Self-adaptive adjustment method and system for full-load peak regulation of coal-fired power generation unit

Abstract

The invention relates to the technical field of peak regulation of coal-fired power units, and discloses a self-adaptive regulation method and system for full-load peak regulation of a coal-fired power unit. The method comprises the steps of collecting unit operation parameters such as boiler temperature, turbine rotating speed and generator voltage, and power grid peak regulation instructions such as real-time load demand and predicted load fluctuation; analyzing a peak regulation state based on the collected data, and dividing the peak regulation state into three types of stable operation, transition regulation and limit peak regulation; then calculating an adjustment characteristic quantity according to the current state and the operation parameters, wherein the characteristic quantity reflects the matching relationship between the unit performance and the peak regulation requirement; generating a plurality of groups of adjustment schemes including combustion optimization, steam adjustment and power control based on the characteristic quantity and the peak regulation instruction; and selecting and executing an optimal scheme by evaluating indexes such as efficiency loss, stability influence and maintenance cost. The method can dynamically respond to the load change of the power grid, accurately adapt to the peak regulation requirement in the full load range, and enable the peak regulation of the unit to be more flexible and harmonious.

Description

Technical Field

[0001] This invention relates to the field of peak shaving technology for coal-fired power units, specifically to an adaptive adjustment method and system for full-load peak shaving of coal-fired power generating units. Background Technology

[0002] The current energy structure transformation is accelerating, with wind power, photovoltaic power, and other new energy power generation technologies developing rapidly and being connected to the grid on a large scale. These new energy power generation technologies have inherent intermittent and fluctuating characteristics, leading to a continuous increase in the amplitude and frequency of fluctuations in the total load of the power grid, placing higher demands on the safe and stable operation of the grid. As the core supporting power source for the power grid, the peak-shaving capacity of coal-fired power generating units directly affects the grid's ability to absorb new energy sources and the reliability of power supply; therefore, full-load peak shaving has become an important operational requirement for coal-fired power generating units.

[0003] In existing peak-shaving regulation methods for coal-fired power generating units, the data acquisition stage often has limitations. Some methods only focus on local operating parameters of the unit or only receive a single real-time load command, lacking a comprehensive consideration of the overall operating status of the unit and the peak-shaving needs of the power grid. This results in insufficient data support for subsequent regulation decisions. Regarding peak-shaving status assessment, existing technologies mostly adopt a general status classification approach, failing to perform refined classification based on factors such as the unit's operating load range and load change rate. This makes it difficult to accurately identify regulation needs under different scenarios such as stable operation, transitional regulation, and extreme peak-shaving, easily leading to a mismatch between regulation strategies and actual conditions.

[0004] In the regulation scheme generation stage, existing methods mostly rely on fixed regulation parameters or preset regulation modes, lacking dynamic analysis of the matching relationship between real-time unit performance and peak-shaving requirements. The generated regulation schemes are not highly targeted and struggle to achieve coordinated optimization across multiple key regulation dimensions such as combustion, steam, and power. In the scheme evaluation stage, most technologies focus only on a single performance indicator or ignore long-term operational factors such as maintenance costs, potentially leading to selected regulation schemes with unreasonable efficiency, insufficient stability, or excessively rapid equipment wear and tear.

[0005] Existing peak-shaving regulation methods generally lack adaptive capabilities. When faced with rapid changes in real-time load demand and predicted load fluctuations, the regulation response is lagging, or there may be over-regulation or under-regulation. This makes it difficult for the units to maintain a good operating condition within the full load range, which not only affects the peak-shaving effect but may also lead to increased energy consumption and increased equipment failure risk. It cannot fully meet the comprehensive requirements of new energy power grids for the flexibility, stability and economy of peak-shaving of coal-fired power generating units. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive adjustment method and system for full-load peak shaving of coal-fired power generating units, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an adaptive adjustment method for full-load peak shaving of coal-fired power generating units, the method comprising: The system collects operating parameters of coal-fired power generating units and grid peak shaving commands. The operating parameters include boiler temperature, turbine speed and generator voltage, while the grid peak shaving commands include real-time load demand and predicted load fluctuations. The collected operating parameters and power grid peak-shaving commands are used to analyze the peak-shaving status and obtain the current peak-shaving status. The current peak-shaving status is divided into stable operation status, transitional regulation status and extreme peak-shaving status. The regulation characteristic is calculated based on the current peak-shaving status and operating parameters. The regulation characteristic represents the matching relationship between the unit performance and the peak-shaving requirements. A set of regulation schemes is generated based on regulation characteristic quantities and grid peak shaving commands. Each regulation scheme involves combustion optimization, steam adjustment and power control. Evaluate the performance metrics of each adjustment scheme, including efficiency loss, stability impact, and maintenance costs, and select the best adjustment scheme. Implement the optimal regulation scheme to achieve adaptive peak-shaving regulation of coal-fired power generating units.

[0008] Preferably, the analysis of peak-shaving status using the collected operating parameters and power grid peak-shaving commands is achieved through the following steps: A state classification model is constructed, which is trained based on the historical operation dataset of coal-fired power generating units. The historical operation dataset contains parameter sequences and corresponding peak-shaving state labels under various operating conditions. An ensemble learning framework is used to train the state classification model. The ensemble learning framework integrates the outputs of multiple base classifiers, including decision trees, nearest neighbor algorithms and Naive Bayes classifiers. The real-time operating parameters are segmented by a sliding window to generate time series segments, and statistical features including mean, variance and trend coefficient are extracted. The statistical characteristics and power grid peak-shaving instructions are input into the state classification model, the model outputs a probability distribution, and the current peak-shaving state is determined by threshold comparison.

[0009] Preferably, the calculation of the adjustment characteristic quantity based on the current peak-shaving status and operating parameters is achieved through the following steps: Establish a baseline parameter template corresponding to the peak-shaving state. The baseline parameter template defines the value range of the ideal operating parameters. Calculate the dynamic deviation between the real-time running parameters and the baseline parameter template. The dynamic deviation is calculated using a combination of Euclidean distance and cosine similarity. An urgency indicator for power grid peak-shaving commands is introduced, which is derived based on load change rate and time constraints. By integrating dynamic deviation and urgency indicators, a moderating feature is generated through weighted summation, and the moderating feature is output as a multi-dimensional vector.

[0010] Preferably, the generation of a set of regulation schemes based on regulation characteristic quantities and power grid peak-shaving commands is achieved through the following steps: Clustering algorithms are used to group the historical adjustment scheme database. The clustering algorithm divides the scheme groups based on adjustment feature quantities and effect indicators. Using the current adjustment feature as the query point, search for nearest neighbor solutions in the solution group to form an initial solution set; The genetic algorithm is applied to optimize and iterate the initial set of solutions. The genetic algorithm generates new solutions through crossover and mutation operations. Select individuals from the new schemes that meet the constraints of the power grid peak-shaving command, and output a set of adjustment schemes.

[0011] Preferably, the performance indicators of each adjustment scheme are evaluated through the following steps: Construct a multi-dimensional evaluation indicator system that covers energy consumption, emissions, and equipment stress; Grey system theory is used to calculate the correlation degree between each adjustment scheme and the ideal scheme. The correlation degree reflects the quality of the scheme. Weighting coefficients are set and dynamically adjusted based on expert knowledge. The relevance is then weighted and fused to obtain a comprehensive score. Compare all the overall scores and select the adjustment scheme with the highest score as the best adjustment scheme.

[0012] Preferably, the training of the state classification model using an ensemble learning framework is achieved through the following steps: Prepare a training sample set, which contains labeled peaking state data; Each base classifier is trained independently, and the accuracy of the base classifiers is evaluated using 10-fold cross-validation. The design incorporates a weighted voting method, with weights allocated based on the accuracy of the base classifiers. Verify the overall performance of the state classification model by adjusting the model parameters using the confusion matrix.

[0013] Preferably, the calculation of the dynamic deviation between the real-time operating parameters and the benchmark parameter template is achieved through the following steps: Standardize the operating parameters to eliminate dimensional differences; The main components of the parameters are extracted, and these main components are obtained by dimensionality reduction using principal component analysis. Calculate the similarity measure between the principal components and the baseline parameter template. The similarity measure is calculated using Mahalanobis distance. The deviation calculation results are dynamically updated by combining a time sliding window.

[0014] Preferably, the grouping of the historical adjustment scheme library using a clustering algorithm is achieved through the following steps: Preprocess historical tuning schemes and extract feature vectors including tuning parameters and runtime context; The density clustering algorithm is applied to divide the scheme group. The density clustering algorithm is based on the neighborhood radius and the minimum number of points parameter. To assess cluster quality, the silhouette coefficient is used to optimize the number of clusters. An index is created for each scheme group to facilitate fast retrieval.

[0015] Preferably, the calculation of the correlation between each adjustment scheme and the ideal scheme using grey system theory is achieved through the following steps: Define the feature sequence of the ideal solution, which is derived from the best historical cases; Normalize the eigenvalues ​​of each adjustment scheme to eliminate the order-of-magnitude effect; Calculate the grey relational coefficient, which reflects the local similarity between the proposed solution and the ideal solution; The overall correlation degree is obtained by aggregating the correlation coefficients, and the correlation degree value range is from 0 to 1.

[0016] Preferably, the present invention also includes an adaptive regulation system for full-load peak shaving of a coal-fired power generating unit, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the adaptive regulation method for full-load peak shaving of a coal-fired power generating unit as described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are: By comprehensively collecting unit operating parameters and grid peak-shaving commands, the limitations of incomplete data collection in existing technologies have been overcome, allowing all subsequent regulation processes to be based on complete and accurate fundamental information. Comprehensive acquisition of operating parameters such as boiler temperature, turbine speed, and generator voltage fully reflects the current operating status of the unit, while the synchronous collection of real-time load demand and predicted load fluctuations fully grasps the immediate needs and future trends of grid peak-shaving. The combination of these two aspects provides comprehensive data support for peak-shaving regulation, making regulation behavior more aligned with actual operating scenarios.

[0018] The refined classification of peak-shaving states clearly categorizes the current state into stable operation, transitional adjustment, and extreme peak-shaving states. This addresses the issue of vague state judgments in existing technologies, enabling precise identification of adjustment priorities under different operating scenarios. Under different peak-shaving states, the operating characteristics and peak-shaving demands of generating units differ significantly. This refined state classification allows adjustment strategies to accurately adapt to various scenarios, avoiding the inapplicability of a single adjustment mode to different states, and making adjustment behavior more targeted and rational.

[0019] The introduction of regulation characteristic quantities establishes a direct correlation between unit performance and peak-shaving requirements, providing a clear direction for the generation of regulation schemes. Through these characteristic quantities, it becomes clear whether the current performance of the unit meets peak-shaving requirements and what adaptation gaps exist. This ensures that the generation of subsequent regulation schemes is no longer arbitrary but rather focuses on bridging adaptation gaps and optimizing matching relationships, thereby improving the scientific rigor and effectiveness of the regulation schemes.

[0020] The generation of multi-dimensional regulation schemes, combined with comprehensive evaluation of multiple indicators, achieves comprehensive optimization of peak-shaving regulation. The regulation schemes cover multiple key dimensions, including combustion optimization, steam adjustment, and power control, enabling coordinated regulation from the core aspects of unit operation. This avoids the limitations of single-dimensional regulation and ensures comprehensive regulation effects. Furthermore, the comprehensive evaluation of multiple indicators, such as efficiency losses, stability impacts, and maintenance costs, comprehensively considers the impact of regulation schemes on the unit's short-term operating performance and long-term operating status. The selected optimal scheme not only meets immediate peak-shaving needs but also takes into account the overall performance of unit operation, achieving a balance between peak-shaving effects and unit operating status.

[0021] The entire regulation process exhibits significant adaptive characteristics, capable of adjusting each stage—including peak-shaving status judgment, regulation characteristic quantity calculation, regulation scheme generation, and evaluation—in real time based on dynamic changes in unit operating parameters and grid peak-shaving commands. This adaptive capability enables the unit to flexibly respond to real-time fluctuations and predicted changes in grid load. Under stable operating conditions, it avoids unnecessary regulation operations, reducing ineffective energy consumption and equipment wear; under transitional regulation conditions, it can quickly respond to load changes, ensuring timely peak-shaving; and under extreme peak-shaving conditions, it can precisely control the regulation amplitude, ensuring safe unit operation.

[0022] This method does not rely on fixed adjustment parameters or preset modes. It can dynamically optimize the adjustment strategy across the entire load operating range, ensuring that the peak-shaving capacity of coal-fired power generating units is always well-matched with grid demand. Through the synergistic effect of each link, this method can improve the operational coordination of the unit during peak shaving, reduce unreasonable losses during the adjustment process, reduce equipment operating pressure, and broaden the effective operating range of the unit. This enables coal-fired power generating units to better play their peak-shaving supporting role in a grid environment where the proportion of renewable energy continues to increase, and adapt to the grid's comprehensive requirements for peak-shaving flexibility, stability, and economy. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the working principle of the adaptive adjustment method for full-load peak shaving of coal-fired power generating units according to the present invention. Figure 2 The flowchart for peak shaving state analysis; Figure 3A flowchart for calculating the adjustment feature quantity; Figure 4 A scatter plot showing the correlation between load demand, dynamic deviation, and peak-shaving status of coal-fired power generating units; Figure 5 A heatmap showing the correlation of key operating parameters for coal-fired power generating units. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1 This invention provides an adaptive regulation method for full-load peak shaving of coal-fired power generating units. The method includes: a data acquisition and control system responsible for collecting operating parameters of the coal-fired power generating unit and grid peak shaving commands. The operating parameters include boiler temperature, turbine speed, and generator voltage. The grid peak shaving commands include real-time load demand and predicted load fluctuations. The system uses the collected operating parameters and grid peak shaving commands to analyze the peak shaving state and obtain the current peak shaving state, which is divided into a stable operating state, a transitional regulation state, and a limit peak shaving state. Regulation characteristic quantities are calculated based on the current peak shaving state and operating parameters. These regulation characteristic quantities represent the matching relationship between unit performance and peak shaving requirements. A set of regulation schemes is generated based on the regulation characteristic quantities and grid peak shaving commands. Each regulation scheme involves combustion optimization, steam adjustment, and power control. The performance indicators of each regulation scheme are evaluated, including efficiency loss, stability impact, and maintenance cost, and the optimal regulation scheme is selected.

[0026] Example 1: See Figure 2In practical implementation, peak-shaving state analysis is achieved by constructing a state classification model. This model is trained on historical operating datasets of coal-fired power generating units. The historical datasets contain parameter sequences under various operating conditions and corresponding peak-shaving state labels. The historical datasets are sourced from long-term archived operating data of the power plant's distributed control system. The parameter sequences are time-series data, recording the values ​​of parameters such as boiler temperature, turbine speed, and generator voltage at historical points in time. Peak-shaving state labels are manually labeled by domain experts based on historical operating records and power grid dispatch logs, categorized into three types: stable operating state, transitional regulation state, and extreme peak-shaving state. An ensemble learning framework is used to train the state classification model. This framework integrates the outputs of multiple base classifiers, including decision trees, nearest neighbor algorithms, and Naive Bayes classifiers. The design goal of the ensemble learning framework is to comprehensively utilize the advantages of different classification algorithms. Decision tree models can generate easily understandable classification rules, nearest neighbor algorithms are suitable for discovering local similarity patterns, and Naive Bayes classifiers classify based on probability principles. Real-time operating parameters are segmented using a sliding window to generate time-series segments, and statistical features, including mean, variance, and trend coefficient, are extracted. The sliding window segmentation uses a fixed-length time window, the length of which is set according to the unit's inertial characteristics. The window slides continuously on the time axis, each slide representing a sampling interval, thus generating continuous time series segments. Statistical feature extraction is performed separately for each operating parameter within the window. The mean is calculated as the arithmetic mean of the data within the window, the variance is calculated as the dispersion of the data within the window, and the trend coefficient is calculated as the slope of the sequence through linear fitting. The statistical features and the grid peak-shaving command are input into the state classification model. The model outputs a probability distribution, and the current peak-shaving state is determined by threshold comparison. The input to the state classification model is a feature vector, the elements of which include the statistical features of all operating parameters and the load demand value in the grid peak-shaving command. The model output is a three-dimensional probability vector, representing the probability of belonging to the stable operation state, the transitional regulation state, and the extreme peak-shaving state, respectively. By comparing the relative magnitudes of the three probability values ​​with preset thresholds, the state corresponding to the highest probability value is determined as the current peak-shaving state.

[0027] In the specific implementation, an ensemble learning framework is used to train the state classification model. A training sample set containing labeled peak-shaving state data is prepared. The training sample set is constructed by randomly selecting representative samples from historical datasets. Each sample contains a parameter sequence within a time window and its corresponding true peak-shaving state label. The number of samples needs to meet the convergence requirements of the model training. Each base classifier is trained independently, and ten-fold cross-validation is used to evaluate the accuracy of the base classifiers. The training process of the base classifiers is parallel. The decision tree classifier constructs a tree structure through recursive partitioning, the nearest neighbor algorithm classifies samples by calculating the distance between them, and the Naive Bayes classifier calculates the posterior probability based on the feature conditional independence assumption. Ten-fold cross-validation randomly divides the training sample set into ten parts, sequentially using nine parts as the training set and one part as the validation set, repeating this process ten times to obtain the average accuracy. A fusion strategy is designed, employing a weighted voting method, with weights allocated based on the accuracy of the base classifiers. In the weighted voting method, the voting weight of each base classifier is proportional to its accuracy obtained in cross-validation. The final output of the state classification model is a weighted average of the outputs of all base classifiers. The overall performance of the state classification model is verified by adjusting the model parameters using a confusion matrix. The confusion matrix is ​​a three-dimensional matrix where rows represent the true state class, columns represent the predicted state class, and matrix elements represent the classification behavior for each class. By analyzing the misclassification patterns in the confusion matrix, the probability threshold of the state classification model can be adjusted, or the base classifiers can be retrained.

[0028] In some embodiments, feature selection is required during the training process of the state classification model. Feature selection involves filtering the subset of features that contribute most to state classification from the original operating parameters. Feature selection methods can employ filtering, wrapping, or embedded methods. Filtering methods rank features based on their correlation with state labels, wrapping methods evaluate feature subsets through classifier performance, and embedded methods integrate feature selection into the model training process. The purpose of feature selection is to reduce feature dimensionality, decrease computational load, and improve the generalization ability of the state classification model. After feature selection, the state classification model needs to be retrained, and its performance should be evaluated using an independent test set. The parameter settings for sliding window segmentation affect the real-time performance and accuracy of state recognition. A window length that is too short introduces excessive noise, while a window length that is too long leads to delays in state recognition. The window length setting needs to comprehensively consider the dynamic response characteristics of the unit and the update frequency of grid peak-shaving commands. Generally, the window length should cover the time constants of the unit's main inertial components. The window sliding step size determines the update rate of state recognition; a smaller step size results in more timely state recognition but also a greater computational load. In practical applications, the window length and sliding step size need to be adjusted according to the specific characteristics of the unit and optimized during the commissioning phase.

[0029] Optionally, statistical features can be further extended to capture richer time series patterns. In addition to mean, variance, and trend coefficient, higher-order statistics such as skewness and kurtosis can be calculated. Skewness reflects the asymmetry of the data distribution, and kurtosis reflects the sharpness of the data distribution; these features help identify abnormal fluctuation patterns in operating parameters. Frequency domain features can also be introduced, transforming the time series to the frequency domain through Fourier transform to extract the amplitude and phase information of the main frequency components. Frequency domain features are more sensitive to periodic fluctuations. Feature extension can improve the ability of state classification models to identify complex operating patterns, but it increases computational complexity and feature dimensionality, requiring a trade-off between model performance and computational efficiency.

[0030] In some embodiments, the base classifier of the integrated learning framework can be dynamically updated. Dynamic updating involves periodically retraining the base classifier based on newly collected operational data. The dynamic update process includes collecting new operational data samples, having experts label the peak-shaving state, adding the new samples to the training set, and re-executing the training of the base classifier and adjusting the fusion strategy. Dynamic updating allows the state classification model to adapt to slow changes in unit performance, such as changes in operating characteristics caused by equipment aging or fuel characteristics. The update cycle can be set according to the unit overhaul cycle or fuel change cycle, typically several months or a year. The output probability distribution of the state classification model can be used to assess the confidence level of state identification. When the probability values ​​of the three states are close, it indicates that the current operating state is at the classification boundary, and the uncertainty of the state identification result is high. In this case, a manual intervention mechanism can be triggered, sending the state identification result to operators for confirmation, or adopting a more conservative adjustment strategy. The entropy of the probability distribution can quantify the uncertainty; a higher entropy value indicates higher uncertainty. By monitoring the entropy value of the probability distribution, online assessment of the reliability of state identification can be achieved.

[0031] Optionally, the state classification model can be established separately for different load ranges. Load ranges are defined based on the typical operating load of the unit, such as low-load, medium-load, and high-load ranges. Historical operating data is collected for each load range to train an independent state classification model. Since the dynamic characteristics of the unit differ across load ranges, load range-specific modeling can improve the accuracy of state classification. The boundaries of the load ranges can be determined based on the unit's design parameters and operating experience; typically, the load value corresponding to the inflection point of the unit's efficiency curve is chosen as the range boundary.

[0032] In practice, the preprocessing of real-time operating parameters includes data cleaning and outlier handling. Data cleaning involves detecting and correcting erroneous values ​​in the collected data, such as jumps or dead values ​​caused by sensor malfunctions. Outlier handling identifies and removes data points that significantly deviate from the normal range. Outlier identification methods include those based on statistical distribution or distance. Only preprocessed data can be used for sliding window segmentation and feature extraction, ensuring the data quality of the input state classification model. Data cleaning rules and outlier judgment thresholds need to be set based on the statistical characteristics of the unit's historical operating data.

[0033] It's understandable that parsing power grid peak-shaving commands requires addressing the standardization of command formats, as different power grid dispatch centers may have different command formats. The command parsing module needs to convert commands of different formats into a standard format recognizable by the state classification model. The standard format includes fields such as load demand value, load change rate, and command validity period. The command parsing module also needs to handle the timeliness of commands, filtering out expired commands and processing only currently valid peak-shaving commands.

[0034] Example 2: See Figure 3In practical implementation, the calculation of regulation characteristic quantities is achieved by establishing benchmark parameter templates corresponding to peak-shaving states. These benchmark parameter templates define the value range of ideal operating parameters. The construction of benchmark parameter templates is based on the design parameters of coal-fired power generating units, historical best operating data, and expert experience. Different benchmark parameter templates are established for stable operating states, transitional regulation states, and extreme peak-shaving states. Each benchmark parameter template contains a set of ideal values ​​or reasonable ranges for operating parameters, including key variables such as boiler temperature, turbine speed, and generator voltage. The benchmark parameter templates are stored in the system in the form of data tables. Rows in the data table correspond to different operating parameters, and columns include parameter name, ideal value, lower threshold, and upper threshold. The dynamic deviation between real-time operating parameters and benchmark parameter templates is calculated using a combination of Euclidean distance and cosine similarity. Euclidean distance measures the absolute numerical difference between the real-time operating parameter vector and the benchmark parameter vector, while cosine similarity measures the directional similarity between the two vectors, i.e., the consistency of parameter change trends. The dynamic deviation is a weighted combination of Euclidean distance and cosine similarity, with the weighting coefficients set according to the importance and dimensions of the parameters. An urgency index for power grid peak-shaving commands is introduced, derived based on the load change rate and time constraints. The load change rate is the change in load demand per unit time, and the time constraint is the remaining time required for generating units to reach the target load. The calculation formula for the urgency index comprehensively considers the magnitude and direction of the load change rate and the urgency of the time constraint. The dynamic deviation and urgency index are integrated, and a regulation characteristic quantity is generated through weighted summation. This regulation characteristic quantity is output as a multi-dimensional vector. The weighting coefficients in the weighted summation are dynamically adjusted according to the current peak-shaving state, biasing towards the dynamic deviation under stable operation and towards the urgency index under extreme peak-shaving conditions. Each dimension in the multi-dimensional vector of the regulation characteristic quantity represents a performance matching aspect, such as the matching degree of the thermal system, the matching degree of the power system, and the matching degree of the electrical system.

[0035] In practice, the dynamic deviation between real-time operating parameters and the baseline parameter template is calculated. Standardization of the operating parameters eliminates dimensional differences. The standardization process uses the Z-score method to convert each operating parameter into a standard normal distribution with a mean of zero and a standard deviation of one. Standardization requires pre-calculating the historical mean and standard deviation of each parameter; historical data comes from long-term unit operating records. The principal components of the parameters are extracted using principal component analysis (PCA). PCA performs eigenvalue decomposition on the standardized operating parameter matrix, selecting the principal components with the largest eigenvalues. Principal components are linear combinations of the original parameters, preserving most of the variation information of the original data. A similarity measure between the principal components and the baseline parameter template is calculated using Mahalanobis distance. Mahalanobis distance is a weighted Euclidean distance between the principal component vectors of the real-time operating parameters and the principal component vectors of the baseline parameter template, with the weight being the reciprocal of the variance of the principal components. Mahalanobis distance considers the correlation between parameters and reflects the true degree of deviation better than Euclidean distance. The deviation calculation results are dynamically updated using a time sliding window. The length of the time sliding window is consistent with the window length in the state classification model. The window slides on the time axis, and each slide is a sampling interval. The dynamic deviation is recalculated at each window position to form a dynamic deviation time series.

[0036] In some embodiments, the baseline parameter template can be designed as an adaptive update mode. Adaptive update dynamically adjusts the ideal values ​​in the baseline parameter template based on the recent operating performance of the unit. The adaptive update algorithm monitors the parameter distribution of the unit under stable operating conditions. When a systematic shift occurs in the parameter distribution, the setpoints of the baseline parameter template are adjusted accordingly. Adaptive update can adapt to slow drifts in unit performance, such as those caused by equipment aging or changes in fuel characteristics. Adaptive update requires setting safety boundaries to prevent erroneous template adjustments due to temporary fluctuations in operating conditions. The choice of dimensionality for principal component analysis (PCA) dimensionality reduction affects the accuracy of dynamic deviation calculation. Too high a dimensionality will retain noise information, while too low a dimensionality will lose useful information. Dimensionality selection typically uses the cumulative contribution rate method, selecting the minimum dimensionality that makes the cumulative contribution rate reach a predetermined threshold. The cumulative contribution rate threshold is generally set between 85% and 95%, and fine-tuned according to the specific operating characteristics of the unit. The PCA model needs to be retrained periodically to adapt to changes in the unit's operating characteristics.

[0037] Optionally, similarity metrics can incorporate dynamic time warping (RTW) to handle phase differences between parameters. RTD is a nonlinear warping technique capable of finding the optimal matching path between two time series. When there is a time delay in the changes of operating parameters, RTD can calculate similarity more accurately. RTD has high computational complexity and is suitable for specific applications requiring high accuracy. The path constraints for RTD need to be set based on the physical correlation between parameters. The calculation of the urgency index can incorporate load forecast uncertainty, which stems from the error range of the power grid load forecasting model. The urgency index considers both the load change rate and time constraints, while also adjusting for forecast uncertainty. When forecast uncertainty is high, the weight of the urgency index should be appropriately reduced to avoid over-adjustment due to forecast errors. The quantification of forecast uncertainty is based on statistical analysis of the historical performance of the load forecasting model.

[0038] It is understandable that weight coefficient tuning in weighted summation is a multi-objective optimization problem, requiring comprehensive consideration of multiple requirements such as adjustment speed, stability, and safety. Weight coefficient tuning can employ a combination of offline simulation optimization and online adaptive adjustment. In the offline phase, the initial values ​​of the weight coefficients are determined through simulation using historical data; in the online phase, the weight coefficients are fine-tuned in real time based on the adjustment effect. The weight coefficient tuning process requires setting constraints to ensure that all performance indicators remain within safe ranges. The multi-dimensional vector of the adjustment characteristic can incorporate time derivative information, which represents the rate of change of each dimension of the adjustment characteristic over time. Time derivative information can reflect the dynamic trend of unit status changes, providing a basis for predictive regulation. The time derivative is calculated using numerical differentiation methods, such as the central difference method or Kalman filtering. The introduction of time derivative information increases the predictability of the adjustment but also amplifies the impact of measurement noise, requiring appropriate filtering.

[0039] In practical implementation, the calculation of dynamic deviation needs to consider the coupling effect between parameters. The coupling effect refers to the chain reaction that a change in one parameter can cause to other parameters. Coupling effect modeling is achieved by analyzing the correlation matrix of parameters in historical data; the correlation matrix characterizes the degree of synchronization of parameter changes. When calculating dynamic deviation, highly coupled parameter groups are processed jointly, rather than calculating the deviation of each parameter individually. Coupling effect modeling improves the ability of dynamic deviation to represent the overall system state. The baseline parameter template should have different setpoints under different load levels. The load segmentation of the baseline parameter template is based on intervals divided according to the typical operating load of the unit. An independent baseline parameter template is established for each load interval, and the ideal values ​​in the template match the optimal operating parameters for that load interval. The granularity of the load interval division affects the accuracy of the baseline parameter template; too coarse a division reduces the applicability of the template, while too fine a division increases the complexity of template maintenance. The load interval division needs to be optimized by combining the unit's design characteristics and operating experience.

[0040] See Figure 4 This graph focuses on the correlation analysis between dynamic deviation and load demand and peak-shaving status. The horizontal axis represents load demand, and the vertical axis represents dynamic deviation. Three types of peak-shaving status are distinguished by color gradients. The graph clearly shows that: stable operation is mostly distributed in the 40%-80% load demand range, with generally low dynamic deviation, indicating small deviations between unit operating parameters and the baseline template, and strong operational stability; transitional regulation covers the 40%-90% load demand range, with a large range of dynamic deviation, reflecting the dynamic adjustment characteristics of unit parameters deviating from the baseline template during load changes; extreme peak-shaving status is concentrated in the extreme range of load demand, with generally high dynamic deviation, reflecting significant deviations between the unit and the ideal baseline state under extreme loads. This correlation analysis is a core supporting link in calculating regulation characteristic quantities. By quantifying the relationship between load demand, dynamic deviation, and peak-shaving status, it provides data basis for the subsequent fusion of dynamic deviation and grid peak-shaving command urgency indicators to generate regulation characteristic quantities. This helps to accurately determine the matching degree between unit performance and peak-shaving requirements, and is one of the key basic analyses for achieving full-load adaptive peak-shaving of coal-fired units.

[0041] Example 3: In specific implementation, the generation of regulation schemes is achieved by grouping the historical regulation scheme database using a clustering algorithm. The clustering algorithm divides the scheme groups based on regulation feature quantities and effect indicators. The historical regulation scheme database comes from successful regulation cases accumulated during the long-term operation of coal-fired power generating units. Each case records a complete peak-shaving regulation process, including the system state before regulation, the regulation measures taken, and the effect evaluation after regulation. Regulation feature quantities are quantitative indicators reflecting the degree of matching between the unit state and peak-shaving requirements. Effect indicators are performance parameters that evaluate the quality of regulation schemes, such as regulation time, overshoot, and coal consumption changes. Before clustering analysis, the historical schemes need to be preprocessed to eliminate the influence of different dimensions and make various indicators comparable. Using the current regulation feature quantity as the query point, the nearest neighbor schemes are searched in the scheme group to form an initial scheme set. The query point is the regulation feature quantity vector corresponding to the current unit state. The nearest neighbor search finds the several historical schemes that are closest to the query point in the multi-dimensional feature space. The distance metric usually uses Euclidean distance or Mahalanobis distance, and the number of nearest neighbors is determined according to the size and diversity of the scheme database. A genetic algorithm is applied to iteratively optimize the initial set of solutions. The genetic algorithm generates new solutions through crossover and mutation operations. Each regulation solution is encoded as a chromosome, with genes on the chromosome representing different regulation parameters, such as coal feed rate, damper opening, and steam valve position. The crossover operation swaps partial gene segments between two parent chromosomes, and the mutation operation randomly changes the values ​​of individual genes within the chromosome. Individuals that satisfy the power grid peak-shaving command constraints are selected from the new solutions, and a set of regulation solutions is output. Constraints include load change rate limits, equipment safety operation boundaries, and environmental emission standards. The selection process checks whether each newly generated solution meets all constraints.

[0042] In practical implementation, a clustering algorithm is used to group the historical regulation scheme database. By preprocessing the historical regulation schemes, feature vectors are extracted, including regulation parameters and operating context. Preprocessing includes data cleaning, missing value imputation, and outlier removal to ensure data quality meets clustering requirements. Feature vectors contain numerical and categorical parameters. Numerical parameters include continuous variables such as load setpoints and main steam pressure, while categorical parameters include discrete variables such as burner combination modes and pulverizing system operating modes. Categorical parameters need to be converted to numerical form using one-heat encoding. A density clustering algorithm is applied to divide the scheme groups. The density clustering algorithm is based on neighborhood radius and minimum number of points. Density clustering can discover clusters of arbitrary shapes and does not require pre-specifying the number of clusters. The neighborhood radius determines the neighborhood range of a sample point, and the minimum number of points specifies the minimum number of samples in a cluster. The clustering quality is evaluated by using a silhouette coefficient to optimize the number of clusters. The silhouette coefficient calculates the average distance between each sample point and other points in the same cluster, as well as the average distance to the nearest point in a different cluster. The coefficient ranges from -1 to 1; a larger value indicates a better clustering effect. By testing different combinations of neighborhood radius and minimum number of points, the parameter settings that maximize the average profile coefficient are selected. An index is built for each scheme group to facilitate fast retrieval. The index structure uses a spatial partitioning tree or hash table, supporting efficient range queries and k-nearest neighbor searches. After the index is built, a new query point can find its cluster and its nearest neighbors in logarithmic time complexity.

[0043] In some embodiments, the fitness function design of the genetic algorithm considers multi-objective optimization, and the fitness function includes multiple evaluation dimensions such as adjustment speed, stability, and economy. The fitness function can be expressed as a weighted sum: in: It is the overall fitness value. It's about adjusting the speed score. It is a stability score. It's an economic score. , , These are the weighting coefficients for each dimension. The adjustment speed score is calculated based on the response time of the scheme to reach the target load, the stability score is assessed based on the load fluctuation amplitude and frequency, and the economic score considers operating costs such as coal consumption and plant power consumption. The weighting coefficients are dynamically adjusted according to the current peak-shaving target, with increased weighting for adjustment speed during emergency peak-shaving and emphasis on economic indicators during normal peak-shaving.

[0044] It's understandable that determining the number of clusters requires balancing intra-cluster similarity and inter-cluster differences. Intra-cluster similarity requires schemes within the same cluster to have similar characteristics, while inter-cluster differences require schemes in different clusters to be significantly different. Silhouette coefficient analysis can help find the optimal number of clusters, but in practical applications, the size of the scheme library and usage requirements must also be considered. For large scheme libraries, the number of clusters can be appropriately increased to improve retrieval accuracy; for small scheme libraries, the number of clusters should be controlled to avoid over-segmentation. Clustering results need expert verification to ensure that the classification results conform to operational experience. Nearest neighbor search can use approximate nearest neighbor algorithms to improve efficiency, which reduce computation through spatial partitioning or hashing techniques. When the scheme library is large, the computational cost of exact nearest neighbor search may not meet real-time requirements. Approximate nearest neighbor algorithms significantly improve search speed within an acceptable error range and are suitable for online application scenarios. The error rate of approximate nearest neighbor algorithms needs to be controlled within an allowable range, generally requiring the overlap between retrieved nearest neighbor schemes and exact results to reach over 90%.

[0045] In some embodiments, the operational parameters of the genetic algorithm need to be carefully set according to the characteristics of the problem. These parameters include population size, crossover probability, and mutation probability. Population size affects the search range of the algorithm and is generally set to 50-200 individuals. Crossover probability controls the frequency of gene exchange and is usually set to 0.6-0.9. Mutation probability determines the probability of gene mutation and is generally set to 0.01-0.1. These parameters need to be determined through experimental tuning to find their optimal values; different systems may require different parameter settings. The parameter tuning process can be systematically optimized using orthogonal experimental design or response surface methodology. Constraint handling is a crucial step in genetic algorithm design, ensuring that the newly generated solution satisfies all operational constraints. Constraint handling methods include penalty function methods and feasible solution retention methods. The penalty function method transforms constraint violations into fitness penalties, guiding the search towards the feasible region. The feasible solution retention method maintains a dedicated set of feasible solutions, prioritizing crossover and mutation operations among feasible solutions. In practical applications, multiple constraint handling methods can be combined to improve the convergence and solution quality of the algorithm.

[0046] In some embodiments, the scheme library needs to establish an update mechanism that periodically incorporates new successful adjustment cases and eliminates outdated schemes. The update cycle is determined based on the frequency of changes in unit operating status, typically quarterly or semi-annually. The addition of new schemes requires effect verification; only adjustment cases that achieve the expected results are added to the library. Outdated schemes are identified based on timestamps and usage frequency; old schemes that have not been used for a long time can be removed. Updating the scheme library maintains the timeliness of knowledge, enabling the system to adapt to changes in unit operating characteristics. The encoding method of the adjustment scheme affects the search efficiency of the genetic algorithm. The encoding method needs to accurately represent the adjustment parameters while ensuring the feasibility of gene operations. Real-number encoding directly uses the actual values ​​of the parameters as gene values, suitable for continuous adjustment parameters. Integer encoding maps discrete operation options to integer values, suitable for switch quantities or mode selection parameters. Hybrid encoding contains both real-number and integer genes, capable of fully representing complex adjustment schemes. The design of the encoding scheme needs to fully consider the coupling relationships between parameters, ensuring that related parameters are arranged adjacently on the chromosome, facilitating meaningful crossover operations.

[0047] Example 4: In practical implementation, the performance indicators of each regulation scheme are evaluated by constructing a multi-dimensional evaluation index system, covering multiple aspects such as energy consumption, emissions, and equipment stress. Energy consumption indicators include quantitative parameters such as power generation coal consumption rate, heating coal consumption rate, and plant power consumption rate. Emission indicators involve environmental parameters such as sulfur dioxide emission concentration, nitrogen oxide emission concentration, and dust emission concentration. Equipment stress indicators cover mechanical and electrical parameters such as turbine rotor thermal stress, boiler tube wall temperature, and generator winding temperature. The establishment of the multi-dimensional evaluation index system is based on the design specifications, operating procedures, and industry standards of coal-fired power generating units. Each indicator has a clear definition and calculation method, and the indicator data comes from the unit's real-time monitoring system and historical operation database. Grey system theory is used to calculate the correlation between each regulation scheme and the ideal scheme; the correlation reflects the superiority or inferiority of the scheme. Grey system theory deals with incomplete information systems, analyzing the degree of influence between factors by generating grey sequences and calculating correlation coefficients. The characteristic sequence of the ideal scheme is extracted from historical best operating cases, and the characteristic values ​​include the optimal or target values ​​of each evaluation index. The characteristic sequences of each evaluation scheme are normalized to eliminate magnitude differences. Normalization methods include initialization or mean normalization. The grey relational coefficient is used to calculate the local similarity between the scheme sequence and the ideal sequence for each indicator. The correlation coefficient ranges from 0 to 1, with larger values ​​indicating a closer approximation to the ideal state. The overall correlation degree is obtained by aggregating the correlation coefficients of each indicator, using either arithmetic mean or weighted average. Weighting coefficients are set and dynamically adjusted based on expert knowledge. A comprehensive score is obtained by weighted fusion of the correlation degrees. The determination of the weighting coefficients considers the importance of the indicators and peak-shaving targets, increasing the weight of emission indicators during periods of grid load emergency and emphasizing equipment stress indicators during equipment maintenance. Expert knowledge is stored in a rule base containing weight allocation strategies for different operating conditions. Weighted fusion uses a linear weighted sum formula; the comprehensive score is the sum of the products of the correlation degree of each dimension and its corresponding weight. All comprehensive scores are compared, and the adjustment scheme with the highest score is selected as the optimal adjustment scheme. The comparison process sorts all candidate schemes, selecting not only the highest-scoring scheme but also retaining the second-best scheme as a backup option to improve the system's robustness.

[0048] In practical implementation, the construction of a multi-dimensional evaluation indicator system requires determining the number and hierarchical structure of indicators. Too few indicators will miss key information, while too many will increase computational complexity. The indicator system adopts a three-level structure: primary indicators are categorized into three main types: energy consumption, emissions, and equipment stress; secondary indicators are specific parameters under each category; and tertiary indicators are detailed measurements of the parameters. Indicator values ​​are obtained in real time through a data acquisition system. For indicators that cannot be directly measured, soft measurement techniques are used for indirect calculation. The application of grey system theory includes establishing a grey relational analysis model. The model input is the characteristic sequence of the scheme, and the output is the correlation degree value. The construction of the characteristic sequence requires a unified time scale, with all indicator values ​​corresponding to the same sampling time or operational stage. The correlation coefficient is calculated using the grey relational degree formula, and the resolution coefficient is set according to the data characteristics, typically set to 0.5. The calculation of the overall correlation degree considers the correlation between indicators to avoid duplicate calculations that could affect the evaluation results.

[0049] In implementation, the dynamic adjustment mechanism for weighting coefficients is based on an expert knowledge base and real-time operational status. The expert knowledge base stores the empirical rules of domain experts, expressed in "if-then" form. For example, if the current air quality index exceeds the standard, then the weight of emission indicators is increased. Real-time operational status includes factors such as unit load rate, ambient temperature, and fuel quality, which affect the importance ranking of indicators. The weighting adjustment frequency is synchronized with the peak-shaving command update cycle to ensure that the weighting settings are consistent with the current peak-shaving target. The weighted fusion process uses matrix operations to improve computational efficiency, and the comprehensive score calculation formula is a linear combination, ensuring the additivity and comparability of scores.

[0050] In practical implementation, referring to Table 1, the process of comparing comprehensive scores requires addressing the identification of solutions with similar scores. When the difference in comprehensive scores among multiple solutions is less than a set threshold, an auxiliary decision-making mechanism is activated. This mechanism considers factors such as the implementation difficulty, risk coefficient, and backup requirements of the solutions, conducting a secondary evaluation. The score comparison algorithm employs quicksort or heapsort to ensure efficient identification of the optimal solution from a large set of solutions. The selection result of the best solution needs to be recorded in a decision log, including the score calculation process, the basis for weight setting, and the comparison results, for post-analysis and system optimization.

[0051] Table 1 illustrates the basic structure of the multi-dimensional evaluation index system. In some embodiments, a dynamic weight allocation algorithm can be introduced into the multi-dimensional evaluation index system. This algorithm automatically adjusts the weights based on the real-time changes in index values. When an index value deviates significantly from the normal range, its weight is automatically increased, enhancing the system's sensitivity to abnormal states. The dynamic weight allocation algorithm is implemented based on sliding window statistics, using the variance and range of index values ​​within the window as the basis for weight adjustment. This mechanism enables the evaluation system to adaptively focus on the most prominent issues. The calculation of the grey relational coefficient can incorporate a time dimension, considering the changing trend of index values ​​rather than just instantaneous values. Calculating the dynamic correlation coefficient between the scheme sequence and the ideal sequence on the time axis better reflects the dynamic performance of the adjustment scheme. The dynamic correlation coefficient is calculated based on grey slope correlation, considering both numerical proximity and consistency of changing trends. This method has high computational complexity and is suitable for scenarios with strict time requirements for the adjustment process.

[0052] Optionally, the weighted fusion of the comprehensive scores can employ nonlinear fusion methods, such as geometric mean or harmonic mean, suitable for situations where there are strong interaction effects between indicators. Nonlinear fusion can better handle the compensation effect between indicators, avoiding over-compensation by other indicators when one indicator performs extremely poorly. The parameters of the nonlinear fusion method need to be determined based on historical data training to ensure the rationality of the fusion results. The characteristic sequence of the ideal solution needs to be updated regularly, and the update of the characteristic sequence should be adjusted according to changes in unit performance and technological advancements. The update cycle is synchronized with the unit overhaul cycle, and the optimal values ​​of each indicator are re-evaluated after each overhaul. The characteristic sequence update process requires expert review to ensure the scientific validity and feasibility of the new benchmark. The update mechanism ensures that the evaluation criteria remain consistent with the actual state of the unit.

[0053] See Figure 5 This heatmap focuses on five core parameters: boiler temperature, turbine speed, generator voltage, load demand, and dynamic deviation. Through color gradients and numerical annotations, it quantifies the correlation coefficients between these parameters. This quantitative correlation analysis provides crucial data support for adaptive peak-shaving regulation of coal-fired power units: strongly correlated parameters can be used as joint features for peak-shaving status identification, while weakly correlated parameters can be modeled separately, thereby improving the accuracy of peak-shaving status classification and regulation scheme generation. It represents a significant visualization achievement of data-driven peak-shaving decision-making at the basic data layer, laying a logical foundation for subsequent peak-shaving status analysis and regulation feature calculation.

[0054] Example 5: In practical implementation, the grey system theory calculates the correlation between each regulation scheme and the ideal scheme by defining the characteristic sequence of the ideal scheme. The characteristic sequence is derived from the best historical cases. The process of constructing the characteristic sequence of the ideal scheme involves analyzing the historical operation database of coal-fired power generating units, selecting cases with the best regulation effect, and extracting the key parameter values ​​of these cases after regulation stabilization as the benchmark of the characteristic sequence. The characteristic sequence contains index values ​​of multiple dimensions, such as boiler main steam temperature, turbine speed, generator power factor, and other operating parameters. Each parameter corresponds to a characteristic value in the characteristic sequence. The values ​​of the characteristic sequence represent the level that each operating parameter should reach under ideal conditions and serve as a reference standard for evaluating other regulation schemes. The characteristic values ​​of each regulation scheme are normalized to eliminate the influence of orders of magnitude. The normalization process is performed separately for each operating parameter, converting the original data of different dimensions into a unified relative value. The normalization method adopts initialization processing, that is, dividing the actual value of each parameter by the benchmark value of the corresponding parameter in the characteristic sequence of the ideal scheme, so that all parameter values ​​are converted into dimensionless values ​​floating around 1. For interval-type parameters, such as the allowable temperature range, mean normalization is used to map the actual values ​​to the 0-1 interval. The grey relational coefficient is calculated, reflecting the local similarity between the proposed solution and the ideal solution. The grey relational coefficient calculation is based on the normalized feature value sequence, calculating the correlation coefficient separately for each parameter dimension. The correlation coefficient calculation formula considers the absolute difference and range between the proposed feature values ​​and the ideal feature values, with a resolution coefficient of 0.5. The correlation coefficient ranges from 0 to 1; a larger value indicates that the proposed solution performs closer to the ideal state in that parameter dimension. The correlation coefficients are aggregated to obtain the overall correlation degree, with a value range of 0 to 1. The aggregation process uses the arithmetic mean method, summing the grey relational coefficients for each parameter dimension and taking the average. The closer the overall correlation degree value is to 1, the closer the overall performance of the adjustment solution is to the ideal solution, and the better the solution is.

[0055] In practical implementation, defining the characteristic sequence of the ideal solution requires establishing case selection criteria, which include multiple performance indicators such as settling time, overshoot, and stability. The selection of optimal historical cases employs a multi-condition sorting method: first, sorting by settling time in ascending order; then, sorting by overshoot in ascending order for cases with the same settling time; and finally, sorting by stable time in ascending order. The top 5% of cases in the overall ranking are selected as the ideal case set, and the average value of each parameter in these cases is calculated as the benchmark value for the characteristic sequence. The characteristic sequence is updated every six months, with each update re-screening the operating cases from the most recent six months to ensure that the characteristic sequence reflects the unit's current optimal performance level. Normalization requires distinguishing between benefit-oriented parameters and cost-oriented parameters. Benefit-oriented parameters, such as thermal efficiency and power generation efficiency, are better the higher they are; cost-oriented parameters, such as coal consumption rate and emission concentration, are better the lower they are. Benefit-oriented parameters are normalized using an upper limit effect measure, while cost-oriented parameters are normalized using a lower limit effect measure, ensuring that all parameters are optimally higher after normalization.

[0056] In practical implementation, the calculation of grey relational coefficients needs to consider the weight differences between parameters, assigning higher weights to important parameters. Weight allocation is based on the degree of impact of parameters on unit operation safety and economy, determined by expert scoring. For example, the weight coefficient for boiler main steam temperature is 0.2, the weight coefficient for turbine speed is 0.15, and the weight coefficient for generator power factor is 0.1. Weighted relational coefficient calculation introduces weighting factors into the original formula, making the relational coefficients of important parameters contribute more to the overall relational degree. The calculation of relational coefficients also needs to handle the impact of outliers. When a parameter value deviates too much from the ideal value, the relational coefficient for that parameter is truncated to avoid excessive impact of abnormal fluctuations in a single parameter on the overall evaluation result. The method for aggregating relational coefficients can be adjusted according to evaluation needs. The arithmetic mean method is suitable when the importance of each parameter is roughly equal. When the differences in parameter importance are significant, the weighted average method is used for aggregation, with the weight coefficients consistent with the weights in the relational coefficient calculation. For evaluation scenarios that need to highlight weaknesses, the minimum relational coefficient is used as the overall relational degree, reflecting the evaluation concept of the "weakest link" principle. The selection of aggregation methods is achieved through configuration parameters, which can be flexibly switched according to different peak-shaving targets.

[0057] In some embodiments, characteristic sequences can be established separately for each load range, which is divided into low load range (30%-50% of rated load), medium load range (50%-80% of rated load), and high load range (80%-100% of rated load). Different load ranges correspond to different ideal operating parameters, and characteristic sequences are established for each of the three load ranges. During evaluation, the corresponding characteristic sequence is selected for calculation based on the target load of the adjustment scheme, improving the accuracy of the evaluation. The boundaries of the load range division can be adjusted according to the actual operating characteristics of the unit, generally selecting the inflection point of the efficiency curve as the range boundary. Grey system theory calculation can introduce a time series similarity metric, which considers the entire process of parameter change, not just the steady-state value. The similarity between the scheme parameter change curve and the ideal change curve is calculated using a dynamic time warping algorithm, and the curve similarity is used as a new feature dimension in the correlation calculation. This method can evaluate the smoothness of the adjustment process and avoid using drastic adjustment methods to pursue the final state. The window size of the dynamic time warping algorithm is set according to the unit's inertial time constant, typically 5-10 sampling periods.

[0058] Optionally, normalization can employ fuzzy membership functions, which map parameter values ​​to membership degrees in fuzzy sets such as "excellent," "good," "medium," and "poor." For example, the degree of belonging to "excellent" for the boiler main steam temperature within ±5℃ of the ideal value is 1, and the degree of belonging to "good" within ±10℃ is 0.7, etc. Fuzzy normalization better reflects the hierarchical characteristics of parameter values ​​and adapts to the evaluation characteristics of operating parameters. The shape and boundary of the fuzzy membership function need to be determined by experts according to operating procedures. A sliding window mechanism can be introduced for correlation calculation. This mechanism calculates the correlation degree at multiple moments during the adjustment process and then aggregates them to obtain the overall correlation degree. The window size is set according to the duration of the adjustment process, typically dividing the entire adjustment process into 5-10 equal windows. Instantaneous correlation degree is calculated for each window, and the final correlation degree is the average or minimum of the correlation degrees from all windows. This method can comprehensively evaluate the performance of the entire adjustment process, rather than just focusing on the final state. A virtual ideal point can be introduced into the characteristic sequence of the ideal scheme. The virtual ideal point is the theoretical optimal value of each parameter, not the historical best value. Virtual ideal points are determined based on unit design parameters, equipment performance limits, and theoretical calculations, representing the theoretically achievable optimal state. Using virtual ideal points can raise evaluation standards and push operational levels closer to theoretical limits. Virtual ideal points need to be reviewed periodically and adjusted accordingly with technological advancements and equipment upgrades.

[0059] It is understandable that the resolution coefficient used in grey system theory to calculate the correlation coefficient affects its discriminative power; the smaller the resolution coefficient, the more sensitive the correlation coefficient is to differences. The resolution coefficient is typically set to 0.5, but in practical applications, it can be adjusted based on the parameter measurement accuracy. Parameters with high measurement accuracy can use a smaller resolution coefficient, while parameters with low measurement accuracy can use a larger resolution coefficient. The resolution coefficient should be set to avoid the correlation coefficients being too concentrated or dispersed, maintaining appropriate evaluation sensitivity.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive regulating method for full load peak shaving of a coal-fired power generating unit, characterized in that, The method realizes the following processes through a data acquisition control system: Collecting operation parameters of a coal-fired generating unit and grid peak regulation instructions, the operation parameters including boiler temperature, turbine speed and generator voltage, and the grid peak regulation instructions including real-time load demand and predicted load fluctuation; Performing peak regulation state analysis using the collected operation parameters and grid peak regulation instructions to obtain a current peak regulation state, the current peak regulation state being divided into a stable operation state, a transition regulation state and a limit peak regulation state; Calculating a regulation characteristic quantity according to the current peak regulation state and the operation parameters, the regulation characteristic quantity representing a matching relationship between unit performance and peak regulation requirements; Generating a set of regulation schemes based on the regulation characteristic quantity and the grid peak regulation instructions, each regulation scheme involving combustion optimization, steam adjustment and power control; Evaluating performance indicators of each regulation scheme, the performance indicators including efficiency loss, stability influence and maintenance cost, and selecting an optimal regulation scheme; Executing the optimal regulation scheme to realize adaptive peak regulation of the coal-fired generating unit.

2. The method of adaptive regulation of full load peaking of a coal-fired power generating unit according to claim 1, characterized in that, The peak regulation state analysis using the collected operation parameters and grid peak regulation instructions is realized through the following steps: A state classification model is constructed, which is trained based on a historical operation data set of the coal-fired generating unit, the historical operation data set containing parameter sequences under multiple working conditions and corresponding peak regulation state labels; An ensemble learning framework is used to train the state classification model, the ensemble learning framework fusing outputs of multiple base classifiers, the base classifiers including decision trees, a nearest neighbor algorithm and a naive Bayes classifier; Real-time operation parameters are segmented by a sliding window to generate time sequence segments, and statistical features including mean, variance and trend coefficient are extracted; The statistical features and the grid peak regulation instructions are input into the state classification model, the model outputting a probability distribution, and the current peak regulation state is determined through threshold comparison.

3. The method of adaptive regulation of full load peaking of a coal-fired power generating unit according to claim 1, characterized in that, The calculation of the regulation characteristic quantity according to the current peak regulation state and the operation parameters is realized through the following steps: A reference parameter template corresponding to the peak regulation state is established, the reference parameter template defining a value range of ideal operation parameters; A dynamic deviation degree of real-time operation parameters from the reference parameter template is calculated, the dynamic deviation degree being calculated by combining Euclidean distance and cosine similarity; An urgency indicator of the grid peak regulation instruction is introduced, the urgency indicator being derived based on a load change rate and a time constraint; The dynamic deviation degree and the urgency indicator are fused to generate the regulation characteristic quantity by weighted summation, the regulation characteristic quantity being output as a multi-dimensional vector.

4. The method of adaptive regulation of full load peaking of a coal-fired power generating unit of claim 1, wherein, The generation of a set of regulation schemes based on the regulation characteristic quantity and the grid peak regulation instruction is realized through the following steps: A clustering algorithm is used to group a historical regulation scheme library, the clustering algorithm dividing scheme groups according to the regulation characteristic quantity and effect indicators; A current regulation characteristic quantity is taken as a query point to search for neighboring schemes in the scheme groups to form an initial scheme set; A genetic algorithm is applied to iteratively optimize the initial scheme set, the genetic algorithm generating new schemes through crossover and mutation operations; Individuals in the new schemes that meet constraints of the grid peak regulation instruction are screened, and a set of regulation schemes is output.

5. The method of adaptive regulation of full load peaking of coal-fired power generating units according to claim 1, characterized by that, The evaluation of performance indicators of each regulation scheme is realized through the following steps: A multi-dimensional evaluation index system is constructed, the index system covering energy consumption, emissions and equipment stress; The grey system theory is used to calculate the correlation degree of each adjustment scheme and an ideal scheme, and the correlation degree reflects the advantages and disadvantages of the scheme. A weight coefficient is set, the weight coefficient is dynamically adjusted based on expert knowledge, and the correlation degrees are weighted and fused to obtain a comprehensive score. All comprehensive scores are compared, and the adjustment scheme with the highest score is selected as the best adjustment scheme.

6. The method of adaptive regulation of full load peaking of a coal-fired power generating unit of claim 2, wherein, The state classification model is trained using an ensemble learning framework, which is implemented by the following steps: Prepare a training sample set containing labeled peak shaving state data. Each base classifier is independently trained, and ten-fold cross-validation is used to evaluate the accuracy of the base classifier. A fusion strategy is designed, which uses a weighted voting method, and the weights are assigned based on the accuracy of the base classifier. The overall performance of the state classification model is verified, and the model parameters are adjusted through the confusion matrix.

7. The adaptive method for full load peaking of coal-fired power generating units as claimed in claim 3 wherein, The dynamic deviation degree of the real-time operation parameter and the benchmark parameter template is calculated, which is implemented by the following steps: Standardize the operation parameters to eliminate dimensional differences. Extract the main components of the parameters, which are obtained by principal component analysis method. Calculate the similarity measure of the main components and the benchmark parameter template using Mahalanobis distance. Combine the time sliding window to dynamically update the deviation degree calculation result.

8. The method of adaptive regulation of full load peaking of a coal-fired power generating unit of claim 4, wherein, The historical adjustment scheme library is grouped using a clustering algorithm, which is implemented by the following steps: Preprocess the historical adjustment scheme to extract the feature vector including the adjustment parameters and the running context. Apply a density clustering algorithm to divide the scheme group, which is based on the neighborhood radius and the minimum point number parameters. Evaluate the clustering quality and optimize the number of clusters using the silhouette coefficient. Index each scheme group for fast retrieval.

9. The method of adaptive regulation of full load peaking of a coal-fired power generating unit of claim 5, wherein, The grey system theory is used to calculate the correlation degree of each adjustment scheme and an ideal scheme, which is implemented by the following steps: Define the feature sequence of the ideal scheme, which is derived from the optimal historical case. Normalize the feature values of each adjustment scheme to eliminate the influence of the order of magnitude. Calculate the grey correlation coefficient, which reflects the local similarity between the scheme and the ideal scheme. Aggregate the correlation coefficients to obtain the overall correlation degree, and the correlation degree value range is 0 to 1.

10. An adaptive regulating system for full load peaking of a coal-fired power generating unit, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor implements the steps of the adaptive adjustment method of the full-load peak shaving of the coal-fired generating unit according to any one of claims 1 to 9 when executing the computer program.

Images