A machine learning-based electric dust removal power supply parameter optimization method and system, electronic device, and storage medium

Abstract

This invention provides a machine learning-based method, system, electronic device, and storage medium for optimizing power supply parameters of an electrostatic precipitator (ESP). The method includes: real-time acquisition and preprocessing of operating status data of the ESP system; inputting the preprocessed operating status data into a pre-trained reinforcement learning agent to obtain an optimized combination of power supply control parameters; wherein the reinforcement learning agent is trained to maximize a preset multi-objective reward function; and controlling the ESP power supply according to the optimized combination of power supply control parameters. The autonomous operation of the entire "perception-decision-execution" closed loop of this invention marks a leap from traditional automation to true intelligence in ESP power supply control, laying a core technological foundation for building an unmanned, self-optimizing "smart dust removal" system.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a method, system, electronic device, and storage medium for optimizing the power supply parameters of an electrostatic precipitator based on machine learning. Background Technology

[0002] Traditional electrostatic precipitators often employ simple closed-loop control (such as spark tracking and critical spark control) based on voltage-current curves (Ul characteristics) for their high-voltage power supplies. However, the electrostatic precipitator process is a typical nonlinear, large-time-lag, multivariable coupled system, and its optimal operating point is dynamically influenced by various factors such as flue gas conditions (temperature, humidity, composition, flow rate), dust characteristics (resistivity, particle size, concentration), and electrode ash accumulation status.

[0003] Existing technologies have the following main shortcomings: 1. Response lag: Parameters are only adjusted after a spark occurs or when the current changes drastically, making it impossible to predict changes in operating conditions. 2. Lack of global optimization: Focusing only on stable operation or maximum dust removal efficiency makes it difficult to achieve a dynamic balance between "efficiency improvement" and "energy saving". 3. Reliance on expert experience: Parameter settings rely heavily on manual experience, making it unable to adapt to complex and changing operating conditions, resulting in high energy consumption or unstable efficiency during long-term operation. Summary of the Invention

[0004] The present invention aims to address the deficiencies in the existing technology by providing a method, system, electronic device and storage medium for optimizing the power supply parameters of electrostatic precipitators based on machine learning, thereby improving the system's adaptability to complex working conditions, reducing manual intervention and achieving long-term stable operation.

[0005] To achieve the above objectives, the present invention provides the following solution: A machine learning-based method for optimizing the power supply parameters of an electrostatic precipitator includes: Real-time acquisition and preprocessing of operating status data of the electrostatic precipitator system; The preprocessed operating condition data is input into a pre-trained reinforcement learning agent to obtain an optimized combination of power control parameters; wherein the reinforcement learning agent is trained with the aim of maximizing a preset multi-objective reward function. The power supply for the electrostatic precipitator is controlled according to the optimized combination of power control parameters.

[0006] Preferably, the method for preprocessing the operating condition data includes: A time-series data fusion layer based on an edge computing gateway is constructed, which integrates dynamic time warping algorithm and adaptive Kalman filtering to uniformly align the operating condition data to an equally spaced timestamp sequence and mark the data confidence level. An unsupervised clustering algorithm is used to perform modal identification on the uniformly aligned working condition data to obtain modal labels; Based on the modal labels, the derived features of the power supply performance indicators are calculated; By utilizing a learnable feature attention layer, the uniformly aligned working condition state data, the modality labels, and the derived features are fused together to obtain the working condition state vector, thus completing the preprocessing of the working condition state data.

[0007] Preferably, the method for calculating the derived features includes: Based on the operating range indicated by the modal label, the sliding window statistical characteristics and trend characteristics of the power performance indicators are calculated to obtain the health status vector; Based on the modal labels, the corresponding mechanism-data fusion model is invoked to estimate the equivalent resistivity of the dust layer and the corona power density distribution between the plate lines, thereby obtaining a virtual feature vector. The derived features are obtained based on the health status vector and the virtual feature vector.

[0008] Preferred methods for constructing reinforcement learning agents include: Based on the historical operating status data of the electrostatic precipitator system and the corresponding power control parameters, a hybrid prediction model integrating prior knowledge and data-driven approaches is constructed. The constructed dust removal efficiency reward, system energy consumption penalty, and spark risk penalty are weighted and summed to obtain a multi-objective reward function; The policy neural network with parameters to be optimized is used as the initial agent; Based on the proximal policy optimization algorithm, a reinforcement learning training framework is constructed that incorporates a secure projection layer and a course learning strategy. Based on the hybrid prediction model, with the goal of maximizing the multi-objective reward function, the initial agent is trained using a reinforcement learning training framework to obtain the reinforcement learning agent.

[0009] Preferred methods for constructing hybrid prediction models that integrate prior knowledge and data-driven approaches include: The electric field of the electrostatic precipitator is simplified into a nonlinear equivalent circuit. Based on historical operating condition data and corresponding power supply control parameters, the predicted secondary voltage and secondary current are calculated. Based on the flue gas flow rate, concentration, and predicted secondary current in the historical operating condition data, the trend of dust removal efficiency change and the estimated value of outlet concentration are estimated. Based on the changing trends of secondary voltage, secondary current, dust removal efficiency, and estimated outlet concentration, a mechanism prediction vector is obtained. A gated recurrent unit network with an attention mechanism is constructed to learn the residual between the response of the real electrostatic precipitator system and the mechanism prediction vector, and obtain the residual prediction vector. The mechanism prediction vector and the residual prediction vector are corrected and fused to obtain a complete prediction state vector, thus completing the construction of the hybrid prediction model.

[0010] The present invention also provides a machine learning-based electrostatic precipitator power supply parameter optimization system for implementing the method, comprising: The data processing module is used to collect and preprocess the operating status data of the electrostatic precipitator system in real time; The parameter optimization module is used to input the preprocessed operating condition data into a pre-trained reinforcement learning agent to obtain an optimized combination of power control parameters; wherein the reinforcement learning agent is trained with the aim of maximizing a preset multi-objective reward function. The control module is used to control the electrostatic precipitator power supply according to the optimized power control parameter combination.

[0011] Preferably, the data processing module includes: The data alignment unit is used to construct a time-series data fusion layer based on an edge computing gateway. It integrates dynamic time warping algorithm and adaptive Kalman filtering to uniformly align the working condition data to an equally spaced timestamp sequence and marks the data confidence level. The modality recognition unit is used to perform modality recognition on the uniformly aligned working condition data using an unsupervised clustering algorithm to obtain modality labels; A derived feature calculation unit is used to calculate derived features of the power performance index based on the modal label; The feature fusion unit is used to utilize a learnable feature attention layer to fuse the uniformly aligned working condition state data, the modal labels, and the derived features to obtain the working condition state vector and complete the preprocessing of the working condition state data.

[0012] Preferably, the derived feature calculation unit includes: The health status vector calculation subunit is used to calculate the sliding window statistical characteristics and trend characteristics of the power performance indicators based on the operating range indicated by the modal label, and obtain the health status vector. The virtual feature vector calculation subunit is used to call the corresponding mechanism-data fusion model based on the modal label to estimate the equivalent resistivity of the dust layer and the corona power density distribution between the plate lines, and obtain the virtual feature vector. The derived feature acquisition subunit is used to obtain the derived features based on the health state vector and the virtual feature vector.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in any one of claims 1 to 5.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when executed, implements the method as described in any one of claims 1 to 5.

[0015] Compared with existing technologies, the advantages of this invention are as follows: Traditional methods often employ control based on fixed rules or local responses (such as spark tracking), which struggles to maintain the optimal operating point under complex and changing conditions. This invention, through continuous optimization of the multi-objective reward function by a reinforcement learning agent, can dynamically calculate the globally optimal power parameter combination based on real-time flue gas conditions, dust characteristics, and equipment status. This enables the system to significantly reduce overall system energy consumption (including high-voltage power supply energy consumption and auxiliary energy consumption such as rapping) while ensuring stable or even better outlet emission concentrations, thus resolving the long-standing industry dilemma of "difficulty in achieving both high efficiency and low energy consumption."

[0016] The operating conditions of electrostatic precipitators fluctuate dramatically due to factors such as coal type, load, and weather. The intelligent agent in this invention, trained on massive amounts of historical data and a digital twin environment, has learned optimization strategies for various typical and boundary conditions. When operating conditions fluctuate during actual operation, the system can quickly and proactively adjust to optimized parameters matching the new conditions based on preprocessed, semantically information-rich state features, rather than passively waiting for system imbalance before correction. This significantly improves the system's robustness and adaptability to complex operating conditions, reducing reliance on manual adjustments. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the electrostatic precipitator power supply parameter optimization method based on machine learning, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0019] Figure label: 1010, Processor; 1020, Memory; 1030, Input / Output Interface; 1040, Communication Interface; 1050, Bus. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: like Figure 1 As shown, a method for optimizing the power supply parameters of an electrostatic precipitator based on machine learning includes: S1: Real-time acquisition and preprocessing of operating status data of the electrostatic precipitator system.

[0023] A further implementation method includes a preprocessing method for the operating condition data, comprising: S11: Construct a time-series data fusion layer based on an edge computing gateway, fusing dynamic time warping and adaptive Kalman filtering to uniformly align operational status data onto a sequence of equally spaced timestamps and label the data confidence levels. Specifically, the time-series data fusion layer based on the edge computing gateway first receives raw data streams from sensors with different sampling frequencies and protocols. For rapidly changing electrical signals and slowly changing process parameters, a dynamic time warping algorithm is used to align the sequences to compensate for variable time delays caused by transmission and processing, forming a preliminary synchronized data sequence. Subsequently, the aligned multi-dimensional data is input into a multivariate adaptive Kalman filter. The observation noise parameters of this filter are not fixed but dynamically adjusted according to the real-time signal quality and historical reliability of each sensor, thereby achieving robust state-optimal estimation and outputting a fused data sequence uniformly spaced on equally spaced timestamps. Simultaneously, a quantified confidence score is generated for each data point in the fused data, which integrates the residuals of the filtering process, sensor health status, and data freshness.

[0024] S12: An unsupervised clustering algorithm is used to perform modal identification on uniformly aligned operating condition data to obtain modal labels. Specifically, a Gaussian mixture model is used to automatically identify multiple inherent steady-state operating modes of the system, and a feature template library is established for each mode. During online operation, a lightweight probabilistic classifier is used to calculate the matching probability between the current data and each modal template in real time, and the most likely operating condition category is output as a semantic modal label, such as high-load normal combustion state or low-load high resistivity state.

[0025] S13: Based on modal tags, calculate derived features of power supply performance indicators; a further implementation method includes: Based on the operating range indicated by the modal label, the sliding window statistical characteristics and trend characteristics of the power supply performance indicators are calculated to obtain the health state vector; based on the modal label, the corresponding mechanism-data fusion model is invoked to estimate the equivalent resistance ratio of the dust layer and the corona power density distribution between the board lines to obtain the virtual feature vector; based on the health state vector and the virtual feature vector, the derived features are obtained.

[0026] Specifically, the process of calculating derived features begins with a deep analysis of modal labels. These labels not only indicate the macroscopic operating status but also implicitly define the normal behavioral boundaries and potential risk types that the equipment should exhibit under the current operating conditions. First, based on the historical data archives corresponding to the modal labels, key power supply performance indicators under the current mode are obtained, such as the tracking error of the secondary voltage setpoint and the typical range and distribution of primary current harmonic content. Based on this, two parallel sliding time windows are dynamically set: a fast window reflecting short-term fluctuations and a slow window characterizing long-term evolution. Within the fast window, the root mean square error, skewness, and kurtosis of the aforementioned performance indicators are calculated in real time to capture the instantaneous accuracy and abnormal pulses of control. Within the slow window, time series decomposition techniques are used to extract the seasonality and trend terms of the indicators, paying particular attention to the slope changes of their linear and nonlinear trends, thereby quantifying the degree of gradual degradation of equipment performance, such as calculating the percentage decrease in corona power efficiency over the past 24 hours. These statistics and trend quantities extracted from different time scales together constitute a health status vector characterizing the real-time health status of the equipment.

[0027] Simultaneously, the system activates the corresponding mechanism-data fusion estimation channel based on the modal label. For the crucial and difficult-to-measure parameter of the equivalent resistivity of the dust layer, the system executes two estimation paths in parallel. The first path is the mechanism-guided path, which, based on the currently measured voltage and current values, substitutes them into a simplified corona discharge current formula to deduce the initial value of the resistivity. The second path is the data compensation path, which inputs current meteorological parameters such as flue gas temperature and humidity, as well as specific features in the aforementioned health state vector, such as current stability indicators, into a pre-trained deep neural network. This network learns the nonlinear mapping relationship between the actual resistivity and the value calculated by the ideal mechanism model under complex operating conditions. Finally, the system fuses the results of the two paths through an adaptive weighting module. The weights are dynamically adjusted according to the confidence level of the current data and the typical characteristics of the modes, thereby outputting a robust resistivity estimate. For the corona power density distribution between the plate and line, the system first establishes a basic three-dimensional electric field simulation model based on the electrical wiring topology and plate geometry parameters, and then quickly solves the model using real-time secondary voltage and current as boundary conditions to obtain the theoretical power density spatial distribution. Next, a convolutional neural network corrects the theoretical distribution map in real time. By learning the differences between the actual corona effect and the theoretical distribution in different modalities in historical data, the network can locally enhance or weaken the theoretical distribution based on the current modal label and the historical location information of local spark events, generating a more realistic corona power density distribution map, and extracting features such as uniformity index and local hot spot intensity from it to form a virtual feature vector.

[0028] Ultimately, instead of simply concatenating the health state vector with the virtual feature vector, a deep integration is achieved through a feature interaction layer. This layer employs a cross-attention mechanism, allowing health state features and virtual features to query and verify each other. For example, the health feature "voltage tracking error trend" can be used as a query to focus on the correlation between "specific resistance estimation" and "corona distribution uniformity" in the virtual feature vector, thereby generating higher-order fused features such as "control difficulty coefficient due to increased specific resistance." The new feature set generated through this interaction process is the final output, a derived feature that deeply reflects the system's intrinsic state and causal relationships.

[0029] S14: Utilizing a learnable feature attention layer, the system fuses uniformly aligned operating condition data, modal labels, and derived features to obtain an operating condition vector, completing the preprocessing of the operating condition data. Specifically, this layer concatenates the basic fused data obtained in the previous steps, the encoded modal labels, and various derived features to form a high-dimensional feature set. During training, this attention layer learns to dynamically assign importance weights to different features based on the current modal labels and data confidence levels, thereby automatically focusing on the most relevant and reliable information related to the current operating state. After this weighted reconstruction, an enhanced operating condition vector with fixed dimensions, highly condensed information, and oriented towards decision optimization tasks is output, thus completing the creative preprocessing of the entire operating condition data.

[0030] S2: Input the pre-processed operating condition data into the pre-trained reinforcement learning agent to obtain an optimized combination of power control parameters; wherein the reinforcement learning agent is trained with the aim of maximizing a preset multi-objective reward function.

[0031] A further implementation method involves constructing a reinforcement learning agent, including: Based on the historical operating status data of the electrostatic precipitator system and the corresponding power control parameters, a hybrid prediction model integrating prior knowledge and data-driven approaches is constructed. A further implementation method involves constructing a hybrid prediction model that integrates prior knowledge and data-driven approaches, including: The electric field of the electrostatic precipitator is simplified into a nonlinear equivalent circuit. Based on historical operating condition data and corresponding power supply control parameters, the predicted secondary voltage and secondary current are calculated. Specifically, the high-voltage electric field of the electrostatic precipitator is first simplified into an equivalent circuit model containing a nonlinear variable resistor and capacitor according to electric field theory and circuit principles. The parameters of this model are obtained by curve fitting and optimization calibration of actual voltage and current data under historical operating conditions. During simulation, the system state vector of the previous moment and the power supply control parameters output by the current agent are used as inputs. By solving the differential equations of the equivalent circuit, the predicted effective value of the secondary voltage waveform and the average value of the secondary current are calculated in real time. At the same time, the circuit topology is automatically adjusted according to the power supply operating mode to adapt to intermittent power supply or pulse power supply.

[0032] Based on historical operating condition data of flue gas flow rate, concentration, and predicted secondary current, the trend of dust removal efficiency and the estimated value of outlet concentration are estimated. Specifically, based on the secondary current value calculated above, combined with the flue gas flow rate and inlet dust concentration synchronously collected in historical operating condition data, the dust classification and collection efficiency is calculated using the DEIC formula corrected by field data, and the instantaneous change rate of the total efficiency is obtained through integral calculation. At the same time, a mass balance differential equation is constructed with the estimated values ​​of flue gas residence time, electric field strength, and dust particle size distribution as inputs, and the dynamic estimated value of outlet dust concentration is solved in real time. The particle size distribution parameter is dynamically selected according to the current operating mode label.

[0033] Based on the trends of secondary voltage, secondary current, dust removal efficiency, and estimated outlet concentration, a mechanism prediction vector is obtained, which represents the baseline state that the system should exhibit under ideal physical laws.

[0034] A gated recurrent unit network with an attention mechanism is constructed to learn the residuals between the response of a real electrostatic precipitator system and the mechanism prediction vector, thereby obtaining the residual prediction vector. Specifically, to compensate for unmodeled dynamics and complex coupling effects in the actual system, a gated recurrent unit network with a dual attention mechanism is constructed as the residual learner. The first layer of this network's attention mechanism is applied to the time dimension, capturing the most critical time segments in the historical state sequence for the current residual prediction; the second layer's attention mechanism is applied to the feature dimension, dynamically adjusting the weights of each input feature channel based on the modal label of the current operating condition. The network input is a preprocessed enhanced operating condition state vector, and the training objective is the difference between the real sensor readings and the aforementioned mechanism prediction vector in each dimension. During training, the network learns the complex mapping relationship between system state, control actions, and physical model residuals, outputting the residual prediction vector corresponding to the mechanism prediction vector dimension.

[0035] The mechanistic prediction vector and the residual prediction vector are corrected and fused to obtain a complete predicted state vector, thus completing the construction of the hybrid prediction model. Specifically, within each simulation time step, the mechanistic prediction vector output by the mechanistic calculation module and the residual prediction vector output by the residual learning network are added element-wise. The concentration component in the summation result is non-negatively processed, and the electrical component is reasonably limited, thereby generating a complete predicted state vector for the system at the next time step. This hybrid architecture allows for continuous correction of model errors through data-driven methods while maintaining physical constraints. Furthermore, the parameters of the residual learning network can be periodically fine-tuned as online running data accumulates, enabling the entire hybrid prediction model to have adaptive evolution capabilities.

[0036] The constructed dust removal efficiency reward, system energy consumption penalty, and spark risk penalty are weighted and summed to obtain a multi-objective reward function. Specifically, in the design of the multi-objective reward function, the dust removal efficiency reward does not directly use the precise efficiency value that is difficult to obtain in real time, but constructs a proxy index based on the effective utilization rate of corona current and the changing trend of inlet and outlet concentrations. The system energy consumption penalty measures both the power consumption of the high-voltage power supply main circuit and the power consumption of auxiliary systems such as rapping after conversion. The spark risk penalty adopts a piecewise nonlinear function design. When the real-time spark rate is lower than the safety threshold, a linear penalty is applied. Once it approaches or exceeds the threshold, an exponentially increasing severe penalty is triggered. The weighting coefficients of the three rewards are not fixed. The weights of the efficiency and energy consumption items can be dynamically adjusted within a preset range through external instructions to achieve different optimization orientations.

[0037] The policy neural network with parameters to be optimized is used as the initial agent. Specifically, the initial agent adopts a policy neural network with a multi-layer fully connected structure. Its input dimension matches the enhanced working condition state vector, and its output dimension corresponds to the adjustable power supply control parameters. The network weights adopt an orthogonal initialization method and inject an appropriate amount of random noise to ensure the initial exploration capability.

[0038] Based on the proximal policy optimization algorithm, a reinforcement learning training framework is constructed that incorporates a safety projection layer and a course learning strategy. Specifically, the constructed reinforcement learning training framework uses the proximal policy optimization algorithm as the main body, and embeds a safety projection layer after its action output layer. This layer strictly constrains the original output action within the dynamic safety parameter space by querying the spark risk prediction model trained by historical accident data and the expert rule base in real time. The course learning strategy is designed in three stages. The first stage trains the basic voltage regulation capability only under limited steady-state conditions. The second stage gradually introduces flue gas parameter disturbances and mode switching to train anti-interference and tracking capabilities. The third stage performs the final policy optimization in the full state space, including start-up, shutdown, and abnormal operating conditions.

[0039] Based on a hybrid prediction model, aiming to maximize the multi-objective reward function, an initial agent is trained using a reinforcement learning training framework to obtain a reinforcement learning agent. Specifically, the training process is iteratively conducted in a digital twin environment constructed by the hybrid prediction model. After collecting a certain amount of interaction data in each training round, the proportion of security violations in the data batch is first evaluated. If it exceeds a threshold, an emergency adjustment of the security projection layer parameters is triggered. When updating the policy network parameters, an adaptive learning rate mechanism is used, dynamically adjusting the update step size according to the smoothness of the recent reward curve. After each course stage is completed, the hybrid prediction model is recalibrated specifically in the state space region explored by the current policy to maintain the fidelity of the environment simulation. The training termination condition considers both the convergence stability of long-term cumulative rewards and the performance of the policy on the independent validation case set. The finally saved policy network parameters constitute a reinforcement learning agent that can be deployed online.

[0040] S3: Control the electrostatic precipitator power supply according to the optimized power control parameter combination. In this embodiment, when controlling the high-voltage power supply of the electrostatic precipitator based on the optimized power control parameter combination, the abstract parameters output by the intelligent agent must first be mapped into specific executable instructions. This process is not a simple transmission of set values, but is completed through a hierarchical instruction interpreter: the first-layer parsing module dynamically interprets the standardized action vector output by the reinforcement learning intelligent agent into control benchmarks with clear physical meaning based on the nameplate parameters and actual operating limits of the current power supply equipment, including the secondary voltage target value, the repetition frequency and duty cycle of pulse power supply, and the conduction cycle ratio of intermittent power supply; the second-layer safety verification module couples these interpreted benchmark values ​​with the boundary conditions such as the real-time collected bus voltage and insulation temperature to ensure that they are within the dynamic feasible domain jointly defined by the equipment safety operation procedures and the real-time risk assessment model. If they exceed the limit, they are scaled proportionally according to preset rules or switched to backup conservative parameters.

[0041] Subsequently, the verified control commands are sent to a digital controller equipped with model prediction capabilities. This controller has a built-in transfer function model of the controlled power supply and performs multi-step forward simulation predictions before executing new commands, focusing on assessing the overcurrent risk or voltage flicker that may be caused by sudden command changes. Based on the prediction results, the controller automatically generates a smooth command transition trajectory. For example, when the target voltage needs to be significantly increased, it uses a buck-boost curve with an adaptive ramp rate instead of a step change. The ramp rate is dynamically adjusted based on the current estimated value of the equivalent capacitance of the electric field, thereby suppressing inrush current.

[0042] At the execution level, control commands are sent to the programmable logic controllers and trigger boards of each power supply unit via a high-speed communication bus. The system employs a timestamp synchronization mechanism to ensure that multiple power supply zones can coordinate their actions with millisecond-level precision, avoiding flue gas disturbances caused by control asynchrony. For complex modes such as pulse or intermittent power supply, the controller pre-calculates and sends complete waveform timing diagrams to the trigger units, which are then executed directly by the local hardware to ensure timing accuracy.

[0043] Simultaneously, the system initiates a parallel closed-loop monitoring and adaptive adjustment thread. This thread compares the expected effects of optimized parameters with actual feedback at millisecond intervals, performing consistency verification by calculating the instantaneous gradients of key performance indicators such as corona power density and spark rate trends in real time, and comparing them with the gradients predicted by the reinforcement learning agent. If a persistent deviation between the actual response and the expected response is found to exceed a threshold, an online fine-tuning mechanism is triggered: on the one hand, the deviation data is packaged, marked, and sent back to the cloud training platform for model iteration; on the other hand, the local controller initiates parameter compensation based on the built-in expert rule base. For example, when insufficient electric field ionization is detected, the voltage reference value is temporarily and slightly increased within a safe range, and this compensation amount is recorded as feedback input for subsequent agent decisions. The entire process must ensure that the key electrical parameters of the power supply output, ripple coefficient and flicker index, always comply with power quality specifications. Ultimately, through this control process that integrates intelligent decision-making, safety verification, smooth execution, and online adaptation, the refined and autonomous operation of the electrostatic precipitator power supply is achieved.

[0044] Example 2 The present invention also provides a machine learning-based electrostatic precipitator power supply parameter optimization system for implementing the method in Embodiment 1, comprising: The data processing module is used to collect and preprocess the operating status data of the electrostatic precipitator system in real time; The parameter optimization module is used to input the pre-processed operating condition data into a pre-trained reinforcement learning agent to obtain an optimized combination of power control parameters; the reinforcement learning agent is trained with the aim of maximizing a preset multi-objective reward function. The control module is used to control the power supply of the electrostatic precipitator based on the optimized combination of power control parameters.

[0045] A further embodiment of the implementation includes a data processing module comprising: The data alignment unit is used to build a time-series data fusion layer based on the edge computing gateway. It integrates the dynamic time warping algorithm and the adaptive Kalman filter to uniformly align the operating condition data to a time stamp sequence with equal intervals and mark the data confidence level. The modality recognition unit is used to perform modality recognition on uniformly aligned working condition data using an unsupervised clustering algorithm to obtain modality labels; The derived feature calculation unit is used to calculate derived features of power supply performance indicators based on modal labels. The feature fusion unit is used to utilize a learnable feature attention layer to fuse uniformly aligned working condition data, modal labels, and derived features to obtain a working condition vector, thus completing the preprocessing of the working condition data.

[0046] A further implementation wherein the derived feature calculation unit includes: The health status vector calculation subunit is used to calculate the sliding window statistical characteristics and trend characteristics of power performance indicators based on the operating range indicated by the modal label, and obtain the health status vector. The virtual feature vector calculation subunit is used to call the corresponding mechanism-data fusion model based on the modal label to estimate the equivalent resistivity of the dust layer and the corona power density distribution between the plate lines, and obtain the virtual feature vector. The derived feature acquisition subunit is used to obtain derived features based on the health state vector and the virtual feature vector.

[0047] Example 3 Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the above embodiments.

[0048] Figure 2 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0049] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0050] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0051] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0052] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB (Universal Serial Bus), network cable, etc.) or wireless means (such as mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0053] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0054] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0055] The system described in the above embodiments is used to implement the corresponding method in any of the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0056] Example 4 Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to perform the methods described in any of the above embodiments.

[0057] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0058] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0059] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0060] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuitry) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0061] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0062] Therefore, the units of the various examples described in the embodiments of this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0063] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for optimizing the power supply parameters of an electrostatic precipitator based on machine learning, characterized in that, include: Real-time acquisition and preprocessing of operating status data of the electrostatic precipitator system; The preprocessed operating condition data is input into a pre-trained reinforcement learning agent to obtain an optimized combination of power control parameters; wherein the reinforcement learning agent is trained with the aim of maximizing a preset multi-objective reward function. The power supply for the electrostatic precipitator is controlled according to the optimized combination of power control parameters.

2. The method according to claim 1, characterized in that, The method for preprocessing the operating condition data includes: A time-series data fusion layer based on an edge computing gateway is constructed, which integrates dynamic time warping algorithm and adaptive Kalman filtering to uniformly align the operating condition data to an equally spaced timestamp sequence and mark the data confidence level. An unsupervised clustering algorithm is used to perform modal identification on the uniformly aligned working condition data to obtain modal labels; Based on the modal labels, the derived features of the power supply performance indicators are calculated; By utilizing a learnable feature attention layer, the uniformly aligned working condition state data, the modality labels, and the derived features are fused together to obtain the working condition state vector, thus completing the preprocessing of the working condition state data.

3. The method according to claim 2, characterized in that, The method for calculating the derived features includes: Based on the operating range indicated by the modal label, the sliding window statistical characteristics and trend characteristics of the power performance indicators are calculated to obtain the health status vector; Based on the modal labels, the corresponding mechanism-data fusion model is invoked to estimate the equivalent resistivity of the dust layer and the corona power density distribution between the plate lines, thereby obtaining a virtual feature vector. The derived features are obtained based on the health status vector and the virtual feature vector.

4. The method according to claim 2, characterized in that, Methods for constructing reinforcement learning agents include: Based on the historical operating status data of the electrostatic precipitator system and the corresponding power control parameters, a hybrid prediction model integrating prior knowledge and data-driven approaches is constructed. The constructed dust removal efficiency reward, system energy consumption penalty, and spark risk penalty are weighted and summed to obtain a multi-objective reward function; The policy neural network with parameters to be optimized is used as the initial agent; Based on the proximal policy optimization algorithm, a reinforcement learning training framework is constructed that incorporates a secure projection layer and a course learning strategy. Based on the hybrid prediction model, with the goal of maximizing the multi-objective reward function, the initial agent is trained using a reinforcement learning training framework to obtain the reinforcement learning agent.

5. The method according to claim 4, characterized in that, Methods for constructing hybrid prediction models that integrate prior knowledge and data-driven approaches include: The electric field of the electrostatic precipitator is simplified into a nonlinear equivalent circuit. Based on historical operating condition data and corresponding power supply control parameters, the predicted secondary voltage and secondary current are calculated. Based on the flue gas flow rate, concentration, and predicted secondary current in the historical operating condition data, the trend of dust removal efficiency change and the estimated value of outlet concentration are estimated. Based on the changing trends of secondary voltage, secondary current, dust removal efficiency, and estimated outlet concentration, a mechanism prediction vector is obtained. A gated recurrent unit network with an attention mechanism is constructed to learn the residual between the response of the real electrostatic precipitator system and the mechanism prediction vector, and obtain the residual prediction vector. The mechanism prediction vector and the residual prediction vector are corrected and fused to obtain a complete prediction state vector, thus completing the construction of the hybrid prediction model.

6. A machine learning-based electrostatic precipitator power supply parameter optimization system, used to implement the method described in any one of claims 1-5, characterized in that, include: The data processing module is used to collect and preprocess the operating status data of the electrostatic precipitator system in real time; The parameter optimization module is used to input the preprocessed operating condition data into a pre-trained reinforcement learning agent to obtain an optimized combination of power control parameters; wherein the reinforcement learning agent is trained with the aim of maximizing a preset multi-objective reward function. The control module is used to control the electrostatic precipitator power supply according to the optimized power control parameter combination.

7. The system according to claim 6, characterized in that, The data processing module includes: The data alignment unit is used to construct a time-series data fusion layer based on an edge computing gateway. It integrates dynamic time warping algorithm and adaptive Kalman filtering to uniformly align the working condition data to an equally spaced timestamp sequence and marks the data confidence level. The modality recognition unit is used to perform modality recognition on the uniformly aligned working condition data using an unsupervised clustering algorithm to obtain modality labels; A derived feature calculation unit is used to calculate derived features of the power performance index based on the modal label; The feature fusion unit is used to utilize a learnable feature attention layer to fuse the uniformly aligned working condition state data, the modal labels, and the derived features to obtain the working condition state vector and complete the preprocessing of the working condition state data.

8. The system according to claim 7, characterized in that, The derived feature calculation unit includes: The health status vector calculation subunit is used to calculate the sliding window statistical characteristics and trend characteristics of the power performance indicators based on the operating range indicated by the modal label, and obtain the health status vector. The virtual feature vector calculation subunit is used to call the corresponding mechanism-data fusion model based on the modal label to estimate the equivalent resistivity of the dust layer and the corona power density distribution between the plate lines, and obtain the virtual feature vector. The derived feature acquisition subunit is used to obtain the derived features based on the health state vector and the virtual feature vector.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method as described in any one of claims 1 to 5.

Images