Multi-modal optimization system for combustion efficiency of thermal power boiler

Abstract

The invention relates to the field of heat energy engineering and automatic control, and discloses a multi-mode optimization system for combustion efficiency of a thermal power boiler. The system comprises a multi-modal data perception and space-time alignment module, a tensor manifold modeling and physical constraint feature extraction module, a space-time coupling dynamic prediction and uncertainty quantification module, a quantum optimization decision and DCS cooperative control module and a combustion state derivative early warning and optimization feedback module. Through multi-modal data space-time alignment, five-order tensor physical constraint modeling, PDE deep network prediction, quantum optimization decision and a closed-loop feedback mechanism, space-time unified fusion and physical feature extraction of combustion data are realized, the reliability of combustion state prediction is improved, an optimal control instruction is efficiently solved, system parameters are dynamically corrected, and the reliability of combustion state prediction is improved. The problems that in the prior art, data fusion is difficult, modeling physical constraints are lacked, optimization real-time performance is poor, and adaptivity is weak are solved, and the combustion efficiency and the intelligent control level of the thermal power boiler are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of thermal energy engineering and automation control technology, and in particular to a multi-modal optimization system for combustion efficiency of thermal power boilers. Background Art

[0002] Current thermal power boiler combustion efficiency optimization technology mainly relies on distributed sensor networks to collect multi-source data, and uses independent modules to process flame images, furnace temperature fields, flue gas composition and other information. The traditional control system then generates adjustment instructions based on simplified mathematical models or empirical formulas. This type of technology usually stores and processes each modal data independently, and only achieves limited optimization through manually preset thresholds or fixed rules. The overall operating characteristics are "data collection, model simplification, and open-loop control."

[0003] Traditional control strategies mostly rely on rule-based logic and empirical models, which are difficult to cope with multi-source disturbances and high-dimensional dynamic characteristics in actual operating conditions. In addition, boiler operation data has strong time-varying, strong coupling, and high-dimensional heterogeneous characteristics. Single modal data cannot fully reflect the combustion state, and there are significant temporal and spatial offsets between physical quantities such as flame images, temperature fields, and flue gas components, which seriously affect the data fusion and model training effects. The neural networks or fuzzy control algorithms introduced in some studies generally have problems with generalization ability and poor physical consistency. Existing optimization control methods are mostly based on classic algorithms such as genetic algorithms and particle swarm optimization. The solution space search efficiency is low and the calculation time is long, which makes it difficult to meet real-time control needs. They are prone to failure in environments with strong dynamics and uncertain boundaries of the combustion process, and may even cause system fluctuations.

[0004] Although some systems have basic data acquisition and DCS linkage functions, they lack online correction and adaptive feedback mechanisms for model errors, resulting in a decrease in control accuracy drifting over time. Since the "black box" characteristics of boiler combustion have not been effectively cracked, actual operation still relies on operator experience and adjustment. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a multimodal optimization system for the combustion efficiency of thermal power boilers, aiming to solve the problems of difficulty in spatiotemporal fusion of multi-source data, lack of physical constraints in combustion modeling, poor real-time performance of optimization algorithms and weak system adaptability.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-modal optimization system for combustion efficiency of thermal power boilers, comprising:

[0007] The multimodal data perception and spatiotemporal alignment module is used to collect multi-source combustion data, align the spatiotemporal references of heterogeneous data based on a spatiotemporal mapping function, and output a multimodal data stream in a unified coordinate system.

[0008] The tensor manifold modeling and physical constraint feature extraction module constructs a fifth-order combustion state tensor based on multimodal data streams, embeds combustion dynamics constraints through pre-trained core tensors, and extracts low-dimensional physical features by updating the factor matrix online;

[0009] The spatiotemporal coupled dynamic prediction and uncertainty quantification module inputs the extracted low-dimensional physical features into the PDE constrained deep network, combines time series modeling with the attention mechanism to predict the combustion state and generate prediction confidence intervals;

[0010] The quantum optimization decision-making and DCS collaborative control module builds a multi-objective optimization model based on prediction results and confidence intervals, solves the optimal control instructions through quantum computing, and realizes the synchronous output of instructions to DCS signals through a protocol converter;

[0011] The combustion state derivative warning and optimization feedback module receives manifold modeling features and DCS feedback data, implements risk warning through feature mutation detection, dynamically modifies modeling parameters and optimization weights, and feeds the adjusted parameters back to the manifold modeling and quantum optimization module to form a closed-loop iterative mechanism.

[0012] Preferably, the multimodal data perception and spatiotemporal alignment module includes:

[0013] The flame vision acquisition unit is used to obtain image information of the burning flame inside the furnace and time-stamp the image frames;

[0014] Temperature field detection unit, used to collect temperature data at multiple points in the furnace and synchronize it with the flame image timestamp;

[0015] Flue gas composition analysis unit, used for real-time detection of O2, CO, NO in exhaust gas x Equal gas content and output corresponding time series data;

[0016] The operating parameter access unit is used to access the boiler operating parameters from the DCS system and unify all sensor data into the time and space alignment processing unit.

[0017] Preferably, the multimodal data perception and spatiotemporal alignment module further includes:

[0018] The time synchronization subunit is used to synchronize data based on the timestamp information of the raw data from various sensors using a precision clock protocol;

[0019] The spatial registration subunit is used to transform the spatial coordinates of the flame image, temperature sampling points and smoke sensor positions according to the preset spatial mapping function to build a unified spatial reference frame;

[0020] The data reconstruction subunit is used to fuse multi-source heterogeneous data that have completed spatiotemporal alignment into a multimodal data stream with a consistent time scale and spatial coordinate system, and output it to subsequent modules via high-speed Ethernet.

[0021] Preferably, the tensor manifold modeling and physical constraint feature extraction module includes:

[0022] A tensor construction unit, configured to reconstruct a multimodal data stream into a fifth-order combustion state tensor, wherein the tensor comprises a spatial dimension, a channel dimension, a combustion feature dimension, and a time dimension;

[0023] Core tensor pre-training unit, used to pre-train core tensors based on computational fluid dynamics simulation data and embed combustion dynamics constraints during the training process;

[0024] The online updating unit is used to update the factor matrix of tensor decomposition according to real-time combustion data and output low-dimensional manifold features representing the combustion state.

[0025] Preferably, the core tensor pre-training unit implements combustion dynamics constraint embedding through the following optimization objectives:

[0026]

[0027] in, is the core tensor after embedding the combustion dynamics constraints, It is a combustion field tensor constructed based on computational fluid dynamics simulation data; is the core tensor to be optimized; U(n) is the n-th order factor matrix; λ is the manifold smoothing coefficient and satisfies 0.1≤λ≤1.0; is the combustion field Laplace matrix; tr(·) represents the matrix trace operation.

[0028] Preferably, the spatiotemporal coupling dynamic prediction and uncertainty quantification module includes:

[0029] A feature input unit, used to receive the low-dimensional physical feature vector output by the tensor manifold modeling module;

[0030] Dynamic prediction unit, which includes a PDE-constrained deep network and a spatiotemporal attention mechanism to predict future combustion state parameters;

[0031] Uncertainty quantification unit, used to generate confidence intervals for prediction results through random dropout strategy.

[0032] Preferably, the PDE constrained deep network is trained using the following composite loss function:

[0033]

[0034] in, is the composite loss function of the deep network; α is the weight coefficient of the prediction error term; is the mean square error between the predicted value and the true value; β is the weight coefficient of the physical constraint term; K is the total number of combustion state parameters; is the predicted value of the kth combustion state parameter; is the time derivative of the prediction parameter; is the discretized differential operator of the combustion field control equation; is the spatial gradient operator; is the Laplace operator.

[0035] Preferably, the quantum optimization decision and DCS collaborative control module includes:

[0036] Optimization model building unit, used to transform prediction results and confidence intervals into multi-objective optimization problems;

[0037] A quantum annealing solver unit, which is used to encode the optimization problem into a QUBO model and solve the optimal control instructions through a quantum annealing machine;

[0038] The protocol conversion unit is used to convert the quantum optimization results into control signals that can be executed by the DCS system.

[0039] Preferably, the objective function construction method of the QUBO model is:

[0040]

[0041] Where H is the quantum annealing optimization objective function; a i is the combustion efficiency optimization coefficient of the i-th air-coal ratio valve; b ij is the coordinated regulation constraint coefficient of the i-th and j-th valves; γ is the balance coefficient; S target is the total number of target valves opened; N is the total number of air-coal ratio regulating valves.

[0042] Preferably, the combustion state derivative warning and optimization feedback module includes:

[0043] Manifold feature receiving unit, used to obtain the low-dimensional manifold feature vector and corresponding timestamp output by the tensor manifold modeling module in real time;

[0044] DCS feedback analysis unit, used to analyze the control instruction execution log and boiler operation status data returned by the DCS system;

[0045] A feature mutation detection unit is used to detect mutation points on the manifold feature vector based on the sliding window T test method, and generate a three-level warning signal when the statistic exceeds the threshold;

[0046] The parameter dynamic correction unit is used to calculate the core tensor update rate of the manifold modeling module and the weight adjustment amount of the quantum optimization module based on the warning level and DCS execution effect evaluation results.

[0047] The present invention provides a multi-modal optimization system for the combustion efficiency of thermal power boilers. It has the following beneficial effects:

[0048] 1. The present invention uses a multimodal data perception and spatiotemporal alignment module to achieve the collection of multi-source combustion data and the unification of spatiotemporal benchmarks, and output a standardized multimodal data stream. This technical solution allows the fusion of different sensor data in the same spatiotemporal coordinate system, solving the problem of data spatiotemporal asynchrony and difficulty in comprehensive utilization in traditional systems, and provides accurate data support for optimizing the combustion efficiency of thermal power boilers.

[0049] 2. The present invention uses tensor manifold modeling and physical constraint feature extraction modules to construct a fifth-order combustion state tensor and embed combustion dynamics constraints. At the same time, the factor matrix is ​​updated online to extract low-dimensional features. This solution enables the model to have physical prior knowledge and can adapt to changes in working conditions in real time. It overcomes the shortcomings of traditional modeling methods such as insufficient physical constraints and difficulty in dynamic updating, and helps to improve the accuracy and real-time performance of combustion state feature extraction of thermal power boilers.

[0050] 3. The present invention adopts a spatiotemporal coupled dynamic prediction and uncertainty quantification module to input low-dimensional physical features into a PDE-constrained deep network, and combines time series modeling with an attention mechanism to predict the combustion state and generate confidence intervals. This technology not only ensures that the prediction results conform to the laws of combustion dynamics, but also quantifies the uncertainty of the prediction. Compared with the problems of weak physical constraints and lack of reliability assessment in traditional prediction methods, it helps to improve the credibility and accuracy of combustion state prediction.

[0051] 4. The present invention adopts quantum optimization decision-making and DCS collaborative control modules to construct a multi-objective optimization model based on the prediction results, and uses quantum computing to quickly solve the optimal control instructions and synchronously output them to the DCS system. This solution can efficiently search for the optimal solution in a large-scale solution space, while achieving collaborative control with the DCS system, solving the problems of low search efficiency and poor linkage with the control system of traditional optimization algorithms, and providing technical support for real-time optimization control of the combustion efficiency of thermal power boilers. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the system architecture of the present invention;

[0053] Figure 2 Schematic diagram of the multimodal data perception and spatiotemporal alignment module of the present invention;

[0054] Figure 3 Schematic diagram of the tensor manifold modeling and physical constraint feature extraction module of the present invention;

[0055] Figure 4 This is a schematic diagram of the spatiotemporal coupling dynamic prediction and uncertainty quantification module of the present invention;

[0056] Figure 5 This is a schematic diagram of the quantum optimization decision and DCS collaborative control module of the present invention;

[0057] Figure 6 Schematic diagram of the combustion state derived warning and optimization feedback module of the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] Please see the attached Figure 1 -Attached Figure 6 An embodiment of the present invention provides a multimodal optimization system for the combustion efficiency of a thermal power boiler, including: a multimodal data perception and spatiotemporal alignment module for collecting multi-source combustion data, aligning the spatiotemporal references of heterogeneous data based on a spatiotemporal mapping function, and outputting a multimodal data stream in a unified coordinate system.

[0060] Specifically, the multimodal data perception and spatiotemporal alignment module includes a flame vision acquisition unit, a temperature field detection unit, a smoke composition analysis unit, an operating parameter access unit, a time synchronization subunit, a spatial registration subunit, and a data reconstruction subunit.

[0061] The flame vision acquisition unit is used to obtain image information of the burning flame inside the furnace and time-stamp the acquired image frames. The unit consists of an industrial-grade infrared high-speed camera device installed on the boiler furnace observation window. It periodically acquires flame image sequences and adds a time stamp t to each frame of the image. i .

[0062] The temperature field detection unit is responsible for collecting multi-point temperature data at different locations in the furnace, using multiple thermocouple sensors. The temperature data is uploaded in real time via the serial port protocol and is accompanied by a timestamp. i .

[0063] The flue gas composition analysis unit is used to detect O2, CO, and NO in exhaust gas in real time x The unit is equipped with an infrared spectrum probe and additional time information to measure gas content and output corresponding time series data.

[0064] The operating parameter access unit is connected to the boiler's DCS system through the OPC-UA protocol to obtain key operating parameters such as air-coal ratio, valve opening, boiler load, and organize them into a unified parameter vector.

[0065] The time synchronization subunit uses the IEEE1588 precision time protocol to achieve data synchronization and correct the timestamps of all sensors to ensure the timing consistency between data. The corrected timestamp is determined by the original timestamp and the offset.

[0066] The function of the spatial registration subunit is to map the data of different sensors to a unified spatial coordinate system, and to convert the position data of each sensor by calculating the preset spatial mapping function to ensure that the information collected by all sensors is in the same reference frame.

[0067] The data reconstruction subunit realizes the spatiotemporal alignment and fusion of different source data, and uses a sliding window mechanism to uniformly process time series data. The integrated data can be expressed as:

[0068] D t ={I t ,T t ,G t ,P t};

[0069] Among them, D t is the integrated data; I t is the flame image data; T t is the temperature data; G t is the detection value of smoke components; P t Boiler operating parameters.

[0070] Finally, the multimodal data flow can be expressed as:

[0071]

[0072] in, For multimodal data streams, is the multimodal data at time t1; n is the total number of data moments.

[0073] The data stream is sent to subsequent modules via industrial Ethernet. Each unit is connected through a switch and uses the ROS protocol for message interaction to ensure real-time performance.

[0074] The tensor manifold modeling and physical constraint feature extraction module constructs a fifth-order combustion state tensor based on multimodal data streams, embeds combustion dynamics constraints through pre-trained core tensors, and extracts low-dimensional physical features by updating the factor matrix online.

[0075] Specifically, this module is used to construct multimodal perception data into a fifth-order combustion state tensor, and pre-train the core tensor in combination with computational fluid dynamics simulation data. It improves the physical consistency of modeling by embedding combustion dynamics constraints, and realizes online update of the model factor matrix to support combustion state analysis and control optimization. This module includes a tensor construction unit, a core tensor pre-training unit, an online update unit and a physical constraint embedding mechanism.

[0076] The tensor construction unit reconstructs the multimodal perception data output by the multimodal data perception and spatiotemporal alignment modules into a fifth-order tensor structure to capture the spatiotemporal evolution and modal associations during the combustion process. The constructed fifth-order tensor is in the following form:

[0077]

[0078] in, is the three-dimensional combustion state tensor; T is the total number of time steps; M is the number of multimodal data channels; S is the number of spatial sampling points; Represents the real field tensor space.

[0079] Tensor construction uses a sliding window mechanism and linear interpolation strategy to ensure synchronous sampling and spatial alignment of data of each modality.

[0080] The core tensor pre-training unit constructs a reference combustion tensor TrefTref based on CFD simulation, and pre-trains the core tensor using Tucker decomposition. The combustion dynamics Laplace regularization term is introduced, and the optimization objective function is as follows:

[0081]

[0082] in, is the core tensor after embedding the combustion dynamics constraints, It is a combustion field tensor constructed based on computational fluid dynamics simulation data; is the core tensor to be optimized; U(n) is the n-th order factor matrix; λ is the manifold smoothing coefficient and satisfies 0.1≤λ≤1.0; is the combustion field Laplace matrix; tr(·) represents the matrix trace operation.

[0083] This unit realizes the modeling a priori capture of the physical structure of the combustion process and provides initialization parameters for subsequent online modeling.

[0084] The online update unit continuously adjusts the pre-trained model based on real-time perception data and updates the factor matrix through an incremental tensor decomposition algorithm to maintain the modeling results' responsiveness to changes in operating conditions. The optimized form of online update is as follows:

[0085]

[0086] Among them, U (n) is the (n)th order factor matrix to be updated; is the new observation tensor data in the current time window; η is the updated learning rate.

[0087] This mechanism ensures that the tensor model can respond dynamically when the data changes, keeping the feature expression consistent with the combustion state.

[0088] The spatiotemporal coupled dynamic prediction and uncertainty quantification module inputs the extracted low-dimensional physical features into the PDE constrained deep network, combines time series modeling with the attention mechanism to predict the combustion state and generate prediction confidence intervals;

[0089] Specifically, the spatiotemporal coupling dynamic prediction and uncertainty quantification module is mainly used to complete the prediction of future combustion states based on the low-dimensional physical features output by the tensor manifold modeling and physical constraint feature extraction module, combined with the deep network constrained by partial differential equations (PDEs), and quantify the uncertainty of the prediction, thereby providing a reliable basis for subsequent quantum optimization decisions and DCS collaborative control. This module includes a feature input unit, a dynamic prediction unit, and an uncertainty quantification unit.

[0090] The feature input unit receives low-dimensional physical feature vectors from the tensor manifold modeling and physical constraint feature extraction modules. The input data includes various physical state characteristics of the boiler combustion process, such as temperature field distribution, flue gas composition changes, and pulverized coal concentration. All feature vectors have a unified timestamp for subsequent spatiotemporal modeling.

[0091] The dynamic prediction unit is based on a PDE-constrained deep network, combined with a spatiotemporal attention mechanism, to predict future combustion states. The PDE constraints in the deep network are used to ensure that the prediction results conform to the physical laws of combustion dynamics. The PDE-constrained deep network is trained using the following composite loss function:

[0092]

[0093] in, is the composite loss function of the deep network; α is the weight coefficient of the prediction error term; is the mean square error between the predicted value and the true value; β is the weight coefficient of the physical constraint term; K is the total number of combustion state parameters; is the predicted value of the kth combustion state parameter; is the time derivative of the prediction parameter; is the discretized differential operator of the combustion field control equation; is the spatial gradient operator; is the Laplace operator.

[0094] The unit also incorporates a spatiotemporal attention mechanism to improve the ability to predict critical moments (such as flame instability periods) and critical areas (such as the center of the furnace), ensuring timely response to abnormal situations.

[0095] The uncertainty quantification unit performs multiple forward propagations by introducing a random dropout strategy, sampling the results of each prediction, and then quantifying the uncertainty of the prediction and calculating the confidence interval. The specific confidence interval formula is as follows:

[0096]

[0097] in, The predicted mean value at the current moment; is the prediction result of the jth forward propagation; is the mean of multiple prediction results; z α is the confidence coefficient under normal distribution; N is the number of sampling times.

[0098] By statistically analyzing the mean and variance of multiple prediction paths, the confidence interval of the prediction results is calculated and used as a constraint condition for the subsequent quantum optimization decision module. This quantization process enhances the stability and robustness of the prediction results and provides an acceptable control space range.

[0099] The quantum optimization decision-making and DCS collaborative control module builds a multi-objective optimization model based on prediction results and confidence intervals, solves the optimal control instructions through quantum computing, and realizes the synchronous output of instructions to DCS signals through a protocol converter;

[0100] Specifically, the quantum optimization decision and DCS collaborative control module constructs a quantum-solvable optimization control model based on the combustion state prediction results and uncertainty confidence intervals output by the spatiotemporal coupling dynamic prediction and uncertainty quantification module. With the help of the quantum annealing calculation method, it outputs the optimal air-coal ratio control instructions and converts them into control signals that can be recognized by the DCS system, realizing closed-loop linkage control of prediction-optimization-execution.

[0101] This module includes an optimization model building unit, a quantum annealing solution unit and a protocol conversion unit.

[0102] The optimization model construction unit receives the combustion parameter prediction values ​​and their confidence intervals output by the spatiotemporal coupling dynamic prediction module and the uncertainty quantification module, and on this basis constructs a multi-objective optimization model with the goal of maximizing combustion efficiency. The model comprehensively considers NO X Emissions, air-coal ratio regulation constraints and control command stability are considered, and the above problems are transformed into the form of quantum binary optimization problem (QUBO), with the objective function as follows:

[0103]

[0104] Where H is the quantum annealing optimization objective function; a i is the combustion efficiency optimization coefficient of the i-th air-coal ratio valve; b ij is the coordinated regulation constraint coefficient of the i-th and j-th valves; γ is the balance coefficient; S target is the total number of target valves opened; N is the total number of air-coal ratio regulating valves.

[0105] Confidence intervals are embedded in the optimization model as optimization constraint boundaries, enabling the quantum annealing algorithm to search for the optimal control solution within a safe and reliable range, balancing combustion efficiency and system stability. This structure introduces both physical and data constraints to the optimization solution, enhancing the interpretability and practical adaptability of the control strategy.

[0106] The quantum annealing solver inputs the aforementioned QUBO structure into the quantum annealing computing platform. Through quantum superposition and tunneling, it efficiently searches for the optimal control bit vector in an exponentially large solution space. The solution objectives are as follows:

[0107]

[0108] in, is the optimal control bit vector obtained by solving; argmin represents the independent variable that makes the objective function take the minimum value; is a vector of binary control variables; Optimize the objective function for quantum physics.

[0109] The unit is compatible with real quantum hardware interfaces and classical simulated annealing alternative mechanisms, ensuring that suboptimal solutions can be obtained and system availability is guaranteed when quantum computing resources are limited.

[0110] The protocol conversion unit is responsible for mapping the bit vector control solution output by the solver into actual control signals recognizable by the DCS system. This process uses an embedded protocol adapter module to convert binary formats into industrial communication protocols (such as Modbus and Profibus). Control commands can generate digital control signals or continuous analog signals on demand and maintain synchronous communication with the actuators of the boiler DCS system. In addition, the unit integrates a feedback comparison mechanism, which collects execution feedback and verifies the consistency of the command results. The feedback information is then transmitted back for model correction and subsequent scheduling optimization, establishing a closed-loop prediction-optimization-control path.

[0111] The combustion state derivative warning and optimization feedback module receives manifold modeling features and DCS feedback data, implements risk warning through feature mutation detection, dynamically modifies modeling parameters and optimization weights, and feeds the adjusted parameters back to the manifold modeling and quantum optimization module to form a closed-loop iterative mechanism.

[0112] Specifically, the combustion state derivative warning and optimization feedback module is used to monitor the tensor manifold modeling characteristics and DCS system feedback data in real time, identify combustion abnormality risks through the feature mutation detection mechanism, and dynamically adjust the manifold modeling parameters and quantum optimization weights to form a closed-loop iterative optimization mechanism to ensure that the system continues to adapt to changes in boiler combustion conditions.

[0113] The module includes a manifold feature receiving unit, a DCS feedback parsing unit, a feature mutation detection unit, a parameter dynamic correction unit and a closed-loop feedback execution unit.

[0114] The manifold feature receiving unit continuously obtains the low-dimensional manifold feature vector formula output by the tensor manifold modeling module:

[0115]

[0116] Among them, F t is the manifold eigenvector at time t; Represents a k-dimensional real space; k is the number of low-dimensional feature dimensions output by the tensor manifold modeling module.

[0117] The DCS feedback parsing unit receives control instruction execution logs and real-time operating status data from the boiler DCS system, including air-coal ratio adjustment, coal feed rate, oxygen feedback value, etc., and generates a structured feedback data set through data cleaning and format conversion. The formula is as follows:

[0118]

[0119] in, Feedback data set for DCS system; u t is the control instruction vector at time t; y t is the actual operating state vector at time t; T is the total length of the data acquisition time window; t=1 represents the starting time index of the data sequence.

[0120] The feature mutation detection unit performs mutation point detection on the manifold feature vector based on the sliding window T test method and calculates the statistical difference of the feature vectors in adjacent time windows. The calculation formula is:

[0121]

[0122] Among them, T score is the T statistic for feature mutation detection; is the manifold feature mean vector of window A; is the manifold feature mean vector of window B; is the manifold feature variance of window A; is the manifold feature variance of window B; n A is the number of samples in window A; n Bis the number of samples in window B.

[0123] The parameter dynamic correction unit calculates the core tensor update rate of the manifold modeling module and the weight adjustment amount of the quantum optimization module according to the warning level At and the DCS execution effect evaluation result Et (the deviation between the quantitative control instruction and the expected state) as follows:

[0124]

[0125] Among them, η g is the core tensor update rate; η0 is the basic learning rate; is the L2 norm of the DCS execution effect evaluation result; σ is the attenuation coefficient; Δα i is the adjustment amount of the i-th quantum optimization weight; γ is the weight adjustment step size; is the comprehensive performance loss function for the i-th weight parameter α i The partial derivative of .

[0126] The correction parameters are synchronously injected into the core tensor storage area of ​​the manifold modeling module and the weight register of the quantum optimization module through an encrypted channel.

[0127] The closed-loop feedback execution unit loads the corrected parameters into the corresponding module in real time and monitors the changes in system performance indicators after the parameters are updated. If the performance improvement rate is greater than or equal to the minimum improvement threshold, the parameter correction is confirmed to be effective; otherwise, the secondary correction process is triggered and back-optimization is performed based on the historical optimal parameters. The judgment formula is as follows

[0128]

[0129] Where ρ is the system performance improvement rate; J t is the comprehensive performance loss function value at the current moment; J t+1 is the comprehensive performance loss function value at the next moment after parameter correction; ρ min The preset minimum performance improvement threshold

[0130] This module uses a closed-loop mechanism of real-time monitoring, early warning, and correction to ensure that the system can maintain high-precision modeling and optimized control capabilities in scenarios such as fluctuations in boiler combustion conditions, changes in fuel characteristics, or equipment aging.

[0131] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The multi-modal optimization system for combustion efficiency of thermal power boilers is characterized by: include: The multimodal data perception and spatiotemporal alignment module is used to collect multi-source combustion data, align the spatiotemporal references of heterogeneous data based on a spatiotemporal mapping function, and output a multimodal data stream in a unified coordinate system. The tensor manifold modeling and physical constraint feature extraction module constructs a fifth-order combustion state tensor based on multimodal data streams, embeds combustion dynamics constraints through pre-trained core tensors, and extracts low-dimensional physical features by updating the factor matrix online; The spatiotemporal coupled dynamic prediction and uncertainty quantification module inputs the extracted low-dimensional physical features into the PDE constrained deep network, combines time series modeling with the attention mechanism to predict the combustion state and generate prediction confidence intervals; The quantum optimization decision-making and DCS collaborative control module builds a multi-objective optimization model based on prediction results and confidence intervals, solves the optimal control instructions through quantum computing, and realizes the synchronous output of instructions to DCS signals through a protocol converter; The combustion state derivative warning and optimization feedback module receives manifold modeling features and DCS feedback data, implements risk warning through feature mutation detection, dynamically modifies modeling parameters and optimization weights, and feeds the adjusted parameters back to the manifold modeling and quantum optimization module to form a closed-loop iterative mechanism.

2. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 1, characterized in that: The multimodal data perception and spatiotemporal alignment module includes: The flame vision acquisition unit is used to obtain image information of the burning flame inside the furnace and time-stamp the image frames; Temperature field detection unit, used to collect temperature data at multiple points in the furnace and synchronize it with the flame image timestamp; Flue gas composition analysis unit, used for real-time detection of O2, CO, NO in exhaust gas x Equal gas content and output corresponding time series data; The operating parameter access unit is used to access the boiler operating parameters from the DCS system and unify all sensor data into the time and space alignment processing unit.

3. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 1, characterized in that: The multimodal data perception and spatiotemporal alignment module also includes: The time synchronization subunit is used to synchronize data based on the timestamp information of the raw data from various sensors using a precision clock protocol; The spatial registration subunit is used to transform the spatial coordinates of the flame image, temperature sampling points and smoke sensor positions according to the preset spatial mapping function to build a unified spatial reference frame; The data reconstruction subunit is used to fuse multi-source heterogeneous data that have completed spatiotemporal alignment into a multimodal data stream with a consistent time scale and spatial coordinate system, and output it to subsequent modules via high-speed Ethernet.

4. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 1, characterized in that: The tensor manifold modeling and physical constraint feature extraction module includes: A tensor construction unit, configured to reconstruct a multimodal data stream into a fifth-order combustion state tensor, wherein the tensor comprises a spatial dimension, a channel dimension, a combustion feature dimension, and a time dimension; Core tensor pre-training unit, used to pre-train core tensors based on computational fluid dynamics simulation data and embed combustion dynamics constraints during the training process; The online updating unit is used to update the factor matrix of tensor decomposition according to real-time combustion data and output low-dimensional manifold features representing the combustion state.

5. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 4, characterized in that: The core tensor pre-training unit achieves combustion dynamics constraint embedding through the following optimization objectives: in, is the core tensor after embedding the combustion dynamics constraints, It is a combustion field tensor constructed based on computational fluid dynamics simulation data; is the core tensor to be optimized; U(n) is the n-th order factor matrix; λ is the manifold smoothing coefficient and satisfies 0.1≤λ≤1.0; is the combustion field Laplace matrix; tr(·) represents the matrix trace operation.

6. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 1, characterized in that: The spatiotemporal coupling dynamic prediction and uncertainty quantification module includes: A feature input unit, used to receive the low-dimensional physical feature vector output by the tensor manifold modeling module; Dynamic prediction unit, which includes a PDE-constrained deep network and a spatiotemporal attention mechanism to predict future combustion state parameters; Uncertainty quantification unit, used to generate confidence intervals for prediction results through random dropout strategy.

7. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 6, characterized in that: The PDE-constrained deep network is trained using the following composite loss function: in, is the composite loss function of the deep network; α is the weight coefficient of the prediction error term; is the mean square error between the predicted value and the true value; β is the weight coefficient of the physical constraint term; K is the total number of combustion state parameters; is the predicted value of the kth combustion state parameter; is the time derivative of the prediction parameter; is the discretized differential operator of the combustion field control equation; is the spatial gradient operator; is the Laplace operator.

8. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 1, characterized in that: The quantum optimization decision and DCS collaborative control module includes: Optimization model building unit, used to transform prediction results and confidence intervals into multi-objective optimization problems; A quantum annealing solver unit, which is used to encode the optimization problem into a QUBO model and solve the optimal control instructions through a quantum annealing machine; The protocol conversion unit is used to convert the quantum optimization results into control signals that can be executed by the DCS system.

9. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 8, characterized in that: The objective function construction method of the QUBO model is: Where H is the quantum annealing optimization objective function; a i is the combustion efficiency optimization coefficient of the i-th air-coal ratio valve; b ij is the coordinated regulation constraint coefficient of the i-th and j-th valves; γ is the balance coefficient; S target is the total number of target valves opened; N is the total number of air-coal ratio regulating valves.

10. The multi-modal optimization system for combustion efficiency of thermal power boilers according to claim 1, characterized in that: The combustion state derivative warning and optimization feedback module includes: Manifold feature receiving unit, used to obtain the low-dimensional manifold feature vector and corresponding timestamp output by the tensor manifold modeling module in real time; DCS feedback analysis unit, used to analyze the control instruction execution log and boiler operation status data returned by the DCS system; A feature mutation detection unit is used to detect mutation points on the manifold feature vector based on the sliding window T test method, and generate a three-level warning signal when the statistic exceeds the threshold; The parameter dynamic correction unit is used to calculate the core tensor update rate of the manifold modeling module and the weight adjustment amount of the quantum optimization module based on the warning level and DCS execution effect evaluation results; The closed-loop feedback execution unit is used to synchronously inject the corrected parameters into the parameter memories of the manifold modeling module and the quantum optimization module through the encrypted channel.

Images