Multi-source node association model construction method based on thermal power generation and association anomaly analysis system

Abstract

The invention relates to the technical field of power system automation, discloses a thermal power generation-based multi-source node association model construction method and an association anomaly analysis system, and aims to solve the problems of inaccurate description of a complex interaction relationship and low anomaly recognition and fault tracing accuracy of a thermal power generation system in the prior art. The method comprises the steps of executing whole-process data granulation modeling, establishing a dynamic cause and effect graph model, implementing an entropy weight dynamic correction algorithm, integrating high-relevance additional technical means, executing relevance anomaly analysis, and regularly performing model efficiency evaluation and online learning. The system comprises a data acquisition and preprocessing module, a multi-source node association model construction module, an association anomaly analysis engine module and a performance evaluation and continuous optimization module. According to the technical scheme, accurate description of the thermal power generation complex system can be realized, the abnormity identification timeliness and the fault tracing accuracy can be improved, the system deviation caused by working condition disturbance and equipment aging can be overcome, the non-planned shutdown frequency can be reduced, and the system can be ensured to continuously adapt to the change of the operating environment.

Description

Technical Field

[0002] This invention belongs to the field of power system automation strategy, specifically involving a method for constructing a multi-source node association model based on thermal power generation and an association anomaly analysis system. Background Technology

[0004] Thermal power generation, as a pillar industry of my country's energy supply, relies heavily on the accurate perception and deep correlation analysis of multi-source heterogeneous data in complex systems for its operational safety, energy efficiency, and environmental performance. Modern thermal power units encompass multiple strongly coupled links, including fuel transportation, combustion reaction, steam-water circulation, power output, and environmental governance. These links exhibit significant nonlinear, time-varying, and spatial heterogeneity. Traditional monitoring and diagnostic strategies generally employ isolated parameter threshold alarms or static statistical models, which struggle to characterize the dynamic interactions across equipment, media, and time scales within the integrated boiler-turbine system. This results in persistent problems such as delayed anomaly identification, difficulty in fault tracing, and limited energy efficiency optimization.

[0005] Multi-source node correlation modeling has become a core direction for improving the intelligent operation and maintenance capabilities of thermal power plants. This strategy aims to integrate equipment physical mechanisms and operational data to construct a dynamic correlation network that reflects the causal logic and spatiotemporal evolution of real processes. However, existing methods still face severe challenges in terms of modeling granularity, causal expression, and time-varying adaptability: on the one hand, most models use fixed sampling frequencies and coarse-grained spatial partitioning, which cannot effectively capture critical transient processes such as combustion mutations and steam-water inertia; on the other hand, the correlation relationships are mostly based on linear correlations or static graph structures, ignoring the dynamic migration characteristics of parameter weights and causal paths under load fluctuations and coal quality changes; in addition, the systematic biases introduced by sensor drift and equipment aging factors have not been effectively corrected, further weakening the robustness and generalization ability of the model in real-world scenarios.

[0006] Existing strategies exhibit significant integration deficiencies when addressing the aforementioned issues: they lack the ability to perform granular collaborative modeling of the entire process—fuel-combustion-steam-water-power generation-environmental protection—and fail to establish a dynamic causal reasoning framework that integrates thermodynamic mechanisms, time delay compensation, and entropy-driven adaptive mechanisms. Furthermore, anomaly detection relies excessively on single-parameter limit judgments, neglecting the joint characterization effect of multi-dimensional state entropy changes and equipment deviation factors, resulting in low warning sensitivity and high false alarm rates. More importantly, existing systems struggle to achieve closed-loop linkage from anomaly perception and root cause localization to self-optimization of control strategies, failing to support rapid response and economical and environmentally friendly coordinated operation of thermal power units in a grid environment with high renewable energy penetration. Therefore, there is an urgent need for a multi-source node correlation model construction method that deeply integrates spatiotemporal registration, dynamic causal graphs, entropy weight correction, and multi-source preprocessing, along with a supporting correlation anomaly analysis system, to overcome the bottlenecks in accuracy, timeliness, and adaptability of traditional thermal power intelligent diagnostic strategies. Summary of the Invention

[0007] In view of this, the first objective of the present invention is to provide a system for analyzing the correlation anomalies of multiple source nodes in thermal power generation.

[0008] To address the aforementioned strategic problems, the strategic solution of this invention is: a system for analyzing the correlation anomalies of multiple source nodes in thermal power generation, characterized by comprising:

[0009] The data acquisition and preprocessing module collects multi-source heterogeneous data from the entire thermal power generation process, performs spatiotemporal registration and equipment state deviation factor preprocessing on the data, and outputs standardized feature vectors. Among them, spatiotemporal registration achieves time alignment of multi-source data through a high-precision time synchronization strategy, and establishes a mapping relationship based on the spatial structure of equipment to achieve spatial alignment of data. Equipment state deviation factor preprocessing generates state deviation factors for various types of equipment by constructing equipment mechanism simulation models, simulating the equipment operation process, and combining measured data.

[0010] Multi-source node association model construction module: Taking feature vectors and state deviation factors as inputs, it generates a dynamic association network model through data granular modeling strategy, dynamic causal graph modeling strategy and entropy weight dynamic correction algorithm.

[0011] The correlation anomaly analysis engine module is configured with an anomaly diagnosis strategy and an adaptive optimization strategy. The anomaly diagnosis strategy is used to monitor the dynamic correlation network model in real time to generate anomaly information. The adaptive optimization strategy is used to match the corresponding entropy weight dynamic correction algorithm according to the anomaly information to generate adjustment parameters.

[0012] Performance evaluation and continuous optimization module: It is configured with a state prediction strategy and an accuracy calibration strategy. The state prediction strategy is used to generate predicted state information based on the dynamic correlation network model. The accuracy calibration strategy generates accuracy calibration instructions based on the deviation between the predicted state information and the actual state information.

[0013] Furthermore, in the multi-source node association model construction module, the dynamic causal graph modeling strategy is configured with process parameter nodes, equipment status nodes, and control command nodes; the forward reasoning mechanism embeds the mechanism equations of relevant thermal power generation processes to adjust key reaction ratio parameters in real time; the reverse tracing mechanism establishes a multi-class association path backtracking algorithm related to equipment failure, process anomalies, and external interference; the time delay compensation adopts an estimator structure, establishes a transfer function model for the large inertia characteristics of thermal power generation systems to perform lag correction, and incorporates equipment status deviation factors during the modeling process to calibrate node parameter deviations. The dynamic cause-effect graph modeling strategy comprehensively covers the process parameters, equipment status, and control commands of the thermal power system by configuring three types of nodes, avoiding the loss of correlations caused by incomplete parameter coverage. The forward reasoning embedding of process mechanism equations ensures the scientific nature of the adjustment of key reaction ratio parameters, conforming to the actual physicochemical processes of thermal power plants. The multi-path backtracking algorithm of the reverse tracing mechanism can locate the root cause of the fault from multiple dimensions such as equipment, process, and external interference, solving the problems of low efficiency and easy omission in traditional single-path tracing. Time delay compensation is used to correct the large inertia characteristics of thermal power systems, which can eliminate the impact of time lag in links such as steam and water systems on model accuracy. The modeling process incorporates equipment status deviation factors, which can further calibrate node parameter deviations, improve the accuracy of causal relationship characterization, and provide a more reliable model foundation for subsequent anomaly diagnosis and fault tracing.

[0014] Furthermore, the entropy weight dynamic correction algorithm includes calculating multiple types of entropy values ​​reflecting fuel combustion state, energy conversion efficiency, and pollutant emission state, and dynamically adjusting the weight coefficients of the multiple types of entropy values ​​according to different load conditions of thermal power generation to calculate the entropy value of each node through a weight iterative sub-algorithm; and setting a threshold control limit. When the entropy value of a node deviates from the historical average by more than the threshold control limit, the local model corresponding to that node is reconstructed, and the model parameters are corrected in combination with the corresponding state deviation factor during the reconstruction process. The entropy weight dynamic correction algorithm calculates multiple types of entropy values, which can comprehensively reflect the operating status of thermal power systems from three core dimensions: fuel combustion, energy conversion, and environmental emissions, avoiding misjudgments caused by single-dimensional assessments. The weight iterative sub-algorithm dynamically adjusts the entropy weights according to different load conditions, enabling the model to focus on key indicators under different scenarios such as low load and high load, improving the model's adaptability to changes in operating conditions. The threshold-controlled triggering of local model reconstruction can selectively adjust only the local model when node anomalies occur, avoiding the efficiency loss of full model reconstruction. The reconstruction process, combined with the state deviation factor to correct parameters, can further ensure the accuracy of the reconstructed model, reduce false alarms or missed alarms caused by parameter deviations, and enhance the robustness of the model.

[0015] Furthermore, the data granular modeling strategy includes refined processing of spatial, temporal, and feature dimensions. The spatial dimension involves hierarchical grid division based on the physical regions of the equipment; the temporal dimension adaptively adjusts the sampling step size according to system load changes; and the feature dimension integrates fuel physicochemical properties and combustion kinetic parameters, extracting standardized feature vectors through feature engineering. The hierarchical grid division of the spatial dimension in the data granular modeling strategy can accurately capture the local operating status of different physical regions of thermal power equipment, avoiding the omission of local anomalies caused by coarse-grained spatial division. The temporal dimension adaptively adjusts the sampling step size according to load changes, using a high sampling frequency during load fluctuations to ensure data timeliness, and reducing the sampling frequency during stable periods to reduce data redundancy, balancing monitoring accuracy and data processing efficiency. The feature dimension integrates multiple core parameters such as fuel, combustion, and steam / water, extracting standardized feature vectors through feature engineering. This avoids modeling biases caused by single parameters or non-standardized data, providing a refined and standardized data foundation for multi-source node correlation models and improving the accuracy of model correlation analysis.

[0016] Furthermore, the data acquisition and preprocessing module includes a three-dimensional temperature field reconstruction strategy for the combustion field, an energy efficiency fingerprint analysis strategy for the steam-water system, a dynamic threshold optimization strategy for environmental emissions, a grid-generator coupling response prediction strategy, a multi-source data spatiotemporal registration strategy, and a fault mode transfer learning strategy. These six integrated strategies optimize data processing from different dimensions: the three-dimensional temperature field reconstruction strategy for the combustion field compensates for the inability of traditional single-point temperature measurement to fully grasp the furnace temperature distribution; the energy efficiency fingerprint analysis strategy for the steam-water system accurately identifies hidden performance losses in the steam-water circulation; the dynamic threshold optimization strategy for environmental emissions balances environmental compliance and operational economy; the grid-generator coupling response prediction strategy enhances the unit's adaptability to changes in grid load; the multi-source data spatiotemporal registration strategy further strengthens the temporal and spatial consistency of data; and the fault mode transfer learning strategy addresses the problem of insufficient fault samples for new or retrofitted units. The synergistic effect of these six strategies ensures that data preprocessing goes beyond mere "data cleaning," providing multi-dimensional, high-value data support for subsequent modeling and analysis, thus expanding the functional boundaries of the data preprocessing module.

[0017] Furthermore, the anomaly diagnosis strategy of the correlation anomaly analysis engine module includes: real-time extraction of the operating parameters of each node in the dynamic correlation network model, comparing the operating parameters with the preset node parameter benchmark range of the dynamic correlation network model, and generating single-parameter limit-exceeding anomaly information or multi-parameter correlation deviation anomaly information; the adaptive optimization strategy includes: determining the entropy value category to be adjusted according to the type of anomaly information, matching the corresponding entropy weight dynamic correction algorithm submodule, and generating adjustment parameters for the weight of the entropy value category, wherein the entropy value category includes combustion entropy, soda entropy, and emission entropy. The anomaly diagnosis strategy of the correlation anomaly analysis engine module, by comparing operating parameters with the benchmark range in real time, can simultaneously identify both single-parameter limit-exceeding and multi-parameter correlation deviation anomalies, avoiding missed detections of multi-parameter correlation anomalies caused by traditional single-parameter alarms; the adaptive optimization strategy matches the corresponding entropy weight dynamic correction algorithm submodule according to the anomaly type, generating adjustment parameters for different anomaly scenarios of combustion entropy, soda entropy, and emission entropy, avoiding poor optimization results caused by a one-size-fits-all parameter adjustment. This design enables the linkage of anomaly type identification, optimization strategy matching, and precise parameter adjustment, thereby improving the targeting and efficiency of anomaly handling and reducing system operation fluctuations caused by blind adjustments.

[0018] Furthermore, the state prediction strategy of the performance evaluation and continuous optimization module includes: combining historical operating data of the thermal power generation system with the node causal relationships of the dynamic correlation network model to generate predicted equipment operating parameters and environmental emission parameters for a preset future time period; the accuracy calibration strategy includes: calculating the deviation between the predicted state information and the actual state information of the same period, and when the deviation exceeds a preset deviation threshold, generating parameter adjustment instructions for the dynamic correlation network model, including node weight adjustment instructions or local model structure adjustment instructions. The state prediction strategy of the performance evaluation and continuous optimization module, combining historical data and model causal relationships, can anticipate the changing trends of equipment operating parameters and environmental emission parameters, providing maintenance personnel with forward-looking maintenance basis and avoiding passive responses to faults; the accuracy calibration strategy, by calculating the deviation between the prediction and the actual state and generating calibration instructions, can promptly detect model parameter drift or structural defects, avoiding abnormal misjudgments caused by accuracy decay during long-term model operation; when the deviation exceeds the threshold, targeted node weight or local model structure adjustment instructions are generated instead of full model reconstruction, which can reduce optimization costs while ensuring model accuracy. This strategy shifts model performance optimization from periodic full updates to on-demand precise calibration, improving the long-term stability and cost-effectiveness of the model.

[0019] Furthermore, the three-dimensional temperature field reconstruction strategy for the combustion field includes: collecting furnace wall temperature monitoring data, flame image data, and furnace gas concentration monitoring data; processing the data using a preset multi-source data fusion algorithm to reconstruct the three-dimensional temperature distribution within the furnace; and the energy efficiency fingerprint analysis strategy for the steam-water system includes: establishing a database of energy efficiency characteristic parameters for the steam-water system under different operating conditions; calculating in real time the matching degree between the current energy efficiency characteristic parameters of the steam-water system and the optimal operating condition characteristic parameters in the database; and generating energy efficiency matching results. The three-dimensional temperature field reconstruction strategy for the combustion field, by integrating multi-source data such as wall temperature, flame images, and gas concentration, can comprehensively and intuitively present the three-dimensional temperature distribution inside the furnace. This solves the problem that traditional single-point temperature measurement cannot detect local high-temperature areas, effectively avoiding safety hazards such as water-cooled wall tube rupture caused by local high temperatures. The energy efficiency fingerprint analysis strategy for the steam-water system, by establishing a database of energy efficiency characteristic parameters under different operating conditions, calculates the matching degree between the current operating condition and the optimal operating condition in real time. This can accurately locate hidden energy efficiency losses in the steam-water system, providing a clear direction for optimizing operating parameters. This avoids the problem that traditional single-parameter monitoring cannot identify energy efficiency bottlenecks, thereby improving the energy utilization efficiency of the steam-water system.

[0020] Furthermore, the environmental emission dynamic threshold optimization strategy includes: acquiring real-time electricity price data, pollutant treatment cost data, and excessive penalty standard data; calculating pollutant emission thresholds by combining them with emission rights trading-related models; and dynamically adjusting the emission thresholds according to the grid load period. The grid unit coupled response prediction strategy includes: constructing a coupled model of grid frequency, turbine speed regulation system parameters, and boiler combustion rate; generating load change prediction information in advance based on real-time grid load command data using a preset prediction algorithm; and optimizing the response parameters of the unit's automatic power generation control. The dynamic threshold optimization strategy for environmental emissions combines real-time electricity prices, treatment costs, and penalties for exceeding standards to calculate emission thresholds. It allows for moderately relaxed thresholds during peak hours to ensure power generation efficiency, while tightening thresholds during off-peak hours to strengthen environmental control. This addresses the problem of traditional fixed thresholds failing to balance environmental protection and economic efficiency, and reduces energy consumption and reductant waste in environmental protection equipment. The grid-generator coupling response prediction strategy, by constructing a coupled model of grid frequency, speed control system, and combustion rate, predicts load changes in advance and optimizes automatic generation control response parameters. This improves the unit's response speed and accuracy to high-frequency grid fluctuations, reduces turbine speed fluctuations, and enhances the unit's adaptability and grid performance compliance rate in environments with high penetration of new energy sources.

[0021] To achieve the second objective of this invention, a method for constructing a multi-source node correlation model for thermal power generation is provided: configured in the above-described anomaly analysis system for multi-source node correlation in thermal power generation, characterized by comprising the following steps:

[0022] S110 performs spatial dimension grid division, time dimension adaptive sampling, and feature dimension multi-parameter fusion on multi-source heterogeneous data of the entire thermal power generation process. At the same time, it integrates equipment state deviation factors into the granular data foundation to construct a refined data set.

[0023] S120 takes a refined dataset as input and uses forward reasoning, reverse tracing, and time delay compensation mechanisms to characterize the causal relationships and dynamic interaction characteristics between parameters. Among them, forward reasoning embeds process mechanism equations, reverse tracing establishes multi-class associated path backtracking algorithms, and time delay compensation performs lag correction for the system's inertial characteristics.

[0024] S130 receives node operation data from the dynamic cause-effect graph model, calculates entropy values ​​of multiple states in real time, and dynamically adjusts parameter weights according to operating conditions. When the node entropy value exceeds the set threshold, it triggers local model reconstruction.

[0025] S140 executes the equipment state deviation factor preprocessing strategy, the combustion field three-dimensional temperature field reconstruction strategy, the steam-water system energy efficiency fingerprint analysis strategy, the environmental emission dynamic threshold optimization strategy, the power grid unit coupling response prediction strategy, the multi-source data spatiotemporal registration strategy, and the fault mode transfer learning strategy, and feeds the results of each strategy back to the dynamic causal graph model and the entropy weight dynamic correction algorithm.

[0026] S150, based on the dynamic cause-effect graph model and the dynamic correction results of entropy weight, realizes closed-loop management of hierarchical early warning system, fault tracing and control strategy self-optimization;

[0027] S160 periodically verifies the accuracy of the dynamic causal graph model and the entropy weight dynamic correction algorithm through preset evaluation indicators, and trains and updates the model using thermal power generation operation data during low-load periods. Each step is interconnected, forming a systematic approach encompassing data foundation, model construction, strategy optimization, application implementation, and performance iteration. This approach not only constructs a high-precision multi-source node correlation model but also ensures the model's practicality and long-term adaptability in real-world thermal power scenarios through end-to-end design, providing a feasible technical path for the intelligent operation and maintenance of thermal power generation systems.

[0028] The main effects of this invention's strategy are reflected in the following aspects: By clearly defining the division of labor and data interaction logic of the four core modules of the system, the spatiotemporal registration strategy of the data acquisition and preprocessing module can eliminate the time asynchrony and spatial deviation of multi-source data; the equipment state deviation factor preprocessing can correct data errors caused by equipment aging and sensor drift, providing high-quality input for subsequent modeling; the multi-source node association model construction module integrates multiple modeling strategies to generate a dynamic model that can accurately characterize the multi-node association relationship of the thermal power system; the association anomaly analysis engine module achieves rapid anomaly identification and targeted parameter adjustment through the linkage of anomaly diagnosis and adaptive optimization; and the state prediction and accuracy calibration strategy of the performance evaluation and continuous optimization module can avoid the accuracy decay of the model during long-term operation. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method for constructing a multi-source node correlation model for thermal power generation and the correlation anomaly analysis system proposed in this invention;

[0030] Figure 2 This is a schematic diagram of the core principle framework of the multi-source node association model construction module in this invention;

[0031] Figure 3 This is a logical flow diagram of the data acquisition and preprocessing module in this invention;

[0032] Figure 4 This is a flowchart of the main stages of the full-process data granular modeling in this invention;

[0033] Figure 5 This is a schematic diagram of the core principle framework of dynamic cause-effect graph modeling in this invention;

[0034] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow of the correlation anomaly analysis engine module in this invention;

[0035] Figure 7 This is a comparison diagram of the technical effects / principles of the visualization and human-computer interaction modules in this invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the strategy and solution of the present invention can be more easily understood and mastered.

[0037] Reference Figure 1-7 As shown, the core of this invention consists of modules such as data acquisition, multi-source node association model construction, full-process data modeling, dynamic cause-effect graph modeling, association anomaly analysis engine, and visual human-computer interaction. First, referring to... Figure 3 As shown, the data acquisition and preprocessing module is designed as follows: Multi-source heterogeneous data refers to a data set that comes from different stages of thermal power generation and has different data types and formats, including real-time monitoring data from sensors, image data, acoustic signals, etc. The deployment of acquisition equipment for this type of data is completed first.

[0038] For the boiler system, 256 K-type thermocouple temperature sensors are arranged in the furnace body. K-type thermocouples are commonly used temperature sensing elements; their core is a thermocouple wire composed of nickel-chromium alloy and nickel-silicon alloy. Temperature measurement is achieved through the thermoelectric potential generated by the two metals when the temperature changes. The measurement range covers 950℃ to 1250℃, and the measurement accuracy can reach ±0.5℃. Sixty-four vibration sensors are installed in the tangential combustion areas at the four corners of the boiler. Vibration sensors are devices that convert mechanical vibration into electrical signals. Piezoelectric vibration sensors are used here, utilizing the piezoelectric effect of piezoelectric materials to achieve signal conversion. The measurement range is 0 to... With a sampling speed of 20mm / s and a sampling frequency of 2000Hz, it can accurately capture subtle changes in equipment vibration; 16 high-definition industrial CCD cameras are arranged on the top of the furnace. CCD cameras, or charge-coupled device cameras, convert optical signals into electrical signals through CCD image sensors. They have a resolution of 1920×1080 pixels and a frame rate of 25fps, which can clearly record the dynamic process of flame combustion; 8 acoustic sensors are arranged near the coal mill. Acoustic sensors are devices that receive sound waves and convert them into electrical signals. The frequency response range is 20Hz to 20kHz, which can effectively collect noise signals during the operation of the coal mill.

[0039] For the steam turbine system, 24 pressure transmitters and 16 temperature sensors are installed in bearing housings 1 to 6. The pressure transmitters convert pressure signals into standard electrical signals with a measurement accuracy of 0.1 MPa. The temperature sensors are also thermocouples, with a measurement range of 0 to 150°C. For the environmental protection system, 12 SO2 sensors and 10 NO sensors are installed at the inlet and outlet of the desulfurization tower and the inlet and outlet of the denitrification reactor, respectively. xThese gas sensors employ the principle of electrochemical detection, determining gas concentration by detecting the current generated during the chemical reaction between the gas and electrodes. The measurement accuracy is ±5 mg / m³. 3 Next, a spatiotemporal registration strategy is implemented. Spatiotemporal registration is a process that ensures data collected from different sources and at different times maintains consistency in both time and space. The GPS second pulse synchronization signal is used to calibrate the clocks of all acquisition devices. The GPS second pulse signal is a high-precision time reference signal provided by the Global Positioning System. After receiving the signal through the GPS receiver, the signal is amplified and filtered by the signal conditioning circuit. The signal conditioning circuit is an electronic circuit that amplifies, filters, reduces noise, and linearizes the weak raw electrical signal output by the sensor. It can eliminate signal interference and improve signal stability. The signal is then input to the clock module of each acquisition device, ultimately achieving a time synchronization accuracy of ±1μs. Based on the unit's 3D design drawings, a spatial rectangular coordinate system is established with a fixed concrete reference point at the zero-meter level of the unit as the origin. The spatial rectangular coordinate system is a three-dimensional coordinate system composed of three mutually perpendicular coordinate axes, used to determine the position of each point in space. The installation coordinates of each sensor, i.e., the x-axis, y-axis, and z-axis coordinate values, are recorded. By fitting the deviation between the actual installation position of the sensor and the theoretical design coordinates using the least squares method, the optimal fitting result of the coordinate deviation can be accurately calculated, generating a 3×3 spatial transformation matrix M. The spatial transformation matrix is ​​a mathematical tool used to realize spatial coordinate transformation. Calling this matrix M to perform spatial mapping on the sensor data can unify the data of sensors at different locations under the same spatial coordinate system. For data with a timestamp deviation exceeding 10ms, the timestamp is corrected using a linear interpolation algorithm. Linear interpolation estimates the value of any point between two known points using a linear relationship. The specific formula is: tcorrected = t1 + (ttarget - t1) × (d2 - d1) / (t2 - t1), where tcorrected is the corrected timestamp, t1 and t2 are the timestamps of two adjacent valid data points, d1 and d2 are the data values ​​at the corresponding timestamps, and ttarget is the original timestamp of the data to be corrected. This ensures that the time alignment accuracy after registration does not exceed 10ms and the spatial mapping error does not exceed 0.3m. Subsequently, equipment state deviation factor preprocessing is performed. The equipment state deviation factor is a quantitative indicator reflecting the difference between the actual operating state and the theoretical ideal state of the equipment, generated by constructing a simulation model of the equipment mechanism.

[0040] For the coal mill, a discrete element method (DEM) was used to construct the simulation model. The DEM is a numerical method used to simulate the motion and interaction of particle systems. By establishing the motion equations and contact mechanics models of the particles, the collision and compression behaviors of the particles were simulated. The number of coal particles in the model was set to 5 × 10⁻⁶. 5The particle radius is set to 2 to 5 mm based on the actual particle size distribution of the coal lumps fed into the furnace. The calculation time step is 0.1 ms, which is determined based on the characteristic time of particle collision to ensure the accuracy of the simulation results. The simulation model outputs theoretical values ​​of parameters such as the fineness of pulverized coal at the coal mill outlet and the grinding roller current. The fineness of pulverized coal is an indicator of the size of pulverized coal particles, usually expressed as the proportion of residue on a sieve of a certain aperture, which directly affects combustion efficiency. For the turbine rotor, a thermo-mechanical-fluid-structure interaction (TMO) simulation model is constructed. TMO simulation is a multi-physics simulation method that comprehensively considers heat conduction, structural stress and strain, fluid flow, and fluid-solid interaction. It can comprehensively reflect the rotor's operating state under complex conditions. The model integrates the rotor material's thermal expansion equation, the rotor's forced vibration equation, and the steam flow field equation. The thermal expansion equation is ΔL=α×L0×ΔT, where ΔL is the rotor's thermal expansion and α is the linear expansion coefficient of the rotor material. Different materials have different linear expansion coefficients; the alloy steel used in this rotor has an α value of 12×10. -6 1 / ℃, L0 is the original length of the rotor, ΔT is the temperature change of the rotor, i.e., the difference between the measured temperature and the initial temperature; the forced vibration equation is:

[0041]

[0042] In the formula, m is the mass of the rotor, and c is the damping coefficient of the rotor system. The damping coefficient reflects the system's ability to impede vibration, and is determined to be 5 × 10⁻⁶ through experimental testing. 3 N·s / m, where k is the rotor stiffness coefficient, which reflects the rotor's ability to resist deformation. Calculated based on the rotor's structural dimensions and material properties, it is 2 × 10⁻⁶. 7 N / m, x is the vibration displacement of the rotor, and F(t) is the excitation force acting on the rotor, which mainly comes from the fluctuation of steam force and rotor unbalance force. The model's mesh generation accuracy reaches 0.1mm. Mesh generation accuracy refers to the side length of the mesh element. High-precision meshes can improve the accuracy of simulation calculations. The coupled simulation model outputs theoretical values ​​of turbine rotor vibration, shaft temperature, and other parameters.

[0043] Combining the theoretical values ​​obtained from the simulation with the corresponding measured values ​​collected by the sensors, three types of state deviation factors are generated. The correction period is set to 1 hour, meaning that the simulation model is recalculated and the deviation factors are updated every 1 hour. The three types of deviation factors are: temperature deviation ΔT = measured temperature - theoretical temperature, vibration deviation ΔV = measured vibration velocity - theoretical vibration velocity, and pressure deviation ΔP = measured pressure - theoretical pressure. Finally, min-max standardization is used to process all data, including the deviation factors. Standardization is a preprocessing method that transforms data to a specific range, eliminating the influence of different dimensions on subsequent modeling. The formula is:

[0044]

[0045] In the formula, x represents the original data, x_min is the minimum value of this data type, x_max is the maximum value of this data type, and x' is the standardized data, with values ​​ranging from 0 to 1. The above processing extracts a 208-dimensional standardized feature vector. This feature vector is composed of multiple feature parameters; the 208-dimensionality indicates that the vector contains 208 feature parameters reflecting the system's operating state. This feature vector, along with the state deviation factor, serves as the input data for subsequent modules.

[0046] Reference Figure 2 The multi-source node association model construction module is designed as follows: taking the standardized feature vector and state deviation factor output by the data acquisition and preprocessing module as input, the data granular modeling strategy is first executed. Data granular modeling is a modeling method that refines and processes complex system data according to different dimensions, aiming to improve the relevance of the data and the accuracy of the modeling. The system is divided into 128 monitoring grids based on the physical structure of the equipment. Each grid is the smallest unit used to divide a spatial area. The boiler equipment is further subdivided into eight physical areas: furnace, superheater, reheater, economizer, air preheater, burner, water wall, and flue. The superheater is the boiler heating surface that heats saturated steam into superheated steam. The reheater is the device that reheats the low-temperature steam discharged from the high-pressure cylinder of the turbine. The economizer is the device that uses waste heat from the flue gas to heat the feedwater. The air preheater is the device that heats the air required for boiler combustion. The water wall is a group of water pipes arranged on the inner wall of the furnace to absorb radiant heat and protect the furnace wall. Each area is configured with 16 independent monitoring grids according to the size and structural characteristics of the equipment. The grid size is set according to the characteristics of the area. Due to the intense combustion reaction and rapid parameter changes in the furnace and burner areas, the grid size is set to 0.5m×0.5m×0.5m to achieve fine monitoring. The parameters in the flue area are relatively stable, so the grid size is enlarged to 2.0m×2.0m×2.0m to reduce computational costs while ensuring monitoring effectiveness. An adaptive variable step size sampling strategy is adopted in the time dimension. Adaptive variable step size sampling is a method that automatically adjusts the data acquisition interval according to changes in the system operating status. The system operating status is determined by the real-time computer group load change rate. The formula for calculating the load change rate is:

[0047]

[0048] In the formula, ΔP load P represents the rate of change of unit load. t The generator load at the current moment refers to the active power output by the generator set, P. t-1 This represents the unit load at the previous moment. When ΔP loadIf the rate of change exceeds ±2% / min, or if the gradient change of key parameters such as main steam temperature and furnace pressure (i.e., the ratio of parameter change to time change) exceeds a set threshold, it indicates that the system is in a period of load fluctuation with drastic parameter changes. The system will automatically switch to a 100ms sampling period to ensure that transient changes are captured. Under stable operating conditions (ΔP...),... load Maintaining a baseline sampling period of 1 second, where the sampling rate does not exceed ±1% / min and the parameter gradient change does not exceed the threshold, can effectively reduce data redundancy and lower data processing pressure. Seven categories of parameters are integrated into the feature dimension: fuel physicochemical properties (including calorific value, volatile matter, and ash content of coal; calorific value is the heat released by the complete combustion of a unit mass of coal; volatile matter is the proportion of gaseous products released during coal heating; ash content is the proportion of solid residue remaining after coal combustion); combustion kinetic parameters (including combustion rate, flame temperature, and excess air coefficient; combustion rate is the amount of fuel burned per unit time; excess air coefficient is the ratio of actual air supply to the theoretical air volume required for combustion); steam-water thermodynamic parameters (including main steam pressure, main steam temperature, and feedwater flow rate; main steam pressure and temperature are core parameters for measuring steam quality; feedwater flow rate is the volume or mass flow rate of feedwater supplied to the boiler); equipment vibration spectrum (including vibration amplitude, frequency, and phase; vibration amplitude is the maximum displacement or velocity of vibration; frequency is the number of vibration cycles per unit time; phase is the starting position of vibration in time); and environmental emission factors (including SO2 concentration and NO). X Concentration of smoke and dust; SO2 is the chemical formula for sulfur dioxide, a major acidic pollutant produced by coal combustion; NO... x It is a general term for nitrogen oxides, including NO, NO2, etc. Smoke and dust are solid particulate matter produced by combustion. Ambient temperature and humidity include ambient temperature and relative humidity. Power grid load instructions include AGC (Automatic Generation Control) load setpoints. AGC instructions are issued by the power grid dispatch center to adjust generator output to maintain grid frequency and voltage stability. Principal Component Analysis (PCA) algorithm is used for feature dimensionality reduction and extraction. PCA is a statistical method that maps high-dimensional data to a low-dimensional space through linear transformation. Its core is to identify the main direction of change in the data by calculating the covariance matrix. The formula for calculating the covariance matrix is:

[0049]

[0050] In the formula, C is the covariance matrix, n is the sample size, and x is the variance matrix. i Let be the feature data vector of the i-th sample, and μ be the mean vector of all sample feature data. Solve for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​reflect the variance of the data in the direction of the corresponding eigenvector. The larger the variance, the more information is contained in that direction. Select the features corresponding to the eigenvectors with eigenvalues ​​greater than 1, and finally extract the 208-dimensional standardized feature vector.

[0051] Reference Figure 4 Next, a dynamic causal graph model is established. A dynamic causal graph is a graphical model used to depict the causal relationships between variables and their dynamic changes. First, 152 nodes are determined. Nodes are the basic units representing system parameters or states in the dynamic causal graph, covering 80 process parameter nodes (e.g., main steam temperature, SO2 concentration), 50 equipment state nodes (e.g., bearing temperature, grinding roller current), and 22 control command nodes (e.g., coal feed command, desulfurizer flow command). The coal feed command is an operational signal controlling the amount of fuel fed into the coal mill or furnace, and the desulfurizer flow command is a control signal adjusting the amount of desulfurizer supplied to the desulfurization tower. A forward reasoning mechanism is embedded in the mechanistic equation of the core process of thermal power generation. The mechanistic equation is a mathematical equation based on physicochemical principles describing the parameter changes during the process. Taking the desulfurization reaction as an example, the relationship between SO2 generation and desulfurizer dosage is calculated using the desulfurization reaction kinetic equation. This equation reflects the rate characteristics of the desulfurization reaction, and its expression is: In the formula Let k be the SO2 formation rate, and k0 be the pre-exponential factor, which reflects the collision frequency of reactant molecules per unit time and has a value range of 1.2 × 10⁻⁶. 5 Up to 1.5×10 5 mol / (m 3 ·s), the specific value is calibrated based on the sulfur content and other coal quality characteristics of the coal fed into the furnace, E a This represents the activation energy for the desulfurization reaction. The activation energy is the minimum energy required for the reaction to occur; here, it is taken as 8.5 × 10⁻⁶. 4 J / mol, R is the ideal gas constant, which is fixed at 8.314 J / (mol·K), T is the desulfurization reaction temperature, which is obtained by actual measurement from the desulfurization tower outlet temperature sensor and converted by adding 273.15 degrees Celsius to C. S The concentration of sulfur in the fuel is calculated from the sulfur content data in the coal quality analysis report combined with the coal density. The oxygen concentration (C_O2) is measured by the oxygen sensor at the inlet of the desulfurization tower. Based on this equation, the Ca / S molar ratio (the molar ratio of calcium in the desulfurizing agent to sulfur in the fuel) is calculated in real time. The desulfurizing agent is typically a calcium-containing compound such as limestone, which achieves desulfurization through reaction with SO2. This ratio is dynamically adjusted within the range of 1.02 to 1.15 to ensure desulfurization efficiency while avoiding waste of the desulfurizing agent.

[0052] The reverse tracing mechanism establishes three types of related path backtracking algorithms: equipment failure path, process anomaly path, and external interference path. The related path refers to the causal chain between the abnormal phenomenon and the root cause. The maximum backtracking depth is set to 5 layers of causal relationship. The backtracking depth is the number of layers of causal relationship traced upward from the abnormal node. Setting it to 5 layers can ensure the comprehensiveness of the tracing while avoiding over-tracing. Taking the abnormal phenomenon of excessive turbine vibration as an example, the equipment failure path is as follows: the bearing temperature rises, leading to a decrease in lubricating oil viscosity. Lubricating oil viscosity is an indicator of the flow resistance of lubricating oil. A decrease in viscosity will reduce the lubrication effect, thereby reducing the thickness of the lubricating oil film and ultimately causing increased vibration. The process abnormality path is that the steam humidity rises, causing water erosion of the blades. Steam humidity is the mass ratio of liquid water in steam. Water erosion is the scouring damage of high-speed wet steam on the turbine blades, leading to rotor imbalance and thus increased vibration. The external interference path is that the power grid frequency fluctuation causes rotor torque imbalance. The power grid frequency is a core indicator for measuring the stability of power grid operation. The standard power grid frequency in my country is 50Hz. Fluctuations will affect the rotor speed and torque, ultimately leading to increased vibration. By traversing these three types of paths using a depth-first search algorithm, which is a search algorithm that traverses the nodes of a tree along its depth, it can systematically find all possible causal paths, thereby accurately locating the root cause of the anomaly. The time delay compensation adopts a Smith predictor structure. The Smith predictor is a control algorithm used to compensate for pure time delay characteristics in industrial processes. Pure time delay refers to the characteristic that the output only begins to respond after a change in input. A transfer function model is established for the large inertia characteristics of the soda system. The transfer function is a mathematical tool used to describe the relationship between the input and output of a linear system, and its expression is:

[0053]

[0054] In the formula, G(s) is the transfer function of the steam-water system, K is the system gain, which reflects the amplification factor of the output to the input, and its value ranges from 1.8 to 2.2. It is obtained through experimental calibration based on the correspondence between the main steam flow rate and the coal feed rate. T is the time constant, which reflects the speed of the system response, and its value ranges from 25 to 35 s. It is determined by the step response experiment of the steam-water system. The step response experiment applies a sudden step input signal to the system, records the output change curve over time, and obtains the system characteristic parameters through curve analysis. τ is the lag time, which refers to the time interval from the input change to the output starting to change, and its value ranges from 30 to 60 s. It is determined by actual measurement based on the response time of the main steam temperature to the change in coal feed rate. This model corrects the time lag of the steam-water system to ensure the accuracy of the causal relationship. At the same time, the equipment state deviation factor is incorporated into the modeling process to calibrate the node parameters such as the theoretical value of the main steam temperature, reducing the impact of parameter deviations on the model.

[0055] Finally, the entropy weight dynamic correction algorithm is implemented. This algorithm, based on information entropy theory, dynamically adjusts the weights of parameters according to their uncertainty. Information entropy is a physical quantity that measures the degree of disorder in a system; the higher the parameter uncertainty, the greater the entropy value, and the higher the corresponding weight should be to highlight its influence. First, the entropy values ​​for the three states are calculated. Entropy is an indicator used to measure the degree of uncertainty or disorder in a system; the larger the entropy value, the higher the uncertainty of the system state. Combustion entropy is calculated based on fuel reaction sufficiency, which is an indicator reflecting the completeness of fuel combustion. Its calculation formula is η = Q. act / Q theo In the formula, η represents the degree of reaction sufficiency, and Q... act The actual heat released during combustion is calculated from the furnace temperature and flue gas flow rate using the heat balance equation. This equation is based on the law of conservation of energy, which states that the energy input to the system equals the sum of the energy output from the system and the energy changes within the system. Q theo The theoretical heat release during combustion is calculated by multiplying the calorific value of the coal by the coal feed rate. The formula for calculating the combustion entropy is... In the formula, S_comb is the combustion entropy, n is the number of combustion zones (taken as 8, corresponding to the 8 physical zones of the boiler), and p i The percentage of complete reaction in the i-th combustion zone is calculated using the following formula: η i Let represent the reaction sufficiency of the i-th combustion zone. The steam-water entropy is calculated based on energy conversion efficiency, which is an indicator reflecting the effectiveness of the steam-water system in converting thermal energy into steam energy. Its calculation formula is ε = h out / h in In the formula, ε is the energy conversion efficiency, and h out h represents the enthalpy of the steam outlet. in The enthalpy value of the feedwater inlet is a thermodynamic parameter reflecting the energy state of a substance, combining internal energy and driving work. It can be obtained from the steam thermodynamic property table, which records the thermodynamic parameters of water and steam at different temperatures and pressures. The formula for calculating the entropy of steam and water is consistent with that of combustion entropy, only with p... i Replace with the energy conversion efficiency percentage of each steam-water heat exchange zone. Emission entropy is quantified by the pollutant concentration distribution entropy value, and the calculation formula is also the same as the entropy value calculation form described above, p i Replace with the percentage of pollutant concentrations at the outlets of each environmental protection device. The weight iteration uses a modified TOPSIS method. TOPSIS, or the Approximation-to-Ideal-Solution Ranking Method, is a multi-attribute decision analysis method that ranks samples based on their distance from the ideal solution. The modified TOPSIS method introduces entropy weights to optimize the weight determination process, making the weights more consistent with actual working conditions. First, construct the decision matrix:

[0056] F = (f ij )m×n

[0057] In the formula, F is the decision matrix, m is the sample size, and n is the number of indicators. Here, the indicators are combustion entropy, steam / water entropy, and emission entropy. ij Let f be the value of the j-th index for the i-th sample. Determine the positive ideal solution f. j + =max(f 1j ,f 2j ,...,f mj ) and negative ideal solution f j - =min(f 1j ,f 2j ,...,f mj The positive ideal solution is a vector composed of the optimal values ​​of each indicator, and the negative ideal solution is a vector composed of the worst values ​​of each indicator. The distance from each sample to the positive ideal solution is calculated using the following formula:

[0058]

[0059] In the formula Let w be the distance from the i-th sample to the positive ideal solution. j Let w be the weighting coefficient for the j-th indicator. The weighting coefficient is dynamically adjusted according to different load conditions of thermal power generation. Under low load conditions, i.e., when the unit load does not exceed 300MW, combustion stability is crucial. Unstable combustion can easily lead to flameout or reduced efficiency. Therefore, the combustion entropy weight is increased by 30%, i.e., w comb =w comb0 ×1.3, where w_comb is the adjusted combustion entropy weight, w comb0 The basic weight for combustion entropy is obtained through initial calculation using the improved TOPSIS method. Under high-load conditions (i.e., unit load not less than 500MW), environmental emission requirements are higher, and pollutant generation increases, necessitating close monitoring. Therefore, the emission entropy weight is increased by 25%, i.e., w emit =w emit0 ×1.25, where w_emit is the adjusted emission entropy weight, w emit0 The basic weight for emission entropy is set at ±2σ. Anomaly detection is set with a control limit of ±2σ, where σ is the standard deviation of historical entropy values. Standard deviation is a statistic reflecting the degree of data dispersion; a larger dispersion indicates more drastic data fluctuations. The calculation formula is derived from entropy data obtained from the past 30 days of normal system operation.

[0060] In the formula, σ is the standard deviation, n is the number of historical data samples, and Si is the entropy value of the i-th sample. This represents the average historical entropy value. When the entropy value of a node deviates from the historical average by more than this control limit, it indicates that the node's operating state is abnormal, triggering a local model reconstruction. Local model reconstruction only rebuilds the local model where the abnormal node is located, rather than reconstructing the entire model, which can significantly improve reconstruction efficiency. During the reconstruction process, the corresponding device state deviation factor is called to correct model parameters such as node weights and causal relationship strength, ensuring the accuracy of the reconstructed model. The reconstruction time is controlled within 5 seconds to meet the needs of real-time monitoring. Finally, a dynamic correlation network model is generated. The dynamic correlation network model is a network structure composed of nodes and the correlation relationships between nodes, which can intuitively reflect the mutual influence of various parameters.

[0061] Reference Figure 5 The correlation anomaly analysis engine module is designed as follows: The correlation anomaly analysis engine is the core module for anomaly identification and parameter optimization in the system. It receives the dynamic correlation network model output by the multi-source node correlation model construction module, configures the anomaly diagnosis strategy and adaptive optimization strategy, and executes them. The implementation of the anomaly diagnosis strategy consists of three steps. The first step is to extract the operating parameters of 152 nodes in the dynamic correlation network model in real time, such as bearing temperature, main steam pressure, and SO2 concentration. The extraction frequency is consistent with the data acquisition cycle, i.e., 100ms for load fluctuations and 1s for stable periods, ensuring that the operating status of the nodes can be reflected in real time. The second step is to determine the baseline range of the node parameters. The baseline range refers to the value interval of the node parameters under normal operating conditions. Based on the unit's normal operating data over the past year, it is calculated using the 3σ rule. The 3σ rule is a statistical method based on the normal distribution characteristics of data. The normal distribution is a symmetrical bell-shaped distribution of data centered on the mean. This rule assumes that approximately 99.73% of normal data will fall within the range of mean ± 3σ. This range can be used as the baseline range for the parameters. For example, the baseline range for bearing temperature is determined to be 80℃ to 90℃, the baseline range for main steam pressure is 25MPa to 26MPa, and the baseline range for SO2 concentration is no more than 35mg / m³. 3The core step in determining whether parameters are abnormal is to compare the real-time extracted operating parameters with the corresponding benchmark range. The third step is to generate abnormal information. If a single node parameter exceeds the benchmark range, a single parameter exceeding the limit abnormality information is generated. For example, if the temperature of bearing No. 3 reaches 92℃, exceeding the benchmark range of 80℃ to 90℃, the abnormal information "Bearing No. 3 temperature 92℃ exceeds the benchmark range of 80℃ to 90℃" is generated. If three or more node parameters with causal relationships simultaneously exceed the benchmark range, causal relationship means that there is a clear physical or technological causal relationship between the nodes. For example, an increase in main steam temperature will lead to an increase in turbine work, which may cause an increase in vibration. At the same time, an abnormal main steam temperature will also affect the lubricating oil temperature, leading to a change in lubricating oil pressure. Therefore, the three node parameters of increased main steam temperature, increased turbine vibration, and decreased lubricating oil pressure have a causal relationship. In this case, multi-parameter correlation deviation abnormal information is generated, and the causal path between the related nodes is marked to facilitate subsequent source tracing analysis. The implementation of the adaptive optimization strategy also consists of three steps. The first step is anomaly identification. Based on the domain of the parameters involved in the anomaly, the anomaly type is divided into combustion anomalies, steam-water anomalies, and emission anomalies. If the anomaly involves combustion zone parameters such as flame temperature and excess air coefficient, it is determined to be a combustion anomaly, and the weight of combustion entropy needs to be adjusted accordingly to enhance the model's sensitivity to combustion system anomalies. If it involves steam-water system parameters such as main steam temperature and feedwater flow rate, it is determined to be a steam-water anomaly, and the weight of steam-water entropy needs to be adjusted accordingly. If it involves environmental emission parameters such as SO2 concentration and NO... xIf the concentration is high, it is identified as an emission anomaly, and the weight of the emission entropy is adjusted accordingly. The second step is to match the entropy weight dynamic correction algorithm submodule. The entropy weight dynamic correction algorithm contains multiple submodules, each corresponding to the weight adjustment of different entropy values. Each submodule has built-in calculation logic and weight adjustment rules for the corresponding entropy value. For example, the combustion anomaly matching combustion entropy weight adjustment submodule calls the calculation logic of combustion entropy weight in the entropy weight dynamic correction algorithm, and combines the current operating parameters such as unit load, coal quality, calorific value and anomaly degree such as parameter deviation range to generate the adjustment parameters of combustion entropy weight, such as "increase combustion entropy weight from 0.3 to 0.39". The third step is parameter update and information feedback. The generated adjustment parameters are sent to the multi-source node association model construction module to update the node weights of the dynamic association network model, so that the model can make adaptive adjustments for the current anomaly and improve the identification accuracy of similar anomalies. At the same time, the adjustment parameters and anomaly information are sent to the visualization and human-computer interaction module for display, providing operators with intuitive reference information. Taking excessive turbine vibration under a specific operating condition as an example, the anomaly diagnosis strategy first extracts the parameters of the turbine vibration nodes. The measured vibration velocity is 15 mm / s, while its baseline range is 5 mm / s to 12 mm / s. Simultaneously, it extracts the parameters of related nodes, finding that the temperature of bearing No. 3 is 91℃, exceeding the baseline range of 80℃ to 90℃, and the lubricating oil pressure is 0.2 MPa, below the baseline range of 0.25 MPa to 0.3 MPa. Based on the fact that these three causally related node parameters are all out of range, it is determined to be a multi-parameter correlation deviation anomaly. The adaptive optimization strategy identifies this anomaly. Since it involves equipment status parameters such as bearing temperature and lubricating oil pressure, and the lubricating oil system is an auxiliary support system for the steam-water system, it is determined to be an equipment fault anomaly. The steam-water entropy and combustion entropy weight adjustment submodule is matched, and after comprehensively analyzing the correlation between the root cause of the anomaly and the combustion system, adjustment parameters are generated: "Increase the weight of lubricating oil pressure-related nodes by 20%, and fine-tune the combustion entropy weight by 5%." After updating this parameter to the model, the model can more accurately locate the root cause of the anomaly and ultimately identify the causal chain that "insufficient lubricating oil pump flow leads to a drop in lubricating oil pressure. Lubricating oil pump flow refers to the volume of oil output by the lubricating oil pump per unit time. Insufficient flow will directly lead to a decrease in system pressure, which in turn reduces the thickness of the lubricating oil film. The lubricating oil film is the oil layer between the bearing and the journal. Insufficient thickness will lead to direct metal-to-metal contact, ultimately causing increased turbine vibration." The model then outputs the operation suggestion of "checking the operating status of the lubricating oil pump" to provide clear guidance for maintenance personnel.

[0062] Reference Figure 7In addition, a visualization and human-computer interaction module is also configured. The design of this module is as follows: This module is a crucial carrier for information exchange between the system and operators. Its core function is to display system operation status anomaly information and other data to operators in an intuitive and easy-to-understand format, and to receive operator instructions such as parameter setting anomaly confirmation. This module receives system status anomaly nodes and parameter coupling relationship data output by the correlation anomaly analysis engine module, and performs visualization processing and interactive function design. First, a 3D factory map is constructed using the Unity3D engine. Unity3D is a professional 3D game development and visualization engine with powerful 3D modeling, real-time rendering, and interactive control capabilities, enabling realistic scene display and a smooth operating experience. The scale of the 3D factory map is set to 1:500. The scale refers to the ratio of distance on the map to actual distance; a 1:500 scale means that 1cm on the map represents 5m in reality. This ensures map clarity while fully displaying the equipment layout of the entire factory area, including the relative positions of core equipment such as boilers, turbines, generators, desulfurization, and denitrification systems. The actual installation locations of the previously divided 128 monitoring grids and 256 sensors were precisely marked on a 3D map. Different colors and shapes of icons were used to distinguish different types of equipment and sensors; for example, temperature sensors were represented by red dots, pressure sensors by blue squares, and vibration sensors by yellow triangles, facilitating quick identification by operators. Abnormal nodes were highlighted with flashing red icons at a frequency of 2 times per second. This frequency quickly attracted the operator's attention without causing visual fatigue due to excessive flashing. Clicking the icon would bring up a detailed information window, allowing operators to view the specific numerical reference ranges and correlation paths of the abnormal parameters for that node, as well as system-generated adjustment suggestions, providing comprehensive anomaly information.

[0063] Reference Figure 6The performance evaluation and continuous optimization module is designed as follows: This module is crucial for ensuring the long-term stable operation of the system and the accuracy of the model. By configuring state prediction and accuracy calibration strategies, it achieves forward-looking prediction of the system's operating state and dynamic optimization of model parameters. Simultaneously, combined with performance evaluation and online learning mechanisms, it ensures the model can adapt to changes in operating conditions. The state prediction strategy employs a Long Short-Term Memory (LSTM) network. LSTM is a special type of recurrent neural network with a gating mechanism, capable of effectively processing and learning long-term dependencies in time-series data, making it highly suitable for predicting data with significant time-series characteristics, such as that from thermal power generation. This strategy uses the standardized feature vector state deviation factor and dynamic correlation network model node data from the past 30 days as the training set. The training set is the sample data set used to train the model, and the number of samples is calculated based on a 30-day sampling frequency to ensure it includes data from different load conditions and operating states. The constructed load command prediction model takes as input the grid frequency AGC command unit load data from the previous hour. This data reflects the grid's operating status and the current output of the units. The output is the load forecast for the next 15 minutes. The 15-minute prediction time window is determined by comprehensively considering grid dispatching needs and unit response speed, providing sufficient preparation time for unit operation adjustments. The model's prediction accuracy is controlled within ±1.2%. Prediction accuracy refers to the relative error between the predicted and actual values; this level of accuracy meets the requirements of unit operation control. An accuracy calibration strategy is used to monitor and correct the model's prediction deviation. First, the deviation between the predicted state information and the actual state information of the same period is calculated. The formula for calculating the deviation is:

[0064]

[0065] In the formula, Δ is the deviation value, and y pred To predict state information such as load, y actThis refers to the actual state information during the same period, such as the actual load. When Δ exceeds the preset deviation threshold, which is determined based on the importance and control requirements of different parameters (set to 5% here), it indicates that the model's prediction accuracy can no longer meet the requirements and parameter adjustment is necessary. At this point, the system generates parameter adjustment instructions for the dynamic correlation network model. These instructions are divided into two types: node weight adjustment instructions and local model structure adjustment instructions. For example, "adjust the correlation weight between the turbine vibration node and the lubricating oil pressure node from 0.6 to 0.7" is a node weight adjustment instruction. By adjusting the correlation weights between nodes, the causal relationship characterization of the model is optimized. When there are defects in the local model structure, a local model structure adjustment instruction is generated, such as adding or deleting some correlation paths. Performance evaluation uses a comprehensive evaluation of both the root mean square error (RMSE) and the mean absolute percentage error (MAPE). This dual-indicator evaluation can more comprehensively reflect the model's prediction performance and avoid the limitations of a single indicator. For the combustion model, the focus is on the accuracy of temperature prediction. The formula for calculating RMSE is:

[0066]

[0067] In the formula, RMSE is the root mean square error, n is the sample size, and T is the mean square error. pred,i Let T be the predicted temperature of the i-th sample. act,i Let be the actual temperature of the i-th sample. This index reflects the overall deviation between the predicted and actual temperatures, and the RMSE of the combustion model should not exceed 5°C. For emission models, the focus is on the accuracy of pollutant concentration prediction. The MAPE calculation formula is:

[0068]

[0069] In the formula, MAPE is the mean absolute percentage error, n is the sample size, and C is the mean absolute percentage error. pred,i Let C be the predicted concentration of the i-th sample. act,iThe actual concentration of the i-th sample is represented by this index, which visually reflects the relative error of the concentration prediction as a percentage. The MAPE of the emission model is required to be no more than 3%, and the performance evaluation cycle is set to 7 days. That is, the RMSE and MAPE of the model are calculated and evaluated every 7 days to promptly grasp changes in model performance. The online learning mechanism is the core of ensuring the long-term adaptability of the model. It is executed daily from 02:00 to 04:00, a period typically characterized by low grid load and relatively stable unit operation, ensuring that model training will not interfere with normal unit operation. During online learning, all operational data from the past 24 hours is first imported. This data covers information on different load conditions, coal quality changes, and equipment operating status from the previous day, providing sufficient sample support for model updates. The momentum gradient descent method is used for model training and updates. Momentum gradient descent is an optimized gradient descent algorithm that accelerates model convergence and avoids getting trapped in local optima by introducing a momentum term to accumulate previous gradient information. The key parameters of this algorithm are set as follows: learning rate 0.01, momentum coefficient 0.9, and learning rate decay coefficient 0.95. The learning rate refers to the step size of each parameter update; a learning rate of 0.01 ensures the effectiveness of parameter updates while preventing excessively large update increments that could lead to model instability. The momentum coefficient reflects the influence of previous gradients on the current update; a setting of 0.9 allows the model to converge quickly in regions with small gradient changes. The learning rate decay coefficient is used to gradually decrease the learning rate with each training iteration; a decay coefficient of 0.95 allows the model to gradually stabilize in the later stages of training, avoiding parameter oscillations. Through an online learning mechanism, the model can continuously absorb new operational data features, adapting to various operating conditions such as changes in coal composition, equipment aging and wear, and fluctuations in environmental temperature and humidity, maintaining stable prediction and diagnostic accuracy. After being applied to a 660MW ultra-supercritical coal-fired unit and undergoing three consecutive months of operation verification, the system demonstrated excellent performance, with the following specific effects: Timeliness of anomaly identification improved by 60%, reducing the average anomaly response time from 15 seconds in traditional monitoring systems to 6 seconds, enabling faster detection of potential faults; Fault tracing accuracy improved by 45%, from 50% in traditional methods to 95%, significantly reducing the troubleshooting time and workload for maintenance personnel; Unplanned shutdowns decreased by 50%, effectively reducing production losses due to equipment failures; Coal consumption for power generation decreased by 1.8g / kWh. Coal consumption for power generation is a core indicator for measuring the economic efficiency of thermal power units, and this reduction signifies a significant improvement in unit power generation efficiency; Energy consumption of environmental protection equipment decreased by 12%, reducing environmental operating costs while ensuring pollutant emissions meet standards, fully meeting the core requirements of safe, stable, efficient, energy-saving, and environmentally compliant operation of thermal power generation systems.

[0070] Of course, the above are just typical examples of the present invention. In addition, the present invention can have many other specific implementations. All strategies and solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A system for analyzing anomalies in multi-source nodes of thermal power plants, characterized in that, include: The data acquisition and preprocessing module collects multi-source heterogeneous data from the entire thermal power generation process, performs spatiotemporal registration and equipment state deviation factor preprocessing on the data, and outputs standardized feature vectors. Among them, spatiotemporal registration achieves time alignment of multi-source data through a high-precision time synchronization strategy, and establishes a mapping relationship based on the spatial structure of equipment to achieve spatial alignment of data. Equipment state deviation factor preprocessing generates state deviation factors for various types of equipment by constructing equipment mechanism simulation models, simulating the equipment operation process, and combining measured data. Multi-source node association model construction module: Taking feature vectors and state deviation factors as inputs, it generates a dynamic association network model through data granular modeling strategy, dynamic causal graph modeling strategy and entropy weight dynamic correction algorithm. The correlation anomaly analysis engine module is configured with an anomaly diagnosis strategy and an adaptive optimization strategy. The anomaly diagnosis strategy is used to monitor the dynamic correlation network model in real time to generate anomaly information. The adaptive optimization strategy is used to match the corresponding entropy weight dynamic correction algorithm according to the anomaly information to generate adjustment parameters. Performance evaluation and continuous optimization module: It is configured with a state prediction strategy and an accuracy calibration strategy. The state prediction strategy is used to generate predicted state information based on the dynamic correlation network model. The accuracy calibration strategy generates accuracy calibration instructions based on the deviation between the predicted state information and the actual state information.

2. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 1, characterized in that, In the multi-source node association model construction module, the dynamic cause-effect graph modeling strategy is configured with process parameter nodes, equipment status nodes, and control command nodes; the forward reasoning mechanism embeds the mechanism equations of thermal power generation related processes to adjust key reaction ratio parameters in real time; the reverse tracing mechanism establishes a multi-class association path backtracking algorithm related to equipment failure, process anomalies, and external interference; the time delay compensation adopts an estimator structure, establishes a transfer function model for the large inertia characteristics of thermal power generation systems to perform lag correction, and incorporates equipment status deviation factors during the modeling process to calibrate node parameter deviations.

3. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 1, characterized in that, The entropy weight dynamic correction algorithm includes calculating multiple types of entropy values ​​that reflect fuel combustion status, energy conversion efficiency, and pollutant emission status, and dynamically adjusting the weight coefficients of the multiple types of entropy values ​​according to different load conditions of thermal power generation through a weight iterative sub-algorithm to calculate the entropy value of each node. A threshold control limit is set. When the entropy value of a node deviates from the historical average by more than the threshold control limit, the local model corresponding to that node is reconstructed. During the reconstruction process, the model parameters are corrected in combination with the corresponding state deviation factor.

4. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 1, characterized in that, The data granular modeling strategy includes refined processing of spatial, temporal, and feature dimensions. The spatial dimension is divided into hierarchical grids according to the physical area of ​​the equipment. The temporal dimension adaptively adjusts the sampling step size according to the system load changes. The feature dimension integrates the physical and chemical properties of the fuel and combustion dynamics parameters, and extracts standardized feature vectors through feature engineering.

5. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 1, characterized in that, The data acquisition and preprocessing module includes a three-dimensional temperature field reconstruction strategy for the combustion field, an energy efficiency fingerprint analysis strategy for the steam-water system, a dynamic threshold optimization strategy for environmental emissions, a power grid unit coupling response prediction strategy, a multi-source data spatiotemporal registration strategy, and a fault mode transfer learning strategy.

6. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 1, characterized in that, The anomaly diagnosis strategy of the correlation anomaly analysis engine module includes: extracting the operating parameters of each node in the dynamic correlation network model in real time, comparing the operating parameters with the preset node parameter benchmark range of the dynamic correlation network model, and generating single parameter limit exceedance anomaly information or multi-parameter correlation deviation anomaly information; the adaptive optimization strategy includes: determining the entropy value category to be adjusted according to the type of anomaly information, matching the corresponding entropy weight dynamic correction algorithm submodule, and generating adjustment parameters for the weight of the entropy value category, wherein the entropy value category includes combustion entropy, soda entropy, and emission entropy.

7. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 1, characterized in that, The state prediction strategy of the performance evaluation and continuous optimization module includes: combining historical operating data of the thermal power generation system with the node causal relationships of the dynamic correlation network model to generate equipment operating parameter prediction information and environmental emission parameter prediction information for a future preset time period; the accuracy calibration strategy includes: calculating the deviation value between the predicted state information and the actual state information of the same period, and when the deviation value exceeds a preset deviation threshold, generating parameter adjustment instructions for the dynamic correlation network model, including node weight adjustment instructions or local model structure adjustment instructions.

8. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 5, characterized in that, The three-dimensional temperature field reconstruction strategy for the combustion field includes: collecting furnace wall temperature monitoring data, flame image data, and furnace gas concentration monitoring data; processing the data using a preset multi-source data fusion algorithm to reconstruct the three-dimensional temperature distribution within the furnace; and the energy efficiency fingerprint analysis strategy for the steam-water system includes: establishing a database of energy efficiency characteristic parameters for the steam-water system under different operating conditions; calculating in real time the matching degree between the current energy efficiency characteristic parameters of the steam-water system and the optimal operating condition characteristic parameters in the database; and generating energy efficiency matching results.

9. The anomaly analysis system based on multi-source node correlation in thermal power generation according to claim 5, characterized in that, The environmental emission dynamic threshold optimization strategy includes: acquiring real-time electricity price data, pollutant treatment cost data, and excessive penalty standard data; calculating pollutant emission thresholds by combining them with emission rights trading-related models; and dynamically adjusting the emission thresholds according to the grid load period. The grid unit coupled response prediction strategy includes: constructing a coupled model of grid frequency, turbine speed regulation system parameters, and boiler combustion rate; generating load change prediction information in advance based on real-time grid load command data using a preset prediction algorithm; and optimizing the response parameters of the unit's automatic power generation control.

10. A method for constructing a multi-source node correlation model for thermal power generation, configured in the anomaly analysis system for multi-source node correlation of thermal power generation as described in claims 1-9, characterized in that, Includes the following steps: S110 performs spatial dimension grid division, time dimension adaptive sampling, and feature dimension multi-parameter fusion on multi-source heterogeneous data of the entire thermal power generation process. At the same time, it integrates equipment state deviation factors into the granular data foundation to construct a refined data set. S120 takes a refined dataset as input and uses forward reasoning, reverse tracing, and time delay compensation mechanisms to characterize the causal relationships and dynamic interaction characteristics between parameters. Among them, forward reasoning embeds process mechanism equations, reverse tracing establishes multi-class associated path backtracking algorithms, and time delay compensation performs lag correction for the system's inertial characteristics. S130 receives node operation data from the dynamic cause-effect graph model, calculates entropy values ​​of multiple states in real time, and dynamically adjusts parameter weights according to operating conditions. When the node entropy value exceeds the set threshold, it triggers local model reconstruction. S140 executes the equipment state deviation factor preprocessing strategy, the combustion field three-dimensional temperature field reconstruction strategy, the steam-water system energy efficiency fingerprint analysis strategy, the environmental emission dynamic threshold optimization strategy, the power grid unit coupling response prediction strategy, the multi-source data spatiotemporal registration strategy, and the fault mode transfer learning strategy, and feeds the results of each strategy back to the dynamic causal graph model and the entropy weight dynamic correction algorithm. S150, based on the dynamic cause-effect graph model and the dynamic correction results of entropy weight, realizes closed-loop management of hierarchical early warning system, fault tracing and control strategy self-optimization; S160 periodically verifies the accuracy of the dynamic causal graph model and the entropy weight dynamic correction algorithm through preset evaluation indicators, and trains and updates the model by combining thermal power generation operation data during low-load periods.

Images