Pulverizing system ADP load optimization control method and device

Abstract

The invention provides a coal pulverizing system ADP load optimization control method and device, and the method employs a three-layer cooperative control architecture of "data driving-algorithm optimization-security constraint" to construct a full-flow solution of coal pulverizing system load optimization. Wherein the data driving layer is responsible for real-time acquisition and preprocessing of operation parameters and provides a high-quality data basis for upper-layer control; the algorithm optimization layer realizes dynamic decision by fusing an advanced intelligent algorithm, and solves the problems of system time lag and model precision; the security constraint layer constructs a dual protection mechanism based on process limit conditions, and ensures the security and stability of the optimization process. The three-layer architecture realizes integrated control of the coal pulverizing system from data perception to intelligent decision making through closed-loop interaction of a data stream and a control instruction.

Description

Technical Field

[0001] This invention relates to the field of power system control technology, and in particular to an ADP load optimization control method and device for a pulverizing system. Background Technology

[0002] As a core component of coal-fired power generating units, the pulverizing system is widely used in the power production field. Its control performance directly affects the unit's peak-shaving capacity, operational economy, and safety. With the development of intelligent control technology, traditional PID control and manual operation have gradually evolved towards data-driven and algorithm optimization. Related technologies utilize a three-layer architecture of multi-source data acquisition, control strategy iteration, and safety constraint coordination to construct a complete control system from parameter perception to decision execution. Specifically, this technology system covers the entire process from air-coal system monitoring and coal quality characteristic analysis to equipment operating status assessment. It includes hardware collaboration such as acoustic sensors, electrostatic detection devices, and PLC controllers, as well as data processing modules based on algorithms such as Z-score normalization and wavelet transform, forming a multi-technology fusion framework covering a 6-dimensional parameter matrix.

[0003] However, existing pulverizing system control methods directly employ single-loop PID control and manual experience-based adjustment, lacking a systematic solution for the strong coupling and time-delay effects of multiple variables. This can lead to technical bottlenecks such as lag in dynamic response, insufficient control accuracy, and decreased system stability. Specifically, in scenarios with abrupt load changes (e.g., 30%~100%), traditional control strategies experience adjustment times of 15-20 seconds due to material transport delays and equipment inertia, significantly lagging behind the grid's AGC requirement of <5 minutes for response time. Furthermore, the strong coupling characteristics of variables such as coal grinding volume, ventilation volume, and hot air temperature often result in overshoot or oscillations during control command execution. Experimental data shows that coal powder fineness deviations can exceed ±5%, leading to reduced combustion efficiency and fluctuations in environmental indicators. Therefore, existing technologies struggle to achieve synergistic optimization of control accuracy and system stability when dealing with complex operating conditions such as varying coal types (e.g., lignite with 25%-40% moisture content) and wide-load operation (e.g., deep peak shaving to 30% output ratio), ultimately resulting in increased equipment failure rates and higher operating costs.

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose an ADP load optimization control method for a pulverizing system.

[0006] Another objective of this invention is to provide an ADP load optimization control device for a pulverizing system.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention provides a method for optimizing and controlling the ADP load of a pulverizing system, comprising: S1: Real-time acquisition of the operating parameters of the flour milling system, outlier removal and feature extraction, and generation of standardized dataset; S2, based on the Smith predictor to predict the system state and combined with the ADHP controller to dynamically correct the control command, forms a time-delay compensation closed-loop control logic. S3 uses a two-layer neural network architecture to update the evaluation network online, achieving decoupled optimization control of multiple variables such as coal grinding volume, ventilation volume, and temperature; S4 executes optimized control commands and simultaneously monitors pipeline pressure and pulverized coal flow rate, triggering a graded compensation strategy based on preset thresholds to maintain system stability.

[0010] In one embodiment of the present invention, S1 includes: S11. The IQR method is used to identify and remove outliers. By calculating the upper and lower quartiles of the parameter data, data that exceed the range of [Q1-1.5×IQR, Q3+1.5×IQR] are marked as outliers and interpolated for correction. S12 uses wavelet transform to extract key indicators such as energy spectral density and peak frequency in the 10-2000Hz frequency band for unstructured signals, thereby achieving quantitative characterization of signal features.

[0011] In one embodiment of the present invention, S2 includes: S21, predict the system state X(k+1) at future time based on the current system state and control commands using the Smith predictor; S22 dynamically corrects the current control command U(k) by utilizing the deviation between the predicted state and the actual feedback information, forming a closed-loop control logic of "prediction-correction-execution".

[0012] In one embodiment of the present invention, S3 includes: S31 employs a two-layer hidden layer structure with 32 and 16 neurons respectively. The activation function is ELU to address the neuron death problem of ReLU. S32, dynamically adjust the discount factor γ according to the load fluctuation amplitude ΔQ: when ΔQ>5%, γ=0.95; when ΔQ<2%, γ=0.8.

[0013] In one embodiment of the present invention, S4 includes: S41, the pressure P and pulverized coal flow velocity v in the pipeline are collected in real time through pressure sensors and flow meters; S42, when P>1.2kPa is detected, negative compensation for the coal feeder speed is performed, and when v<18m / s, compensation for the opening of the damper is performed.

[0014] To achieve the above objectives, a second aspect of the present invention provides an ADP load optimization control device for a pulverizing system, comprising: The data acquisition and preprocessing module is used to collect the operating parameters of the flour milling system in real time, remove outliers and extract features, and generate a standardized dataset. The Smith predictor and AHDP control module is used to predict the system state based on the Smith predictor and combine it with the AHDP controller to dynamically correct the control commands, forming a time-delay compensated closed-loop control logic. The dual-layer neural network optimization module is used to update the evaluation network online through a dual-layer neural network architecture, thereby achieving decoupled optimization control of multiple variables such as coal grinding rate, ventilation rate, and temperature. The execution and compensation monitoring module is used to execute optimized control commands and simultaneously monitor pipeline pressure and pulverized coal flow rate, and trigger graded compensation strategies according to preset thresholds to maintain system stability.

[0015] The present invention discloses an ADP load optimization control method and device for a pulverizing system, which effectively solves the time delay effect and multivariate coupling problem of the pulverizing system under sudden load changes, significantly improves the dynamic response speed and control accuracy, shortens the response time to less than 3 seconds, and controls the coal powder fineness deviation to within ±2%.

[0016] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory to implement an ADP load optimization control method for a pulverizing system as described in the first aspect embodiment.

[0017] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an ADP load optimization control method for a pulverizing system as described in the first aspect embodiment.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] Figure 1 This is a flowchart of an ADP load optimization control method for a pulverizing system according to an embodiment of the present invention; Figure 2 This is a distributed measurement system architecture diagram according to an embodiment of the present invention; Figure 3 This is a structural diagram of an ADP load optimization control device for a pulverizing system according to an embodiment of the present invention; Figure 4 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes an ADP load optimization control method and apparatus for a pulverizing system according to an embodiment of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of an ADP load optimization control method for a pulverizing system according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1 collects the operating parameters of the flour milling system in real time, performs outlier removal and feature extraction, and generates a standardized dataset.

[0024] Specifically, the system collects operating parameters of the pulverizing system in real time, performs outlier removal and feature extraction, and generates a standardized dataset. Its technical implementation principle is based on a synchronous acquisition and intelligent preprocessing mechanism of multi-source heterogeneous data. The specific operation includes: first, real-time acquisition of key parameters during the pulverizing system's operation is performed through a distributed sensor network (such as acoustic sensors, electrostatic sensors, flow detection devices, and PLC controllers). This covers 18 parameters across six categories: air-coal system, fuel characteristics, equipment operation, pulverized coal characteristics, combustion status, and unit load, forming a parameter matrix. The acquisition frequency is [missing information]. This ensures the continuity of data in time and the real-time reflection of system status.

[0025] Furthermore, the IQR (interquartile range) method is used to identify and remove outliers. Specifically, the upper quartile (Q3) and lower quartile (Q1) of each parameter are calculated, and the range of outliers is defined as follows: Data points outside this range are marked as anomalies and corrected using linear interpolation or moving average methods to eliminate the impact of sensor noise or transient disturbances on subsequent modeling. For unstructured signals (such as acoustic signals from pulverized coal flow and charged particle signals), wavelet transform techniques are used to extract... The energy spectral density and peak frequency within the frequency band are used to achieve a structured characterization of the signal.

[0026] Furthermore, Z-score normalization was performed on the cleaned data to uniformly map parameters of different dimensions to the [-1,1] interval, improving the stability and generalization ability of model training. For feature selection, Pearson correlation coefficient analysis and random forest feature importance assessment were combined to remove redundant features with a correlation of less than 0.3 with the load optimization objective, retaining parameters that have a significant impact on the control strategy.

[0027] Furthermore, this step is widely applicable in practical applications to the pulverizing systems of coal-fired units, especially in bituminous coal (volatile matter content...). ) and lignite (moisture content) The model was designed to operate under various coal types and load conditions. By constructing a comprehensive dataset containing over 100,000 samples and using a sliding time window method (with a window length of 60 sampling periods) to divide the training and testing sets, the adaptability and robustness of the model under complex operating conditions were ensured.

[0028] Furthermore, the technical effect of this step is to provide high-quality, low-noise, and structured input data for the upper-level ADHDP control algorithm, thereby significantly improving the model's training efficiency and control accuracy. Experiments show that the dataset processed in this step can control the modeling error of the control model within a certain range. This lays a solid foundation for achieving adaptive optimization control of the pulverizing system load.

[0029] Furthermore, S1 includes: S11. The IQR method is used to identify and remove outliers. By calculating the upper and lower quartiles of the parameter data, data that exceed the range of [Q1-1.5×IQR, Q3+1.5×IQR] are marked as outliers and interpolated for correction.

[0030] Specifically, in the data acquisition and preprocessing module of this invention, the IQR method is used to identify and correct outliers in the acquired operating parameters of the pulverizing system. This method is based on statistical principles, calculating the lower quartile (Q1) and upper quartile (Q3) of the data, and defining the interquartile range. Further construct outlier identification intervals Sample points outside the specified range are identified as outliers and interpolated.

[0031] Furthermore, the system first processes the collected parameters (such as primary air flow, pulverizer current, and coal powder fineness) into separate channels, calculating Q1 and Q3 for each parameter. Q1 and Q3 can be calculated by sorting the parameters and taking the 25th and 75th percentiles, or by using an approximate algorithm (such as Tukey's method) for rapid estimation. Subsequently, the system uses the formula... Determine if each data point is an outlier. satisfy or If so, mark it as an exception.

[0032] Furthermore, to avoid model training bias or control strategy distortion caused by outliers, the system employs linear interpolation or spline interpolation methods to correct outlier data. The interpolation operation is based on time series characteristics, using several normal samples (usually the first three and the last three) before and after the outlier to fit the data, ensuring that the corrected data maintains temporal continuity and reasonableness. The interpolation accuracy requirement is that the mean squared error (MSE) between the corrected data and the original normal data does not exceed [a certain value]. To ensure the reliability of the data.

[0033] Furthermore, by eliminating outliers caused by sensor malfunctions, transient disturbances, or system anomalies, the stability and representativeness of the dataset can be significantly improved, providing high-quality input data for the ADHDP control module, thereby enhancing the model's generalization ability and control accuracy. Especially in complex operating environments with multiple coal types and operating conditions, the IQR method can effectively filter noise and improve the system's adaptability to changes in coal quality and load fluctuations.

[0034] S12 uses wavelet transform to extract key indicators such as energy spectral density and peak frequency in the 10-2000Hz frequency band for unstructured signals, thereby achieving quantitative characterization of signal features.

[0035] Specifically, in the data acquisition and preprocessing module of this invention, wavelet transform technology is used in the feature extraction step for unstructured signals (such as the acoustic wave signal of coal powder flow and the charged signal of particles) to achieve signal... Quantitative characterization of indicators such as energy spectral density and peak frequency within the frequency band.

[0036] Furthermore, either discrete wavelet transform (DWT) or continuous wavelet transform (CWT) can be used, the specific choice depending on the time-varying characteristics of the signal and computational resource constraints. For non-stationary signals with transient characteristics, such as the acoustic wave signal of pulverized coal flow, CWT is more advantageous, as it utilizes a scaling function... Multi-scale analysis of the signal, including For scale parameters, For translation parameters, The mother wavelet function is used. In this invention, Morlet wavelets or the Daubechies wavelet family are preferred to balance frequency domain resolution and computational efficiency.

[0037] Furthermore, after wavelet transform, the signal is extracted. The energy spectral density within a frequency band is calculated by summing the squares of the wavelet coefficients within that band, i.e.: in Indicates frequency The wavelet coefficients at that point. The peak frequency is then achieved by searching for the frequency point corresponding to the maximum energy value within that frequency band, i.e.:

[0039] Optionally, to improve feature robustness, the system can also calculate the mean and variance of the energy distribution within the frequency band, using these as auxiliary feature inputs to the ADHDP control module. In practical applications, this step is deployed in the signal processing chain of the acoustic and electrostatic sensors in the pulverizing system, operating at a frequency of [frequency missing]. The sampling period is This ensures real-time response to dynamic operating conditions.

[0040] Furthermore, the originally difficult-to-analyze unstructured signals are transformed into frequency domain features with physical meaning, providing quantifiable input variables for intelligent control algorithms. This is achieved by extracting... By analyzing the energy spectral density and peak frequency within the frequency band, the system can effectively identify the flow state of pulverized coal, particle distribution characteristics, and equipment malfunctions, thereby improving the accuracy and adaptability of load optimization control.

[0041] S2, based on the Smith predictor to predict the system state and combined with the ADHDP controller to dynamically correct the control commands, forms a time-delay compensation closed-loop control logic.

[0042] Specifically, the technical principle of "predicting system state based on Smith predictor and dynamically correcting control commands based on ADHDP controller to form time delay compensation closed-loop control logic" in this invention is based on the fusion control strategy of system identification and time delay compensation, which aims to solve the control lag problem caused by material transmission delay, equipment inertia and other factors in the pulverizing system.

[0043] Furthermore, the ADHDP controller dynamically corrects the current control command based on the deviation between the predicted state and the actual feedback state. The ADHDP controller employs a two-layer nested neural network architecture, including an evaluation network and a control network, and is updated online through a reinforcement learning mechanism. During training, the system uses a loss function... ,in As a discount factor, based on the load fluctuation range Dynamic adjustment: when hour, To increase the weight of the current action; when hour, This mechanism prioritizes long-term benefits. Through this mechanism, the controller can adjust its control strategy in real time, improving the system's robustness to time-delay disturbances.

[0044] Furthermore, the prediction step size of the Smith predictor is one sampling period. The ADHDP controller's neural network structure consists of two hidden layers (32 / 16 neurons), using ELU as the activation function and Adam as the optimizer. The initial learning rate is 0.001, decaying to 10% of the current value every 5000 iterations. Under typical load disturbance scenarios, this step can reduce the system response lag time from 15-20 seconds to less than 3 seconds, improving response speed by approximately 80% and significantly enhancing the system's dynamic performance.

[0045] Furthermore, it is applicable to the control of pulverizing systems in coal-fired power units under complex operating conditions such as variable load operation, coal type switching, and equipment aging. Through the synergistic effect of time delay compensation and the ADHDP controller, the system can achieve rapid response and precise control of key parameters such as mill outlet temperature, pulverized coal fineness, and air volume, thereby improving overall operating efficiency and economy. Its technical value lies in effectively solving the problems of regulation lag and overshoot caused by time delay in traditional control strategies, providing a guarantee for the stable and efficient operation of the pulverizing system.

[0046] Furthermore, S2 includes: S21, predict the system state X(k+1) at future time based on the current system state and control commands using the Smith predictor.

[0047] Specifically, in the time delay compensation module of this invention, the Smith predictor predicts the system state at future times based on the current system state and control commands. This step involves constructing a prediction mechanism based on a system model to compensate for control lag issues caused by factors such as material transport delays and equipment inertia in the pulverizing system.

[0048] Furthermore, the Smith predictor establishes a dynamic model of the system and calculates the system state vector at the current moment. With control input vector Input into the model to predict the system state at the next time step. This model is typically based on state-space equations:

[0049] in, Here is the state transition matrix. To control the input matrix, The perturbation input matrix, This indicates external disturbances (such as changes in coal quality, fluctuations in air volume, etc.). For systematic random error, The error coefficient is denoted as . This model is trained using historical data and can accurately reflect the dynamic behavior of the system under different operating conditions.

[0050] Furthermore, the Smith predictor incorporates a time delay compensation mechanism during the prediction process, that is, it explicitly models the system time delay in the model. Based on the deviation between the predicted results and the actual feedback, the control commands are dynamically adjusted. Specifically, the system first uses the current state and control input to predict... Then, the predicted value is compared with the actual measured value, and the prediction error is calculated. This information is then fed back to the ADHDP controller to adjust the control strategy, thereby compensating for the time delay effect.

[0051] Furthermore, the prediction step size of the Smith predictor is typically consistent with the system sampling period, i.e. This ensures that the prediction results are synchronized with the dynamic changes of the system. The training data of the model comes from the operation data of multiple units, multiple coal types (such as bituminous coal and lignite), and multiple operating conditions (such as variable load and variable coal quality). A comprehensive dataset containing more than 100,000 samples is constructed to improve the generalization ability and prediction accuracy of the model.

[0052] Furthermore, it is particularly suitable for sudden changes in load (such as...) This is particularly useful in scenarios involving abrupt changes or significant fluctuations in coal quality. By predicting the system state in advance, the controller can respond earlier, avoiding overshoot or lag caused by time delays, thereby significantly improving the system's dynamic performance and control accuracy. Experimental data shows that this predictive mechanism can control the time delay compensation error within a certain range. Within 3 seconds, the response lag time is reduced from 15-20 seconds in traditional control to within 3 seconds, and the response speed is improved by about 80%, providing key technical support for the stable and efficient operation of the pulverizing system.

[0053] S22 dynamically corrects the current control command U(k) by utilizing the deviation between the predicted state and the actual feedback information, forming a closed-loop control logic of "prediction-correction-execution".

[0054] Specifically, in the "time delay compensation module" of this invention, the step of "dynamically correcting the current control command U(k) by utilizing the deviation between the predicted state and the actual feedback information to form a closed-loop control logic of 'prediction-correction-execution'" is the core link in realizing the dynamic response optimization of the system. This step is based on the fusion architecture of Adaptive Dynamic Programming (ADHDP) and Smith predictor. By predicting the system state in real time and combining the feedback deviation to correct the control command online, it effectively alleviates the control lag problem caused by material transmission delay, equipment inertia and other factors in the pulverizing system.

[0055] Furthermore, the system first uses the Smith predictor to estimate the state at the next time step. The prediction model is based on the current state. With control input The linear or nonlinear mapping relationship is then established. Subsequently, the ADHDP controller compares the predicted state with the actual feedback state to calculate the state deviation. This deviation is input to the control network to dynamically adjust the control commands. This is to compensate for prediction errors caused by system time delays. The correction process of control commands follows the gradient descent strategy in reinforcement learning, minimizing the loss function. To achieve online optimization of the control strategy, among which... This is a dynamic discount factor, based on the load fluctuation range. Adaptive adjustments are made to balance current and long-term control effects.

[0056] Furthermore, the system is set to control the time delay compensation error within... Within this range, ensure the stability of control accuracy. The prediction window length and sampling frequency of the Smith predictor... To match, the ADHDP controller's neural network structure consists of three hidden layers (64 / 32 / 16 neurons), and the output layer uses the Sigmoid function to normalize the control input. The Adam optimizer's learning rate is initially set at 0.001, and decays to 10% of the current value every 5000 iterations to balance model convergence speed and control accuracy.

[0057] Furthermore, this step is widely applicable to the control of pulverizing systems in coal-fired units under complex operating conditions such as variable load and variable coal quality. This is especially true for situations involving sudden load changes (such as...). When switching between coal types (such as alternating between bituminous coal and lignite), the system can quickly correct control commands to avoid overshoot and oscillation caused by time delay, ensuring stable output of key parameters such as coal mill outlet temperature and coal powder fineness.

[0058] Furthermore, experimental data shows that after integrating this closed-loop correction mechanism, the system response lag time is reduced from 15-20 seconds under traditional control to less than 3 seconds, improving the response speed by approximately 80%. Simultaneously, through deviation-driven control command correction, the system maintains high robustness and stability in the face of external disturbances and internal time delays, providing key technical support for the efficient and safe operation of the pulverizing system.

[0059] S3 uses a two-layer neural network architecture to update the evaluation network online, achieving decoupled optimization control of multiple variables such as coal grinding volume, ventilation volume, and temperature.

[0060] Specifically, the evaluation network is updated online through a two-layer neural network architecture to achieve decoupled optimization control of multiple variables such as coal grinding volume, ventilation volume and temperature. This step is based on adaptive dynamic programming (ADP) theory and combined with reinforcement learning mechanism to construct a two-layer neural network model with online learning capability, so as to achieve collaborative optimization of multiple strongly coupled variables in the pulverizing system.

[0061] Furthermore, this two-layer neural network architecture consists of an evaluation network and a control network, wherein the evaluation network is used to estimate the performance index function of the system in the current state. The control network is used to generate the optimal control input. The network structure employs two hidden layers with 32 and 16 neurons respectively. ELU is chosen as the activation function to mitigate the neuron death problem of ReLU in the negative input region and improve the model's non-linear fitting ability. The loss function is defined as:

[0062] in, This is the reward signal for the current moment. Let V be the discount factor, and V^(X(k+1)) be the optimal performance estimate for the next time step. In this invention, Employ a dynamic adjustment strategy: when load fluctuates hour, To increase the weight of the current action; when hour, To prioritize long-term gains, the optimizer uses the Adam algorithm with an initial learning rate of 0.001, which decays to 10% of the current value every 5000 iterations to balance the model's convergence speed and stability.

[0063] Furthermore, the system constructs a state vector by collecting 12 key parameters in real time, including coal mill current, outlet temperature, and coal powder fineness. and in combination with the current control input The weight matrix of the evaluation network is continuously updated through a reinforcement learning mechanism. This process uses a training cycle of 5000 iterations, and the modeling error is controlled within a certain range. Within this range, ensure consistency between the model output and the actual system response.

[0064] Furthermore, the control platform for pulverizing systems, primarily deployed in coal-fired units, is applicable to various coal types, including bituminous coal and lignite. Especially in scenarios involving sudden load changes (such as a 30%–100% step change), it can significantly shorten dynamic adjustment time and improve system response speed. Through an online update mechanism, the system can adapt in real time to nonlinear disturbances such as changes in coal quality and equipment aging, achieving decoupled control of variables such as pulverization rate, ventilation volume, and temperature. This avoids the regulation lag or overshoot problems caused by variable coupling in traditional PID control, thereby improving the operating efficiency and control accuracy of the pulverizing system.

[0065] Furthermore, S3 includes: S31 employs a two-layer hidden layer structure with 32 and 16 neurons respectively. The activation function chosen is ELU to address the neuron death problem of ReLU.

[0066] Specifically, in the ADHDP core control module of this invention, the neural network structure adopts a two-layer hidden layer design with 32 and 16 neurons respectively. The activation function is ELU (Exponential Linear Unit) to solve the neuron "death" problem that may occur during ReLU training. This structure has a clear mathematical modeling foundation and engineering parameter configuration in its implementation, ensuring the stability and generalization ability of the model under complex working conditions.

[0067] Furthermore, the hidden layer structure is designed based on the modeling requirements of the multivariate coupling characteristics of the pulverizing system. The first hidden layer contains 32 neurons, used to extract the nonlinear mapping relationship of the input features, enhancing the model's ability to perceive the system state; the second hidden layer contains 16 neurons, used to further abstract features and approximate the optimal control strategy. The output of each neuron undergoes a nonlinear transformation through the ELU function, the mathematical expression of which is:

[0068] in, The slope parameter for the negative interval is usually set to... This enhances the model's expressive power when the input is negative, avoiding the vanishing gradient problem caused by ReLU's output always being zero in the negative input region. The ELU function has continuous derivatives at zero, which helps improve the convergence speed and stability during training.

[0069] Furthermore, the input layer of the neural network receives 12 key parameters from the data acquisition and preprocessing module, including the pulverizer current, outlet temperature, pulverized coal fineness, primary air volume, and hot air damper opening, which are then normalized to the [-1,1] range using Z-score. The output layer employs the Sigmoid activation function to normalize the control quantity to the [0,1] range, facilitating subsequent interface adaptation with the actuator.

[0070] Furthermore, the optimizer uses the Adam algorithm, with an initial learning rate set to... The loss function is decayed to 10% of its current value after every 5000 iterations to balance the model's rapid convergence with its fine-tuning capabilities. The loss function takes the following form:

[0071] in, As a discount factor, based on the load fluctuation range Dynamic adjustments are made to enhance the model's adaptability to both short-term and long-term benefits. This neural network architecture plays a crucial role in the load optimization of the pulverizing system. Through its nonlinear modeling capabilities, it effectively improves the robustness and real-time response performance of the control strategy, providing a solid foundation for achieving adaptive control over a wide load range.

[0072] S32, dynamically adjust the discount factor γ according to the load fluctuation amplitude ΔQ: when ΔQ>5%, γ=0.95.

[0073] Specifically, in the ADHDP core control module of this invention, based on the load fluctuation amplitude... Dynamically adjust discount factor The core technical principle of this step lies in adjusting the weight distribution between the current control action and future benefits through the discount factor mechanism in reinforcement learning, thereby achieving dynamic optimization of the control strategy under different load change scenarios.

[0074] Furthermore, when the load fluctuation amplitude At that time, the system will apply the discount factor. Set to 0.95. This setting is based on the principle of balancing immediate rewards and long-term benefits in reinforcement learning: under conditions of drastic load fluctuations, the system tends to prioritize responding to changes in the current state to quickly adjust the control output and avoid system instability or overshoot due to delayed response. In this case, the control network assigns a higher weight to the reward signal at the current moment, thereby improving the system's dynamic response capability and robustness. Conversely, when... hour, Setting it to 0.8 emphasizes the accumulation of long-term benefits and is suitable for the stable operation phase of the load, in order to optimize overall energy consumption and system stability.

[0075] Furthermore, this step is implemented based on differential calculation of real-time load data, that is, by comparing the current load... Compared with the previous time load Calculate its relative fluctuation range .when When the threshold is exceeded by 5%, the system triggers. The update mechanism adjusts the discount factor in the control network to 0.95 to enhance its sensitivity to the current control action. This adjustment process is embedded in the online learning process of the ADHDP controller, achieving real-time optimization of the control strategy by dynamically updating the performance index function of the evaluation network.

[0076] Furthermore, this dynamic adjustment mechanism can shorten the system's dynamic adjustment time by more than 40% in scenarios involving sudden load changes (such as a step change of 30% to 100%), while maintaining the coal powder fineness control deviation within a certain range. Within this range, the regulation performance is significantly better than that of traditional PID control. Furthermore, this mechanism supports adaptive control for different coal types (such as bituminous coal and lignite) and operating conditions, ensuring the stability and economy of the system under complex operating conditions.

[0077] Furthermore, this step is widely applicable to coal-fired power units in operating scenarios such as deep peak shaving and rapid load changes. For example, when grid AGC commands change frequently, the system dynamically adjusts... This mechanism can quickly respond to load demands, reduce the number of coal mill start-ups and shutdowns, and improve operating efficiency. Working in conjunction with the time-delay compensation module and the safety constraint module, it forms a closed-loop control system, providing support for the intelligent optimization of the pulverizing system.

[0078] In summary, this step introduces load fluctuation-based... The dynamic adjustment mechanism effectively improves the adaptability and control accuracy of the ADHDP controller under different operating conditions, and is an important technical means to achieve multi-objective collaborative optimization of the pulverizing system.

[0079] S4 executes optimized control commands and simultaneously monitors pipeline pressure and pulverized coal flow rate, triggering a graded compensation strategy based on preset thresholds to maintain system stability.

[0080] Specifically, in the ADP load optimization control method for the pulverizing system of the present invention, the step of executing the optimized control command and simultaneously monitoring the pipeline pressure and pulverized coal flow rate ensures that the system parameters are always within the safe and economical operating range under complex operating conditions through the synergistic effect of the real-time feedback mechanism and the graded compensation strategy.

[0081] Furthermore, the system utilizes pressure sensors and pulverized coal flow meters deployed at key pipeline nodes to... Real-time data is acquired at a sampling frequency. The pressure sensor uses a differential pressure measurement principle, with a measurement range of [range missing]. With an accuracy class of 0.5%, it is used to monitor pressure fluctuations during the transportation of pulverized coal in pipelines; the flow meter is based on laser Doppler or ultrasonic principles, and its measurement range is... resolution It is used to assess the flow state of pulverized coal in pipelines. The collected data is uploaded to the control center in real time via PLC or DCS system and processed synchronously with the optimized control commands output by the ADHDP controller.

[0082] Furthermore, the system sets the pipeline pressure threshold to be [value missing]. When the real-time monitored value exceeds this threshold, it is determined to be an abnormal pressure, triggering the first-level compensation strategy, which involves adjusting the speed of the coal feeder. This allows for dynamic adjustment of the pulverized coal supply to reduce the pressure of material buildup within the pipeline. Simultaneously, a lower limit for the pulverized coal flow rate is set. When the flow rate is lower than this value, it is determined that the conveying capacity is insufficient, triggering a secondary compensation strategy, which adjusts the opening of the primary air damper. Increase air volume to enhance pulverized coal carrying capacity. The compensation strategy follows a tiered response mechanism to ensure differentiated control actions are taken under different levels of anomalies, avoiding over-adjustment or delayed response.

[0083] Furthermore, this step is widely applicable to the operation and control of coal-fired power plant pulverizing systems under complex conditions such as variable load, variable coal quality, and start-up and shutdown of coal mills. Especially during low-load peak shaving or coal type switching, the system effectively addresses risks such as unstable pulverized coal transportation, pipeline blockage, or overpressure through synchronous monitoring and graded compensation, ensuring continuous and stable system operation.

[0084] Furthermore, the technical benefits of this step are reflected in a significant improvement in the system's dynamic response capability and operational safety. Through the combination of real-time monitoring and a tiered compensation strategy, the system can... The load adjustment was completed within minutes, while the fineness of the pulverized coal was controlled within [specific range]. The economic range effectively prevents equipment failure and reduced operating efficiency caused by parameter imbalance, providing a reliable guarantee for the intelligent and efficient operation of the pulverizing system.

[0085] Furthermore, S4 includes: S41 uses a pressure sensor and a flow meter to collect the pressure P and pulverized coal flow velocity v in the pipeline in real time.

[0086] Specifically, in the "safety constraint module" of this invention, the step "real-time acquisition of pressure within the pipeline via pressure sensor and flow meter" and pulverized coal flow rate The technology behind this step is based on the collaborative deployment of distributed sensor networks and high-precision measurement equipment, ensuring real-time perception and feedback of key operating parameters within the pipeline under complex working conditions.

[0087] Furthermore, the pressure sensor is typically a differential pressure or capacitive sensor, installed at key pipeline nodes between the coal mill outlet and the burner inlet to measure instantaneous pressure values ​​during pulverized coal conveying. Its sampling frequency is To ensure timely response to pressure fluctuations, the flow meter is preferably an ultrasonic Doppler flow meter or a hot-wire anemometer. Its measurement principle is based on the signal frequency shift or heat loss changes caused by pulverized coal particles in the airflow, thereby deriving the pulverized coal flow velocity. The measurement range of a flow meter is typically set to... This is to cover the changes in conveying speed under different load conditions.

[0088] Furthermore, the measurement accuracy of the pressure sensor should reach [a certain level]. The range is This is to accommodate pressure fluctuations caused by changes in pulverized coal concentration within the pipeline. The measurement error of the flow meter should be controlled within [specific range]. Within this range, ensure low flow rates (such as...) It can still provide reliable feedback signals even when [the data is collected from the sensor]. The acquired data must meet real-time requirements, that is, the time delay from sensor acquisition to the data processing center should be less than [a certain value]. This is to support rapid response from subsequent control modules.

[0089] Furthermore, this step is widely used in the pulverizing systems of coal-fired power plants, especially under conditions of frequent boiler load changes, coal type switching, or potential system blockage risks. This is achieved through real-time data acquisition. and The system can dynamically determine the coal powder conveying status within the pipeline, providing a basis for subsequent graded compensation strategies. For example, when it detects... When the system determines that the pipeline pressure is close to the safety limit, it triggers negative compensation for the feeder speed; and when If the current condition is not met, it is determined that the conveying capacity is insufficient, triggering positive compensation for the opening of the primary windshield.

[0090] Furthermore, high-precision, high-frequency real-time data acquisition provides crucial assurance for the safe operation of the pulverizing system. Its role is twofold: first, it provides dynamic feedback signals to the system, supporting real-time adjustments to subsequent control algorithms; second, through dual-parameter joint monitoring, it effectively identifies potential operational risks, such as pipeline blockage, coal powder accumulation, or insufficient conveying capacity, thereby improving the stability and safety of system operation. The implementation of this step lays a solid data foundation for realizing the three-layer collaborative control architecture of "data-driven - algorithm optimization - safety constraints," and is a crucial guarantee for the invention to achieve wide-load, high-response, and low-risk operation.

[0091] S42, when P>1.2kPa is detected, negative compensation of coal feeder speed is performed (reduction ΔN), and when v<18m / s, compensation of the damper opening (Δθ) is performed.

[0092] Specifically, in the ADP load optimization control method for the pulverizing system of the present invention, when the pressure inside the pipeline is monitored... At that time, the system will perform negative compensation for the feeder speed (i.e., reduce the speed). When the pulverized coal flow rate At that time, the system will compensate for the windshield opening (i.e., increase the opening). This step, through real-time monitoring and dynamic response mechanisms, ensures the safe and stable operation of the pulverizing system under complex working conditions.

[0093] Furthermore, this step involves real-time data acquisition of pipeline pressure and pulverized coal flow rate using a distributed sensor network, with a data sampling frequency of [missing information]. This ensures the system's ability to dynamically sense key parameters. When the pressure sensor detects... When the system determines that the pressure of coal powder accumulation in the pipeline is too high, which may cause safety risks such as vibration, leakage, or even explosion, it immediately triggers the negative compensation mechanism of the coal feeder speed, reducing the speed. This reduces the amount of pulverized coal supplied, thereby relieving pipeline pressure. When the pulverized coal flow rate sensor detects... If the system determines that the pulverized coal conveying capacity is insufficient and there is a risk of pulverized coal accumulation or blockage, it will automatically adjust the opening of the primary air damper. By increasing the air volume, the carrying capacity of pulverized coal is improved, ensuring the continuity and stability of the transportation process.

[0094] Furthermore, pressure threshold This is a safety threshold value set based on pipeline design pressure and historical operating data; exceeding this value triggers a compensation mechanism. Flow velocity threshold. It is determined based on the balance point between pulverized coal conveying efficiency and blockage risk; a value below this indicates insufficient conveying capacity. Compensation Amount and The adjustment range needs to be dynamically calculated based on the current system state and control objectives. A proportional-integral (PI) control strategy is usually adopted, combined with the output of the ADHDP controller for fine-tuning to ensure the smoothness of the compensation process and the stability of the system.

[0095] Furthermore, this step is widely used in the pulverizing systems of coal-fired power generating units, especially under complex operating conditions such as variable load operation, coal quality fluctuations, or equipment aging. For example, during low-load operation, the pulverized coal flow rate tends to decrease; the system can address this by increasing the opening of the primary air damper. To maintain conveying capacity; while under high load or when the risk of pulverized coal accumulation is high, the system reduces the feeder speed. This mechanism controls pressure and prevents the system from operating under overpressure. It is applicable to different coal types (such as bituminous coal and lignite) and different coal mill models, and has good versatility and adaptability.

[0096] Furthermore, this step, through a dual-parameter collaborative monitoring and graded compensation strategy, effectively avoids misjudgments or blind spots that might occur with single-parameter control. Experimental data shows that this mechanism can control pipeline pressure within a safe range, preventing equipment failures caused by overpressure; simultaneously, by maintaining the pulverized coal flow rate at... The above significantly reduces the risk of transport congestion and improves the continuity and safety of system operation. This step, as the core control logic of the safety constraint module, provides crucial real-time response and safety assurance capabilities for the entire ADP load optimization control method.

[0097] The load optimization control method for the pulverizing system of this invention effectively solves the problems of time delay response and multivariable coupling in the load adjustment of the pulverizing system, improves control accuracy and dynamic response speed, and achieves stable and efficient operation under a wide load range.

[0098] Example 2 The system flow of this invention is as follows: Figure 2 The system employs a dual-track collaborative design of "data flow-control flow," achieving load optimization of the pulverizing system through deep coupling of real-time data acquisition and dynamic closed-loop control. The overall process uses a 1 Hz sampling frequency as the time base, integrating innovative mechanisms such as time delay compensation and segmented optimization. Its core follows the state equation X(k+1) = A·X(k) + B·U(k) + C·D(k) + ε (where X is the system state vector, U is the control input, D is the disturbance term, and ε is the random error), forming a complete control closed loop of "perception-decision-execution-feedback." Through deep collaboration between "data flow and control flow," the system can respond quickly to load changes (e.g., the adjustment time for matching coal mill output to boiler demand is <5 minutes) and control the fineness of pulverized coal within the economic range (deviation ≤ ±2%), significantly improving the load adaptability and operational economy of the pulverizing system.

[0099] In one embodiment of the present invention, the data acquisition and preprocessing module includes: This data acquisition and preprocessing module is a fundamental component of the present invention, and its core function is to provide high-quality data input for subsequent model training and optimization decisions. This module ensures the integrity, accuracy, and effectiveness of the data through data acquisition strategies and a systematic preprocessing process, laying the foundation for the generalization ability and optimization accuracy of the control algorithm.

[0100] Furthermore, the multi-source data acquisition system includes: to achieve a comprehensive characterization of the complex operating conditions of the pulverizing system, the module adopts a multi-unit, multi-condition collaborative acquisition strategy, covering different coal types such as bituminous coal and lignite, as well as typical operating scenarios such as variable load and variable coal quality. Acquisition equipment includes acoustic sensors, electrostatic sensors, flow detection devices, and PLC controllers, etc., achieving real-time monitoring of key parameters through distributed deployment. The specific acquired parameters can be summarized into a 6-dimensional parameter matrix, the composition of which is as follows: Parameters of the air-coal system: primary air flow rate, hot air damper opening, cold air damper opening, and coal mill outlet temperature; Fuel characteristic parameters: coal ash content, volatile matter, moisture content on a dry basis, and moisture content of raw coal; Equipment operating parameters: coal feeder speed, coal mill current, oil pump pressure, valve status; Pulverized coal characteristic parameters: pulverized coal concentration, flow rate, particle charge signal intensity, and flow acoustic signal; Combustion status parameters: in-furnace flame detection signal, flue gas NOx concentration, CO concentration, and fly ash carbon content; Unit load parameters: generator power, main steam pressure, number of pulverizing systems in operation, and real-time total coal consumption.

[0101] Further, data preprocessing includes: Outliers: Extreme outliers are identified and removed using the IQR (interquartile range) method. By calculating the upper quartile (Q3) and lower quartile (Q1) of the parameter data, data exceeding the range [Q1-1.5×IQR, Q3+1.5×IQR] are marked as outliers and interpolated for correction. Feature extraction: For unstructured signals (such as coal powder flow acoustic signals and particle charged signals), spectral decomposition technology is used to convert the original time-domain signals into frequency-domain features. Wavelet transform is used to extract key indicators such as energy spectral density and peak frequency in the 10-2000Hz frequency band, achieving quantitative characterization of signal features. Data standardization and feature selection: The cleaned data is Z-score normalized to uniformly map different dimensional parameters to the [-1,1] interval. Redundant features weakly correlated with the load optimization objective are removed through Pearson correlation coefficient analysis and random forest feature importance assessment.

[0102] Furthermore, this module employs a multi-unit, cross-condition data fusion strategy to collect operational data from over 10 different types of coal mills under bituminous coal (volatile matter 15%-30%) and lignite (moisture 25%-40%) conditions, constructing a comprehensive dataset containing over 100,000 samples. Data partitioning utilizes a sliding time window method (window length 60 sampling periods), allocating the training set (historical samples with completed temperature control) and the test set (real-time samples requiring optimized control) in a 7:3 ratio to ensure the model's adaptability to varying coal types and load scenarios.

[0103] In one embodiment of the present invention, the ADHDP core control module includes: The ADHDP core control module, centered on an "intelligent decision-making hub," achieves adaptive optimization control of the pulverizing system load through several key technological innovations. Its technical implementation path mainly includes three levels: ADHDP selection strategy, customized network parameter design, and dynamic γ coefficient optimization. Furthermore, systematic network training ensures control accuracy and robustness.

[0104] This invention employs a two-layer nested neural network architecture to implement ADHDP control. The network training uses a reinforcement learning mechanism with 5000 iterations, and the modeling error ≤ ±2% is used as the convergence criterion. During training, the system continuously collects 12 key parameters, including the coal mill current, outlet temperature, and coal powder fineness. Through the dynamic interaction between the evaluation network and the control network, the control strategy weight matrix is ​​continuously corrected until the deviation between the model's predicted output and the actual system response stabilizes within the threshold range.

[0105] A two-layer hidden layer structure is used, with 32 and 16 neurons respectively. The activation function chosen is ELU (Exponential Linear Unit) to address the neuron death problem associated with ReLU. The loss function is:

[0106] The discount factor γ employs a dynamic adjustment mechanism: when the load fluctuation ΔQ > 5%, γ = 0.95 (enhancing the weight of the current action); when ΔQ < 2%, γ = 0.8 (emphasizing long-term returns). A three-layer hidden layer (64 / 32 / 16 neurons) is designed, and the output layer uses a Sigmoid activation function to normalize the control. The optimizer is Adam, with an initial learning rate of 0.001, which decays to 10% of the current value every 5000 iterations to ensure rapid convergence in the early stages and fine-tuning in the later stages.

[0107] This module's ability to handle multivariate coupling is reflected in its optimal decoupling framework based on adaptive dynamic programming. By constructing an evaluation neural network to estimate the optimal performance index function J*, and then feeding its estimate into the calculation of the control variable u(t) in real time, the evaluation network can be updated online and the control strategy can be adjusted adaptively. This mechanism can effectively decouple strongly coupled variables such as coal grinding rate, ventilation rate, and hot air temperature in the pulverizing system, avoiding the regulation lag or overshoot problems caused by dependent variable correlation in traditional PID control. Especially in scenarios with sudden load changes (such as a 30%~100% load step change), the dynamic adjustment time can be shortened by more than 40%.

[0108] In one embodiment of the present invention, the time delay compensation module includes: This module is a core component of the present invention, and its core function is to eliminate the effects of time delay by dynamically correcting control commands to improve the timeliness and accuracy of system response. To achieve this goal, the module innovatively adopts a fusion architecture of Smith predictor and ADHDP. In the specific implementation process, the module first predicts the system state X(k+1) at a future time based on the current system state and control commands using the Smith predictor; subsequently, the ADHDP controller uses the deviation between the predicted state and the actual feedback information to dynamically correct the current control command U(k), forming a closed-loop control logic of "prediction-correction-execution". This mechanism can effectively offset the time delay effect caused by factors such as material transmission and equipment inertia in the pulverizing system, ensuring that the actual effect of the control command is consistent with the theoretical expectation. From the perspective of compensation effect, experimental data shows that the module can strictly control the time delay compensation error within ±3%, ensuring the stability of control accuracy. Compared with traditional control methods that lack a dedicated time delay compensation mechanism, this module significantly improves the dynamic response performance of the system: under typical load disturbance scenarios, the response lag time caused by time delay in traditional control can reach 15-20 seconds, while after integrating the time delay compensation module, the response lag time is shortened to less than 3 seconds, and the response speed is improved by about 80%. This effectively avoids problems such as overshoot and oscillation caused by lag, and provides key technical support for the stable and efficient operation of the pulverizing system.

[0109] In one embodiment of the present invention, the safety constraint module includes: This module innovatively adopts a dual-parameter collaborative monitoring mechanism to simultaneously collect the pressure (P) and pulverized coal flow velocity (v) within the pipeline, forming a safety protection network and effectively avoiding misjudgments or blind spots that may result from single-parameter monitoring. Dual-parameter monitoring and graded compensation strategy: The module collects the pressure (P) and pulverized coal flow velocity (v) within the pipeline in real time through a pressure sensor and a flow meter, and triggers graded compensation control based on preset thresholds: Pressure anomaly control: When the pipeline pressure P > 1.2 kPa is detected, the system determines that it is approaching the pipeline safety limit and immediately executes negative compensation (reduction ΔN) of the coal feeder speed. By reducing the pulverized coal supply, the system reduces the material accumulation pressure within the pipeline, preventing pipeline vibration, leakage, or even bursting risks caused by overpressure. Flow velocity anomaly control: When the pulverized coal flow velocity v < 18 m / s, it is determined that the pulverized coal conveying capacity is insufficient, which may cause pulverized coal accumulation or blockage within the pipe. At this time, the system automatically compensates for the primary damper opening (Δθ), increasing the airflow to enhance the pulverized coal carrying capacity and ensure conveying stability.

[0110] In one embodiment of the present invention, the load segmentation adaptive module includes: The core function of the load segmentation adaptive module is to achieve wide load adaptation of the pulverizing system. It dynamically adjusts the control strategy to adapt to the operating characteristics of different load ranges, which is particularly important for optimizing low-load conditions under deep peak shaving scenarios. In engineering implementation, the module achieves precise control by establishing load segmentation logic. Referring to the output characteristics of the pulverizing system per unit time period, the load range is divided into high-load and low-load segments, with corresponding differentiated control parameters: the first output ratio is set to 90% (high-load range), and the second output ratio is set to 30% (low-load range), forming a tiered adaptation strategy. This segmentation logic not only considers the output limit of the pulverizing system but also combines the power consumption characteristics under different loads. By establishing a load scheduling model to assess the adjustable potential, it achieves segmented resource control, ensuring the consistency of control objectives within each range.

[0111] The embodiments of this invention also have the following technical effects: Improved grid response and peak-shaving capabilities: The dynamic optimization control algorithm reduces the response delay from 15 seconds to 8 seconds, meeting the rapid response requirements of the grid's AGC (Automatic Generation Control), reducing the load peak-valley difference by more than 15%, and the decoupling design of the coal mill and boiler enhances adaptability to grid fluctuations. Reduced energy consumption and improved operational economy: Combining permanent magnet motor direct drive and dynamic airflow adaptation, pulverizing unit consumption is reduced by 8%-12%, saving approximately 3 million yuan in annual coal costs for a 300MW unit. The small-capacity pulverizer combines the advantages of both types of pulverizing systems, reducing infrastructure and maintenance costs and improving pulverizing efficiency and control accuracy. Pulverizing efficiency: Through a stepped predictive controller and machine learning, efficiency is increased by 10%-15%, with air-coal concentration deviation ≤3%, coal powder fineness fluctuation ±2%, air-coal deviation ≤5%, and boiler efficiency improved by 0.5%-0.8%. Enhanced coal adaptability and environmental benefits: Lignite blending ratio increased to 30% (from 15%), coking rate reduced by 40%, water-cooled wall corrosion rate reduced by 0.1 mm / year, and nitrogen oxide emissions reduced by 10%-15 mg / Nm³. System stability and fault response capabilities: Mill inlet air temperature deviation reduced from ±8℃ to ±3℃, single mill failure output loss ≤5%, cold start success rate increased from 85% to 99%, and warm mill outlet temperature accuracy ±5℃.

[0112] Example 3 To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides an ADP load optimization control device 10 for a pulverizing system, comprising: The data acquisition and preprocessing module 100 is used to collect the operating parameters of the flour milling system in real time, perform outlier removal and feature extraction, and generate a standardized dataset. The Smith predictor and AHDP control module 200 is used to predict the system state based on the Smith predictor and combine it with the AHDP controller to dynamically correct the control commands, forming a time-delay compensated closed-loop control logic. The dual-layer neural network optimization module 300 is used to update the evaluation network online through a dual-layer neural network architecture, thereby achieving decoupled optimization control of multiple variables such as coal grinding rate, ventilation rate, and temperature. The execution and compensation monitoring module 400 is used to execute optimized control commands and simultaneously monitor pipeline pressure and pulverized coal flow rate, and trigger a graded compensation strategy according to preset thresholds to maintain system stability.

[0113] Furthermore, the data acquisition and preprocessing module 100 is also used for: The IQR method is used to identify and remove outliers by calculating the upper and lower quartiles of the parameter data. Data within the specified range is marked as abnormal and interpolated for correction. For unstructured signals, wavelet transform is used to extract key indicators such as energy spectral density and peak frequency in the 10-2000Hz frequency band to achieve quantitative characterization of signal features.

[0114] Furthermore, the Smith prediction and AHDP control module 200 is also used for: The Smith predictor forecasts the future system state based on the current system state and control commands. ; Dynamically correct the current control command by utilizing the deviation between the predicted state and the actual feedback information. This forms a closed-loop control logic of "prediction-correction-execution".

[0115] An embodiment of the present invention provides an ADP load optimization control device for a pulverizing system, which effectively solves the time delay effect and multivariate coupling problem of the pulverizing system under sudden load changes, significantly improves dynamic response speed and control accuracy, shortens the response time to less than 3 seconds, and controls the coal powder fineness deviation to within ±2%.

[0116] Example 4 To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the ADP load optimization control method for a pulverizing system described above.

[0117] Example 5 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an ADP load optimization control method for a pulverizing system as described in the foregoing embodiments.

[0118] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0119] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for optimizing and controlling the ADP load of a milling system, characterized in that, include: S1: Real-time acquisition of the operating parameters of the flour milling system, outlier removal and feature extraction, and generation of standardized dataset; S2, based on the Smith predictor to predict the system state and combined with the ADHP controller to dynamically correct the control command, forms a time-delay compensation closed-loop control logic. S3 uses a two-layer neural network architecture to update the evaluation network online, achieving decoupled optimization control of multiple variables such as coal grinding volume, ventilation volume, and temperature; S4 executes optimized control commands and simultaneously monitors pipeline pressure and pulverized coal flow rate, triggering a graded compensation strategy based on preset thresholds to maintain system stability.

2. The method as described in claim 1, characterized in that, S1 includes: S11. The IQR method is used to identify and remove outliers. By calculating the upper and lower quartiles of the parameter data, data that exceed the range of [Q1-1.5×IQR, Q3+1.5×IQR] are marked as outliers and interpolated for correction. S12 uses wavelet transform to extract key indicators such as energy spectral density and peak frequency in the 10-2000Hz frequency band for unstructured signals, thereby achieving quantitative characterization of signal features.

3. The method as described in claim 1, characterized in that, The S2 includes: S21, predict the system state X(k+1) at future time based on the current system state and control commands using the Smith predictor; S22 dynamically corrects the current control command U(k) by utilizing the deviation between the predicted state and the actual feedback information, forming a closed-loop control logic of "prediction-correction-execution".

4. The method as described in claim 1, characterized in that, The S3 includes: S31 employs a two-layer hidden layer structure with 32 and 16 neurons respectively. The activation function is ELU to address the neuron death problem of ReLU. S32, dynamically adjust the discount factor γ according to the load fluctuation amplitude ΔQ: when ΔQ>5%, γ=0.95; when ΔQ<2%, γ=0.

8.

5. The method as described in claim 1, characterized in that, The S4 includes: S41, the pressure P and pulverized coal flow velocity v in the pipeline are collected in real time through pressure sensors and flow meters; S42, when P>1.2kPa is detected, negative compensation for the coal feeder speed is performed, and when v<18m / s, compensation for the opening of the damper is performed.

6. An ADP load optimization control device for a milling system, characterized in that, include: The data acquisition and preprocessing module is used to collect the operating parameters of the flour milling system in real time, remove outliers and extract features, and generate a standardized dataset. The Smith predictor and AHDP control module is used to predict the system state based on the Smith predictor and combine it with the AHDP controller to dynamically correct the control commands, forming a time-delay compensated closed-loop control logic. The dual-layer neural network optimization module is used to update the evaluation network online through a dual-layer neural network architecture, thereby achieving decoupled optimization control of multiple variables such as coal grinding rate, ventilation rate, and temperature. The execution and compensation monitoring module is used to execute optimized control commands and simultaneously monitor pipeline pressure and pulverized coal flow rate, and trigger graded compensation strategies according to preset thresholds to maintain system stability.

7. The apparatus as claimed in claim 6, characterized in that, The data acquisition and preprocessing module is also used for: The IQR method is used to identify and remove outliers by calculating the upper and lower quartiles of the parameter data. Data within the specified range is marked as abnormal and interpolated for correction. For unstructured signals, wavelet transform is used to extract key indicators such as energy spectral density and peak frequency in the 10-2000Hz frequency band to achieve quantitative characterization of signal features.

8. The apparatus as claimed in claim 6, characterized in that, The Smith prediction and AHDP control module is also used for: The Smith predictor forecasts the future system state based on the current system state and control commands. ; Dynamically correct the current control command by utilizing the deviation between the predicted state and the actual feedback information. This forms a closed-loop control logic of "prediction-correction-execution".

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the ADP load optimization control method for a pulverizing system as described in any one of claims 1-5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements an ADP load optimization control method for a pulverizing system as described in any one of claims 1-5.

Images