AI intelligent amplitude modulation method for frequency conversion power supply of electric dust removal under flue gas disturbance

Abstract

This invention relates to the field of electrostatic precipitator (ESP) control technology and discloses an AI-based intelligent amplitude modulation method for ESP variable frequency power supplies under flue gas disturbance. The method includes: acquiring thermal and electrical feedback data; executing a dynamic time warping algorithm to align the real-time thermal variable sequence with a historical thermal disturbance template, calculating the pure time delay, and establishing a feedforward control time window; injecting high-frequency probe pulses within the feedforward control time window to extract the voltage and current phase difference, calculating the filtered electric field load deviation index and its rate of change; generating a joint feature vector, selecting a Gaussian process regression model based on the aforementioned rate of change, and outputting the predicted value and prediction variance of the spark discharge critical voltage; and controlling the power conversion module through multi-mode control. This invention predicts the electric field state through the feedforward time window and dynamically increases the main switching frequency to increase the series AC impedance by combining the prediction variance, thereby limiting the transient discharge current and ensuring the stable operation of the variable frequency power supply under operating disturbances.

Description

Technical Field

[0001] This invention relates to the field of electrostatic precipitator control technology, specifically to an AI-based intelligent amplitude modulation method for electrostatic precipitator variable frequency power supplies under flue gas disturbance. Background Technology

[0002] During the operation of coal-fired power generating units, boiler load regulation or soot blowing system operation can alter the flue gas state, creating thermal disturbances that are transmitted to the downstream electrostatic precipitator (ESP). The variable frequency drive (VFD), as the core power supply for the ESP, needs to adjust its high-voltage DC output according to the flue gas conditions. Most existing VFD amplitude modulation strategies are based on closed-loop feedback of the voltage and current on the secondary side of the electric field, representing a reactive adjustment mechanism dependent on sudden changes in electrical quantities. When thermal disturbances cause an increase in dust concentration within the electric field, the control system can only detect the abnormal feedback state and execute amplitude modulation after a physical breakdown and spark discharge occurs. This reactive adjustment logic makes it difficult to predict the evolution of the electric field load state in advance, and the accompanying spark discharges and output voltage fluctuations can hinder the overall dust removal efficiency.

[0003] In existing schemes attempting to introduce feedforward control, static delay parameters fixed within the controller are often used to match the flue gas transmission process. However, in actual fluid transport environments, the pure time delay of thermal disturbances from their origin to the electrostatic precipitator's electric field dynamically changes with unit load and flue gas system operating conditions. Using fixed delay parameters to trigger feedforward regulation causes a time misalignment between the controller's action and the actual arrival time of the thermal disturbance in the electric field, making it difficult to achieve the expected control. Existing regulation methods primarily rely on limiting the duty cycle to address deteriorating electric field conditions, lacking designs for proactive intervention in the discharge precursor stage combined with the power module's hardware topology characteristics. Furthermore, they cannot utilize the AC impedance characteristics of the underlying circuitry to limit primary-side current and electrical stress. Summary of the Invention

[0004] The technical problem addressed by this invention is that existing electrostatic precipitator (ESP) frequency converters, when faced with flue gas disturbances such as boiler load regulation and sootblower operation, primarily rely on a single closed-loop feedback of electrical quantities for control strategies, failing to anticipate drastic changes in the load state within the electric field. Because the transmission of thermal disturbances from the generating end to the ESP electric field involves a complex and dynamic process with pure time delay, traditional adjustment methods based on fixed delay parameters or fixed thresholds suffer from control response lag. During periods of rapid increase in flue gas dust, this can easily trigger frequent spark discharges, leading to significant fluctuations in the electric field output voltage, thereby reducing dust removal efficiency and affecting the operational stability of the power conversion module.

[0005] To address the above problems, the present invention provides the following technical solution: This invention provides an AI-based intelligent amplitude modulation method for electrostatic precipitator inverter power supplies under flue gas disturbance, comprising: Obtain the thermal variable vector and thermal data acquisition timestamp, current main switch frequency, current duty cycle and electrical feedback data; The dynamic time warping algorithm is executed to align the real-time thermal variable sequence composed of thermal variable vectors with the time domain of the historical thermal disturbance template, calculate the pure time delay of the process, and establish a feedforward control time window. Inject high-frequency detection pulses within the feedforward control time window, extract the voltage and current phase difference, and calculate the filtered electric field load deviation index and the rate of change of the electric field load deviation index. The thermal variable vector, the filtered electric field load deviation index, the rate of change of the electric field load deviation index, the current main switch frequency, the current duty cycle, and the electrical feedback data are combined to form a joint feature vector, which is then standardized to generate the model input vector. Based on the rate of change of the electric field load deviation index, a Gaussian process regression model is selected to output the predicted value of the spark discharge critical voltage and the prediction variance. Based on the predicted value and variance of the spark discharge critical voltage, the power conversion module is controlled through voltage amplitude adjustment mode, frequency boosting and impedance enhancement mode, and frequency recovery mode. This scheme overcomes the hysteresis effect of operating condition disturbance transmission by using feedforward thermal data correlation and high-frequency impedance detection, combined with the prediction uncertainty output by the Gaussian process regression model to perform multi-mode control.

[0006] Furthermore, the acquisition of the thermal variable vector and the thermal data acquisition timestamp includes: Receive thermal operation commands, state variables, and thermal data acquisition timestamps from the distributed control system; According to the preset variable order, the thermal operation instructions and status variables corresponding to the same thermal data acquisition timestamp are arranged sequentially to generate a thermal variable vector; The thermal variable vector is associated with the thermal data acquisition timestamp and written into static random access memory. The above steps establish a unified time base, ensuring data synchronization between the lower-level electrical control unit and the upper-level distributed control system.

[0007] Furthermore, the execution of the dynamic time warping algorithm aligns the real-time thermal variable sequence composed of thermal variable vectors with the historical thermal disturbance template in the time domain, and calculates the pure time delay of the process, including: Range standardization was performed on the real-time thermal variable series and each historical thermal perturbation template. Calculate the Euclidean distance between the real-time thermal variable sequence and each data point in the historical thermal disturbance template, and recursively establish the cumulative distance matrix; The normalized path is obtained by tracing back from the end point of the cumulative distance matrix to the starting point, and the historical thermal disturbance template with the smallest normalized dynamic time normalized distance and lower than the preset template matching threshold is selected. The time scaling factor is calculated based on the start and end indices of the matching segments in the regularized path. This time scaling factor is then multiplied by the baseline lag time recorded in the selected historical thermal disturbance template, and upper and lower limits are applied to obtain the pure process lag time. The time scaling factor is introduced to correct the nonlinear time stretching or compression deviation caused by disturbances of different amplitudes on the thermal state propagation time.

[0008] Furthermore, establishing the feedforward control time window includes: The normalized Euclidean distance is calculated using the standardized difference of adjacent thermal variable vectors; When the normalized Euclidean distance reaches a preset value for multiple consecutive refresh cycles, the first refresh time that reaches the preset value is determined as the start time of the thermal disturbance. The lower limit of the time is obtained by subtracting the preset control response margin time from the pure delay time of the process, and the upper limit of the time is obtained by taking the pure delay time of the process. Using the onset time of the thermal disturbance as a benchmark, and combining the lower and upper time limits, a feedforward control time window is established. The control action range is then established by combining the onset time with the calculated time limit value to prevent control intervention from being too early or too late.

[0009] Furthermore, the step of injecting high-frequency detection pulses within the feedforward control time window to extract the voltage-current phase difference includes: During the high-frequency probe pulse injection, the probe window is activated for high-speed continuous sampling to acquire primary side high-frequency voltage feedback data and primary side high-frequency current feedback data. The sampled data is subjected to digital bandpass filtering with a center frequency equal to the repetition frequency of the probe pulse, and then single-frequency discrete Fourier transform is performed to obtain the complex components of the primary side voltage and the complex components of the primary side current. The phase angle difference is obtained by complex phase angle calculation, and the steady-state reference phase obtained based on three-dimensional linear interpolation is subtracted to obtain the actual voltage and current phase difference. High-frequency detection signals are separated and power frequency interference is filtered out through bandpass filtering and frequency domain calculation to obtain the phase shift characteristics of the equivalent impedance of the electric field.

[0010] Furthermore, the calculation of the filtered electric field load deviation index and the rate of change of the electric field load deviation index includes: The phase deviation term is calculated based on the ratio of the voltage-current phase difference to the steady-state reference phase and the effective value of the primary side high-frequency current. The deviation from the steady-state reference complex impedance is calculated based on the ratio of the complex components of the primary voltage to the complex components of the primary current, and then converted into an impedance deviation term. The electric field load deviation index is obtained by combining the phase deviation term and the impedance deviation term in a weighted manner. The electric field load deviation index is filtered to obtain a filtered electric field load deviation index, and the rate of change of the electric field load deviation index is calculated using the difference between adjacent prediction calculation periods. The deviation index is calculated by combining both phase and impedance characteristics to quantify the degree of load characteristic deviation caused by sudden changes in dust concentration and resistivity.

[0011] Furthermore, the selection of the Gaussian process regression model based on the exponential rate of change of the electric field load includes: When the magnitude of the electric field load deviating from the exponential rate of change reaches the preset rate of change threshold, the Gaussian process regression model selects the Matrn 3 / 2 kernel Gaussian process regression model. When the magnitude of the electric field load deviating from the exponential rate of change falls back to within the preset rate of change exit threshold and continues for a preset period, the Gaussian process regression model selects the squared exponential kernel Gaussian process regression model. When the magnitude of the electric field load deviating from the exponential rate of change is between the rate of change exit threshold and the rate of change entry threshold, the Gaussian process regression model selected in the previous prediction calculation cycle is maintained. The kernel function is switched for different stages of drastic changes and smooth recovery of the operating conditions to address the problem that a single model cannot simultaneously meet the requirements of local abrupt response and global smooth prediction.

[0012] Furthermore, the predicted value and prediction variance of the output spark discharge critical voltage include: The mean and standard deviation saved offline are used to perform standardization on the joint feature vector to generate the model input vector; Calculate the covariance between the model input vector and the representative training sample set; By combining the Cholesky decomposition results saved during the training phase and the preset observation noise variance, the predicted value and prediction variance of the spark discharge critical voltage are obtained through forward substitution and back substitution operations. The matrix decomposition preprocessing results are used to reduce the computational cost of inverting the covariance matrix during the operation phase, meeting the real-time computational requirements of the underlying control module and synchronously outputting the prediction uncertainty.

[0013] Furthermore, the power conversion module is controlled via voltage amplitude adjustment mode, frequency boosting and impedance enhancement mode, and frequency recovery mode, wherein the voltage amplitude adjustment mode includes: The target given voltage is obtained by subtracting the preset safety margin voltage from the predicted value of the spark discharge critical voltage, and then subtracting the product of the preset standard deviation margin coefficient and the predicted standard deviation. The predicted standard deviation is the square root of the predicted variance. The deviation between the target given voltage and the feedback value of the secondary high voltage DC voltage is input into the control loop, and the duty cycle command is output. When the prediction variance reaches the preset threshold for frequency boosting and impedance amplification, the system is triggered to switch to frequency boosting and impedance amplification mode. An adaptive voltage drop mechanism based on the prediction standard deviation is introduced into the voltage setpoint calculation to dynamically adjust the safety margin according to the model prediction confidence level.

[0014] Furthermore, the frequency boosting and impedance amplification mode includes: Keep the target given voltage from rising further, and limit the upper limit of the duty cycle to the preset ratio of the current duty cycle before entering the frequency boosting and impedance boosting mode; Based on the predicted variance, linear interpolation is performed between the up-frequency and up-resistance entry threshold and the preset upper limit of the frequency mapping variance to determine the target main switch frequency; Increase the current main switch frequency by a fixed step size until the target main switch frequency is reached, thereby increasing the AC impedance of the series resonant network. Once the prediction variance falls below the preset frequency increase and impedance reduction exit threshold and meets the preset recovery conditions, the system switches to frequency recovery mode. In frequency recovery mode, the current main switch frequency is gradually reduced to the preset reference main switch frequency. When the prediction uncertainty is high, the switching frequency is increased to increase the physical impedance of the internal resonant network, limiting the transient discharge current at the high-voltage end.

[0015] Furthermore, the method also includes: When communication interruption, model data verification failure, modulus sampling abnormality, or memory read abnormality occurs, the derating operation mode is triggered. In derating operation mode, the current duty cycle is fixed to a preset derating duty cycle, the current main switch frequency is fixed to a preset reference main switch frequency, and hardware protection is enabled. The execution priority of hardware protection is higher than that of control logic based on regression prediction results. A boundary safety judgment mechanism is set for abnormal states to prevent abnormal input data from causing the control module output to exceed limits, ensuring the basic operational safety of the system.

[0016] This invention provides an AI-based intelligent amplitude modulation method for electrostatic precipitator frequency converter power supplies under flue gas disturbance. It offers the following advantages: 1. This invention obtains thermal variable vectors and executes a dynamic time warping algorithm to align the real-time thermal variable sequence with the historical thermal disturbance template in the time domain to calculate the pure time delay of the process and establish a feedforward control time window. This overcomes the response lag defect caused by relying on fixed delay parameters in traditional control strategies, and makes the triggering node of subsequent adjustment commands match the objective time when the thermal disturbance arrives at the electric field, thus improving the timing of feedforward adjustment of the frequency converter when facing changes in dust removal conditions.

[0017] 2. This invention extracts the phase difference between voltage and current by actively injecting high-frequency detection pulses within the feedforward control time window, and calculates the electric field load deviation index and its rate of change by combining the complex components of voltage and current. Based on this rate of change, it switches between Gaussian process regression models with different kernel functions to output the predicted value and prediction variance of the spark discharge critical voltage. Before the breakdown of conventional low-frequency electrical quantities, the physical evolution of dust concentration and space charge inside the electric field is quantified, providing quantitative data feature input for subsequent anti-discharge control actions.

[0018] 3. This invention combines the prediction variance with the resonant physical characteristics of the power conversion module. When the prediction variance reaches the threshold for frequency increase and impedance increase, the main switching frequency is increased by a fixed step size. The increase in inductive reactance caused by the main switching frequency crossing the inherent resonant frequency is used to increase the AC impedance of the series resonant network. The rise rate of the primary current is limited by the hardware impedance at the bottom layer of the circuit. In the precursor stage of spark discharge, the short-circuit current is suppressed to maintain the continuity of equipment operation. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a flowchart of the multi-source data synchronous acquisition process of the present invention; Figure 4 This is a flowchart illustrating the time-domain alignment and time windowing of thermal and electrical feedback data in this invention. Figure 5 This is an online schematic diagram of the high-frequency detection pulse injection and electric field load characteristics of the present invention; Figure 6 This is a flowchart of the Gaussian process regression prediction and kernel function of the present invention; Figure 7 This is a schematic diagram of the control mode switching and frequency boosting / resistance amplification principle of the present invention; Figure 8 This is a timing diagram of the prediction variance and main switch frequency of the present invention; Figure 9 This is a comparison chart showing the improvement ratio of operating indicators under different control methods of the present invention. Detailed Implementation

[0020] The technical solutions in 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] See attached document Figure 1The electrostatic precipitator variable frequency power supply AI intelligent amplitude modulation system under flue gas disturbance includes a distributed control system, a variable frequency power supply controller, a signal feedback module, and a power conversion module.

[0022] The distributed control system collects thermal operation commands and status variables from the boiler side and the flue gas system, and combines the various thermal data at the same acquisition time into a thermal variable vector. This thermal variable vector, along with the thermal data acquisition timestamp, is transmitted to the variable frequency power controller via Modbus-TCP industrial Ethernet. The thermal data refresh cycle is 1 second.

[0023] The distributed control system and the variable frequency power controller synchronize their local clocks via the IEEE 1588 precision time protocol. The thermal data acquisition timestamp is generated based on the synchronized local clock and is used to mark the acquisition time of thermal data within the distributed control system.

[0024] The variable frequency power supply controller includes a digital signal processor, a field-programmable gate array, a static random access memory, and a non-volatile memory.

[0025] Digital signal processors (DSPs) are used to perform thermal and electrical feedback data processing, dynamic time warping calculations, Gaussian process regression calculations, control mode determination, and closed-loop regulation. Field-programmable gate arrays (FPGAs) are used to perform analog-to-digital sampling triggering, pulse width modulation (PWM), pulse frequency modulation (PWM), dead-time control, and hardware protection.

[0026] The non-volatile memory stores historical thermal disturbance templates, steady-state reference phase tables, steady-state reference impedance tables, control thresholds, square exponential kernel Gaussian process regression model parameters, and Matrn 3 / 2 kernel Gaussian process regression model parameters.

[0027] The signal feedback module includes a primary-side high-frequency feedback channel and a secondary-side high-voltage DC feedback channel.

[0028] The primary-side high-frequency feedback channel connects the voltage sampling component and current transformer on the primary side of the high-frequency transformer, and is used for isolation, conditioning, and anti-aliasing filtering of the primary-side voltage and current signals. The processed analog signal is input to the high-speed analog-to-digital conversion channel group of the frequency converter controller.

[0029] The secondary-side high-voltage DC feedback channel is connected to the output terminal of the high-voltage output rectifier component. Through the high-voltage voltage divider component and the DC current sampling component, the secondary-side high-voltage DC voltage feedback signal and the secondary-side DC current feedback signal are obtained.

[0030] Digital bandpass filtering, single-frequency discrete Fourier transform calculation, phase calculation, and impedance characteristic calculation are all performed by a digital signal processor.

[0031] The power conversion module includes an input rectifier assembly, a DC bus capacitor, a full-bridge inverter, a resonant inductor, a resonant capacitor, a high-frequency transformer, and a high-voltage output rectifier assembly.

[0032] The input rectifier converts the AC input voltage to the DC bus voltage. The full-bridge inverter converts the DC bus voltage to a high-frequency AC voltage. The high-frequency transformer boosts the high-frequency AC voltage. The high-voltage output rectifier converts the boosted high-frequency AC voltage into a high-voltage DC voltage applied to the electric field of the electrostatic precipitator.

[0033] The resonant inductor and resonant capacitor form a series resonant network. The reference main switching frequency is set above the inherent resonant frequency of the series resonant network.

[0034] The variable frequency power controller changes the amplitude of the secondary side high voltage DC output by adjusting the duty cycle of the drive pulse of the full-bridge inverter, and increases the AC impedance of the series resonant network by increasing the main switching frequency of the full-bridge inverter.

[0035] The frequency converter power controller is equipped with overcurrent protection, overvoltage protection, spark discharge detection, bridge arm shoot-through protection, high-frequency transformer temperature protection, switching device temperature protection, maximum duty cycle limit, and maximum main switch frequency limit.

[0036] Spark discharge detection uses a combined criterion of the drop in secondary high-voltage DC voltage and the rise in secondary DC current. Within a 1ms detection time window, when the drop in secondary high-voltage DC voltage reaches 10% of the voltage value before detection and the rise in secondary DC current reaches 20% of the current value before detection, the frequency converter power supply controller outputs a spark discharge detection signal.

[0037] Hardware protection acts directly on the full-bridge inverter drive signal, and its execution priority is higher than that of control logic based on regression prediction results.

[0038] See attached document Figure 2 The AI ​​intelligent amplitude modulation method for electrostatic precipitator inverter power supply under flue gas disturbance determines the voltage amplitude adjustment mode and frequency increase and resistance enhancement mode of the inverter power supply based on the transmission delay of thermal disturbance and electric field load state, including the following steps: S100, the variable frequency power controller obtains the thermal variable vector and thermal data acquisition timestamp sent by the distributed control system, and at the same time obtains the primary side high-frequency voltage feedback data, primary side high-frequency current feedback data, secondary side high-voltage DC voltage feedback data and secondary side DC current feedback data. S200, the variable frequency power controller executes a dynamic time warping algorithm to align the real-time thermal variable sequence with the historical thermal disturbance template in the time domain, calculate the pure time delay of the process, and establish a feedforward control time window; S300: After the control clock enters the feedforward control time window, the frequency converter power controller injects high-frequency detection pulses into the full-bridge inverter drive sequence, extracts the voltage and current phase difference at the repetition frequency of the detection pulses, and calculates the electric field load deviation index and its rate of change. The S400 variable frequency power controller uses the thermal variable vector, the filtered electric field load deviation index, the electric field load deviation index change rate, the current main switch frequency, the current duty cycle, the secondary side high voltage DC voltage feedback value, and the secondary side DC current feedback value to form the model input vector. Based on the electric field load deviation index change rate, it selects a Gaussian process regression model and outputs the spark discharge critical voltage prediction value and prediction variance. The S500 variable frequency power controller switches control modes based on predicted variance and controls the power conversion module through voltage amplitude adjustment mode, frequency boosting and resistance increase mode, and frequency recovery mode.

[0039] See attached document Figure 3 Step S100 specifically includes the following sub-steps to achieve data acquisition, time alignment, and vectorization processing at different time scales: The S110 variable frequency power controller receives thermal operation commands, status variables, and thermal data acquisition timestamps from the distributed control system via Modbus-TCP industrial Ethernet. The thermal operation commands and status variables include boiler system coal feed rate step commands, induced draft fan damper opening, sootblower start-up status flag, and secondary damper opening. The thermal data acquisition timestamps indicate the actual acquisition time of each thermal data point within the distributed control system.

[0040] S120: The variable frequency power controller arranges the coal feed step command, induced draft fan damper opening, soot blower start status flag, and secondary damper opening corresponding to the same thermal data acquisition timestamp in a preset variable order, generating a thermal variable vector. The variable frequency power controller associates the thermal variable vector with the corresponding thermal data acquisition timestamp and writes it into the static random access memory for subsequent dynamic time warping calculations, Gaussian process regression calculations, and control mode determination.

[0041] The S130, a field-programmable gate array, controls the high-speed analog-to-digital converter channel group to perform conventional feedback sampling and high-speed continuous sampling of the probe window. Conventional feedback sampling is triggered by the peak value of the pulse width modulation counter and is performed once per main switching cycle to obtain the secondary side high-voltage DC feedback value and the secondary side DC current feedback value.

[0042] High-speed continuous sampling of the probe window is initiated during the injection of high-frequency probe pulses. The field-programmable gate array (FPGA) uses an independent high-speed sampling clock to continuously trigger the analog-to-digital converter (ADC) channel corresponding to the primary side high-frequency feedback channel. The sampling frequency is 16 times the probe pulse repetition frequency.

[0043] Each high-frequency probe pulse group consists of 8 probe cycles. Within each high-frequency probe pulse group, the high-speed analog-to-digital converter (ADC) channel group acquires 128 primary-side high-frequency voltage sampling points and 128 primary-side high-frequency current sampling points. After the high-frequency probe pulse group ends, the high-speed ADC channel group resumes regular feedback sampling.

[0044] S140, the variable frequency power controller establishes a circular queue buffer in the static random access memory. The thermal variable vector, primary side high-frequency voltage feedback data, primary side high-frequency current feedback data, secondary side high-voltage DC voltage feedback data, and secondary side DC current feedback data are all timestamped using a unified clock reference.

[0045] The thermal variable vector is written to the circular queue buffer according to the thermal data acquisition timestamp. The primary side high-frequency feedback data and the secondary side high-voltage DC feedback data are written to the circular queue buffer according to the local hardware sampling time.

[0046] The variable frequency power supply controller employs a zero-order hold mode, mapping low-frequency updated thermal variables to the electrical feedback characteristic sampling time. Primary-side high-frequency feedback data extracts phase and impedance characteristics based on high-frequency detection pulse groups. Secondary-side high-voltage DC feedback data extracts voltage and current values ​​according to the prediction calculation cycle. All data are matched using timestamps, forming data records under a unified time coordinate.

[0047] See attached document Figure 4 Step S200 specifically includes the following sub-steps to calculate the pure time delay of the thermal disturbance being transmitted to the electric field of the electrostatic precipitator, and to determine the feedforward control time window: S210, the variable frequency power controller reads multiple historical thermal disturbance templates from non-volatile memory. These templates include templates for changes in coal feed rate, induced draft fan damper opening, sootblower start-up, and secondary damper opening. Each template records the trajectory of the corresponding thermal parameters and marks the disturbance start time, the moment the electric field load deviation index first reaches the response threshold, and the basic lag time.

[0048] During the historical thermal disturbance template establishment phase, one high-frequency probe pulse group is injected in each prediction calculation cycle. The electric field load deviation index response threshold is taken as the steady-state sample mean plus three times the steady-state sample standard deviation. The basic lag time is the duration from the start of the historical disturbance to the moment when the electric field load deviation index first reaches the response threshold.

[0049] The variable frequency power controller extracts the thermal variable vector within the current sliding time window, forming a vector of length [missing information]. The real-time thermal variable sequence. The length of each historical thermal perturbation template is... , and All are not less than 2.

[0050] S220, the variable frequency power controller performs range standardization on the real-time thermal variable sequence and each historical thermal disturbance template. When the maximum and minimum values ​​of a single thermal variable are equal within a time window, the standardized value for that variable is set to 0. The variable frequency power controller calculates the Euclidean distance between each data point in the real-time thermal variable sequence and the historical thermal disturbance template, and establishes a cumulative distance matrix. The cumulative distance matrix is ​​recursively derived according to the following formula: ; In the formula, Represents the sequence of real-time thermal variables. The data points and the historical thermal disturbance template The cumulative distance between data points; This represents the Euclidean distance between two data points; The value range is 1~ ; The value range is 1~ ; Indicates the length of the real-time thermal variable sequence; This represents the length of the historical thermal disturbance template. The starting point of the cumulative distance matrix is... = The first row and first column are calculated recursively based on adjacent already calculated elements.

[0051] S230, the frequency converter backtracks from the endpoint of the cumulative distance matrix to the starting point to obtain the normalized path with the minimum cumulative distance. The frequency converter divides the endpoint value of the cumulative distance matrix by the product of the number of points in the normalized path and the dimension of the thermal variables to obtain the normalized dynamic time normalized distance, and selects the historical thermal disturbance template with the minimum normalized dynamic time normalized distance that is lower than the template matching threshold.

[0052] The variable frequency power controller calculates the time scaling factor, process pure time delay, and feedforward control time window based on the start and end indices of the matching segments of the real-time thermal variable sequence in the normalized path, as well as the start and end indices of the matching segments of the historical thermal disturbance template. ; ; ; In the formula, Indicates the time scaling factor; and These represent the starting and ending indices of the regularized path within the matching segment of the real-time thermal variable sequence, respectively. and These represent the starting and ending indices of the regularized path in the historical thermal disturbance template matching segment, respectively; Indicates the sampling period of the real-time thermal variable sequence; Indicates the sampling period of the historical thermal disturbance template; Indicates the base lag time of the selected historical thermal disturbance template record; Indicates the pure time delay of the process; This indicates the lower limit of the pure time delay of the process; This represents the upper limit of the pure time delay of the process; This indicates that the calculation results will be restricted to... ~ Within the range; Indicates the feedforward control time window; Indicates the onset time of the thermal disturbance; This indicates the control response margin time.

[0053] Thermal disturbance onset time The normalized Euclidean distance between adjacent thermal variable vectors is used for determination. The normalized Euclidean distance is calculated by taking the square root of the sum of the squares of the differences between the standardized thermal variables and dividing by... get.

[0054] When the normalized Euclidean distance reaches 0.1 for three consecutive refresh cycles, the first refresh time to reach 0.1 is determined as the start time of the thermal disturbance.

[0055] Lower limit of pure time delay in process Take 80% of the minimum historical baseline lag time as the upper limit of the pure process lag time. Take 120% of the maximum value of all historical baseline lag times.

[0056] S240, controls response margin time Take the pure time delay of the process 10%. The frequency converter controller writes the feedforward control time window into the finite state machine. When the minimum normalized dynamic time warp distance in each historical thermal disturbance template is greater than or equal to the template matching threshold, the frequency converter controller uses the median of the pure time delay of the most recent 10 effective processes to establish the feedforward control time window.

[0057] During the system commissioning phase, the median of all historical baseline delay times is written into non-volatile memory as the initial pure process delay time. When the cumulative number of effective pure process delay times is less than 10, the frequency converter uses the initial pure process delay time to establish a feedforward control time window.

[0058] See attached document Figure 5 Step S300 specifically includes the following sub-steps to extract the phase and impedance deviation characteristics of the electrostatic precipitator load at the detection frequency: The S310 variable frequency power supply controller monitors the local control clock. Once the control clock enters the feedforward control time window, the field-programmable gate array (FPGA) initiates the online detection program for electric field load characteristics. High-frequency probe pulses are injected symmetrically with positive and negative half-cycles. The probe pulse repetition frequency is fixed at 2.5 times the reference main switching frequency, with a duty cycle of 3%. Each high-frequency probe pulse group is injected continuously for 8 probe cycles.

[0059] The permissible short-time pulse frequency of the switching device is no less than three times the reference main switching frequency. The high-frequency probe pulse width is greater than the minimum reliable on-time of the switching device, and the bridge arm drive retains a set dead time. The high-frequency probe pulse is generated by an independent probe clock. After entering the frequency boosting and impedance amplification mode, the probe pulse repetition frequency remains unchanged. When the control clock is within the feedforward control time window, the field-programmable gate array injects one high-frequency probe pulse group per prediction calculation cycle.

[0060] After the control clock passes the end of the feedforward control time window, the field-programmable gate array (FPGA) continues to inject five high-frequency probe pulse groups. In voltage amplitude regulation mode, when the control clock is outside the feedforward control time window, the FPGA injects one high-frequency probe pulse group every 10 prediction calculation cycles to update the electric field load deviation index.

[0061] During the injection of the high-frequency probe pulse group, the voltage amplitude closed-loop regulator maintains the integral state of the previous prediction calculation cycle. After the high-frequency probe pulse group ends, the field-programmable gate array resumes the normal pulse width modulation sequence.

[0062] During the injection of the high-frequency probe pulse group, the S320 frequency converter power controller acquires the voltage feedback data and current feedback data of the primary side of the high-frequency transformer, performs digital bandpass filtering on the sampled data with the probe pulse repetition frequency as the center frequency, and then performs single-frequency discrete Fourier calculation.

[0063] The complex components of the primary side voltage, the complex components of the primary side current, and the phase difference between the voltage and current at the repetition frequency of the probe pulse are expressed as follows: ; ; ; In the formula, This represents the complex component of the primary side voltage at the frequency of the probe pulse repetition. This represents the complex component of the primary current at the frequency of the probe pulse repetition. This indicates the number of sampling points within the detection and analysis time window; Indicates the first One primary side voltage sample value; Indicates the first One primary side current sampling value; Indicates the repetition frequency of the probe pulse; Indicates the high-speed analog-to-digital sampling period; Represents the imaginary unit, and =-1; Represents the natural constant; Represents pi; Represents complex number phase angle operations; This means that the phase is restricted to the interval [-π, π). Indicates the steady-state reference phase; Indicates the current main switch frequency; Indicates the current duty cycle; This represents the effective value of the primary side high-frequency voltage; This represents the phase difference between the probe frequency voltage and current after reference phase compensation.

[0064] The S330 variable frequency power controller calculates the electric field load deviation index based on the phase difference between the detected frequency voltage and current after reference phase compensation, the effective value of the primary side high-frequency current, the complex components of the primary side voltage, the complex components of the primary side current, and the steady-state reference complex impedance.

[0065] The electric field load deviation index, filter value, and rate of change are expressed as follows: ; ; ; In the formula, Indicates the first Electric field load deviation index for each prediction calculation cycle; This represents the weight of the phase deviation term, which is set to 0.6. This represents the weight of the impedance deviation term, which is set to 0.4. This represents the reference phase, which is 1 rad. Indicates the rated current on the primary side; This represents the effective value of the primary side high-frequency current; To represent the lower limit of the denominator of the current, take... 5%; This indicates taking the larger value; Represents the steady-state reference complex impedance; The vertical line indicates the primary side rated voltage; the vertical line indicates the complex modulus. This represents the electric field load deviation index after filtering; Pick ; Indicates the prediction calculation period; This represents the rate of change of the electric field load deviation exponentially; Indicates the current main switch frequency; Indicates the repetition frequency of the probe pulse; Indicates the current duty cycle; This represents the effective value of the primary side high-frequency voltage; This represents the phase difference between the probe frequency voltage and current after reference phase compensation.

[0066] S340, the steady-state reference phase table and steady-state reference impedance table are established during the equipment commissioning phase. The frequency converter sets a frequency node at 1% of the reference main switching frequency within the range of the reference main switching frequency to the maximum main switching frequency; a duty cycle node at 5% within the maximum duty cycle limit range; and a voltage node at 5% of the primary side rated voltage within the range of 20% to 100% of the primary side rated voltage.

[0067] The variable frequency power controller records the steady-state reference phase and steady-state reference complex impedance at each node combination, forming a steady-state reference phase table and a steady-state reference impedance table. During online operation, the variable frequency power controller uses a fixed probe pulse repetition frequency as a reference and performs three-dimensional linear interpolation based on the current main switch frequency, current duty cycle, and primary side high-frequency voltage RMS value to obtain... and .

[0068] The variable frequency power controller writes the filtered electric field load deviation index, the electric field load deviation index change rate, the current main switch frequency, the current duty cycle, the secondary side high voltage DC voltage feedback value, and the secondary side DC current feedback value into the static random access memory.

[0069] See attached document Figure 6 Step S400 specifically includes the following sub-steps, which calculate the spark discharge critical voltage and its prediction variance based on the current thermal state and electric field load characteristics: S410, the variable frequency power controller extracts the thermal variable vector, the filtered electric field load deviation index, the electric field load deviation index change rate, the current main switch frequency, the current duty cycle, the secondary side high voltage DC voltage feedback value and the secondary side DC current feedback value, and splices them together in a fixed order to form a joint feature vector.

[0070] The variable frequency power controller uses the mean and standard deviation saved during the training phase to perform Z-score standardization on the joint feature vector to generate the model input vector. The order, mean, and standard deviation of each input variable are stored in non-volatile memory and remain unchanged during online prediction.

[0071] The S420 quadratic exponential kernel Gaussian process regression model is used for stationary changes in the electric field load when the rate of change deviating from the exponential is below the threshold. The Matrn 3 / 2 kernel Gaussian process regression model is used for transient changes in the electric field load when the rate of change deviating from the exponential reaches the threshold. Both Gaussian process regression models use the standardized joint eigenvector as the independent variable and the spark discharge critical voltage as the training label.

[0072] The joint feature vector of each training sample is taken from the prediction calculation cycle one period before the spark discharge detection signal is triggered. The training label is the peak value of the secondary side high-voltage DC voltage at the moment the spark discharge detection signal is triggered. During the training phase, training samples are grouped according to the rate of change threshold. Training samples whose magnitude of the electric field load deviation from the exponential rate of change is lower than the rate of change threshold are used to train the quadratic exponential kernel Gaussian process regression model. Training samples whose magnitude of the electric field load deviation from the exponential rate of change reaches the rate of change threshold are used to train the Matrn 3 / 2 kernel Gaussian process regression model.

[0073] The hyperparameters of the quadratic exponential kernel Gaussian process regression model include signal variance, feature length scale, and observation noise variance. The hyperparameters of the Matrn 3 / 2 kernel Gaussian process regression model also include signal variance, feature length scale, and observation noise variance. During the training phase, the conjugate gradient method is used to solve for the model hyperparameters that maximize the logarithmic marginal likelihood function.

[0074] During the equipment commissioning phase, at least 128 training samples are collected for each Gaussian process regression model. Each Gaussian process regression model uses K-means clustering to divide the training samples into 128 clusters, and selects the training samples closest to the cluster center from each cluster to form a representative training sample set. The digital signal processor performs Cholesky decomposition on the sum of the covariance matrix and the observation noise matrix of the representative training sample set, and stores the representative training sample set, training labels, kernel function hyperparameters, observation noise variance, and Cholesky decomposition results in non-volatile memory.

[0075] During the online prediction phase, the digital signal processor uses forward substitution and back substitution operations to complete the prediction calculation, without performing online inversion of the covariance matrix.

[0076] S430, Variable Frequency Power Controller Setting Change Rate Enters Threshold and rate of change exit threshold ,and Less than . Take the 95th percentile value of the magnitude of the rate of change of electric field load deviation under the stable operating condition of the offline verification. Pick 60%.

[0077] The magnitude of the electric field load deviating from the exponential rate of change reaches At that time, the variable frequency power controller selected the Matrn 3 / 2 kernel Gaussian process regression model. The magnitude of the electric field load deviating from the exponential rate of change fell back to... After five consecutive prediction calculation cycles, the variable frequency power supply controller selects a squared exponential kernel Gaussian process regression model. The magnitude of the electric field load deviating from the exponential rate of change is within... ~ When within the specified range, the variable frequency power controller maintains the Gaussian process regression model selected in the previous prediction calculation cycle.

[0078] S440, the selected Gaussian process regression model calculates the covariance between the model input vector and the representative training sample set, and outputs the predicted value and prediction variance of the spark discharge critical voltage: ; ; In the formula, This represents the predicted critical voltage value for spark discharge, and Corresponding to the secondary side high voltage DC voltage; Indicates the prediction variance; This represents the covariance vector between the current model input vector and the training sample set. This represents the covariance matrix among the training sample sets; Indicates the variance of observation noise; Represents the covariance matrix Identity matrices of the same order; This represents the column vector of training labels corresponding to the training sample set; This represents the covariance between the current model input vector and itself. Online prediction calculations are performed using the Cholesky decomposition results saved during the training phase.

[0079] See attached document Figure 7 , Figure 7 This is a control mode switching and frequency boosting / resistance increase control diagram according to an embodiment of the present invention. Step S500 specifically includes the following sub-steps to complete voltage amplitude adjustment, frequency boosting / resistance increase, and main switch frequency recovery based on the predicted variance: S510, the variable frequency power controller reads the predicted value of the spark discharge critical voltage, the prediction variance, and the threshold for frequency boosting and impedance increase. Frequency increase and impedance increase exit threshold Safety margin voltage, standard deviation margin factor, reference main switching frequency, and maximum main switching frequency. Frequency boosting and impedance increase exit threshold. Below the threshold for frequency upsampling and impedance increase The standard deviation of the prediction is taken as the square root of the variance of the prediction. A finite state machine manages the voltage amplitude adjustment mode, the frequency boosting and impedance amplification mode, and the frequency recovery mode.

[0080] The threshold for increasing frequency and impedance is determined during the offline verification phase. During the offline verification phase, the allowable prediction error is set to 5% of the rated output voltage, and the verification samples are arranged in ascending order of prediction variance.

[0081] If, among the validation samples whose prediction variance is no higher than the candidate threshold, the proportion of samples whose prediction error is no higher than the allowable prediction error reaches 95%, then the largest candidate threshold satisfying this condition is determined as the up-conversion and impedance-increasing entry threshold. The up-conversion and impedance-increasing exit threshold is taken as 60% of the up-conversion and impedance-increasing entry threshold.

[0082] After system startup, the finite state machine enters the voltage amplitude regulation mode. In the voltage amplitude regulation mode, when the prediction variance is lower than the frequency boosting and impedance increase threshold, the finite state machine maintains the voltage amplitude regulation mode; when the prediction variance reaches the frequency boosting and impedance increase threshold, the finite state machine enters the frequency boosting and impedance increase mode.

[0083] In the frequency upsampling and impedance upsampling mode, when the prediction variance is higher than the frequency upsampling and impedance upsampling exit threshold, the finite state machine maintains the frequency upsampling and impedance upsampling mode; after the prediction variance falls back to within the frequency upsampling and impedance upsampling exit threshold and continues to reach the recovery hold time, the finite state machine enters the frequency recovery mode.

[0084] In the S520 voltage amplitude regulation mode, the frequency converter power controller subtracts the safety margin voltage from the predicted value of the spark discharge critical voltage, and then subtracts the product of the standard deviation margin coefficient and the predicted standard deviation to obtain the target given voltage.

[0085] The safety margin voltage is taken as 5% of the rated output voltage, the standard deviation margin factor is taken as 2, and the minimum operating voltage is taken as 20% of the rated output voltage. The frequency converter power controller subtracts the safety margin voltage from the predicted spark discharge critical voltage, and then subtracts the product of the standard deviation margin factor and the predicted standard deviation to obtain the calculated target setpoint voltage. When the calculated target setpoint voltage is lower than the minimum operating voltage, the minimum operating voltage is used; when the calculated target setpoint voltage is higher than the rated output voltage, the rated output voltage is used; when the calculated target setpoint voltage is neither lower than the minimum operating voltage nor higher than the rated output voltage, the calculated value is used.

[0086] The variable frequency power supply controller inputs the deviation between the target setpoint voltage and the secondary side high-voltage DC voltage feedback value into the proportional-integral-derivative (PID) control loop. The PID control loop outputs a duty cycle command, and the field-programmable gate array (FPGA) generates the full-bridge inverter drive signal according to the duty cycle command.

[0087] The proportional-integral-derivative (PID) control parameters are tuned during the equipment commissioning phase based on the step response of the secondary high-voltage DC voltage and stored in non-volatile memory.

[0088] S530: Once the prediction variance reaches the threshold for frequency boosting and resistance enhancement, the finite state machine enters the frequency boosting and resistance enhancement mode. The frequency converter controller maintains the target setpoint voltage to prevent further increases and limits the upper limit of the duty cycle to 90% of the duty cycle before entering the frequency boosting and resistance enhancement mode.

[0089] The variable frequency power controller sets the target main switch frequency corresponding to the frequency increase and resistance increase threshold as the reference main switch frequency, sets the larger of the maximum prediction variance of the offline validation set and twice the frequency increase and resistance increase threshold as the upper limit of the frequency mapping variance, and sets the target main switch frequency corresponding to the upper limit of the frequency mapping variance as the maximum main switch frequency.

[0090] When the prediction variance is not lower than the frequency ramp-up threshold and not higher than the upper limit of the frequency mapping variance, the variable frequency power controller uses linear interpolation to determine the target main switching frequency. When the prediction variance is higher than the upper limit of the frequency mapping variance, the target main switching frequency is the maximum main switching frequency. The field-programmable gate array increases the current main switching frequency by a fixed step of 1% of the reference main switching frequency in each prediction calculation cycle until the current main switching frequency reaches the target main switching frequency.

[0091] The inherent resonant frequency and AC impedance of the S540 series resonant network are expressed as follows: ; ; ; In the formula, This represents the inherent resonant frequency of the series resonant network; This represents the AC impedance of a series resonant network. This represents the equivalent series resistance of the power conversion module; Indicates the current main switch frequency; Indicates the reference main switch frequency; Indicates resonant inductance; Indicates the resonant capacitance; It represents pi (π).

[0092] The reference main switching frequency is taken as 1.1 times the natural resonant frequency of the series resonant network, and the maximum main switching frequency is taken as 1.35 times the natural resonant frequency of the series resonant network. When the current main switching frequency is above the natural resonant frequency of the series resonant network, the increase in the main switching frequency increases the inductive reactance component of the series resonant network, and the AC impedance of the series resonant network increases accordingly.

[0093] In the S550's frequency boosting and impedance boosting mode, the field-programmable gate array injects one high-frequency probe pulse group per prediction calculation cycle, with each high-frequency probe pulse group consisting of eight detection cycles. The frequency converter power controller updates the electric field load deviation index, the rate of change of the electric field load deviation index, the predicted value of the spark discharge critical voltage, and the prediction variance based on each detection result.

[0094] During the high-frequency probe pulse injection period, the voltage amplitude closed-loop regulator maintains the integral state of the previous prediction calculation cycle. The finite state machine enters frequency recovery mode after five prediction calculation cycles: the prediction variance falls below the frequency increase and resistance increase exit threshold; the magnitude of the electric field load deviation from the exponential rate of change falls below the rate of change exit threshold; the secondary-side DC current feedback value is lower than the recovery current threshold; and the spark discharge detection signal remains untriggered. The recovery current threshold is set to 80% of the rated DC output current.

[0095] In frequency recovery mode, when the prediction variance reaches the frequency boosting and impedance increase threshold again, the finite state machine returns to the frequency boosting and impedance increase mode. When the prediction variance remains below the frequency boosting and impedance increase threshold, the field-programmable gate array reduces the reference main switching frequency by a fixed step of 1% per prediction calculation cycle, restoring the current main switching frequency to the reference main switching frequency.

[0096] Once the current main switching frequency reaches the reference main switching frequency, the finite state machine returns to the voltage amplitude regulation mode. In any control mode, overcurrent protection, overvoltage protection, spark discharge detection, bridge arm shoot-through protection, and temperature protection all directly limit the full-bridge inverter drive signal.

[0097] Communication interruption, model data verification failure, analog-to-digital sampling anomaly, and memory read anomaly trigger derating operation mode. In derating operation mode, the duty cycle is fixed at the preset derating duty cycle, the main switch frequency is fixed at the reference main switch frequency, and hardware protection remains enabled. The preset derating duty cycle is 50% of the maximum duty cycle limit.

[0098] In this embodiment, the prediction calculation period is 10ms, the template matching threshold is 0.15, the control response margin time is 10% of the pure delay time of the process, the probe pulse repetition frequency is 2.5 times the reference main switch frequency, the high-frequency probe pulse duty cycle is 3%, the high-speed continuous sampling frequency of the probe window is 16 times the probe pulse repetition frequency, and a single high-frequency probe pulse group includes 8 probe cycles.

[0099] The above parameters are written to non-volatile memory. During the equipment commissioning phase, the steady-state reference phase table, steady-state reference impedance table, historical thermal disturbance template, quadratic exponential kernel Gaussian process regression model, and Matrn 3 / 2 kernel Gaussian process regression model are calibrated.

[0100] Specific application examples: This embodiment is applied to the final stage electric field of an electrostatic precipitator in a 300MW coal-fired power unit. The rated DC output voltage of the frequency converter is 80kV, and the rated DC output current is 1A. The unit load increases from 240MW to 270MW, and the soot blowers are started simultaneously. The frequency converter controller calculates the pure time delay of the process from flue gas disturbance to the electrostatic precipitator electric field based on the thermal variable sequence. Before the flue gas disturbance arrives, it initiates online detection of the electric field load characteristics and adjusts the secondary high-voltage DC voltage and the main switch frequency based on the predicted value and variance of the spark discharge critical voltage.

[0101] In this embodiment, the initial stable value of the secondary side high-voltage DC voltage is 68.0 kV, and the initial stable value of the secondary side DC current is 0.72 A. During the unit load adjustment process, the coal feed rate increases from 132 t / h to 148 t / h, the induced draft fan damper opening increases from 61% to 67%, the secondary damper opening increases from 46% to 51%, and the sootblower start-up status flag changes from 0 to 1. The first refresh time when the thermal variables undergo a step change is recorded as... .

[0102] The variable frequency power controller reads historical thermal disturbance templates corresponding to sootblower startup and unit load increase conditions from non-volatile memory. These historical thermal disturbance templates record the basic lag time. The sampling period is 18.4 s. The real-time thermal variable sequence matching segment contains 104 sampling intervals, and the historical thermal disturbance template matching segment contains 90 sampling intervals. Sampling period of historical thermal disturbance template All times are 1 second.

[0103] Lower limit of pure time delay in process The time delay is 14.4s, which is the upper limit of the pure time delay of the process. The control response margin time is 24.0s. Take 10% of the pure time delay of the process. The time scaling factor, the pure time delay of the process, and the feedforward control time window are calculated using the following formula. Substituting the parameters of this embodiment, we get: ; ; ; The recording time when the electric field load deviation index first reaches the response threshold is +21.40s, the calculated pure lag time differs from the recorded result by 0.13s. When a fixed base lag time of 18.4s is used, the fixed time differs from the recorded result by 3.00s.

[0104] After the control clock enters the feedforward control time window, the field-programmable gate array injects one high-frequency probe pulse group for each prediction calculation cycle. The prediction calculation cycle is 10ms, the probe pulse repetition frequency is 2.5 times the reference main switch frequency, the high-frequency probe pulse duty cycle is 3%, each high-frequency probe pulse group includes 8 detection cycles, and the high-speed continuous sampling frequency of the detection window is 16 times the probe pulse repetition frequency.

[0105] Before the flue gas disturbance reaches the electric field, the filtered electric field load deviates exponentially. Maintain between 0.058 and 0.071. Control clock reaches... +19.0s later, The values ​​successively increased to 0.116, 0.181, and 0.247. Electric field load deviation exponential change rate. exist +21.0s to reach 1.84s -1 A rate of change exceeding the threshold =1.20s -1 The variable frequency power controller is selected using the Matrn 3 / 2 kernel Gaussian process regression model.

[0106] exist At +21.0s, the predicted value of the spark discharge critical voltage output by the Gaussian process regression model. The voltage is 67.8 kV, and the prediction variance is... 3.24kV 2 The prediction standard deviation is 1.80kV. Frequency upsampling and impedance increase reach the threshold. 2.25kV 2 Frequency increase and impedance increase exit threshold 1.35kV 2 Once the prediction variance reaches the threshold for frequency upsampling and impedance increase, the finite state machine enters the frequency upsampling and impedance increase mode.

[0107] The safety margin voltage is taken as 5% of the rated output voltage, i.e., 4.0kV; the standard deviation margin factor is taken as 2. The target given voltage is obtained by subtracting the safety margin voltage and twice the predicted standard deviation from the predicted spark discharge critical voltage value; the result calculated in this case is 60.2kV. After entering the frequency boosting and impedance amplification mode, the upper limit of the duty cycle is limited to 90% of the duty cycle before entering this mode.

[0108] In this embodiment, the resonant inductor The resonant capacitance is 65μH. The equivalent series resistance is 1.0 μF. The impedance is 1.8Ω. The inherent resonant frequency and AC impedance of the series resonant network are calculated using the following formula, substituting... =65μH and =1.0μF, thus obtaining the natural resonant frequency of the series resonant network. The reference main switching frequency is 21.71kHz, at which the AC impedance of the series resonant network is 2.37Ω.

[0109] Based on the linear mapping between the prediction variance and the target main switching frequency, the target main switching frequency is 23.46 kHz. After the main switching frequency increases to 23.46 kHz, the AC impedance of the series resonant network increases to 3.33 Ω. The main switching frequency adjustment lasts for 8 prediction calculation cycles, with a peak secondary DC current of 1.21 A and a minimum secondary high-voltage DC voltage of 60.6 kV.

[0110] exist At +24.60s, the prediction variance decreased to 1.18kV. 2 The rate of change of the electric field load deviation index decreased to 0.43s. -1 The secondary DC current drops to 0.76A. After this state persists for 5 prediction calculation cycles, the finite state machine enters the frequency recovery mode. The main switch frequency recovers to 21.71kHz after 8 prediction calculation cycles, and the secondary high-voltage DC voltage recovers to above 65.5kV within 0.61s.

[0111] See attached document Figure 8 , Figure 8 A rectangular coordinate graph with two vertical axes is used. The horizontal axis represents the time after the thermal disturbance occurs, in seconds; the left vertical axis represents the prediction variance, in kV. 2 The right vertical axis represents the main switching frequency, in kHz.

[0112] Figure 8 The gray area in the text represents +19.14s~ The feedforward control time window is +21.27s. The solid line represents the prediction variance, and the stepped dashed line represents the main switching frequency. The two horizontal reference lines represent the frequency upsampling and impedance upsampling entry threshold and the frequency upsampling and impedance upsampling exit threshold, respectively. The two vertical reference lines represent the arrival time of the flue gas disturbance and the time of entering the frequency recovery mode, respectively. The legend is located inside the Cartesian coordinate system and has a border.

[0113] Figure 8 The prediction variance curve is plotted using conformal interpolation based on discrete sampled values. This interpolation is only used to improve the continuity of the curve display; the controller still performs threshold determination based on the original discrete sampled values. The prediction variance reaches 2.25kV. 2 Subsequently, the main switching frequency increased from 21.71kHz to 23.46kHz; the prediction variance decreased to 1.35kV. 2 Once the recovery conditions are met, the main switching frequency recovers to 21.71kHz.

[0114] To verify the control effect, three sets of comparative tests were set up: conventional voltage closed-loop control, voltage feedforward control, and the control method of this invention. All three sets of tests used the same initial voltage, initial current, thermal disturbance amplitude, and electric field load conditions, with each control method repeated 30 times. Statistical indicators included the peak value of the secondary side DC current, voltage recovery time, number of spark discharges within 10 minutes, and outlet dust concentration.

[0115] Conventional voltage closed-loop control employs a fixed main switching frequency and closed-loop regulation of the secondary side high-voltage DC voltage. Voltage feedforward control reduces the target setpoint voltage in advance based on the pure time lag of the process, while maintaining a constant main switching frequency. The control method of this invention simultaneously performs voltage amplitude regulation and frequency-increase impedance control.

[0116] Under conventional voltage closed-loop control, the peak DC current on the secondary side is 2.46±0.18A, the voltage recovery time is 1.42±0.08s, the number of spark discharges within 10 minutes is 13.2±1.7, and the outlet dust concentration is 22.7±1.1mg / Nm³. 3 .

[0117] Under voltage feedforward control, the peak DC current on the secondary side is 1.82±0.13A, the voltage recovery time is 0.98±0.06s, the number of spark discharges within 10 minutes is 8.1±1.2, and the outlet dust concentration is 17.6±0.8mg / Nm³. 3 .

[0118] Under the control method of this invention, the peak DC current on the secondary side is 1.21±0.09A, the voltage recovery time is 0.61±0.04s, the number of spark discharges within 10 minutes is 4.3±0.8, and the outlet dust concentration is 13.9±0.6mg / Nm³. 3 .

[0119] See attached document Figure 9 , Figure 9 Grouped bar charts were used, with the horizontal axis including peak secondary DC current, voltage recovery time, number of spark discharges within 10 minutes, and outlet dust concentration, and the vertical axis representing the improvement ratio of the indicators compared to conventional voltage closed-loop control, in units of %.

[0120] Figure 9 The light gray bars represent voltage feedforward control, and the dark gray bars represent the control method of this invention. The corresponding improvement percentage is indicated above each bar, and the measured mean and standard deviation of 30 repeated tests are indicated inside each bar. The legend is located within a rectangular coordinate system and has a border.

[0121] Compared with conventional voltage closed-loop control, voltage feedforward control reduces the peak DC current on the secondary side by 26.0%, shortens the voltage recovery time by 31.0%, reduces the number of spark discharges within 10 minutes by 38.6%, and reduces the outlet dust concentration by 22.5%.

[0122] Compared with conventional voltage closed-loop control, the control method of this invention reduces the peak value of secondary DC current by 50.8%, shortens the voltage recovery time by 57.0%, reduces the number of spark discharges within 10 minutes by 67.4%, and reduces the outlet dust concentration by 38.8%.

Claims

1. An AI-based intelligent amplitude modulation method for electrostatic precipitator frequency converter power supply under flue gas disturbance, characterized in that, include: Obtain the thermal variable vector and thermal data acquisition timestamp, current main switch frequency, current duty cycle and electrical feedback data; The dynamic time warping algorithm is executed to align the real-time thermal variable sequence composed of the thermal variable vectors with the time domain of the historical thermal disturbance template, calculate the pure time delay of the process, and establish a feedforward control time window. Inject high-frequency detection pulses within the feedforward control time window, extract the voltage and current phase difference, and calculate the filtered electric field load deviation index and the rate of change of the electric field load deviation index. The thermal variable vector, the filtered electric field load deviation index, the electric field load deviation index change rate, the current main switch frequency, the current duty cycle, and the electrical feedback data are combined into a joint feature vector, which is then standardized to generate a model input vector. A Gaussian process regression model is selected based on the electric field load deviation index change rate to output the predicted value of the spark discharge critical voltage and the prediction variance. Based on the predicted value of the spark discharge critical voltage and the predicted variance, the power conversion module is controlled through voltage amplitude adjustment mode, frequency boosting and resistance increasing mode, and frequency recovery mode.

2. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 1, characterized in that, The acquisition of the thermal variable vector and the thermal data acquisition timestamp includes: Receive thermal operation commands, state variables, and thermal data acquisition timestamps from the distributed control system; According to the preset variable order, the thermal operation instructions and the state variables corresponding to the same thermal data acquisition timestamp are arranged sequentially to generate the thermal variable vector; The thermal variable vector is associated with the thermal data acquisition timestamp and written into a static random access memory.

3. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 1, characterized in that, The execution of the dynamic time warping algorithm aligns the real-time thermal variable sequence composed of the thermal variable vectors with the historical thermal disturbance template in the time domain, and calculates the pure time delay of the process, including: Range standardization is performed on the real-time thermal variable sequence and each of the historical thermal disturbance templates; Calculate the Euclidean distance between the real-time thermal variable sequence and each data point in the historical thermal disturbance template, and recursively establish the cumulative distance matrix; The normalized path is obtained by tracing back from the end point of the cumulative distance matrix to the starting point, and the historical thermal disturbance template with the smallest normalized dynamic time normalized distance and lower than the preset template matching threshold is selected. The time scaling factor is calculated based on the start and end indices of the matching segments in the regularized path. The time scaling factor is then multiplied by the base lag time recorded in the selected historical thermal disturbance template and subjected to upper and lower limit processing to obtain the pure lag time of the process.

4. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 3, characterized in that, The establishment of the feedforward control time window as described in claim 1 includes: The normalized Euclidean distance is calculated using the standardized difference of adjacent thermal variable vectors; When the normalized Euclidean distance reaches a preset value for multiple consecutive refresh cycles, the first refresh time that reaches the preset value is determined as the thermal disturbance start time. The lower limit of the time is obtained by subtracting the preset control response margin time from the pure delay time of the process, and the upper limit of the time is obtained by taking the pure delay time of the process. Using the start time of the thermal disturbance as a reference, and combining the lower time limit and the upper time limit, the feedforward control time window is established.

5. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 1, characterized in that, The step of injecting high-frequency detection pulses within the feedforward control time window and extracting the voltage-current phase difference includes: During the high-frequency detection pulse injection, the detection window is activated for high-speed continuous sampling to acquire primary-side high-frequency voltage feedback data and primary-side high-frequency current feedback data. The sampled data is subjected to digital bandpass filtering with a center frequency equal to the repetition frequency of the probe pulse, and then single-frequency discrete Fourier transform is performed to obtain the complex components of the primary side voltage and the complex components of the primary side current. The phase angle difference is obtained by performing complex phase angle operations and then subtracting the steady-state reference phase obtained based on three-dimensional linear interpolation to obtain the voltage-current phase difference.

6. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 5, characterized in that, The calculation of the filtered electric field load deviation index and the rate of change of the electric field load deviation index as described in claim 1 includes: The phase deviation term is calculated based on the ratio of the voltage-current phase difference to the steady-state reference phase and the effective value of the primary side high-frequency current. The deviation from the steady-state reference complex impedance is calculated based on the ratio of the complex component of the primary voltage to the complex component of the primary current, and converted into an impedance deviation term. The electric field load deviation index is obtained by combining the phase deviation term and the impedance deviation term in a weighted manner. The electric field load deviation index is filtered to obtain the filtered electric field load deviation index, and the rate of change of the electric field load deviation index is calculated using the difference between adjacent prediction calculation periods.

7. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 1, characterized in that, The step of selecting a Gaussian process regression model based on the rate of change of the electric field load deviation includes: When the magnitude of the electric field load deviating from the exponential rate of change reaches a preset rate of change threshold, the Gaussian process regression model selects the Matern 3 / 2 kernel Gaussian process regression model. When the magnitude of the electric field load deviating from the exponential rate of change falls back to within the preset rate of change exit threshold and continues for a preset period, the Gaussian process regression model selects the square exponential kernel Gaussian process regression model. When the magnitude of the electric field load deviating from the exponential rate of change is between the rate of change exit threshold and the rate of change entry threshold, the Gaussian process regression model selected in the previous prediction calculation cycle is maintained.

8. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 7, characterized in that, The predicted value and prediction variance of the output spark discharge critical voltage as described in claim 1 include: The mean and standard deviation stored offline are used to perform standardization on the joint feature vector to generate the model input vector; Calculate the covariance between the model input vector and the representative training sample set; By combining the Cholesky decomposition results saved during the training phase and the preset observation noise variance, the predicted value of the spark discharge critical voltage and the predicted variance are obtained through forward substitution and back substitution operations.

9. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 1, characterized in that, The power conversion module is controlled via voltage amplitude adjustment mode, frequency boosting and impedance enhancement mode, and frequency recovery mode. The voltage amplitude adjustment mode includes: The target given voltage is obtained by subtracting the preset safety margin voltage from the predicted value of the spark discharge critical voltage, and then subtracting the product of the preset standard deviation margin coefficient and the predicted standard deviation, wherein the predicted standard deviation is the square root of the predicted variance. The deviation between the target given voltage and the secondary side high voltage DC voltage feedback value is input into the control loop, and the duty cycle command is output. When the predicted variance reaches the preset up-frequency and down-resistance threshold, the switch to the up-frequency and down-resistance mode is triggered.

10. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 9, characterized in that, The frequency boosting and impedance enhancement mode includes: The target given voltage is kept from rising further, and the upper limit of the duty cycle is limited to a preset ratio of the current duty cycle before entering the frequency boosting and impedance amplification mode; Based on the predicted variance, a linear interpolation is performed between the frequency up-conversion and impedance up-conversion threshold and the preset upper limit of the frequency mapping variance to determine the target main switch frequency; The current main switch frequency is increased by a fixed step size until the target main switch frequency is reached, thereby increasing the AC impedance of the series resonant network. When the predicted variance falls back to within the preset frequency increase and impedance reduction exit threshold and meets the preset recovery conditions, the frequency recovery mode is switched to, and the current main switch frequency is gradually reduced to the preset reference main switch frequency in the frequency recovery mode.

11. The AI ​​intelligent amplitude modulation method for electrostatic precipitator frequency converter under flue gas disturbance as described in claim 1, characterized in that, The method further includes: When communication interruption, model data verification failure, modulus sampling abnormality, or memory read abnormality occurs, the derating operation mode is triggered. In the derating operation mode, the current duty cycle is fixed to a preset derating duty cycle, the current main switch frequency is fixed to a preset reference main switch frequency, and hardware protection is enabled. The execution priority of the hardware protection is higher than that of the control logic based on regression prediction results.

Images