A virtual power plant compressed air energy storage system anti-disturbance control method and system

Abstract

This invention relates to a disturbance suppression control method and system for a virtual power plant compressed air energy storage system. It employs a two-level modeling strategy—equipment-level modeling and module-level integration—to construct an AA-CAES system dynamic model. A virtual power plant disturbance simulation model is constructed and coupled with the constructed dynamic model. Time-series data under various disturbance scenarios are collected as a training sample set to train a neural network prediction model, enabling it to possess time-series prediction capabilities. The real-time state collected by the system dynamic model is used to correct the disturbance deviation of the trained neural network output prediction state, resulting in a corrected prediction output. Based on the system setpoint, the corrected prediction output, and the disturbance deviation, optimal control signals for multiple future time periods are obtained. These optimal control signals are executed step-by-step, and the system undergoes dynamic rolling control. Compared with existing technologies, this invention, by identifying system disturbance deviations in real time and embedding a reference trajectory generation stage, can quickly suppress overshoot caused by disturbances.

Description

Technical Field

[0001] This invention relates to the field of virtual power plant control technology, and in particular to a disturbance rejection control method and system for a virtual power plant compressed air energy storage system. Background Technology

[0002] Advanced adiabatic compressed air energy storage (AA-CAES) systems have become core energy storage units in virtual power plants due to their advantages such as high efficiency, lack of geographical limitations, fast response speed, high safety, and zero pollution. They play a crucial role in scenarios such as peak shaving and valley filling, renewable energy integration, and the construction of independent power systems. In existing technologies, the dynamic modeling of AA-CAES systems is mainly based on thermodynamic and fluid dynamics mechanisms. Mathematical models of core equipment such as compressors, heat exchangers, and air tanks are constructed through modular modeling methods, and model parameters are obtained using inverse problem-solving methods to achieve simulation of the system's dynamic characteristics.

[0003] During the operation of a virtual power plant, the AA-CAES system often faces various disturbances, including fluctuations in input power during the energy storage phase, changes in output power demand during the energy release phase, and fluctuations in the mass flow rate of air in the gas storage tank. These disturbances can lead to a deterioration in the system's dynamic response characteristics, such as speed overshoot, power fluctuations, and increased steady-state error.

[0004] To address the aforementioned issues, existing control methods largely employ traditional PID control. However, due to the nonlinear, hysteretic, and dead-zone characteristics of the AA-CAES system, traditional PID control is prone to over / under-regulation leading to instability when deviating from rated operating conditions, exhibiting poor robustness. Furthermore, it relies on manual parameter tuning, making it difficult to adapt to the complex and variable disturbance scenarios of virtual power plants. In addition, while existing technologies have conducted disturbance simulation studies on the AA-CAES system, clarifying the impact of disturbances on the system's dynamic characteristics, they lack targeted optimization control strategies. Summary of the Invention

[0005] The purpose of this invention is to provide a disturbance rejection control method and system for a virtual power plant compressed air energy storage system in order to address the various disturbances that the AA-CAES system often faces.

[0006] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, a disturbance rejection control method for a virtual power plant compressed air energy storage system is provided, comprising the following steps: A two-level modeling strategy of device-level modeling and module-level integration is adopted to construct the dynamic model of the AA-CAES system; A virtual power plant disturbance simulation model is constructed and coupled with the constructed dynamic model. Time series data under various disturbance scenarios are collected as training sample sets to train a neural network prediction model to enable it to have time series prediction capabilities. The perturbation deviation of the predicted output state of the trained neural network is corrected by using the real-time state collected by the dynamic model of the AA-CAES system to obtain the corrected predicted output. Based on the system setpoint, the corrected predicted output, and the disturbance deviation, the optimal control signal for multiple future time periods is obtained. The optimal control signal is executed step by step, and the system is dynamically rolled over.

[0007] As a preferred technical solution, the construction of the dynamic model of the AA-CAES system is as follows: The AA-CAES system is divided into five equipment sub-modules: compressor, heat exchanger for energy storage / release stage, gas storage / release stage storage tank, throttle valve, and turbine. Mathematical models are established for each equipment sub-module. Simulations were performed on each equipment sub-module separately, and the model parameters were adjusted until the error between the module's output data and the actual operating data was less than the set threshold. The compressor module, which was completed by trial and error as a single module, was combined with the heat exchanger module, and the trial and error process was repeated to adjust the model parameters. Then, the gas storage tank and turbine modules were added, and the model parameters were adjusted by trial and error until the error between the simulation data and the actual data of the whole system was less than the set threshold.

[0008] As a preferred technical solution, the collection of the training sample set is specifically as follows: Set the parameters for the energy storage stage, energy release stage, and disturbance scenarios in the energy release stage; inject the signal output by the disturbance simulation module into the corresponding device sub-module through the interface of the AA-CAES system dynamic model; Input the power disturbance during the energy storage stage into the compressor module power interface; Input the flow fluctuation during the energy release phase into the flow interface between the gas storage tank and the throttle valve module; Input the power disturbance during the energy release phase into the power demand interface of the turbine module; The dynamic model of the coupled AA-CAES system is started, and a training sample set is formed by synchronously collecting time-series data of disturbance signal, control input and system state output. The sample includes timestamps, disturbance parameters, control variables and state parameters of each device.

[0009] As a preferred technical solution, the input of the neural network model includes time features, disturbance parameters, current control variables, and system state parameters at historical moments; the output is the system state parameters at future moments, including: rotational speed, power, pressure, and temperature.

[0010] As a preferred technical solution, the dynamic rolling control of the system is specifically as follows: The system acquires AA-CAES system state data in real time, and inputs the current control input, disturbance signal, and system state parameters from the previous three time points into a trained neural network prediction model to generate future predictions. Initial prediction output for each time period; Collect the real-time status and current disturbance signals of each device sub-module in the dynamic model of the AA-CAES system, and calculate the deviation between the actual status and the initial predicted output; The deviation is decomposed into local errors of each device. Each local error is multiplied by its corresponding correction coefficient, and then added to the initial prediction output of the prediction model to obtain the corrected prediction output. Based on system settings Current actual status And the deviation between the actual state and the initial predicted output. Generate the future Reference trajectory for each time period: With reference trajectory and the corrected prediction output As input, within each sampling period, a rolling solution is performed to obtain the objective function that minimizes the difference between the reference trajectory and the corrected predicted output target, along with the weighted sum of the control increments, resulting in the future... The optimal control signal for each time period; Each device is given the current optimal control signal in stages, and the control device tracks the reference trajectory. After entering the next sampling period, the above steps are repeated to update the prediction model, reference trajectory and control signal based on the new system state and disturbance data, so as to realize dynamic rolling control.

[0011] As a preferred technical solution, the generation of the reference trajectory is specifically represented as follows: in, Indicating the future Reference trajectory values ​​for each time period, , It is the time step. This indicates the time it takes for the equipment to reach stable operation; This is the disturbance compensation coefficient; This represents the deviation between the actual state and the initial predicted output.

[0012] As a preferred technical solution, the objective function is as follows: in, The weighting coefficient for the output deviation. Indicating the future Reference trajectory for each time period, Indicating the future The predicted output after correction for each time period To predict the time domain, The weighting coefficient for the increment of the control quantity. To control the time domain, This is the control vector.

[0013] As a preferred technical solution, the control vector includes: turbine regulating valve opening; compressor inverter frequency; heat exchanger medium flow rate; and gas tank flow control valve opening.

[0014] As a second aspect of the present invention, a disturbance rejection control system for a virtual power plant compressed air energy storage system is provided. The control system executes the disturbance rejection control method for the virtual power plant compressed air energy storage system as described above, specifically including: The AA-CAES system dynamic modeling module adopts a two-level modeling strategy of device-level modeling and module-level integration to construct the dynamic model of the AA-CAES system. The virtual power plant disturbance simulation module is used to simulate typical disturbances faced by the AA-CAES system in a virtual power plant. The neural network model predictive control module uses the real-time state collected by the dynamic model of the AA-CAES system to correct the disturbance deviation of the predicted output state of the trained neural network. Based on the system setpoint, the corrected predicted output and the disturbance deviation, the optimal control signal for multiple future time periods is obtained. The sensor data acquisition module and actuator, including the compressor frequency converter, turbine regulating valve, and gas tank flow control valve, receive control signals output by the neural network model predictive control module and adjust the operating status of the equipment.

[0015] As a preferred technical solution, the neural network model prediction and control module includes four sub-units: prediction model, feedback correction, reference trajectory, and rolling optimization, and the specific design is as follows: Predictive model: The input is the valve opening and disturbance signal, and the output is the key state parameters of the AA-CAES system. The input and output data of the dynamic model of the AA-CAES system are collected offline as training samples, and the network weights and thresholds are iteratively optimized by gradient descent method. Feedback correction: Calculate the error between the actual output and the predicted output, and correct the predicted value at future time points using the error correction coefficient; Reference trajectory: Design an exponential reference trajectory to achieve a smooth transition of the system output; Rolling optimization: With the goal of minimizing the weighted sum of output prediction error and control increment, an optimization performance index function is constructed, and the optimal control signal is solved by gradient descent.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention proposes a coupled modeling method for disturbances in the AA-CAES system and a virtual power plant. It constructs a collaborative mechanism of neural network prediction, disturbance perception, and dynamic compensation of the reference trajectory. The predictive model anticipates the impact of disturbances in advance, identifies system disturbance deviations in real time, and embeds them into the reference trajectory generation process. Optimal control signals are generated through rolling optimization. This enables rapid perception and trajectory adaptation to disturbances such as grid load fluctuations and gas storage tank pressure changes. Compared to traditional fixed trajectory generation methods, the disturbance response delay is reduced by more than 30%.

[0017] 2) This invention combines neural network model predictive control with AA-CAES system anti-disturbance control, and proposes a rolling optimization strategy adapted to virtual power plants. The dynamic output of the reference trajectory is used as the real-time target of rolling optimization. Combined with the fast solution characteristics of gradient descent method, the control signal is re-optimized in each sampling period. The optimal control command can be output within one sampling period after the disturbance occurs, and the speed / power overshoot caused by the disturbance can be quickly suppressed.

[0018] 3) This invention adopts hierarchical execution control signal logic. The controller only executes the optimal control signal for the current time period. At the same time, it dynamically updates the prediction model and reference trajectory based on new disturbance data. This ensures a fast response to disturbances and avoids system oscillations caused by the superposition of instructions from multiple time periods, thereby improving the robustness of the CAES energy release process under disturbances. Attached Figure Description

[0019] Figure 1 This is a flowchart of the anti-disturbance control method for a virtual power plant compressed air energy storage system based on neural network model predictive control according to the present invention.

[0020] Figure 2 This is a schematic diagram of the anti-disturbance control system of the virtual power plant compressed air energy storage system based on neural network model predictive control according to the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0022] Example 1 This invention provides a disturbance rejection control method for a virtual power plant AA-CAES system based on neural network model predictive control. By constructing a dynamic model of the AA-CAES system and a disturbance simulation model of the virtual power plant, a neural network model predictive controller is designed to achieve early prediction and precise suppression of disturbances, reduce system overshoot, shorten adjustment time, and improve robustness. This ensures that the AA-CAES system operates stably and efficiently in complex disturbance scenarios of the virtual power plant, providing support for the optimized scheduling of the virtual power plant.

[0023] like Figure 1 As shown, the steps of the disturbance rejection control of the virtual power plant AA-CAES system of the present invention are as follows: S1. Based on thermodynamics, fluid dynamics, and the laws of energy and mass conservation, a two-level modeling strategy of "equipment-level modeling - module-level integration" is adopted to construct a high-precision dynamic model of the AA-CAES system. The specific steps are as follows: S1.1 Device-level modular decomposition and modeling.

[0024] The AA-CAES system is divided into five core equipment modules: compressor, heat exchanger for energy storage / release stage, gas storage / release stage storage tank, throttle valve, and turbine. Mathematical models are established for each module. Compressor model: Ignoring the internal gas volume and treating it as an adiabatic process, the outlet temperature equation is: in, , The air temperatures at the compressor inlet and outlet are... The specific heat ratio of air. For insulation efficiency, This refers to the pressure ratio.

[0025] Heat exchanger module: Considered as an open energy system, the air energy balance equation is: in, For the air quality inside the heat exchanger, The mass flow rate of air. The specific heat capacity of air at constant pressure. , These are the inlet and outlet air temperatures of the heat exchanger. The energy released into the air.

[0026] Gas storage tank model: Considered as a constant volume process, the pressure change equation is: in, For import flow, The temperature of the gas inside the storage tank. The heat transfer coefficient, This refers to the heat exchange surface area of ​​the gas storage tank. For ambient temperature, The specific heat capacity at constant volume of air. The volume of the gas storage tank. is the gas constant.

[0027] Throttling valve model: Treated as an adiabatic process, air is considered an ideal gas with a constant specific heat. The flow equation is: in, The mass flow rate of air at the outlet of the throttle valve is kg / min; The fluid compressibility coefficient, For valve admittance, For the throttle valve characteristic function, The air inlet density is expressed in kg / m³. This represents the pressure difference across the throttle valve.

[0028] Turbine model: Considered as the reverse process of a compressor, the outlet temperature equation is: in, This refers to the turbine's outlet temperature. The inlet temperature of the turbine; The specific heat ratio of air; This refers to the airflow rate of the turbine.

[0029] S1.2 module parameter inversion optimization.

[0030] The key parameters of the model are obtained by using an inversion problem-solving method of "single module trial and error + module-by-module trial and error". Single-module trial and error: Taking a heat exchanger module as an example, fix the constant coefficient values ​​of the module and assume parameters that cannot be calculated (such as heat exchange efficiency). Input the actual inlet operating data of the TICC-500 power plant (air flow rate 1668 kg / h, inlet pressure 0.099). Start the simulation at 6.041MPa and adjust the unknown parameters until the error between the module's output data and the actual operating data is ≤±2%.

[0031] Module-by-module trial and error: The completed compressor module is connected to the heat exchanger module, and the trial and error process is repeated. The gas storage tank and turbine module are added in turn until the error between the simulation data and the actual data of the whole system is ≤ ±2%.

[0032] S2. Construct a virtual power plant disturbance simulation model and couple it with the S1 dynamic model.

[0033] S2.1 sets the parameter range for three types of disturbance scenarios.

[0034] Input power disturbance during the energy storage phase: initial power 6461.91kW, rising to 7045.15kW at 200min, and decreasing to 5991.79kW at 350min; Output power perturbation during the energy release phase: initial power 1.0 pu, stepping to 0.85 pu, 0.7 pu, 0.85 pu, and 1.0 pu in sequence, with each phase lasting 50 minutes; Flow fluctuation during the energy release phase: base flow rate 1765 kg / min, superimposed with sinusoidal fluctuations of ±5% and ±10%, with a fluctuation period of 60 s.

[0035] S2.2 Disturbance signal injection.

[0036] The signal output from the disturbance simulation module is injected into the corresponding device sub-module through the interface designed in S1, for example: Energy storage stage input power disturbance → compressor module power interface; Energy release phase flow fluctuation → gas storage tank + throttle valve module flow interface; Power disturbance during energy release phase → power demand interface for turbine module.

[0037] S2.3 Data Acquisition and Sample Construction.

[0038] After the coupled model is started, the time-series data of "disturbance signal - control input - system status output" is collected synchronously through the sensor data acquisition module (≥3000 sets of data are collected for each disturbance scenario) to form the training sample set required by S3. The sample includes timestamps, disturbance parameters, control variables, and status parameters of each device to ensure that the sample covers the dynamic response of the entire disturbance process.

[0039] S3. Train a neural network prediction model with time series prediction capabilities.

[0040] S3.1 Sample preprocessing.

[0041] The time-series samples collected by S2 were normalized and divided into a training set (80%) and a validation set (20%). The sample input dimension included: time features (sampling time). The system parameters include: the identifiers of the previous three historical moments, disturbance parameters (such as input power value, flow fluctuation amplitude), current control variables (regulating valve opening, inverter frequency, etc.), and system status parameters of the previous three moments (speed, power, etc.).

[0042] S3.2 Neural Network Structure Design.

[0043] Set to "8 inputs - 12 hidden layers - 4 outputs". The input layer contains time characteristics, disturbance parameters, control variables, and historical states. The output layer contains system state parameters (speed, power, pressure, temperature) for the next time step.

[0044] S3.3 model training.

[0045] The learning rate was set to 0.01, the number of iterations was 1000, and the gradient descent method was used to optimize the weight threshold. The training was continued until the prediction error on the validation set was ≤2%, ensuring that the model had the ability to make time-series predictions.

[0046] S4. Start the neural network model predictive control.

[0047] S4.1 Neural Network Prediction Model Invocation.

[0048] Based on the current time t The control input, disturbance signal, and system state parameters from the previous three time points (from the real-time output of the S1 model) are used to predict the future using a trained model. =System state parameters for 7 time periods .

[0049] S4.2 system real-time status and disturbance perception.

[0050] By collecting real-time status (compressor speed, heat exchanger temperature, gas tank pressure, etc.) and current disturbance signals of each device submodule in the S1 model using sensors, the deviation between the actual state and the predicted state is calculated. The deviation term is decomposed into the local errors of each equipment sub-module (such as speed error and temperature error).

[0051] S4.3 Feedback Correction.

[0052] Will Decomposed into local errors of each equipment sub-module (such as compressor speed error) Heat exchanger temperature error Gas tank pressure error ), multiplied by the corresponding correction factor respectively =0.2、 =0.15、 =0.25, thus obtaining the corrected prediction output. This improves prediction accuracy.

[0053] S4.4 Reference trajectory generation.

[0054] Input value: System setting (e.g., rated speed, target power), current actual status (From S1 model), perturbation bias S1 model equipment characteristic parameters (turbine response time) =15min, compressor stabilization time =25min).

[0055] Core formula: in, Indicating the future Reference trajectory values ​​for each time period, , It is the time step. This indicates the time it takes for the equipment to reach stable operation, adapted according to the type of disturbance (power disturbance is taken as...). =15min, flow disturbance taken =20min); The disturbance compensation coefficient (adjusted according to equipment type: speed-related) is taken as follows: =0.8, pressure-related value =0.6); Output: Future Reference trajectory values ​​for each time period The trajectory adapts to the dynamic response characteristics of the S1 model, avoiding overshoot.

[0056] S4.5 Scrolling Optimization Solution.

[0057] by To optimize the target, a performance metric function is constructed: in, The weighting coefficient for the output deviation. The weighting coefficient for the increment of the control quantity. To predict the time domain, To control the time domain, The control vector includes: turbine regulating valve opening (controlling speed / power); compressor inverter frequency (controlling input power); heat exchanger medium flow rate (controlling inlet and outlet temperatures); and gas receiver flow control valve opening (controlling outlet flow rate).

[0058] The objective function is minimized by using gradient descent in each sampling period to obtain the future... The optimal control signal sequence for each time period Only the first control signal is taken. Used for current control, achieving "local optimization + real-time adaptation".

[0059] S4.6 control signal execution.

[0060] The controller executes control signals according to the logic of "real-time acquisition - correction prediction - optimization solution - step-by-step execution - dynamic update". The specific steps are as follows: Signal Acquisition: The controller acquires the compressor / expander speed / power (actual output) of the AA-CAES system in real time. ), Current opening degree of regulating valve (current control quantity) Data on disturbance sources such as power grid load and gas storage tank pressure; Prediction Correction: Input the collected data into the neural network prediction model of S4.1 to generate an initial prediction output. And obtained through feedback correction ; Trajectory Generation: Based on the corrected prediction output, setpoint, and disturbance deviation, the reference trajectory generation logic in S4.3 is executed, and the output is... ; Optimization solution: with and Using this as input, perform rolling optimization in S4.4 to obtain the optimal control signal sequence. ; Step-by-step execution: The controller only sends the current optimal control signal to the regulating valve actuator. (Such as valve opening adjustment commands), control the compressor / expander speed / power to track the reference trajectory; Dynamic update: After entering the next sampling period, the controller repeats steps 1-5 to update the prediction model, reference trajectory and control signal based on the new system state and disturbance data, so as to realize real-time tracking and rapid response to disturbances.

[0061] S4.7 Closed-loop iteration.

[0062] Repeat S4.1 Step S4.6, with a cycle of 100ms, achieves dynamic scrolling control.

[0063] Compared with traditional PID control, the predictive control method proposed in this application has the following advantages: 1) Significantly improved anti-disturbance performance: The neural network model predictive control of this invention can reduce the maximum speed overshoot of the AA-CAES system by 3.2% and the maximum power overshoot by 7.2%. In the ±10% flow fluctuation scenario, the turbine output power fluctuation range is reduced from ±9.3% to ±3.5%, effectively suppressing the impact of virtual power plant disturbances on the system. 2) Faster dynamic response speed: By predicting disturbances in advance through predictive models and generating control signals through rolling optimization, the system's stabilization time during input power disturbances is shortened from 40 minutes to 15 minutes, and the power regulation stabilization time during the energy release phase is shortened from 25 minutes to 10 minutes, thereby improving the system's responsiveness to virtual power plant dispatch commands. 3) Enhanced robustness: The neural network model has good nonlinear mapping capability, which can adapt to the characteristic changes of the AA-CAES system under different disturbance intensities and different operating conditions, avoid the instability problem caused by improper parameter tuning in traditional PID control, and ensure the long-term stable operation of the system in the complex scenario of virtual power plant. 4) Reduced energy efficiency loss: By precisely controlling the system to reduce the non-steady-state operating time caused by disturbances, the steady-state operating efficiency of the AA-CAES system is improved by 2% to 3%, which provides support for the overall energy efficiency optimization of the virtual power plant and reduces operating costs.

[0064] Example 2 As another embodiment of the present invention, this embodiment also provides a disturbance rejection control system for a virtual power plant AA-CAES system based on neural network model predictive control, including an AA-CAES system dynamic modeling module, a virtual power plant disturbance simulation module, a neural network model predictive control module, a sensor data acquisition module, and an actuator. The modules are interconnected via a data bus. Figure 2 The specific architecture shown is as follows: The AA-CAES system dynamic modeling module establishes mathematical models of core equipment such as compressors, heat exchangers in the energy storage / release stage, gas tanks, turbines, and throttling valves based on thermodynamic and fluid dynamics mechanisms. It adopts a modular modeling method to encapsulate the models into independent sub-modules and builds the overall dynamic model of the system through the MATLAB / Simulink platform. The model parameters are optimized and obtained through the inversion problem solving method to ensure that the steady-state operation error is within ±2%.

[0065] Virtual power plant disturbance simulation module: Simulates typical disturbances faced by the AA-CAES system in a virtual power plant, including: input power disturbances during the energy storage stage, output power disturbances during the energy release stage, and flow fluctuations during the energy release stage.

[0066] The neural network model prediction and control module comprises four sub-units: prediction model, feedback correction, reference trajectory, and rolling optimization. The specific design is as follows: Prediction model: Constructed using a BP neural network, with inputs including valve opening and disturbance signals (power / flow rate changes), and outputs including key state parameters of the AA-CAES system (compressor speed, turbine power, and gas tank pressure). The input and output data of the AA-CAES system dynamic model are collected offline as training samples, and the network weights and thresholds are iteratively optimized using gradient descent to ensure that the prediction error is ≤2%. Feedback correction: Calculate the error between the actual output and the predicted output, and correct the predicted value at future time by using the error correction coefficient h (value range of 0.1~0.3); Reference trajectory: Design an exponential reference trajectory to achieve a smooth transition of the system output; Rolling optimization: With the goal of minimizing the weighted sum of output prediction error and control increment, an optimization performance index function is constructed, and the optimal control signal is solved by gradient descent.

[0067] Sensor data acquisition module: Deploys pressure sensor (accuracy ±0.01MPa), temperature sensor (accuracy ±0.5℃), speed sensor (accuracy ±1r / min), and power sensor (accuracy ±0.5%FS) to collect the operating parameters of each device in the AA-CAES system in real time, with a sampling frequency of 100Hz.

[0068] Actuator module: includes compressor frequency converter, turbine regulating valve, and gas tank flow control valve. It receives control signals output by neural network model predictive control module to achieve precise adjustment of equipment operating status.

[0069] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A disturbance rejection control method for a virtual power plant compressed air energy storage system, characterized in that the steps include: include: A two-level modeling strategy of device-level modeling and module-level integration is adopted to construct the dynamic model of the AA-CAES system; A virtual power plant disturbance simulation model is constructed and coupled with the constructed dynamic model. Time series data under various disturbance scenarios are collected as training sample sets to train a neural network prediction model to enable it to have time series prediction capabilities. The perturbation deviation of the predicted output state of the trained neural network is corrected by using the real-time state collected by the dynamic model of the AA-CAES system to obtain the corrected predicted output. Based on the system setpoint, the corrected predicted output, and the disturbance deviation, the optimal control signal for multiple future time periods is obtained. The optimal control signal is executed step by step, and the system is dynamically rolled over.

2. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 1, characterized in that, The construction of the dynamic model of the AA-CAES system is as follows: The AA-CAES system is divided into five equipment sub-modules: compressor, heat exchanger for energy storage / release stage, gas storage / release stage storage tank, throttle valve, and turbine. Mathematical models are established for each equipment sub-module. Simulations were performed on each equipment sub-module separately, and the model parameters were adjusted until the error between the module's output data and the actual operating data was less than the set threshold. The compressor module, which was completed by trial and error as a single module, was combined with the heat exchanger module, and the trial and error process was repeated to adjust the model parameters. Then, the gas storage tank and turbine modules were added, and the model parameters were adjusted by trial and error until the error between the simulation data and the actual data of the whole system was less than the set threshold.

3. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 1, characterized in that, The collection of the training sample set is as follows: Set the parameters for the energy storage stage, energy release stage, and disturbance scenarios in the energy release stage; inject the signal output by the disturbance simulation module into the corresponding device sub-module through the interface of the AA-CAES system dynamic model; Input the power disturbance during the energy storage stage into the compressor module power interface; Input the flow fluctuation during the energy release phase into the flow interface between the gas storage tank and the throttle valve module; Input the power disturbance during the energy release phase into the power demand interface of the turbine module; The dynamic model of the coupled AA-CAES system is started, and a training sample set is formed by synchronously collecting time-series data of disturbance signal, control input and system state output. The sample includes timestamps, disturbance parameters, control variables and state parameters of each device.

4. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 1, characterized in that, The input of the neural network model includes time features, disturbance parameters, current control variables, and system state parameters at historical moments; the output is the system state parameters at future moments, including: rotational speed, power, pressure, and temperature.

5. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 1, characterized in that, The specific details of the dynamic rolling control of the system are as follows: The system acquires AA-CAES system state data in real time, and inputs the current control input, disturbance signal, and system state parameters from the previous three time points into a trained neural network prediction model to generate future predictions. Initial prediction output for each time period; Collect the real-time status and current disturbance signals of each device sub-module in the dynamic model of the AA-CAES system, and calculate the deviation between the actual status and the initial predicted output; The deviation is decomposed into local errors of each device. Each local error is multiplied by its corresponding correction coefficient, and then added to the initial prediction output of the prediction model to obtain the corrected prediction output. Based on system settings Current actual status And the deviation between the actual state and the initial predicted output. Generate the future Reference trajectory for each time period: With reference trajectory and the corrected prediction output As input, the objective function, which aims to minimize the weighted sum of the output prediction error and the control increment, is solved continuously within each sampling period to obtain the future... The optimal control signal for each time period; Each device is given the current optimal control signal in stages, and the control device tracks the reference trajectory. After entering the next sampling period, the above steps are repeated to update the prediction model, reference trajectory and control signal based on the new system state and disturbance data, so as to realize dynamic rolling control.

6. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 5, characterized in that, The generation of the reference trajectory is specifically represented as follows: in, Indicating the future Reference trajectory values ​​for each time period, , It is the time step. This indicates the time it takes for the equipment to reach stable operation; This is the disturbance compensation coefficient; This represents the deviation between the actual state and the initial predicted output.

7. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 5, characterized in that, The specific objective function is as follows: in, The weighting coefficient for the output deviation. Indicating the future Reference trajectory for each time period, Indicating the future The predicted output after correction for each time period To predict the time domain, The weighting coefficient for the increment of the control quantity. To control the time domain, This is the control vector.

8. The anti-disturbance control method for a virtual power plant compressed air energy storage system according to claim 5, characterized in that, The control vector includes: turbine regulating valve opening; compressor inverter frequency; heat exchanger medium flow rate; and gas tank flow control valve opening.

9. A disturbance rejection control system for a virtual power plant compressed air energy storage system, characterized in that, The control system executes the anti-disturbance control method for the virtual power plant compressed air energy storage system as described in any one of claims 1-8, specifically including: The AA-CAES system dynamic modeling module adopts a two-level modeling strategy of device-level modeling and module-level integration to construct the dynamic model of the AA-CAES system. The virtual power plant disturbance simulation module is used to simulate typical disturbances faced by the AA-CAES system in a virtual power plant. The neural network model predictive control module uses the real-time state collected by the dynamic model of the AA-CAES system to correct the disturbance deviation of the predicted output state of the trained neural network. Based on the system setpoint, the corrected predicted output and the disturbance deviation, the optimal control signal for multiple future time periods is obtained. The sensor data acquisition module and actuator, including the compressor frequency converter, turbine regulating valve, and gas tank flow control valve, receive control signals output by the neural network model predictive control module and adjust the operating status of the equipment.

10. The anti-disturbance control system for a virtual power plant compressed air energy storage system according to claim 9, characterized in that, The neural network model prediction and control module comprises four sub-units: prediction model, feedback correction, reference trajectory, and rolling optimization, with the following specific design: Predictive model: The input is the valve opening and disturbance signal, and the output is the key state parameters of the AA-CAES system. The input and output data of the dynamic model of the AA-CAES system are collected offline as training samples, and the network weights and thresholds are iteratively optimized by gradient descent method. Feedback correction: Calculate the error between the actual output and the predicted output, and correct the predicted value at future time points using the error correction coefficient; Reference trajectory: Design an exponential reference trajectory to achieve a smooth transition of the system output; Rolling optimization: With the goal of minimizing the weighted sum of output prediction error and control increment, an optimization performance index function is constructed, and the optimal control signal is solved by gradient descent.

Images