Coal-fired boiler furnace flow field prediction and combustion abnormity diagnosis method, system and equipment and medium

Abstract

The invention provides a coal-fired boiler furnace flow field prediction and combustion abnormity diagnosis method, system and device and a medium. According to the method, computational fluid mechanics simulation is carried out by using the operation condition to obtain a flow field feature vector, a flow field mode is marked on the flow field feature vector, the operation condition, the flow field feature vector and the flow field mode form a sample training set, and the constructed prediction diagnosis model is trained, so that the real-time operation condition can be used as input under the variable load condition, and the real-time operation condition can be used as input. And directly utilizing the prediction diagnosis model to output a flow field feature vector, performing clustering analysis according to the flow feature vector to obtain a flow field mode, and further performing combustion abnormity diagnosis according to the flow field mode. According to the method, the real-time or quasi-real-time prediction capability can be provided, the requirement of online intelligent combustion control for the future state can be met, meanwhile, the capability of restraining and diagnosing operation safety on the hearth flow field level is achieved, and flow field state prediction and combustion abnormity diagnosis can be achieved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of boiler combustion process monitoring and intelligent control technology in thermal power generation engineering, and particularly relates to a coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method, system, device and medium. BACKGROUND

[0002] The flow field state inside the coal-fired boiler furnace includes the position and volume of the backflow area, the vortex intensity distribution, the mixing uniformity of air and fuel, the jet impact position and diffusion form, etc. The above flow field state directly affects the flame stability, the ignition and burnout position, the formation of local oxygen-rich or oxygen-poor areas, thereby determining the boiler thermal efficiency, the nitrogen oxide generation level, and the coking and high-temperature corrosion risk.

[0003] Large coal-fired boilers face the following key technical pain points under variable load operation conditions. First, the three-dimensional flow field structure inside the furnace belongs to a typical high-temperature, strong-turbulence, and combustion-coupled environment, which includes the position and volume fraction change of the main backflow area, the impact range and diffusion form of the combustion jet, the interference effect of the upper and lower staged air, the mixing uniformity of the coal powder and air, the velocity non-uniformity of each characteristic section, and other key parameters. These physical quantities have a direct impact on combustion stability, coking tendency, flame deviation, and heating surface safety. However, existing power plant online monitoring means are still limited to a small number of temperature, pressure, oxygen content, and other scalar measurement points arranged on the wall, which can only indirectly infer the complex three-dimensional structure inside the furnace, making it difficult to identify abnormal patterns such as flow field deviation, backflow area enlargement, mixing deterioration, or local turbulence enhancement in a timely manner, and unable to make forward-looking judgments on the combustion state. Second, although computational fluid dynamics (CFD) technology can provide complete velocity field, temperature field, and species distribution under typical design conditions, and obtain high-precision description of backflow structure and mixing state, its solving process requires a large amount of computing resources and a long solving time, and can only be used as an offline analysis tool for design stage or a few fixed-condition diagnosis scenarios. For dynamic conditions such as load fluctuations, fuel quality changes, and differences in air-coal regulation strategies frequently occurring in actual operation of the unit, traditional CFD cannot provide real-time or quasi-real-time prediction capability, and it is even more difficult to support the needs of online intelligent combustion control for future states.

[0004] Existing intelligent combustion optimization systems mainly build prediction and optimization models around scalar indicators such as boiler efficiency, flue gas temperature, and nitrogen oxide emissions, and the optimization objectives and constraints are mostly overall thermal performance and emission indicators, lacking constraints and diagnosis capabilities for operation safety at the furnace flow field level. In the case of only optimizing scalar performance, even if the overall efficiency and emission level reaches the optimization target, engineering risks such as flame center deviation, local strong backflow leading to coking and high-temperature corrosion may still occur.

[0005] Therefore, there is an urgent need for a technical solution that can combine limited high-fidelity flow field simulation data and a large amount of online operation data to predict the evolution of the furnace flow field state under variable load conditions and identify abnormal combustion patterns, so as to provide forward-looking constraints and diagnostic basis at the flow field level for operation adjustment and intelligent combustion optimization. SUMMARY

[0006] To solve the above problems of the prior art, the present application provides a coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method, system, device and medium.

[0007] The present application is realized by the following technical solutions: In a first aspect, the present application provides a coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method, comprising: obtaining a real-time operating condition vector of the boiler; inputting the real-time operating condition vector into a trained prediction and diagnosis model to output a predicted flow field feature vector and a flow field mode category; and performing combustion abnormality diagnosis according to the flow field mode category; wherein the prediction and diagnosis model is obtained by the following method: obtaining an operating condition time series of the boiler, performing CFD simulation to obtain a flow field feature vector corresponding to the operating condition of each time window, performing cluster analysis on the flow field feature vector to obtain different flow field mode categories and performing semantic annotation, grouping each operating condition and the corresponding flow field feature vector and flow field mode category into a time series supervised learning sample set, training a prediction and diagnosis model based on a time series neural network using the time series supervised learning sample set, and obtaining a trained prediction and diagnosis model.

[0008] Preferably, the operating condition time series of the boiler is obtained by: obtaining historical operating data from the boiler distributed control system, using a time window-based operating data screening mechanism to screen operating data within a time window with a stability meeting a preset requirement from the historical operating data, performing outlier rejection, consistency correction and normalization on the screened operating data, and organizing the operating data in a time series form to obtain the operating condition time series.

[0009] Preferably, the CFD simulation is performed to obtain a flow field feature vector corresponding to the operating condition of each time window, specifically including: taking the operating condition of each time window in the operating condition time series as input, obtaining a three-dimensional field quantity through CFD simulation, extracting a flow field feature reflecting the structure of the furnace flow field from the CFD three-dimensional field quantity, and grouping the flow field feature into a flow field feature vector corresponding to the operating condition of each time window; wherein the extracted flow field features include backflow region volume fraction, backflow region barycenter coordinates, jet widening coefficient, velocity non-uniformity coefficient of a preset key section, and vorticity integral of a preset easy coking area.

[0010] Preferably, the CFD simulation is performed to obtain the flow field feature vector corresponding to the operating condition of each time window, specifically including: performing CFD simulation on the operating condition of a preset part of the time window to obtain the corresponding flow field feature vector; for each time window that has not been simulated by CFD, the following operations are performed: taking the average of the operating data in the time window to form a condition vector , selecting the K nearest CFD samples to the condition vector in the CFD simulation database by weighted Euclidean distance measurement, and taking the reciprocal of the Euclidean distance as the similarity weight, and performing weighted fusion on the flow field feature vectors corresponding to the K CFD samples to obtain the flow field feature vector corresponding to the operating condition of the time window; wherein the CFD simulation database stores the operating condition of a preset part of the time window and the three-dimensional field quantity obtained by CFD simulation; the CFD sample is composed of the operating condition of the time window and the corresponding three-dimensional field quantity; the three-dimensional field quantity includes three-dimensional velocity field, temperature field and flow field feature vector.

[0011] Preferably, the time series neural network is a long short-term memory network, a gated recurrent unit network or a convolutional time series network.

[0012] Preferably, the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method further comprises: retrieving the K nearest CFD samples in the CFD simulation database according to the predicted flow field feature vector, and then taking the reciprocal of the flow field feature space distance as the weight to perform weighted average on the three-dimensional velocity field and temperature field of the K CFD samples to obtain the three-dimensional field quantity consistent with the predicted flow field feature vector, thereby completing the reconstruction of the three-dimensional flow field of the furnace; wherein the CFD simulation database stores the operating condition of each time window and the three-dimensional field quantity obtained by CFD simulation; the CFD sample is composed of the operating condition of the time window and the corresponding three-dimensional field quantity; the three-dimensional field quantity includes three-dimensional velocity field, temperature field and flow field feature vector.

[0013] Preferably, the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method further comprises: constructing a combustion risk index according to the predicted flow field mode, and performing alarm when the future risk index presents a continuous rising trend. Taking the amount of each layer of secondary air, the proportion of primary air, and the coal load of each coal feeder as the optimization variables, taking the improvement of the future flow field structure as the objective function, and adopting the particle swarm optimization algorithm to solve the multi-objective problem, the amount of primary air and secondary air is adjusted, and the coal load correction scheme of each coal feeder is output.

[0014] In a second aspect, the present application provides a coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis system, comprising: A data acquisition module is configured to acquire a real-time operating condition vector of a boiler. The prediction diagnosis model takes the real-time operation condition vector as input and outputs a predicted flow field feature vector and a flow field mode category; combustion abnormality diagnosis is performed according to the flow field mode category; wherein the prediction diagnosis model is obtained by the following method: obtaining an operation condition time sequence of the boiler and performing CFD simulation to obtain a flow field feature vector corresponding to the operation condition of each time window, performing cluster analysis on the flow field feature vector to obtain different flow field mode categories and perform semantic annotation, grouping each operation condition and the corresponding flow field feature vector and flow field mode category into a time sequence supervised learning sample set, training the prediction diagnosis model based on the time sequence neural network using the time sequence supervised learning sample set, and obtaining the trained prediction diagnosis model.

[0015] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method as described above when executing the computer program.

[0016] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and wherein the computer program is executable by a processor to implement the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method as described above.

[0017] Compared with the prior art, the present application has the following advantages: The present application uses operation conditions to perform computational fluid dynamics (CFD) simulation to obtain flow field feature vectors, and annotates flow field modes for the flow field feature vectors, groups operation conditions, flow field feature vectors, and flow field modes into a sample training set, trains the constructed prediction diagnosis model, so that the flow field feature vector can be directly output by the prediction diagnosis model using real-time operation conditions as input under variable load conditions, and the flow field mode can be obtained by cluster analysis according to the flow field feature vector, and the combustion abnormality can be diagnosed according to the flow field mode. The present application can provide real-time or quasi-real-time prediction capability without real-time CFD solution, can support the demand for future state of online intelligent combustion control, and has the constraint and diagnosis capability of operation safety at the level of furnace flow field, and can realize flow field state prediction and combustion abnormality diagnosis.

[0018] Further, the present application uses a three-dimensional field interpolation reconstruction technology based on neighborhood CFD to map the predicted flow field feature vector to the corresponding three-dimensional flow field distribution, and provides an internal flow structure estimation with engineering interpretability.

[0019] Further, the present application constructs a combustion risk index based on a predicted flow field mode, and takes the bias flow degree, backflow intensity, mixing condition and the like as quantitative indexes for optimization target, adopts a particle swarm optimization (PSO) algorithm to perform multi-objective adjustment search on air volume distribution and coal supply amount, forms a future-oriented combustion adjustment suggestion, and realizes early identification and intervention on the flow field anomaly. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 The figure is a schematic diagram of the overall structure of the present application prediction and combustion anomaly diagnosis method.

[0022] Figure 2 The figure is a schematic diagram of the overall structure of the present application prediction and combustion anomaly diagnosis method.

[0023] Figure 3 The figure is a schematic diagram of the overall structure of the present application prediction and combustion anomaly diagnosis method. DETAILED DESCRIPTION

[0024] The embodiments of the present application will be described below through specific concrete examples, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. The present application can also be implemented or applied through other different specific embodiments, and each detail in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application.

[0025] It should be understood that the process equipment or device not specifically indicated in the following examples all uses the conventional equipment or device in the art.

[0026] It should be noted that the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices. Moreover, unless otherwise specified, the numbering of each method step is only a convenient tool for identifying each method step, and is not a limitation on the arrangement order of each method step or a limitation on the scope of the present application. Changes or adjustments of the relative relationship, without substantial changes in the technical content, are also considered as the scope of the present application.

[0027] This invention proposes a method for predicting furnace flow field and diagnosing combustion anomalies in coal-fired boilers based on time-series machine learning. This method uses historical operating data and furnace flow field simulation data as its foundation. By constructing a nonlinear mapping between the operating condition time series and the furnace flow field feature vector, it achieves dynamic prediction of the furnace flow field state under varying load conditions and identifies flow field patterns related to combustion anomalies within the flow field feature space. The method of this invention specifically includes the following parts (…). Figure 1 ).

[0028] 1. Construction of operating condition time series and selection of time windows corresponding to stable operating conditions. This invention first acquires historical operating data covering a long timescale from the boiler distributed control system. This historical operating data includes, but is not limited to, unit load, total air volume and stratified primary and secondary air volumes, coal feed rate and start / stop status of each coal feeder, furnace outlet flue gas temperature, exhaust gas oxygen content, temperature of key parts of the heating surface, and other operating data that can reflect the furnace flow state. Because this type of operating data is characterized by large fluctuations, high noise, and a large proportion of non-steady-state periods during acquisition, directly using it as input to a predictive diagnostic model would make it difficult to establish a stable mapping between furnace flow field feature labels and the operating condition time series. Therefore, this invention adopts an operating data filtering mechanism based on time windows, that is, dividing continuous historical operating data into several time windows with a fixed step size, and extracting representative operating data within each time window. (e.g., unit load or total air volume), calculate the maximum value of this operating data within the time window. Minimum value Compared with the average Based on this, the stability factor is defined:

[0029] when Less than the preset threshold If the stability of the operating conditions within the specified time window is within a certain range, it indicates that the fluctuation range is limited and the stability meets the preset requirements, thus representing a single stable operating condition. Otherwise, the time window is considered to be in a rapidly changing state and is discarded. Through this screening method, the operating condition time series subsequently learned by the model consists only of time windows whose stability meets the preset requirements, which helps improve the learnability of the prediction and diagnostic model and the accuracy of label alignment.

[0030] Subsequently, the time window passing through the screening is cleaned, that is, the outlier rejection and consistency correction. The outlier rejection is based on the joint distribution relationship between the running data to carry out statistical test, and the points higher than 3 times of standard deviation or contradictory to the similar running data are rejected; the consistency correction is to restore a small amount of missing values caused by sensor fluctuation through linear interpolation or KNN interpolation. In order to keep the running data of different dimensions consistent in scale in the input of the prediction and diagnosis model, the present application normalizes all running data, that is, adopts the standardization mode of zero mean and unit variance, so that each dimension of running data conforms to similar scale distribution. Finally, the running data passing through screening, cleaning and normalization are organized in time sequence form to form the running condition time sequence as the basis for subsequent construction of the input of the prediction and diagnosis model.

[0031] The specific implementation of this step includes: (1) Collecting multi-source channels (such as coal feeding rate, primary / secondary air volume, tuyere differential pressure, furnace outlet temperature, etc.) from DCS at a fixed sampling frequency, and constructing a time window according to a sliding window length , (Examples: 30-300 sampling points, determined according to the sampling frequency and control time limit); ; (2) Calculating the stability criterion for each time window, and determining it as a stable time window and retaining when the stability of all key variables in the time window is lower than a threshold value (Examples: 0.05-0.2); otherwise, marking the time window as a transition state or abnormal window and temporarily not using it as a training sample; (3) Using neighbor interpolation or linear interpolation based on adjacent time windows to complete the missing values in the time window, rejecting the abnormal points according to or manual checking, and standardizing the training set statistics for each channel; (4) Writing the time window passing through the screening into the sample pool and recording the corresponding time stamp and condition signature for alignment with the CFD label.

[0032] 2. Three-dimensional flow field simulation and flow field feature vector construction ​​After obtaining the screened, cleaned and normalized operation condition time series, in order to construct the "condition-flow field structure" mapping relationship that the prediction and diagnosis model can learn, the application further constructs a furnace flow field simulation database (referred to as CFD simulation database) based on the computational fluid dynamics (CFD) method. Taking a plurality of representative operation conditions (i.e. the operation conditions of the preset partial time window) in the operation condition time series covering the typical load interval, different front and back wall coal feeding combinations and different stratified air distribution schemes as the input of the CFD simulation model, the CFD simulation is performed, and through the steady-state solving or weak non-steady-state solving, the three-dimensional velocity field, temperature field, turbulent kinetic energy field, combustion reaction scalar and vortex scale of the furnace interior are obtained. The operation condition and the corresponding three-dimensional field quantity form a CFD sample stored in the CFD simulation database.

[0033] The implementation of the CFD simulation database includes: (1) Set the database field structure, including the operation condition signature (such as coal feeding rate, primary air volume, secondary air volume, air ratio, boiler load, inlet air temperature, fuel property, etc.), three-dimensional field file reference (velocity field, temperature field, main species mass fraction) and flow field feature vector obtained by CFD calculation; (2) Obtain real-time or historical operation conditions from the DCS according to time or condition triggers, and automatically initiate CFD or approximate rapid simulation calculation when the operation condition significantly deviates from the existing database record or reaches the preset sampling strategy (for example, condition distance threshold or time interval); that is, the CFD simulation database is a constantly updated database; (3) Use RANS (k-ε or k-ω) or LES solver for high-fidelity CFD, locally encrypt the grid at the burner outlet, recirculation zone and heating surface, and save the instantaneous or time-averaged field; for online rapid simulation, a simplified combustion model or a pre-trained approximate solver can be used to reduce the calculation time; (4) For physical processes with system response time lag, search for time offset parameters in the historical window range, and determine the optimal alignment delay by minimizing the input-output correlation error, so as to maintain causal consistency; (4) Post-process the CFD simulation results each time, extract and store the flow field feature vector and operation condition signature, establish an index (such as KD-tree or neighbor index based on the operation condition signature), and record the simulation timestamp and simulation accuracy metadata for subsequent retrieval and confidence assessment.

[0034] Although the three-dimensional field quantity obtained by the CFD simulation can theoretically completely describe the flow structure inside the furnace, if directly used as the label of the prediction and diagnosis model, it will lead to problems such as too large data volume, unstable training, and difficult structure learning due to the extremely high data dimension. Therefore, the present application extracts a group of low-dimensional flow field feature vectors with clear physical meaning and capable of stably reflecting the flow field structure of the furnace from the three-dimensional field quantity obtained by the CFD simulation.

[0035] In order to accurately describe the size and position of the main recirculation zone of the furnace, the recirculation area is determined by the sign of the velocity component, and the mathematical expression is:

[0036] Among them, is the volume fraction of the recirculation zone; is the CFD calculation domain of the furnace; is the position of the velocity component along the jet direction; is the volume element; is an indicator function, which is 1 when the velocity component direction of the point is reversed (i.e. less than zero), otherwise 0, which is used to clearly select the area where the velocity is reversed. In order to reduce the influence of numerical noise on the recirculation determination, a small negative threshold (e.g. =10 -3 ~10 -2 m / s, adjusted according to the grid and solver accuracy) is introduced in the criterion, that is, the is used as a more robust recirculation determination condition. Through the volume fraction, the overall size of the recirculation structure can be intuitively reflected, and when increases, it indicates that the volume of the recirculation zone expands, and vice versa.

[0037] At the same time, in order to characterize the spatial offset of the recirculation structure, according to the definition of the geometric center of the recirculation zone, the recirculation zone barycentric coordinates are obtained by averaging the position vector field of the recirculation zone. Let the position vector be , then the recirculation zone barycentric coordinates are expressed as:

[0038] Among them, is the barycentric coordinate of the vector recirculation zone; is the spatial coordinate of the center position of the calculation unit. The change of the barycentric coordinate of the recirculation zone can be used to identify whether the furnace has a left-right flow deviation trend, which is of great significance to the judgment of non-uniform heating of the heating surface.

[0039] Regarding jet-related characteristics, this invention determines the jet impact location and jet broadening degree by tracking the velocity distribution along the centerline of the nozzle jet. To characterize the jet broadening scale, this invention performs root-mean-square statistics on the transverse velocity components on a vertical section along the jet penetration direction to construct the jet broadening:

[0040] in, This is the jet broadening factor; The area of ​​the selected jet characteristic section; Let be any point on the cross section; Let these be the velocity components along the horizontal and vertical directions at that point, respectively. The area is represented by a micro-element. The jet broadening characterizes the jet's penetration and dispersion effect, and this parameter is directly significant for judging the mixing uniformity and the flame impact zone.

[0041] Furthermore, to reflect whether the velocity distribution within the furnace is uniform, this invention calculates the velocity non-uniformity at several key cross-sections (such as the main combustion zone cross-section, the near-wall cross-section, and the deflection zone cross-section):

[0042] in, The coefficient for velocity non-uniformity; This represents the number of sampling points for the cross section. The local velocity magnitude at the i-th sampling point; This is the average velocity across the entire cross-section. This indicator can be used to identify whether the nozzle jet is deflected to one side, whether high-speed channels are formed in the near-wall region of the flue gas, and whether there are abnormal distributions of strong backflow coupling in the deflection zone. Therefore, it is of great significance in combustion stability diagnosis and operational adjustment decisions.

[0043] To further describe the mixing condition, this invention uses mixing uniformity to evaluate the uniformity of mass fraction or velocity at different cross sections, reflecting the degree of mixing between flue gas and fuel. It is defined as follows:

[0044] in, It is the mixing uniformity index; For the first A mixed scalar at each sampling point, such as oxygen concentration, fuel mass fraction, or temperature; This is the average scalar value of the corresponding cross section; The reference value used for normalization can be the peak deviation across the entire cross-section or the maximum deviation statistically obtained from various operating conditions. When When the value is close to 1, it indicates a highly uniform mixing level; when... The decrease indicates that the mixing deteriorates and the stratification is obvious, and the index is of decisive significance for judging the mixing performance and flame stability of the furnace.

[0045] In addition, in order to describe the development degree of the local turbulent structure of the furnace, the application introduces a local vorticity integral index to describe the local strong turbulent structure in the high-risk area (such as the corner recirculation zone, the turning zone, and the nozzle collision zone). That is, the integral of the velocity curl modulus in the specified volume region is calculated:

[0046] Among them, is the local vorticity integral; The specified local high-risk volume region. The change of the vorticity integral can be used to judge the trend of forming a local strong recirculation or stagnation zone.

[0047] Finally, the application combines the above-mentioned multiple flow field characteristics (recirculation zone volume fraction , recirculation zone barycentric coordinates , jet widening , velocity non-uniformity coefficient , mixing uniformity index , vortex intensity ) into a low-dimensional flow field characteristic vector in a fixed order, and establishes a mapping relationship between each operating condition and its corresponding flow field characteristic vector to form a representative "operating condition-flow field characteristic vector" sample training set that can be used for supervised learning. The obtained flow field characteristic vector is also added to the CFD sample.

[0048] Each of the flow field characteristics is scaled (such as z-score or MinMax) before being combined into a vector , and is stored as a label vector in a predetermined order.

[0049] 3. Time window and CFD flow field characteristic vector label alignment Since the operating conditions that can be covered by CFD simulation are limited, and the actual operating conditions present continuous and multi-dimensional characteristics, it is not realistic to directly map the time window into the CFD flow field characteristic vector label. The application proposes a time window-flow field characteristic vector label alignment method based on operating condition similarity, performs CFD simulation on part of the representative operating conditions, finds the CFD sample that best matches the operating condition of the time window that has not been simulated by CFD, and generates continuous and smooth flow field characteristic vector labels through interpolation, thereby greatly expanding the number of samples that can be used for time series supervised learning.

[0050] Specifically, the application firstly averages the operation data in each operation window without CFD simulation to form an operation condition vector representing the steady state of the time window . Then an operation condition vector of a CFD sample is selected in the CFD simulation database , and importance weight is set for different dimension data according to the sensitivity of the operation condition to the flow field , the similarity between the two is measured by weighted Euclidean distance , , and the similarity is obtained as follows:

[0051] The smaller the Euclidean distance is, the closer the operation condition of the time window is to the corresponding operation condition . According to the method, K CFD samples with the closest Euclidean distance to the operation condition of the time window are selected, and the corresponding flow field feature vectors are recorded as , and the continuous flow field feature label corresponding to the time window is constructed as follows:

[0052] , wherein is the normalized weight with Euclidean distance reciprocal as the similarity:

[0053] The significance of this interpolation method is that even if CFD only provides limited samples, through the mapping of condition similarity, the time window can also obtain continuous, smooth and physically consistent flow field feature vector labels, so that the scale of the training set of the time series supervised learning sample is expanded from hundreds of thousands to millions, greatly improving the learnability and robustness of the prediction and diagnosis model.

[0054] The implementation of this step specifically includes: (1) calculating the operation condition vector for the time window, which includes the average coal feed rate, average primary / secondary air volume, air ratio, boiler load, inlet air temperature, etc. of the time window; and constructing the operation condition vector for each CFD sample; (2) calculating the weighted Euclidean distance between and :

[0055] and taking the minimum distance or several adjacent samples meeting the distance threshold as the candidate CFD sample; (3) when the interpolation condition point is located in the condition range of multiple adjacent CFD samples, the corresponding CFD flow field feature vectors are interpolated by condition distance weighting to obtain continuous labels:

[0056] (5) Finally, the aligned or interpolated labels with the corresponding time window The training sample pairs are written into the training set.

[0057] 4. Flow field feature space clustering and semanticization After obtaining the time series supervised learning sample training set by the above method, the flow field feature vectors of the CFD samples are classified to identify flow field patterns with typical structural characteristics in the hearth. The flow field feature vectors can be divided into several categories, including normal stable mode, bias flow mode, strong backflow mode, mixed deterioration mode, and flame upward movement mode, etc. According to historical coking events, smoke temperature deviation statistics, and operation personnel experience, the present application semantically labels each flow field mode, so as to establish a corresponding relationship between each flow field mode and a specific operation phenomenon. For other time windows with flow field feature vectors obtained by interpolation, the flow field mode is determined according to the flow field mode to which the CFD sample with the nearest Euclidean distance belongs. Thus, the flow field feature vector corresponding to each time window is classified into a flow field mode, and the flow field mode corresponding to the flow field feature vector is added to the time series supervised learning sample.

[0058] 5. Training of prediction and diagnosis model based on time series neural network Figure 2 After obtaining a large number of time window and flow field feature vector label pairs and flow field modes, the present application organizes them into time series supervised learning samples. The input sequence is composed of the operation condition vectors at consecutive time points in the operation condition time series, and the output is the flow field feature vector at the current or future time points and the flow field mode corresponding to the flow field feature vector , thereby constructing a prediction and diagnosis model of "historical operation condition time series→current / future flow field structure→flow field mode", which is represented as:

[0059] In order to accurately learn the evolution law of the hearth flow field, the prediction and diagnosis model of the present application selects a time series neural network with time memory capability, including a long short-term memory network (LSTM), a gated recurrent unit network (GRU), or a convolutional time series network (TCN). The model is trained by back propagation to minimize the weighted mean square error between the predicted features and the true labels, wherein the key features such as backflow scale and velocity non-uniformity are given higher weights, so that the model can focus on learning the flow characteristics that have a decisive influence on combustion safety.

[0060] ​​Due to the significant differences in the working condition change speed, flow field structure stability and mixing path in different boiler load intervals, a single fixed structure of neural network cannot take into account all working conditions. Therefore, the particle swarm optimization algorithm (PSO) is further introduced into the super parameter selection process, and the optimal combination is searched in the high-dimensional parameter space of network layer, hidden node size, time window length and learning rate through group search, so that the prediction and diagnosis model has high generalization performance and stability in the full load range.

[0061] The prediction and diagnosis model first predicts the flow field feature vector according to the input operating condition, and then performs cluster analysis on the flow field feature vector to output the mode category and the corresponding risk level. Through K-means, DBSCAN or Gaussian mixture clustering, the flow field feature vector can be divided into several categories, including normal stable mode, partial flow mode, strong backflow mode, mixed deterioration mode and flame upward mode. The prediction and diagnosis model of the application can automatically classify the future flow field mode, so that once the predicted flow field feature vector shows a trend of deviating to a high-risk mode, an abnormal risk warning can be given in advance.

[0062] The specific implementation of the prediction and diagnosis model construction and training includes: (1) Model structure: adopting a stacked LSTM or Seq2Seq (encoder-decoder) structure, the encoder receives a multi-channel input with a window length of , and the decoder directly outputs the flow field feature vector of the future steps in multiple steps ; an attention mechanism can be used between the encoder and the decoder to improve the long-term dependence capturing ability; (2) Loss and training: using a weighted mean square error loss function , wherein higher weights are given to key engineering quantities (such as and ); the Adam optimizer, learning rate decay and early stopping are used for training; training, validation and testing are time-series divided to avoid information leakage; (3) Super parameter search: model super parameters (L, H, number of LSTM layers, number of hidden units, loss weight, etc.) can be automatically searched for the best configuration on the validation set through particle swarm optimization (PSO) or Bayesian optimization; (4) Online deployment: after training, the model is exported as an inference service, the inference period (example 30s-300s) is set, the latest window is input into the model in real time to obtain the future H-step prediction, and the prediction confidence (such as based on model ensemble variance or MC-dropout) is recorded for subsequent decision-making; at the same time, periodic offline retraining or incremental fine-tuning is realized to cope with concept drift.

[0063] 6. Fast reconstruction of 3D flow field based on nearest neighbor interpolation in feature space In order to make the low-dimensional flow field feature vector of time series prediction be interpreted as visualized 3D velocity field and temperature field, the present invention reconstructs 3D field quantities based on nearest neighbor search in feature space. When the flow field feature vector is predicted , the K nearest CFD samples in the CFD simulation database are searched, and then the 3D velocity field and temperature field of these CFD samples are weighted averaged with the inverse of the flow field feature vector spatial distance as the weight, so as to obtain the reconstructed 3D field quantities consistent with the current predicted flow field feature vector. This method does not need to solve the CFD equation again, and the operation speed is improved by tens of thousands of times compared with traditional CFD solution, realizing near real-time reconstruction of 3D flow field in the furnace, and providing visual and intuitive internal flow structure information for the operating personnel.

[0064] The specific implementation of 3D flow field reconstruction (interpolation reconstruction): according to the position of the predicted flow field feature vector in the feature space, the K nearest CFD samples (according to the Euclidean distance of the operating condition or flow field feature vector) in the CFD simulation database are searched, and the 3D field quantities 、 、 of the K CFD samples are linearly weighted interpolated point by point in space according to the weight (for example ) or radial basis function interpolation to obtain the reconstructed 3D flow field:

[0065] Confidence evaluation: based on matching distance , interpolation condition number or energy preservation (if using POD mode) when reconstruction to calculate the reconstruction confidence.

[0066] 7. Future combustion risk deduction and optimal wind-coal adjustment strategy generation based on PSO Finally, the present invention constructs a combustion risk index according to the predicted flow field mode at multiple future time steps, with the bias flow intensity, recirculation zone volume, mixing non-uniformity and high-risk vortex intensity as the risk quantification indicators. When the future risk index shows a continuous upward trend, the system can deduce in advance that the combustion structure may evolve towards a high-risk mode, such as strong recirculation, deterioration of bias flow or upward movement of flame center, etc.

[0067] The implementation of the combustion risk assessment includes: (1) Define risk mapping rules or probability models for each semantic flow field mode, and the rules are based on key feature thresholds (for example 、 、 , cross-section temperature difference, etc.) and the failure probability obtained from historical failure sample statistics; (2) Constructing the comprehensive combustion risk index , which is the weighted sum of several standardized sub-indexes:

[0068] wherein, is the standardized risk score of a single physical quantity, is the engineering weighting coefficient; according to , the combustion risk is divided into several levels (for example: low, medium, and high); (3) In the risk assessment, the evolution trend of the risk (such as continuous N steps exceeding the threshold in the future H steps) is considered to improve the robustness and advance of the early warning.

[0069] On this basis, the present application takes the distribution of secondary air at each level, the proportion of primary air, and the coal feeding load of each coal feeder as the optimization variables, takes the improvement of the future flow field structure as the objective function, and uses the particle swarm optimization algorithm for multi-objective solution. The PSO algorithm finds the optimal adjustment combination in the adjustable parameter space by simulating the group collaboration search mechanism, so that the risk indicators such as bias flow, backflow, and poor mixing are reduced as much as possible, and finally outputs specific air volume adjustment suggestions, coal feeding load correction schemes, and air-coal collaborative optimization strategies, realizing the early intervention of abnormal combustion trends.

[0070] As shown in Figure 3 , the implementation process of using the particle swarm optimization (PSO) algorithm for multi-objective optimization and adjustment decision-making includes: (1) Decision variables: adjustment amount of air damper opening at each level , adjustment proportion of coal feeding amount in each partition , etc. (2) Optimization target: construct a multi-objective function, the target at least includes reducing the degree of bias flow , reducing the backflow intensity , improving the mixing uniformity , and limiting the temperature deviation , which can be converted into a single target by weighted combination:

[0071] wherein, is the target weight, which is set by the engineering priority; (3) Constraint condition: physical and safety constraints (such as maximum and minimum opening of air valve, allowable adjustment range of coal feeder, system response time, combustion safety boundary, emission regulation constraints, etc.); if the automatic execution condition is met, the control instruction is output, otherwise the suggestion is output for manual confirmation; (4) PSO configuration: particle coding contains the above decision variables, population size example 20-50, iteration number example 30-100, velocity / position update adopts inertia weight linear decay strategy and combines local optimal / global optimal double search mechanism; evaluation function is based on the simulation evaluation or rapid index estimation of the reconstructed three-dimensional flow field characteristics or the future flow field characteristic vector predicted by the model; (5) output and implementation: the optimal solution returned by PSO is used as a recommended adjustment scheme, and the implementation cost, expected risk reduction and confidence of the scheme are calculated, and the scheme is issued or displayed in the form of structured command or suggestion, if set to automatic execution and meet the safety prior check, the execution command is automatically issued and the effect is continuously monitored to determine whether to back off or fine-tune.

[0072] Through the above technical solutions, the furnace flow field can be predicted without real-time CFD solving, and the information at the flow field level can be directly involved in intelligent combustion optimization decision-making. The method can significantly improve the combustion stability, energy saving and operation safety of large coal-fired boilers under variable load operation conditions, and provides key technical support for efficient, low-carbon and intelligent combustion control.

[0073] Embodiment In combination with an engineering application scene of a three-hundred-megawatt subcritical opposed firing pulverized coal boiler, the furnace flow field prediction and combustion abnormality diagnosis method based on time series machine learning proposed by the present application is further described. Those skilled in the art can understand that, without changing the essential concept of the present application, the boiler capacity, furnace structure, burner arrangement and specific unit parameters can be appropriately adjusted according to actual needs, which all belong to the implementable range of the present application.

[0074] In the unit described in the present embodiment, the unit load, main steam parameters, primary air and secondary air flow, and the opening degree of each layer damper, the coal supply load and start-stop state of each coal feeder, the furnace outlet flue gas temperature, the reheater inlet and outlet temperature, the air preheater flue gas temperature and pressure difference, and the exhaust gas oxygen content can be synchronously collected through the distributed control system, and the data sampling period is set to one minute. In order to ensure the representativeness and stability of the model training data, in the continuous three-month commercial operation data, firstly, the stability of the original operation data is selected, the running stability is determined by using a time window of every thirty minutes, the maximum value, minimum value and average value of the load are calculated, if the load fluctuation amplitude accounts for less than the set threshold value of the average load, the time window is regarded as a stable window and all operation data in the time window is retained. Then, the abnormal points in the time window are removed by using the box plot method, the missing data is filled by using the adjacent time interpolation method, and all operation data is normalized according to its dimension range, thereby forming the cleaned operation data set which can be directly used to construct the time series sample.

[0075] To establish the mapping relationship between operating conditions and furnace flow field characteristics, different front and back wall coal blending ratios and multiple sets of secondary air ratio distribution schemes in each layer are selected in three typical load intervals (60%, 80%, and 100% load near) in this embodiment, and a simulation database covering representative operating condition combinations is constructed through CFD. The furnace geometry model is constructed based on the actual structure, and the burner arrangement, nozzle form, corner area structure, and screen-type heating surface position are restored. The furnace grid is composed of nearly 1.5 million unstructured units, and is encrypted in the nozzle area to capture the backflow core, jet expansion, and turbulent large-scale structure. The standard k-ε model is selected for the turbulent model, the second-order upwind scheme is used for the discrete scheme, and the SIMPLE algorithm is used for the pressure-velocity coupling. The pulverized coal is simulated by the equivalent continuous phase method, the combustion process is regarded as a volume heat source release, and an empirical reaction rate model is introduced to approximate the burnout behavior. Through steady or weak non-steady state solving, the three-dimensional velocity field, temperature field, turbulent kinetic energy field and combustion reaction scalar distribution can be obtained.

[0076] In order to enable the three-dimensional flow field information to be effectively learned by the prediction and diagnosis model, the low-dimensional flow field feature vector with physical meaning is constructed from the three-dimensional field obtained by CFD simulation according to the above feature vector extraction method. First, the backflow area is defined by the negative axial velocity in the velocity field, the volume fraction of the area is calculated, and the gravity coordinates are obtained by volume weighting to describe the size and position distribution of the main backflow structure. Second, the velocity distribution is counted on the typical cross section downstream of the burner, the jet expansion coefficient is calculated to describe the diffusion and expansion of the jet, and the penetration of the jet and the position of the flame impact area are reflected. Third, the mixing uniformity index is obtained by normalizing the mean square deviation of the oxygen concentration or reaction scalar, which is used to measure the spatial mixing degree of fuel and air. At the same time, in the corner area, the local vortex intensity index is formed by volume integrating the vorticity model in the easy ash accumulation position such as the turning area, to reflect the potential strong backflow, stagnation area and high temperature aggregation phenomenon. Finally, the above multiple flow field characteristics are combined into a flow field feature vector z in a fixed order, and the corresponding operating condition vector is formed as a "operating condition-flow field characteristic" pairing sample, which is used as the label of the subsequent prediction and diagnosis model.

[0077] Since there is a distribution difference between the real operation data and the CFD working conditions, the embodiment further adopts working condition similarity mapping to realize the alignment of the time window and the CFD flow field feature vector label. For each quasi-steady time window, the working condition vector is formed by averaging the operation data in the time window, and the similarity matching is performed with the CFD working condition vector according to the weighted Euclidean distance, the K nearest CFD samples are selected, and the flow field feature vectors are interpolated and fused in the distance reciprocal weighting mode, so as to obtain the flow field feature vector label adapted to the time window. The method can obtain continuous, smooth and physically consistent flow field feature vector labels on the basis of limited CFD samples, so that the scale of the training set is expanded from hundreds of levels to thousands of levels or more.

[0078] When constructing the time series supervised learning sample, the embodiment adopts a sliding window with a length of 30 minutes, takes the working condition vector of the past 30 minutes as the model input, and takes the flow field feature vector after five minutes as the output label to construct the advanced prediction diagnosis model. The long short-term memory network is adopted as the basic structure, the network includes an LSTM layer with 64 memory units and a fully connected layer with 32 nodes, and the weighted mean square error is taken as the loss function, and higher weight is given to the key features such as reflux scale, reflux center of gravity and mixing uniformity, so as to improve the prediction ability of the combustion safety related structure. In order to make the network structure have strong generalization in different load regions, the particle swarm optimization algorithm is introduced to perform global search on the time window length, the size of the hidden layer, the learning rate and other hyperparameters, and the final network configuration is selected according to the minimum error principle of the verification set.

[0079] The flow field mode is divided into several typical categories such as normal stable mode, left flow mode, right flow mode and strong reflux high coking risk mode. The operation record, wall temperature deviation history, flame monitoring image and other information are combined to semantically explain each flow field mode category, so that each flow field mode corresponds to a clear combustion operation phenomenon. After the prediction of the flow field feature vector is completed, the clustering method is adopted to perform clustering analysis on the flow field feature vector, and the flow field mode is output, and the prediction diagnosis model is trained, so that the system can automatically judge the current or future five minutes of the flow field mode and the risk level.

[0080] In order to realize the visualization of the flow field prediction result, the embodiment retrieves a plurality of CFD samples most similar to the predicted flow field feature vector from the CFD simulation database according to the K nearest neighbor retrieval method of the flow field feature space, and performs weighted fusion on the corresponding three-dimensional velocity field according to the similarity as the weight, so as to realize the reconstruction of the three-dimensional flow structure inside the furnace in near real time without CFD solving, so that the operation personnel can directly observe the possible formation of the flow, strong reflux or mixing deterioration form in the future.

[0081] In actual operation, the embodiment deploys the above prediction diagnosis model on a unit computing platform, and the system updates the prediction window every minute. If the prediction shows that the flow field may evolve into a strong backflow high-coking risk mode in the next five minutes, the system automatically generates a warning message and gives wind adjustment, coal adjustment direction and possible cause prompt. Further, the system can search for the optimal combination that can reduce the risk in the adjustable wind-coal ratio parameter space based on the particle swarm optimization algorithm by constructing a multi-objective optimization function including the flow deviation index, backflow scale and mixing uniformity, generate a recommended wind-coal adjustment scheme for reference by the operation personnel.

[0082] It can be known from the embodiment that the method provided by the application can realize forward-looking prediction and abnormal identification of the furnace flow field without online CFD solution, can provide constraint information at the flow field level for boiler combustion optimization under variable load operation conditions, and improve the safety, stability and economy of boiler operation. It can be understood by those skilled in the art that if the boiler type, burner arrangement, fuel property or equipment capacity is adjusted, the equivalent function can be realized by regenerating the CFD simulation database, updating the feature definition or parameterizing and adapting the model structure under the general framework of the application, which all belong to the protection scope of the application.

[0083] The following is an apparatus embodiment of the application, which can be used to execute the method embodiments of the application. For details not mentioned in the apparatus embodiment, please refer to the method embodiments of the application.

[0084] In another embodiment of the application, a coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis system is provided, comprising: A data acquisition module is configured to acquire a real-time operation condition vector of the boiler. A prediction diagnosis model is configured to take the real-time operation condition vector as input, output a predicted flow field feature vector and a flow field mode category, and perform combustion abnormality diagnosis according to the flow field mode category. The prediction diagnosis model is constructed by the following method: acquiring a time series of operation conditions of the boiler, performing CFD simulation to obtain a flow field feature vector corresponding to each time window of the operation conditions, performing cluster analysis on the flow field feature vector to obtain different flow field mode categories and perform semantic labeling, combining each operation condition and the corresponding flow field feature vector and flow field mode category to form a time series supervised learning sample set, training a prediction diagnosis model based on a time series neural network using the time series supervised learning sample set, and obtaining the trained prediction diagnosis model.

[0085] In another embodiment of the present application, a computer device is provided, which comprises a processor and a memory, the memory being configured to store a computer program, the computer program comprising program instructions, and the processor being configured to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are particularly suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method process or a corresponding function. The processor in the embodiment of the present application can be used for the operation of the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method.

[0086] In another embodiment of the present application, the present application further provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in the computer device, and is configured to store programs and data. It can be understood that the computer readable storage medium herein can include the built-in storage medium in the computer device, and of course can also include the expansion storage medium supported by the computer device. The computer readable storage medium provides a storage space, and the storage space stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a random access memory (RAM), and can also be a non-volatile memory, for example, at least one disk memory. One or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method in the above-mentioned embodiments.

[0087] The embodiment further provides a computer program product, which comprises a computer program, and when the computer program is executed by the processor, the corresponding steps of the coal-fired boiler furnace flow field prediction and combustion abnormality diagnosis method in the above-mentioned embodiments are implemented.

[0088] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage etc.) embodying computer readable program code.

[0089] The present application is described in reference to the flowchart and / or block diagrams of the method, the system (device) and the computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagrams, and the combination of flows and / or blocks in the flowchart and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing device, generate a means for implementing the functions specified in the flowchart and / or block diagrams of the flowchart and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the functions specified in the flowchart and / or block diagrams of the flowchart and / or block diagrams.

[0090] These computer program instructions can also be stored in a computer readable storage medium that can direct the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a product including instruction means, which implement the functions specified in the flowchart and / or block diagrams of the flowchart and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the functions specified in the flowchart and / or block diagrams of the flowchart and / or block diagrams.

[0091] These computer program instructions can also be loaded to the computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a means for implementing the functions specified in the flowchart and / or block diagrams of the flowchart and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the functions specified in the flowchart and / or block diagrams of the flowchart and / or block diagrams.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by the same, without departing from the spirit and scope of the present application, and any modification or replacement, which does not depart from the spirit and scope of the present application, should be covered in the protection scope of the claims of the present application.

Claims

1. A method for predicting the flow field in a coal-fired boiler furnace and diagnosing combustion anomalies, characterized in that, include: Obtain the real-time operating condition vector of the boiler; The real-time operating condition vector is input into the trained prediction and diagnosis model, which outputs the predicted flow field feature vector and flow field mode category. Diagnose combustion anomalies based on flow field pattern categories; The predictive diagnostic model is constructed as follows: the boiler's operating condition time series is obtained and CFD simulation is performed to obtain the flow field feature vectors corresponding to the operating conditions of each time window. Cluster analysis is performed on the flow field feature vectors to obtain different flow field mode categories and semantic annotation is performed. Each operating condition and its corresponding flow field feature vector and flow field mode category are combined into a time series supervised learning sample set. The time series supervised learning sample set is used to train the predictive diagnostic model based on the time series neural network to obtain the trained predictive diagnostic model.

2. The method for predicting the furnace flow field and diagnosing combustion anomalies in a coal-fired boiler according to claim 1, characterized in that, The acquisition of the boiler's operating condition time series includes: Historical operating data is obtained from the boiler distributed control system. An operating data filtering mechanism based on time windows is adopted to filter operating data within a time window whose stability meets the preset requirements from the historical operating data. The filtered operating data is then subjected to outlier removal, consistency correction and normalization, and organized in the form of time series to obtain the operating condition time series.

3. The method for predicting the furnace flow field and diagnosing combustion anomalies in a coal-fired boiler according to claim 1, characterized in that, The CFD simulation is performed to obtain the flow field feature vector corresponding to the operating conditions of each time window. Specifically, this includes: taking the operating conditions of each time window in the operating condition time series as input, obtaining three-dimensional field quantities through CFD simulation, extracting flow field features reflecting the furnace flow field structure from the CFD three-dimensional field quantities, and assembling the flow field features into flow field feature vectors corresponding to the operating conditions of each time window; wherein, the extracted flow field features include the volume fraction of the recirculation zone, the centroid coordinates of the recirculation zone, the jet broadening coefficient, the velocity non-uniformity coefficient of the preset key section, and the vorticity integral of the preset easily coking region.

4. The method for predicting the furnace flow field and diagnosing combustion anomalies in a coal-fired boiler according to claim 1, characterized in that, The step of performing CFD simulation to obtain flow field feature vectors corresponding to the operating conditions of each time window specifically includes: performing CFD simulation on the operating conditions of a preset portion of the time windows to obtain the corresponding flow field feature vectors; for each remaining time window where CFD simulation was not performed, the following operation is performed: averaging the operating data within the time window to form an operating condition vector. Distance case vectors in the CFD simulation database are selected using a weighted Euclidean distance metric. The K most recent CFD samples are used, and the flow field feature vectors corresponding to the K CFD samples are weighted and fused using the reciprocal of the Euclidean distance as the similarity weight to obtain the flow field feature vector corresponding to the operating conditions of the time window. The CFD simulation database stores the operating conditions of a preset time window and the three-dimensional field quantities obtained by CFD simulation. The CFD sample consists of the operating conditions of the time window and the corresponding three-dimensional field quantities. The three-dimensional field quantities include the three-dimensional velocity field, temperature field and flow field feature vector.

5. The method for predicting the furnace flow field and diagnosing combustion anomalies in a coal-fired boiler according to claim 1, characterized in that, The time-series neural network is a long short-term memory network, a gated recurrent unit network, or a convolutional time-series network.

6. The method for predicting the furnace flow field and diagnosing combustion anomalies in a coal-fired boiler according to claim 1, characterized in that, Also includes: Based on the predicted flow field feature vector, the K closest CFD samples are retrieved from the CFD simulation database. Then, using the reciprocal of the spatial distance of the flow field features as the weight, the three-dimensional velocity field and temperature field of the K CFD samples are weighted and averaged to obtain the three-dimensional field quantity consistent with the predicted flow field feature vector, thus completing the reconstruction of the furnace three-dimensional flow field. The CFD simulation database stores the operating conditions of each time window and the three-dimensional field quantity obtained by CFD simulation. The CFD sample consists of the operating conditions of the time window and the corresponding three-dimensional field quantity. The three-dimensional field quantity includes the three-dimensional velocity field, temperature field, and flow field feature vector.

7. The method for predicting the furnace flow field and diagnosing combustion anomalies in a coal-fired boiler according to claim 1, characterized in that, Also includes: A combustion risk index is constructed based on the predicted flow field pattern. An alarm is triggered when the risk index shows a continuous upward trend in the future. Using the secondary air volume distribution, primary air ratio, and coal feeder load of each layer as optimization variables, and the objective function of improving the future flow field structure, the particle swarm optimization algorithm is used to solve the multi-objective problem, and outputs the air volume adjustment scheme of primary and secondary air and the coal feeder load correction scheme of each coal feeder.

8. A system for predicting the flow field and diagnosing combustion anomalies in a coal-fired boiler furnace, characterized in that, include: The data acquisition module is used to acquire the real-time operating condition vector of the boiler; The predictive diagnostic model takes the real-time operating condition vector as input and outputs the predicted flow field feature vector and flow field mode category. Combustion anomaly diagnosis is performed based on flow field pattern categories. The predictive diagnosis model is constructed as follows: the boiler's operating condition time series is obtained and CFD simulation is performed to obtain flow field feature vectors corresponding to the operating conditions of each time window. Cluster analysis is performed on the flow field feature vectors to obtain different flow field pattern categories and semantic annotation is performed. Each operating condition and its corresponding flow field feature vector and flow field pattern category are combined into a time series supervised learning sample set. The time series supervised learning sample set is used to train the predictive diagnosis model based on a time series neural network to obtain a trained predictive diagnosis model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for predicting the flow field and diagnosing combustion anomalies in a coal-fired boiler furnace as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for predicting the flow field and diagnosing combustion anomalies in a coal-fired boiler furnace as described in any one of claims 1 to 7.

Images