architecture diagram of the system based on online identification of calorific value of coal entering the furnace and optimization of air-coal ratio provided in the embodiments of this application is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this 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 this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0021] Example 1: like Figure 1-2 As shown, this embodiment provides a method for online identification of the calorific value of coal entering the furnace and optimization of the air-coal ratio. This method aims to solve the problems of output conflict, information loss, and inability to adapt to changes in operating conditions caused by the simple serial combination of mechanistic and data models in existing technologies. By constructing a dual-channel deeply coupled architecture and a feature-driven integrated optimization mechanism, it achieves precise decision-making for combustion optimization control.

[0022] Specifically, the method includes the following steps: Step 100: Collect real-time data from the DCS system of the coal-fired power generation unit, perform wavelet denoising and time alignment, and output feature vectors.

[0023] The DCS system's real-time data covers key parameters during the operation of coal-fired power generating units, such as coal feed rate, mill current, mill inlet and outlet differential pressure, mill inlet air temperature, mill outlet air temperature, primary air volume, secondary air volume, flue gas oxygen content, exhaust gas temperature, and unit load. Due to the complex on-site environment, the raw data often contains a large amount of high-frequency noise and suffers from transmission lag; direct use would lead to a decrease in model accuracy. Therefore, this step first uses wavelet denoising technology to filter out high-frequency noise components from the signal, retaining low-frequency features reflecting combustion trends. Then, time-series alignment technology is used to compensate for the transmission lag from pulverized coal preparation to combustion response, ensuring the temporal consistency of the input data. Finally, a time-synchronized feature vector is output.

[0024] The process of performing wavelet denoising and time alignment to output feature vectors specifically includes: first, constructing a multi-scale wavelet denoising unit; performing an 8-level decomposition of the original signal based on the Daubechies-4 wavelet basis; extracting low-frequency components from the 5th to the 8th levels for reconstruction; and outputting a purified signal sequence. Then, using a cross-correlation function, the optimal delay time between the coal feeding command and the combustion response is calculated, and the purified signal sequence is time-aligned according to the optimal delay time to compensate for the transmission lag from coal powder preparation to combustion response, outputting a time-synchronized feature vector. For wavelet denoising, this embodiment selects the Daubechies-4 (db4) wavelet basis as the mother wavelet because it has a good balance between compact support and regularity: Compact support ensures the computational efficiency of the algorithm, making it suitable for processing real-time data sampled at high frequencies in DCS systems.

[0025] Regularity ensures the smoothness of the reconstructed signal and avoids signal distortion.

[0026] In practice, the multi-scale wavelet denoising unit decomposes the original signal into eight layers (the decomposition process divides the signal frequency into different frequency bands). Since the key features of the combustion process are mainly concentrated in the low-frequency band, while on-site noise is mostly high-frequency random interference, this embodiment extracts the low-frequency approximate components from layers 5 to 8 for signal reconstruction, filtering out high-frequency detail components to output a purified signal sequence. This processing method effectively removes high-frequency noise introduced by sensor jitter, electromagnetic interference, etc., while retaining effective information reflecting the changing trend of the calorific value of the coal fed into the furnace.

[0027] To address the timing alignment process, due to the physical transmission lag between the issuance of the coal feeding command and the combustion response in the furnace of a coal-fired power plant, directly using coal feed data and combustion product data (such as flue gas oxygen content) collected at the same time as model input will lead to a severe timing mismatch problem. This embodiment uses a cross-correlation function to quantify this lag time. Specifically, a sequence of coal feeding commands and a sequence of combustion responses (such as main steam pressure or furnace temperature) under steady-state operating conditions are selected, and the cross-correlation coefficient between the two at different time offsets is calculated. When the cross-correlation coefficient reaches its maximum value, the corresponding time offset is the optimal delay time τ. According to field measurement data, the optimal delay time τ for units of different capacities and pulverizing systems is typically between 30 and 120 seconds. For example, for a 300MW unit using a medium-speed coal mill, τ may be approximately 45 seconds. Based on the calculated optimal delay time τ, the system performs a translation operation on the purification signal sequence, delaying the input signals such as coal feed by τ sampling periods, thereby achieving precise alignment of the input features and output response on the time axis and outputting a time-synchronized feature vector. This step ensures the causal consistency of input and output data over physical time during subsequent model training and inference, significantly improving the model's recognition accuracy.

[0028] The final output feature vector contains the key parameters required for combustion optimization. Coal feed rate, mill current, and mill inlet / outlet differential pressure directly reflect the amount of coal fed into the furnace and the energy consumption level of the pulverizing system; mill inlet air temperature, mill outlet air temperature, and primary air volume are used to calculate the heat balance of the pulverizing system; secondary air volume, flue gas oxygen content, and exhaust gas temperature reflect the combustion state in the furnace; and unit load serves as an operating condition indicator to distinguish combustion characteristics at different operating points. These parameters together constitute a multi-dimensional feature vector, providing a high-quality data foundation for subsequent deep coupling calculations across two channels. Step 200: Construct the mechanism model channel and the BP neural network channel. The mechanism model channel calculates the coarse identification calorific value and outputs dynamic parameters. The BP neural network channel takes the feature vector as input to obtain the decomposed dynamic characteristic error components and coal quality characteristic error components, and embeds physical consistency constraints in the loss function of the BP neural network channel.

[0029] The mechanistic model channel is a static mechanistic model built based on the heat balance equation of the pulverizing system. It is used to calculate the coarse calorific value of the coal fed into the furnace and output the dynamic parameters of the first-order inertial plus pure hysteresis model. The coarse calorific value is a preliminary estimate of the calorific value of the coal fed into the furnace, obtained by back-calculation from the heat balance equation. The dynamic parameters include the inertial time constant and the pure delay time. The inertial time constant reflects the response speed of the pulverizing system to changes in input, while the pure delay time reflects the transmission lag of pulverized coal from preparation to combustion. The BP neural network channel is a data-driven model that uses feature vectors as input and obtains dynamic characteristic error components and coal quality characteristic error components through a dual-output structure. The dynamic characteristic error components characterize the deviation between the parameters of the first-order inertial plus pure hysteresis model of the mechanistic model and the actual dynamic characteristics of the pulverizing system. The coal quality characteristic error components characterize the steady-state error of the mechanistic model caused by unconventional coal quality changes in the coal fed into the furnace. The physical consistency constraint refers to embedding a penalty term in the loss function of the BP neural network channel to force the fused calorific value to remain within the preset physical calorific value boundary.

[0030] This step is one of the core innovations of this embodiment. Unlike the serial architecture in existing technologies where "the calculation results of the mechanism model are directly used as the input of the BP network," this embodiment constructs a dual-channel architecture in parallel with the mechanistic model channel and the BP neural network channel. Specifically: The mechanism model channel is based on physical laws such as the heat balance equation of the pulverizing system. By collecting data such as the inlet air temperature, outlet air temperature, primary air volume and coal feed rate of the coal mill, it back-calculates the coarse calorific value of the coal entering the furnace. At the same time, it calculates the inertial time constant and pure delay time of the first-order inertial plus pure lag model as dynamic parameters to reflect the dynamic response characteristics of the pulverizing system.

[0031] The input layer of the BP neural network channel receives feature vectors, and the hidden layer adopts a three-layer structure. It can utilize its powerful nonlinear fitting ability to explore the complex relationships hidden in the feature vectors. Meanwhile, its output layer is designed as a dual-output layer containing a first output node and a second output node, which outputs the dynamic characteristic error component and the coal quality characteristic error component respectively, providing a clear physical guide for subsequent fine correction.

[0032] In particular, this embodiment does not simply add the results of the two channels together. Instead, it embeds a physical consistency constraint (composed of a weighted sum of a data fitting loss term, a physical consistency penalty term, and an output smoothing regularization term) into the loss function of the BP neural network channel. This forces the output of the neural network to be within a physically reasonable boundary range, thereby achieving deep coupling between the two channels at the algorithm level and effectively avoiding the problem of output divergence in sparse sample regions of a purely data-driven model.

[0033] Step 300: Establish a two-way correction mechanism. In each control cycle, the dynamic parameters are used as input features of the BP neural network channel to correct the error components online with dynamic characteristics, and multi-source uncertainties are fused to output a calorific value feature information package.

[0034] The bidirectional correction mechanism refers to establishing a forward propagation channel and a backward correction channel to achieve closed-loop interaction between the dynamic parameters of the mechanistic model and the error components of the BP neural network. Specifically, the forward propagation channel uses the dynamic parameters of the first-order inertial plus pure time-delay model calculated from the mechanistic model as the input features of the BP neural network channel. The backward correction channel receives the dynamic characteristic error components output by the BP neural network channel, and after filtering and gain processing, updates the inertial time constant of the mechanistic model channel online. Simultaneously, multiple uncertainties include input spatial distance uncertainties, model parameter identification uncertainties, and sensor noise uncertainties.

[0035] In some implementations, a forward propagation channel can be established first, using the inertial time constant and pure delay time from the first-order inertial plus pure time-delay model calculated by the mechanistic model channel as input features to the BP neural network channel. Then, a reverse correction channel is established to receive the dynamic characteristic error components output by the BP neural network channel. After filtering by a first-order low-pass filter, the correction step size is obtained by multiplying it by the feedback gain coefficient, and the inertial time constant of the mechanistic model channel is updated online using a recursive formula. Within each control cycle, the updated inertial time constant is input again into the BP neural network channel through the forward propagation channel, forming a closed bidirectional interactive loop.

[0036] Simultaneously, multiple uncertain factors are integrated: the spatial distance uncertainty is calculated based on the Mahalanobis distance between the current feature vector and the training sample set; the model parameter identification uncertainty is calculated based on the variance estimation of the output layer weights; and the sensor noise uncertainty is calculated based on the real-time signal-to-noise ratio estimation. After weighted fusion of the three factors to obtain the comprehensive uncertainty, the coarse identification calorific value output by the mechanism model channel is added to the total error correction amount output by the BP neural network channel to obtain the final fused calorific value. The calorific value feature information package is then output by combining the comprehensive uncertainty, physical consistency deviation, dynamic feature intensity, and coal quality feature change intensity.

[0037] It is important to note that this step further deepens the interaction between the two channels. Since the bidirectional correction mechanism includes both forward propagation and backward correction processes, in the forward propagation, the dynamic parameters calculated by the mechanistic model channel are used as one of the input features of the BP neural network channel. This allows the neural network to perceive the dynamic response characteristics of the current pulverizing system, thereby more accurately separating the error components. Simultaneously, in the backward correction, the dynamic characteristic error components output by the BP neural network channel are used to correct the parameters of the mechanistic model online, achieving adaptive adjustment of the model parameters to equipment state drift.

[0038] Furthermore, this step integrates multiple uncertainties such as input spatial distance, model parameter identification, and sensor noise to generate a calorific value feature information package containing the final fused calorific value, comprehensive uncertainty, physical consistency deviation, dynamic characteristic intensity, and coal quality characteristic change intensity. This calorific value feature information package not only provides the numerical value of the calorific value but also its reliability and operating condition characteristics, becoming a key link between the identification layer and the decision-making layer.

[0039] Step 400: Calculate the risk aversion coefficient based on the calorific value characteristic information package to obtain the efficiency prediction uncertainty.

[0040] The risk aversion coefficient is a dynamically calculated coefficient based on the physical consistency deviation, dynamic characteristic intensity, and coal quality characteristic change intensity in the calorific value characteristic information package. It is used to quantify the degree of conservatism the decision-making level should adopt when the current calorific value identification result is unreliable. Efficiency prediction uncertainty refers to the degree of uncertainty in the predicted boiler efficiency value, obtained by multiplying the calorific value uncertainty by the partial derivative of the efficiency model with respect to the calorific value. Physical consistency deviation characterizes whether the current fused calorific value is close to or exceeds the physically reasonable range. Dynamic characteristic intensity characterizes the severity of the dynamic response under the current operating conditions. Coal quality characteristic change intensity characterizes the degree of change in the quality of the coal entering the furnace.

[0041] In some implementations, this step breaks away from the limitations of traditional deterministic optimization. It extracts three feature components from the calorific value feature information package: physical consistency deviation, dynamic feature intensity, and coal quality feature change intensity. These components are multiplied by preset weighting coefficients, summed, and then multiplied by a baseline risk aversion coefficient to calculate the risk aversion coefficient for the current control cycle. When uncertainty is high or physical consistency deviation is large, the risk aversion coefficient increases accordingly, indicating lower reliability of the current calorific value identification result, thus introducing greater conservatism in subsequent optimizations.

[0042] Simultaneously, the comprehensive uncertainty (characterizing the reliability of the current calorific value identification result) in the calorific value feature information package is obtained. Using the boiler efficiency model, the partial derivative of the efficiency prediction value with respect to the calorific value is calculated. The comprehensive uncertainty is multiplied by the square of this partial derivative, and the square root is taken to obtain the efficiency prediction uncertainty. This obtained efficiency prediction uncertainty will then serve as the input to the optimization objective function in subsequent optimization decisions, enabling the decision-making level to perceive the impact of the uncertainty in the calorific value identification result on efficiency prediction.

[0043] Step 500: Based on the uncertainty of efficiency prediction, set optimization objectives and adaptively adjust the optimization cycle and search space constraint boundaries.

[0044] The optimization objective refers to the objective function that we hope to maximize or minimize during the air-coal ratio optimization process. In this step, the optimization objective is the difference between the expected value of boiler efficiency and the risk aversion coefficient multiplied by the efficiency prediction uncertainty. The optimization cycle refers to the time interval between two air-coal ratio optimization calculations, which is adaptively adjusted according to the intensity of dynamic characteristics. The search space constraint boundary refers to the feasible value range of each decision variable during the air-coal ratio optimization process, and its tightness is adaptively adjusted according to the physical consistency deviation. Decision variables include total air volume, secondary damper opening, burnout air opening, and excess air coefficient.

[0045] In some implementations, the optimization objective is first set as maximizing the expected value of boiler efficiency minus the risk aversion coefficient multiplied by the efficiency prediction uncertainty, while simultaneously minimizing the nitrogen oxide emission concentration, thus constituting a multi-objective optimization problem. The optimization period can then be calculated based on the dynamic characteristic intensity in the calorific value characteristic information package. The baseline optimization period is then divided by the product of the dynamic characteristic intensity and the first sensitivity coefficient; in this case, the greater the dynamic characteristic intensity, the shorter the optimization period.

[0046] Then, the search space constraint boundary is adjusted according to the physical consistency deviation in the calorific value characteristic information package. The nominal constraint boundary is multiplied by the product of the physical consistency deviation and the third sensitivity coefficient. The larger the physical consistency deviation, the looser the constraint boundary is to reserve a safety margin.

[0047] Finally, the search space range is adjusted based on the intensity of coal quality characteristic changes in the calorific value characteristic information package. The baseline search range is multiplied by the product of the intensity of coal quality characteristic changes and the second sensitivity coefficient. The greater the intensity of coal quality characteristic changes, the wider the search range is to cover more possible optimal solutions.

[0048] It should be noted that this step embodies the feature-driven integrated optimization approach. The optimization strategy is adaptively adjusted based on the uncertainty of efficiency prediction. For example, when uncertainty is high, the optimization cycle can be automatically shortened to accelerate response speed, or the search space constraint boundary can be narrowed to ensure operational safety. Furthermore, the setting of the optimization objective incorporates a risk aversion coefficient, balancing high efficiency with risk control, achieving a synergistic match between the optimization strategy and the heat value identification status.

[0049] Step 600: Use a multi-objective optimization algorithm to solve for the optimal air-coal ratio parameter set and output the execution command.

[0050] Among them, the multi-objective optimization algorithm refers to an optimization algorithm that can simultaneously handle multiple conflicting optimization objectives and solve for the Pareto optimal solution set. This step adopts the improved NSGA-II algorithm. The optimal air-fuel ratio parameter set refers to the optimal combination of total air volume, secondary damper opening, burnout air opening, and excess air coefficient selected from the Pareto front solution set.

[0051] In some implementations, total air volume, secondary damper opening, burnout air opening, and excess air coefficient are used as decision variables, with their values ​​set within the adaptively adjusted search space constraint boundary. This allows for maximizing the expected boiler efficiency, subtracting the risk aversion coefficient, multiplying by the efficiency prediction uncertainty, and minimizing nitrogen oxide emission concentration as two optimization objectives. Therefore, the improved NSGA-II algorithm can be used for multi-objective optimization.

[0052] The algorithm initializes the population size to 100 individuals, each representing a set of wind-coal ratio parameters. After 50 generations of evolution, the mutation probability is dynamically adjusted in each generation based on the dynamic feature intensity. The Pareto front solution set is preserved through non-dominated sorting and crowding distance calculation.

[0053] After optimization, a compromise solution is selected from the Pareto front solution set as the optimal air-coal ratio parameter set. The command values ​​of total air volume, secondary damper opening, burnout air opening and excess air coefficient are output to the control system of the coal-fired power generation unit through incremental distribution to complete closed-loop control.

[0054] Based on the above technical solutions, a complete process from data acquisition, dual-channel deep coupling identification, feature-driven decision-making to closed-loop execution was constructed. On the one hand, by constructing a dual-channel deep coupling architecture of mechanism model and BP neural network, physical consistency constraints are embedded in the neural network training process, effectively avoiding model output divergence and direction conflict, and solving the problem of model output conflict. On the other hand, through calorific value feature information package, the uncertainty information of the identification layer is transmitted to the decision layer, realizing integrated optimization based on risk perception, effectively solving the problem of information gaps in the decision layer. At the same time, combined with bidirectional correction mechanism and closed-loop feedback, the system can adaptively calibrate the dynamic parameter drift caused by equipment aging, improving the combustion optimization control accuracy and system robustness of coal-fired power generation units under complex operating conditions such as coal quality fluctuations and variable loads.

[0055] Example 2: like Figure 3-5 As shown, based on the above embodiments, this embodiment elaborates on the specific implementation process of the dual-channel deep coupling mechanism. This mechanism is the core of the present invention for achieving the co-evolution of the mechanistic model and the data model. Through the forward propagation of dynamic parameters and the reverse correction of error components, a closed bidirectional interactive loop is constructed.

[0056] Specifically, in step 200, the process of constructing the mechanism model channel and the BP neural network channel includes the following detailed steps: The mechanistic model channel constructs a static mechanistic model based on the heat balance equation of the pulverizing system. This model utilizes the principle of energy conservation to perform inverse calculations of the heat balance equation from real-time data of the DCS system. Specifically, by collecting parameters such as the inlet and outlet air temperatures and primary air volume of the coal mill, and combining them with the coal feed rate and the power consumption of the coal mill, the coarse calorific value of the coal entering the furnace is calculated. Simultaneously, the mechanistic model channel also calculates the dynamic parameters of the first-order inertial plus pure lag (FOPDT) model within the pulverizing system. These dynamic parameters include the inertial time constant and the pure delay time. Here, the inertial time constant reflects the response speed of the pulverizing system to changes in input, while the pure delay time reflects the transport lag of pulverized coal from preparation to combustion.

[0057] Among them, the inverse calculation of the heat balance equation refers to the method of calculating the calorific value of coal fed into the furnace by substituting real-time measurable thermal parameters into the heat balance equation. The first-order inertial plus pure time delay model is a dynamic characteristic transfer function used to describe the time delay and slow change process of the output response of a pulverizing system after a change in input. The inertial time constant is the time required for the system output response to reach 63.2% of its steady-state value. The pure time delay is the lag time between the issuance of the input command and the start of the system output response.

[0058] It should be noted that the calculation of this static mechanism model relies on the pulverizing system being in a relatively stable thermodynamic state. The thermal balance assumptions may deviate during the start-up and shutdown of the coal mill or during significant load changes. The coarsely identified calorific value serves as a physical baseline output, providing a constraint reference for subsequent data models. The dynamic parameters of the first-order inertial plus pure time-delay model are initially calibrated offline using a least-squares identification algorithm based on step response test or normal operation data.

[0059] The BP neural network channel adopts a dual-output structure. Within the current control cycle, the dynamic parameters (inertia time constant and pure delay time) calculated by the mechanistic model channel are used as input features of the BP neural network channel, enabling the neural network to perceive the dynamic response characteristics of the current pulverizing system. The output layer of its BP neural network channel includes a first output node and a second output node, which output the dynamic characteristic error component and the coal quality characteristic error component, respectively. The dynamic characteristic error component is used to characterize the deviation between the first-order inertia plus pure lag model parameters of the mechanistic model channel and the actual dynamic characteristics of the pulverizing system. For example, when the wear of the steel balls in the coal mill causes an increase in system inertia, this component can sensitively capture this shift in dynamic characteristics. The coal quality characteristic error component is used to characterize the steady-state error of the mechanistic model caused by unconventional changes in the coal quality of the coal fed into the furnace, such as the calorific value deviation caused by sudden changes in coal type.

[0060] Among them, error component decomposition refers to forcibly separating the total calorific value identification error into two parts, dynamic response deviation and coal quality characteristic deviation, in the output layer according to the different generation mechanisms.

[0061] In some implementations, feature dimension expansion can be performed on the input layer of the BP neural network channel. The inertial time constant and pure delay time calculated by the mechanism model channel within the current control cycle are used as new input features and concatenated with the preprocessed feature vector to form the complete input vector of the network, enabling the neural network to perceive the dynamic response characteristics of the current pulverizing system.

[0062] A dual-output structure with a first output node and a second output node is set up in the output layer. The output of the first output node is used to characterize the dynamic characteristic error component, while the output of the second output node is used to characterize the coal quality characteristic error component. Simultaneously, during the network training phase, the difference between the true value of the coal quality analysis and the coarsely identified calorific value is used as the total error target. An error backpropagation algorithm is employed to synchronously update the network weight parameters corresponding to the two output nodes, giving the error component decomposition a clear physical direction.

[0063] It should be noted that the key to the dual-output structure lies in guiding the two output nodes to converge to error subspaces with clear physical meanings during the training phase through a loss function. The dynamic characteristic error component is sensitive to the dynamic response deviation of the pulverizing system, but tends to zero under steady-state conditions; the coal quality characteristic error component is sensitive to changes in coal type switching or blending ratios, but remains relatively stable during dynamic adjustments. This separation of the two components allows the subsequent bidirectional correction mechanism to specifically utilize the dynamic characteristic error component to back-calibrate the mechanism model parameters.

[0064] During the training phase of the BP neural network channel, the difference between the true value of coal quality analysis and the coarsely identified calorific value output by the mechanism model channel is used as the total error target. The error backpropagation algorithm is used to update the weight parameters corresponding to the first and second output nodes.

[0065] Among them, the true value of coal quality testing refers to the standard measured value of the received basis lower heating value of coal obtained through offline sampling and laboratory analysis. The total error target refers to the total deviation between the expected output value and the actual output value that the BP neural network needs to fit during training. The error backpropagation algorithm is a supervised learning algorithm that propagates the output layer error layer by layer along the network connection weights and updates the weight parameters of each layer using the gradient descent method. Weight parameters are adjustable coefficients connecting neurons in each layer of the neural network, determining the transmission strength and transformation mode of the input signal in the network.

[0066] In some implementations, a training sample set is constructed, and real-time DCS data and feature vectors corresponding to the coal quality test sampling time during historical runtime are collected. At the same time, the coarse identification calorific value calculated by the mechanism model channel at that time is recorded, and the difference between the true value of coal quality test and the coarse identification calorific value is calculated as the total error target of the BP neural network channel.

[0067] During the forward propagation of the network, the feature vectors and dynamic parameters passed from the mechanism model are input into the network. After nonlinear transformation in the hidden layer, the first and second output nodes in the output layer generate dynamic characteristic error components and coal quality characteristic error components, respectively. The sum of these two components constitutes the total error correction of the network prediction. The mean squared error loss between the total error correction of the network output and the total error target is then calculated. An error backpropagation algorithm is used to propagate the loss gradient backward from the output layer to the input layer. According to the chain rule, the gradient values ​​of the corresponding connection weights of the first and second output nodes are calculated respectively. The weight parameters of each layer are updated according to a preset learning rate so that the network output gradually approximates the true error distribution.

[0068] It is important to note that the key to the training phase lies in the fact that the total error objective simultaneously includes both dynamic characteristic deviation and coal quality characteristic deviation. The network, through a dual-output structure driven by the loss function, automatically learns to separate these two components. After training, the network can output both error components synchronously based solely on the real-time feature vectors during the inference phase, without requiring the true coal quality analysis data.

[0069] During the inference stage of the BP neural network channel, the feature vector is input into the BP neural network channel, and the dynamic characteristic error component and the coal quality characteristic error component are output simultaneously. The sum of the two is used as the total error correction output, which is added to the coarse identification calorific value to obtain a more accurate fused calorific value.

[0070] Among them, the fused calorific value refers to the final calorific value identification result obtained by adding the coarse identification calorific value output by the mechanism model channel and the total error correction amount output by the BP neural network channel.

[0071] During the inference phase, the feature vector within the current control cycle is concatenated with the inertial time constant and pure delay time transmitted through the mechanism model channel to form a complete network input. Inside the trained BP neural network channel, forward propagation calculation is performed based on the concatenated input vector. After receiving the vector, the input layer passes it forward layer by layer through weighted summation of each neuron in the hidden layer and nonlinear transformation with the activation function, ultimately reaching the output layer containing the first and second output nodes.

[0072] The first output node calculates and outputs the dynamic characteristic error component based on the weight parameters learned during its training phase. The second output node simultaneously calculates and outputs the coal quality characteristic error component. The output values ​​of the two nodes are then algebraically added to obtain the total error correction. The total error correction is then added to the coarsely identified calorific value output by the mechanism model channel to obtain the fused calorific value, which serves as the final identification result of the calorific value of the coal fed into the furnace in the current cycle and is output to the calorific value feature information package.

[0073] Furthermore, this embodiment establishes a bidirectional correction mechanism to achieve adaptive calibration of model parameters. This mechanism includes a forward propagation channel and a backward correction channel. The forward propagation channel uses the dynamic parameters (including the inertial time constant and pure delay time) calculated by the mechanistic model channel as input features to the BP neural network channel, a process described in the above steps. The backward correction channel receives the dynamic characteristic error components output by the BP neural network channel, filters them through a first-order low-pass filter, multiplies them by the feedback gain coefficient to obtain the correction step size, and updates the inertial time constant of the mechanistic model channel online.

[0074] Specifically, the reverse correction channel first inputs the dynamic characteristic error component into a first-order low-pass filter for filtering to remove high-frequency noise interference. The filtered error component is then multiplied by a feedback gain coefficient to obtain the correction step size. This feedback gain coefficient is typically small, for example, between 0 and 0.1, to ensure the stability of the correction process.

[0075] The inertial time constant of the mechanism model channel is updated online using this corrected step size. The update formula is:

[0076] in The updated inertial time constant, The inertial time constant before the update. For feedback gain coefficient, This indicates a low-pass filter operation. This represents the dynamic characteristic error component.

[0077] In this way, when the characteristics of the powder-making system change (such as equipment aging or wear), the dynamic characteristic error components captured by the BP neural network channel can drive the parameter adjustment of the mechanism model channel in reverse, making it closer to the real physical process.

[0078] Finally, before the next control cycle, the updated inertial time constant is input again into the BP neural network channel through the forward propagation channel, forming a closed bidirectional interactive loop. This loop enables the coordinated adaptive adjustment of dynamic parameters and neural network input features under conditions of coal type switching or changes in pulverizing system characteristics.

[0079] For example, when the wear of the steel balls in the coal mill leads to an increase in the system's inertial time constant, the reverse correction channel will gradually increase the inertial time constant parameter in the mechanism model. At the same time, the updated parameter serves as the input feature of the BP neural network channel, helping the neural network to more accurately separate dynamic errors from coal quality errors, thereby improving the overall identification accuracy.

[0080] Based on the above technical solution, by constructing a composite loss function with physical consistency constraints, during the training process of the BP neural network channels, the data fitting loss term, physical consistency penalty term, and output smoothing regularization term are weighted and summed using preset weight coefficients. This achieves a soft constraint on the physical boundary of the neural network output, effectively solving the technical problem of output divergence in sparse training sample regions such as coal type switching, and conflict with the output direction of the mechanism model, which is a problem of traditional pure data-driven models. Specifically, when the fused calorific value exceeds the preset physical calorific value boundary, the hinge loss function generates a penalty gradient proportional to the excess amount, forcing the network weights to update in the direction of reducing the excess. At the same time, the output smoothing regularization term, by penalizing the L2 norm of the time gradient of the error correction amount, suppresses the drastic time jump of the calorific value identification result. Ultimately, the error correction output of the two channels is always kept within a physically reasonable range, eliminating non-physical oscillation phenomena and improving the accuracy and stability of calorific value identification.

[0081] Example 3: like Figure 6 As shown, based on the above embodiments, this embodiment elaborates in detail the specific form of the composite loss function embedding physical consistency constraints in the BP neural network channel: A composite loss function for the BP neural network channels is constructed. The composite loss function consists of a data fitting loss term, a physical consistency penalty term, and an output smoothing regularization term, which are weighted and summed using preset weight coefficients.

[0082] Traditional BP neural network training typically focuses solely on data fitting accuracy, i.e., minimizing the error between predicted and true values. However, in calorific value identification scenarios, purely data-driven models are prone to outputting divergent results in areas with insufficient training sample coverage (such as coal type mutation conditions), leading to final calorific values ​​exceeding physically reasonable ranges (e.g., calculating negative calorific values ​​or calorific values ​​far exceeding the coal type's limits), thus creating a directional conflict with the output of the mechanistic model.

[0083] This embodiment solves this problem fundamentally by constructing a composite loss function and introducing soft constraints from physical laws during training. The mathematical expression of the composite loss function is:

[0084] in, This is the composite loss function for the channels of the BP neural network, which is the total loss target that needs to be minimized during training. The preset weighting coefficient for the physical consistency penalty term is used to adjust the influence of physical constraints on the total loss. This is a physical consistency penalty term used to constrain the fusion heat value to remain within a preset physical boundary; The preset weighting coefficients for the output smoothing regularization term are used to adjust the influence of temporal smoothing constraints on the total loss. This is a smoothing regularization term for the output, used to suppress drastic fluctuations in the network output over time. The data fitting loss term is expressed in the form of mean squared error (MSE), and its calculation formula is as follows: ,in This refers to the number of samples in the training batch, i.e., the number of input feature vectors used in this loss calculation. This represents the true value of the coal quality analysis. This is the fused calorific value (i.e., the sum of the coarsely identified calorific value and the total error correction amount of the BP network output). This term ensures the model's ability to fit historical data.

[0085] The data fitting loss term is calculated based on the mean square error between the fused calorific value output by the BP neural network channel and the true value of coal quality analysis.

[0086] Among them, the true value of coal quality testing refers to the standard measured value of the received basis lower calorific value of coal obtained through offline sampling and laboratory analysis.

[0087] In some implementations, this step is performed during the training phase of the BP neural network channel. First, feature vectors and coarse-identified calorific values ​​corresponding to the coal quality analysis sampling time within the historical runtime are collected. The difference between the true coal quality analysis value and the coarse-identified calorific value is then calculated as the total error target. The feature vectors are input into the BP neural network channel, and after nonlinear transformation in the hidden layer, dynamic characteristic error components and coal quality characteristic error components are generated in the output layer. The sum of these two components is used as the total error correction amount, which is added to the coarse-identified calorific value to obtain the fused calorific value. Then, the mean square error between the fused calorific value and the true coal quality analysis value is calculated. This is achieved by squaring the difference for each sample in the batch, summing the results, and then dividing by the number of samples. Finally, this mean square error is used as the data fitting loss term and incorporated into the composite loss function. The gradient is then calculated layer by layer using the error backpropagation algorithm, and the network weights are updated accordingly.

[0088] The physical consistency penalty term is calculated based on the degree to which the sum of the coarse identification calorific value output by the mechanistic model channel and the error correction amount output by the BP neural network deviates from the preset physical calorific value boundary. It adopts the form of hinge loss function. When the fused calorific value is within the preset physical calorific value boundary, the penalty term is zero. When the fused calorific value exceeds the preset physical calorific value boundary, the penalty term is proportional to the amount of excess.

[0089] The preset physical calorific value boundary refers to a reasonable calorific value range pre-set based on the received basis lower heating value range of commonly used coal types in coal-fired power generating units. The hinge loss function is a piecewise linear loss function characterized by zero loss when the input value is within the specified range, and a loss proportional to the amount of excess when the input value exceeds the range. The amount of excess refers to the absolute value of the difference between the fused calorific value and the upper or lower limit of the preset physical calorific value boundary. The penalty gradient refers to the backpropagation gradient signal of the physical consistency penalty term on the update of the BP neural network channel weights.

[0090] The physical consistency penalty term is the core innovation of this embodiment. This term adopts the form of a hinge loss function, and its calculation logic is as follows: First, a preset physical calorific value boundary is set. This boundary is determined based on the physical properties of coal commonly used in coal-fired power generation units, such as setting a lower limit. It is 18 MJ / kg, with an upper limit. It is 28 MJ / kg. When the fusion calorific value... Falling on [ , When the output is within the specified interval, it is assumed to conform to physical laws, and a penalty term is applied. It is zero, of which To integrate calorific value, To roughly identify the calorific value, This is the total error correction amount; When the fusion heat value When the value exceeds this range, the penalty is proportional to the amount exceeding the range, and the calculation formula is as follows: This is equivalent to building a "soft wall" at the physical boundary: when the neural network tries to output a correction amount that causes the heat value to go out of bounds, the loss function will generate a huge penalty gradient, forcing the network weights to be updated in the direction of reducing the degree of going out of bounds during backpropagation.

[0091] Compared to simple hard truncation, the hinged loss function preserves the continuity of gradients, making network training smoother and easier to converge.

[0092] The output smoothing regularization term is calculated based on the L2 norm of the time gradient of the error correction amount of the BP neural network channel output.

[0093] Here, the time gradient refers to the difference between the total error correction of the BP neural network channel output in the current control cycle and the output value of the previous control cycle. The L2 norm refers to the Euclidean norm of the time gradient vector, which is the square root of the sum of the squares of its elements.

[0094] A smoothing regularization term is introduced to prevent drastic jumps in the output of the BP neural network over time. Since the calorific value of the coal fed into the furnace is a continuously changing physical quantity and will not experience instantaneous abrupt changes, this constraint is included. Its calculation formula is as follows: That is, to calculate the L2 norm of the gradient of the error correction amount output by the BP neural network in the time dimension.

[0095] in, This refers to the number of samples in the training batch, i.e., the number of input feature vectors involved in this loss calculation; The time gradient of the error correction is specifically defined as the total error correction amount output by the BP neural network channel in the current control cycle. The L2 norm symbol represents the gradient over time. Find the Euclidean norm.

[0096] This method filters out high-frequency noise interference by penalizing drastic changes in output values ​​at adjacent time points, thus ensuring the temporal stability of the calorific value identification results.

[0097] During the forward and backward propagation iterations of the BP neural network channels, the composite loss function is calculated synchronously, and the preset weight coefficients are updated with the goal of minimizing the composite loss function.

[0098] Forward propagation refers to the computational process where the input feature vector is passed from the input layer of the BP neural network through the hidden layers, weighted and summed layer by layer, and nonlinearly transformed by the activation function, finally reaching the output layer. Backpropagation refers to the process of passing the gradient of the composite loss function with respect to the network output layer by layer from the output layer back to the input layer, and calculating the gradient of the connection weights of each layer according to the chain rule. Simultaneous computation refers to the complete process in which forward propagation is performed first in the same iteration to obtain the values ​​of each loss term, and then weighted and summed to form the composite loss function, before backpropagation is performed to update the network weights.

[0099] In each training iteration of the BP neural network channel, the feature vector is first input into the network, weighted and summed in the input layer, and then passed to the hidden layer. A nonlinear transformation is performed through the activation function, and the vector is then passed to the output layer to obtain the dynamic characteristic error component and the coal quality characteristic error component. Subsequently, based on the output value, the data fitting loss term, the physical consistency penalty term, and the output smoothing regularization term are calculated simultaneously. These three losses are then weighted and summed using preset weight coefficients to form a composite loss function.

[0100] The goal is to minimize the composite loss function. The gradient of the composite loss function with respect to each node of the output layer is calculated. Then, the gradient is backpropagated to the hidden layer and the input layer by layer using the chain rule. The gradient values ​​of the connection weights of each layer are calculated respectively. Then, the connection weights of each layer are updated along the negative gradient direction according to the preset learning rate, so that the value of the composite loss function gradually decreases.

[0101] It should be noted that the preset weighting coefficients and The importance of this factor in balancing various losses can be adjusted according to actual operating conditions. For example, in operating conditions with frequent fluctuations in coal quality, the factor can be appropriately increased. To strengthen physical constraints; in noisy operating conditions, the strength can be appropriately increased. This enhances smoothness. Through the constraints of this composite loss function, the BP neural network channels can not only fit the data but also follow physical laws, effectively avoiding model output divergence and non-physical oscillations, thus achieving deep coupling between the mechanistic model and the data model.

[0102] Based on the above technical solutions, a composite loss function is constructed, comprising a data fitting loss term, a physical consistency penalty term, and an output smoothing regularization term. This function is embedded in the BP neural network training process to address the technical problem of output divergence and conflict with the output direction of the mechanistic model in purely data-driven models during sparse sample regions such as coal type switching. Specifically, when the fused calorific value exceeds the preset physical calorific value boundary, the hinge loss function generates a penalty gradient proportional to the excess amount, forcing the network weights to update in the direction of reducing the out-of-bounds, thereby constraining the output within a physically reasonable range and avoiding non-physical oscillations. Simultaneously, the output smoothing regularization term suppresses drastic temporal jumps in the calorific value identification results by penalizing the L2 norm of the time gradient of the error correction amount. These technical means together achieve deep coupling between the mechanistic model and the data model, improving the accuracy and stability of calorific value identification and ensuring that the system can still output a reliable calorific value benchmark under complex operating conditions.

[0103] Example 4: like Figure 7-8 As shown, based on the above embodiments, this embodiment elaborates in detail the fusion process of multi-source uncertain factors and the specific method for generating calorific value feature information packages. This is a key step in realizing the leap from "data-driven" to "risk-awareness-driven" in this invention, providing a risk basis for subsequent optimization decisions by quantifying the credibility of the identification results.

[0104] Specifically, the process of integrating multiple uncertainties includes: input spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty.

[0105] For input spatial distance uncertainty factors, the spatial distance uncertainty value is calculated based on the Mahalanobis distance between the current feature vector and the training sample set of the BP neural network channel, which is used to characterize the sparsity of the current working condition in the training sample space.

[0106] It should be noted that, for uncertainties in the input space distance, this embodiment calculates the distance based on the Mahalanobis distance between the current feature vector and the training sample set of the BP neural network channels. This Mahalanobis distance can consider the correlation between features and effectively measure the position of the current operating point in the historical training sample space. The specific calculation formula is as follows:

[0107] in It is an exponential function. It is used to map the square of the Mahalanobis distance onto the interval (0,1], such that the greater the distance, the closer the uncertainty value is to 1. The square of the Mahalanobis distance from the current feature vector to the center of the training sample set. The standard deviation squared (variance) of the training sample set. σdist represents the spatial distance uncertainty. When this distance is large, it indicates that the current working condition is relatively rare and is located in a sparse region of training samples. In this case, σdist increases, which indicates that the spatial distance uncertainty increases and the reliability of the model output decreases.

[0108] For identifying uncertainties in model parameters, the uncertainty value of model parameter identification is calculated based on the variance estimation of the weights of the output layer of the BP neural network channel, which is used to characterize the confidence level of the model parameters themselves.

[0109] It should be noted that, for identifying uncertainties in model parameters, this embodiment calculates the variance based on the variance estimation of the output layer weights of the BP neural network channel.

[0110] Furthermore, during neural network training, the variance of the weights reflects the degree of certainty the model has regarding the mapping relationship of specific features. The specific calculation formula is as follows: ,in It is a variance operator used to quantify the dispersion or uncertainty of the weights in the output layer of a neural network; This is the weight matrix of the output layer of the BP neural network. This matrix contains all the connection weights from the last hidden layer to the output layer. Identify uncertainties in the model parameters. This value characterizes the confidence level of the BP neural network model parameters themselves. If this value is large, it indicates that the output layer weights of the model fluctuate significantly, and the mapping relationship between the model and the current input features is not sufficiently certain. In this case, the reliability of the heat value identification result is low.

[0111] For sensor noise uncertainty, the sensor noise uncertainty value is calculated based on the real-time signal-to-noise ratio estimation of real-time data from the DCS system, which is used to characterize the reliability of the measurement data.

[0112] It should be noted that, regarding the uncertainty of sensor noise, this embodiment calculates the signal-to-noise ratio based on real-time data from the DCS system. Field sensors are affected by electromagnetic interference, wear, and other factors, and their measurement data often contains noise. The specific calculation formula is as follows: ; σ_sensor = σ_noise / σ_signal, where σ_noise is the estimated noise standard deviation and σ_signal is the signal standard deviation. When the signal-to-noise ratio decreases, the sensor noise uncertainty increases, indicating a decline in the reliability of the original data. This is the sensor noise uncertainty value. This value characterizes the reliability of the measurement data. A larger value indicates stronger noise in the signal and lower reliability of the original data; a smaller value indicates a clearer signal and more reliable data. The estimated noise standard deviation is a quantitative estimate of the intensity of random noise mixed into the real-time data of the DCS system, and is a statistic used to quantify the amplitude of such fluctuations. The standard deviation of the signal is a quantitative estimate of the variation amplitude of the effective signal itself in the real-time data of the DCS system. It can reflect the natural fluctuation range of the measured physical quantity (such as coal feed, air volume, temperature, etc.) under normal operating conditions. The uncertainties of spatial distance, model parameter identification, and sensor noise are weighted and fused to obtain the comprehensive uncertainty. The weighting coefficients are adaptively adjusted according to the unit load change rate to enhance the perception of uncertainty under dynamic operating conditions.

[0113] In some implementations, the weighting coefficients are not fixed but are adaptively adjusted according to the rate of change of unit load. For example, under dynamic operating conditions where the unit load changes rapidly, the system automatically increases the weight of the uncertain value of the input spatial distance to enhance the sensitivity to operating condition drift; Under steady-state conditions, the weight of sensor noise uncertainty is appropriately increased to address fluctuations in the measurement data. This adaptive weighting strategy effectively enhances the ability to perceive uncertainties under dynamic conditions.

[0114] After calculating the overall uncertainty, this embodiment generates a calorific value characteristic information package. Specifically, the calorific value characteristic information package includes the final fused calorific value, overall uncertainty, physical consistency deviation, dynamic characteristic intensity, and coal quality characteristic change intensity.

[0115] The final fused calorific value is obtained by adding the coarsely identified calorific value output from the mechanistic model channel to the total error correction amount output from the BP neural network channel, representing the best estimate of the calorific value of the coal fed into the furnace at the current moment.

[0116] The overall uncertainty, obtained by weighted fusion of input spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty, represents the confidence interval of the estimate.

[0117] Physical consistency deviation is calculated based on the sum of the coarsely identified calorific value and the total error correction amount, indicating the degree of deviation from the preset physical calorific value boundary. It is used to characterize whether the current fused calorific value is close to or exceeds the physically reasonable range.

[0118] Specifically: if the fusion calorific value is within the preset boundary (e.g., 18-28 MJ / kg), the deviation is zero or a small value; if it is close to or exceeds the boundary, the deviation value increases. This component is used to characterize whether the current fusion calorific value is close to or exceeds the physically reasonable range, and is a key indicator for judging whether the model output is "out of line".

[0119] The dynamic characteristic intensity is the absolute value of the dynamic characteristic error component output by the BP neural network channel, and is used to characterize the severity of the dynamic response under the current operating condition.

[0120] The dynamic characteristic intensity is taken as the absolute value of the dynamic characteristic error component output by the BP neural network channel. As in the aforementioned embodiment, the dynamic characteristic error component reflects the deviation of the dynamic response of the pulverizing system, and therefore its absolute value directly characterizes the severity of the dynamic response under the current operating condition.

[0121] For example, this intensity value will increase significantly when the coal mill is started or stopped or when the coal feed rate is adjusted significantly.

[0122] The intensity of coal quality characteristic change is taken as the absolute value of the coal quality characteristic error component output by the BP neural network channel, and is used to characterize the degree of change in the coal quality entering the furnace.

[0123] When the coal type is switched or the blending ratio changes, this intensity value will rise rapidly, indicating that the coal quality characteristics of the system have changed significantly.

[0124] The calorific value feature information package generated through the above process is not merely a package of data, but rather makes explicit implicit information such as confidence level, physical compliance, and dynamic characteristics from the identification process. This information package serves as a unifying link, completely transmitting the underlying calorific value identification results to the upper-level optimization decision-making module. This enables the optimization process to perceive risks and identify operating conditions, thereby making safer and more reasonable optimization decisions.

[0125] Example 5: like Figure 9 As shown, based on the above embodiments, this embodiment elaborates on the closed-loop feedback and feature-driven model self-calibration process. This embodiment aims to solve the problem of model accuracy degradation caused by equipment aging or long-term operating condition drift, and achieves long-term adaptive evolution of model parameters by establishing a feedback mechanism based on actual operating results.

[0126] Specifically, after the execution instruction is output in step 600, the system does not end the process, but enters the closed-loop feedback stage.

[0127] After executing the command, the system collects the actual boiler efficiency and calculates the efficiency prediction deviation. The actual boiler efficiency can be calculated using the inverse balancing method, and the efficiency prediction deviation is the absolute value of the difference between the predicted efficiency and the actual efficiency. This deviation directly reflects the degree to which the current model fits the actual combustion process.

[0128] Based on efficiency prediction deviations and calorific value characteristic information, the BP neural network channels are selectively updated. This selective triggering mechanism avoids system oscillations caused by frequent updates, intervening only when necessary. Specifically, it includes the following two triggering logics: The first triggering mechanism targets physical consistency deviation: when the physical consistency deviation exceeds the first preset threshold and the duration exceeds the first preset time window, incremental retraining of the BP neural network channel is triggered.

[0129] For example, if the first preset threshold is set to 0.5 MJ / kg and the first preset time window is 10 minutes, if the physical consistency deviation in the calorific value feature information package exceeds 0.5 MJ / kg for 10 consecutive minutes, it indicates that the output of the current BP neural network continues to deviate from the physical reasonable range, and online correction alone is no longer sufficient to correct it.

[0130] At this point, the system will automatically extract recent operational data and combine it with the latest coal quality test results (if available) to incrementally retrain the BP neural network, adjusting the network weights to reconverge it to a physically reasonable range. This effectively solves the problem of model failure caused by long-term changes in coal type.

[0131] The second triggering mechanism targets dynamic characteristic drift: when the moving average of the dynamic characteristic intensity continuously exceeds the second preset time window and is non-zero, and the efficiency prediction deviation exceeds the second preset threshold, the recursive least squares online correction of the inertial time constant of the motor model channel is triggered.

[0132] For example, if the second preset time window is set to 30 minutes and the second preset threshold is 1.5%, and the dynamic characteristic intensity remains non-zero for 30 minutes (meaning there is a continuous deviation in the dynamic characteristics of the pulverizing system), and the efficiency prediction deviation exceeds 1.5%, it indicates that the inertial time constant in the mechanism model can no longer match the actual response characteristics of the current equipment (e.g., increased inertia due to wear of the steel balls in the coal mill). In this case, the system will activate the Recursive Least Squares (RLS) algorithm, using the dynamic characteristic intensity as input, to identify and correct the inertial time constant online, achieving adaptive calibration for gradual drift such as equipment aging. The correction formula is:

[0133] in, The inertial time constant is updated after the kth control cycle; Let be the inertial time constant for the (k-1)th control cycle; The Kalman gain vector for the kth period is used to determine the correction weight of the new measurement on the parameter estimate, balancing the influence of old information and new measurement. This is the actual output measurement value for the k-th cycle; The parameters of the previous period Prediction of the current output.

[0134] When the triggering condition is not met, keep the current parameters of the BP neural network channel and the mechanism model channel unchanged.

[0135] This reflects the principle of system stability. When the operating conditions are stable and the model accuracy meets the requirements, unnecessary parameter adjustments should be avoided to prevent the introduction of human disturbance.

[0136] Thus, through the aforementioned closed-loop feedback and feature-driven self-calibration mechanism, this embodiment ensures that the system can maintain high-precision identification and optimization capabilities during long-term operation, solving the problem of the decline in accuracy of traditional open-loop models over time.

[0137] Example 6: like Figure 10 As shown, this embodiment provides a system for online identification of the calorific value of coal fed into the furnace and optimization of the air-coal ratio. This system is applied to the method for online identification of the calorific value of coal fed into the furnace and optimization of the air-coal ratio in any of the above embodiments. The system specifically includes a communication unit and a processing unit.

[0138] The communication unit is the interface between the system and the field physical equipment. It communicates with the DCS system of the coal-fired generator set to collect real-time data from the DCS system of the coal-fired generator set and transmit the real-time data from the DCS system to the processing unit.

[0139] In practical industrial applications, the communication unit can take the form of an industrial Ethernet card, an OPC (OLE for Process Control) client interface, or a serial communication interface module. It can establish a data connection with the distributed control system (DCS) of a coal-fired power plant via fieldbus or industrial Ethernet, and read analog signals such as coal feed rate, mill current, mill inlet and outlet differential pressure, mill inlet air temperature, mill outlet air temperature, primary air volume, secondary air volume, flue gas oxygen content, exhaust gas temperature, and unit load in real time according to a preset sampling period (e.g., once every 5 seconds).

[0140] The acquired raw industrial signals are then converted into a digital signal format that the processing unit can recognize, and preliminary protocol parsing and verification are performed to ensure the integrity and real-time performance of data transmission.

[0141] It should be understood that communication units can be wired connections, and in some renovation projects, wireless communication modules that support wireless transmission protocols (such as LoRa and 5G) can also be used, as long as the bandwidth and latency requirements for data transmission can be met.

[0142] The processing unit is the core computing carrier of the system, used to construct the mechanistic model channel and the BP neural network channel. The mechanistic model channel calculates the coarse identification calorific value and outputs dynamic parameters. The BP neural network channel takes the feature vector as input to obtain the decomposed dynamic characteristic error components and coal quality characteristic error components, and embeds physical consistency constraints into the loss function of the BP neural network channel. A two-way correction mechanism is established. In each control cycle, the dynamic parameters are used as input features of the BP neural network channel to correct the error components online with dynamic characteristics, and multi-source uncertainties are fused to output the calorific value feature information package. The risk aversion coefficient is calculated based on the calorific value feature information package to obtain the efficiency prediction uncertainty. Based on the efficiency prediction uncertainty, the optimization objective is set, and the optimization cycle and search space constraint boundary are adaptively adjusted. A multi-objective optimization algorithm is used to solve for the optimal air-coal ratio parameter set and outputs the execution instructions.

[0143] Specifically, the processing unit typically consists of hardware devices such as high-performance industrial control computers, programmable logic controllers (PLCs), or edge computing servers. It integrates a memory and a processor; the memory stores the computer program, and the processor executes this program to implement the aforementioned method steps.

[0144] To more efficiently handle complex algorithmic logic, the processing unit is logically divided into multiple collaborative modules: The mechanism model calculation module is used to execute the functions of the mechanism model channel in the above embodiments. This module incorporates the heat balance equations of the pulverizing system and the first-order inertial plus pure delay (FOPDT) model algorithm. It receives real-time data transmitted from the communication unit, calculates the coarse calorific value of the coal fed into the furnace, and extracts dynamic parameters such as the inertial time constant and pure delay time. This module provides a digital representation of the physical mechanism, offering the system a benchmark output that conforms to physical laws.

[0145] The neural network calculation module performs the functions of the BP neural network channel in the above embodiments. This module deploys a trained BP neural network model, with its output layer designed as a dual-output structure. It can effectively receive feature vectors and dynamic parameters output by the mechanism model calculation module, and calculate the dynamic characteristic error components and coal quality characteristic error components through forward propagation.

[0146] In particular, this module is also responsible for calculating the composite loss function during training and inference, and constraining the network output through a physical consistency penalty term to ensure that the results conform to the physical boundaries.

[0147] The uncertainty fusion module performs the function of fusing multiple uncertainties in the above embodiments. This module calculates the input spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty in real time, and adaptively adjusts the weighting coefficients according to the unit load change rate, outputting the comprehensive uncertainty. This module is key to the system's "risk perception" capability, making implicit errors in the underlying calculation process explicit.

[0148] The optimization decision-making module is used to execute the integrated optimization decision-making function described in the above embodiments. This module receives a calorific value feature information packet jointly generated by the neural network calculation module and the uncertainty fusion module. Based on components such as physical consistency deviation and dynamic feature intensity, it calculates a risk aversion coefficient, and then dynamically adjusts the optimization objective function and search space constraint boundaries. Finally, the module calls a multi-objective optimization algorithm (such as the improved NSGA-II algorithm) to solve for the optimal wind-coal ratio parameter set, and sends execution instructions to the DCS system via the communication unit.

[0149] Through the deep integration of the aforementioned hardware architecture and software modules, this embodiment constructs a closed-loop control system with perception, cognition, and decision-making capabilities. The communication unit ensures the breadth and real-time nature of data sources, while the modular design within the processing unit enables efficient decoupling and parallel computation of complex algorithmic logic. This architecture not only supports the stable operation of the dual-channel deep coupling and feature-driven optimization method described in the previous embodiment but also possesses excellent scalability. For example, when a new optimization objective (such as reducing NOx emissions) needs to be introduced, only the corresponding objective function logic needs to be added to the optimization decision module without modifying the underlying data acquisition and identification module, demonstrating the flexibility and adaptability of the system architecture.

[0150] Example 8: To more intuitively demonstrate the practical application effect of the technical solution of this invention, this embodiment takes a 300MW coal-fired power generating unit as the application object and describes in detail an operational example under typical complex conditions such as coal type switching and equipment aging. This unit is equipped with a medium-speed coal mill, and the coal type it usually burns is high-volatile bituminous coal, but it needs to be blended or switched to lean coal during specific periods.

[0151] Scenario 1: Stability verification of calorific value identification under coal type switching conditions.

[0152] During unit operation, when dispatch instructions require switching from high-volatile bituminous coal to lean coal, in the traditional approach, due to the lag in the response of the mechanistic model and the divergence of the BP neural network output in the sample sparse region, the calorific value identification results exhibit severe "sawtooth" fluctuations within 5-8 minutes after the switch, with fluctuation amplitudes as high as 18%, seriously interfering with the optimal wind-coal ratio decision-making.

[0153] In applying the method of this embodiment, in step 200, a physical consistency constraint is embedded in the loss function of the BP neural network channel. When the BP network attempts to output a correction amount exceeding the preset physical calorific value boundary (18-28 MJ / kg), the hinge loss function generates a penalty gradient, forcing the network output to revert to the physically reasonable range. Thus, during the same coal type switching process, the final fused calorific value output by this invention transitions smoothly, the calorific value fluctuation amplitude is compressed, and non-physical oscillations are effectively eliminated, providing a stable and reliable calorific value benchmark for subsequent optimization.

[0154] Scenario 2: Verification of the model's adaptive capability under aging equipment conditions.

[0155] After three months of continuous operation, wear on the steel balls in the coal mill caused a drift in the inertial time constant of the pulverizing system. In traditional solutions, fixed model parameters lead to a gradual increase in calorific value identification errors. However, with the method described in this embodiment, the bidirectional correction mechanism in step 300 comes into play. This allows the dynamic characteristic error component output by the BP neural network channel to accurately detect deviations in the dynamic response of the pulverizing system. Furthermore, by using a recursive formula, the inertial time constant of the mechanistic model channel is corrected online through a reverse correction channel. After a period of adaptive adjustment, the dynamic parameters of the mechanistic model successfully tracked the equipment state drift, stabilizing the calorific value identification error. This solves the problem of decreased accuracy due to equipment aging and avoids frequent manual intervention by operators.

[0156] Scenario 3: Adaptive adjustment and verification of optimization strategies based on uncertainty.

[0157] During periods of rapid load fluctuations in the generating unit, traditional solutions often suffer from mismatched air-fuel ratio adjustments due to lag in calorific value identification, resulting in a phenomenon of "black smoke escaping followed by high oxygen levels." By applying the method of this embodiment, step 300 integrates multi-source uncertainties to output a calorific value characteristic information package, where the overall uncertainty changes in real time with fluctuations in the operating conditions. Simultaneously, in steps 400 and 500, the optimization decision module dynamically adjusts its strategy based on this overall uncertainty. This means that when uncertainty is high, the risk aversion coefficient is automatically increased, the search space constraint boundary is narrowed, and the optimization cycle is shortened to accelerate the response speed. Therefore, under varying load conditions, the control strategy can adaptively adjust based on the calorific value reliability, reducing the decline in boiler efficiency during load changes and improving the system's robustness and safety.

[0158] It should be understood that the above-mentioned 300MW unit operation example is only used to explain the specific implementation process and effects of the technical solution of the present invention, and does not constitute a limitation on the scope of protection of the present invention. Those skilled in the art can apply the present invention to coal-fired power generating units of different capacities such as 600MW and 1000MW, or to other solid fuel combustion control scenarios, and can achieve similar technical effects.

[0159] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention, such as replacing the BP neural network with other types of deep learning models, or making equivalent transformations to the specific functional form of the physical consistency constraint, should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method based on online identification of the calorific value of coal fed into the furnace and optimization of the air-coal ratio, characterized in that, include: Real-time data from the DCS system of coal-fired power generating units is collected, wavelet denoising and time-series alignment are performed, and feature vectors are output. A mechanism model channel and a BP neural network channel are constructed. The mechanism model channel calculates the coarse identification calorific value and outputs dynamic parameters. The BP neural network channel takes the feature vector as input to obtain the decomposed dynamic characteristic error component and coal quality characteristic error component, and embeds physical consistency constraints in the loss function of the BP neural network channel. A two-way correction mechanism is established. In each control cycle, the dynamic parameters are used as input features of the BP neural network channel. The error components are corrected online using the dynamic characteristics. Multi-source uncertainties are fused together to output a calorific value feature information package. The risk aversion coefficient is calculated based on the calorific value characteristic information package to obtain the efficiency prediction uncertainty. Based on the aforementioned efficiency prediction uncertainty, an optimization objective is set, and the optimization cycle and search space constraint boundaries are adaptively adjusted. A multi-objective optimization algorithm is used to solve for the optimal air-coal ratio parameter set and output the execution instructions.

2. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The process of performing wavelet denoising and temporal alignment to output feature vectors specifically includes: A multi-scale wavelet denoising unit is constructed to decompose the original signal based on the Daubechies-4 wavelet basis, extract low-frequency components for reconstruction, and output a purified signal sequence. The optimal delay time between the coal feeding command and the combustion response is calculated using a cross-correlation function. The purification signal sequence is then time-aligned based on the optimal delay time to compensate for the transmission lag from coal powder preparation to combustion response, and a time-synchronized feature vector is output. The feature vector includes at least: coal feed rate, pulverizer current, pulverizer inlet and outlet differential pressure, pulverizer inlet air temperature, pulverizer outlet air temperature, primary air volume, secondary air volume, flue gas oxygen content, exhaust gas temperature, and unit load.

3. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The calculation process of the mechanism model channel specifically includes: A static mechanism model based on the heat balance equation of the pulverizing system is constructed. The heat balance equation is back-calculated on the real-time data of the DCS system to obtain the coarse calorific value of the coal fed into the furnace. The dynamic parameters of the first-order inertial plus pure delay model in the pulverizing system are calculated. The dynamic parameters include the inertial time constant and the pure delay time. Within the current control cycle, the dynamic parameters are used as input features of the BP neural network channel, and the dynamic characteristic error components from the previous control cycle are received. Based on the dynamic characteristic error components of the previous control cycle, the inertial time constant is corrected online using a recursive formula, wherein the correction step size is determined by the product of the feedback gain coefficient and the dynamic characteristic error components processed by the low-pass filter.

4. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The specific form of the loss function for the physical consistency constraint includes: Construct a composite loss function for the BP neural network channel. The composite loss function consists of a data fitting loss term, a physical consistency penalty term, and an output smoothing regularization term, which are weighted and summed using preset weight coefficients. The data fitting loss term is calculated based on the mean square error between the fused calorific value output by the BP neural network channel and the true value of the coal quality test. The physical consistency penalty term is calculated based on the degree to which the sum of the coarse identification calorific value output by the mechanism model channel and the error correction amount output by the BP neural network channel deviates from the preset physical calorific value boundary. It adopts the form of a hinge loss function. When the fused calorific value is within the preset physical calorific value boundary, the penalty term is zero. When the fused calorific value exceeds the preset physical calorific value boundary, the penalty term is proportional to the amount of excess. The output smoothing regularization term is calculated based on the L2 norm of the time gradient of the error correction amount of the BP neural network channel output; During the forward and backward propagation iterations of the BP neural network channel, the composite loss function is calculated synchronously, and the preset weight coefficients are updated with the goal of minimizing the composite loss function.

5. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The process of decomposing the dynamic characteristic error component and the coal quality characteristic error component specifically includes: The output layer of the BP neural network channel is a dual-output layer containing a first output node and a second output node. The first output node outputs the dynamic characteristic error component, and the second output node outputs the coal quality characteristic error component. The dynamic characteristic error component is used to characterize the deviation between the first-order inertial plus pure hysteresis model parameters of the mechanism model channel and the dynamic characteristics of the actual pulverizing system. The coal quality characteristic error component is used to characterize the steady-state error of the mechanism model caused by unconventional coal quality changes in the coal fed into the furnace. During the training phase of the BP neural network channel, the difference between the true value of coal quality analysis and the coarsely identified calorific value output by the mechanism model channel is used as the total error target, and the error backpropagation algorithm is employed. The weight parameters corresponding to the first output node and the second output node are updated. During the inference stage of the BP neural network channel, the feature vector is input into the BP neural network channel, and the dynamic characteristic error component and the coal quality characteristic error component are output synchronously, and the sum of the two is output as the total error correction amount.

6. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The bidirectional correction mechanism specifically includes: A forward propagation channel is established, and the dynamic parameters of the first-order inertial plus pure time delay model calculated by the mechanism model channel are used as the input features of the BP neural network channel. The dynamic parameters include the inertial time constant and the pure time delay. A reverse correction channel is established to receive the dynamic characteristic error component output by the BP neural network channel. After filtering by a first-order low-pass filter, the correction step size is obtained by multiplying it by the feedback gain coefficient, and the inertial time constant of the mechanism model channel is updated online. Before the next control cycle, the updated inertial time constant is input again into the BP neural network channel through the forward transmission channel to form a closed bidirectional interactive loop, which is used to realize the coordinated adaptive adjustment of dynamic parameters and neural network input features under the conditions of coal type switching or changes in the characteristics of the pulverizing system.

7. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The process of integrating multi-source uncertainties specifically includes: Multi-source uncertainties include input spatial distance uncertainties, model parameter identification uncertainties, and sensor noise uncertainties; For the input spatial distance uncertainty factor, the spatial distance uncertainty value is calculated based on the Mahalanobis distance between the current feature vector and the training sample set of the BP neural network channel, which is used to characterize the sparsity of the current working condition in the training sample space. For the uncertainties in identifying the model parameters, the uncertainty value of identifying the model parameters is calculated based on the variance estimation of the weights of the output layer of the BP neural network channel, which is used to characterize the confidence level of the model parameters themselves. For the sensor noise uncertainty factor, the sensor noise uncertainty value is calculated based on the real-time signal-to-noise ratio estimation of the real-time data of the DCS system, which is used to characterize the reliability of the measurement data. The uncertainties in spatial distance, model parameter identification, and sensor noise are weighted and fused to obtain a comprehensive uncertainty. The weighting coefficients are adaptively adjusted according to the unit load change rate to enhance the perception of uncertainty under dynamic operating conditions.

8. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, The calorific value characteristic information package specifically includes the final fused calorific value, comprehensive uncertainty, physical consistency deviation, dynamic characteristic intensity, and coal quality characteristic change intensity. The final fused calorific value is obtained by adding the coarse identification calorific value output by the mechanism model channel to the total error correction amount output by the BP neural network channel; The overall uncertainty is obtained by weighted fusion of input spatial distance uncertainty, model parameter identification uncertainty, and sensor noise uncertainty. The physical consistency deviation is calculated based on the sum of the coarse identification calorific value and the total error correction amount, and the degree to which it deviates from the preset physical calorific value boundary. It is used to characterize whether the current fused calorific value is close to or exceeds the physical reasonable range. The dynamic feature intensity is taken as the absolute value of the dynamic characteristic error component output by the BP neural network channel, and is used to characterize the intensity of the dynamic response under the current working condition. The intensity of the change in coal quality characteristics is taken as the absolute value of the coal quality characteristic error component output by the BP neural network channel, which is used to characterize the degree of change in the coal quality entering the furnace.

9. The method for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio according to claim 1, characterized in that, Also includes: After executing the execution command, the actual boiler efficiency is collected, and the efficiency prediction deviation is calculated. Based on the efficiency prediction deviation and the calorific value feature information package, the BP neural network channel update is selectively triggered; Specifically, when the physical consistency deviation exceeds a first preset threshold and the duration exceeds a first preset time window, incremental retraining of the BP neural network channel is triggered. When the moving average of the dynamic feature intensity continuously exceeds the second preset time window non-zero, and the efficiency prediction deviation exceeds the second preset threshold, the recursive least squares online correction of the inertial time constant of the mechanism model channel is triggered. When the triggering condition is not met, the current parameters of the BP neural network channel and the mechanism model channel remain unchanged.

10. A system for online identification of calorific value of coal fed into the furnace and optimization of the air-coal ratio, characterized in that, The system applied to the method for online identification of calorific value of coal fed into the furnace and optimization of air-coal ratio as described in any one of claims 1-9 specifically includes: a communication unit and a processing unit; The communication unit is connected to the DCS system of the coal-fired power generator set and is used to collect real-time data from the DCS system of the coal-fired power generator set and transmit the real-time data of the DCS system to the processing unit. The processing unit is used to construct a mechanism model channel and a BP neural network channel. The mechanism model channel calculates the coarse identification calorific value and outputs dynamic parameters. The BP neural network channel takes the feature vector as input to obtain the decomposed dynamic characteristic error component and coal quality characteristic error component, and embeds physical consistency constraints in the loss function of the BP neural network channel. A two-way correction mechanism is established. In each control cycle, the dynamic parameters are used as input features of the BP neural network channel. The error components are corrected online using the dynamic characteristics. Multi-source uncertainties are fused together to output a calorific value feature information package. The risk aversion coefficient is calculated based on the calorific value characteristic information package to obtain the efficiency prediction uncertainty. Based on the aforementioned efficiency prediction uncertainty, an optimization objective is set, and the optimization cycle and search space constraint boundaries are adaptively adjusted. A multi-objective optimization algorithm is used to solve for the optimal air-coal ratio parameter set and output the execution instructions.

Images