A power plant AGC load response optimization system based on deep learning

Abstract

The application belongs to the technical field of wind power generation, and particularly relates to a power plant AGC load response optimization system based on deep learning, which comprises a multi-source data acquisition module, a data preprocessing and feature engineering module, a multi-scale deep learning prediction module, a deep reinforcement learning decision module, a dynamic execution and feedback correction module, and a visual monitoring and evaluation module; the multi-scale deep learning prediction module and the deep reinforcement learning decision module are designed in a fusion mode, so that the dependence on an accurate model is eliminated; the prediction module adopts an LSTM-Transformer hybrid architecture to accurately predict wind power output and AGC instruction changes; the decision module adopts an improved DDPG algorithm to dynamically generate a control strategy, effectively solves the problems of response lag and overshoot, and improves tracking accuracy.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation technology, specifically a power plant AGC load response optimization system based on deep learning. Background Technology

[0002] As the proportion of wind power generation in the power system continues to increase, the volatility and uncertainty of wind power pose significant challenges to grid frequency stability and wind power AGC control. Traditional wind farm AGC control methods mainly employ technologies such as PID control and model predictive control (MPC), which achieve tracking and response to wind power AGC commands by establishing a fixed control model.

[0003] In large-scale onshore and offshore wind farms, it is particularly suitable for the grid AGC regulation needs in scenarios with a high proportion of wind power connected to the grid, specifically covering core application scenarios such as wind farm AGC command tracking, wind power output fluctuation prediction and smoothing, wind turbine group differentiated control, and grid frequency coordinated stability control.

[0004] Existing technologies still face several challenges in optimizing wind power AGC load response: First, traditional AGC control methods rely heavily on precise mathematical models, making it difficult to adapt to the strong nonlinearity and rapid fluctuations in wind power output caused by meteorological factors such as wind speed and direction. In scenarios with a high proportion of wind power connected to the grid, problems such as response lag and regulation overshoot can easily occur, failing to meet the rapid tracking requirements of AGC commands. Second, the application of existing deep learning technologies in the field of wind power AGC is mostly focused on single-module optimization, lacking a complete integrated design of "prediction-decision-execution". Prediction models do not fully integrate multi-scale spatiotemporal features and meteorological forecast data, and decision models do not take into account wind turbine operating constraints and regulation economy, making it difficult to achieve a balanced optimization of AGC response performance and wind turbine losses.

[0005] Therefore, this invention provides a power plant AGC load response optimization system based on deep learning. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by the present invention to solve its technical problem is: a power plant AGC load response optimization system based on deep learning, comprising a multi-source data acquisition module, a data preprocessing and feature engineering module, a multi-scale deep learning prediction module, a deep reinforcement learning decision module, a dynamic execution and feedback correction module, and a visualization monitoring and evaluation module. The multi-source data acquisition module is used to acquire multi-dimensional operational data of wind farms and power grids in real time, including but not limited to AGC commands, wind speed, wind direction, wind turbine output, power grid frequency, tie line power, ACE value, and wind turbine operating status parameters. The data preprocessing and feature engineering module is used to clean, normalize, fill missing values, and handle outliers of the collected multi-source data, and extract spatiotemporal features, trend features, and correlation features. The multi-scale deep learning prediction module is used to predict the AGC command change trend, wind power output fluctuation range and grid frequency response characteristics within a preset time period based on preprocessed feature data. The deep reinforcement learning decision module is used to dynamically generate the optimal AGC load response strategy, including the total regulating power allocation scheme and the control commands of each wind turbine, by taking the prediction results as input and combining them with the wind power AGC control constraints. The dynamic execution and feedback correction module is used to send the optimal control command to each wind turbine controller and monitor the execution effect in real time, and adjust the control strategy online through a closed-loop feedback mechanism. The visualization monitoring and evaluation module is used to display the system's operating status, control effect, and performance indicators in real time, and to provide historical data analysis and evaluation reports.

[0008] Preferably, the multi-scale deep learning prediction module adopts a hybrid neural network architecture that integrates LSTM and Transformer, wherein LSTM is used to capture the long-term and short-term dependencies of time series, and Transformer is used to extract global correlation features between multiple variables and output prediction results at multiple time scales.

[0009] Preferably, the deep reinforcement learning decision module adopts an improved DDPG algorithm to construct a Markov decision process model that includes a state space, an action space, and a reward function. The state space includes wind power output prediction, AGC commands, grid frequency deviation, tie-line power deviation, and wind turbine operating constraints. The action space includes power adjustment commands for each wind turbine. The reward function comprehensively considers power tracking accuracy, adjustment rate, frequency stability, and wind turbine loss cost.

[0010] Preferably, the data preprocessing and feature engineering module further includes a feature selection unit, which uses the SHAP value analysis method to quantify the influence weight of each input feature on the AGC response performance, and selects key features to reduce model complexity and improve generalization ability.

[0011] Preferably, the multi-scale deep learning prediction module also integrates meteorological forecast data and dynamically adjusts the weight coefficients of meteorological data and historical operational data through an attention mechanism to improve prediction accuracy.

[0012] Preferably, the deep reinforcement learning decision module further includes a constraint processing unit, which is used to verify in real time whether the control strategy meets physical and operational constraints such as the maximum adjustment rate of the wind turbine, the minimum stable output, and the pitch angle limit, and to automatically correct strategies that do not meet the constraints.

[0013] Preferably, the dynamic execution and feedback correction module adopts a hierarchical control architecture, including a field-level AGC controller and a wind turbine-level controller. The field-level controller is responsible for the overall regulation power allocation, while the wind turbine-level controller is responsible for quickly tracking and executing control commands and feeding back real-time output data.

[0014] Preferably, the system further includes a transfer learning module, which is used to transfer the trained model parameters to a new wind farm scenario. Model adaptation can be achieved by fine-tuning with a small amount of field data, reducing training costs and shortening the deployment cycle.

[0015] Preferably, the visualization monitoring and evaluation module provides multi-dimensional performance indicators, including AGC adjustment accuracy, response time, frequency deviation pass rate, fan fatigue damage degree, and economic benefit evaluation index; The system supports seamless integration with wind farm energy management systems and power grid dispatch automation systems, enabling data sharing and command interaction.

[0016] The beneficial effects of this invention are as follows: 1. The power plant AGC load response optimization system based on deep learning described in this invention integrates a multi-scale deep learning prediction module and a deep reinforcement learning decision module to eliminate the dependence on accurate models. The prediction module adopts an LSTM-Transformer hybrid architecture to accurately predict wind power output and AGC command changes. The decision module uses an improved DDPG algorithm to dynamically generate control strategies, effectively solving the problems of response lag and overshoot, and improving tracking accuracy.

[0017] 2. The power plant AGC load response optimization system based on deep learning described in this invention achieves collaboration among modules by constructing a closed-loop optimization architecture; the prediction module integrates multi-source data and filters key features through SHAP value analysis; the decision module sets up a constraint processing unit and combines it with a comprehensive reward function to achieve multi-objective balanced optimization, thus solving the problem of poor module collaboration.

[0018] 3. The power plant AGC load response optimization system based on deep learning described in this invention achieves differentiated power allocation through a hierarchical control architecture of dynamic execution modules; the plant-level controller allocates power based on the wind turbine's response capability and fatigue damage level; the wind turbine-level controller adopts a PID + feedforward compensation algorithm to ensure efficient response and reduce wind turbine losses.

[0019] 4. The power plant AGC load response optimization system based on deep learning described in this invention achieves two major design breakthroughs: first, the dynamic execution module constructs a closed-loop feedback mechanism to update model parameters in real time and adapt to changes in operating conditions; second, the transfer learning module adopts a fine-tuning strategy to achieve rapid adaptation to new scenarios, reducing deployment costs and time. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is the overall system architecture diagram of this invention; Figure 2 This is a flowchart of the multi-scale deep learning prediction module in this invention; Figure 3 This is a flowchart of the multi-scale deep learning prediction module in this invention; Figure 4 This is a flowchart of the deep reinforcement learning decision module in this invention; Figure 5 This is a flowchart of the dynamic execution, feedback correction, and visual monitoring module in this invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0023] like Figures 1 to 5 As shown in the embodiment of the present invention, a power plant AGC load response optimization system based on deep learning includes a multi-source data acquisition module, a data preprocessing and feature engineering module, a multi-scale deep learning prediction module, a deep reinforcement learning decision module, a dynamic execution and feedback correction module, and a visualization monitoring and evaluation module. The aforementioned multi-source data acquisition module is used to acquire multi-dimensional operational data of wind farms and power grids in real time, including but not limited to AGC commands, wind speed, wind direction, wind turbine output, grid frequency, tie line power, ACE value, and wind turbine operating status parameters. The aforementioned data preprocessing and feature engineering module is used to clean, normalize, fill missing values, and handle outliers of the collected multi-source data, and to extract spatiotemporal features, trend features, and correlation features. The aforementioned multi-scale deep learning prediction module is used to predict the trend of AGC command changes, wind power output fluctuation range, and grid frequency response characteristics within a preset time period based on preprocessed feature data. The aforementioned deep reinforcement learning decision module is used to dynamically generate the optimal AGC load response strategy, including the total regulating power allocation scheme and the control commands of each wind turbine, by taking the prediction results as input and combining them with the wind power AGC control constraints. The aforementioned dynamic execution and feedback correction module is used to send the optimal control command to each wind turbine controller and monitor the execution effect in real time, and adjust the control strategy online through a closed-loop feedback mechanism; The aforementioned visualization monitoring and evaluation module is used to display the system's operating status, control effects, and performance indicators in real time, and to provide historical data analysis and evaluation reports.

[0024] like Figures 1 to 5 As shown, the multi-scale deep learning prediction module adopts a hybrid neural network architecture that integrates LSTM and Transformer. LSTM is used to capture the long-term and short-term dependencies of time series, while Transformer is used to extract global correlation features between multiple variables and output prediction results at multiple time scales.

[0025] The aforementioned hybrid neural network architecture includes a feature fusion layer, which is used to weight and fuse the time-series features output by the LSTM with the global correlation features output by the Transformer. The weighting coefficients are dynamically adjusted through an attention mechanism. The multiple time scales include short scale (1-5 minutes), medium scale (5-30 minutes), and long scale (30 minutes-2 hours), which correspond to the prediction requirements of different AGC adjustment scenarios. The output frequency of the prediction results is consistent with the highest acquisition frequency of the multi-source data acquisition module to ensure the real-time performance of the prediction. The core algorithm of the prediction model is expressed by the following formula: in, : The predicted output value at any given time is specifically the predicted value of AGC command, the predicted value of wind power output, or the predicted value of grid frequency, corresponding to the three core prediction targets of multi-scale prediction. : The input feature vector at any given time is composed of key features output by the data preprocessing and feature engineering modules, including historical operational data features, meteorological features, etc. : The network's output, i.e., the input feature vector. go through The time series long-term and short-term dependency features obtained after layer processing; : The network's output, i.e., the input feature vector. go through Multivariate global correlation features obtained after layer processing; : respectively Output features The weighting coefficients of the output features are dynamically adjusted by the attention mechanism, and their values ​​are all in the range [0,1], and satisfy the following conditions: ; Attention mechanism fusion function, used to fuse two types of weighted features to highlight features that contribute highly to the prediction results; : Model bias term, used to correct prediction bias, with a value range of [-0.1, 0.1], obtained through adaptive learning during model training.

[0026] like Figures 1 to 5 As shown, the aforementioned deep reinforcement learning decision module uses an improved DDPG algorithm to construct a Markov decision process model that includes a state space, an action space, and a reward function. The state space includes wind power output predictions, AGC commands, grid frequency deviations, tie-line power deviations, and wind turbine operating constraints. The action space includes power regulation commands for each wind turbine. The reward function comprehensively considers power tracking accuracy, regulation rate, frequency stability, and wind turbine loss costs.

[0027] The improvements to the DDPG algorithm mentioned above are as follows: a residual connection layer is added to the Actor network to reduce gradient vanishing; a dual-Critic network structure is introduced to the Critic network to reduce overestimation bias; the optimizer uses the AdamW optimizer with a learning rate set to 0.001-0.0001 and adaptively decaying according to the number of training iterations; the wind turbine operation constraints specifically include the maximum adjustment rate of the wind turbine, the minimum stable output, and the pitch angle adjustment range; the weight coefficients of each index in the reward function can be dynamically configured according to the grid dispatch requirements, with a configuration range of 0.1-0.4, to ensure the balance of multi-objective optimization; The core algorithm of the reward function is expressed by the following formula: in, : At any given time, the system is in a state. Execution of actions The total reward value obtained is the core indicator for evaluating the quality of the action, and its value ranges from [-10, 10]. : The system state vector at time t corresponds to the state space of the Markov decision process. ; : The system action vector at time t corresponds to the action space of the Markov decision process. (n is the number of wind turbines); The weight coefficients of each reward item are all within the range of [0.1, 0.4], and satisfy the following condition: 1. It can be dynamically configured according to the power grid dispatch priority; The power point tracking accuracy bonus is negatively correlated with the AGC command tracking error, and is calculated using the following formula: The value range is [0,1]. Frequency stability bonus, negatively correlated with the absolute value of grid frequency deviation, is calculated using the following formula: ( For adjustment coefficients, The value range is [0,1]. The efficiency reward for adjustment is positively correlated with the rate of adjustment, and its calculation formula is as follows: ( for Adjust the speed constantly. (where is the maximum adjustment rate), and its value ranges from [0,1]. The wind turbine loss penalty is positively correlated with the wind turbine fatigue damage degree, and the calculation formula is as follows: ( for Constant-time fan fatigue damage level, (where is the cumulative distribution function), and its value range is [0,1].

[0028] like Figures 1 to 5 As shown, the data preprocessing and feature engineering module also includes a feature selection unit, which uses the SHAP value analysis method to quantify the influence weight of each input feature on the AGC response performance, and selects key features to reduce model complexity and improve generalization ability.

[0029] The SHAP value analysis method described above uses the TreeExplainer interpreter to calculate the SHAP value of each feature for the preprocessed feature data. The selection criterion is to retain features with a mean absolute SHAP value greater than 0.05. The key features after selection need to undergo a correlation test to remove redundant features with an absolute Pearson correlation coefficient greater than 0.8 to ensure the independence and effectiveness of the features. At the same time, the key features retained will be synchronously fed back to the multi-scale deep learning prediction module as the core feature set for model input. The core algorithm for calculating SHAP values ​​is expressed by the following formula: in, : No. Each input feature The value is used to quantify the influence weight of this feature on the prediction result of AGC response performance. The value range is [-0.5, 0.5]. The larger the absolute value, the higher the influence weight of the feature. : except for the first All other features besides the one feature ( Take the expected value; Includes all input features At that time, the prediction function of AGC response performance (i.e., the output function of the multi-scale deep learning prediction module); Remove the first After one feature, only the remaining features are included. At that time, the prediction function for AGC response performance; The set of all input features. ( (Total number of input features). Remove the first Features Then, the set of remaining input features, .

[0030] like Figures 1 to 5 As shown, the multi-scale deep learning prediction module also integrates meteorological forecast data and dynamically adjusts the weight coefficients of meteorological data and historical operational data through an attention mechanism to improve prediction accuracy.

[0031] The aforementioned meteorological forecast data includes short-term wind speed forecast, wind direction forecast, temperature forecast, and air pressure forecast. The data comes from the local meteorological monitoring station of the wind farm and the regional meteorological forecast platform, and is updated once per minute. The attention mechanism adopts a multi-head attention structure with 4-8 heads. The weight coefficients are adjusted based on the predictive contribution of the features. The contribution is calculated through the mutual information value between the features and the prediction target. The weight coefficients of the meteorological data are adjusted in the range of 0.2-0.6 to ensure the reasonable integration of meteorological data with historical operational data.

[0032] like Figures 1 to 5 As shown, the aforementioned deep reinforcement learning decision module also includes a constraint processing unit, which is used to verify in real time whether the control strategy meets physical and operational constraints such as the maximum adjustment rate of the wind turbine, the minimum stable output, and the pitch angle limit, and to automatically correct strategies that do not meet the constraints.

[0033] The verification frequency of the above constraint processing unit is consistent with the decision frequency of the decision module, which is 100ms / time. The maximum adjustment rate of the wind turbine does not exceed 0.2 times the rated power / second, the minimum stable output is not less than 0.1 times the rated power, and the pitch angle is limited to 0°-90°. The correction logic adopts the "peak shaving and valley filling" strategy, which attenuates the adjustment command that exceeds the constraints, and at the same time redistributes the excess adjustment power to the wind turbines that meet the constraints to ensure that the total adjustment power is consistent with the AGC command. The corrected control strategy needs to be verified again until all constraints are met before it can be output to the dynamic execution and feedback correction module.

[0034] like Figures 1 to 5 As shown, the above-mentioned dynamic execution and feedback correction module adopts a hierarchical control architecture, including a field-level AGC controller and a wind turbine-level controller. The field-level controller is responsible for the overall regulation power allocation, while the wind turbine-level controller is responsible for quickly tracking and executing control commands and feeding back real-time output data.

[0035] The aforementioned field-level AGC controller and wind turbine-level controller communicate via industrial Ethernet, with a communication delay of no more than 50ms. The power allocation of the field-level controller is based on the real-time response capability and fatigue damage degree of each wind turbine. The response capability is determined by statistical analysis of the historical adjustment rate of the wind turbine, and the fatigue damage degree is calculated by the wind turbine running time and vibration parameters. The wind turbine-level controller adopts a PID + feedforward compensation control algorithm to ensure that the tracking error of the adjustment command does not exceed ±1%. The feedback frequency of real-time output data is 50ms / time. The feedback data includes core operating parameters such as actual wind turbine output, pitch angle, and speed, which are used for parameter adjustment of the feedback correction mechanism. The core formula of the PID control algorithm for wind turbine-level controllers is as follows: in, : The control commands output by the wind turbine stage controller at all times (i.e., wind turbine pitch angle adjustment commands or speed adjustment commands). The proportional gain of the PID controller, with a value range of [1.0, 5.0], is used to adjust the control response speed. The integral coefficient of the PID controller, with a value range of [0.01, 0.1], is used to eliminate steady-state error; The derivative coefficient of the PID controller, with a value range of [0.1, 1.0], is used to suppress overshoot and improve control stability. : Adjust the instruction tracking error continuously. ( for Adjust command power at all times. for (Actual output of the wind turbine at any time) From 0 to The integral term of the tracking error is used to accumulate the steady-state error and make corrections. : The differential term of the error is tracked at all times to reflect the rate of change of the error and to suppress overshoot in advance; : The feedforward compensation value is taken as the control command of the previous time step. It is used to improve the tracking speed of control commands.

[0036] like Figures 1 to 5 As shown, the system also includes a transfer learning module, which is used to transfer the trained model parameters to new wind farm scenarios. Model adaptation can be achieved by fine-tuning with a small amount of field data, reducing training costs and shortening the deployment cycle.

[0037] The aforementioned transfer learning module employs a fine-tuning transfer strategy. The transferred model parameters include the feature extraction layer parameters of the multi-scale deep learning prediction module and the basic network parameters of the deep reinforcement learning decision module. Only the output layer and fully connected layer of the model are fine-tuned. The small amount of field data consists of 1-2 months of wind farm operation data. Fine-tuning uses a mini-batch gradient descent algorithm with a batch size of 32 and 20-50 training rounds. The fine-tuned model needs to be validated with field data. The validation criteria are prediction error ≤3% and adjustment accuracy of the decision strategy ≥95%, ensuring that the model is adapted to the new wind farm scenario. The core formula of the model fine-tuning loss function is expressed as follows: in, The loss function value during model fine-tuning is used to measure the deviation between the model's predicted value and the actual value. The value range is [0, +∞). The smaller the loss value, the better the model's fit. The set of parameters that can be updated during model fine-tuning includes only the parameters of the model output layer and fully connected layers, while the parameters of the feature extraction layer remain unchanged. : Fine-tune the sample size of the field data used, corresponding to 1-2 months of wind farm operation data. ; Model for the first The predicted output value of each sample (AGC command prediction value, wind power output prediction value, or control command prediction value). : No. The actual observed value of each sample, and One-to-one correspondence; : No. The mean squared error term for each sample is used to measure the prediction bias of a single sample. L2 regularization coefficient, with a value range of [0.0001, 0.001], is used to prevent overfitting during model fine-tuning. The model can update the L2 norm squared of the parameters to constrain the range of parameter values ​​and suppress overfitting.

[0038] like Figures 1 to 5 As shown, the above-mentioned visualization monitoring and evaluation module provides multi-dimensional performance indicators, including AGC adjustment accuracy, response time, frequency deviation pass rate, wind turbine fatigue damage degree, and economic benefit evaluation index. The specific definitions and calculation methods for each performance indicator are as follows: AGC adjustment accuracy = (1 - |actual adjustment power - AGC command power| / AGC command power) × 100%, with a pass rate of ≥95%; Response time refers to the time from receiving the AGC command to the actual output of the fan reaching the command requirement, with a pass rate of ≤2s; Frequency deviation pass rate = (duration of frequency deviation within ±0.2Hz / total running time) × 100%; Fan fatigue damage is calculated using Miner's linear cumulative damage theory, with a damage degree ≤0.8 being considered passable; Economic benefit evaluation index = (energy saving benefits after implementing this system + reduced maintenance costs) / system deployment cost, with an evaluation cycle of 1 year. Historical data analysis and evaluation reports are generated monthly, including the changing trends and optimization suggestions for each performance indicator; The formulas for calculating core performance indicators are expressed as follows: AGC adjustment precision: in, : AGC adjustment accuracy (%) The actual regulated power (kW) Power specified by AGC command (kW); Frequency deviation pass rate: in, : This represents the frequency deviation pass rate (%). The operating time (h) is within the frequency deviation range of ±0.2Hz. Total runtime (h); Fan fatigue damage degree: in, : This refers to the fatigue damage degree of the fan. For the wind turbine in the first The actual number of cycles under the stress level. For the wind turbine in the first The fatigue limit number of cycles under the highest stress level. Number of stress levels; Economic benefit assessment index: in, : is an economic benefit assessment index. The annual energy saving benefit (in yuan) after implementing this system. The annual reduction in maintenance costs (in yuan) after the implementation of this system. Total system deployment cost (RMB); The above systems support seamless integration with wind farm energy management systems (EMS) and grid dispatch automation systems (SCADA) to achieve data sharing and command interaction; The system employs standardized communication protocols, with the EMS system using the IEC61850-7-4 standard protocol and the SCADA system using the ModbusTCP protocol. Data sharing includes raw data from the multi-source data acquisition module, prediction results from the multi-scale deep learning prediction module, control strategies from the deep reinforcement learning decision module, and execution data from the dynamic execution and feedback correction module. The data exchange frequency is 100ms / time. Command interaction includes AGC commands issued by the power grid dispatch center, wind turbine operating status uploaded by the system, and regulation execution feedback information, ensuring the system's compatibility and collaborative operation capabilities with existing wind farms and power grid dispatch systems.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A deep learning-based power plant AGC load response optimization system, characterized in that, include: The multi-source data acquisition module is used to acquire multi-dimensional operational data of wind farms and power grids in real time, including but not limited to AGC commands, wind speed, wind direction, wind turbine output, grid frequency, tie line power, ACE value, and wind turbine operating status parameters. The data preprocessing and feature engineering module is used to clean, normalize, fill missing values, and handle outliers of the collected multi-source data, and extract spatiotemporal features, trend features, and correlation features. The multi-scale deep learning prediction module is used to predict the AGC command change trend, wind power output fluctuation range and grid frequency response characteristics within a preset time period based on preprocessed feature data. The deep reinforcement learning decision module is used to dynamically generate the optimal AGC load response strategy, including the total regulating power allocation scheme and the control commands of each wind turbine, by taking the prediction results as input and combining them with the wind power AGC control constraints. The dynamic execution and feedback correction module is used to send the optimal control commands to each wind turbine controller and monitor the execution effect in real time, and adjust the control strategy online through a closed-loop feedback mechanism. The visualization monitoring and evaluation module is used to display the system's operating status, control effects, and performance indicators in real time, and to provide historical data analysis and evaluation reports.

2. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The multi-scale deep learning prediction module adopts a hybrid neural network architecture that integrates LSTM and Transformer. LSTM is used to capture the long-term and short-term dependencies of time series, while Transformer is used to extract global correlation features between multiple variables and output prediction results at multiple time scales.

3. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The deep reinforcement learning decision module adopts an improved DDPG algorithm to construct a Markov decision process model that includes a state space, an action space, and a reward function. The state space includes wind power output prediction, AGC commands, grid frequency deviation, tie-line power deviation, and wind turbine operating constraints. The action space includes power adjustment commands for each wind turbine. The reward function comprehensively considers power tracking accuracy, adjustment rate, frequency stability, and wind turbine loss cost.

4. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The data preprocessing and feature engineering module also includes a feature selection unit, which uses the SHAP value analysis method to quantify the influence weight of each input feature on the AGC response performance, and selects key features to reduce model complexity and improve generalization ability.

5. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The multi-scale deep learning prediction module also integrates meteorological forecast data and dynamically adjusts the weight coefficients of meteorological data and historical operational data through an attention mechanism to improve prediction accuracy.

6. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The deep reinforcement learning decision module also includes a constraint processing unit, which is used to verify in real time whether the control strategy meets physical and operational constraints such as the maximum adjustment rate of the wind turbine, the minimum stable output, and the pitch angle limit, and to automatically correct strategies that do not meet the constraints.

7. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The dynamic execution and feedback correction module adopts a hierarchical control architecture, including a field-level AGC controller and a wind turbine-level controller. The field-level controller is responsible for the overall regulation power allocation, while the wind turbine-level controller is responsible for quickly tracking and executing control commands and feeding back real-time power output data.

8. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The system also includes a transfer learning module, which is used to transfer the trained model parameters to new wind farm scenarios. Model adaptation can be achieved by fine-tuning with a small amount of field data, reducing training costs and shortening the deployment cycle.

9. The power plant AGC load response optimization system based on deep learning according to claim 1, characterized in that: The visualization monitoring and evaluation module provides multi-dimensional performance indicators, including AGC adjustment accuracy, response time, frequency deviation pass rate, wind turbine fatigue damage degree, and economic benefit evaluation index. The system supports seamless integration with wind farm energy management systems and power grid dispatch automation systems, enabling data sharing and command interaction.

Images