An Optimization Method for Water Level Control in Thermal Power Boilers Based on Particle Swarm Optimization

Abstract

This invention belongs to the field of thermal power boiler technology and discloses an optimization method for water level control in thermal power boilers based on particle swarm optimization (PSO). By collecting equipment health parameters such as the thickness of scale buildup on the water-cooled wall and the pressure difference between the inlet and outlet of the feedwater pump, the method calculates the characteristics of heat exchange efficiency attenuation coefficient and head attenuation coefficient. An equipment health correction term is introduced into the dynamic hybrid model, and an equipment attenuation compensation mode is designed to solve the problem of unquantified equipment attenuation in traditional control. Compared with the shortcomings of traditional models where deviations gradually increase due to equipment aging, this method can correlate equipment health status with the water level model in real time, distinguishing between disturbance effects and equipment attenuation effects, and ensuring that the deviation between the model's calculated value and the actual water level is stably controlled within ±3mm. Cross-system parameters such as grid frequency and turbine valve opening are collected, and a disturbance propagation path matrix is ​​constructed through "cross-correlation analysis + transfer entropy." Then, a GNN is used to predict the magnitude of the implicit disturbance chain impact within the next 30 seconds and trigger predictive feedforward control. This design improves the water level compliance rate under extreme operating conditions, reduces the risk of maloperation of water level protection during isolated grid operation, and ensures safe and stable operation of the unit under extreme scenarios such as deep peak shaving and grid failures.

Description

Technical Field

[0001] This invention belongs to the field of thermal power boiler technology, specifically a method for optimizing water level control in thermal power boilers based on particle swarm optimization algorithm. Background Technology

[0002] In thermal power generation, the boiler, as the core energy conversion equipment, relies heavily on its drum water level control as a crucial link in the industrial control system to ensure the safe and stable operation of the unit. Precise control of the drum water level parameter not only directly relates to equipment safety but also affects the overall operating efficiency of the industrial control system. Excessively high water levels can lead to water carryover in the steam, threatening turbine safety and disrupting the closed-loop steam-water quality control system. Conversely, excessively low water levels can cause dry burning of the water-cooled walls, leading to equipment failure and disrupting the coordinated balance between power generation efficiency and grid power quality in the industrial control system. Currently, traditional boiler water level control in the industry largely relies on classic industrial control algorithms such as PID, constructing a single-loop regulation system based on local parameters. However, under complex operating conditions, this type of industrial control solution still faces the following technical bottlenecks:

[0003] Long-term operation of boilers can lead to equipment aging problems such as scaling on water-cooled walls and wear on feedwater pump impellers. Traditional control systems do not link the health status of the equipment with the dynamic water level model, and only design control logic based on the parameters of the new equipment. As the operating time goes on, the deviation between the model calculation value and the actual water level gradually widens, and it is impossible to distinguish between "disturbance effects" and "equipment attenuation effects", resulting in inaccurate adjustment of control parameters.

[0004] Boiler water level is significantly affected by the implicit transmission of disturbances across the "power grid-steam turbine-boiler" system. Such disturbances can indirectly cause sudden deviations in water level through multiple links. However, traditional control only monitors local boiler parameters and does not identify the transmission path of implicit disturbances. This not only results in delayed control actions but also makes it easy to misjudge implicitly transmitted disturbances as local random disturbances, exacerbating water level fluctuations.

[0005] Under extreme conditions such as deep peak shaving and isolated grid operation, boilers may experience special problems such as weakened water circulation at low load and a sudden drop in steam flow leading to amplified false water levels. Traditional control systems do not have dedicated logic designed for extreme conditions. For example, during deep peak shaving, traditional PID control can easily cause frequent fluctuations and accelerated wear of feedwater regulating valves. During isolated grid operation, traditional feedforward compensation cannot cope with false water level problems and can easily trigger water level protection shutdown. Summary of the Invention

[0006] The purpose of this invention is to provide an optimization method for water level control in thermal power boilers based on particle swarm optimization, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an optimization method for water level control in thermal power boilers based on particle swarm optimization, the specific steps of which are as follows:

[0008] Data Acquisition: Collect core parameters, device health parameters, and cross-system related parameters, and use a data priority caching mechanism to ensure that high-priority data is not lost;

[0009] Data preprocessing: First, high-frequency noise is removed from the collected data threshold; then, the denoised data is dynamically smoothed to eliminate random errors; finally, cross-system latent disturbance features and equipment attenuation features are extracted.

[0010] Hybrid model construction: A mechanistic model is set up, based on the boiler thermal balance equation and hydrodynamic characteristics, and equipment health correction terms and cross-system coupling terms are introduced; a data-driven model is set up, using preprocessed data as input, to compensate for the output error of the mechanistic model;

[0011] Disturbance prediction and ranking: In the disturbance prediction stage, the impact magnitude of the hidden disturbance chain is predicted; in the priority ranking stage, a judgment matrix is ​​constructed, which includes 4 evaluation indicators.

[0012] Dynamic optimization of PID control parameters: The boiler operating conditions are divided into 5 intervals, and 3 sets of PID reference parameters are trained in advance; in the PSO particle swarm optimization stage, the PID reference parameters are adjusted with 4 objective functions.

[0013] Control strategy and execution: Predictive feedforward control initiates adjustments before disturbances occur, adaptive PID control achieves water level regulation, emergency fault-tolerant control can switch modes and alarm when sensor failure occurs, and extreme working condition emergency mode is triggered when deep peak shaving and isolated network operation are detected, and equipment attenuation compensation mode automatically adjusts water supply flow commands.

[0014] Effect evaluation and model optimization: Collect indicators, including water level deviation and overshoot, set evaluation thresholds, and judge whether the control effect meets the standard by comparing the indicators. If it does not meet the standard, update the dynamic mixing model parameters and PID parameter optimization rules.

[0015] Preferably, the specific steps of the data acquisition are as follows:

[0016] A distributed data acquisition system based on the Industrial Internet of Things (IIoT) is set up. The data collected by the acquisition system includes core parameters, equipment health parameters, and cross-system correlation parameters. The core parameters include boiler drum water level, feedwater flow rate, etc.; the equipment health parameters include water-cooled wall scaling thickness, etc.; and the cross-system correlation parameters include power grid frequency, etc.

[0017] The core parameter is 10Hz, and the device health parameter and the cross-system correlation parameter are 1Hz; at the same time, the data priority caching mechanism of the edge computing node is adopted to ensure that high-priority data is not lost, thus solving the problem of missing data between the device and the cross-system.

[0018] Preferably, the specific steps of the data preprocessing are as follows:

[0019] For electromagnetic interference in industrial settings, wavelet transform is used to denoise the collected data threshold, removing high-frequency noise; then Kalman filtering is used to dynamically smooth the denoised data, eliminating random errors and improving data stability.

[0020] By calculating the correlation coefficient through cross-correlation analysis and the transfer entropy algorithm, a disturbance propagation path matrix is ​​established, thereby extracting cross-system latent disturbance characteristics; the attenuation coefficient is calculated based on equipment health parameters, thereby extracting equipment attenuation characteristics and using them as independent feature vectors.

[0021] Formula for the attenuation coefficient of heat exchange efficiency of water-cooled walls:

[0022]

[0023] In the formula: The value represents the attenuation coefficient of the water-cooled wall heat exchange efficiency, which is a core characteristic of the equipment attenuation. The smaller the value, the lower the water-cooled wall heat exchange efficiency and the more severe the attenuation. When it exceeds the threshold, the control strategy and the equipment attenuation compensation mode will be triggered.

[0024] This represents the actual heat flux density of the water-cooled wall, expressed in W / m². The thickness of scale on the water-cooled wall, collected by the ultrasonic thickness gauge in the data acquisition process, is calculated by combining the heat conduction formula, reflecting the actual heat transfer capacity of the water-cooled wall at present.

[0025] This indicates the rated heat flux density of the water-cooled wall, in W / , is the design rated parameter of the equipment at the time of manufacture, representing the ideal heat transfer capacity of the water-cooled wall without attenuation;

[0026] This indicates the actual scale thickness of the water-cooled wall, in mm. It is directly collected by the ultrasonic thickness gauge in the data acquisition system, with a range of 0-5 mm. It is the core physical parameter reflecting the attenuation of the water-cooled wall.

[0027] This indicates the rated allowable scale thickness of the water-cooled wall, in mm. It is a standard threshold for equipment maintenance. For example, if it is usually set to 1 mm, scale thickness below this value is considered to have no significant degradation.

[0028] This indicates the maximum allowable scale thickness of the water-cooled wall, in mm. The upper limit of the range of the ultrasonic thickness gauge (5 mm) in the data acquisition is the upper limit of the scale thickness for safe operation of the equipment. If this value is exceeded, the machine must be stopped for cleaning.

[0029] Preferably, the specific steps for constructing the hybrid model are as follows:

[0030] A dynamic hybrid model of boiler water level is established by integrating the mechanism model and the data-driven model. The mechanism model is based on the boiler heat balance equation and hydrodynamic characteristics, combined with the equipment health status and cross-system disturbance effects, and introduces equipment health correction terms and cross-system coupling terms to ensure that it conforms to the physical nature of the boiler, while reflecting the impact of equipment degradation and cross-system disturbances on water level.

[0031] Boiler water level dynamic mixing model output formula:

[0032]

[0033] In the formula: This represents the actual water level in the boiler drum at time t, in mm. It is the core controlled variable of the control system and must be maintained within this range to ensure boiler safety.

[0034] This indicates the initial reference value of the boiler water level, in mm. It is the initial water level value set when the equipment is started, such as the usual setting of 500mm.

[0035] This represents the water level-to-flow conversion factor, in mm. h / t, a mechanism coefficient derived from physical parameters such as boiler water volume and steam drum pressure, reflects the degree of influence of the difference between feedwater flow rate and steam flow rate on water level changes.

[0036] This represents a time variable, with the unit being hours (h), corresponding to the real-time time series of the data acquisition, and is the time dimension for integral calculation.

[0037] express The water flow rate at all times, in t / h, is collected by the electromagnetic flow meter in the data acquisition system, with a range of 0-1000 t / h;

[0038] express The steam flow rate at any time, in t / h, is collected by the vortex flow meter in the data acquisition system, with a range of 0-800 t / h.

[0039] This represents the equipment health correction term at time t, in mm, derived from the water-cooled wall heat exchange efficiency attenuation coefficient. The equipment health parameters, such as the pump head attenuation coefficient, are calculated to compensate for the impact of equipment attenuation on the water level.

[0040] The term represents the cross-system coupling term at time t, in mm. It is calculated from cross-system parameters such as grid frequency deviation and turbine valve opening deviation, and is used to reflect the impact of cross-system latent disturbances on water level.

[0041] The data-driven error compensation term at time t is expressed in mm and is output by the LSTM-attention mechanism network. It is used to correct the nonlinear error of the mechanism model and improve the accuracy of water level calculation. 30% of the weight is allocated to the equipment attenuation feature and the latent disturbance feature.

[0042] The data-driven model uses an LSTM-attention mechanism network, taking pre-processed multi-dimensional data (water level, water flow, etc.) as input to compensate for the output error of the mechanism model; through the attention mechanism, 30% of the weight is allocated to the equipment attenuation features and latent disturbance features, and the influence of the two types of features on the water level is learned first.

[0043] Preferably, the specific steps for disturbance prediction and ranking are as follows:

[0044] In the disturbance prediction stage, parameters such as grid frequency and turbine valve opening are used as GNN nodes, and the correlation coefficient between these parameters is used as edge weights. Through the GNN node message passing mechanism, the impact amplitude of the hidden disturbance chain within the next 30 seconds is predicted, and cross-system hidden disturbances are captured in advance.

[0045] GNN formula for predicting the magnitude of the impact of latent perturbations across systems:

[0046]

[0047] In the formula: Indicates the future The magnitude of the impact of the latent disturbance on the boiler water level at any given time is measured in mm. If the absolute value is large, such as exceeding 5 mm, the control strategy and the predictive feedforward control will be triggered to adjust the feedwater flow in advance to offset the disturbance.

[0048] This represents the prediction time window, measured in seconds. The plan sets it to 30 seconds to ensure sufficient time for feedforward adjustments to be initiated and to prevent water level overshoot.

[0049] This indicates the number of disturbance source nodes, corresponding to the number of cross-system correlation parameters and equipment health parameters in the data acquisition, such as power grid frequency, turbine valve opening, feedwater pump pressure difference, etc., with n=3 in common.

[0050] This indicates the number of disturbed parameter nodes, corresponding to the number of core boiler parameters in the data acquisition, such as steam pressure, steam drum water level, etc., with m=2 in common.

[0051] Indicates the source node of the disturbance To the disturbed parameter node The propagation weights are learned through the "node message passing" mechanism of GNN, reflecting the strength of the propagation of the disturbance from the source node to the disturbed node.

[0052] Indicates the source node of the disturbance With the disturbed parameter node The correlation coefficient is calculated by the "cross-correlation analysis + transfer entropy" algorithm of the data preprocessing. When the value is greater than 0.6, it is considered a strongly correlated perturbation;

[0053] Indicates the source node of the disturbance at time t. The actual deviation value is consistent with the corresponding disturbance source parameter, such as the power grid frequency deviation in Hz and the turbine valve opening deviation in %, which is calculated from the real-time parameters and rated values ​​of the data acquisition.

[0054] In the priority ranking process, a judgment matrix is ​​constructed based on FAHP. The evaluation indicators include the degree of impact of disturbance on water level, probability of occurrence, duration, and robustness impact weight. Disturbances under extreme conditions (such as a sudden drop in steam flow during isolated network operation) are assigned a weight coefficient of 1.5 times to ensure that disturbances under extreme conditions are dealt with first, thus solving the problem of low priority for disturbances under extreme conditions in traditional ranking.

[0055] Preferably, the specific steps for the dynamic optimization of the adaptive PID control parameters are as follows:

[0056] In terms of load range division, the boiler operating conditions are divided into five ranges: deep peak shaving, low load, medium load, high load and isolated grid operation. Three sets of PID benchmark parameters are pre-trained for each range to provide an initial basis for parameter optimization.

[0057] In the PSO particle swarm optimization stage, the objective functions are to minimize water level control deviation, minimize adjustment time, optimize control robustness, and adapt equipment attenuation compensation. The PID baseline parameters are adjusted in real time through the PSO particle swarm algorithm. For example, the derivative coefficient is reduced during deep peak shaving to avoid frequent valve fluctuations, and the integral coefficient is increased during isolated network operation to offset the influence of false water levels and adapt to different operating conditions.

[0058] PSO (Physical Optimization of PID) fitness function formula:

[0059]

[0060] In the formula: This represents the fitness function value of the Particle Swarm Optimization (PSO) algorithm, where the optimization objective is to minimize... This will output the optimal PID parameters, including the proportional gain. Integral coefficient Differential coefficients ;

[0061] , , Denotes the weight coefficients of the fitness function, and satisfies Standard experience is set as =0.5, =0.3, =0.2 is used to achieve three objectives: minimizing the equilibrium water level deviation, minimizing the adjustment time, and optimizing the control robustness.

[0062] This indicates the control assessment time window, measured in seconds. The plan sets it to 60 seconds, which covers a complete water level disturbance response cycle and ensures comprehensive assessment results.

[0063] The actual water level in the boiler drum at time t is expressed in mm. It is acquired by the differential pressure transmitter in the data acquisition system and used to compare with the target water level to calculate the deviation.

[0064] This indicates the target water level in the boiler drum, in mm. It is a reference value for the water level set according to the boiler's operating process, such as 500 mm, which is the target value for water level control.

[0065] This indicates the water level adjustment time, measured in seconds (s), which refers to the time it takes for the water level to recover from a deviation from the target value to within ±3 mm. This is required in the effect evaluation and model optimization. ≤10s;

[0066] This indicates the robustness index of the control, determined by the water level overshoot. The calculation yields the formula R = 1 - / 100, the overshoot value is required in the effect evaluation and model optimization. ≤5%, therefore ≥0.95, The larger the value, the stronger the control robustness.

[0067] Preferably, the control strategy and the specific steps of its execution are as follows:

[0068] Predictive feedforward control calculates the water flow adjustment amount in advance based on the disturbance prediction results and inputs it into the water supply regulating valve, and starts the adjustment before the disturbance occurs to offset the impact; adaptive PID control uses optimized PID parameters for closed-loop control during the disturbance stability period or in low-priority disturbance scenarios to achieve precise water level regulation; emergency fault-tolerant control estimates the fault parameters based on a dynamic hybrid model when a sensor fails, switches to model estimation + redundant sensor mode and alarms.

[0069] In the extreme operating condition emergency mode, when deep peak shaving is detected, the feedwater pump frequency converter is linked to reduce the feedwater flow regulation accuracy and reduce the PID differential gain; when islanded operation is detected, the turbine and boiler coordinated control is started to reduce the valve opening change rate; in the equipment attenuation compensation mode, when the attenuation coefficient exceeds the threshold, the feedwater flow command is automatically adjusted to offset the equipment attenuation effect.

[0070] Preferably, the specific steps for effect evaluation and model optimization are as follows:

[0071] In the control effect evaluation stage, indicators such as water level deviation and overshoot are collected in real time, and evaluation thresholds are set, including water level deviation ±3mm and overshoot ≤5%. The control effect is judged by comparing the indicators.

[0072] In the model iteration and optimization stage, if the control effect is not up to standard, such as the water level deviation exceeding the threshold three times in a row, the DQN-transfer learning algorithm is used to update the dynamic hybrid model parameters and PID parameter optimization rules. The model adaptability is optimized with the control effect compliance rate as the reward function. The model is retrained monthly with the latest equipment health data. When new extreme working condition data is added, transfer learning is used to quickly adapt to new scenarios to ensure that the model adapts to the aging trend of equipment in the long term.

[0073] The beneficial effects of this invention are as follows:

[0074] 1. This invention collects equipment health parameters such as the thickness of scale on the water-cooled wall and the pressure difference between the inlet and outlet of the feed water pump, calculates the characteristics of the heat exchange efficiency decay coefficient and the head decay coefficient, introduces equipment health correction terms into the dynamic hybrid model, and designs equipment decay compensation modes to solve the problem of unquantified equipment decay in traditional control. Compared with the shortcomings of traditional models where deviations gradually increase due to equipment aging, this invention can correlate equipment health status with the water level model in real time, distinguish between the effects of disturbances and the effects of equipment decay, ensure that the water level control accuracy does not decrease with equipment aging during long-term boiler operation, extend equipment service life and reduce maintenance costs.

[0075] 2. This invention collects cross-system parameters such as grid frequency and turbine valve opening, constructs a disturbance propagation path matrix through "cross-correlation analysis + transfer entropy", and then uses GNN to predict the impact amplitude of implicit disturbance chains within the next 30 seconds, triggering predictive feedforward control. This process can capture the implicit disturbance propagation of "grid-turbine-boiler" in advance, avoiding the action lag and disturbance misjudgment problems of traditional control, offsetting the sudden deviation of water level caused by cross-system disturbances, reducing the frequency and amplitude of water level fluctuations, and enabling the water level to quickly recover to the target range under cross-system disturbances, improving the unit's operational stability, and reducing process fluctuations and energy waste caused by water level fluctuations.

[0076] 3. This invention designs a dedicated emergency mode: during deep peak shaving, the feedwater pump frequency converter is linked and the PID differential gain is reduced to avoid frequent valve fluctuations and wear caused by traditional PID; during islanded grid operation, "turbine-boiler coordinated control" is activated to reduce the rate of change of the regulating valve opening to suppress false water level amplification; at the same time, the PSO algorithm can optimize PID parameters for different extreme operating conditions, solving the problems of insufficient robustness and coordination of traditional control; this design improves the water level compliance rate under extreme operating conditions, reduces the risk of water level protection malfunction during islanded grid operation, and ensures that the unit can still operate safely and stably under extreme scenarios such as deep peak shaving and grid failure. Attached Figure Description

[0077] Figure 1 This is a flowchart of the method for optimizing water level control in a thermal power boiler according to the present invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] like Figure 1 As shown in the figure, this invention provides an optimization method for water level control in thermal power boilers based on particle swarm optimization. The specific steps of this method are as follows:

[0080] Data acquisition: Collect three types of data: core parameters, device health parameters, and cross-system related parameters, and adopt a data priority caching mechanism for edge computing nodes;

[0081] Data preprocessing: Wavelet transform is used to threshold and denoise the collected data, and Kalman filtering is used to dynamically smooth the denoised data; cross-system latent disturbance features and equipment attenuation features are extracted;

[0082] Hybrid model construction: The mechanistic model is based on the boiler heat balance equation and hydrodynamic characteristics, and introduces equipment health correction terms and cross-system coupling terms; the data-driven model adopts an LSTM-attention mechanism network to compensate for the output error of the mechanistic model;

[0083] Perturbation prediction and ranking: GNN uses relevant parameters as nodes and correlation coefficients as edges to predict the impact magnitude of hidden perturbation chains; a judgment matrix containing four evaluation indicators is constructed based on FAHP, and perturbations under extreme conditions are assigned a weight coefficient of 1.5 times.

[0084] Adaptive PID control parameter dynamic optimization: The boiler operating conditions are divided into 5 intervals, and 3 sets of PID benchmark parameters are pre-trained for each interval; in the PSO algorithm optimization stage, the PID parameters are adjusted in real time with 4 objective functions.

[0085] Control strategy and execution: predictive feedforward control predicts results, adaptive PID control uses optimized PID parameters for closed-loop control, emergency fault-tolerant control estimates fault parameters, and in the event of deep peak shaving or triggering extreme operating conditions during isolated network operation, the equipment attenuation compensation mode automatically adjusts the water supply flow command.

[0086] Performance evaluation and model optimization: Real-time data collection of indicators, setting of evaluation thresholds, judgment of control effectiveness, and updating of dynamic hybrid model parameters and PID parameter optimization rules if the targets are not met.

[0087] Example: Implementation of Optimized Boiler Water Level Control under Medium Load and Conventional Operating Conditions

[0088] I. Data Collection Stage

[0089] Establish a distributed data acquisition system based on the Industrial Internet of Things (IIoT), with three data acquisition nodes set up according to the "boiler body - turbine side - power grid interface":

[0090] Key parameters: Rosemount 3051CD differential pressure transmitter is used to acquire steam drum water level (range 0-1000mm, accuracy ±0.5mm, acquisition frequency 10Hz), LDG-M9 electromagnetic flowmeter is used to acquire feed water flow (range 0-1000t / h, frequency 10Hz), and vortex flowmeter (model: LUGB-200) is used to acquire steam flow (range 0-800t / h, frequency 10Hz).

[0091] Equipment health parameters: The thickness of scale on the water-cooled wall was measured using a DM5E ultrasonic thickness gauge (range 0-5mm, frequency 1Hz), and the pressure difference between the inlet and outlet of the feed water pump was measured using a pressure transmitter (model: EJA530A) (range 0-2.5MPa, frequency 1Hz).

[0092] Cross-system correlation parameters: The power grid frequency is collected using a BD-4000 frequency transmitter (range 48-52Hz, frequency 1Hz), and the turbine valve opening is collected using an angular displacement sensor (model: WDD35D4) (range 0-100%, frequency 1Hz).

[0093] Data caching is implemented using the EC600S edge gateway, with priority set according to "core parameters > device health parameters > cross-system parameters". Core parameters are written to DRAM, and the rest are written to Flash to avoid loss of high-priority data.

[0094] II. Data Preprocessing Stage

[0095] Denoising and smoothing: The collected water level and flow rate data are decomposed into three levels using the db4 wavelet basis, and high-frequency coefficients are denoised using a soft threshold (threshold = 1.2 times the noise standard deviation); then, the Kalman filter formula is applied.

[0096] Equations of state: (W represents process noise, with a variance of 0.01);

[0097] Observation equation: (V represents observation noise, with a variance of 0.008), eliminating random errors;

[0098] Feature extraction: Using a 60s sliding window, the correlation coefficient between power grid frequency and water level was calculated to be r=0.72 through cross-correlation analysis, and the disturbance propagation entropy value was calculated to be 0.35 using the propagation entropy algorithm, constructing a 3×3 disturbance propagation path matrix; based on the water-cooled wall scaling thickness δ=0.8mm (rated... =1mm, maximum =5mm), according to formula α w = ×(1− Calculate the attenuation coefficient α w =0.92, which is used as the device attenuation feature vector.

[0099] III. Hybrid Model Construction Phase = 0.6 × 0.92 × 500 + 0.4 × 0.95 × 500 = 463 mm

[0100] Mechanism model: Based on the boiler heat balance equation, initial water level =500mm, water level-flow conversion factor ;0.7×(50-50)×2+0.3×(50-50)×1.5=0mm

[0101] Equipment health correction items =0.6×0.92×500+0.4×0.95×500=463mm (Water pump head attenuation coefficient) =0.95);

[0102] Cross-system coupling terms =0.7×(50-50)×2+0.3×(50-50)×1.5=0mm (no deviation in power grid frequency and valve opening); Data-driven model: LSTM-Attention Network is used, input layer dimension = 8 (8 types of parameters such as water level and water supply flow), 2 hidden layers (64 units per layer), dropout = 0.2; Attention mechanism according to formula:

[0103] ( , It is a 64×32 matrix. =32), 30% weight allocated to Combined with the perturbation propagation entropy value, the error of the compensation mechanism model is calculated, and the final output of the hybrid model is:

[0104] (error compensation item) =-2.1mm).

[0105] IV. Disturbance Prediction and Ranking

[0106] Disturbance prediction: The GNN nodes are set as grid frequency, valve opening, feedwater pump differential pressure, steam pressure, and steam drum water level (5 nodes), and the edge weights are the correlation coefficients between nodes (e.g., frequency and water level weights are 0.72). A GCN network (32-dimensional hidden layer, ReLU activation, 20 iterations) is used to predict the magnitude of the implicit disturbance within the next 30 seconds. (If the thickness is less than 5mm, feedforward will not be triggered);

[0107] Priority ranking: A judgment matrix is ​​constructed based on FAHP (indicators: impact level, probability of occurrence, duration, robustness, with scales of 5, 3, 2, and 4 respectively), and the calculated weights are 0.4, 0.25, 0.2, and 0.15. Since there are no extreme operating conditions, there is no need to amplify the weights; the disturbance priority ranking is "feed pump differential pressure fluctuation > power grid frequency deviation > valve opening fluctuation".

[0108] V. Dynamic Optimization of PID Control Parameters

[0109] Operating condition determination: Current load 60% (medium load range: 4096-70% of rated load), call pre-trained PID baseline parameters. ,

[0110] PSO optimization: 30 particles, 50 iterations, inertia weight 0.9 Learning factors C1=C2=2;

[0111] objective function The optimized PID parameters are as follows: , .

[0112] VI. Control Strategies and Implementation Stages

[0113] Adaptive PID control: because (Low-priority disturbance) Optimized PID parameter closed-loop control is adopted with a control cycle of 0.5s to adjust the opening of the water supply regulating valve in real time;

[0114] No emergency trigger: Sensors are fault-free, non-extreme operating conditions, fault-tolerant and emergency modes are not activated; equipment attenuation coefficient (If the threshold of 0.8 is not exceeded), attenuation compensation is not triggered.

[0115] VII. Effectiveness Evaluation

[0116] Data collected over 1 hour: maximum water level deviation ±1.8mm (≤±3mm), overshoot 3.296 (≤596), adjustment time The control effect meets the requirements; the model does not require iterative optimization.

[0117] The specific steps for data collection are as follows:

[0118] A distributed data acquisition system based on the Industrial Internet of Things (IIoT) was built to construct a full-dimensional data matrix and collect three types of core parameters. These core parameters include boiler drum water level (accuracy ±0.5mm) and feedwater flow rate (range 0-1000t / h); equipment health parameters include water-cooled wall scale thickness (range 0-5mm); and cross-system related parameters include power grid frequency (range 48-52Hz).

[0119] Core parameter acquisition uses a boiler drum water level differential pressure transmitter (model: Rosemount3051CD) and a feedwater flow electromagnetic flowmeter (model: LDG-M9, range 0-1000t / h); equipment health parameter acquisition uses an ultrasonic thickness gauge for water-cooled wall scale thickness (model: DM5E, range 0-5mm); and cross-system correlation parameter acquisition uses a power grid frequency transmitter (model: BD-4000, range 48-52Hz). Edge computing nodes utilize EC600S edge gateways, with a caching strategy prioritizing core parameters (10Hz) written to DRAM, and equipment health and cross-system parameters (1Hz) written to Flash, ensuring no loss of high-priority data.

[0120] The distributed data acquisition system adopts the Modbus-TCP communication protocol. Acquisition nodes are deployed in zones according to "boiler body - turbine side - grid interface", with a total of 8 acquisition nodes (3 for the boiler body, 3 for the turbine side, and 2 for the grid interface). Core parameters are collected once every 0.1 seconds, while equipment health parameters and cross-system related parameters are collected once every 1 second. Edge nodes prioritize the collected data according to "core parameters > equipment health parameters > cross-system parameters". When low-priority data is full, non-critical data older than 72 hours is overwritten first.

[0121] Differentiated acquisition frequencies are set, with the core parameter at 10Hz and device health parameters and cross-system correlation parameters at 1Hz; at the same time, a data priority caching mechanism for edge computing nodes is adopted to ensure that high-priority data is not lost, thus solving the problem of missing data between devices and cross systems.

[0122] The specific steps of data preprocessing are as follows:

[0123] To address electromagnetic interference in industrial settings, wavelet transform is performed using a db4 wavelet basis for three-level decomposition. Soft thresholding (with a threshold set to 1.2 times the noise standard deviation) is applied to high-frequency coefficients to remove high-frequency noise. The Kalman filter state equation is set as X(k) = 1.02X(k-1) + 0.05U(k) + W(k) (where X is the data state variable, U is the control variable, W is the process noise, and the variance is 0.01), and the observation equation is set as Z(k) = 0.98*X(k) + V(k) (where Z is the observed value, V is the observation noise, and the variance is 0.008). The denoised data is dynamically smoothed to eliminate random errors.

[0124] Two types of feature extraction were performed based on a 60-second sliding time window: cross-correlation analysis and the transfer entropy algorithm were used to calculate the correlation coefficient and construct the disturbance propagation path matrix to extract cross-system latent disturbance features; the attenuation coefficient was calculated based on equipment health parameters to extract equipment attenuation features and use them as independent feature vectors.

[0125] The specific steps for constructing the hybrid model are as follows:

[0126] A hybrid model of boiler water level dynamics is constructed by integrating mechanistic model and data-driven hybrid model. The mechanistic model is based on the boiler heat balance equation and hydrodynamic characteristics, combined with the equipment health status and cross-system disturbance effects, and introduces equipment health correction terms and cross-system coupling terms to ensure that it conforms to the physical nature of the boiler, while reflecting the impact of equipment degradation and cross-system disturbances on water level.

[0127] The data-driven model uses an LSTM-attention mechanism network, taking pre-processed multi-dimensional data (water level, water flow, etc.) as input to compensate for the output error of the mechanism model; through the attention mechanism, 30% of the weight is allocated to the equipment attenuation features and latent perturbation features, and the influence of the two types of features on the water level is learned first.

[0128] The specific steps for disturbance prediction and ranking are as follows:

[0129] In the disturbance prediction stage, parameters such as grid frequency and turbine valve opening are used as GNN nodes, and the correlation coefficient between parameters is used as edge weight. Through the GNN node message passing mechanism, the impact magnitude of the hidden disturbance chain in the next 30 seconds is predicted, and cross-system hidden disturbances are captured in advance.

[0130] In the priority ranking process, a judgment matrix is ​​constructed based on FAHP. The evaluation indicators include the degree of impact of disturbance on water level, probability of occurrence, duration, and robustness impact weight. Disturbances under extreme conditions (such as a sudden drop in steam flow during isolated network operation) are assigned a weight coefficient of 1.5 times to ensure that disturbances under extreme conditions are dealt with first, thus solving the problem of low priority for disturbances under extreme conditions in traditional ranking.

[0131] The specific steps for dynamic optimization of PID control parameters are as follows:

[0132] A PID parameter self-optimization mechanism is constructed. In terms of load range division, the boiler operating conditions are divided into five ranges: deep peak shaving (10%-20% of rated load), low load, medium load, high load, and isolated grid operation. Three sets of PID benchmark parameters are pre-trained for each range to provide an initial basis for parameter optimization.

[0133] In the PSO algorithm optimization phase, the objective functions are to minimize water level control deviation, minimize adjustment time, optimize control robustness, and adapt to equipment attenuation compensation. The PSO algorithm (30 particles, 50 iterations) is used to adjust PID parameters in real time. For example, during deep peak shaving, the derivative coefficient is reduced to avoid frequent valve fluctuations, and the integral coefficient is increased to offset the influence of false water levels during isolated network operation, thus adapting to different operating conditions.

[0134] The specific steps of the control strategy and its execution are as follows:

[0135] Predictive feedforward control calculates the water flow adjustment amount in advance based on the disturbance prediction results and inputs it into the water supply regulating valve, and starts the adjustment before the disturbance occurs to offset the impact; adaptive PID control uses optimized PID parameters for closed-loop control during the disturbance stability period or in low-priority disturbance scenarios to achieve precise water level regulation; emergency fault-tolerant control estimates the fault parameters based on a dynamic hybrid model when a sensor fails, switches to model estimation + redundant sensor mode and alarms.

[0136] In the extreme operating condition emergency mode, when deep peak shaving (load <20%) is detected, the feedwater pump frequency converter is linked to reduce the feedwater flow regulation accuracy (±0.5t / h to ±1t / h) and reduce the PID differential gain; when islanded operation (grid frequency fluctuation >±0.5Hz) is detected, the turbine and boiler coordinated control is started to reduce the valve opening change rate (5% / s to 2% / s) to suppress false water levels; in the equipment attenuation compensation mode, when the attenuation coefficient exceeds the threshold, the feedwater flow command is automatically adjusted to offset the effect of equipment attenuation.

[0137] The specific steps for effect evaluation and model optimization are as follows:

[0138] In the control effect evaluation stage, indicators such as water level deviation and overshoot are collected in real time, and evaluation thresholds are set, including water level deviation ±3mm and overshoot ≤5%. The control effect is judged by comparing the indicators.

[0139] In the model iteration and optimization phase, if the control effect is not up to standard, such as the water level deviation exceeding the threshold three times in a row, the DQN-transfer learning algorithm is used to update the dynamic hybrid model parameters and PID parameter optimization rules. The model's adaptability is optimized using the control effect compliance rate as the reward function. The model is retrained monthly with the latest equipment health data. When new extreme working condition data is added, transfer learning is used to quickly adapt to new scenarios, ensuring that the model adapts to the aging trend of equipment in the long term.

[0140] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An optimization method for water level control in thermal power boilers based on particle swarm optimization, characterized in that: The specific steps of this method are as follows: Data Acquisition: Collect core parameters, device health parameters, and cross-system related parameters, and use a data priority caching mechanism to ensure that high-priority data is not lost; Data preprocessing: First, wavelet transform is used to denoise the collected data to remove high-frequency noise. Then, the denoised data is dynamically smoothed to eliminate random errors. Finally, cross-system latent disturbance features and equipment attenuation features are extracted. Hybrid model construction: Set up a mechanistic model based on the boiler thermal balance equation and hydrodynamic characteristics, and introduce equipment health correction terms and cross-system coupling terms; Set up a data-driven model, using preprocessed data as input, to compensate for the output error of the mechanistic model; Disturbance prediction and ranking: In the disturbance prediction stage, the impact magnitude of the hidden disturbance chain is predicted; in the priority ranking stage, a judgment matrix is ​​constructed, which includes 4 evaluation indicators. Dynamic optimization of PID control parameters: The boiler operating conditions are divided into 5 intervals, and 3 sets of PID benchmark parameters are pre-trained for each interval; in the PSO algorithm optimization stage, the PID benchmark parameters are adjusted with 4 objective functions; Control strategy and execution: Predictive feedforward control initiates adjustments before disturbances occur, adaptive PID control achieves water level regulation, emergency fault-tolerant control can switch modes and alarm when sensor failure occurs, and extreme working condition emergency mode is triggered when deep peak shaving and isolated network operation are detected, and equipment attenuation compensation mode automatically adjusts water supply flow commands. Effect evaluation and model optimization: Collect indicators, including water level deviation and overshoot, set evaluation thresholds, and judge whether the control effect meets the standard by comparing the indicators. If it does not meet the standard, update the dynamic mixing model parameters and PID parameter optimization rules.

2. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The specific steps for data collection are as follows: Set up a distributed data acquisition system based on the Industrial Internet of Things. The data collected by the acquisition system includes core parameters, equipment health parameters, and cross-system correlation parameters. The core parameters are collected at a frequency of 10Hz, and the device health parameters and cross-system related parameters are collected at a frequency of 1Hz; at the same time, the data priority caching mechanism of the edge computing node is adopted to ensure that high-priority data is not lost.

3. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The specific steps of the data preprocessing are as follows: For electromagnetic interference in industrial settings, wavelet transform is used to denoise the collected data to remove high-frequency noise; then Kalman filtering is used to dynamically smooth the denoised data to eliminate random errors. By calculating the correlation coefficient through cross-correlation analysis and the transfer entropy algorithm, a disturbance propagation path matrix is ​​established, thereby extracting cross-system latent disturbance characteristics. The attenuation coefficient is calculated based on the equipment health parameters, thereby extracting the equipment attenuation features and using them as independent feature vectors.

4. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The specific steps for constructing the hybrid model are as follows: A dynamic hybrid model of boiler water level is established by integrating the mechanism model and the data-driven model. The mechanism model is based on the boiler heat balance equation and hydrodynamic characteristics, and combines the equipment health status and cross-system disturbance effects to introduce equipment health correction terms and cross-system coupling terms. The data-driven model uses an LSTM-attention mechanism network, taking preprocessed multi-dimensional data as input to compensate for the output error of the mechanism model; through the attention mechanism, 30% of the weight is allocated to the device attenuation features and latent perturbation features.

5. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The specific steps for perturbation prediction and ranking are as follows: In the disturbance prediction stage, the GNN node includes parameters such as grid frequency and turbine valve opening. The correlation coefficient between these parameters serves as the edge weight. Through the GNN node message passing mechanism, the impact magnitude of the implicit disturbance chain within the next 30 seconds is predicted. In the priority ranking process, a judgment matrix is ​​constructed based on FAHP, and the evaluation indicators include the degree of impact of disturbance on water level, probability of occurrence, duration and robustness impact weight; disturbances under extreme conditions are assigned a weight coefficient of 1.5 times.

6. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The specific steps for dynamically optimizing the PID control parameters are as follows: In terms of load range division, the boiler operating conditions are divided into five ranges: deep peak shaving, low load, medium load, high load and isolated grid operation. Three sets of PID benchmark parameters are pre-trained for each range. In the PSO algorithm optimization stage, the objective functions are to minimize water level control deviation, minimize adjustment time, optimize control robustness, and adapt equipment attenuation compensation. The PID reference parameters are adjusted in real time through the PSO algorithm.

7. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The control strategy and its specific execution steps are as follows: Predictive feedforward control calculates the water flow adjustment amount input to the water supply regulating valve in advance based on the disturbance prediction results; adaptive PID control uses optimized PID parameters for closed-loop control during the disturbance stability period or in low-priority disturbance scenarios; emergency fault-tolerant control estimates the fault parameters based on a dynamic hybrid model when a sensor fails, switches to model estimation + redundant sensor mode and alarms. In the extreme operating condition emergency mode, when deep peak shaving is detected, the feedwater pump frequency converter is linked to reduce the feedwater flow regulation accuracy and reduce the PID differential gain; when islanded operation is detected, the turbine and boiler coordinated control is started to reduce the valve opening change rate; in the equipment attenuation compensation mode, the feedwater flow command is automatically adjusted when the attenuation coefficient exceeds the threshold.

8. The method for optimizing water level control in thermal power boilers based on particle swarm optimization as described in claim 1, characterized in that: The specific steps for effect evaluation and model optimization are as follows: In the control effect evaluation stage, indicators including water level deviation and overshoot are collected, evaluation thresholds are set, and the control effect is judged by comparing the indicators. In the model iteration and optimization stage, if the control effect does not meet the standard, the DQN-transfer learning algorithm is used to update the dynamic hybrid model parameters and PID parameter optimization rules, and the model adaptation capability is optimized by using the control effect compliance rate as the reward function; when adding extreme working condition data, the new scenario is quickly adapted through transfer learning.

Images