Multi-formula self-adaptive electric precipitation intelligent variable-frequency power supply regulation and control method

Abstract

The invention discloses an electric precipitation intelligent variable-frequency power supply regulation and control method with a multi-formula self-adaptive function, and the method comprises the steps: obtaining a fire coal type parameter, a unit load parameter, a flue gas humidity parameter and a dust concentration parameter in real time, and constructing a dynamic feature vector containing six types of working condition features; performing cosine similarity matching on the basis of the dynamic feature vector and a working condition label in a preset formula library, and selecting an optimal formula above a similarity threshold as an initial regulation and control parameter; performing reinforcement learning dual-objective optimization according to the initial regulation and control parameters, and dynamically balancing dust emission to reach the standard and minimizing energy consumption through small-step parameter adjustment and real-time effect evaluation; and periodically updating the formula library, converting the optimization parameter combination with the operation effect score reaching a preset threshold value into a new formula for storage, and eliminating a low-efficiency formula. According to the invention, the self-adaptive regulation and control of the electric precipitation power supply under the multi-dimensional working condition can be realized, the emission stability and the energy-saving effect are obviously improved, the manual intervention requirement is reduced, and the system response speed and the operation and maintenance efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent variable frequency power supply control systems for electrostatic precipitators, and more particularly to a method and apparatus for controlling intelligent variable frequency power supply for electrostatic precipitators with multi-formula adaptive function. Background Technology

[0002] In the field of intelligent variable frequency power supply control systems for electrostatic precipitators, some related technologies have been developed and applied. However, these existing technologies still have many shortcomings and urgently need improvement. From the perspective of power circuit structure, some existing technologies aim to increase output voltage. For example, some patents use a combination of primary and secondary boost circuits to obtain higher high-voltage DC current to improve dust removal efficiency. However, this method of simply relying on circuit-level voltage boosting makes the power system structure complex, increasing equipment cost and maintenance difficulty. Furthermore, it is difficult to flexibly and accurately control output parameters when facing frequent changes in operating conditions, failing to meet the diverse needs for power output characteristics under different operating conditions. Regarding the level of intelligence in power supply control, existing technologies are significantly insufficient. Although some high-voltage variable frequency power supplies for electrostatic precipitators attempt to use deep learning-based artificial intelligence technology to adjust the power frequency based on data collected by dust concentration monitoring sensors, in practical applications, their algorithms often focus only on a single dust concentration dimension, ignoring the comprehensive impact of other key operating factors such as coal type, unit load, and flue gas humidity on the electrostatic precipitator effect and power supply control requirements. This results in a lack of comprehensiveness and precision in power supply regulation, making it difficult to adapt to complex and ever-changing actual operating environments and failing to fully utilize the optimal performance of electrostatic precipitators. Regarding the handling of back corona, some existing technologies use control chips to monitor current and adjust circuit phase to avoid it. However, this control method has a slow response speed; when sudden changes in operating conditions cause rapid and abnormal current changes, it cannot adjust effectively in a timely manner, leading to recurring back corona and severely impacting dust removal efficiency. Furthermore, this control logic is relatively simple and does not consider the complex mechanisms of back corona generation under different operating conditions, failing to fundamentally solve the problem and only providing some degree of mitigation. Existing technologies also lack multi-recipe management and self-optimization adaptation functions. Most electrostatic precipitator power supply regulation systems can only store a small number of fixed operating parameter combinations, unable to flexibly recall or automatically generate suitable parameter recipes based on different operating conditions. When faced with diverse operating conditions in actual production, such as high-ash coal, low-load operation, and high-humidity flue gas, it is impossible to quickly switch to the optimal operating mode. It can only rely on manual adjustment of power parameters, which is not only inefficient, but also difficult to guarantee the adjustment accuracy. It is easy to have unreasonable parameter settings, which leads to problems such as unstable dust removal effect and increased energy consumption. Summary of the Invention

[0003] The main objective of this invention is to provide a method for controlling an intelligent variable frequency power supply for electrostatic precipitators with multi-formula adaptive function. This method enables adaptive control of the electrostatic precipitator power supply under multi-dimensional operating conditions, significantly improving emission stability and energy-saving effect, reducing the need for manual intervention, and improving system response speed and operation and maintenance efficiency.

[0004] Another objective of this invention is to propose an intelligent variable frequency power supply control device for electrostatic precipitators with multi-formula adaptive function.

[0005] To achieve the above objectives, a first aspect of the present invention proposes a method for controlling an intelligent variable frequency power supply for an electrostatic precipitator with multi-formula adaptive function, comprising: S1 acquires real-time parameters of coal type, unit load, flue gas humidity, and dust concentration, and constructs a dynamic feature vector containing six types of operating conditions. S2, based on the dynamic feature vector and the working condition labels in the preset formula library, perform cosine similarity matching, and select the optimal formula above the similarity threshold as the initial control parameter; S3, perform reinforcement learning bi-objective optimization based on the initial control parameters, and dynamically balance dust emission compliance and energy consumption minimization through small step parameter adjustment and real-time effect evaluation; S4 periodically updates the recipe library, converting optimized parameter combinations that have reached a preset threshold in performance evaluation into new recipes for storage, and eliminating inefficient recipes.

[0006] In one embodiment of the present invention, the real-time acquisition of coal type parameters, unit load parameters, flue gas humidity parameters, and dust concentration parameters, and the construction of a dynamic feature vector containing six types of operating condition characteristics, further includes: S11, Inlet dust concentration data are collected by a laser scattering dust concentration meter, the range of which is 0-1000mg / m³ and the accuracy is ±5mg / m³; S12, X-ray fluorescence analyzer is used to obtain coal ash parameters. The sampling frequency of the ash parameters is 1Hz and the data update cycle is 5 minutes.

[0007] In one embodiment of the present invention, the cosine similarity matching based on the dynamic feature vector and the working condition labels in the preset recipe library further includes: S21, Calculate the cosine similarity between the real-time operating condition feature vector and the recipe label vector, using the following formula:

[0008] in Let i be the i-th parameter of the real-time operating condition feature vector. The i-th parameter of the recipe label vector; S22, set the similarity threshold to 85%. When the matching similarity is lower than this threshold, use an interpolation algorithm to generate a temporary recipe parameter combination.

[0009] In one embodiment of the present invention, the step of performing reinforcement learning bi-objective optimization based on the initial control parameters further includes: S31, update the Q-table using the Q-Learning algorithm, the formula is:

[0010] in For learning rate, Discount factor; S32, when performing the action, limits the secondary voltage adjustment step size to ±0.5kV, the secondary current adjustment step size to ±0.05A, and the pulse frequency adjustment step size to ±2Hz.

[0011] In one embodiment of the present invention, the periodically updated formula library further includes: S41, the formula is scored based on three indicators: compliance rate, energy consumption level, and stability. The compliance rate has a weight of 40%, the energy consumption level has a weight of 30%, and the stability has a weight of 30%. S42, perform elimination operation on formulations with a score below 60, and perform solidification operation on temporary formulations with a score of 90 or above and a running time of more than 30 days.

[0012] In one embodiment of the present invention, it further includes: S5, execute the back corona warning and rectifier fault diagnosis steps: S51, when the voltage fluctuation is ≥15%, the current surge is ≥20%, and the electric field impedance drops by ≥30% for 30 seconds, it is determined to be a precursor to back corona and an emergency callback action is triggered. S52 monitors the secondary current ripple coefficient. When the ripple coefficient is >0.5 and lasts for 1 minute, it will warn of rectifier IGBT damage and locate the faulty module.

[0013] To achieve the above objectives, a second aspect of the present invention provides an intelligent variable frequency power supply control device for electrostatic precipitators with multi-formula adaptive function, comprising: The dynamic feature vector construction module is used to acquire coal type parameters, unit load parameters, flue gas humidity parameters, and dust concentration parameters in real time, and construct dynamic feature vectors containing six types of operating condition features. The formula matching and selection module is used to perform cosine similarity matching between the dynamic feature vector and the working condition labels in the preset formula library, and select the optimal formula above the similarity threshold as the initial control parameter. The reinforcement learning optimization module is used to perform reinforcement learning bi-objective optimization based on the initial control parameters. Through small step parameter adjustment and real-time effect evaluation, it dynamically balances dust emission compliance and energy consumption minimization. The recipe library update module is used to periodically update the recipe library, converting optimized parameter combinations that have reached a preset threshold in performance evaluation into new recipes for storage, and eliminating inefficient recipes.

[0014] The electrostatic precipitator intelligent frequency conversion power supply control method and device with multi-formula adaptive function in the embodiments of the present invention can realize the adaptive control of the electrostatic precipitator power supply under multi-dimensional operating conditions, significantly improve emission stability and energy saving effect, and reduce manual intervention and operation and maintenance costs. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for controlling an intelligent variable frequency power supply for an electrostatic precipitator with multi-formula adaptive function, provided in an embodiment of the present invention; Figure 2 This is an architecture diagram of an intelligent variable frequency power supply control system for electrostatic precipitators with multi-formula adaptive function, provided in an embodiment of the present invention. Figure 3 A detailed logic diagram of a multi-formula adaptive electrostatic precipitator intelligent variable frequency power supply control algorithm provided in an embodiment of the present invention; Figure 4 This is a closed-loop control logic diagram provided in an embodiment of the present invention; Figure 5 This is a structural diagram of the first type of intelligent variable frequency power supply control device for electrostatic precipitators with multi-formula adaptive function provided in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0018] The following description, with reference to the accompanying drawings, describes a method and apparatus for regulating an intelligent variable frequency power supply for electrostatic precipitators with multi-formula adaptive function, according to an embodiment of the present invention.

[0019] Example 1 This embodiment provides a method for controlling an intelligent variable frequency power supply for electrostatic precipitators with multi-formula adaptive function. For example... Figure 1 As shown, the method includes the following steps: S1 acquires real-time parameters of coal type, unit load, flue gas humidity, and dust concentration, and constructs a dynamic feature vector containing six types of operating conditions.

[0020] Specifically, this step achieves comprehensive perception and feature extraction of the operating environment of the electrostatic precipitator system through the collaborative work of a multi-source sensor network and an edge computing controller, providing a data foundation for subsequent formula matching and parameter optimization.

[0021] This process relies on multi-dimensional sensors deployed at the coal-fired boiler outlet, electrostatic precipitator inlet, and power module. Coal type parameters (such as ash and sulfur content) are acquired in real time using an X-ray fluorescence analyzer and an infrared sulfur analyzer, with accuracies of ±1% and ±0.05%, respectively. Unit load parameters are obtained through the DCS system, with an accuracy of ±0.5%. Flue gas humidity parameters are measured by a capacitive humidity sensor, with a range of 0–40% RH and an accuracy of ±3%. Dust concentration parameters are acquired by a laser scattering dust concentration meter, with a range of 0–1000 mg / m³ and an accuracy of ±5 mg / m³. All parameters are synchronously sampled using a 24-bit ADC chip at a sampling frequency of 1 kHz to ensure data real-time performance and consistency.

[0022] Furthermore, the system integrates the above four types of parameters with the operating status of the electrostatic precipitator (such as secondary voltage and secondary current) to construct a six-dimensional operating condition feature vector. Where C_{out} is the outlet dust concentration. It is a secondary voltage. It is a secondary current. The pulse frequency, This represents the real-time power of the power supply. This is the temperature of the power module. This vector serves as the state input for the reinforcement learning algorithm, used to determine the current operating condition in real time and perform similarity matching with preset operating condition labels in the recipe library (e.g., using a cosine similarity algorithm). The matching threshold is set to... This is to ensure the accuracy and applicability of the formula application.

[0023] This step is widely used in the real-time control of electrostatic precipitators (ESPs) in thermal power plants. Especially under complex operating conditions such as unit load fluctuations, coal type switching, and changes in flue gas humidity, it can quickly construct feature vectors and match the optimal operating formula, thereby achieving a dynamic balance between dust removal efficiency and energy consumption. For example, in a scenario where the unit load drops from 800MW to 500MW, the system can identify the change in operating conditions within 5 minutes and activate the low-load formula to ensure that the outlet dust concentration remains stable below 25 mg / m³, while simultaneously reducing power consumption.

[0024] By constructing high-precision, multi-dimensional dynamic feature vectors, reliable data support is provided for subsequent reinforcement learning optimization and adaptive formula switching, significantly improving the system's response speed and control accuracy. This is an important technical foundation for achieving the goal of "less manned operation and intelligent maintenance".

[0025] Furthermore, S1 includes: S11, the inlet dust concentration data is collected by a laser scattering dust concentration meter, the range of which is 0-1000mg / m³ and the accuracy is ±5mg / m³.

[0026] Specifically, the technical implementation of this step is based on the principle of optical scattering. By measuring the scattering intensity of dust particles on the laser beam, the dust mass concentration per unit volume is derived, providing real-time and accurate environmental input data for the subsequent multi-formula self-optimization algorithm.

[0027] The laser scattering dust concentration meter uses a measurement system consisting of a laser emitter and a photodetector. Its working principle is as follows: when a laser beam passes through dust-laden gas, dust particles scatter the light. The intensity of the scattered light is positively correlated with the number and size distribution of the particles. By acquiring the scattered light signal through a high-sensitivity photodetector and combining it with a preset calibration curve of scattering intensity versus mass concentration, the system can calculate the inlet dust concentration in real time. Its range is The accuracy is This meets the high-precision monitoring requirements for dust concentration in flue gas treatment of thermal power plants.

[0028] The dust concentration meter's sampling frequency is synchronized with the system's main control unit to ensure real-time data acquisition and processing. Its measurement accuracy is within... Maintain within range It meets the accuracy requirements for dust concentration measurement equipment in GB / T 16157-1996 "Methods for Determination of Particulate Matter and Sampling of Gaseous Pollutants in Exhaust Gas from Stationary Sources". Furthermore, the equipment has an IP65 protection rating, making it suitable for harsh working environments with high temperature, high humidity, and high dust concentrations, ensuring long-term stable operation.

[0029] The dust concentration meter is deployed in the inlet flue of the electrostatic precipitator, working in conjunction with sensors such as ultrasonic flow meters and infrared humidity meters to construct a multi-dimensional operating condition sensing network. The data it collects serves as one of the key parameters in the "operating condition label," used to match the optimal operating strategy from the recipe library in real time. For example, under high-ash coal combustion conditions, an increase in inlet dust concentration will trigger the system to call the "high-ash-high-load" recipe, and combine it with reinforcement learning algorithms to fine-tune the parameters to suppress back corona and maintain the outlet concentration at the standard.

[0030] This step plays a fundamental role in the entire technical solution and is a prerequisite for realizing the "multi-formula self-optimization" control logic. Through high-precision and high-response dust concentration data acquisition, the system can accurately identify the characteristics of the current operating condition, providing a reliable basis for subsequent formula matching, parameter optimization, and fault early warning, thereby significantly improving the operating efficiency and environmental performance of the electrostatic precipitator system.

[0031] S12, X-ray fluorescence analyzer is used to obtain coal ash parameters. The sampling frequency of the ash parameters is 1Hz and the data update cycle is 5 minutes.

[0032] Specifically, its technical principle is based on X-ray fluorescence spectroscopy (XRF) technology. By exciting elements in a sample and detecting their characteristic X-ray radiation, it achieves non-contact, high-precision online detection of ash content in coal. The X-ray fluorescence analyzer irradiates the coal sample with a high-energy X-ray source (such as an Rh target X-ray tube), exciting the elemental atoms in the sample and causing them to emit characteristic X-rays. These characteristic X-ray signals are acquired by a high-resolution detector (such as a silicon drift detector, SDD), and combined with an elemental database for spectral analysis, ultimately outputting the ash content (usually expressed as a mass percentage, e.g., ...). or This serves as a key basis for the system to determine the characteristics of coal types.

[0033] In this step, the X-ray fluorescence analyzer's sampling frequency is set to 1Hz, meaning it collects ash data once per second to ensure the system can quickly respond to changes in coal type. The data update cycle is 5 minutes, meaning the system integrates and updates the ash data every 5 minutes to trigger the formula matching and parameter optimization process. This setting ensures real-time performance while also considering data stability and system processing efficiency, meeting the real-time requirements for data acquisition and processing in industrial settings (such as the response time requirements for industrial control systems in the IEC 61131-3 standard).

[0034] This step is typically deployed in the coal conveying system or the coal quality testing stage before combustion, working in conjunction with sensors such as infrared sulfur analyzers and laser dust analyzers to construct a multi-dimensional operating condition sensing network. For example, when the ash content of coal increases from 20% to 35%, the X-ray fluorescence analyzer can promptly report the change in ash content. Based on this, the system identifies it as "high ash coal" and triggers the invocation of a "high ash-high load" formula. At the same time, a reinforcement learning algorithm is activated to dynamically optimize parameters such as voltage and current to suppress back corona and maintain emission compliance.

[0035] The technical value of this step lies in providing key coal type input parameters for the "multi-formula self-optimization" system. This enables the system to classify operating conditions and match formulations based on ash content, thereby achieving precise control of the electrostatic precipitator power supply. Through real-time acquisition and periodic updates of ash parameters, the system can effectively identify coal quality change trends, improve the accuracy and response speed of formulation matching, and provide reliable data support for subsequent intelligent decision-making and parameter optimization.

[0036] S2, based on the dynamic feature vector and the working condition labels in the preset formula library, perform cosine similarity matching, and select the optimal formula above the similarity threshold as the initial control parameter.

[0037] Specifically, the system first collects six core parameters in real time through a multi-dimensional sensor network, including coal ash content, sulfur content, unit load, flue gas humidity, and inlet dust concentration, to construct a dynamic feature vector. ,in The dust concentration at the outlet (mg / m³) is the concentration of particulate matter. It is the secondary voltage (kV). It is the secondary current (A). The pulse frequency (Hz) This represents the real-time power of the power supply (kW). The value is the power module temperature (°C). The cosine similarity between this vector and the operating condition label vector of each recipe in the recipe library is calculated using the following formula:

[0038] in For recipe tag vectors, The angle between the feature vector and the formula vector. This represents the similarity value, ranging from [0, 1]. The system sets the similarity threshold to [value]. That is, when At that time, it was considered that the current working conditions were highly compatible with the formula and it could be used as a candidate formula.

[0039] Furthermore, among all recipes with a matching degree higher than the threshold, the system selects the recipe with the highest similarity as the initial control parameter. If no recipe has a matching degree higher than the threshold, the system generates temporary control parameters based on the 2-3 most similar recipe parameters using an interpolation algorithm, ensuring that the system can maintain stable operation even under sudden changes in operating conditions. This step has a significant advantage in terms of system response time; under normal operating conditions, the matching and recall time is less than 1 second, and under complex operating conditions, the response time is controlled within 5 minutes, meeting the high requirements of thermal power plants for real-time performance and stability.

[0040] In application scenarios, this step is widely used in typical operating conditions such as unit load changes, coal type switching, and flue gas humidity fluctuations. For example, when the unit load drops from 800MW to 500MW, the system quickly identifies the "low load - low sulfur" condition through cosine similarity matching and calls the corresponding formula to reduce the secondary voltage from 68kV to 60kV, avoiding excessive power supply and energy waste under low load. Through intelligent matching of structured formula tags and real-time operating condition vectors, the system's identification accuracy and response efficiency for complex operating conditions are significantly improved, providing high-quality initial parameters for subsequent parameter self-optimization algorithms, thereby achieving the core goals of "precise control, energy saving and consumption reduction, and reduced manpower."

[0041] Furthermore, S2 includes: S21, Calculate the cosine similarity between the real-time operating condition feature vector and the recipe label vector, using the following formula:

[0042] in Let i be the i-th parameter of the real-time operating condition feature vector. is the i-th parameter of the recipe label vector.

[0043] Specifically, this step first constructs two vectors: a real-time operating condition feature vector. With recipe tag vector The real-time operating condition feature vector consists of operating parameters collected by multi-dimensional sensors, including coal ash content, sulfur content, unit load, flue gas humidity, and inlet dust concentration, typically a 6-dimensional vector. The recipe label vector consists of applicable operating condition parameters corresponding to each recipe in the recipe library, and has the same dimensions. In some implementations, each dimension of the vector is normalized to eliminate dimensional differences and ensure the accuracy of similarity calculation. Cosine similarity measures the directional consistency of two vectors by calculating the cosine of the angle between them, using the formula:

[0044] in, Represents the vector dot product. This represents the Euclidean norm of the vector. This formula is used to quantify the degree of matching between real-time operating conditions and the applicable scenarios of the formulation.

[0045] The calculation of cosine similarity requires meeting a certain matching threshold. The system sets the similarity threshold to be [value missing]. Specifically, when the similarity between a real-time operating condition and a certain recipe label reaches or exceeds a certain value, the system determines that the recipe is the appropriate solution for the current operating condition. If no recipe meets the threshold, the system uses an interpolation algorithm to fuse the parameters of the 2-3 most similar recipes to generate a temporary control solution. In addition, the vector normalization process uses the min-max normalization method to map the parameters of each dimension to the [0, 1] interval to improve computational stability and matching accuracy.

[0046] This step is widely used in the operation and control of electrostatic precipitators (ESPs) in thermal power plants, especially under complex conditions such as sudden changes in unit load, coal type switching, and fluctuations in flue gas humidity. For example, when the unit load drops from 800MW to 500MW, the system collects current operating parameters, calculates their similarity to "low load" recipes in the recipe library, quickly identifies and calls the optimal combination of operating parameters, ensuring emissions meet standards while reducing energy consumption. In some implementations, this step can be completed in milliseconds on an ARM Cortex-A53 processor, supporting a data processing capability of 1000 points per second.

[0047] This step transforms the system from "human experience-based judgment" to "data-driven intelligent matching," significantly improving the accuracy and response speed of operating condition identification. In actual operation, this method can reduce recipe retrieval time to less than 1 second and increase matching accuracy to [missing information]. This effectively reduces emissions exceeding standards and energy waste caused by delayed operating condition identification. Furthermore, this step provides reliable initial parameters for subsequent parameter self-optimization algorithms, serving as a fundamental step in realizing closed-loop control of the "multi-formula self-optimization" control system.

[0048] S22, set the similarity threshold to 85%. When the matching similarity is lower than this threshold, use an interpolation algorithm to generate a temporary recipe parameter combination.

[0049] Specifically, in some implementations, when the matching similarity is lower than a set threshold of 85%, the system will use an interpolation algorithm to generate a temporary formula parameter combination to address situations where the current operating condition does not perfectly match the preset formulas in the formula library. The interpolation algorithm performs linear or nonlinear interpolation operations based on the currently collected operating condition feature vector and the 2-3 sets of formula parameters with the highest similarity in the formula library. Specifically, the system first uses a cosine similarity algorithm to match the real-time collected operating condition parameters (such as coal type and ash content, unit load, flue gas humidity, etc.) with the operating condition labels of each formula in the formula library to calculate the similarity value. If the similarity of the highest matching formula is lower than 85%, the system enters the interpolation process, selects the top 3 formulas with the highest similarity as the interpolation source, and generates a set of temporary parameter combinations using a weighted average or spline interpolation method based on their parameter distribution characteristics. This combination will serve as the initial control parameters under the current operating condition for further optimization by subsequent reinforcement learning algorithms.

[0050] The interpolation algorithm's input parameters include six core operating condition parameters: coal ash content (X-ray fluorescence analyzer), unit load (DCS signal), and flue gas humidity (capacitive hygrometer). Its similarity calculation is based on normalized feature vectors. The parameter combinations generated by interpolation must satisfy equipment safety boundary conditions, such as secondary voltage. Secondary current pulse frequency The output accuracy requirement of the interpolation algorithm is: voltage accuracy. Current accuracy Frequency accuracy This is to ensure that the generated parameters have sufficient finesse for adjustment.

[0051] This step is mainly used to address parameter transition issues when operating conditions change abruptly or when a new operating condition occurs for the first time. For example, when the unit load drops from 800MW to 500MW, and there is no perfectly matching "medium load - low sulfur" formula in the formula library, the system will interpolate based on two sets of formula parameters, "high load - low sulfur" and "low load - medium sulfur," to generate a temporary voltage-current-frequency combination. This ensures that the system can maintain emission compliance and stable operation until the new formula is fully optimized.

[0052] The technical effect of this step is that it significantly improves the system's response capability and operational stability under non-preset operating conditions. Through interpolation algorithms, the system can quickly generate reasonable parameters even without matching formulations, avoiding control failures or excessive emissions due to missing parameters. Simultaneously, this temporary parameter combination provides favorable initial conditions for subsequent reinforcement learning optimization, helping to shorten the optimization convergence time and improve overall control efficiency. Therefore, this step plays a dual role in this invention: "formula library expansion" and "transitional control assurance," serving as a crucial supporting link for realizing the "adaptive and self-optimizing" control logic.

[0053] S3. Based on the initial control parameters, perform reinforcement learning bi-objective optimization, and dynamically balance dust emission compliance and energy consumption minimization through small step parameter adjustment and real-time effect evaluation.

[0054] Specifically, this step, based on the reinforcement learning (Q-Learning) algorithm, constructs an adaptive optimization mechanism with "emission compliance" and "lowest energy consumption" as dual objectives, thereby achieving efficient and energy-saving operation of the electrostatic precipitator system under complex operating conditions. This step first loads initial parameters, such as secondary voltage, from the formula library to match the current operating conditions. Secondary current Pulse frequency This serves as the initial state for reinforcement learning. The system uses an edge computing controller to collect data in real time, including the outlet dust concentration. Power supply electric field temperature State parameters, including those mentioned above, are used to construct a state vector. Subsequently, the algorithm employs a frequency of once per second. - Greedy strategy ( From the action space Select the adjustment action, where , , These represent the adjustment step sizes for voltage, current, and frequency, respectively, with step size ranges of... , , This is to ensure the precision of parameter adjustments and system stability.

[0055] Reward function of reinforcement learning It consists of three parts: environmental protection reward R env Energy consumption bonus R energy and stability penalty R stable Their weights are respectively , , Environmental protection rewards will prioritize ensuring export dust concentrations. If the target is exceeded, a significant negative reward will be given (e.g.) Energy rewards are obtained by comparing power changes before and after an action. This is used to quantify energy-saving effects; stability penalty suppresses drastic fluctuations in parameters and avoids system oscillations.

[0056] This procedure is widely applicable to the regulation of electrostatic precipitators (ESPs) in thermal power plants under different loads, coal types, and flue gas humidity conditions. For example, during low-load operation, the system gradually reduces voltage and current to minimize energy consumption while ensuring emissions meet standards. When a change in coal type causes a sudden change in inlet dust concentration, the system uses reinforcement learning to quickly adjust the pulse frequency and charging ratio to suppress back corona discharge and maintain dust removal efficiency.

[0057] This step employs a closed-loop mechanism of "small-step trial—effect evaluation—strategy adjustment" to enable continuous energy consumption optimization while meeting stringent environmental constraints. During the iteration process, if energy consumption changes after 10 consecutive adjustments... The optimization process terminates when the maximum number of iterations reaches 50, and the current optimal parameters are output. This method significantly improves the system's response speed and operational economy, transforming the electrostatic precipitator power supply control from "passive response" to "active optimization," and providing solid technical support for achieving "less manned, intelligent operation and maintenance" environmental protection systems in thermal power plants.

[0058] Furthermore, S3 includes: S31, update the Q-table using the Q-Learning algorithm, the formula is:

[0059] in For learning rate, Discount factor; Specifically, the Q-table stores the expected reward value for each state-action pair, and its dimension is the number of states multiplied by the number of actions. The state space consists of six key parameters, including the outlet dust concentration C. out Secondary voltage Secondary current Pulse frequency Power supply and module temperature These parameters are acquired in real time through a sensor network and discretized to reduce computational complexity. The action space defines the adjustment step sizes for voltage, current, and frequency, for example... , , After each action is performed, the system collects the new state. And calculate the reward value based on the difference before and after the state transition. .

[0060] The update of the Q table follows the classic formula of Q-Learning:

[0061] in, The learning rate controls the update magnitude; This is a discount factor used to balance the importance of current rewards and future rewards. This formula ensures that at each step, the system can dynamically adjust the value estimate in the Q-table based on the current reward and the expected optimal future action, thereby gradually approaching the optimal control strategy.

[0062] In practical applications, the Q-table update process is executed cyclically with a 1-second cycle, using... - Greedy strategy ( The system selects actions with a 90% probability of choosing the action with the highest value in the current Q-table and a 10% probability of randomly exploring new actions. In this way, the system achieves a balance between stable operation and strategy exploration, avoiding getting trapped in local optima.

[0063] This step plays a crucial role in the entire technical solution, serving as the core algorithm supporting the "multi-recipe self-optimization" function. Through continuous updates to the Q-table, the system can dynamically adjust power parameters based on real-time operating conditions, thereby minimizing energy consumption while meeting environmental emission standards. Experimental data shows that this algorithm can reduce energy consumption by more than 10% under low-load conditions, while controlling emission concentration fluctuations within... Within this timeframe, the system's operating efficiency and stability have been significantly improved.

[0064] S32, when performing the action, limits the secondary voltage adjustment step size to ±0.5kV, the secondary current adjustment step size to ±0.05A, and the pulse frequency adjustment step size to ±2Hz.

[0065] Specifically, this step size constraint is a key component of the action space in reinforcement learning algorithms. Within the Q-Learning framework, the system discretizes the power supply parameters using a finite set of actions to reduce algorithm complexity and improve convergence efficiency. Specifically, the voltage adjustment step size is ±0.5kV, ensuring fine control of the output electric field strength without affecting electric field stability; the current adjustment step size is ±0.05A, corresponding to the fine-tuning of the corona current in the electrostatic precipitator, helping to maintain constant current output characteristics and preventing uneven dust accumulation on the electrode plates or back corona phenomena caused by excessive current fluctuations; the pulse frequency adjustment step size is ±2Hz, used to adjust the pulse characteristics of corona discharge, thereby optimizing dust charging efficiency and electric field stability.

[0066] The step size settings strictly adhere to the operating boundary conditions and control accuracy requirements of the electrostatic precipitator power supply. For example, the secondary voltage adjustment step size is ±0.5kV, corresponding to the adjustable range of the secondary voltage in the electrostatic precipitator (50-72kV), ensuring that each adjustment does not exceed the minimum allowable control unit of the system; the secondary current step size is ±0.05A, conforming to the sampling accuracy of the Rogowski coil sensor (±0.01A), ensuring the system's sensitivity to current changes; and the pulse frequency step size is ±2Hz, matching the frequency adjustment capability of the FPGA coprocessor in the system, supporting parameter iterative optimization at the 1-second level.

[0067] This step size limit is widely applicable to various operating condition switching scenarios in thermal power plants, such as load changes, coal type switching, and flue gas humidity fluctuations. During low-load operation, the system can gradually reduce voltage and current to avoid over-supply. When operating with high-ash coal, small-step frequency adjustments improve the stability of corona discharge and suppress back corona phenomena. Furthermore, this step size setting supports a hybrid control strategy of "recipe calling + parameter fine-tuning," ensuring that even after matching the optimal recipe, small-step optimization can still further approach the energy consumption minimization target.

[0068] This step size limit improves system response speed while effectively avoiding drastic oscillations during parameter adjustment, thus enhancing system robustness. With a voltage adjustment of ±0.5kV, the system can complete parameter switching from high to low load within 10 minutes; a current adjustment of ±0.05A helps maintain current fluctuations within 0.1A / minute, ensuring stable electric field operation; and a frequency adjustment of ±2Hz improves charging efficiency while preventing corona current instability caused by frequency abrupt changes. In summary, this step size setting is a crucial foundation for achieving the dual objectives of "emission compliance + minimum energy consumption," providing reliable technical support for the intelligent control of electrostatic precipitator power supplies.

[0069] S4 periodically updates the recipe library, converting optimized parameter combinations that have reached a preset threshold in performance evaluation into new recipes for storage, and eliminating inefficient recipes.

[0070] Specifically, this step involves statistical analysis of daily operational data to score the effectiveness of each formula in the formula library. The scoring mechanism comprehensively considers three dimensions: emission compliance rate, energy consumption level, and system stability. The emission compliance rate is determined using a formula... (Duration of achieving the target / Total running time) A quantitative assessment is conducted, with a weighting of 40%; energy consumption level is based on the ratio to the historical best energy consumption under the same operating conditions, with a weighting of 30%; stability indicators are determined through voltage standard deviation (e.g., ...). The system will evaluate the scores, with a weighting of 30%. Based on the scoring results, the system will assign a score... The optimized parameter combinations are converted into formal recipes and stored in the recipe library, while the scores are evaluated. Formulas that have not been used for 30 consecutive days will be discarded.

[0071] The formula score is calculated using a weighted scoring model to ensure the evaluation results are scientifically sound and reasonable. Formula library capacity requirements. The system supports structured storage and fast retrieval. Recipes are updated daily, with parameter adjustment steps controlled within a specified range. , , Within this range, the system's continuity and stability are ensured. The elimination mechanism is based on the dual criteria of "inefficiency" and "low utilization," ensuring that the formula library always retains the optimal operating strategy.

[0072] Through this step, the system achieves a closed-loop management mechanism of "recipe self-learning, self-updating, and self-elimination," significantly improving the adaptability and practicality of the recipe library. On the one hand, the system can solidify the optimization experience generated during operation into new recipes, realizing knowledge accumulation; on the other hand, by eliminating inefficient recipes, parameter redundancy and system performance degradation are avoided. This mechanism, in synergy with reinforcement learning algorithms and recipe matching strategies, enables the electrostatic precipitator power supply control system to have continuous evolution capabilities, thereby ensuring environmental compliance while minimizing energy consumption and improving operational stability. This is a crucial support for achieving the "less manned, intelligent operation and maintenance" goal of this invention.

[0073] Furthermore, S4 includes: S41 scores the formula based on three indicators: compliance rate, energy consumption level, and stability. The compliance rate has a weight of 40%, the energy consumption level has a weight of 30%, and the stability has a weight of 30%.

[0074] Specifically, the scoring mechanism employs a weighted scoring model, assigning weights of 40%, 30%, and 30% to compliance rate, energy consumption level, and stability, respectively, to reflect their priority in the operation of the electrostatic precipitator system. Compliance rate, as the primary indicator, directly relates to whether environmental emissions meet national or local standards; its calculation formula is as follows:

[0075] Among them, the compliance time refers to the outlet dust concentration The cumulative time, and the total runtime is the total system runtime. The high weight of the compliance rate (40%) ensures that the system always prioritizes emission compliance and avoids environmental non-compliance due to parameter optimization.

[0076] Energy consumption levels are determined by comparing the energy consumption per unit of dust removal during the current formulation's operation. Compared with the historical best energy consumption under the same operating conditions The ratio is evaluated, and its calculation method is as follows:

[0077] The higher the score, the better the current formulation is in terms of energy consumption control. The stability score is based on the fluctuation range of key parameters such as voltage and current, for example, the voltage standard deviation. Compared to the set threshold, the smaller the fluctuation, the higher the stability and the higher the score.

[0078] In practical applications, this scoring mechanism operates within the system's daily recipe evaluation process, quantitatively scoring all recipes in the recipe library. The scoring results are used for dynamic recipe management: Scoring Formulas that meet the criteria can be included in the official formula library and rated. Formulas that fail to meet the requirements are automatically eliminated, thus ensuring that the formula library always contains efficient, stable, and adaptable operating schemes.

[0079] This step plays a crucial role in the entire technical solution. By introducing a multi-dimensional scoring mechanism, the system can not only identify the operational effectiveness of the current formula, but also continuously optimize the formula library during long-term operation, improving the system's self-learning ability and operational economy. Its technical value lies in realizing the transformation from "experience-driven" to "data-driven," enhancing the intelligence level of the electrostatic precipitator system, and providing solid support for thermal power plants to achieve the operational goal of "less manpower operation and high energy efficiency."

[0080] S42, perform elimination operation on formulations with a score below 60, and perform solidification operation on temporary formulations with a score of 90 or above and a running time of more than 30 days.

[0081] Specifically, in the multi-formula self-optimization system of this invention, eliminating formulas with scores below 60 and solidifying temporary formulas with scores of 90 and a running time exceeding 30 days are key steps in achieving continuous system optimization and dynamic updates to the formula library. This step, based on a quantitative evaluation of the formula's performance and combined with time-based stability verification, ensures that the system always operates under optimal parameter configurations, improving overall control efficiency and system robustness.

[0082] This step involves daily statistical analysis of the operational data of all recipes in the recipe library to calculate their comprehensive score. The scoring system consists of three indicators: emission compliance rate (weight 40%), energy consumption level (weight 30%), and system stability (weight 30%). The formula for calculating the emission compliance rate is as follows: Energy consumption level is assessed by comparing it to the historical best energy consumption under the same operating conditions, while stability is evaluated based on the fluctuation range of parameters such as voltage and current (e.g., voltage standard deviation). The system quantifies the scores. For recipes with scores below 60, the system will automatically remove them from the recipe library to avoid inefficient parameters interfering with system performance. For temporary recipes with scores of 90 or higher and a runtime of more than 30 days, the system will convert them into formal recipes and include them in the recipe library as candidate solutions for subsequent operating condition matching.

[0083] The elimination threshold is set at 60 points to ensure that only formulas with good performance are retained; the solidification condition is a scoring system. And runtime The system ensures the formulation possesses sufficient stability and adaptability in actual operation. The statistical period for runtime is continuous operation time, not cumulative time, to eliminate the interference of intermittent operation on the evaluation results. A weighted average method is used for score calculation to ensure that the evaluation results comprehensively reflect the formulation's overall performance in terms of environmental protection, energy consumption, and stability.

[0084] This step applies to scenarios where the electrostatic precipitator (ESP) system in a thermal power plant requires periodic maintenance and updates to its formula library during long-term operation. For example, under "high humidity + low sulfur" conditions, if a temporary formula achieves 100% emission compliance, 15% energy consumption reduction compared to the historical best, and good stability with a score of 92 after one month of operation, the system will automatically solidify it as a formal formula. Conversely, older formulas, such as those under "high load + high sulfur" conditions, will be phased out by the system if they have high energy consumption, poor stability, and a score of only 58, thus freeing up formula library space and improving system operating efficiency.

[0085] This step, through quantitative evaluation and time-based verification mechanisms, effectively filters out inefficient formulations and retains and solidifies high-performance formulations, thereby improving the overall quality of the formulation library and the system's adaptability. In actual operation, this mechanism ensures that the system consistently uses the optimal parameter combination when facing complex operating conditions, achieving emission compliance rates. Reduced energy consumption The goal is to achieve this. Furthermore, this step also supports the system in realizing a "self-learning-self-updating" closed loop for the formula without human intervention, significantly improving the intelligence level and operational economy of the electrostatic precipitator system.

[0086] The electrostatic precipitator intelligent variable frequency power supply control method with multi-formula adaptive function in this invention can realize adaptive control of the electrostatic precipitator power supply under complex operating conditions, significantly improve emission stability and energy saving effect, and reduce manual intervention and operation and maintenance costs.

[0087] Furthermore, it also includes: S5, execute the back corona warning and rectifier fault diagnosis steps: Specifically, in the intelligent variable frequency power supply control system for electrostatic precipitators of the present invention, back corona warning and rectifier fault diagnosis are key steps to ensure stable system operation and efficient dust removal. This step, through multi-dimensional sensor data acquisition and real-time analysis, combined with a pre-set anomaly detection model and fault diagnosis knowledge base, enables early identification and response to back corona phenomena and rectifier faults, thereby improving the safety and reliability of system operation.

[0088] The anti-corona warning mechanism is based on real-time monitoring of the electric field's operating status. The system collects key parameters such as secondary voltage, current, and electric field impedance through Hall voltage sensors and Rogowski coil current sensors. When a voltage fluctuation is detected... Sudden increase in current Electric field impedance decreases And duration At a certain time, the system detects a precursor to back corona and immediately triggers a warning signal. After the warning, the system can automatically reduce the voltage, adjust the pulse frequency, or switch to an anti-back corona formula to suppress further deterioration of the corona effect. Rectifier fault diagnosis is performed by monitoring the ripple coefficient of the secondary current (normal range is...). When the ripple coefficient And duration At the specified time, the system determined that the rectifier IGBT module was malfunctioning and pushed a fault warning to the operation and maintenance terminal. At the same time, it automatically switched to the redundant module to maintain system operation.

[0089] The determination of back corona warning is based on a comprehensive judgment of three parameters: voltage fluctuation, current surge, and impedance drop, with thresholds of 15%, 20%, and 30%, respectively. The response time... Seconds. The key parameter for rectifier fault diagnosis is the current ripple coefficient, and its warning threshold is... The duration threshold is Minutes. The system supports multi-channel synchronous sampling (sampling frequency). (kHz), ensuring data real-time performance and accuracy. Furthermore, the system has the function of automatically switching redundant modules, supporting... Redundant configuration, switching time in case of single module failure Seconds ensure continuous system operation.

[0090] This step is widely used in the operation monitoring of electrostatic precipitators in thermal power plants. Especially under conditions of high coal ash content, high flue gas humidity, or frequent unit load fluctuations, the back corona warning can effectively prevent electric field breakdown and avoid a sudden drop in dust removal efficiency. Rectifier fault diagnosis plays an important role in scenarios such as power module aging and IGBT damage, enabling early detection and location of faults, reducing downtime and maintenance costs.

[0091] By implementing this step, the system can reduce the occurrence rate of back corona by 70%, and the average detection time for rectifier faults can be reduced from traditional manual inspections. hours shortened to Hourly fault warning accuracy This feature significantly enhances the system's self-diagnostic capabilities and operational stability, providing a solid guarantee for achieving the goal of "less manpower and intelligent operation and maintenance".

[0092] S51, when voltage fluctuation ≥15%, current surge ≥20%, and electric field impedance decrease ≥30% for 30 seconds, it is determined as a precursor to back corona and an emergency callback action is triggered. Specifically, this step involves real-time acquisition of key parameters such as secondary voltage, secondary current, and electric field impedance through a multi-dimensional sensor network deployed within the electrostatic precipitator system. Voltage fluctuation detection is based on the relative deviation between the current voltage value and a set reference value. A sudden increase in current is assessed by comparing the current current with the average value of the previous moment to determine if the relative change exceeds 20%. A decrease in electric field impedance is assessed by monitoring the rate of change of the electric field impedance to determine if it exceeds 30%. All three conditions must be met simultaneously, and the duration must reach a certain threshold. It takes several seconds for the system to determine if a back corona is a precursor. Once the determination is made, the system immediately triggers an "emergency callback" action, which means increasing the secondary voltage within 1 second. Secondary current boost This is to enhance the electric field strength and suppress the further development of the back corona phenomenon.

[0093] The voltage fluctuation threshold is set to 15%, the current surge threshold to 20%, the electric field impedance drop threshold to 30%, and the duration window is [not specified]. Seconds. These parameters are set based on extensive operational data and analysis of the back corona generation mechanism to ensure that a response is triggered in the early stages of anomalies, preventing deterioration of the electric field performance. Furthermore, the parameter adjustment step size for the emergency callback action (e.g., seconds) , Simulation verification shows that it can operate without exceeding the equipment's safety boundaries (such as...). , Under the premise of […], the electric field stability can be quickly restored.

[0094] This procedure is applicable to real-time monitoring and emergency response of electrostatic precipitators in thermal power plants under complex operating conditions such as high-ash coal types, low-load operation, and high-humidity flue gas. Through a multi-parameter joint judgment mechanism, this procedure significantly improves the sensitivity and accuracy of anti-corona warning, avoiding the lag and false alarm rates of traditional single-current monitoring methods.

[0095] S52 monitors the secondary current ripple coefficient. When the ripple coefficient is >0.5 and lasts for 1 minute, it will warn of rectifier IGBT damage and locate the faulty module.

[0096] Specifically, in some implementations, monitoring the secondary current ripple coefficient is a key step in the fault warning and self-diagnosis system of this invention. Its technical principle is based on real-time analysis of the rectifier output current waveform to identify potential damage risks to the IGBT module. The specific operation involves high-frequency sampling of the secondary current (sampling frequency of 1kHz) using a Rogowski coil current sensor and calculating its ripple coefficient. The ripple coefficient is defined as the ratio of the current fluctuation amplitude to the average current, and its mathematical expression is:

[0097] in, and These represent the maximum and minimum values ​​of the current within the sampling period, respectively. This is the average current value. Under normal operating conditions, this coefficient should be maintained at [value missing]. Within the specified range, this indicates that the rectifier output is stable and the IGBT module is working normally. When the ripple coefficient... And duration At the specified time, the system determined that the rectifier IGBT module was malfunctioning and triggered the early warning mechanism.

[0098] Further, this step uses the real-time signal processing module in the edge computing controller to perform sliding window analysis (window length is 1 minute) on the collected current data, calculating the average ripple coefficient within the current window. If the average exceeds a threshold... If the system detects a fault, it will immediately generate an early warning message and push the fault module number (such as "A phase group 3 IGBT") to the monitoring platform via the Modbus TCP / IP protocol to help maintenance personnel quickly locate the fault point.

[0099] This method requires the current sensor to have The bandwidth must be sufficient to ensure accurate capture of high-frequency ripple; simultaneously, the controller must possess... This reduces response delays to enable timely identification and handling of abnormal operating conditions. This step complies with the real-time and accuracy requirements for power equipment condition monitoring in the IEC 61850 standard.

[0100] In practical applications, this technology is widely used for monitoring the operation of power modules in electrostatic precipitator (ESP) systems of thermal power plants. Especially in the early stages of faults such as IGBT module aging, short circuits, or breakdowns, it can provide early warnings through abnormal changes in current ripple, preventing ESP system shutdowns or excessive emissions due to rectifier failure. This step plays a crucial role in fault prevention and system stability assurance within the overall technical solution, significantly improving equipment availability and operational efficiency.

[0101] The multi-dimensional adaptive control method for electrostatic precipitator power supply in this embodiment of the invention, by real-time monitoring of voltage fluctuations, current surges and changes in electric field impedance, combined with secondary current ripple coefficient analysis, can accurately identify precursors of back corona and rectifier IGBT damage faults, and realize rapid location and emergency callback of faulty modules, further improving the reliability and maintenance efficiency of electrostatic precipitator power supply operation.

[0102] Example 2 like Figure 2-4 As shown, in view of the shortcomings of existing electrostatic precipitator power supply control technology, this invention aims to solve the following key technical problems through the core function of "multi-formula self-optimization": 1) Existing technologies have poor adaptability to various operating conditions and cannot match complex and ever-changing operational scenarios. Existing technologies can only adjust power output based on a single parameter (such as dust concentration) or a fixed mode, ignoring the coupled effects of multiple dimensions of operating conditions, such as coal type (ash content, sulfur content), unit load, and flue gas humidity. For example, high-ash coal types are prone to back corona discharge, and low-load conditions require reduced energy consumption, but existing systems lack targeted parameter configurations, leading to fluctuations in dust removal efficiency (such as instantaneous exceedances) or energy waste. This invention aims to achieve accurate identification and adaptive parameter matching for diverse operating conditions.

[0103] 2) Lack of multi-recipe management capabilities; parameter adjustment relies on manual intervention and is inefficient: Existing technologies can only store a small number of fixed parameter combinations, and recipe switching requires manual adjustment of parameters such as voltage and current one by one, resulting in a delayed response (e.g., when operating conditions change abruptly, 24 electric field parameters require more than 30 minutes of debugging). For common sub-conditions in thermal power plants, such as "high load-high sulfur content" and "low load-high humidity," the optimal parameters cannot be quickly retrieved. This invention needs to build a scalable recipe database (≥10 sets), supporting one-click retrieval and dynamic updates, replacing manual operation.

[0104] 3) Energy consumption and emissions are difficult to optimize adaptively and in a coordinated manner, resulting in poor operational economy: Existing technologies mostly take "achieving emission standards" as the single objective, ensuring dust removal efficiency through redundant power supply (such as fixed high voltage), leading to high energy consumption (e.g., some systems have an energy saving rate of less than 10%). There is a lack of parameter iterative optimization mechanisms aimed at "minimizing energy consumption," making it impossible to dynamically reduce the power consumption per unit of dust removal while meeting standards. This invention aims to achieve coordinated control of the dual objectives of "emission compliance + minimum energy consumption."

[0105] 4) Weak fault early warning and self-diagnosis capabilities, resulting in high maintenance costs: Existing technologies rely on manual monitoring of panel parameters to detect faults (such as power supply breakdown and abnormal electric field), lacking early warning mechanisms for early anomalies such as voltage fluctuations and sudden current drops. This invention aims to achieve early fault warning (such as identification of back corona precursors) through real-time parameter analysis and provide accurate diagnosis and maintenance guidance.

[0106] By solving the above problems, this invention ultimately achieves "self-identification of operating conditions, self-calling of formulas, self-optimization of parameters, and self-early warning of faults" for electrostatic precipitator power supplies, meeting the core needs of thermal power plants for "high-efficiency dust removal, energy saving and consumption reduction, and reduced manpower operation".

[0107] Specifically, the system hardware architecture of this invention includes the following components: Intelligent variable frequency constant current power supply body: It adopts a three-phase bridge rectifier + IGBT high-frequency inverter + multi-winding transformer boost topology to achieve adjustable secondary voltage of 0-72kV and constant current output of 0-2.0A. Innovations include dual closed-loop control: the inner loop uses current PID control (response time <10ms), and the outer loop adjusts the voltage through a fuzzy algorithm, solving the oscillation problem of traditional pulse power supplies; modular design, with power modules supporting "N + 1" redundant configuration, automatically switching in case of single-module failure to ensure system availability; and a multi-dimensional sensor network: the sensing units include flue gas parameter sensors (laser scattering dust concentration meter, ultrasonic flow meter, infrared humidity meter) and power parameter sensors (Rogowski coil current sensor, Hall voltage sensor, temperature sensor). Data acquisition uses a 24-bit ADC sampling chip with a sampling frequency of 1kHz, supporting simultaneous sampling of multi-channel data. The edge computing controller system adopts an ARM Cortex-A53 quad-core processor (1.8GHz) + FPGA coprocessor architecture, supporting real-time data processing (1000 points / second), multi-recipe management (storing ≥100 sets of parameters), and reinforcement learning algorithm execution (iteration time <500ms). Communication interfaces include dual-redundant Ethernet (Modbus TCP / IP), RS485, and 4G / 5G wireless backup, ensuring reliable and flexible data transmission. This invention achieves adaptive control based on operating conditions through hardware architecture innovation, algorithm optimization, and system integration. The innovative hardware architecture design, combined with a high-efficiency edge computing controller and a multi-dimensional sensor network, provides a solid foundation for stable system operation and intelligent control, improving the overall performance and reliability of the system.

[0108] Furthermore, the method for implementing the "multi-formula self-optimization" function of the present invention is as follows: The "multi-recipe self-optimization" function implementation method of this invention covers multiple levels. First, at the basic level, the patent involves the design of a recipe database, which is scalable and can store no fewer than 10 sets of operating recipes. Each set of recipes consists of two parts: operating condition tags and core parameters. The operating condition tags can accurately define the scenarios in which the recipe is applicable, such as "high ash coal (ash content ≥30%) + high load (600 - 1000MW) + flue gas humidity (20 - 30% RH)" and "low sulfur coal (sulfur content ≤1%) + low load (300 - 600MW)," covering a variety of typical operating condition combinations in thermal power plants. The core parameters include key control parameters of the electrostatic precipitator power supply, such as secondary voltage (50 - 72kV), secondary current (0.5 - 1.2A), pulse frequency (20 - 50Hz), and charging ratio (1:2 - 1:5). These parameters are accurate to 0.1kV / 0.01A, ensuring precise control. For example, in a formula setting for "high ash content + high load" conditions, the secondary voltage is set to 68kV, the current to 1.0A, the frequency to 40Hz, and the charging ratio to 1:3. This method suppresses back corona discharge and improves dust collection efficiency, thus adapting to this scenario. Furthermore, a rapid formula retrieval mechanism is supported, allowing formulas to be retrieved through "operating condition tag search" and "parameter similarity matching." When the unit load, coal type and other operating parameters are clear, the corresponding recipe can be directly retrieved through the tag search (response time is less than 1 second); when the operating conditions are more ambiguous, the optimal recipe is automatically recommended by calculating the similarity between real-time parameters (such as ash content and load) and recipe tags (using cosine similarity algorithm) (the matching accuracy can reach 85% or more).

[0109] In the adaptation layer, the "multi-recipe self-optimization" function implementation method of this invention mainly involves multi-dimensional operating condition perception and adaptive switching due to sudden changes in operating conditions. Multi-dimensional operating condition perception collects six core parameters in real time through a sensor network, with a sampling frequency of 1Hz, constructing an operating condition feature vector. These core parameters include coal combustion characteristics (ash content detected by X-ray fluorescence analyzer, sulfur content detected by infrared sulfur analyzer), operating parameters (unit load acquired by DCS signal, flue gas volume measured by vortex flow meter), and environmental parameters (flue gas humidity detected by capacitive hygrometer, inlet dust concentration detected by laser scattering instrument). Regarding adaptive switching due to sudden changes in operating conditions, when the change in operating parameters exceeds a set threshold (e.g., load fluctuation ≥20%, ash content change ≥5%), the system triggers a "recipe switching process". If the new operating condition matches a preset formula (similarity ≥ 85%), the formula is directly called and the parameters are fine-tuned (e.g., voltage adjustment range ±1kV). If no matching formula exists, a temporary formula is generated based on the parameters of the 2-3 most similar formulas using an interpolation algorithm (e.g., combining parameter characteristics of "high ash" and "low load" formulas) to ensure stable emissions during the operating condition transition (concentration fluctuation controlled within ≤ 5mg / m³). Taking a load reduction from 800MW (high load) to 500MW (medium load) as an example, the system can identify the change in operating condition within 5 minutes, call the "medium load formula," and reduce the voltage from 68kV to 60kV, thereby avoiding energy waste under low load conditions.

[0110] At the core layer, the "multi-formula self-optimization" function of this invention is mainly implemented through reinforcement learning algorithms to achieve a dynamic balance between "emission compliance + minimum energy consumption". During formulation execution, constraints are first set, including hard constraints (outlet dust concentration ≤30mg / m³, secondary voltage ≤72kV, current fluctuation ≤0.1A / min) and optimization objectives (minimizing power consumption per unit of dust removal). Then, parameter iterative adjustments are performed, starting with the initial formulation parameters and following the logic of "small-step trial—effect evaluation—direction correction". Each time parameters are adjusted (e.g., voltage adjustment range ±0.5kV, current adjustment range ±0.05A), the dust concentration and energy consumption within one minute after the adjustment are recorded. If the concentration meets the standard and energy consumption decreases, the adjustment is retained and optimization continues in that direction; if the concentration exceeds the standard or energy consumption increases, the adjustment is reversed and the step size is reduced (e.g., the adjustment range becomes ±0.2kV). The iteration terminates when the energy consumption change is ≤0.5% after 10 consecutive adjustments, or when the maximum number of iterations (50 times) is reached. Taking a certain "low-load formula" as an example, its initial parameters are voltage 55kV and current 0.6A. During the optimization process, it was found that when the voltage dropped to 54kV, the concentration remained stable at 25mg / m³ and the energy consumption was reduced by 3%. Finally, the parameters were updated to the optimal solution.

[0111] At the evolutionary level, the "multi-formula self-optimization" function of this invention is mainly implemented through periodic self-learning and formula library updates. The system evaluates the operational data of all formulas daily and dynamically updates the formula library, thereby achieving a process of "experience accumulation—continuous optimization." Regarding formula effectiveness evaluation, formulas are scored based on three indicators (out of 100): emission compliance rate (weight 40%) is calculated as (compliance time / total operating time) × 100%; energy consumption level (weight 30%) is the ratio to the historical best energy consumption under the same operating conditions; stability (weight 30%) is reflected in parameter fluctuation amplitude, for example, the voltage standard deviation should be ≤2kV. Regarding dynamic updates to the formula library, temporary formulas with a score ≥90 are converted into formal formulas and added to the database; while inefficient formulas that have not been used for 30 consecutive days or have a score <60 are automatically eliminated; for frequently used formulas (e.g., "high ash content + medium load"), parameters are fine-tuned based on the latest data (e.g., optimized once per quarter). Taking a temporary formula as an example, after operating for one month under "high humidity + low sulfur" conditions, its emission compliance rate reached 100%, energy consumption decreased by 15%, and it ultimately scored 92 points, thus being included in the official formula library. Meanwhile, another "old, high-load formula" was eliminated due to its high energy consumption (score of 58 points). In summary, the core logic of "multi-formula self-optimization" is a three-layer architecture of "preset formula foundation + real-time data optimization + historical experience iteration," allowing the system to quickly adapt to known operating conditions (through formula invocation) and autonomously learn about unknown operating conditions (through parameter optimization and formula updates), ultimately achieving the goal of "continuously matching complex operating conditions without manual intervention, balancing environmental protection and energy consumption." Compared to traditional technologies, this function has the advantages of a response speed improved to the minute level (traditional manual adjustment requires more than 30 minutes), energy consumption reduced by ≥35%, and emission compliance rate increased to 99.9%.

[0112] Furthermore, a dual-objective reinforcement learning framework: State space: S = {dust concentration, voltage, current, frequency, energy consumption, temperature}; Operating space: A = {voltage ±0.5kV, current ±0.05A, frequency ±5Hz}.

[0113] Further, initialization: load initial parameters from the recipe library; Iterative optimization: Perform the current action and collect the new state; Calculate the reward value; Update the Q-table (Q-Learning algorithm); Choose the next action (ε-greedy strategy); Convergence criteria: The reward change is less than 0.1% for 10 consecutive iterations or the maximum number of iterations (50) is reached.

[0114] The parameter self-optimization algorithm of this invention uses reinforcement learning (Q-Learning) as its core framework, combining the nonlinear characteristics of electrostatic precipitators with dual objective requirements (emission compliance + minimum energy consumption) to achieve dynamic iterative optimization of parameters. The algorithm autonomously finds the optimal parameter combination through a closed-loop logic of "state perception - action decision - reward feedback - policy update," and the specific implementation method is as follows: The essence of parameter self-optimization algorithm is to minimize power consumption through trial and error learning under the premise of meeting environmental protection hard constraints. Its core elements include three parts: "state space, action space, and reward function", which constitute the basic model of reinforcement learning.

[0115] The state space is a quantitative description of the real-time operating characteristics of an electrostatic precipitator system, containing six key parameters (forming the state vector S), specifically:

[0116] This content describes the state space S used in a parametric self-optimization algorithm, which consists of five key parameters that collectively define the current operating state of the electrostatic precipitator system. Specifically: Cout: Represents the outlet dust concentration, measured in milligrams per cubic meter (mg / m³). This is a core environmental indicator used to measure whether the amount of dust emitted meets environmental standards. U: Represents the secondary voltage, measured in kilovolts (kV). This is a core parameter of the power supply output and directly affects the efficiency of the electrostatic precipitator system. I: Represents the secondary current, measured in amperes (A). Also a core parameter of the power supply output, it, along with the voltage, determines the power supply's output power. f: Represents the pulse frequency, measured in hertz (Hz). This parameter affects the intensity of the corona discharge, thus affecting the dust collection efficiency. P: Represents the real-time power of the power supply, measured in kilowatts (kW). This is a direct indicator of energy consumption, used to measure the energy consumption level during system operation. T: Represents the power module temperature, measured in degrees Celsius (°C). This is a safety constraint to ensure the power module operates within a safe temperature range.

[0117] These parameters collectively constitute the state space S, which is the basis for decision-making in reinforcement learning algorithms. By monitoring and adjusting these parameters, the algorithm can optimize the performance of the electrostatic precipitator system, achieving lower energy consumption and higher stability.

[0118] State discretization: To reduce computational complexity, continuous parameters are discretized into intervals (e.g., U is divided into intervals of 50-55kV, 55-60kV, etc., with a step size of 0.5kV). The final state space size is controlled between 1000-5000, balancing accuracy and efficiency.

[0119] The action space consists of the parameter adjustment operations that the algorithm can execute (forming action vector A). Based on the control characteristics of the electrostatic precipitator power supply, three types of core actions are designed:

[0120] Action constraints: All actions must meet the equipment safety boundaries (e.g., U≤72kV, I≤1.2A), and actions exceeding the boundaries are prohibited from being executed.

[0121] Reward function: Defines the optimization objective and constraints. The reward function acts as the algorithm's "command stick," guiding parameters towards optimization towards "emission compliance + lowest energy consumption" through quantitative feedback. The design is as follows: Reward function in parameter self-optimization algorithm R ( S , A , S The function consists of three parts: environmental rewards. R env, energy reward R energy and stability penalties R stable. The weighting coefficients are respectively α , β and γ This is used to determine the relative importance of each part in the total reward.

[0122] Total reward function:

[0123] in, α , β and γ It is a weighting coefficient used to adjust the proportion of each part of the reward in the total reward.

[0124] Environmental protection awards R env: The environmental incentive program is designed to prioritize ensuring emissions compliance, and the specific calculation method is as follows:

[0125] Here, Cout′ represents the new outlet concentration after the action is performed.

[0126] Energy consumption reward : The energy consumption reward aims to encourage the reduction of energy consumption per unit of dust removal. Its calculation formula is:

[0127] Where: P is the power before the action is executed. P' is the power after execution. If (P' < P), that is, the energy consumption is reduced, the reward is positive, and the more the energy consumption is reduced, the higher the reward.

[0128] Stability penalty : The stability penalty is used to suppress the drastic fluctuations of parameters to avoid system oscillation. Its calculation formula is:

[0129] The greater the parameter adjustment amplitude, the heavier the penalty.

[0130] The weight coefficients are used to determine the priorities of environmental protection, energy consumption, and stability in the total reward: (Environmental protection first), (Energy consumption second), (Stability as assistance).

[0131] These weight coefficients ensure that the target priorities meet the actual requirements, that is, in the optimization process, the environmental protection goal is the most important, followed by energy consumption, and finally the stability of the system.

[0132] Furthermore, the parameter self-optimization algorithm is executed through three steps: "Initialization - Iterative Optimization - Convergence Judgment". The specific process is as follows: Initialization (Initialization): Initialization of the Q table: The Q table is used to store the matrix of "state-action" values. Its dimension is the number of states multiplied by the number of actions, and the initial value is set to 0, which will be updated iteratively later.

[0133] Initial parameter setting: Call the initial parameters matching the current working condition from the recipe library (for example, U0 = 60 kV, I0 = 0.8 A, f0 = 30 Hz) as the starting point for optimization. Parameter boundary definition: Set the hard constraints for parameter adjustment (such as U ∈ [50, 72] kV, I ∈ [0.5, 1.2] A), and actions beyond the boundary will not be executed.

[0134] Iterative Optimization (Iterative Optimization): The algorithm loops and executes the following operations at a frequency of 1 second / time until convergence: State perception: Collect the current state St = [Cout, U, I, f, P, T]; Action selection: Adopt ( Action selection: 90% probability: select the action with the highest value in the current Q table (exploitation, utilizing known experience); 10% probability: randomly select an action (exploration, exploring new strategies); Action execution: Execute the selected action At and adjust the parameters (e.g., Ut+1=Ut+ΔU). State transition: Collect the new state St+1 after the action is executed; Reward calculation: Calculate the reward Rt based on St, At, and St+1; Q-table update: Optimization strategy using Q-Learning update formula:

[0135] Where α=0.1 is the learning rate and γ=0.9 is the discount factor, balancing current and future rewards.

[0136] Constraint check: If the new state middle If the value exceeds the limit, an "emergency callback" action will be enforced (e.g., U increases by 1kV and I increases by 0.1A) to ensure environmental compliance.

[0137] Convergence Criterion: The algorithm terminates its iteration and outputs the current parameters as the optimal solution when any of the following conditions are met: 1. Reward value fluctuation ≤ 1% over 20 consecutive iterations: This indicates that the strategy has stabilized, and further optimization has little impact on the reward value. 2. Cumulative iterations reach 50: To avoid over-optimization leading to system oscillations, a maximum iteration limit is set. 3. Energy consumption index P drops to within 95% of the historical lowest value under similar operating conditions: This indicates that it is close to the theoretical optimal value, and the space for further optimization is limited.

[0138] Furthermore, taking the "low load (500MW) + low sulfur coal (sulfur content = 0.8%)" operating condition of Huaneng Jinling Power Plant as an example, the algorithm execution process is explained: Initial state: "Low load formula" is activated with parameters U=55kV, I=0.6A, f=30Hz. At this time, Cout=28mg / m3 and P=120kW. Iterations 1-5: Attempt to reduce U to 54.5kV (action) With Cout remaining at 28 mg / m³ and P decreasing to 115 kW, the bonus calculation is as follows:

[0139] The value of the "state-action" pair in table Q is updated from 0 to 62.25.

[0140] Iterations 6-15: Continue to reduce U to 54kV, I to 0.55A, Cout = 27mg / m3, P = 108kW, and calculate the reward as follows:

[0141] After multiple iterations, the action value corresponding to "U=54kV,I=0.55A,f=30Hz" in the Q table is the highest.

[0142] Convergence Termination: In the 20th iteration, the parameters stabilize at U=54kV, I=0.55A, f=30Hz, the outlet dust concentration Cout=27mg / m3 (meets the standard), the energy consumption P=108kW (10% lower than the initial value), and the reward fluctuation ≤1%, at which point the algorithm terminates.

[0143] Furthermore, the fault early warning and self-diagnosis system includes an anomaly detection model specifically designed to identify and warn of potential faults in electrostatic precipitator (ESP) systems. The system primarily focuses on two types of early warnings: First, back corona warning, which extracts features by monitoring voltage fluctuations (when fluctuations reach or exceed 15%), sudden current increases (increases exceeding 20%), and significant decreases in electric field impedance (decreases exceeding 30%). If these conditions are simultaneously met and persist for 30 seconds, the system will determine it as a precursor to back corona and issue an early warning. Second, rectifier fault warning, which mainly monitors the ripple coefficient of the secondary current. The normal range for this coefficient should be between 0.1 and 0.3. If the ripple coefficient exceeds 0.5 and persists for 1 minute, the system will issue an early warning, indicating "the rectifier IGBT may be damaged." This early warning mechanism helps to identify and handle faults promptly, thereby ensuring the normal operation of the ESP system and maintaining its long-term stability.

[0144] The fault diagnosis knowledge base and diagnostic rules define a fault database and a diagnostic function to identify potential faults based on abnormal characteristics. The fault database is a dictionary containing two types of faults and their corresponding feature lists: Large voltage fluctuations: possibly caused by cable insulation degradation, electric field short circuits, or controller PID parameter misalignment. Abnormally high current: possibly caused by corona wire breakage, ash hopper blockage, or rapping system malfunction. The diagnostic function `diagnose` accepts an parameter `abnormal characteristics`, representing the abnormal conditions detected by the system. The function aims to determine the possible faults corresponding to these abnormal characteristics. The function's steps are as follows: Initialize an empty list `possible faults` to store possible faults. Iterate through each fault in the fault database and its corresponding feature list. For each fault, check if the passed `abnormal characteristics` contain all the characteristics of that fault in the fault database. If all characteristics match, add the fault to the `possible faults` list. Finally, the function returns the `possible faults` list. This diagnostic system helps identify possible fault causes by comparing abnormal characteristics with the feature list in the fault database, thereby enabling fault warnings and self-diagnosis.

[0145] Furthermore, in a sudden change in operating conditions (high load → low load): when the unit load drops from 800MW to 500MW and the flue gas volume decreases by 30%, the system can identify the load drop within 5 minutes and trigger the "low load condition". The system automatically calls "Formula 2" (voltage 55kV, current 0.6A) and, through a self-optimization algorithm, reduces the current to 0.55A within 30 minutes, reducing energy consumption by 8%, while ensuring that the emission concentration remains stable at 22mg / m³. The next day, the self-learning program updates the "low load current" parameter from 0.6A to 0.55A to adapt to the new operating condition requirements.

[0146] Coal type switching scenario (low ash content → high ash content): When the ash content of the coal increases from 20% to 35%, and the inlet dust concentration increases from 30g / m³ to 50g / m³, the system identifies the coal type change within 10 minutes using dust concentration and ash content sensors. An emergency adjustment triggers a temporary "anti-corona discharge" formula, reducing the voltage to 60kV and increasing the frequency to 50Hz. The charging ratio is dynamically adjusted using a reinforcement learning algorithm, stabilizing the outlet concentration at 28mg / m³ within 20 minutes. After 24 hours of operation, the new formula achieved a comprehensive score of 92 points and was included in the official formula library.

[0147] Fault warning scenario (rectifier failure): Monitoring data shows that the secondary current ripple coefficient suddenly increases from 0.2 to 0.6 and lasts for 2 minutes. The system pushes a "rectifier diode damaged" warning to maintenance personnel, along with a fault location diagram (e.g., "phase A, group 3 diodes") and repair steps. To ensure continuous system operation, it automatically switches to the backup rectifier module.

[0148] The key technical indicators of the system include: operating condition response time of less than 5 minutes (normal operating conditions) and less than 30 minutes (complex operating conditions); power saving rate of at least 35% (compared to traditional power frequency power supply); emission compliance rate of at least 99.9%; fault early warning accuracy rate of at least 95%; and formula optimization cycle of daily automatic updates with parameter adjustment step size not exceeding 5%.

[0149] Through the above technical solution, this patent realizes intelligent and adaptive control of the electrostatic precipitator power supply, significantly improving its adaptability to operating conditions, reducing energy consumption, and minimizing manual intervention. This promotes the transformation of environmental protection equipment in thermal power plants towards "less manned operation and intelligent maintenance," improving operational efficiency and environmental performance.

[0150] This invention achieves adaptive control of operating conditions through innovative hardware architecture, algorithm optimization, and system integration, thereby enhancing the intelligence level of electrostatic precipitator systems. Through self-learning and self-optimization mechanisms, the system can quickly respond to changes in operating conditions, optimize operating parameters, ensure emissions meet standards, and reduce energy consumption, providing strong support for environmental protection and energy efficiency management in thermal power plants.

[0151] In summary, the present invention has the following technical effects: This invention significantly improves the stability of environmental emissions through its core technology of "multi-formula self-optimization," completely solving the problem of instantaneous exceedances. Existing technologies have poor adaptability to operating conditions, resulting in large fluctuations in dust concentration (±10mg / m³) and an average of ≥15 exceedances per year. In contrast, this invention, through closed-loop control of "operating condition labeling - formula matching - parameter optimization," controls the fluctuation of outlet dust concentration within ±3mg / m³, increasing the emission compliance rate from 95% to 99.9%. For example, at Huaneng Jinling Power Plant, under high-ash coal (30% ash) operating conditions, traditional power supplies caused an average of 3 exceedances per month due to back corona discharge, while using this invention's technology resulted in zero exceedances for 6 consecutive months, meeting the latest ultra-low emission standards (≤30mg / m³). This invention also demonstrates significant advantages in energy consumption control, creating direct economic benefits. Existing technologies rely on redundant power supply (such as fixed high voltage), with energy savings generally below 10% and unit dust removal power consumption ≥150kWh / t. This invention employs a dual-objective reinforcement learning algorithm to achieve "on-demand power supply" while ensuring compliance with standards, resulting in a power saving rate of 15-35% and a unit power consumption reduction to 80-120 kWh / t. Taking a 1030MW unit as an example, annual power consumption can be reduced by 300,000-500,000 kWh. Calculated at 0.3 yuan / kWh, this translates to annual electricity cost savings of 90,000-150,000 yuan per unit, and for the entire power plant with two units, annual electricity cost savings of 180,000-300,000 yuan.

[0152] This invention improves response speed by 6 times, adapting to complex and ever-changing operating scenarios. Existing technologies require manual parameter adjustments when operating conditions change, resulting in a lag (30-60 minutes) and potential emission fluctuations or energy waste during transitions. This invention's dynamic formula library and similarity matching algorithm enable operating condition identification and parameter switching within 5 minutes, coupled with 1-second-level parameter iteration optimization, ensuring stable system operation during sudden changes in operating conditions (such as a 20% load drop or coal type switching). For example, when a unit reduces its load from 800MW (high load) to 500MW (medium load), traditional technology requires 40 minutes of adjustment, while this invention completes parameter adjustments within 5 minutes, with a maximum concentration fluctuation of only 4mg / m³ during the transition, reducing energy consumption by 12%.

[0153] This invention significantly improves operational efficiency and reduces maintenance costs. Existing technologies rely on manual inspections and parameter adjustments, requiring two dedicated maintenance personnel per unit, with an average annual maintenance time of ≥1000 hours and a fault handling time of ≥4 hours per incident. This invention reduces manual operation by 80%: it automatically completes recipe recall and parameter optimization, requiring only system status monitoring by operators without manual adjustments; it also improves fault warning and diagnosis efficiency: multi-sensor fusion warning accuracy is ≥95%, fault location time is reduced from 4 hours to 1.5 hours, and average annual maintenance time is reduced to below 400 hours; and it features parameter copying: recipe parameters are "plug and play" via SD card, reducing controller replacement and debugging time from 3 days to 2 hours.

[0154] This invention enhances the robustness of the system and adapts to extreme operating conditions. Existing technologies are prone to problems such as power supply oscillation and back corona under extreme conditions (e.g., high humidity flue gas, blending of inferior coal), resulting in equipment failure rates ≥8 times / year. The anti-back corona formula of this invention effectively suppresses back corona phenomena in high-ash coal types by reducing the average voltage and increasing the pulse frequency, reducing the number of electric field breakdowns by 70%. The modular power supply design supports "N+1" redundancy, automatically switching in case of single-module failure, increasing system availability from 98% to 99.9%. The wide-temperature design (-20℃ to +60℃) adapts to harsh environments, ensuring stable operation in low winter temperatures or high summer temperatures. In summary, this invention represents a technological leap from "passive control" to "active adaptation." By combining "experience-based formulations" with "intelligent algorithms," it upgrades the electrostatic precipitator power supply from traditional "fixed parameter adjustment" to an intelligent system capable of "self-sensing, self-decision-making, and self-optimization." This not only meets the stringent requirements of thermal power plants for "environmental compliance and energy conservation," but also aligns with the industry trend of "intelligent transformation of auxiliary control systems with fewer personnel," thus possessing broad application value.

[0155] Example 3 This invention also provides a multi-formula adaptive intelligent variable frequency power supply control device 10 for electrostatic precipitators, such as... Figure 5 As shown, the device 10 includes: The dynamic feature vector construction module 100 is used to acquire coal type parameters, unit load parameters, flue gas humidity parameters and dust concentration parameters in real time, and construct a dynamic feature vector containing six types of operating condition features. The formula matching and selection module 200 is used to perform cosine similarity matching between the dynamic feature vector and the working condition labels in the preset formula library, and select the optimal formula above the similarity threshold as the initial control parameter. The reinforcement learning optimization module 300 is used to perform reinforcement learning bi-objective optimization based on the initial control parameters. Through small step parameter adjustment and real-time effect evaluation, it dynamically balances dust emission compliance and energy consumption minimization. The recipe library update module 400 is used to periodically update the recipe library, converting optimized parameter combinations that have reached a preset threshold in performance evaluation into new recipes for storage, and eliminating inefficient recipes.

[0156] Furthermore, the dynamic feature vector construction module is also used for: Inlet dust concentration data are collected using a laser scattering dust concentration meter, which has a range of 0-1000 mg / m³ and an accuracy of ±5 mg / m³. The ash content parameters of coal were obtained using an X-ray fluorescence analyzer. The sampling frequency of the ash content parameters was 1 Hz and the data update cycle was 5 minutes.

[0157] Furthermore, the recipe matching selection module is also used for: The cosine similarity between the real-time operating condition feature vector and the recipe label vector is calculated using the following formula:

[0158] in Let i be the i-th parameter of the real-time operating condition feature vector. The i-th parameter of the recipe label vector; A similarity threshold of 85% is set. When the matching similarity is lower than this threshold, an interpolation algorithm is used to generate a temporary combination of recipe parameters.

[0159] Furthermore, the reinforcement learning optimization module is also used for: The Q-table is updated using the Q-Learning algorithm, with the following formula:

[0160] in For learning rate, Discount factor; When performing the operation, the adjustment step size for the secondary voltage is limited to ±0.5kV, the adjustment step size for the secondary current is limited to ±0.05A, and the adjustment step size for the pulse frequency is limited to ±2Hz.

[0161] Furthermore, the recipe library update module is also used for: Formulas are scored based on three indicators: compliance rate, energy consumption level, and stability. The compliance rate has a weight of 40%, the energy consumption level has a weight of 30%, and the stability has a weight of 30%. Formulas with a score below 60 are eliminated, while temporary formulas with a score of 90 or higher and a running time of more than 30 days are solidified.

[0162] Furthermore, it also includes: The anti-corona warning module is used to determine the anti-corona precursor and trigger an emergency callback action when the voltage fluctuation is ≥15%, the current surge is ≥20%, and the electric field impedance drops by ≥30% for 30 seconds. The rectifier fault diagnosis module is used to monitor the secondary current ripple coefficient. When the ripple coefficient is >0.5 and lasts for 1 minute, it will warn of rectifier IGBT damage and locate the fault module.

[0163] An embodiment of the present invention provides an intelligent variable frequency power supply control device for electrostatic precipitators with multi-formula adaptive function. By real-time monitoring of voltage fluctuations, current surges, and changes in electric field impedance, combined with secondary current ripple coefficient analysis, the system can accurately identify precursors of back corona and rectifier IGBT damage faults, and achieve rapid location and emergency reversal of faulty modules, further improving the reliability and maintenance efficiency of the electrostatic precipitator power supply.

[0164] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0165] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0166] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for regulating an intelligent variable frequency power supply for an electrostatic precipitator with multi-formula adaptive function, characterized in that, include: S1 acquires real-time parameters of coal type, unit load, flue gas humidity, and dust concentration, and constructs a dynamic feature vector containing six types of operating conditions. S2, based on the dynamic feature vector and the working condition labels in the preset formula library, perform cosine similarity matching, and select the optimal formula above the similarity threshold as the initial control parameter; S3, perform reinforcement learning bi-objective optimization based on the initial control parameters, and dynamically balance dust emission compliance and energy consumption minimization through small step parameter adjustment and real-time effect evaluation; S4 periodically updates the recipe library, converting optimized parameter combinations that have reached a preset threshold in performance evaluation into new recipes for storage, and eliminating inefficient recipes.

2. The method as described in claim 1, characterized in that, The real-time acquisition of coal type parameters, unit load parameters, flue gas humidity parameters, and dust concentration parameters, and the construction of a dynamic feature vector containing six types of operating condition characteristics, also includes: S11, Inlet dust concentration data are collected by a laser scattering dust concentration meter, the range of which is 0-1000mg / m³ and the accuracy is ±5mg / m³; S12, X-ray fluorescence analyzer is used to obtain coal ash parameters. The sampling frequency of the ash parameters is 1Hz and the data update cycle is 5 minutes.

3. The method as described in claim 1, characterized in that, The cosine similarity matching based on the dynamic feature vector and the working condition labels in the preset recipe library also includes: S21, Calculate the cosine similarity between the real-time operating condition feature vector and the recipe label vector, using the following formula: in Let i be the i-th parameter of the real-time operating condition feature vector. The i-th parameter of the recipe label vector; S22, set the similarity threshold to 85%. When the matching similarity is lower than this threshold, use an interpolation algorithm to generate a temporary recipe parameter combination.

4. The method as described in claim 1, characterized in that, The step of performing reinforcement learning bi-objective optimization based on the initial control parameters further includes: S31, update the Q-table using the Q-Learning algorithm, the formula is: in For learning rate, Discount factor; S32, when performing the action, limits the secondary voltage adjustment step size to ±0.5kV, the secondary current adjustment step size to ±0.05A, and the pulse frequency adjustment step size to ±2Hz.

5. The method as described in claim 1, characterized in that, The periodically updated formula library also includes: S41, the formula is scored based on three indicators: compliance rate, energy consumption level, and stability. The compliance rate has a weight of 40%, the energy consumption level has a weight of 30%, and the stability has a weight of 30%. S42, perform elimination operation on formulations with a score below 60, and perform solidification operation on temporary formulations with a score of 90 or above and a running time of more than 30 days.

6. The method as described in claim 1, characterized in that, Also includes: S5, execute the back corona warning and rectifier fault diagnosis steps: S51, when the voltage fluctuation is ≥15%, the current surge is ≥20%, and the electric field impedance drops by ≥30% for 30 seconds, it is determined to be a precursor to back corona and an emergency callback action is triggered. S52 monitors the secondary current ripple coefficient. When the ripple coefficient is >0.5 and lasts for 1 minute, it will warn of rectifier IGBT damage and locate the faulty module.

7. A multi-formula adaptive intelligent variable frequency power supply control device for electrostatic precipitators, characterized in that, include: The dynamic feature vector construction module is used to acquire coal type parameters, unit load parameters, flue gas humidity parameters, and dust concentration parameters in real time, and construct dynamic feature vectors containing six types of operating condition features. The formula matching and selection module is used to perform cosine similarity matching between the dynamic feature vector and the working condition labels in the preset formula library, and select the optimal formula above the similarity threshold as the initial control parameter. The reinforcement learning optimization module is used to perform reinforcement learning bi-objective optimization based on the initial control parameters. Through small step parameter adjustment and real-time effect evaluation, it dynamically balances dust emission compliance and energy consumption minimization. The recipe library update module is used to periodically update the recipe library, converting optimized parameter combinations that have reached a preset threshold in performance evaluation into new recipes for storage, and eliminating inefficient recipes.

8. The apparatus as claimed in claim 7, characterized in that, The dynamic feature vector construction module is also used for: Inlet dust concentration data are collected using a laser scattering dust concentration meter, which has a range of 0-1000 mg / m³ and an accuracy of ±5 mg / m³. The ash content parameters of coal were obtained using an X-ray fluorescence analyzer. The sampling frequency of the ash content parameters was 1 Hz and the data update cycle was 5 minutes.

9. The apparatus as claimed in claim 7, characterized in that, The recipe matching and selection module is also used for: The cosine similarity between the real-time operating condition feature vector and the recipe label vector is calculated using the following formula: in Let i be the i-th parameter of the real-time operating condition feature vector. The i-th parameter of the recipe label vector; A similarity threshold of 85% is set. When the matching similarity is lower than this threshold, an interpolation algorithm is used to generate a temporary combination of recipe parameters.

10. The apparatus as claimed in claim 7, characterized in that, The reinforcement learning optimization module is also used for: The Q-table is updated using the Q-Learning algorithm, with the following formula: in For learning rate, Discount factor; When performing the operation, the adjustment step size for the secondary voltage is limited to ±0.5kV, the adjustment step size for the secondary current is limited to ±0.05A, and the adjustment step size for the pulse frequency is limited to ±2Hz.

Images