Thermal power generating unit AGC performance index optimization method and system based on deep learning

Abstract

The invention relates to the technical field of thermal power generating unit automatic power generation control performance optimization, and discloses a thermal power generating unit AGC performance index optimization method and system based on deep learning, and the method comprises the steps: obtaining and processing AGC adjustment process time sequence data, and obtaining a load change rate; grading and parameter definition are carried out based on the load change rate, and an adjustment process grading result, a core performance index and a rapid adjustment process segmentation result are obtained; constructing a model based on grading and segmentation results, and learning a mapping relation to output a starting time point of a deceleration section under rapid adjustment; and carrying out data preprocessing and dynamic optimization training on the basis of model training requirements and parameter definition to finish accurate optimization of AGC performance indexes. According to the method, the technical problems of AGC performance index distortion, low model generalization precision and weak section recognition lagging caused by inaccurate segmentation, unscreened abnormal data and unassociated deceleration section starting point and rate grade of the existing model are solved.

Description

Technical Field

[0001] This invention relates to the field of automatic power generation control performance optimization technology for thermal power units, specifically a method and system for optimizing AGC performance indicators of thermal power units based on deep learning. Background Technology

[0002] Optimization of automatic generation control performance of thermal power units is an important technology. Specifically, it is applied to the precise calculation of response rate and regulation accuracy indicators in the AGC regulation process of thermal power units. The core is to improve the accuracy of indicator calculation by adapting to the segmented characteristics of the regulation process and data quality screening, so as to meet the stringent evaluation requirements of the power grid for the AGC response capability of the unit. Current AGC performance index calculation models for thermal power units struggle to adapt to the segmented dynamic characteristics and data quality differences of rapid regulation processes, leading to index distortion and insufficient model generalization accuracy. Existing AGC performance index calculation models fail to divide the regulation process into acceleration, constant speed, and deceleration phases based on load change rate. They include data from the deceleration phase, where synchronicity significantly decreases, in the overall response rate average calculation, resulting in a large deviation between the calculated response rate and the actual acceleration phase rate. Regulation accuracy calculations focus only on static deviation values, failing to recognize the dynamic trend of deviation convergence in the deceleration phase, thus underestimating regulation accuracy. Furthermore, the lack of correlation thresholds for load command values, actual load values, and valve opening values ​​leads to the inclusion of numerous latent anomalies with missing synchronicity in model training, reducing model generalization accuracy in rapid regulation scenarios. Additionally, the failure to establish a correlation between the deceleration phase start time and load change rate level results in the misjudgment and lag of weak dynamic intervals in rapid regulation. These combined problems prevent AGC performance indexes from accurately reflecting the unit's actual response capability, impacting the grid's accurate assessment of unit regulation performance and its judgment of unit optimization directions. Summary of the Invention

[0003] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for optimizing the AGC performance index of thermal power units based on deep learning, so as to solve the technical problems of AGC performance index distortion, low model generalization accuracy and lagging weak section identification caused by the lack of accurate segmentation, failure to screen abnormal data, and failure to associate the deceleration section start point with the speed level in the existing model.

[0004] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for optimizing the AGC performance index of thermal power units based on deep learning, comprising: The timing data of the AGC regulation process is acquired and processed to obtain the load change rate; Based on the load change rate, the classification and parameter definition are performed to obtain the classification results of the regulation process, core performance indicators and segmentation results of the rapid regulation process. A model is built based on the hierarchical and segmented results, and the mapping relationship is learned to output the starting time point of the deceleration segment under rapid adjustment. Based on the model training requirements and parameter definitions, data preprocessing and dynamic optimization training are performed to achieve precise optimization of AGC performance indicators.

[0005] Preferably, the process of acquiring and processing the timing data of the AGC regulation process to obtain the load change rate includes: real-time acquisition of the timing data of the load command value, actual load value, valve opening value and timestamp during the AGC regulation process, and real-time calculation of the load change rate by the derivative of the load command value and the timestamp. The process of calculating the load change rate in real time by taking the derivative of the load command value and the timestamp includes the following steps: A first-order difference operation is performed on the load command value sequence with a fixed time window, and the load change per unit time is calculated by combining the timestamp interval. Measurement noise is eliminated by moving average filtering to generate a smooth load change rate curve, which is used as the input benchmark for the three-level rate division.

[0006] Preferably, the specific process for classifying and defining parameters based on the load change rate to obtain the classification results of the regulation process, core performance indicators, and segmented results of the rapid regulation process includes: The regulation process is classified into three levels based on the load change rate, including slow, medium and fast. K1 is defined as the unit response rate index and K2 is defined as the regulation accuracy index. The timing data of the fast regulation process is divided into acceleration, constant speed and deceleration stages based on the load change rate. The starting point of the deceleration stage is the time point when the load change rate decreases from the peak. The regulation process is classified into slow, medium, and fast based on the load change rate. The classification is based on thresholds, including: slow corresponds to a load change rate not exceeding the first rate threshold, medium corresponds to a load change rate between the first and second rate thresholds, and fast corresponds to a load change rate higher than the second rate threshold. The first and second rate thresholds are dynamically calibrated based on the statistical quantiles of the unit's historical regulation data.

[0007] Furthermore, the definitions of K1 and K2 specifically include using the average actual load change rate of the acceleration and constant speed sections as the unit response rate index K1, and using the integral average of the absolute value of the deviation between the actual load value and the command value during the entire adjustment process as the adjustment accuracy index K2; among them, the calculation of K1 needs to exclude the deceleration section data, and the calculation of K2 needs to distinguish the dynamic weight.

[0008] Furthermore, based on the load change rate, the time series data of the rapid adjustment process is divided into acceleration, constant speed, and deceleration segments, specifically including: The extreme point of the load change rate curve is identified as the starting point of the deceleration section. The interval in which the load change rate continues to rise before this starting point is the acceleration section. The plateau interval in which the load change rate fluctuates within a set threshold is the uniform speed section. The interval in which the load change rate monotonically decreases to zero is the deceleration section.

[0009] Preferably, the model is constructed based on the hierarchical and segmented results, and the mapping relationship is learned to output the specific process at the starting time point of the rapid adjustment deceleration segment, including: A bidirectional long short-term memory network is constructed. The network learns the functional mapping relationship between the load change rate level and the starting point of the deceleration segment, and outputs the starting time point of the deceleration segment during the rapid adjustment process. The construction and training process of the bidirectional long short-term memory network includes: The system takes the load change rate level encoding vector and the historical deceleration segment start time series as inputs, and controls the memory strength of the temporal features through forget gate and input gate. The output layer uses a fully connected network to map to the predicted timestamp of the deceleration segment start. The network training uses the mean square error loss function with time alignment constraints.

[0010] Furthermore, the functional mapping relationship between the load change rate level and the deceleration phase start point specifically includes: concatenating the load change rate level encoding vector with the load command change pattern of the previous N seconds of the current adjustment cycle, extracting the positive and negative temporal dependency features through the hidden layer of the bidirectional long short-term memory network, focusing on the rate drop threshold by the attention mechanism, and finally generating the absolute time offset of the deceleration phase start point through the output layer. This offset is then superimposed on the start timestamp of the current adjustment cycle to obtain the deceleration phase start point.

[0011] Preferably, based on model training requirements and parameter definitions, the specific processes for data preprocessing and dynamic optimization training include: Before model training, samples with missing synchronization are removed. Based on the predefined correlation coefficient threshold, samples with missing synchronization of load command value, actual load value and valve opening value are removed. A dynamic compensation module is constructed using a gated loop unit to optimize the nonlinear characteristics of the deceleration section in the time series data of the rapid adjustment process. Among them, the samples with missing synchronization between load command value, actual load value and valve opening value are specifically included in the range where the load change rate is greater than the second rate threshold and the Pearson correlation coefficient between the load command value and the valve opening value is lower than the predefined threshold. These samples are characterized by the valve opening value changing lagging behind the load command value.

[0012] Furthermore, in the optimization training of the nonlinear characteristics of the deceleration phase in the time series data of the rapid adjustment process, the K1 index is calculated by selecting only the load command value, actual load value and valve opening value of the acceleration phase and the constant speed phase time series data and extracting highly correlated dynamic features through a one-dimensional convolutional neural network to output the real unit response rate. In the calculation of the K1 index, the input of the one-dimensional convolutional neural network is a data pair of load command value, actual load value and valve opening value sampled at equal intervals in the acceleration section and the constant speed section. The convolutional kernel slides along the time axis to extract local dynamic coupling features, the pooling layer filters the first relevant feature fragment, and the fully connected layer regresses to output the real unit response rate value. During training, the weight of the deceleration section data is fixed at zero. For the K2 index calculation, weights are dynamically allocated based on the deviation convergence trend during the deceleration phase, specifically including: During the deceleration phase, the first derivative of the deviation convergence trend is used as the basis for adjusting the weight coefficients. When the absolute value of the deviation decreases with time, the first weight is assigned, and when the deviation fluctuation increases or fails to converge, the second weight is assigned. The weight coefficients and timestamps are jointly input into the gated loop unit to generate the weighted cumulative value of the deviation integral.

[0013] Secondly, the present invention also provides a deep learning-based AGC performance index optimization system for thermal power units, comprising: The data processing module is used to acquire and process the timing data of the AGC regulation process to obtain the load change rate; The classification definition module is used to classify and define parameters based on the load change rate, and obtain the classification results of the regulation process, core performance indicators and segmentation results of the rapid regulation process. The mapping output module is used to build a model based on the hierarchical and segmented results, learn the mapping relationship, and output the starting time point of the deceleration segment under rapid adjustment. The optimization calculation module is used to perform data preprocessing and dynamic optimization training based on model training requirements and parameter definitions, so as to achieve accurate optimization of AGC performance indicators.

[0014] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a deep learning-based method for optimizing the AGC performance index of thermal power units. The method divides the AGC adjustment process into three levels—slow, medium, and fast—based on the load change rate. For the fast adjustment process, the time-series data is further subdivided into acceleration, constant speed, and deceleration segments. A bidirectional long short-term memory network is used to accurately locate the starting point of the deceleration segment to avoid lag identification. Before model training, abnormal samples with missing synchronization of load command values, actual load values, and valve opening values ​​are removed based on correlation coefficient thresholds. A one-dimensional convolutional neural network is then used to extract the dynamic coupling features of the acceleration and constant speed segments to calculate the response rate index. A gated recurrent unit is used to dynamically allocate weights based on the deceleration segment deviation convergence trend to calculate the adjustment accuracy index. Overall, this method significantly reduces the deviation between the response rate and the actual unit response rate, accurately reflects the adjustment accuracy without underestimation, improves the model's generalization accuracy in fast adjustment scenarios, and eliminates lag identification in weak dynamic intervals. This effectively solves the problems of existing models in the background technology, which suffer from distorted AGC performance indexes, low model generalization accuracy, and lag identification of weak intervals due to inaccurate segmentation, failure to filter abnormal data, and failure to correlate the starting point of the deceleration segment with the rate level. Attached Figure Description

[0015] Figure 1 This is a flowchart of the AGC performance index optimization method for thermal power units in an embodiment of the present invention; Figure 2 This is a schematic diagram of the thermal power unit AGC performance index optimization system in an embodiment of the present invention; In the diagram: 1. Data processing module; 2. Hierarchical definition module; 3. Mapping output module; 4. Optimization calculation module. Detailed Implementation

[0016] 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.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] The purpose of this invention is to provide a method and system for optimizing the AGC performance index of thermal power units based on deep learning, so as to solve the technical problems of existing models, such as distorted AGC performance index, low model generalization accuracy and lagging weak zone identification, due to inaccurate segmentation, failure to screen abnormal data, and failure to associate the deceleration section start point with the speed level.

[0019] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 See Figure 1 In one embodiment of the present invention, a method for optimizing the AGC performance index of thermal power units based on deep learning is provided, specifically including: Step 1: Obtain and process the timing data of the AGC regulation process to obtain the load change rate; Specifically, the process includes: real-time acquisition of time-series data such as load command value, actual load value, valve opening value, and timestamp during the AGC regulation process; and real-time calculation of the load change rate by the derivative of the load command value and the timestamp. The process of calculating the load change rate in real time by taking the derivative of the load command value and the timestamp includes the following steps: A first-order difference operation is performed on the load command value sequence with a fixed time window, and the load change per unit time is calculated by combining the timestamp interval. Measurement noise is eliminated by moving average filtering to generate a smooth load change rate curve, which is used as the input benchmark for the three-level rate division.

[0020] Step 2: Classify and define parameters based on the load change rate to obtain the classification results of the regulation process, core performance indicators, and segmentation results of the rapid regulation process; Specifically, the specific process includes: The regulation process is classified into three levels based on the load change rate, including slow, medium and fast. K1 is defined as the unit response rate index and K2 is defined as the regulation accuracy index. The timing data of the fast regulation process is divided into acceleration, constant speed and deceleration stages based on the load change rate. The starting point of the deceleration stage is the time point when the load change rate decreases from the peak. The regulation process is classified into slow, medium, and fast based on the load change rate. The classification is based on thresholds, including: slow corresponds to a load change rate not exceeding the first rate threshold, medium corresponds to a load change rate between the first and second rate thresholds, and fast corresponds to a load change rate higher than the second rate threshold. The first and second rate thresholds are dynamically calibrated based on the statistical quantiles of the unit's historical regulation data.

[0021] The average actual load change rate during the acceleration and constant speed phases is used as the unit response rate index, denoted as K1. The integral average of the absolute value of the deviation between the actual load value and the command value during the entire adjustment process is used as the adjustment accuracy index, denoted as K2. K1 calculation needs to exclude deceleration phase data, and K2 calculation needs to distinguish dynamic weights.

[0022] For the time-series data of the rapid adjustment process, it is divided into acceleration, constant speed, and deceleration segments based on the load change rate, specifically including: The extreme point of the load change rate curve is identified as the starting point of the deceleration section. The interval in which the load change rate continues to rise before this starting point is the acceleration section. The plateau interval in which the load change rate fluctuates within a set threshold is the uniform speed section. The interval in which the load change rate monotonically decreases to zero is the deceleration section.

[0023] The bidirectional long short-term memory network takes the load change rate level encoding vector and the historical deceleration phase start time series as inputs. The memory strength of the temporal features is controlled by the forget gate and the input gate. The output layer is mapped to the predicted timestamp of the deceleration phase start using a fully connected network. The network training uses the mean square error loss function with time alignment constraints.

[0024] Step 3: Build a model based on the hierarchical and segmented results, and learn the mapping relationship to output the starting time point of the deceleration segment under rapid adjustment; The specific process includes: A bidirectional long short-term memory network is constructed. The network learns the functional mapping relationship between the load change rate level and the starting point of the deceleration segment, and outputs the starting time point of the deceleration segment during the rapid adjustment process. The construction and training process of the bidirectional long short-term memory network includes: The system takes the load change rate level encoding vector and the historical deceleration segment start time series as inputs, and controls the memory strength of the temporal features through forget gate and input gate. The output layer uses a fully connected network to map to the predicted timestamp of the deceleration segment start. The network training uses the mean square error loss function with time alignment constraints.

[0025] The functional mapping relationship between the load change rate level and the deceleration phase start point specifically includes: concatenating the load change rate level encoding vector with the load command change pattern of the previous N seconds of the current adjustment cycle, extracting the positive and negative temporal dependency features through the hidden layer of the bidirectional long short-term memory network, focusing on the rate drop threshold by the attention mechanism, and finally generating the absolute time offset of the deceleration phase start point through the output layer. This offset is then superimposed on the start timestamp of the current adjustment cycle to obtain the deceleration phase start point.

[0026] Step 4: Based on the model training requirements and parameter definitions, perform data preprocessing and dynamic optimization training to achieve precise optimization of AGC performance indicators.

[0027] Specifically, the specific process includes: Before model training, samples with missing synchronization are removed. Based on the predefined correlation coefficient threshold, samples with missing synchronization of load command value, actual load value and valve opening value are removed. A dynamic compensation module is constructed using a gated loop unit to optimize the nonlinear characteristics of the deceleration section in the time series data of the rapid adjustment process. Among them, the samples with missing synchronization between load command value, actual load value and valve opening value are specifically included in the range where the load change rate is greater than the second rate threshold and the Pearson correlation coefficient between the load command value and the valve opening value is lower than the predefined threshold. These samples are characterized by the valve opening value changing lagging behind the load command value.

[0028] In the optimization training of the nonlinear characteristics of the deceleration section in the time series data of the rapid adjustment process, the K1 index is calculated by selecting only the load command value, actual load value and valve opening value of the acceleration section and the constant speed section. The time series data are then processed by a one-dimensional convolutional neural network to extract highly correlated dynamic features and output the real unit response rate. In the calculation of the K1 index, the input of the one-dimensional convolutional neural network is a data pair of load command value, actual load value and valve opening value sampled at equal intervals in the acceleration section and the constant speed section. The convolutional kernel slides along the time axis to extract local dynamic coupling features, the pooling layer filters the first relevant feature fragment, and the fully connected layer regresses to output the real unit response rate value. During training, the weight of the deceleration section data is fixed at zero. For the K2 index calculation, weights are dynamically allocated based on the deviation convergence trend during the deceleration phase, specifically including: During the deceleration phase, the first derivative of the deviation convergence trend is used as the basis for adjusting the weight coefficients. When the absolute value of the deviation decreases with time, the first weight is assigned, and when the deviation fluctuation increases or fails to converge, the second weight is assigned. The weight coefficients and timestamps are jointly input into the gated loop unit to generate the weighted cumulative value of the deviation integral.

[0029] In this embodiment, after real-time acquisition of the load command value, actual load value, valve opening value, and corresponding timestamp during the AGC regulation process, in order to accurately classify the regulation rate level, it is necessary to first calculate the load change rate in real time based on the derivative of the load command value and the timestamp. The core is to transform the discrete load command sequence into a continuous and discernible rate curve through a standardized data processing flow, avoiding calculation deviations caused by noise interference and data discreteness. The specific implementation method is as follows: The first step in the calculation is to set a fixed time window and perform a first-order difference operation. A fixed time window refers to selecting a uniform time interval as the calculation unit for load changes. Considering the response characteristics of AGC regulation in thermal power units, this time window is set to 1 second. This covers the instantaneous fluctuations of rapid regulation while avoiding calculation distortion caused by excessively sparse data in slow regulation. In practice, the continuously collected load command value sequence is first segmented into 1-second time windows. The average value of multiple load command values ​​within each time window is then taken to reduce random errors at individual data points, forming a load command average value sequence arranged in chronological order. Subsequently, this average value is... The value sequence undergoes a first-order difference operation. This operation involves subtracting the average load command of the previous time window from the average load command of the subsequent time window to obtain the load change within two adjacent time windows. Simultaneously, the timestamps corresponding to these two time windows are extracted, and the time interval between them, i.e., the timestamp difference, is calculated. Since the time window is fixed at 1 second, the normal interval is 1 second. If there is a data acquisition delay, the actual timestamp difference is used. Then, the calculated load change is divided by the corresponding timestamp interval to obtain the load change per unit time. This value is the initial load change rate data. However, due to… Sensor measurement noise is unavoidable during data acquisition, causing irregular fluctuations in the initial rate data. Directly using this data for rate classification can lead to misjudgments. Therefore, moving average filtering is necessary to eliminate this measurement noise. Moving average filtering is a noise reduction method that smooths local fluctuations by averaging multiple consecutive data points. Here, the filtering window size is set to 5 time windows, or 5 seconds. Specifically, for each initial rate data point, the arithmetic mean of the five adjacent rate data points (two before and two after) is calculated. This average value is then used to replace the initial rate value of the original center data point. This process is repeated point by point in chronological order. By iterating through all initial rate data and completing the full sequence filtering process, the original discrete and noisy load command sequence is transformed into a smooth and continuous load change rate curve through the complete process of fixed time window segmentation, first-order difference operation, time interval conversion, and moving average filtering. This curve can truly reflect the dynamic change trend of the load during AGC regulation, retaining the rate peak during fast regulation and the stable characteristics during slow regulation, while eliminating meaningless noise interference. Thus, it serves as the input benchmark for the subsequent classification of slow, medium, and fast rates, providing a reliable guarantee for the accuracy and stability of the classification.

[0030] After obtaining a smooth load change rate curve, the AGC regulation process needs to be divided into three levels: slow, medium, and fast, based on this curve. If a fixed threshold is used, the division will be biased due to different unit operating conditions. Therefore, this method dynamically calibrates the threshold using the statistical quantiles of the unit's historical regulation data to ensure that the division is adapted to the actual regulation characteristics of the unit. The specific implementation method is as follows: The rate ranges for slow, medium, and fast are clearly defined. Slow rate corresponds to a load change rate not exceeding the first rate threshold; that is, all intervals in the rate curve below this threshold are classified as slow regulation. This type of regulation often occurs during small load adjustments when the unit is operating stably. Medium rate corresponds to a load change rate between the first and second rate thresholds, commonly seen in medium-amplitude load changes in the unit. Fast rate corresponds to a load change rate exceeding the second rate threshold, mainly occurring when there are sudden changes in grid load demand. In this case, the unit needs to respond fully to the command. The core lies in the dynamic calibration of the first and second rate thresholds. The calibration is based on the statistical quantiles of the unit's historical regulation data. Statistical quantiles divide a set of ordered data into multiple equally probable parts. Here, the 25th percentile is chosen as the first rate threshold, and the 75th percentile as the second rate threshold. This selection ensures that approximately 25% of the rate values ​​in the historical data fall into the slow range, 50% into the medium range, and 25% into the fast range, conforming to the balance and practicality of the three-level classification. The specific calibration process is as follows: Historical AGC (Automatic Control) data of the generating unit over the past three months is collected, covering regulation scenarios in different load ranges and seasons. Abnormal data is removed, as these data rates far exceed the normal range. Typically, a rate exceeding three times the historical average is considered abnormal. The remaining valid data are then sorted by load change rate from smallest to largest. Statistical tools are used to calculate the 25th and 75th percentiles of this sequence, which are used as the first and second rate thresholds for the current period, respectively. To adapt to changes in generating unit performance, a monthly threshold calibration cycle is set. Each calibration recalculates the quantiles and updates the thresholds based on the latest three months of historical data, ensuring that the thresholds always match the unit's current regulation capacity.

[0031] After completing the classification of regulation rate levels, the AGC regulation performance needs to be evaluated through quantitative indicators. The unit response rate indicator K1 focuses on regulation speed, reflecting the unit's ability to respond quickly to load commands. The regulation accuracy indicator K2 focuses on accuracy, reflecting the degree of deviation between the actual load and the command. Both need to be calculated in conjunction with the segmented characteristics of the regulation process to avoid interference from irrelevant data. The specific implementation method is as follows: The unit response rate index K1 is calculated based on the average actual load change rate during the acceleration and constant speed phases. The key to this index is excluding data from the deceleration phase, as this phase represents the unit's speed reduction after completing a load command and cannot reflect the unit's true response capability to the load command. The specific calculation process is as follows: First, the time intervals for the acceleration and constant speed phases are extracted from the predefined adjustment segments. For example, in the time series data of a rapid adjustment process, the acceleration phase is 0-10 seconds, the constant speed phase is 10-25 seconds, and the deceleration phase is 25-35 seconds. Therefore, only the actual load change rate data for the 0-25 second period is taken. The actual load change rate is calculated based on the collected actual load value and its corresponding timestamp, which differs from the rate based on the load command value and better reflects the actual operating status of the unit. The calculation method is consistent with the load change rate, using a fixed time window, first-order difference, and moving average filtering. Then, the arithmetic mean of all actual load change rate data points within the extracted acceleration and constant speed phases is taken. The average value is K1. For example, if the acceleration speed data is 3MW / min, 4MW / min, and 5MW / min, and the constant speed speed data is 5MW / min and 5MW / min, the average value is (3+4+5+5+5) / 5 = 4.4MW / min, i.e., K1 = 4.4MW / min. The larger the K1 value, the faster the unit responds to the load command. The regulation accuracy index K2 is calculated as the integral average of the absolute values ​​of the deviations between the actual load value and the command value throughout the entire regulation process. Dynamic weights need to be distinguished because the deviations in different regulation stages have different effects on the overall accuracy. Fixed weights will mask the deviation problems in critical stages. The specific calculation process is as follows: First, the total time range of the adjustment process is determined, from the start of the load command change until the actual load stabilizes within ±1% of the command value. Within this time range, the absolute value of the deviation between the actual load value and the load command value is calculated every second. Then, the absolute value of the deviation at all time points is integrated over time, that is, the product of the absolute value of the deviation and the time interval is accumulated in chronological order, with a time interval of 1 second. The integration result reflects the total cumulative deviation of the entire adjustment process. The integration result is divided by the total adjustment time to obtain the integral average value of the absolute value of the deviation, which is the base value of K2. At the same time, it is clear that the calculation of K2 needs to distinguish dynamic weights. Subsequently, different weights will be assigned to different stages based on the deviation convergence trend to avoid excessive deviation or fluctuations in the deceleration phase from misleading K2, ensuring that K2 can truly reflect the overall accuracy of the unit's adjustment. The smaller the K2 value, the smaller the deviation between the actual load and the command, and the higher the adjustment accuracy.

[0032] After completing the three-level rate division of the AGC regulation process, for the time-series data of the rapid regulation process, i.e., when the load change rate is higher than the second rate threshold, its internal regulation characteristics exhibit significant stage differences. The acceleration stage reflects the unit's ability to increase speed in response to commands, the constant speed stage reflects the stable response state, and the deceleration stage is the deceleration process after the load approaches the command value. If the data of each stage is confused, it will lead to the calculation deviation of the K1 and K2 indicators. Therefore, it is necessary to accurately divide the three intervals based on the load change rate curve. The specific implementation method is as follows: The core first step in the classification is to identify the extreme points of the load change rate curve as the starting point of the deceleration phase. An extreme point is the turning point in the rate curve where the trend changes from "upward" to "downward," and it is also the time point when the rate reaches its peak. For example, if the rate increases from 5 MW / min to 8 MW / min and then first drops to 7.8 MW / min, the time point corresponding to this 8 MW / min is the extreme point. The specific identification process is as follows: A 3-second sliding window is used to iterate through the smoothed load change rate curve. The rate difference between two adjacent time points within each window is calculated. When the rate difference between two consecutive windows changes from a "positive value" (the rate at the previous time point is less than the rate at the next time point, and the curve rises) to a "negative value" (the rate at the previous time point is greater than the rate at the next time point, and the curve falls), and the absolute value of the rate difference in the second window is greater than that in the first window, it indicates that the downward trend has been established. The starting time point of the second window is then determined as the extreme point, i.e., the starting point of the deceleration phase. To avoid misjudgment due to small fluctuations, such as if the rate drops from 8 MW / min to 7.9 MW / min and then rises again to 8.1 MW / min, it is necessary to additionally verify whether the rate continues to decrease within 3 seconds after the extreme point. If it continues to decrease, the starting point is confirmed; otherwise, the iteration is repeated to ensure the accuracy of extreme point identification.

[0033] Using this extreme point as a boundary, the interval in which the load change rate continuously increases before it is the acceleration phase. Continuous increase means that from the start of the adjustment process to the extreme point, the rate difference between all adjacent time points is positive. A single time point difference is allowed to be zero, but the difference between two consecutive time points must not be negative, and the overall rate must increase from the initial value, such as 0 MW / min at the start of adjustment, to the peak value (extreme point rate). Specifically, starting from the start of adjustment, the rate difference with the next time point is calculated point by point. If the consecutive differences are all non-negative and eventually point to the extreme point... If the adjustment starts at 0 seconds with a rate of 0 MW / min, 1 second with a rate of 1 MW / min, 2 seconds with a rate of 2.5 MW / min, and so on until 10 seconds with a rate of 8 MW / min (extreme point), then 0-10 seconds is the acceleration segment. If the rate difference is zero once during the acceleration process, such as a rate of 5 MW / min at 5 seconds, still 5 MW / min at 6 seconds, but rising to 5.5 MW / min at 7 seconds, it is still considered an acceleration segment and is regarded as a continuation of the rise after a brief period of stabilization. The acceleration segment determination is terminated only when there are two consecutive negative differences.

[0034] Between the acceleration phase and the extreme point (the starting point of the deceleration phase), if there exists a plateau interval where the load change rate fluctuation range is less than a set threshold, this interval is considered a uniform speed phase. A plateau interval refers to a period where the rate fluctuates stably within a small range, without a significant upward or downward trend, reflecting the unit's stable rate response to load commands. The set threshold is determined based on the statistical analysis of uniform speed phase fluctuations from the unit's historical rapid adjustment data, typically ±5% of the average uniform speed rate during rapid adjustment. The specific identification process is as follows: Within the interval between the end of the acceleration phase and the extreme point, the maximum and minimum speed values ​​within a 5-second window are calculated using a sliding calculation. If the difference between the two values ​​is less than or equal to a set threshold, and the duration of the window is greater than or equal to 3 seconds to avoid misjudgment of instantaneous stability, then this window interval is determined to be a uniform speed segment. For example, if 7.6 MW / min lasts for 4 seconds, and the fluctuation range is 0.2 MW / min less than or equal to 0.75 MW / min, then this 4-second interval is a uniform speed segment. If the end of the acceleration phase directly connects to the extreme point, and there is no platform interval that meets the conditions, then a uniform speed segment is not divided, and only the acceleration and deceleration phases are retained to ensure that the segmentation closely matches the actual adjustment process. After the extreme point, the interval where the load change rate monotonically decreases to zero is the deceleration phase. Monotonically decreasing means that the rate shows an overall downward trend, allowing a slight rebound in the rate at a single time point, with the rebound amplitude less than or equal to 5% of the peak value, but the rate does not show an upward trend for 3 consecutive time points. Decreasing to zero means that the rate drops to a stable value close to zero, usually set to less than or equal to 0.1 MW / min, which is considered as the unit stopping rate change and the load approaching the command value. The specific determination process is as follows: Starting from the extreme point, monitor the rate change point by point. If the rate continues to decrease and eventually stabilizes in the range of less than or equal to 0.1 MW / min, then the range from the extreme point to the stable point is the deceleration section. If a single rate rebound occurs during the decline, it is still considered a monotonic decline. If two consecutive rebounds occur, recheck whether the extreme point was misjudged to ensure that the deceleration section reflects the actual deceleration process of the unit.

[0035] After accurately dividing the three intervals of rapid adjustment, traditional manual identification or fixed threshold methods are prone to the problem of delayed identification of deceleration segment, such as determining the extreme point only 3 seconds after it occurs, which leads to the K1 index missing some acceleration segment data. Therefore, a bidirectional long short-term memory network is introduced to achieve real-time prediction of the starting point by learning the temporal correlation between the load change rate and the starting point of the deceleration segment. The specific implementation method is as follows: The input to the bidirectional long short-term memory network is a load change rate level encoding vector and a historical deceleration phase start time series. Together, they provide the model with current rate characteristics and historical time series experience, ensuring that the prediction is both real-time and regular. The load change rate level encoding vector is a digital representation of the rate level of the time series data of the current rapid adjustment process after further subdivision. Rapid adjustment is divided into rapid level 1 based on the rate peak, with the peak value being the second rate threshold plus 3 MW / min. Rapid level 2 has a peak value of the second rate threshold plus 3 MW / min plus the second rate threshold and 6 MW / min. One-hot encoding is used to convert the sub-levels into vectors, while incorporating the initial rate and rate rise slope of the current adjustment cycle to form an encoding vector with a dimension of sub-level encoding dimension + 2. The historical deceleration phase start time series is composed of the deceleration phase start timestamps of the previous 10 rapid adjustments in the current adjustment cycle, arranged in chronological order to form a time series with a dimension of 10, and standardized as an offset relative to the start time of the adjustment cycle.

[0036] The input layer concatenates the encoded vector with the historical time sequence into a 15-dimensional input tensor, which is then fed into the hidden layer of the bidirectional long short-term memory network. This ensures that the model simultaneously captures the current rate features and historical time sequence patterns. The model controls the memory strength of time sequence features through a forget gate and an input gate, avoiding the gradient vanishing problem of traditional recurrent neural networks. At the same time, it accurately retains key features for predicting the starting point of the deceleration phase. The forget gate is responsible for determining how many historical time sequence features to retain. It uses a sigmoid function to output a gate value between 0 and 1. When the gate value is close to 1, most historical features are retained; when it is close to 0, irrelevant history is forgotten. In practice, the forget gate receives the input tensor and the output of the hidden layer at the previous time step, calculates the gate value, and multiplies it with the historical feature vector to achieve feature filtering.

[0037] The input gate determines how much of the current rate feature to update. It also uses a sigmoid function to output a gate value, which is then weighted and combined with candidate feature vectors generated by the tanh function to obtain the feature information to be updated. Subsequently, the historical features filtered by the forget gate are fused with the current features processed by the input gate to update the hidden layer state at the current time step. This ensures that the model only remembers temporal features useful for prediction. The bidirectional structure of the model extracts features simultaneously through a "forward long short-term memory network" (from the start of regulation to the current time step) and a "reverse long short-term memory network" (from the current time step to the end of the regulation period), further improving the ability to capture temporal dependencies and avoiding the unidirectional model from missing potential trends in later rate changes. The output layer uses a fully connected network to map the hidden layer features to the prediction timestamp of the deceleration phase start, realizing the extraction of features from the output layer. The transformation from vector to specific time is achieved by using two hidden layers in a bidirectional long short-term memory network (LSTM) network, each containing 256 neurons. The outputs of the hidden layers of the forward and backward LSM networks (each 256-dimensional) are concatenated to form a 512-dimensional fused feature vector. This vector is first input to the first fully connected layer (128 neurons, ReLU activation function), where it undergoes linear transformation and nonlinear activation to filter key features. It is then input to the second fully connected layer (64 neurons, ReLU activation) for further dimensionality compression. Finally, the third fully connected layer (1 neuron, no activation function) outputs the absolute time offset relative to the start time of the current adjustment cycle. This offset is then superimposed on the start timestamp of the current adjustment cycle to obtain the predicted timestamp of the deceleration phase start (10:00:11), enabling real-time prediction of the start time and solving the problem of delayed recognition.

[0038] The network training employs a mean squared error loss function with time alignment constraints to ensure precise temporal alignment between predicted and actual timestamps, avoiding temporal misalignment caused by simply minimizing error. The mean squared error loss function is based on the squared error between the predicted and actual timestamps, reflecting their numerical deviation. The time alignment constraint adds a time-series deviation penalty term to the basic loss. When the predicted timestamp is more than 2 seconds earlier or 3 seconds later than the actual timestamp (the penalty threshold is set according to the fast-adjustment period, typically 20-30 seconds; this threshold prevents severe time-series misalignment), the penalty term is weighted at twice the deviation duration. If the deviation is within ±2 seconds... Within seconds, the penalty term is 0, and only the basic mean squared error is calculated. In specific training, the dataset uses time series data of the unit's rapid adjustment process over the past 6 months, divided into training and validation sets in a 7:3 ratio. The optimizer is Adam, with an initial learning rate of 0.001, which decays to 0.8 times the previous rate every 5 epochs. The training iteration is 50 epochs. If the validation set loss does not decrease for 8 consecutive epochs, an early stop mechanism is activated to ensure model convergence and prevent overfitting. Through training with this loss function, the model can not only reduce prediction error but also strictly follow the time series pattern, ensuring that the output deceleration phase start time stamp is highly aligned with the actual adjustment process.

[0039] The specific implementation of the bidirectional long short-term memory network function mapping relationship: After clarifying the input composition and gating mechanism of the bidirectional long short-term memory network, its core function mapping relationship needs to be further refined. This mapping needs to simultaneously integrate the level characteristics of the current load change rate and the dynamic change pattern of recent load commands. By extracting bidirectional time-series features and focusing on key nodes through an attention mechanism, the time offset of the deceleration phase start point is accurately output, ensuring that the prediction results match the actual adjustment rhythm of the unit. The specific implementation method is as follows: First, the load change rate level encoding vector is concatenated with the load command change pattern of the previous N seconds of the current adjustment cycle. This provides the model with a two-dimensional input of the current rate attribute and the recent command trend. The current adjustment cycle refers to the complete time period from the start of the change in the current AGC load command to the point where the command value stabilizes and no longer fluctuates. The value of N seconds needs to be determined based on the historical data statistics of the unit's rapid adjustment, and is usually set to 5 seconds. This is because the average duration of the initial acceleration phase of the unit's rapid adjustment is about 6 seconds, and the command change in the first 5 seconds can most intuitively reflect the rate increase trend of this adjustment. This avoids missing key trends due to an N value that is too small or introducing redundant data due to an N value that is too large. The process of extracting the load command change pattern is as follows: Starting from the beginning of the current adjustment cycle, load command values ​​for the previous 5 seconds are collected at fixed intervals of 1 second, resulting in 5 consecutive command data points. Then, the command change between two adjacent data points is calculated. These 5 changes are arranged in chronological order to form a 5-dimensional command change pattern vector. This vector quantifies the recent increase and change rhythm of commands. The load change rate level encoding vector continues the sub-level division logic. During splicing, the rate level encoding vector (2-dimensional) and the command change pattern vector (5-dimensional) are directly spliced ​​in the order of rate feature first and command feature second to form a 7-dimensional fusion input vector. During the splicing process, it is necessary to ensure that the timestamps of the two are strictly aligned. The time range corresponding to the rate level encoding vector is completely consistent with the time range of the command change pattern of the previous 5 seconds, which is 0-5 seconds of the current adjustment cycle, to avoid feature confusion due to time sequence misalignment.

[0040] After feature concatenation, the 7-dimensional fused input vector is fed into the hidden layer of the bidirectional long short-term memory network. Through the synergistic effect of the forward and reverse long short-term memory networks, bidirectional temporal dependency features are extracted, ensuring that the model does not miss any key patterns in any temporal direction. Forward temporal dependency features are extracted by the forward long short-term memory network in chronological order, i.e., adjustment period 0 seconds → 1 second → 2 seconds → 3 seconds → 4 seconds → 5 seconds. The focus is on capturing the gradual evolution of rate level and command changes. For example, when the command change decreases from 8MW to 6MW, how does the rate level transition from "fast level 1" to "fast level 2," and the implicit slowing signal of the acceleration trend in this transition? Reverse temporal dependency features are extracted by the reverse long short-term memory network in reverse temporal order, i.e., 5 seconds → 4 seconds → 3 seconds → 2 seconds → 1 second → 0 seconds. Focusing on the potential impact of later-stage command changes on earlier-stage rate trends, and thus determining whether the rate is about to enter its peak phase, the hidden layer employs a two-layer structure, with each layer containing 128 neurons. The first layer, consisting of a forward long short-term memory network and a reverse long short-term memory network, performs preliminary feature extraction on the fused input vector, outputting their respective 128-dimensional feature vectors. The second layer further deepens feature mining based on the output of the first layer, using the tanh activation function to perform nonlinear transformations on the features, enhancing the model's ability to fit complex temporal correlations. Finally, the output feature vectors of the second-layer forward and reverse long short-term memory networks are concatenated step by step according to time steps to form a 256-dimensional bidirectional temporal fusion feature vector, which fully contains the bidirectional temporal correlation information between rate and command within 0-5 seconds.

[0041] After obtaining the bidirectional temporal fusion feature vector, to avoid the model over-focusing on non-critical temporal features, an attention mechanism needs to be introduced to focus on the rate drop threshold. The rate drop threshold refers to the moment when the rate changes from a continuous increase to an impending decrease, usually corresponding to the time when the command change begins to decrease continuously and the rate level is close to the peak. The features of this node are crucial for predicting the start of the deceleration phase. The specific working process of the attention mechanism is as follows: First, a rate descent feature template is constructed, generated based on the time-series features of the 1-2 seconds before the start of the deceleration phase in the unit's historical rapid adjustment data. Then, the cosine similarity between the features of each time step in the bidirectional time-series fusion feature vector and the template is calculated. Higher similarity indicates that the time step is closer to the rate descent threshold, corresponding to a higher attention weight. The feature vector of each time step is multiplied by its attention weight and summed to obtain a 128-dimensional key feature vector focused on the critical node. This vector retains only the most valuable information for predicting the start of the deceleration phase, filtering out irrelevant steady-state features. Finally, the 128-dimensional key feature vector is input to the output layer, mapped to the absolute time offset of the deceleration phase start through a fully connected network, and superimposed on the start timestamp of the current adjustment cycle to obtain the final prediction result. The output layer adopts... A three-layer fully connected structure is used. The first layer has 64 neurons, which perform dimensionality compression and nonlinear transformation on key feature vectors through the ReLU activation function. The second layer has 32 neurons to further filter core features. The third layer has only one neuron with no activation function and directly outputs the absolute time offset. This offset refers to the time interval from the start of the current adjustment cycle to the start of the deceleration phase. After obtaining the offset, it needs to be superimposed with the start timestamp of the current adjustment cycle. The calculation method is to add the absolute time offset to the start timestamp to obtain the predicted timestamp of the deceleration phase start. The whole process integrates the current rate level, recent command changes, bidirectional temporal patterns and critical node focusing to achieve a precise function mapping from feature input to time output, effectively solving the problem of delayed identification of the deceleration phase start.

[0042] Before model training, to avoid missing synchronous samples interfering with the model's learning of the coupling relationship between load command values, actual load values, and valve opening values, it is necessary to accurately identify and remove such samples. The core of the judgment revolves around the degree of linear correlation between the two within the rapid adjustment range. The specific implementation method is as follows: Samples lacking synchronization between load command values, actual load values, and valve opening values ​​specifically refer to the rapid adjustment range where the load change rate exceeds the second rate threshold. This range is defined by the dynamically calibrated second rate threshold, representing a high-intensity response state for the unit. Samples with a Pearson correlation coefficient between the load command value and the valve opening value below a predefined threshold are also considered. The Pearson correlation coefficient measures the degree of linear correlation between two variables, ranging from -1 to 1. The closer the coefficient is to 1, the stronger the positive linear correlation. Valve opening changes synchronously with the load command; the closer it is to 0 or a negative value, the weaker the linear correlation and the poorer the synchronization. The predefined threshold needs to be determined based on historical correlation coefficient statistics during normal rapid adjustment of the unit, and is usually set to 0.6. Analysis of nearly three months of fault-free rapid adjustment data revealed that the correlation coefficient under normal operating conditions is mostly above 0.7. A threshold of 0.6 effectively eliminates samples with poor synchronization without mistakenly deleting normal samples with slight fluctuations. The specific judgment process is as follows: First, extract all time intervals where the load change rate is greater than the second rate threshold from the AGC regulation timing data. For example, if the rate is 9 MW / min or greater than 8 MW / min during the 10-25 second period of a certain regulation process, this period is the target interval. Then, within this interval, load command values ​​and valve opening values ​​are synchronously collected at 1-second intervals to form multiple sets of corresponding data pairs. Statistical tools are used to calculate the Pearson correlation coefficient of these data pairs. If the coefficient is less than 0.6, it is determined to be a sample with missing synchronization. The typical manifestation of this type of sample is that the change in valve opening lags behind the change in load command. For example, if the load command has risen to 320 MW at 10 seconds, the valve opening will not rise from 40% to 45% until 12 seconds. The two changes are not synchronized and cannot reflect the true response capability of the unit, so they need to be removed from the training set.

[0043] After identifying the samples lacking synchronization, the calculation of the unit response rate index K1 needs to focus on the effective data of the acceleration and constant speed phases. Since it is clear that the deceleration phase needs to be excluded from K1, a one-dimensional convolutional neural network (1D-CNN) is used to extract the dynamic coupling features of the load command value, actual load value, and valve opening value to avoid the subjective bias of manual calculation. The specific implementation method is as follows: For the calculation of the K1 index, the input to the one-dimensional convolutional neural network consists of data pairs of load command values, actual load values, and valve opening values ​​sampled at equal intervals within the acceleration and constant speed sections. The sampling interval needs to be adapted to the time resolution of the unit's regulation. Combined with the acquisition frequency of AGC regulation data, the sampling interval is set to 1 second, which preserves sufficient timing details without increasing the computational burden due to excessive data volume. The intervals between the acceleration and constant speed sections are determined by the segmented results of the timing data during the rapid regulation process, and these two intervals need to be accurately extracted from the timing data. The data pairs refer to the combination of load values ​​and valve opening values ​​at the same time stamp. These data pairs are arranged in chronological order to form an input tensor with a dimension of sampling point number × 2. For example, with 25 sampling points (10 seconds of acceleration and 15 seconds of constant speed), the input tensor dimension is 25 × 2. Simultaneously, the data is normalized, mapping both load values ​​and valve opening values ​​to the 0-1 interval to avoid the impact of numerical magnitude differences on feature extraction. The convolution kernel slides along the time axis to extract local dynamic coupling features. A kernel is a parameter matrix in a one-dimensional convolutional neural network used to capture local temporal features. Here, three convolutional kernels of different sizes are designed, with each kernel containing 16 kernels. The kernel size corresponds to the length of the local time window. For example, a kernel of size 3 covers three consecutive sampling points, i.e., dynamic changes within 3 seconds. Multiple kernel sizes can simultaneously capture short, medium, and long-term coupled features. Sliding along the time axis means that the convolutional kernel slides along the time dimension of the input tensor with a step size of 1 sampling point, i.e., 1 second. Each time it slides, the dot product between the kernel and the data in the current local window is calculated, generating a feature. The eigenvalue extraction process can extract the local dynamic coupling relationship between load command value, actual load value, and valve opening value. For example, within a 3-second window, when the load increases from 301MW to 305MW, the valve opening increases from 31% to 34%, showing a synchronous change. Or, when the load suddenly increases, the valve opening follows the load. These features directly reflect the synchronicity and speed of the unit's response. The output features of all convolution kernels are spliced ​​together to form a feature map with a dimension of "(number of sampling points - kernel size + 1) × (16 × 3)", providing a rich feature base for subsequent screening.

[0044] The pooling layer filters the first relevant feature fragment. The pooling layer is used to compress feature dimensions and retain key features. Here, max pooling is used, with a pooling window size of 3 corresponding to 3 consecutive feature values ​​and a stride of 2. The core logic for filtering the first relevant feature fragment is that max pooling selects the point with the largest feature value within each pooling window as the output. These local windows corresponding to the largest feature values ​​are precisely the fragments with the strongest coupling between the load command value, the actual load value, and the valve opening value. Through pooling, not only is the feature map dimension compressed, but redundant features with weak coupling are also eliminated, ensuring that subsequent fully connected layers are calculated only based on highly relevant features, improving the accuracy of the K1 index. The fully connected layer regresses and outputs the actual unit response rate value, and the weight of the deceleration segment data is fixed at zero during training. The fully connected layer adopts a two-layer structure: The first layer has 64 neurons, using the ReLU activation function. It receives highly correlated features from the pooling layer output and maps these features to 64 dimensions through linear transformation and the activation function. The second layer has one neuron, which further maps the 64-dimensional features to a single numerical value, which is the actual unit response rate (K1). During training, the weights of the deceleration segment data are fixed at zero. In the model training dataset, all load command values, actual load values, and valve opening values ​​for all deceleration phases are assigned a weight coefficient of 0. By multiplying the prediction error corresponding to the deceleration phase data by 0 in the loss function calculation, the model completely ignores the influence of the deceleration phase data during training and learns features based solely on the acceleration and constant speed phase data. This ensures that the K1 index calculation completely excludes the interference of the deceleration phase and conforms to its definition of reflecting the true response speed of the unit.

[0045] Compared to the K1 index, which focuses on speed, the adjustment accuracy index K2 pays more attention to deviation control throughout the adjustment process. Especially in the deceleration phase, the deviation convergence trend directly affects the fairness of the accuracy assessment. Therefore, it is necessary to dynamically allocate weights based on the deviation convergence trend. The specific implementation method is as follows: For the K2 index calculation, the first derivative of the deviation convergence trend is used as the basis for adjusting the weight coefficient during the deceleration phase. The deviation convergence trend refers to the changing trend of the deviation between the actual load value and the load command value during the deceleration phase. The first derivative is the rate of change calculated by the absolute value of the deviation between two adjacent time points, used to quantify the speed and direction of the convergence trend. For example, if the absolute value of the deviation is 4MW at time t1 and 3MW at time t2 (t1+1 seconds), the first derivative is (3-4) / 1=-1. A negative derivative represents that the absolute value of the deviation decreases (converges) over time, and a positive derivative represents that the absolute value of the deviation increases. The larger the absolute value of the derivative, the faster the convergence or divergence speed. When the absolute value of the deviation decreases over time, the first weight is assigned. The first weight is set to 1.2. This weight is lower than the weight when there is fluctuation or no convergence because the decrease in deviation represents that the unit is gradually approaching the command value, and the negative impact on accuracy is small. There is no need to excessively increase its weight. The specific judgment and assignment process is as follows: The first derivative of the absolute value of the deviation is calculated every second during the deceleration phase. If the derivative is less than 0 for two consecutive time steps, the absolute value of the deviation is considered to be decreasing over time, and all deviation data within this time period are assigned a weight of 1.2. If the derivative is less than 0 for only a single time step and then turns positive, it is considered a short-term fluctuation and is not assigned the first weight to avoid misjudging the convergence trend. When the deviation fluctuation increases or fails to converge, a second weight is assigned. Increased deviation fluctuation refers to the first derivative of the absolute value of the deviation frequently switching between positive and negative, with a fluctuation amplitude greater than 0.3 MW / s. Failure to converge means that the absolute value of the deviation remains stable at a certain value for a long period, with a change of less than 0.5 MW within 5 consecutive seconds. Both indicate that the unit has weak control over the deviation and require close attention, therefore a higher second weight is assigned. The specific assignment process is as follows: If the first derivative is detected to switch between positive and negative values ​​more than twice within three consecutive time steps, indicating increased fluctuation, or if the absolute value of the deviation changes by less than 0.5MW for five consecutive seconds, then the deviation data for the corresponding time period is assigned a weight of 1.5. This high weight amplifies the impact of such deviations on the K2 index, ensuring that the accuracy assessment reflects the true shortcomings of the regulation. The weighting coefficient and the timestamp are jointly input into the gated recurrent unit to generate a weighted cumulative value of the deviation integral. The gated recurrent unit is a simplified recurrent neural network that captures long-term dependencies in time-series data through update and reset gates. It is suitable for handling dynamic changes in deviations during deceleration. The specific process is as follows: The absolute value of the deviation at each time step, the corresponding weight coefficient, and the timestamp are combined to form a three-dimensional input vector with a dimension of deceleration duration × 3. After this vector is input into the gated loop unit, the update gate determines whether to retain the weighted deviation information of the previous time step, while the reset gate forgets irrelevant short-term fluctuation information. The output of the gated loop unit is the weighted cumulative deviation value for each time step. Finally, the cumulative values ​​of all time steps are summed to obtain the weighted cumulative value of the deviation integral for the entire deceleration phase. This value is combined with the deviation integral values ​​of the acceleration and constant speed phases and divided by the total adjustment duration to obtain the final K2 index, ensuring that K2 can objectively reflect the impact of deviations at different stages on the overall accuracy.

[0046] This invention collects load command values, actual load values, valve opening values, and timestamps in real time during the AGC (Automatic Generation Control) adjustment process. It calculates the load change rate and classifies it into levels. The time-series data of the rapid adjustment process is divided into acceleration, constant speed, and deceleration stages. A bidirectional long short-term memory network is used to learn the mapping relationship between the rate level and the starting point of the deceleration stage, and the starting point timestamp is output. Before model training, samples with missing synchronization of load command values, actual load values, and valve opening values ​​are removed. For the unit response rate index, a one-dimensional convolutional neural network is used to extract coupling features and output the true response rate. For the adjustment accuracy index, weights are dynamically allocated based on the deviation convergence trend in the deceleration stage. A weighted deviation integral value is generated through a gated loop unit, thereby improving the accuracy of AGC performance index calculation.

[0047] Example 2 according to Figure 2 As shown, this embodiment also provides a deep learning-based AGC performance index optimization system for thermal power units, including: Data processing module 1 is used to acquire and process the timing data of the AGC regulation process to obtain the load change rate; The classification definition module 2 is used to classify and define parameters based on the load change rate, and obtain the classification results of the regulation process, core performance indicators and segmentation results of the rapid regulation process. Mapping output module 3 is used to build a model based on the hierarchical and segmented results, learn the mapping relationship, and output the starting time point of the deceleration segment under rapid adjustment. The optimization calculation module 4 is used to perform data preprocessing and dynamic optimization training based on model training requirements and parameter definitions, so as to achieve accurate optimization of AGC performance indicators.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for optimizing the AGC performance index of thermal power units based on deep learning, characterized in that, include: The timing data of the AGC regulation process is acquired and processed to obtain the load change rate; Based on the load change rate, the classification and parameter definition are performed to obtain the classification results of the regulation process, core performance indicators and segmentation results of the rapid regulation process. A model is built based on the hierarchical and segmented results, and the mapping relationship is learned to output the starting time point of the deceleration segment under rapid adjustment. Based on the model training requirements and parameter definitions, data preprocessing and dynamic optimization training are performed to achieve precise optimization of AGC performance indicators.

2. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 1, characterized in that, The process of acquiring and processing the timing data of the AGC regulation process to obtain the load change rate includes: real-time acquisition of the timing data of the load command value, actual load value, valve opening value and timestamp during the AGC regulation process; and real-time calculation of the load change rate by the derivative of the load command value and the timestamp. The process of calculating the load change rate in real time by taking the derivative of the load command value and the timestamp includes the following steps: A first-order difference operation is performed on the load command value sequence with a fixed time window, and the load change per unit time is calculated by combining the timestamp interval. Measurement noise is eliminated by moving average filtering to generate a smooth load change rate curve, which is used as the input benchmark for the three-level rate division.

3. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 1, characterized in that, The specific process of classifying and defining parameters based on the load change rate to obtain the classification results of the regulation process, core performance indicators, and segmented results of the rapid regulation process includes: The regulation process is classified into three levels based on the load change rate, including slow, medium and fast. K1 is defined as the unit response rate index and K2 is defined as the regulation accuracy index. The timing data of the fast regulation process is divided into acceleration, constant speed and deceleration stages based on the load change rate. The starting point of the deceleration stage is the time point when the load change rate decreases from the peak. The regulation process is classified into slow, medium, and fast based on the load change rate. The classification is based on thresholds, including: slow corresponds to a load change rate not exceeding the first rate threshold, medium corresponds to a load change rate between the first and second rate thresholds, and fast corresponds to a load change rate higher than the second rate threshold. The first and second rate thresholds are dynamically calibrated based on the statistical quantiles of the unit's historical regulation data.

4. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 3, characterized in that, The definitions of K1 and K2 specifically include using the average actual load change rate of the acceleration and constant speed sections as the unit response rate index K1, and using the integral average of the absolute value of the deviation between the actual load value and the command value during the entire adjustment process as the adjustment accuracy index K2; among them, the calculation of K1 needs to exclude the deceleration section data, and the calculation of K2 needs to distinguish the dynamic weight.

5. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 3, characterized in that, Based on the load change rate, the time series data of the rapid adjustment process is divided into acceleration, constant speed, and deceleration segments, specifically including: The extreme point of the load change rate curve is identified as the starting point of the deceleration section. The interval in which the load change rate continues to rise before this starting point is the acceleration section. The plateau interval in which the load change rate fluctuates within a set threshold is the uniform speed section. The interval in which the load change rate monotonically decreases to zero is the deceleration section.

6. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 1, characterized in that, A model is constructed based on the hierarchical and segmented results, and the mapping relationship is learned to output the specific process at the beginning time point of the rapid adjustment deceleration segment, including: A bidirectional long short-term memory network is constructed. The network learns the functional mapping relationship between the load change rate level and the starting point of the deceleration segment, and outputs the starting time point of the deceleration segment during the rapid adjustment process. The construction and training process of the bidirectional long short-term memory network includes: The system takes the load change rate level encoding vector and the historical deceleration segment start time series as inputs, and controls the memory strength of the temporal features through forget gate and input gate. The output layer uses a fully connected network to map to the predicted timestamp of the deceleration segment start. The network training uses the mean square error loss function with time alignment constraints.

7. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 6, characterized in that, The functional mapping relationship between the load change rate level and the deceleration phase start point specifically includes: concatenating the load change rate level encoding vector with the load command change pattern of the previous N seconds of the current adjustment cycle, extracting the positive and negative temporal dependency features through the hidden layer of the bidirectional long short-term memory network, focusing on the rate drop threshold by the attention mechanism, and finally generating the absolute time offset of the deceleration phase start point through the output layer. This offset is then superimposed on the start timestamp of the current adjustment cycle to obtain the deceleration phase start point.

8. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 1, characterized in that, Based on model training requirements and parameter definitions, the specific processes of data preprocessing and dynamic optimization training include: Before model training, samples with missing synchronization are removed. Based on the predefined correlation coefficient threshold, samples with missing synchronization of load command value, actual load value and valve opening value are removed. A dynamic compensation module is constructed using a gated loop unit to optimize the nonlinear characteristics of the deceleration section in the time series data of the rapid adjustment process. Among them, the samples with missing synchronization between load command value, actual load value and valve opening value are specifically included in the range where the load change rate is greater than the second rate threshold and the Pearson correlation coefficient between the load command value and the valve opening value is lower than the predefined threshold. These samples are characterized by the valve opening value changing lagging behind the load command value.

9. The method for optimizing the AGC performance index of thermal power units based on deep learning according to claim 8, characterized in that, In the optimization training of the nonlinear characteristics of the deceleration section in the time series data of the rapid adjustment process, the K1 index is calculated by selecting only the load command value, actual load value and valve opening value of the acceleration section and the constant speed section. The time series data are then processed by a one-dimensional convolutional neural network to extract highly correlated dynamic features and output the real unit response rate. In the calculation of the K1 index, the input of the one-dimensional convolutional neural network is a data pair of load command value, actual load value and valve opening value sampled at equal intervals in the acceleration section and the constant speed section. The convolutional kernel slides along the time axis to extract local dynamic coupling features, the pooling layer filters the first relevant feature fragment, and the fully connected layer regresses to output the real unit response rate value. During training, the weight of the deceleration section data is fixed at zero. For the K2 index calculation, weights are dynamically allocated based on the deviation convergence trend during the deceleration phase, specifically including: During the deceleration phase, the first derivative of the deviation convergence trend is used as the basis for adjusting the weight coefficients. When the absolute value of the deviation decreases with time, the first weight is assigned, and when the deviation fluctuation increases or fails to converge, the second weight is assigned. The weight coefficients and timestamps are jointly input into the gated loop unit to generate the weighted cumulative value of the deviation integral.

10. A deep learning-based AGC performance index optimization system for thermal power units, characterized in that, include: The data processing module is used to acquire and process the timing data of the AGC regulation process to obtain the load change rate; The classification definition module is used to classify and define parameters based on the load change rate, and obtain the classification results of the regulation process, core performance indicators and segmentation results of the rapid regulation process. The mapping output module is used to build a model based on the hierarchical and segmented results, learn the mapping relationship, and output the starting time point of the deceleration segment under rapid adjustment. The optimization calculation module is used to perform data preprocessing and dynamic optimization training based on model training requirements and parameter definitions, so as to achieve accurate optimization of AGC performance indicators.

Images