Performance prediction method for frequency modulation of thermal power unit AGC

Abstract

This invention belongs to the field of AGC frequency regulation technology for thermal power units, and discloses a performance prediction method for AGC frequency regulation of thermal power units. It constructs a multi-source heterogeneous data acquisition system covering the entire frequency regulation chain, achieves multi-scale data synchronization through timestamp alignment, and completes data preprocessing by combining physical constraints. Simultaneously, it employs the CCM-PCMCI causal filtering algorithm with physical constraints of the unit's thermal system to accurately distinguish between causal and pseudo-correlation relationships, eliminate redundant features, and constructs a multi-scale time-series feature adaptive extraction system to form a full-dimensional feature library, comprehensively characterizing the dynamic characteristics of frequency regulation, improving the model's generalization ability and interpretability, and avoiding prediction bias caused by feature redundancy or missing information. It optimizes the density peak clustering method to achieve adaptive partitioning of typical operating condition clusters for unit AGC frequency regulation, and constructs a TFT model with an adaptive attention mechanism and a mechanism constraint layer for each operating condition cluster, which can dynamically focus on core features and key time segments.

Description

Technical Field

[0001] This invention belongs to the field of AGC frequency regulation technology for thermal power units, specifically a performance prediction method for AGC frequency regulation of thermal power units. Background Technology

[0002] Automatic generation control (AGC), as a core technology for secondary frequency regulation of the power grid, is a key support for maintaining the grid frequency within the rated range and ensuring the safe and stable operation of the grid. Thermal power units, with their strong regulation capabilities and high response reliability, remain the main resource for AGC frequency regulation in my country's power grid. Their AGC frequency regulation performance not only directly determines the upper limit of the grid's frequency regulation capacity but also directly affects the unit's frequency regulation revenue and performance evaluation results in the power ancillary services market.

[0003] As a complex thermodynamic system with strong correlation, large lag, and strong nonlinearity, the AGC frequency regulation process of thermal power units is affected by multiple factors such as boiler combustion, turbine regulation, coal quality fluctuations, and equipment aging. Existing AGC frequency regulation performance prediction methods for thermal power units are mainly divided into three categories: mechanistic modeling-based methods rely on precise unit physical parameters, which are difficult to fully characterize the strong nonlinearity and multi-factor correlation characteristics of the unit, resulting in poor adaptability; traditional data-driven methods (such as BP neural networks and support vector regression) do not require complex mechanistic modeling, but they suffer from weak generalization ability, low prediction accuracy under small sample conditions, and susceptibility to getting trapped in local optima. While deep learning methods (such as LSTM) have improved the ability to fit time series data, they still have core shortcomings: the data source is singular, focusing only on real-time operating data of the unit and failing to fully cover multiple influencing factors such as AGC commands, coal quality, and equipment status, resulting in missing feature information; feature engineering is based solely on correlation analysis, failing to distinguish between causal relationships and spurious correlations, which easily leads to feature redundancy and insufficient generalization ability; it is not adapted to the multi-condition characteristics of thermal power units, and lacks the ability to predict small sample conditions such as deep peak shaving and extreme coal quality; and it does not consider the slow time-varying characteristics of unit equipment aging and coking of heating surfaces, which easily leads to a decline in prediction accuracy after long-term model operation. Summary of the Invention

[0004] The purpose of this invention is to provide a performance prediction method for AGC frequency regulation of thermal power units, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a performance prediction method for AGC frequency regulation of thermal power units, comprising: Preferably, the data preprocessing stage constructs a multi-source heterogeneous data acquisition system covering all dimensions of AGC frequency regulation. The data sources include second-level real-time unit operation data, millisecond-level AGC scheduling command timing and grid frequency deviation data, minute-level online coal quality detection data, daily-level unit equipment status and slow time-varying data, historical frequency regulation assessment and performance data, and environmental parameter data, to achieve full coverage of the influencing factors of the entire frequency regulation chain. To address the characteristics of multi-source data with different sampling frequencies and time scales, a time-stamp alignment and multi-scale data synchronization method based on AGC frequency modulation event triggering is proposed. Taking AGC command change events as the benchmark, the inertial time constant of the unit's thermal system is introduced as the synchronization weight. Combined with the correlation response characteristics of the boiler and turbine, the time dimension synchronization of data with different sampling frequencies is achieved. A preprocessing workflow based on the physical constraints of frequency regulation operating conditions is constructed. An improved 3σ criterion combined with the physical boundary threshold of unit operation is used to remove abnormal data that do not conform to the operating law of the unit's thermal system. For missing values, a similar sample filling method based on clustering under the same operating condition is adopted to preserve the operating condition characteristics of the data. Abnormal data is classified and labeled by tracing the source of abnormal data. Finally, standardization processing is adopted by operating condition, and the operating condition clusters of different load segments and coal qualities are standardized separately.

[0006] Preferably, the feature construction stage, based on preprocessed multi-source heterogeneous data, clarifies the core prediction target set of AGC frequency regulation performance, including core performance indicators for power grid assessment, frequency regulation energy consumption increment indicators, and frequency regulation revenue correlation indicators, thereby achieving a deep binding between prediction targets and actual application value; A causal feature screening framework based on the CCM-PCMCI causal screening algorithm is adopted. The physical constraints of the unit's thermal system are used as the prior conditions for causal screening. The threshold for conditional independence test is optimized to distinguish between causal relationships and pseudo-correlation relationships between features and prediction targets. Redundant features are eliminated to screen out the core causal feature set that has a direct driving effect on AGC frequency regulation performance. To address the strong temporal dynamics of AGC frequency regulation, a multi-scale adaptive extraction system for temporal features is constructed. Based on different time sliding windows, temporal features including the amplitude, slope, rate of change, duration, and frequency of reverse regulation of AGC commands are extracted. The lag time, inertia time, and temporal evolution features of the unit's frequency regulation response are also extracted. In addition, the disturbance temporal features of grid frequency deviation and the energy storage-assisted frequency regulation response features are combined. Finally, a full-dimensional feature library is constructed, including static features, slow time-varying features, dynamic temporal features, and energy storage synergy features, to comprehensively characterize the full-dimensional dynamics of AGC frequency regulation.

[0007] Preferably, the model construction stage is based on the constructed full-dimensional feature library, and adopts an adaptive partitioning method for AGC frequency regulation conditions based on density peak clustering. The distance calculation method of density peak clustering is optimized. The five core clustering features are unit load rate, coal calorific value, AGC command characteristics, ambient temperature, and slow time-varying factor. The cluster centers are corrected in combination with the physical constraints of unit operation. Unsupervised adaptive clustering is performed on historical frequency regulation condition samples to automatically divide typical condition clusters including deep peak regulation low load condition, rated load condition, coal quality large fluctuation condition, frequent command change condition, and extreme environment condition, so as to realize the division of the unit's full operating conditions. For each typical operating condition cluster, a basic model for frequency regulation performance prediction based on the TFT model and with built-in operating condition adaptive attention mechanism is constructed. It can dynamically adjust the feature attention weight according to the current operating condition type, focus on the features and time segments that have the greatest impact on frequency regulation performance under the current operating condition, and embed a simplified unit mechanism model as a constraint layer to capture the nonlinear dynamic characteristics of the AGC frequency regulation process. A cross-unit transfer learning mechanism is adopted. For the baseline operating condition cluster with sufficient sample size, the basic model is pre-trained. For the small sample operating condition cluster, including deep peak regulation and extreme coal quality, the feature extraction layer parameters of the baseline operating condition pre-trained model are transferred. At the same time, historical operating condition data of the same type of unit is introduced as auxiliary transfer samples. The output layer of the model is fine-tuned and adapted only based on the small sample operating condition data.

[0008] Preferably, the dynamic correction stage constructs a quantitative index system and a real-time identification module for the slow time-varying characteristics of the thermal power unit. The quantitative index system includes equipment aging factors based on the unit's equivalent operating time and overhaul cycle, coal quality fluctuation and heating surface coking factors based on the deviation between the actual efficiency and design efficiency of the boiler, and actuator aging factors based on the deviation between the actual flow characteristics and design characteristics of the regulating valve. The real-time identification module, combined with digital twin technology, constructs a virtual simulation model of the unit to achieve real-time quantitative characterization and dynamic updating of the slow time-varying characteristics. A two-level adaptive dynamic correction module is designed. The first level is a feedforward correction based on a slow time-varying factor. The quantized slow time-varying factor is input into the frequency regulation performance prediction basic model. Combined with the prediction results of the unit virtual simulation model, the prediction output of the basic model is compensated by feedforward. The second level is a sliding window online feedback correction. A sliding time window is set on a daily basis. Based on the latest frequency regulation performance measured data and abnormal data tracing results within the window, the parameters of the basic model are finely adjusted using an incremental learning method. At the same time, the quantized value of the slow time-varying factor and the operating condition cluster center are updated. A model accuracy warning and triggering mechanism is set up so that when the prediction error of the prediction model exceeds the preset threshold, the full incremental training of the basic model is automatically triggered, and the working condition cluster center, model core parameters and feature selection threshold are automatically updated. A model failure fallback switching logic is set up so that when the correction module fails, it automatically switches to the simplified mechanism model for auxiliary prediction.

[0009] Preferably, in the timing optimization stage, an AGC instruction timing embedding module is constructed based on the dynamically corrected prediction model. The AGC instruction sequence and the power grid frequency deviation sequence issued by the power grid dispatching agency are embedded into the encoder of the prediction model. The attention mechanism of the AGC instruction timing embedding module highlights key instruction segments. The design module for real-time identification and dynamic compensation of the frequency regulation response lag characteristics of the unit is designed. Based on real-time operating data, the recursive least squares method combined with the forgetting factor is used to identify the dynamic characteristic parameters of the unit, including the pure lag time, inertial time constant, and rise rate, online. The compensation coefficient is dynamically adjusted according to the identification results to adaptively compensate the prediction output of the prediction model. The prediction model can simultaneously output AGC frequency modulation performance prediction results at different time steps, realize the performance prediction of the entire response process of a single AGC command, and display the deviation trend between the predicted value and the measured value in real time through the prediction result visualization module.

[0010] Preferably, the energy consumption prediction stage combines the time-series optimized AGC frequency regulation performance prediction results to construct a collaborative prediction framework for frequency regulation performance, energy consumption, and revenue. Under this collaborative prediction framework, the core performance indicators of the power grid assessment are predicted. At the same time, a sub-model for predicting the incremental frequency regulation energy consumption during the AGC frequency regulation process is built. Based on the power supply coal consumption characteristic curve of the unit operating under different conditions, combined with real-time coal quality fluctuation data and energy storage-assisted frequency regulation energy consumption data, the incremental power supply coal consumption of the unit under different AGC frequency regulation commands and different regulation rates is quantified, thereby achieving a three-dimensional collaborative prediction of frequency regulation response performance, frequency regulation energy consumption cost, and frequency regulation revenue. The incremental energy consumption cost and revenue under different frequency regulation scenarios are quantified. Combined with the predicted frequency regulation performance indicators, a multi-objective optimization function is constructed with the goal of minimizing energy consumption cost, maximizing frequency regulation revenue, and meeting the power grid performance assessment requirements. This provides a quantitative basis for optimizing frequency regulation control parameters and balancing frequency regulation performance and energy consumption cost. A multi-scenario frequency regulation performance simulation and prediction module was built, which can simulate frequency regulation performance and energy consumption changes under different AGC command sequences, different load conditions, different coal quality conditions, and different energy storage coordination strategies.

[0011] Preferably, the deployment and support phase constructs an online deployment architecture that coordinates the edge, cloud, and digital twin terminals. The edge side is deployed on the plant-level monitoring information system or distributed control system side of the unit to realize real-time data collection, preprocessing, second-level real-time performance prediction, and local output. The cloud deploys a model training and optimization platform to realize the aggregation and storage of multi-source data, iterative updates of the operating condition database, pre-training and incremental training of the model, and knowledge transfer and sharing among multiple units. The digital twin terminal deploys a virtual simulation model of the unit to realize simulation verification of prediction results and operating condition simulation. A model parameter iterative optimization mechanism is constructed to feed back the measured frequency regulation performance data, actual energy consumption data, and abnormal data tracing results after the unit actually executes AGC commands to the cloud model optimization platform in real time. The data results are used to optimize feature screening results, operating condition clustering centers, model parameters, and correction modules. A model reliability fallback module is set up to verify the physical boundary rationality of the prediction model output results in real time. When the prediction results exceed the physical limit range of unit operation, or the prediction error of the prediction model continuously exceeds the threshold, an early warning is automatically triggered, and the system switches to a fallback prediction mode based on a simplified mechanism model of the unit. A data security encryption module is also set up to ensure the security of multi-source data transmission and storage.

[0012] The beneficial effects of this invention are as follows: 1. This invention constructs a multi-source heterogeneous data acquisition system covering the entire frequency regulation chain, achieves multi-scale data synchronization through timestamp alignment, and completes data preprocessing by combining physical constraints; at the same time, it adopts the CCM-PCMCI causal screening algorithm with physical constraints of the unit's thermal system to accurately distinguish between causal and pseudo-correlation relationships, eliminate redundant features, and also constructs a multi-scale time series feature adaptive extraction system to form a full-dimensional feature library, comprehensively characterizing the dynamic characteristics of frequency regulation, improving the model's generalization ability and interpretability, and avoiding prediction bias caused by feature redundancy or missing information.

[0013] 2. This invention optimizes the density peak clustering method to achieve adaptive partitioning of typical operating condition clusters for unit AGC frequency regulation. For each operating condition cluster, a TFT model with an adaptive attention mechanism and a mechanism constraint layer is constructed, which can dynamically focus on core features and key time segments. At the same time, a cross-unit transfer learning mechanism is adopted to complete the model pre-training using the benchmark operating condition. For small sample operating conditions such as deep peak regulation, only the output layer is fine-tuned, which solves the problem of low prediction accuracy of traditional methods for small sample operating conditions. The mechanism constraint layer can also prevent the prediction results from exceeding the physical limit, thus achieving accurate prediction of all operating conditions of the unit.

[0014] 3. This invention constructs a quantitative index system for the slow time-varying characteristics of thermal power units and a two-level adaptive dynamic correction module. It combines digital twins to realize dynamic updates of slow time-varying characteristics, and offsets the accuracy decay caused by equipment aging through feedforward compensation and online feedback fine-tuning. It also ensures the long-term stability of the model by combining accuracy early warning and fault fallback mechanisms. At the same time, it designs a time-series embedding and lag characteristic dynamic compensation module to realize multi-time step rolling prediction. It also constructs a collaborative prediction framework for performance, energy consumption and benefits to quantify energy consumption and benefits under different scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of the AGC frequency regulation performance prediction process for thermal power units according to the present invention; Figure 2 This is a flowchart of the multi-source heterogeneous data preprocessing process of the present invention. Detailed Implementation

[0016] 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.

[0017] like Figures 1 to 2 As shown, this embodiment of the invention provides a performance prediction method for AGC frequency regulation of thermal power units, including: The data preprocessing stage involves constructing a multi-source heterogeneous data acquisition system covering all dimensions of AGC frequency regulation. The data sources include second-level real-time unit operation data (core parameters of boiler, turbine, generator, and actuator), millisecond-level AGC scheduling command timing and grid frequency deviation data, minute-level online coal quality monitoring data, daily-level unit equipment status and slow time-varying data (flow efficiency, coking degree of heating surface, valve wear characteristics, equivalent running time, etc.), historical frequency regulation assessment and performance data, and environmental parameter data, achieving full coverage of all factors affecting the entire frequency regulation chain. To address the characteristics of multi-source data with different sampling frequencies and time scales, a time-stamp alignment and multi-scale data synchronization method based on AGC frequency modulation event triggering is proposed. Taking AGC command change events as the benchmark, the inertial time constant of the unit's thermal system is introduced as the synchronization weight. Combined with the correlation response characteristics of the boiler and turbine, the time dimension synchronization of data with different sampling frequencies is achieved. A preprocessing workflow based on the physical constraints of frequency regulation conditions is constructed. An improved 3σ criterion combined with the physical boundary threshold of unit operation is adopted to remove abnormal data that does not conform to the operating law of the thermal system. The improved 3σ criterion is a dynamic adaptive σ correction criterion based on the unit's frequency regulation conditions. The value of the σ coefficient is dynamically adjusted according to the data analysis results of different load segments and coal quality condition clusters. For abnormal data detection in the sudden change stage of frequency regulation conditions, the σ coefficient is amplified to reduce the false rejection rate. For the steady-state frequency regulation stage, the σ coefficient is reduced to improve the accuracy of abnormal data identification. At the same time, the physical boundary threshold of the unit's thermal system is used as a hard constraint for the judgment of abnormal data. When the data exceeds the physical boundary, it is judged as abnormal data regardless of whether it conforms to the 3σ statistical law. For missing values, a similar sample imputation method based on clustering under the same operating conditions is adopted to preserve the operating condition characteristics of the data. Specifically, the similar sample imputation method based on clustering under the same operating conditions uses the unit load rate, coal calorific value, and AGC instruction type as key matching dimensions to calculate the feature similarity between the missing value sample and other complete samples within the same operating condition cluster. The top 5 complete samples with the highest similarity are selected as reference samples. Weights are assigned according to the similarity between each reference sample and the missing value sample, and a weighted average method is used to imput the missing data. For time-series continuous missing data, a smoothing correction is performed based on the weighted average and the changing trend of the time-series data before and after to ensure the time-series continuity and operating condition consistency of the imputed data. The abnormal data is traced back to its source and then classified and labeled, including equipment failure, measurement error, and sudden change in operating conditions. Finally, standardization is carried out by operating condition, and the operating condition clusters for different load ranges and coal qualities are standardized separately.

[0018] The source tracing of abnormal data adopts a combination of data correlation analysis and operation log verification. First, the abnormal data is correlated and matched with the operating status of the unit equipment, changes in AGC commands, and grid frequency fluctuations during the same period. Then, the source is determined by combining the operation logs of the unit's DCS system and equipment alarm information. The criteria for judging abnormal data of equipment failure are that the equipment corresponding to the time when the abnormal data occurs has clear alarm information and the equipment operating parameters deviate from the normal range. The criteria for judging abnormal data of measurement error are that only a single measuring point data is abnormal, the data of the same type of related measuring points is not abnormal, and there is no equipment alarm information. The criteria for judging abnormal data of sudden change in operating conditions are that the abnormal data is caused by external operating condition changes such as a large change in AGC commands or a sudden change in grid frequency, and multiple measuring point data show a regular deviation simultaneously.

[0019] The feature construction stage, based on preprocessed multi-source heterogeneous data, clarifies the core prediction target set for AGC frequency regulation performance, including core performance indicators for power grid assessment, frequency regulation energy consumption increment indicators, and frequency regulation benefit-related indicators, thus achieving a deep binding between prediction targets and practical application value. The core performance indicators for power grid assessment include regulation rate, regulation accuracy, response time, frequency regulation mileage, and comprehensive frequency regulation performance K-value. The frequency regulation benefit-related indicators are the theoretical benefits and penalty losses calculated based on frequency regulation assessment rules. A causal feature selection framework based on the CCM-PCMCI causal screening algorithm is adopted. The physical constraints of the unit's thermal system (such as boiler thermal inertia and turbine regulation characteristics) are used as the prior conditions for causal screening. The threshold for conditional independence test is optimized to distinguish between causal and pseudo-correlation relationships between features and prediction targets. Redundant features that only have correlation but no direct causal relationship are eliminated. The core causal feature set that has a direct driving effect on AGC frequency regulation performance is selected, reducing feature dimensionality, avoiding the curse of dimensionality, and improving the generalization ability and interpretability of the model. The optimization method for the conditional independence test threshold is as follows: based on the physical constraints of the unit's thermal system, the causal relationship between the feature and the prediction target is determined to establish a basic threshold. Then, the basic threshold is iteratively corrected through cross-validation of historical frequency regulation operating condition samples. The accuracy and recall of causal relationship identification are used as dual evaluation indicators. When both indicators reach the preset optimal value, the final test threshold is determined. For features that have a direct physical causal relationship with frequency regulation performance, the test threshold is appropriately lowered to avoid omissions. For features that do not have a direct physical causal relationship, the test threshold is increased to eliminate spurious features.

[0020] To address the strong temporal dynamics of AGC frequency regulation, a multi-scale adaptive extraction system for time-series features is constructed. Based on different time sliding windows, the window size is adaptively matched with the AGC command change cycle to extract time-series features such as amplitude, slope, rate of change, duration, and reverse adjustment frequency of AGC commands. The system also extracts the temporal evolution features of the unit's frequency regulation response, including lag time, inertia time, and regulation deviation. Specifically, the adaptive matching rule between the time sliding window size and the AGC command change cycle involves first performing feature recognition on the AGC command sequence to determine the change of a single AGC command. The window size is set to 1.2 times the AGC instruction change cycle for the period and instruction duration. For complex AGC instruction sequences such as continuous instruction changes and reverse adjustment, the window size is dynamically adjusted to 1.5-2 times the instruction change cycle. For a single AGC instruction sequence in the steady-state frequency modulation stage, the window size is adjusted to 1 times the instruction change cycle. At the same time, the window step size is adaptively set according to the data sampling frequency. The window step size is 1s for second-level sampled data and 50ms for millisecond-level sampled data to ensure the completeness and timeliness of time series feature extraction. Simultaneously, by combining the disturbance timing characteristics of grid frequency deviation with the energy storage-assisted frequency regulation response characteristics, a full-dimensional feature library is finally constructed, including static features, slow time-varying features, dynamic timing features, and energy storage synergy features, to comprehensively characterize the full-dimensional dynamic characteristics of AGC frequency regulation.

[0021] Static characteristics are inherent unit characteristics that do not change during frequency regulation, including unit capacity, boiler rated evaporation capacity, turbine regulation stages, and rated power of energy storage devices. Slow time-varying characteristics are characteristics that change slowly over time, including equipment aging factors, heating surface coking factors, actuator aging factors, and long-term average coal calorific value. Dynamic time-series characteristics are time-series characteristics that change in real time during frequency regulation, including AGC command time-series characteristics, unit frequency regulation response time-series characteristics, and grid frequency deviation disturbance time-series characteristics. Energy storage coordination characteristics are real-time operating characteristics of energy storage devices participating in frequency regulation, including energy storage charging and discharging power, state of charge, charging and discharging response time, and the proportion of energy storage-assisted frequency regulation. These four types of characteristics complement each other, achieving a full-dimensional and multi-dimensional characterization of the AGC frequency regulation process.

[0022] In the model construction phase, based on the constructed full-dimensional feature library, an adaptive partitioning method for AGC frequency regulation conditions based on density peak clustering is adopted. The distance calculation method of density peak clustering is optimized. The five core clustering features are unit load rate, coal calorific value, AGC command characteristics, ambient temperature, and slow time-varying factor. The cluster centers are corrected in combination with physical constraints of unit operation. Unsupervised adaptive clustering is performed on historical frequency regulation condition samples to automatically divide typical condition clusters such as deep peak regulation low load condition, rated load condition, coal quality large fluctuation condition, frequent command change condition, and extreme environment condition, so as to realize the division of the unit's full operating conditions. The distance calculation method for density peak clustering is optimized to a weighted Euclidean distance calculation method. Different weight coefficients are assigned according to the influence of five core clustering features on AGC frequency regulation conditions. The weight coefficients for unit load rate and AGC command characteristics are both set to 0.25, the weight coefficient for coal calorific value is set to 0.2, and the weight coefficients for ambient temperature and slow time-varying factor are both set to 0.15. The weight coefficients can be fine-tuned according to the thermal system characteristics of different units and actual frequency regulation conditions. At the same time, the clustering distance is corrected in combination with the physical constraints of unit operation. When the feature data exceeds the physical boundary, its clustering distance is appropriately increased to avoid the interference of abnormal feature data on the cluster center.

[0023] For each typical operating condition cluster, a basic model for frequency regulation performance prediction based on the TFT model and with built-in operating condition adaptive attention mechanism is constructed. It can dynamically adjust the feature attention weight according to the current operating condition type, focus on the features and time segments that have the greatest impact on frequency regulation performance under the current operating condition, and embed a simplified unit mechanism model as a constraint layer to avoid the prediction results from exceeding the physical operating limits of the unit and capture the nonlinear dynamic characteristics of the AGC frequency regulation process. The operating condition adaptive attention mechanism adopts a soft attention weight adjustment algorithm based on operating condition features. First, it identifies the core features of the current operating condition cluster, determines the ranking of the impact of each feature on frequency regulation performance, and then assigns initial attention weights to each feature according to the impact. During model training and inference, the predictive contribution of each feature is calculated in real time. The attention weight of features with high predictive contribution is dynamically increased, while the attention weight of features with low predictive contribution is gradually decreased. For key time segments (such as AGC command mutations and large fluctuations in grid frequency), attention weight enhancement in the time dimension is used to focus on the feature data within this time period, thereby improving the model's ability to capture features during key frequency regulation stages.

[0024] The simplified mechanistic model of the unit is constructed based on the core physical laws of the thermal system and frequency regulation of thermal power units. It eliminates the complex coupling relationships of secondary auxiliary systems, retains the core mechanistic relationships of boiler combustion, turbine regulation, and generator power output, and uses the unit load rate, coal calorific value, and main steam parameters as inputs, and the unit frequency regulation rate and response time as outputs to construct a mechanistic constraint model that conforms to the physical operating boundary of the unit. The integration of this model with the TFT model is to use the output results of the simplified mechanistic model as the hard constraint conditions of the TFT model. When the prediction results of the TFT model exceed the physical boundary of the simplified mechanistic model, the prediction results are automatically corrected to the range of the physical boundary. When the prediction results of the TFT model are within the range of the physical boundary, the prediction values ​​of the TFT model are slightly corrected in combination with the output results of the mechanistic model, thereby realizing the integration of data-driven and mechanistic modeling.

[0025] A cross-unit transfer learning mechanism is adopted to solve the prediction problem of small sample operating conditions. For the benchmark operating condition cluster with sufficient sample size, such as the rated load operating condition, the basic model is pre-trained. For the small sample operating condition cluster, such as deep peak shaving and extreme coal quality, the feature extraction layer parameters of the benchmark operating condition pre-trained model are transferred. At the same time, historical operating condition data of the same type of unit are introduced as auxiliary transfer samples. The output layer of the model is fine-tuned and adapted only based on the small sample operating condition data, which improves the model prediction accuracy and generalization ability under small sample operating conditions and achieves accurate prediction of the entire operating condition range of the unit.

[0026] The selection criteria for historical operating condition data of similar units are as follows: units with the same installed capacity, thermal system structure, and frequency regulation range as the target unit are selected as similar units. Their historical operating condition data must meet the following requirements: the load rate and coal calorific value deviations from the target unit's small sample operating conditions are within ±10%, and there are no abnormal data such as equipment failures or sudden changes in operating conditions. The fusion method for assisted migration is to perform feature alignment and standardization processing on the selected historical operating condition data of similar units and the target unit's small sample operating condition data, and merge them into fine-tuning training samples at a ratio of 7:3. During the fine-tuning process of the model output layer, 70% of the training weight is allocated to the target unit's small sample data, and 30% of the training weight is allocated to the historical data of similar units to ensure the model's adaptability to the target unit's operating conditions.

[0027] The dynamic correction stage constructs a quantitative index system and a real-time identification module for the slow time-varying characteristics of thermal power units. The quantitative index system includes equipment aging factors based on the equivalent operating time and overhaul cycle of the unit, coal quality fluctuation and coking factors based on the deviation between the actual efficiency and design efficiency of the boiler, and actuator aging factors based on the deviation between the actual flow characteristics and design characteristics of the regulating valve. The real-time identification module, combined with digital twin technology, constructs a virtual simulation model of the unit to achieve real-time quantitative characterization and dynamic updating of the slow time-varying characteristics. The quantification method of equipment aging factor is as follows: taking the unit after major overhaul as the initial state (factor value is 0), it is linearly calculated based on the ratio of the unit's equivalent operating time to the designed overhaul cycle. When the equivalent operating time reaches the designed overhaul cycle, the factor value is 1. After exceeding the overhaul cycle, it increases slowly in a non-linear manner. The coal quality fluctuation and heating surface coking factors are calculated based on the deviation value between the actual efficiency and the design efficiency of the boiler. When the deviation value is 0, the factor value is 0. For every 1% increase in the deviation value, the factor value increases by 0.15. The actuator aging factor is calculated based on the deviation rate between the actual flow characteristics and the design characteristics of the regulating valve. When the deviation rate is 0, the factor value is 0. When the deviation rate reaches 20%, the factor value is 1. All factor values ​​are normalized to the 0-1 range to achieve a quantitative characterization of slow time-varying characteristics.

[0028] A two-level adaptive dynamic correction module is designed. The first level is a feedforward correction based on a slow time-varying factor. The quantized slow time-varying factor is input into the frequency regulation performance prediction basic model. Combined with the prediction results of the unit virtual simulation model, the prediction output of the basic model is fed forward to compensate for the impact of slow time-varying characteristics on frequency regulation performance. The second level is a sliding window online feedback correction. A sliding time window is set on a daily basis. Based on the latest measured frequency regulation performance data and the source tracing results of abnormal data within the window, the parameters of the basic model are finely adjusted using an incremental learning method. At the same time, the quantized value of the slow time-varying factor and the operating condition cluster center are updated. The incremental learning method employs a mini-batch gradient descent approach to update model parameters. Real-world data within a daily sliding window is input into the model in batches. The learning rate for incremental learning is set to 1 / 10 of the base model's training learning rate. Only the parameters of the fully connected layers and output layer are updated, while the parameters of the feature extraction layer and attention mechanism layer remain fixed. After each parameter update, accuracy is verified using a validation set. If the updated model's prediction accuracy improves, the updated parameters are retained; otherwise, the update is abandoned, and the learning rate is appropriately reduced for the next update. Simultaneously, the gradient information corresponding to outlier data is removed based on the outlier data's source tracing results to prevent interference from outlier data in parameter updates.

[0029] A model accuracy warning and triggering mechanism is set up so that when the prediction error of the prediction model exceeds a preset threshold, the full incremental training of the base model is automatically triggered, and the working condition cluster center, model core parameters and feature selection threshold are automatically updated to avoid the continuous decay of model prediction accuracy over time. A model failure fallback switching logic is set up so that when the correction module fails, it automatically switches to a simplified mechanism model for auxiliary prediction to ensure the stability and accuracy of the model in long-term operation.

[0030] The preset threshold for prediction model accuracy warning is determined based on the allowable deviation range of the power grid AGC frequency regulation assessment indicators. For core assessment indicators such as regulation rate and regulation accuracy, the preset relative error threshold is 5%. For comprehensive indicators such as frequency regulation mileage and comprehensive frequency regulation performance K value, the preset relative error threshold is 8%. The criterion for continuous prediction error exceeding the threshold is that if the prediction error exceeds the corresponding preset threshold in 8 or more out of 10 consecutive prediction results, it is judged as continuous threshold exceeding, and the full incremental training of the basic model is automatically triggered.

[0031] The criteria for determining the failure of the correction module are as follows: both feedforward correction and online feedback correction fail simultaneously. Feedforward correction failure is determined by the deviation of the real-time identification result of the slow time-varying factor continuously exceeding ±0.2. Online feedback correction failure is determined by the model prediction accuracy decreasing instead of increasing after the incremental learning parameter update and failing to improve after three consecutive updates, or by hardware or software failures such as communication interruption or program errors in the correction module. The specific execution process of model switching is as follows: when the correction module failure is determined, a switching command is immediately triggered, the prediction output of the TFT model is paused, and the simplified mechanism model is started at the same time. The real-time collected unit operation data is input into the simplified mechanism model, and the auxiliary prediction result is quickly output. At the same time, the correction module failure alarm information is sent to the unit monitoring system. After the correction module recovers, it automatically switches back to the TFT model and compensates and corrects the model by combining the measured data from the fallback prediction stage.

[0032] In the timing optimization stage, an AGC command timing embedding module is constructed based on the dynamically corrected prediction model. The AGC command sequence and the power grid frequency deviation sequence issued by the power grid dispatching agency are embedded into the encoder of the prediction model. The attention mechanism of the AGC command timing embedding module highlights key command segments, such as large load change commands and continuous reverse adjustment commands, so as to achieve accurate perception of the timing characteristics of AGC commands. The quantitative criteria for determining a large load change command are: the load adjustment amplitude of a single AGC command exceeds 5% of the unit's rated load, or the load adjustment rate exceeds 80% of the unit's maximum designed adjustment rate; the quantitative criteria for determining a continuous reverse adjustment command are: three or more AGC adjustment commands in opposite directions occur within 5 minutes, and the load adjustment interval between two adjacent reverse commands does not exceed 30 seconds; in addition to the above two types of commands, when an AGC command shows five or more consecutive small adjustments and the total adjustment amplitude exceeds 3% of the unit's rated load, it is also determined as a critical command segment, and the feature capture weight of this type of command is increased through an attention mechanism.

[0033] The module for real-time identification and dynamic compensation of the frequency regulation response lag characteristics of the generator unit is designed. Based on real-time operating data, the module uses the recursive least squares method combined with the forgetting factor to identify dynamic characteristic parameters of the generator unit such as pure lag time, inertial time constant, and rise rate online. The compensation coefficient is dynamically adjusted according to the identification results to adaptively compensate the prediction output of the prediction model and offset the prediction deviation caused by the large lag characteristics of the generator unit. In the recursive least squares method, the basic value range of the forgetting factor is 0.95-0.99. During the steady-state stage of unit frequency regulation, the forgetting factor is set to 0.99 to enhance the identification weight of historical data and improve the stability of parameter identification. During the dynamic stage of sudden changes in unit AGC commands and large fluctuations in grid frequency, the forgetting factor is dynamically adjusted to 0.95-0.97 to reduce the identification weight of historical data, increase the proportion of real-time data, and ensure the real-time performance and accuracy of dynamic characteristic parameter identification. The adjustment of the forgetting factor is triggered by the grid frequency deviation and the AGC command change rate. When the frequency deviation exceeds ±0.05Hz or the command change rate exceeds 2% / s, it automatically switches to dynamic value.

[0034] The prediction model can simultaneously output AGC frequency modulation performance prediction results with different time steps such as 10s, 30s, and 60s, realizing performance prediction of the entire response process of a single AGC command, and displaying the deviation trend between the predicted value and the measured value in real time through the prediction result visualization module.

[0035] The selection of multiple time steps of 10s, 30s, and 60s is based on the assessment time scale of power grid AGC frequency regulation and the frequency regulation response characteristics of generating units. 10s is the fast response stage of generating unit frequency regulation, which is the key time point for rapid adjustment of power grid frequency. 30s is the intermediate response stage of generating unit frequency regulation, and 60s is the steady-state response stage of generating unit frequency regulation, covering the entire response cycle of a single AGC command. The rolling prediction window update mechanism is to slide the prediction window in seconds. Every 1 second of real-time operating data is acquired, the prediction window is slid forward by 1 second, and the prediction results of each time step are updated. At the same time, historical data that exceeds the prediction cycle is removed to ensure the real-time performance and effectiveness of rolling prediction.

[0036] In the energy consumption prediction stage, the AGC frequency regulation performance prediction results after time-series optimization are combined to construct a collaborative prediction framework for frequency regulation performance, energy consumption and revenue. Under the collaborative prediction framework, the core performance indicators of the power grid assessment are predicted. At the same time, a sub-model for predicting the incremental frequency regulation energy consumption during the AGC frequency regulation process is built. Based on the power supply coal consumption characteristic curve of the unit under different operating conditions, combined with real-time coal quality fluctuation data and energy storage-assisted frequency regulation energy consumption data, the incremental power supply coal consumption of the unit under different AGC frequency regulation commands and different regulation rates is quantified, thereby realizing the three-dimensional collaborative prediction of frequency regulation response performance, frequency regulation energy consumption cost and frequency regulation revenue. The specific quantitative integration method for the increase in coal consumption for power supply is as follows: First, based on the coal consumption characteristic curve of the unit operating under varying conditions, the benchmark value of coal consumption for power supply under the basic AGC frequency regulation command and regulation rate is determined; then, based on real-time data of coal quality fluctuations, the coal consumption correction coefficient corresponding to the deviation of coal calorific value is calculated. For every 100 kcal / kg decrease in coal calorific value, the coal consumption correction coefficient is increased by 1.02 times; at the same time, combined with the power and operating time of energy storage-assisted frequency regulation, the reduction in unit coal consumption corresponding to energy storage-assisted frequency regulation is calculated. The final increase in coal consumption for power supply is the benchmark value of coal consumption multiplied by the coal consumption correction coefficient, and then the reduction in coal consumption due to energy storage-assisted frequency regulation is subtracted, thus achieving accurate quantification of the increase in coal consumption under different frequency regulation conditions.

[0037] The incremental energy consumption cost and revenue under different frequency regulation scenarios are quantified. Combined with the predicted frequency regulation performance indicators, a multi-objective optimization function is constructed with the goal of minimizing energy consumption cost, maximizing frequency regulation revenue, and meeting the power grid performance assessment requirements. This provides a quantitative basis for optimizing frequency regulation control parameters and balancing frequency regulation performance and energy consumption cost. The objective weights of the multi-objective optimization function are dynamically set according to the grid frequency regulation assessment requirements and unit operation needs. During the strict grid frequency regulation assessment phase, the weight for meeting the grid performance assessment requirements is set to 0.4, and the weights for minimizing energy consumption costs and maximizing frequency regulation benefits are both set to 0.3. During the phase with higher unit energy conservation and consumption reduction requirements, the weight for minimizing energy consumption costs is set to 0.4, and the weights for meeting the grid performance assessment requirements and maximizing frequency regulation benefits are both set to 0.3. The quantification rules for the constraints are as follows: the grid performance assessment indicators must meet the following requirements: regulation rate ≥ grid requirement value, regulation accuracy ≤ grid allowable deviation, and comprehensive frequency regulation performance K value ≥ grid benchmark value. The constraints on energy consumption costs and frequency regulation benefits are within a reasonable range based on the historical operating data of the units, ensuring the actual executability of the optimization results.

[0038] A multi-scenario frequency regulation performance simulation and prediction module was built, which can simulate frequency regulation performance and energy consumption changes under different AGC command sequences, different load conditions, different coal quality conditions, and different energy storage coordination strategies.

[0039] The specific implementation of multi-scenario frequency regulation performance simulation and prediction is as follows: a scenario parameter configuration module is built based on a full-dimensional feature library. This module supports custom settings or selection of typical scenario parameters for parameters such as the amplitude, rate, and frequency of AGC command sequences; the load rate range and trend of load conditions; the calorific value and volatile matter content of coal; and the charging and discharging power, intervention timing, and auxiliary frequency regulation ratio of energy storage synergy strategies. During simulation, the configured scenario parameters are input into the time-series optimized prediction model, and multiple rounds of simulation are performed in combination with the simplified mechanism model of the unit. At the same time, the frequency regulation performance indicators and energy consumption data of each round of simulation are recorded. The core output of the simulation results includes frequency regulation performance indicator curves, energy consumption increment change curves, and performance-energy consumption correlation analysis reports for each scenario. It also outputs optimal frequency regulation control parameter suggestions for each scenario, providing quantitative references for on-site operation and control.

[0040] The deployment and assurance phase constructs an online deployment architecture that coordinates the edge, cloud, and digital twin terminals. The edge side is deployed on the plant-level monitoring information system (SIS) / distributed control system (DCS) side of the unit to realize real-time data collection, preprocessing, second-level real-time performance prediction and local output, so as to meet the real-time requirements of on-site online applications. The latency control standard for edge-side second-level real-time performance prediction is that the maximum end-to-end allowable latency from the completion of real-time data acquisition to the output of performance prediction results is 500ms, of which the latency of data preprocessing does not exceed 200ms, the latency of model inference prediction does not exceed 250ms, and the latency of local output of results does not exceed 50ms. By using edge-side local data caching and lightweight model inference engine deployment, the latency of each link is strictly controlled to meet the requirements of real-time frequency regulation control of thermal power units' AGC.

[0041] A cloud-based model training and optimization platform enables the aggregation and storage of multi-source data, iterative updates of the operating condition database, pre-training and incremental training of models, and knowledge transfer and sharing among multiple units. A digital twin terminal deploys a virtual simulation model of the unit to realize the simulation verification of prediction results and operating condition simulation. The three work together to improve the reliability and practicality of the deployment. A model parameter iterative optimization mechanism is constructed, which feeds back the measured frequency regulation performance data, actual energy consumption data, and abnormal data tracing results after the unit actually executes AGC commands to the cloud model optimization platform in real time. The data results are used to optimize feature screening results, operating condition clustering centers, model parameters and correction modules, so as to realize the continuous self-optimization and self-adaptation of the model. The iterative optimization mechanism for model parameters is divided into daily micro-iterations and monthly full iterations. Daily micro-iterations are based on daily measured data and feedback results to fine-tune the parameters of the model's correction module, with an iteration cycle of 24 hours. Monthly full iterations are based on monthly measured data statistics to comprehensively optimize feature selection results, operating condition cluster centers, and core model parameters, with an iteration cycle of 30 days. The effective sample selection rules are as follows: samples corresponding to abnormal data such as equipment failure, measurement error, and sudden changes in operating conditions are removed, and effective samples are selected from the steady-state stage of frequency modulation and the normal adjustment stage of commands. Furthermore, the prediction error of the effective samples must be within a reasonable range to ensure the sample quality of iterative optimization.

[0042] A model reliability fallback module is set up to verify the physical boundary rationality of the prediction model output results in real time. When the prediction results exceed the physical limits of unit operation, or the prediction error of the prediction model continuously exceeds the threshold, an early warning is automatically triggered, and the system switches to a fallback prediction mode based on a simplified mechanism model of the unit. A data security encryption module is set up to ensure the security of multi-source data transmission and storage, and to avoid risks to unit operation and grid dispatch caused by model anomalies and data leakage.

[0043] The specific indicators for verifying the physical boundary rationality of the prediction results are the core physical operation indicators of the unit's frequency regulation, including the unit's load regulation range, main steam pressure, main steam temperature, and turbine regulating valve opening. The verification threshold for each indicator is the physical operating limit range designed for the unit, namely, the unit's load regulation range does not exceed the range from the minimum stable combustion load to the rated load, the main steam pressure deviation does not exceed ±0.5MPa of the design value, the main steam temperature deviation does not exceed ±10℃ of the design value, and the turbine regulating valve opening does not exceed the 0-100% range. When any indicator in the prediction results exceeds the above physical boundary threshold, a physical boundary rationality warning is immediately triggered.

[0044] The data security encryption module covers the entire lifecycle of multi-source data, including data transmitted between the edge, cloud, and digital twin terminals, as well as locally stored data in the cloud and edge. The core encryption objects are sensitive data such as real-time core operating parameters of the unit, AGC scheduling instruction data, frequency regulation assessment and revenue data, and core model parameters. The core encryption method uses an industrial-grade symmetric encryption algorithm to encrypt the data transmission process. At the same time, sensitive data stored locally is encrypted in blocks and access permissions are controlled. Multi-level operation permissions are set, and only authorized personnel can access encrypted data. Communication between the edge, cloud, and digital twin terminals uses a dedicated encrypted communication protocol to prevent interception and tampering during data transmission. Data is also backed up and encrypted regularly to ensure data integrity and security.

[0045] 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.

[0046] 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. A performance prediction method for AGC frequency regulation of thermal power units, characterized in that, include: Data preprocessing stage: Construct a multi-source heterogeneous data acquisition system for AGC frequency modulation, covering all influencing factors in the frequency modulation chain, and achieve multi-scale data synchronization by aligning the timestamps triggered by AGC frequency modulation events. Physical constraint preprocessing is used to remove abnormal data, fill missing values, and implement standardization by working condition. Feature construction stage: Based on the preprocessed multi-source heterogeneous data, the core prediction targets related to AGC frequency regulation performance, energy consumption and revenue are identified. The CCM-PCMCI causal screening algorithm with physical constraints of the unit's thermal system is used to screen core causal features. A multi-scale time series feature adaptive extraction system is constructed to form a full-dimensional feature library to characterize the frequency regulation dynamic characteristics. Model building phase: Based on the full-dimensional feature library, the optimized density peak clustering method is used to divide the typical operating conditions of the unit's AGC frequency regulation into clusters. For each operating condition cluster, a TFT basic model with an operating condition adaptive attention mechanism and a mechanism constraint layer is built. Dynamic correction phase: Construct a quantitative index system and real-time identification module for the slow time-varying characteristics of thermal power units, combine digital twin technology to realize dynamic updates of slow time-varying characteristics, design a two-level adaptive dynamic correction module to perform feedforward compensation and online feedback fine-tuning, and set up accuracy warning, full incremental training and fault fallback mechanism. Timing optimization stage: Based on the dynamically corrected prediction model, an AGC command timing embedding module is constructed to identify the frequency regulation lag characteristics of the unit and perform dynamic compensation, so as to realize multi-time step rolling prediction and visualization of prediction results; Energy consumption prediction stage: Combining the AGC frequency regulation performance prediction results after time series optimization, a collaborative prediction framework for frequency regulation performance, energy consumption and revenue is constructed. The incremental energy consumption and revenue under different scenarios are quantified and a multi-objective optimization function is constructed to provide multi-scenario simulation prediction function. Deployment and assurance phase: Construct an online deployment architecture that integrates edge computing, cloud computing, and digital twins, establish a model parameter iterative optimization mechanism, and set up a reliability safety net module and a data security encryption module.

2. The performance prediction method for AGC frequency regulation of thermal power units according to claim 1, characterized in that, The multi-source heterogeneous data acquisition system in the data preprocessing stage covers all dimensions of frequency regulation impact. The data sources include second-level real-time unit operation data, millisecond-level AGC scheduling instructions and time-series grid frequency deviation data, minute-level online coal quality monitoring data, daily-level unit equipment status and slow time-varying data, as well as historical frequency regulation assessment data and environmental parameter data. Multi-scale data time synchronization is achieved based on AGC instruction change events, and the synchronization accuracy is corrected by combining the characteristics of the unit's thermal system. Physical constraint preprocessing removes abnormal data and fills in missing values ​​based on the physical boundaries of unit operation, and finally implements sub-condition standardization according to different load segments and coal quality condition clusters.

3. The performance prediction method for AGC frequency regulation of thermal power units according to claim 2, characterized in that, The core prediction target set in the feature construction stage includes core performance indicators of power grid assessment, incremental frequency regulation energy consumption indicators, and frequency regulation revenue-related indicators. The CCM-PCMCI causal screening algorithm uses the physical constraints of the unit's thermal system as a priori conditions, optimizes the test threshold to distinguish between causal relationships and pseudo-correlation relationships, and eliminates redundant features. The multi-scale time-series feature adaptive extraction system extracts time-series features related to AGC commands, unit frequency regulation responses, power grid frequency deviations, and energy storage-assisted frequency regulation based on different time sliding windows, ultimately forming a full-dimensional feature library containing static features, slow time-varying features, dynamic time-series features, and energy storage synergy features.

4. The performance prediction method for AGC frequency regulation of thermal power units according to claim 3, characterized in that, The density peak clustering in the model construction phase uses unit load rate, coal calorific value, AGC command characteristics, ambient temperature, and slow time-varying factor as core clustering features. Combined with physical constraints of unit operation, the cluster centers are corrected to adaptively divide typical operating condition clusters, including deep peak shaving and low load, rated load, large fluctuations in coal quality, frequent command changes, and extreme environment. The TFT basic model corresponding to each operating condition cluster dynamically adjusts the feature weights through an operating condition adaptive attention mechanism and embeds a simplified unit mechanism model as a constraint layer. A cross-unit transfer learning mechanism is adopted, which utilizes the pre-trained model parameters of the baseline operating condition cluster to fine-tune only the output layer of the small sample operating condition cluster model to adapt to the operating condition cluster.

5. The performance prediction method for AGC frequency regulation of thermal power units according to claim 4, characterized in that, The quantitative index system for slow time-varying characteristics in the dynamic correction stage includes equipment aging factor, coal quality fluctuation and coking factor of heating surface, and actuator aging factor. The real-time identification module, combined with the digital twin virtual simulation model, realizes the real-time quantification and dynamic updating of slow time-varying characteristics. In the dual-level adaptive dynamic correction module, the feedforward correction is based on the slow time-varying factor and the results of the virtual simulation model for compensation, while the online feedback correction uses a daily sliding window combined with incremental learning to fine-tune the model parameters. The accuracy warning mechanism triggers full incremental training when the prediction error continuously exceeds the preset threshold, and the fault fallback mechanism switches to a simplified mechanism model to assist prediction when the correction module fails.

6. The performance prediction method for AGC frequency regulation of thermal power units according to claim 5, characterized in that, The timing optimization phase of the AGC command timing embedding module embeds the AGC command sequence and the grid frequency deviation sequence into the model encoder, and highlights key command segments through an attention mechanism; the hysteresis characteristic identification adopts the recursive least squares method combined with the forgetting factor to identify the core dynamic characteristic parameters of the unit online and dynamically adjust the compensation coefficient; the prediction model can output rolling prediction results of AGC frequency regulation performance at multiple time steps, realize the performance prediction of the entire response process of a single AGC command, and display the deviation trend between the predicted value and the measured value through the visualization module.

7. The performance prediction method for AGC frequency regulation of thermal power units according to claim 6, characterized in that, The collaborative prediction framework in the energy consumption prediction stage synchronously predicts frequency regulation performance, frequency regulation energy consumption increment, and revenue. The frequency regulation energy consumption increment prediction is based on the power supply coal consumption characteristic curve of the unit operating under different conditions, and combines coal quality fluctuations and energy storage-assisted frequency regulation energy consumption data to quantify the power supply coal consumption increment. The multi-objective optimization function aims to minimize energy consumption cost, maximize frequency regulation revenue, and meet grid assessment requirements, providing a basis for optimizing frequency regulation control parameters. The multi-scenario simulation module can simulate performance and energy consumption changes under different AGC commands, load conditions, coal quality conditions, and energy storage collaborative strategies.

8. The performance prediction method for AGC frequency regulation of thermal power units according to claim 7, characterized in that, In the collaborative deployment architecture of the deployment and protection phase, the edge side realizes real-time data processing and second-level prediction, the cloud side realizes data aggregation, model training and knowledge sharing, and the digital twin end realizes simulation verification and working condition simulation. The model parameter iterative optimization mechanism feeds back the measured data and anomaly tracing results to the cloud for optimizing feature selection, working condition clustering and model parameters. The reliability fallback module verifies the physical rationality of the prediction results, triggers an early warning and switches to the fallback prediction mode when the threshold is exceeded, and the data security encryption module ensures the security of data transmission and storage.

Images