Control method for synchronous grid connection of thermal energy storage system and power grid

Abstract

The invention provides a control method for synchronous grid connection of a thermal energy storage system and a power grid, belongs to the technical field of thermal energy storage systems, and realizes rapid synchronous tracking by constructing a self-adaptive phase-locked loop controller adopting a fuzzy control algorithm and combining with a feed-forward compensation mechanism. A heat storage power prediction model based on a heat transfer differential equation is established to provide a prospective control basis, a sliding mode variable structure power controller is designed, power fluctuation is eliminated by adopting a reaching law method, and a neural network control model based on progressive network growth and element gradient optimization is constructed to realize intelligent multi-target coordination optimization. Meanwhile, power factor compensation control and an electric energy quality comprehensive treatment strategy are established to guarantee grid-connected electric energy quality, and finally safe and reliable operation of the system is ensured through a system operation state evaluation mechanism and a standby control strategy, so that the technical problems of response lag and insufficient control precision during synchronous grid connection of the thermal energy storage system and the power grid are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of thermal energy storage systems, and more specifically, relates to a control method for synchronously connecting a thermal energy storage system with the power grid. Background Technology

[0002] With the rapid development of renewable energy power generation technology and the increasing demand for grid peak shaving, thermal energy storage systems, as important energy storage and regulation devices, are widely used in power systems for peak shaving, valley filling, and frequency regulation. Traditional thermal energy storage grid-connected control technology mainly uses fixed-parameter phase-locked loop synchronous control and linear proportional-integral power regulators to achieve electrical connection with the grid. By detecting grid frequency signals and phase angle information, the system's output power and phase relationship are adjusted to enable the thermal energy storage device to maintain synchronous operation with the grid and stably transmit power. However, in practical engineering applications, traditional control methods show obvious limitations when facing complex operating conditions such as grid frequency fluctuations, load abrupt changes, and voltage disturbances. Due to the use of fixed control parameters and simple linear control algorithms, the system cannot quickly adapt to changes in grid operating conditions, leading to a decline in synchronous tracking performance. At the same time, traditional control strategies lack in-depth modeling of the thermodynamic characteristics and dynamic response process of the thermal energy storage medium, making it impossible to accurately predict power output change trends, resulting in control decisions often lagging behind the actual needs of the system. The control systems in the current technology lack adaptive learning and parameter optimization capabilities. The controller parameters are mainly set by engineering experience, which makes it difficult to dynamically adjust them for different operating conditions. Therefore, under the complex and ever-changing conditions of the power grid, problems such as control response lag and low synchronization accuracy are likely to occur, which affects the stability and reliability of the grid-connected operation of the thermal energy storage system. Summary of the Invention

[0003] In view of this, the present invention provides a control method for synchronous grid connection of a thermal energy storage system with the power grid, which can solve the technical problems of response lag and insufficient control accuracy in the prior art when the thermal energy storage system is synchronously connected with the power grid.

[0004] This invention is implemented as follows: It provides a control method for synchronous grid connection of a thermal energy storage system with the power grid. This method establishes a multi-parameter real-time monitoring network for the thermal energy storage system, collecting operating parameters such as thermal energy storage medium temperature distribution, thermal storage capacity, heat release power, grid frequency, phase angle, voltage amplitude, power factor, and harmonic content. An adaptive phase-locked loop (PLL) controller is constructed, which, combined with a feedforward compensation algorithm, detects changes in grid frequency and phase angle data. A fuzzy control algorithm is used to dynamically adjust the PLL parameters based on the grid frequency and phase angle data. A thermal energy storage power prediction model is established to predict the power output curve for future periods. A sliding mode variable structure power control model is designed. The system includes a power controller that eliminates power output fluctuations based on predicted power and heat release power data; a grid-connected optimized control model for thermal energy storage that outputs optimized control commands; a power factor compensation control strategy that improves energy utilization efficiency by adjusting power factor correction circuit parameters when the power factor is below 0.9; comprehensive power quality management that activates a multi-stage active filter system when harmonic content exceeds 5%; and a system operation status evaluation mechanism and backup control strategy that automatically switch to backup control when synchronization accuracy error is ≥0.1%, power fluctuation coefficient is ≥5%, or power quality index is <90 points.

[0005] Specifically, the step of establishing a multi-parameter real-time monitoring network for the thermal energy storage system involves using a distributed sensor network to achieve millisecond-level data updates, thereby obtaining data on the thermal energy storage medium temperature distribution, thermal storage capacity, heat release power, grid frequency, phase angle, voltage amplitude, power factor, and harmonic content.

[0006] Specifically, the adaptive phase-locked loop controller adopts a second-order generalized integrator structure to achieve fast synchronous tracking. The feedforward compensation algorithm predicts the phase change trend through the differential signal of the grid voltage. The fuzzy control algorithm dynamically adjusts the proportional-integral parameters according to the frequency deviation and phase error.

[0007] Specifically, the thermal energy storage power prediction model is a power prediction model established based on the variation law of temperature distribution data and thermal storage capacity data of the thermal energy storage medium and the thermodynamic characteristics of the thermal storage medium. It is based on the heat transfer differential equation of the thermal storage medium, where the predicted power is the product of the correction coefficient, the temperature difference, and the mass flow rate, multiplied by the rated power.

[0008] Specifically, the sliding mode variable structure power controller employs a reaching law method and an exponential reaching law. The reaching speed is adaptively adjusted according to the power error amplitude. Combined with power smoothing filtering technology, the power fluctuation amplitude is controlled within a set threshold range. The power smoothing filtering is implemented using a moving average algorithm combined with Kalman filtering.

[0009] Specifically, the thermal energy storage grid-connected optimization control model is a multi-layer feedforward neural network structure, which includes an input layer, three hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer outputs control commands through the Sigmoid function. The inputs are grid frequency data, phase angle data, and stable power control signals, and the output is optimized control commands.

[0010] Specifically, the power factor compensation control strategy maintains the existing circuit configuration without change when the power factor data is in the range [0.9, 1.0]. The power factor correction circuit parameters include the capacitor bank capacity and the timing of the switching operation. The power factor correction control signal is used to adjust the switching state of the capacitor bank.

[0011] Specifically, the comprehensive power quality management involves adopting a single-stage active filtering scheme if the harmonic content data is controlled within the range (0, 5%). This scheme uses an imbalance compensation algorithm to adjust the three-phase voltage balance and eliminate harmonic interference. The multi-stage active filtering system includes a fundamental wave filtering unit, a low-order harmonic filtering unit, and a high-order harmonic filtering unit. The single-stage active filtering scheme uses only the fundamental wave filtering unit.

[0012] Specifically, the system operation status evaluation mechanism comprehensively evaluates the synchronization control signal, stable power control signal, power factor correction control signal, and power quality control signal, and calculates the synchronization accuracy error, power fluctuation coefficient, and power quality index. The unbalance compensation algorithm calculates the negative sequence current based on the symmetrical component method and performs real-time compensation.

[0013] In the off-grid operation mode, the electrical connection with the power grid is cut off, the thermal energy storage system independently supplies power to the local load, and islanding detection technology is used to monitor the power grid status. When the power grid returns to normal, it automatically reconnects to the grid.

[0014] The process includes several steps, prior to establishing the optimal control model for grid-connected thermal energy storage. This involves collecting grid operation data and thermal energy storage system response data under different operating conditions to create a dataset encompassing 1000 typical scenarios, including grid frequency fluctuations, load changes, and thermal storage temperature variations. Before establishing the optimal control model, the process also includes model training. The Adam optimizer is used for parameter updates, with a learning rate of 0.001, a batch size of 256, and 500 training epochs. Cross-validation is employed to evaluate model performance.

[0015] Furthermore, the thermal energy storage grid-connected optimization control model also includes a dynamic architecture expansion mechanism based on progressive network growth and a second-order optimization framework based on meta-gradients. The dynamic architecture expansion mechanism based on progressive network growth specifically monitors the convergence of the loss function and the gradient change trend during network training. When insufficient network capacity is detected, new neurons and connection layers are automatically added to achieve dynamic expansion of the network structure. The second-order optimization framework based on meta-gradients specifically utilizes the approximate calculation of the Hessian matrix to obtain the curvature information for parameter updates, and adjusts the parameter update direction and step size by calculating the gradient. The second-order optimization framework based on meta-gradients uses the L-BFGS algorithm to approximate the second-order derivative information, avoiding the huge computational overhead of directly calculating the Hessian matrix, and simultaneously utilizes historical gradient information to construct an approximate curvature model.

[0016] This invention constructs a multi-level coordinated control system based on adaptive phase-locked loop (PLL) and intelligent predictive control. It employs fuzzy control algorithms to dynamically adjust control parameters, combined with a feedforward compensation mechanism and a thermal energy storage power prediction model, effectively solving the problems of response lag and insufficient accuracy in traditional control methods. This invention utilizes fuzzy control theory to optimize PLL parameters in real time based on grid frequency deviation and phase error, enabling the system to quickly track grid state changes. By predicting grid parameter change trends through a feedforward compensation algorithm, it transforms passive response into active predictive control, significantly improving the system's dynamic response speed. Simultaneously, the power prediction model based on the differential equation of heat transfer in the thermal energy storage medium accurately predicts the power output curve for future periods, providing precise reference signals for the sliding mode variable structure controller and effectively eliminating control decision lag. The progressive neural network architecture and meta-gradient second-order optimization algorithm introduced in this invention endow the control system with self-learning and self-optimization capabilities, enabling dynamic adjustment of control strategies according to the complexity of the operating environment. Compared to traditional fixed-parameter control methods, it exhibits higher control accuracy and faster response speed. In summary, this invention solves the technical problems of response lag and insufficient control accuracy mentioned in the background art when thermal energy storage systems are synchronously connected to the grid. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention.

[0018] Figure 2 The diagram shows the power control response characteristics of the thermal energy storage system in this embodiment.

[0019] Figure 3 This is a graph showing the changes in system operation status evaluation indicators in the embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0021] like Figure 1 The diagram shown is a flowchart of a control method for synchronizing a thermal energy storage system with the power grid, provided by this invention. This method includes the following steps: S01. Establish a multi-parameter real-time monitoring network for the thermal energy storage system, collect operating parameters such as thermal energy storage medium temperature distribution, thermal storage capacity, heat release power, grid frequency, phase angle, voltage amplitude, power factor, and harmonic content, and achieve millisecond-level data updates through a distributed sensor network to obtain thermal energy storage medium temperature distribution data, thermal storage capacity data, heat release power data, grid frequency data, phase angle data, voltage amplitude data, power factor data, and harmonic content data; S02. Construct an adaptive phase-locked loop controller, combine a feedforward compensation algorithm to detect changes in the power grid frequency data and phase angle data, and use a fuzzy control algorithm to dynamically adjust the phase-locked loop parameters according to the power grid frequency data and phase angle data to achieve fast synchronous tracking and output a synchronous control signal. S03. Establish a thermal energy storage power prediction model. Based on the variation law of the temperature distribution data and thermal storage capacity data of the thermal energy storage medium and the thermodynamic characteristics of the thermal storage medium, predict the power output curve for future periods, provide a forward-looking basis for power control, and output predicted power data. S04. Design a sliding mode variable structure power controller, use the approach law method to eliminate power output fluctuations based on the predicted power data and heat release power data, and combine power smoothing filtering technology to control the power fluctuation amplitude within the set threshold range, and output a stable power control signal. S05. Construct a grid-connected optimization control model for thermal energy storage, input the grid frequency data, phase angle data, and stable power control signal, and achieve multi-objective coordinated optimization control through a dynamic architecture expansion mechanism based on progressive network growth and a second-order optimization framework based on meta-gradient, and output optimization control commands. S06. Establish a power factor compensation control strategy. When the power factor data is lower than 0.9, improve the power utilization efficiency by adjusting the power factor correction circuit parameters and output a power factor correction control signal. When the power factor data is in the range of [0.9, 1.0], maintain the existing circuit configuration without making any changes. S07. Implement comprehensive power quality management. When the harmonic content data exceeds 5%, start a multi-stage active filter system to eliminate harmonic interference. If the harmonic content data is controlled within the range of (0, 5%), a single-stage active filter scheme can meet the requirements. Adjust the three-phase voltage balance through the imbalance compensation algorithm and output a power quality control signal. S08. Establish a system operation status evaluation mechanism and backup control strategy. Based on the synchronization control signal, stable power control signal, power factor correction control signal and power quality control signal, a comprehensive evaluation is performed to calculate the synchronization accuracy error, power fluctuation coefficient and power quality index. When the synchronization accuracy error is ≥0.1% or the power fluctuation coefficient is ≥5% or the power quality index is <90 points, the system automatically switches to the backup control strategy. The backup control strategy includes a power reduction operation mode and an off-grid operation mode.

[0022] The adaptive phase-locked loop controller employs a second-order generalized integrator structure. The feedforward compensation algorithm predicts the phase change trend using the differential signal of the grid voltage. The fuzzy control algorithm dynamically adjusts the proportional-integral parameters based on the frequency deviation and phase error. The thermal energy storage power prediction model is based on the heat transfer differential equation of the thermal storage medium, and the prediction formula is expressed as follows: ,in The predicted power unit is kW. For correction factor, The unit of temperature difference is , The mass flow rate is expressed in kg / s, and the subscript ref indicates a reference value.

[0023] The sliding mode variable structure power controller employs an exponential reaching law, with the reaching speed adaptively adjusted according to the power error amplitude. The power smoothing filter utilizes a moving average algorithm combined with Kalman filtering. The power fluctuation amplitude is controlled within 5% of the rated power, and the response time does not exceed 100ms. The power factor correction circuit parameters include capacitor bank capacity and switching sequence, and the power factor correction control signal is used to adjust the switching state of the capacitor bank. The multi-stage active filtering system includes a fundamental frequency filter unit, a low-order harmonic filter unit, and a high-order harmonic filter unit; the single-stage active filtering scheme uses only the fundamental frequency filter unit. The unbalance compensation algorithm calculates the negative sequence current based on the symmetrical component method and performs real-time compensation.

[0024] The power reduction operation mode in the backup control strategy adjusts the output power of the thermal energy storage system to 50% of its rated power. Power control is achieved by reducing the flow rate of the thermal storage medium and adjusting the heat exchange area. This mode is suitable for operating conditions where grid frequency fluctuations are large but still within acceptable limits. The off-grid operation mode disconnects the electrical connection with the grid, allowing the thermal energy storage system to independently supply power to local loads. Islanding detection technology is used to monitor the grid status, and the system automatically reconnects to the grid when it returns to normal. The formula for calculating the synchronization accuracy error is as follows: ,in For synchronization accuracy error, The phase angle of the power grid is expressed in rad. The system phase angle is in rad. The formula for calculating the power fluctuation coefficient is as follows: ,in For power fluctuation coefficient, The power at the i-th sampling point is in kW. The average power is expressed in kW, and N is the total number of sampling points.

[0025] The specific structure of the thermal energy storage grid-connected optimization control model is a multi-layer feedforward neural network, including an input layer, three hidden layers, and an output layer. The input layer receives grid parameters and thermal energy storage system state parameters. The hidden layers use the ReLU activation function, and the output layer outputs control commands through the Sigmoid function. The total number of network parameters is 15,000, and batch normalization technology is used to improve training stability. The steps for establishing the training dataset of the thermal energy storage grid-connected optimization control model specifically include collecting grid operation data and thermal energy storage system response data under different operating conditions, establishing a dataset containing 1,000 typical scenarios such as grid frequency fluctuations, load changes, and thermal storage temperature changes. Each scenario contains 24 hours of continuous operation data, with a data sampling interval of 1 second, forming a training set with a total of 86,400,000 sample points. The specific steps for training the thermal energy storage grid-connected optimized control model include using the Adam optimizer for parameter updates, setting the learning rate to 0.001, the batch size to 256, the number of training rounds to 500, using cross-validation to evaluate model performance, and using an early stopping mechanism to prevent overfitting. The final model achieves an accuracy of 95.6% on the test set.

[0026] The dynamic architecture expansion mechanism based on progressive network growth monitors the convergence of the loss function and the gradient trend during network training. When insufficient network capacity is detected, new neurons and connection layers are automatically added to achieve dynamic expansion of the network structure. The mechanism starts with a simple network containing 50 neurons in a single hidden layer and gradually increases to a final structure of 150 neurons in 3 hidden layers, depending on the complexity of the training data. During network growth, the weights of newly added neurons are initialized using the Xavier method to maintain parameter compatibility with existing networks. The second-order optimization framework based on meta-gradients uses the approximate calculation of the Hessian matrix to obtain curvature information for parameter updates, and adjusts the parameter update direction and step size by calculating the gradient of the gradient. The framework uses the L-BFGS algorithm to approximate the second derivative information, avoiding the huge computational overhead of directly calculating the Hessian matrix, and simultaneously utilizes historical gradient information to construct an approximate curvature model.

[0027] The progressive network growth dynamic architecture expansion mechanism and the meta-gradient-based second-order optimization framework bring significant technological advancements to the thermal energy storage grid-connected control system. The progressive network growth dynamic architecture expansion mechanism enables the control model to dynamically adjust its structure according to the complexity of the actual operating environment, avoiding the insufficient adaptability of traditional fixed network structures when facing varying operating conditions. This mechanism learns basic control laws starting from a simple structure, gradually increasing network complexity as training progresses to handle more refined control requirements. This progressive learning method improves the model's generalization ability and robustness, enabling the control system to maintain good control performance even when facing unknown operating conditions. The meta-gradient-based second-order optimization framework improves the accuracy and efficiency of parameter updates by introducing curvature information. Compared to traditional first-order optimization methods, this framework can more accurately estimate the parameter update direction, avoiding the problem of getting trapped in local optima. The meta-gradient-based second-order optimization framework considers the curvature characteristics of the loss function, using a larger step size for rapid convergence in flat regions and a smaller step size to ensure stability in steep regions. This adaptive step size adjustment mechanism significantly improves model training efficiency. The synergistic effect of the two mechanisms enables the thermal energy storage grid-connected control system to have self-learning and self-optimization capabilities, improving control accuracy by 23% and response speed by 35% compared to traditional methods. At the same time, it reduces the complexity of system debugging and maintenance, laying a solid technical foundation for the widespread application of thermal energy storage technology in power systems.

[0028] Alternatively, the present invention also provides a method implemented by a computer to form a synchronous grid-connected control system for a thermal energy storage system, wherein the computer is provided with a readable storage medium, the readable storage medium storing program instructions, and the program instructions are used to execute the above-described method when the computer is run.

[0029] The specific implementation methods of the above steps are described in detail below.

[0030] The specific implementation of step S01 involves establishing a multi-parameter real-time monitoring system using a distributed sensor network architecture. This step provides accurate basic data support for the entire control system. First, temperature sensor arrays are deployed at different locations along the thermal energy storage medium. These sensors employ thermocouples or infrared temperature measurement principles, achieving a temperature measurement accuracy of ±0.1℃ and a response time of less than 1 second. Through multi-point deployment, a three-dimensional reconstruction of the temperature field is achieved, obtaining the temperature distribution characteristics of the thermal energy storage medium. Next, a thermal energy storage capacity detection device is installed. Based on the heat balance equation, the current stored heat is calculated. Combined with temperature distribution data and the specific heat capacity parameters of the medium, the system's thermal energy storage capacity status is estimated in real time. Then, a power measurement module is deployed, using current transformers and voltage transformers to acquire real-time power information. The measurement accuracy is no less than 0.5%, and the sampling frequency reaches above 10kHz, ensuring the ability to capture instantaneous power changes. Simultaneously, grid parameter monitoring equipment is configured. The grid frequency is detected using a phase-locked loop principle, achieving a frequency measurement accuracy of ±0.01Hz, a phase angle measurement accuracy of ±0.1 degrees, and a voltage amplitude measurement accuracy of ±0.2%. Finally, a harmonic analysis module was established, which uses the fast Fourier transform algorithm to perform spectrum analysis on the grid voltage and current, and can detect the content of each harmonic up to the 50th order, providing a data basis for power quality assessment.

[0031] The specific implementation of step S02 involves constructing an adaptive phase-locked loop (PLL) controller based on a second-order generalized integrator structure. The purpose of this step is to achieve precise synchronization between the thermal energy storage system and the power grid. First, a second-order generalized integrator is used as the core filtering unit. This integrator has excellent frequency selection characteristics, effectively suppressing grid harmonic interference and extracting the fundamental component for phase detection. Next, a feedforward compensation algorithm is implemented. By performing differential operations on the grid voltage signal, the phase change trend is predicted. When the grid frequency undergoes a step change, the feedforward compensation can adjust the control quantity of the PLL in advance, reducing transient errors in phase tracking. Then, a fuzzy control algorithm is used to dynamically adjust the PLL parameters. Frequency deviation and phase error are used as input variables for the fuzzy controller, and the output is the adjustment coefficients of the proportional gain and integral gain. The fuzzy rule base contains 25 control rules, covering various operating conditions. The membership function of the fuzzy controller adopts a trigonometric function. The domain of the input variables is within the range of frequency deviation ±2Hz and phase error ±10 degrees, and the output adjustment coefficient ranges from 0.5 to 2.0 times. When the frequency deviation is greater than 1 Hz or the phase error is greater than 5 degrees, the controller will increase the proportional gain to improve the response speed. When the system is close to the synchronization state, the integral gain will be increased to reduce the steady-state error.

[0032] The specific implementation of step S03 is to establish a power prediction model based on the heat transfer differential equation of the thermal storage medium. This step provides forward-looking guidance for power control. First, a mathematical model of heat transfer in the thermal storage medium is established. Based on Fourier's heat transfer law and the principle of energy conservation, the coupling effects of conduction, convection, and radiation are considered to construct a three-dimensional unsteady-state heat transfer differential equation system. Next, the finite difference method or finite element method is used to numerically solve the heat transfer differential equations. The thermal storage medium is divided into several control volumes, and an energy balance equation is established within each control volume. The spatiotemporal distribution of the temperature field is obtained through iterative calculation. Then, according to the working principle of the thermal energy storage system, the product of temperature difference and mass flow rate is used as the main variable for power prediction. A correction coefficient is introduced to consider the influence of actual factors such as heat exchange efficiency and pipeline losses. The input parameters of the prediction model include the current temperature distribution, thermal storage capacity, ambient temperature, and expected load demand. The output is a power prediction curve for the next 15 minutes at 1-minute intervals. The model uses a sliding window mechanism to update the prediction results every 30 seconds. The prediction accuracy is verified by historical data, and the average absolute error is controlled within 5%.

[0033] The specific implementation of step S04 involves designing a sliding mode variable structure power controller based on an exponential reaching law. The purpose of this step is to eliminate power output fluctuations and achieve precise power control. First, a sliding surface function is defined, using the power error and its integral as state variables to construct a linear sliding surface. The parameters of the sliding surface are designed based on the dynamic characteristics of the system to ensure good stability. Next, an exponential reaching law is used to design the control law. The reaching speed is adaptively adjusted according to the magnitude of the power error. When the power error is large, a large reaching speed is used to quickly approach the sliding surface; when the error is small, the reaching speed is reduced to decrease chattering. Then, power smoothing filtering is implemented. A moving average algorithm is used for initial filtering of the power command, with a window length of 10 sampling points. Kalman filtering is then used to further eliminate noise interference. The state variables of the Kalman filter are the power value and its rate of change, and the observation noise variance is set to the square of the measurement accuracy. The switching function of the sliding mode controller uses a continuous function instead of a sign function, and the boundary layer thickness is set to 1% of the rated power, effectively reducing high-frequency chattering of the control signal. The power fluctuation amplitude is controlled within 5% of the rated power through real-time monitoring. When the fluctuation exceeds this threshold, the controller will increase the control gain, and the response time will not exceed 100 milliseconds.

[0034] The specific implementation of step S05 involves constructing a thermal energy storage grid-connected optimization control model based on progressive network growth and meta-gradient optimization. The purpose of this step is to achieve multi-objective coordinated optimization control. First, a multi-layer feedforward neural network architecture is established. The network includes an input layer, three hidden layers, and an output layer. The input layer receives 14 parameters, including grid frequency, phase angle, voltage amplitude, power factor, harmonic content, thermal storage temperature, and thermal storage capacity. The number of neurons in each hidden layer is 50, 60, and 40, respectively. The output layer contains 5 neurons corresponding to different control commands. The hidden layers use a modified linear unit activation function, and the output layer uses a sigmoid activation function to limit the output to between 0 and 1. Next, a progressive network growth mechanism is implemented. The initial network starts with a single hidden layer of 50 neurons. By monitoring the convergence speed of the loss function and the change in the gradient norm during training, new neurons and connection layers are automatically added when insufficient network capacity is detected. The weights of the newly added neurons are initialized using the Xavier method. Then, a second-order optimization framework based on meta-gradients is adopted. The Hessian matrix information is approximated using the L-BFGS algorithm, and a curvature approximation model is constructed using historical gradient information. A larger learning rate is used in flat regions, and a smaller learning rate is used in steep regions, adaptively adjusting the direction and step size of parameter updates. Batch normalization is applied before each hidden layer to accelerate training convergence and improve model stability.

[0035] The specific implementation of step S06 involves establishing a power factor compensation control strategy. This step aims to improve energy utilization efficiency and power quality. First, the system's power factor is monitored in real time. When the power factor falls below the 0.9 threshold, a power factor correction program is initiated, determining the required compensation capacity by calculating reactive power demand. Next, the switching status of the capacitor banks is controlled. The capacitor banks are configured in a tiered manner, including large-capacity and small-capacity capacitors, enabling refined reactive power compensation. Thyristors or contactors are used for switching, with a response time of less than 20 milliseconds. Then, the compensation capacity is dynamically adjusted according to load changes. A fuzzy control algorithm is used to determine the switching strategy based on the power factor deviation and rate of change, avoiding grid impact caused by frequent switching. The compensation capacitor capacity is configured in a 1:2:4 ratio, enabling seven different compensation combinations with a compensation accuracy of ±0.02. When the power factor is between 0.9 and 1.0, the system maintains the existing configuration without adjustment, avoiding power factor reduction due to overcompensation. The compensation controller also has an overvoltage protection function, which automatically disconnects the corresponding capacitor bank when the capacitor terminal voltage exceeds 110% of the rated voltage.

[0036] The specific implementation of step S07 involves implementing comprehensive power quality management measures. The purpose of this step is to eliminate harmonic interference and maintain power grid quality. First, the harmonic content level is detected. When the total harmonic distortion rate exceeds 5%, a multi-stage active power filter system is activated. This system includes a fundamental frequency filter unit, a low-order harmonic filter unit, and a high-order harmonic filter unit, each targeting harmonics in different frequency ranges. The fundamental frequency filter unit uses a control algorithm based on instantaneous reactive power theory to calculate harmonic current commands in real time. The low-order harmonic filter unit focuses on suppressing the 3rd, 5th, and 7th harmonics, while the high-order harmonic filter unit handles high-frequency harmonics above the 11th order. Next, when the harmonic content is within the range of 0 to 5%, a single-stage active power filter scheme is adopted. Activating only the fundamental frequency filter unit is sufficient to meet the harmonic management requirements, reducing system complexity and operating costs. Then, a three-phase imbalance compensation algorithm is implemented. The symmetrical component method is used to decompose the three-phase current into positive-sequence, negative-sequence, and zero-sequence components. The magnitude and phase of the negative-sequence current are calculated, and the inverter outputs corresponding compensation current to balance the three-phase load. The imbalance is calculated based on the ratio of negative-sequence current to positive-sequence current. When the imbalance exceeds 2%, the compensation function is activated to control the imbalance within 1%. The switching frequency of the active filter is set to 10kHz, and the DC-side voltage is maintained at 1.2 times the rated voltage to ensure sufficient control margin.

[0037] The specific implementation of step S08 involves establishing a system operation status assessment mechanism and a backup control strategy. This step aims to ensure the safe and stable operation of the system under abnormal conditions. First, a comprehensive evaluation index system is established. Synchronization accuracy error is calculated based on the ratio of the difference between the grid phase angle and the system phase angle. When the error reaches or exceeds 0.1%, it is considered a synchronization anomaly. The power fluctuation coefficient is calculated using the ratio of the standard deviation to the mean. The fluctuation coefficient is calculated by continuously sampling 100 data points. When the fluctuation coefficient reaches or exceeds 5%, it is considered power instability. Power quality indicators include multiple sub-indicators such as harmonic content, power factor, and voltage deviation. A weighted average method is used to calculate the comprehensive score, with the weights allocated as follows: harmonic content 40%, power factor 35%, and voltage deviation 25%. When the comprehensive score is below 90 points, the power quality is considered unqualified. Then, when any indicator exceeds the threshold, the system automatically switches to the backup control strategy. The power reduction operation mode adjusts the output power to 50% of the rated power. This power reduction is achieved by reducing the flow rate of the heat storage medium and adjusting the effective heat exchange area of ​​the heat exchanger. This mode is suitable for operating conditions where the grid frequency fluctuates significantly but remains within the allowable range of 49.5Hz to 50.5Hz. Next, the off-grid operation mode disconnects the electrical connection with the grid and uses islanding detection technology to continuously monitor the grid status. Detection methods include frequency offset detection and phase change detection. When the grid frequency returns to the normal range and remains stable for more than 5 minutes, the system automatically initiates the reconnection procedure. First, synchronization detection is performed to confirm that the frequency, phase, and voltage amplitude all meet the grid connection conditions before closing the grid connection switch.

[0038] Further explanation is needed regarding the thermal energy storage grid-connected optimization control model, which employs a multi-layer feedforward neural network structure. Specifically, it includes a five-layer architecture: an input layer, three hidden layers, and an output layer. The input layer has 14 neurons, corresponding to parameters such as grid frequency, phase angle, voltage amplitude, power factor, harmonic content, five key temperature values ​​of the thermal storage medium temperature distribution, thermal storage capacity, heat release power, and ambient temperature. Each input parameter is normalized, mapping its value to a uniform range of 0 to 1. The first hidden layer contains 50 neurons, the second 60 neurons, and the third 40 neurons. Each hidden layer uses a modified linear unit as the activation function, and batch normalization, including learnable scaling and offset factors, is applied before each hidden layer. The output layer has 5 neurons, corresponding to synchronization control commands, power control commands, power factor correction commands, harmonic mitigation commands, and system protection commands. A sigmoid activation function is used to ensure that the output value is between 0 and 1. The network has a total of 15,000 parameters, including weight parameters and bias parameters. The weights are initialized using the He initialization method.

[0039] The detailed steps for establishing the training dataset first involve building a basic database through simulation platform and actual system operation data collection, covering operational data under different seasons, load levels, and power grid conditions. The data collection scope includes power grid frequency fluctuations between 49Hz and 51Hz, load variations from 20% to 120% of rated capacity, thermal storage temperature variations from 300℃ to 800℃, and seasonal environmental temperature variations from -20℃ to 45℃, forming 1000 typical operating scenarios. Each scenario contains 24 hours of continuous operational data, with a data sampling interval of 1 second, generating 86,400 data samples per scenario. In the data preprocessing stage, the raw data undergoes quality checks, removing outliers and missing values. Three-standard-deviation methods are used to identify outlier data points, and missing data is filled using linear interpolation or nearest-neighbor interpolation. Next, data standardization is performed, standardizing the input features and output labels with zero mean and unit variance to ensure that the data in each dimension have the same numerical range and distribution characteristics. The dataset was then divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter learning, the validation set was used for hyperparameter tuning and model selection, and the test set was used for final performance evaluation, resulting in a complete training dataset of 86,400,000 sample points.

[0040] Model training utilizes the Adam optimizer for parameter updates, combining the advantages of momentum and adaptive learning rate. The beta1 parameter is set to 0.9, the beta2 parameter to 0.999, and the epsilon parameter to 1e-8. The initial learning rate is 0.001, employing an exponential decay strategy, multiplying the learning rate by 0.95 every 100 epochs. The batch size is set to 256, ensuring both gradient estimation accuracy and computational efficiency. The number of training epochs is set to 500, employing an early stopping mechanism to prevent overfitting; training stops when the validation set loss shows no improvement for 20 consecutive epochs. Cross-validation uses a 5-fold cross-validation method to evaluate the model's generalization performance, calculating accuracy, precision, recall, and F1 score on the validation set after each fold. The loss function uses a weighted combination of mean squared error and cross-entropy loss with a weight ratio of 7:3, considering both numerical accuracy for regression problems and decision accuracy for classification problems. Regularization techniques include L2 weight decay and Dropout. The L2 regularization coefficient is set to 0.0001, and the Dropout probability is set to 0.2, applied between hidden layers. Gradient clipping is used during training to prevent gradient explosion, with a clipping threshold set to 5.0. The model achieves an overall accuracy of 95.6% on the test set, meeting the accuracy requirements for practical engineering applications.

[0041] It should be noted that the key technical ideas of this invention are mainly reflected in four aspects. First, a dynamic architecture expansion mechanism based on progressive network growth. This mechanism breaks through the limitations of traditional fixed network structures. By monitoring the convergence state and gradient changes during training, it automatically adjusts the network complexity, enabling the control model to dynamically optimize its own structure according to the complexity of the actual operating environment. This avoids underfitting caused by insufficient network capacity and overfitting caused by excessive capacity, significantly improving the model's adaptability and generalization performance. Second, a second-order optimization framework based on meta-gradient. This framework improves the accuracy and directionality of parameter updates by introducing curvature information. Compared with traditional first-order gradient optimization methods, it can more accurately estimate the optimal parameter update path. It uses a larger step size in the flat region of the loss function to achieve fast convergence, and a smaller step size in the steep region to ensure training stability, effectively avoiding the problem of getting trapped in local optima, and significantly improving the model training efficiency and final performance. Thirdly, it employs a collaborative mechanism of multi-parameter real-time monitoring and adaptive phase-locked loop (PLL) control. By acquiring millisecond-level system state information through a distributed sensor network, and combining feedforward compensation and fuzzy control algorithms, it achieves rapid response and precise tracking for grid synchronization. Compared to traditional fixed-parameter PLLs, it exhibits stronger robustness and faster dynamic response under grid disturbance conditions. Fourthly, it integrates and optimizes sliding mode variable structure power control with comprehensive power quality management. It effectively suppresses power fluctuations using exponential reaching laws and power smoothing filtering techniques, and combines multi-stage active filtering and power factor compensation to achieve a comprehensive improvement in power quality. Compared to traditional single control strategies, it demonstrates superior control accuracy and stability in complex grid environments.

[0042] The synergistic effect of these four key technological approaches forms the overall technological advantage of the thermal energy storage grid-connected control system. The incremental network growth mechanism provides the system with strong learning capabilities and adaptability; the meta-gradient optimization framework ensures rapid convergence and global optimality of the control strategy; multi-parameter monitoring and adaptive phase-locked loops guarantee precise synchronization with the grid; and sliding mode power control and power quality management achieve stable and reliable power output. The deep integration of these four technologies endows the control system with intelligent characteristics of self-learning, self-optimization, and self-adaptation. It can maintain stable control performance in the face of various uncertainties such as changes in grid parameters, load fluctuations, and changes in thermal energy storage conditions. Compared with traditional control schemes based on fixed control parameters and single algorithms, it significantly improves control accuracy, response speed, robustness, and adaptability, providing reliable technical support for the large-scale application of thermal energy storage technology in complex power system environments.

[0043] It should be noted that this invention also solves the following technical problems: During the grid-connected operation of a thermal energy storage system, due to the nonlinear characteristics of the power electronic converter and the power pulsation caused by temperature fluctuations in the thermal storage medium, harmonic currents are easily injected into the grid, resulting in a decrease in the power factor. Traditional control systems lack measures to address power quality issues, leading to power quality that does not meet standard requirements during grid-connected operation. This invention establishes a hierarchical comprehensive power quality management strategy. When the power factor is below 0.9, power factor compensation control is automatically activated to adjust the switching status of the capacitor bank. When the harmonic content exceeds 5%, a multi-stage active filter system is activated to process the fundamental, low-order, and high-order harmonic components respectively. Simultaneously, an unbalanced compensation algorithm based on the symmetrical component method is used to adjust the three-phase voltage balance in real time, ensuring that the thermal energy storage system can provide high-quality power to the grid under various operating conditions. Furthermore, thermal energy storage systems face multiple risk factors during grid-connected operation, such as grid disturbances, equipment failures, and parameter drift. Existing technologies lack a comprehensive fault identification and emergency response mechanism; once an abnormal situation occurs, the system often has to be passively shut down, failing to guarantee continuous and stable operation. This invention constructs a system operation status evaluation mechanism based on multi-index comprehensive evaluation, calculates synchronization accuracy error, power fluctuation coefficient and power quality index in real time, and immediately triggers corresponding backup control strategies when any index exceeds the normal range, including a power reduction operation mode that reduces the output power to 50% of the rated power and an off-grid operation mode that completely disconnects the grid connection. At the same time, it integrates islanding detection technology to continuously monitor the grid recovery status, realizes safe operation under fault conditions and automatic reconnection to the grid after the fault is eliminated, and significantly improves the reliability and self-healing capability of system operation.

[0044] Specifically, the principle of this invention is as follows: The reason this invention can effectively solve the problems of response lag and insufficient control accuracy in existing technologies lies in establishing a three-layer progressive control architecture integrating adaptive synchronous control, predictive power management, and intelligent optimization decision-making. Through the synergistic effect of multiple technical components, it achieves precise and rapid control of the grid connection process of the thermal energy storage system. At the synchronous control level, the adaptive phase-locked loop controller abandons the traditional fixed-parameter design approach and adopts a fuzzy control algorithm to dynamically adjust the proportional-integral parameters according to the real-time changes in grid frequency deviation and phase error. This enables the system to maintain optimal synchronous tracking performance under different operating conditions. Simultaneously, the integrated feedforward compensation mechanism predicts phase change trends by analyzing the differential signal of the grid voltage, transforming the traditional lag response into predictive control and significantly shortening the system's response time. At the power control level, the power prediction model constructed based on the differential equation of heat transfer in the thermal energy storage medium can accurately calculate the power output capacity for future periods based on the changes in temperature distribution and mass flow rate. This provides an accurate forward-looking reference signal for the sliding mode variable structure power controller, upgrading power regulation from traditional error feedback control to active control based on predictive information, effectively avoiding power output fluctuations and control lag. At the optimization decision-making level, the neural network control model adopts a progressive architecture expansion mechanism, starting from a simple structure and gradually increasing the network complexity to adapt to different control requirements. This avoids the insufficient adaptability of traditional fixed-structure networks when facing complex operating conditions. The second-order optimization framework based on meta-gradients uses the curvature information of the loss function to guide the parameter update direction and step size. Compared with the traditional first-order optimization method, it can converge to the optimal control strategy more quickly and accurately. This multi-level coordinated control mechanism ensures that the thermal energy storage system can achieve rapid response and precise control under various power grid operating conditions, fundamentally solving the technical problems of response lag and insufficient control accuracy.

[0045] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0046] The specific implementation of step S01 involves establishing a multi-parameter real-time monitoring network for the thermal energy storage system, collecting system operating parameters, and performing data fusion processing. Temperature distribution data acquisition utilizes a multi-point temperature sensor array, and the temperature field reconstruction formula is expressed as follows: ; In the formula, Normalized temperature field distribution; For the first Each sensor measures temperature, in °C. For reference temperature, the empirical value is 500℃; This is the weighting coefficient, typically ranging from 0.1 to 1.0; For the first Spatial basis functions of each sensor; This represents the total number of sensors. The real-time calculation formula for thermal storage capacity is as follows: ; In the formula, Normalized thermal storage capacity; For the first The density of the medium in a control volume, in units of ; Specific heat capacity, unit: ; To control volume, the unit is... ; For the first The temperature of a controlled volume is expressed in °C. This refers to the ambient temperature, expressed in °C. To control the total volume.

[0047] The specific implementation of step S02 is to construct an adaptive phase-locked loop controller to achieve grid synchronization tracking. The transfer function of the second-order generalized integrator is expressed as follows: ; In the formula, The transfer function for a second-order generalized integrator; This is the damping coefficient, with a default value of 1.414. The center angular frequency, in units of ; This is the Laplace operator. The formula for estimating the frequency of a phase-locked loop is as follows: ; In the formula, To estimate the angular frequency, the unit is . ; The rated angular frequency is, empirically, [value missing]. ; This is the integral gain, in units of The value is typically between 50 and 200. This is the proportional gain, in units of The value is typically between 10 and 50. Phase error, in units of The formula for adjusting fuzzy control parameters is expressed as follows: ; ; In the formula, These are the initial proportional control parameters, in units of The default value is 20; These are the initial integral control parameters, in units of The default value is 100; and This is the output value of the fuzzy membership function, ranging from 0 to 2.

[0048] The specific implementation of step S03 involves establishing a thermal energy storage power prediction model to predict power output based on the heat transfer characteristics of the thermal storage medium. The power prediction formula is expressed as follows: ; In the formula, For predicted power, the unit is kW; This is a correction factor, typically ranging from 0.85 to 0.95; This is the temperature difference, expressed in °C. For reference temperature difference, the empirical value is 300℃; Mass flow rate, in kg / s; For reference mass flow rate, the unit is kg / s; This represents the rated power, expressed in kW. The differential equation for heat transfer in the thermal storage medium is as follows: ; In the formula, Thermal conductivity, in units of ; Density of the thermal storage medium, in units of ; Specific heat capacity of the heat storage medium, in units of ; For internal heat sources, the unit is... .

[0049] The specific implementation of step S04 involves designing a sliding mode variable structure power controller to eliminate power output fluctuations. The sliding mode surface function is defined as follows: ; In the formula, For normalized sliding surface functions; This is the power error, in kW. This is the parameter for the sliding surface, with a default value of 5, and the unit is... The exponential reaching law control law is expressed as follows: ; In the formula, To normalize the control output; As an equivalent control term, its calculation is based on the inverse dynamics of the system's nominal model, and it is typically taken as a value of ; The reaching-law gain is typically set to a value between 2 and 5. The reaching law exponent is set to 0.5 by default. The power smoothing filter algorithm is described as follows: ; In the formula, This represents the power after normalized filtering. This is the length of the moving average window, which defaults to 10. The sampling interval is expressed in units of 1 / 2. ; This is the Kalman filter gain, typically ranging from 0.1 to 0.5. This is the Kalman filter order, which defaults to 3. These are the filter weight coefficients, typically ranging from 0.2 to 0.8. Noise power, in kW; Historical power values, in kW.

[0050] The specific implementation of step S05 involves constructing a grid-connected optimization control model for thermal energy storage based on progressive network growth and meta-gradient optimization. In the dynamic architecture expansion mechanism of progressive network growth, the network capacity judgment function is expressed as follows: ; In the formula, This is an indicator of network capacity adequacy. For the first The loss function value for each training round is dimensionless. The default interval between training rounds is 10. Let L be the 2-norm of the gradient of the loss function. The network growth trigger condition is stated as follows: ; In the formula, This is a sign of network growth; The capacity threshold is 0.95, based on experience. To ensure convergence accuracy, the calculation method is as follows: ; The convergence threshold is typically set to a value of [value to be filled in]. The initialization statement for the weights of newly added neurons is as follows: ; In the formula, For the first The weight matrix of newly added neurons in the layer; For the first Number of input neurons in the layer; These are random numbers distributed according to a standard normal distribution. In the meta-gradient optimization framework, the approximate calculation of the second-order gradient is expressed as follows: ; In the formula, It is an approximation of the Hessian matrix; This is the search direction vector; For perturbation parameters, the default value is... ; This is the network parameter vector; Let be the gradient vector of the loss function with respect to the parameters. The L-BFGS update formula is expressed as follows: ; In the formula, For the first Approximation of the inverse Hessian matrix in the next iteration; For parameter variation vectors; The gradient change vector; This is the scaling factor; Let be the identity matrix. The adaptive learning rate adjustment is described as follows: ; In the formula, For the first The learning rate for each iteration; The initial learning rate is 0.001 (empirical value). This is the learning rate decay coefficient, which is usually set to 0.1.

[0051] The specific implementation of step S06 is to establish a power factor compensation control strategy to improve power utilization efficiency. The reactive power demand calculation formula is expressed as follows: ; In the formula, To compensate for reactive power, the unit is... ; Active power, measured in kW; The target power factor is empirically set at 0.95. The current power factor; The rated power factor is typically 1.0. The capacitor switching control algorithm is described below: ; In the formula, Total installed capacity, in units of ; This refers to the number of capacitor banks; For the first The switching status of the capacitor bank, with a value of 0 or 1; For the first Capacitor capacitance, in units of ; Rated voltage, unit: ; This is the actual voltage, in units of... .

[0052] The specific implementation of step S07 involves comprehensive power quality management to eliminate harmonic interference and regulate voltage balance. The formula for calculating the total harmonic distortion rate is as follows: ; In the formula, Total harmonic distortion (THD); For the first The effective value of the second harmonic current, in units of ; This is the effective value of the fundamental current, in units of . ; The highest harmonic order for analysis is set to 50 by default. The formula for calculating the three-phase unbalance using the symmetrical component method is as follows: ; In the formula, This refers to the three-phase imbalance. The complex number of negative sequence current, in units of . ; The positive-sequence current is a complex number, with units of . The formula for calculating negative sequence current is as follows: ; In the formula, , , These are the complex numbers of the three-phase currents A, B, and C, respectively, with units of . ; For complex number operators.

[0053] The specific implementation of step S08 involves establishing a system operation status evaluation mechanism and a backup control strategy. The formula for calculating synchronization accuracy error is as follows: ; In the formula, This refers to synchronization accuracy error; The phase angle of the power grid, in units of ; The system phase angle is expressed in units of 1 / 2π. The formula for calculating the power fluctuation coefficient is as follows: ; In the formula, The power fluctuation coefficient; For the first The power at each sampling point is expressed in kW. Average power, in kW; This represents the total number of sampling points, defaulting to 100. The formula for calculating the comprehensive power quality evaluation index is as follows: ; In the formula, Power quality indicators; , , These are the weighting coefficients, with default values ​​of 0.4, 0.35, and 0.25 respectively. The harmonic limit is typically 5%; The unbalance limit is 2%, which is an empirical value.

[0054] It should be explained that the temperature field reconstruction formula is based on spatial interpolation theory and a weighted averaging method. It achieves numerical reconstruction of the continuous temperature field by fusing data from multiple discrete temperature measurement points. Compared to traditional single-point temperature monitoring methods, this approach obtains complete temperature distribution information of the thermal storage medium, providing a reliable data foundation for accurate thermal storage capacity calculation and power prediction. The real-time thermal storage capacity calculation formula is based on the thermodynamic law of conservation of energy. It divides the thermal storage medium into multiple control volumes and calculates the heat storage capacity of each volume. The total thermal storage capacity is obtained by summing these volumes. Compared to traditional average temperature estimation methods, this approach has higher calculation accuracy and can accurately reflect the impact of non-uniform temperature distribution of the thermal storage medium on the thermal storage capacity.

[0055] Second-order generalized integrator transfer function Based on resonator theory, it selectively amplifies signals of a certain frequency and suppresses harmonics. Compared with traditional first-order filters, it has stronger fundamental frequency extraction capability and better anti-interference performance in power grid harmonic environments. Phase-locked loop frequency estimation formula. Employing a proportional-integral (PI) control structure, rapid frequency tracking is achieved through proportional and integral feedback of the phase error. Compared to traditional methods using a fixed frequency, this approach adaptively tracks grid frequency changes, improving the accuracy and stability of grid synchronization. The fuzzy control parameter adjustment formula, based on fuzzy logic reasoning, dynamically adjusts control parameters according to the phase error and its rate of change, exhibiting better adaptability and robustness under grid disturbance conditions compared to fixed-parameter control.

[0056] Power prediction formula Based on the principles of heat transfer and energy conversion, this study takes the temperature difference and mass flow rate of the thermal storage medium as the main influencing factors. Through normalization, it achieves universality under different operating conditions, offering a stronger physical foundation and higher prediction accuracy compared to traditional empirical formulas. The differential equation for heat transfer in the thermal storage medium, based on Fourier's law of heat transfer and the principle of energy conservation, describes the temperature transfer process within the thermal storage medium. Compared to simplified lumped parameter models, it can more accurately reflect the heat transfer characteristics and temperature distribution evolution of the thermal storage medium.

[0057] Sliding surface function Based on sliding mode control theory, the power error and its integral linear combination are used as the switching surface, ensuring that the system state reaches the sliding surface and moves along it within a finite time. Compared to traditional PID control, this method exhibits stronger robustness and faster response speed. (Exponential reaching law control law) A variable exponential reaching law design is adopted, which adjusts the reaching speed through an exponential function. A large reaching speed is used when far from the sliding surface, and the reaching speed is reduced when approaching the sliding surface. Compared with a constant reaching law, this effectively reduces chattering and improves control accuracy. The power smoothing filtering algorithm combines the advantages of moving average filtering and Kalman filtering, which can eliminate random noise while maintaining the dynamic characteristics of the signal. Compared with a single filtering method, it achieves a better balance between noise suppression and signal tracking.

[0058] Network capacity assessment function Based on a comprehensive evaluation of the convergence trend of the loss function and the magnitude of gradient changes, the training convergence speed is assessed through the loss ratio term, and the parameter update activity is assessed through the gradient norm ratio term. Compared to a fixed network structure, this allows for dynamic sensing of the network's learning capacity saturation, enabling adaptive expansion of the network capacity. A new neuron weight initialization formula has been added. An improved version based on Xavier initialization theory uses variance scaling by taking the inverse square root of the number of input neurons to ensure that the activation values ​​and gradients of newly added neurons are within a reasonable range. Compared with random initialization, this method can better maintain the stability and convergence of network training.

[0059] Second-order gradient approximation formula The finite difference method is used to approximate the product of the Hessian matrix and the vector, avoiding the huge computational overhead of directly calculating the Hessian matrix. Compared with first-order optimization methods, it can obtain curvature information to guide the parameter update direction. L-BFGS update formula. Based on the quasi-Newton method, the inverse Hessian matrix approximation is recursively updated using historical gradient information and parameter change information, where the parameter change vector... and gradient change vector The scaling factor forms the basis of curvature information. This ensures the numerical stability of the update, exhibiting lower computational complexity and better numerical stability compared to directly calculating the second derivative. Adaptive learning rate adjustment formula. The learning rate is dynamically adjusted based on curvature information. The step size is adjusted according to the curvature characteristics of the current parameter space through an exponential decay function. A larger learning rate is used in flat regions to accelerate convergence, while a smaller learning rate is used in steep regions to ensure stability. Compared with a fixed learning rate strategy, this can significantly improve optimization efficiency and final performance.

[0060] Reactive power demand calculation formula Based on the power triangle theory, the required reactive power compensation is calculated by the difference between the target power factor and the current power factor, which has higher calculation accuracy and a stronger theoretical foundation compared to traditional empirical estimation methods. The capacitor switching control algorithm considers the impact of voltage changes on the reactive power output of the capacitor and compensates for it using the square of the voltage term. Compared to the fixed switching method, it can maintain a stable reactive power compensation effect under voltage fluctuations.

[0061] Total Harmonic Distortion Rate Calculation Formula Based on Fourier analysis theory, the degree of harmonic pollution is quantified by the ratio of the root mean square value of each harmonic component to the fundamental component. Compared with traditional single-harmonic analysis, this method can comprehensively assess the power quality. The formula for calculating three-phase unbalance using the symmetrical component method is also provided. Based on the theory of symmetrical components, the three-phase unbalanced system is decomposed into positive-sequence, negative-sequence, and zero-sequence components. The degree of imbalance is quantified by the ratio of negative-sequence to positive-sequence current. Compared with the traditional amplitude difference method, it has a stronger theoretical basis and higher analytical accuracy.

[0062] Synchronization accuracy error calculation formula Using the relative error of the phase angle difference to evaluate synchronization performance offers better versatility and comparability compared to absolute error. (Power fluctuation coefficient calculation formula) Using the concept of coefficient of variation, power stability is quantified by the ratio of standard deviation to mean, exhibiting better normalization characteristics compared to absolute fluctuation amplitude. The comprehensive power quality evaluation index uses a weighted average method to combine multiple power quality indicators into a single evaluation value. Compared to evaluation by a single indicator, it can more comprehensively reflect the power quality status of the system, providing a quantitative basis for operational decisions.

[0063] It should be noted that the variables involved in this invention are explained in detail in Tables 1 and 2 below.

[0064] Table 1. Variable Explanation Table (Part 1)

[0065] Table 2. Variable Explanation Table (Part Two)

[0066] To better understand and implement this invention, a specific application scenario is provided below as Example 2: An industrial park is equipped with a molten salt thermal energy storage system. The thermal storage medium is a sodium nitrate-potassium nitrate mixed salt, with an operating temperature range of 290–565℃ and a total thermal storage capacity of 450 MWh. The park's power grid environment is complex, containing various electrical loads, with a grid frequency fluctuation range of 49.8–50.2 Hz, a voltage level of 35 kV, and a power factor requirement of no less than 0.92. The main challenge faced by the technical team is to ensure the stable grid-connected operation of the thermal energy storage system in this complex power grid environment, while simultaneously meeting power quality standards and power control accuracy requirements.

[0067] The technical team first established a multi-parameter real-time monitoring network covering the entire thermal energy storage system. 128 temperature sensors were deployed at different heights of the storage tank, forming a three-dimensional temperature distribution monitoring network with a data sampling frequency of 10Hz, enabling real-time acquisition of the molten salt temperature distribution. Thermal storage capacity monitoring is achieved by calculating the integral value of the storage medium's temperature; combined with data from mass flow sensors, the system can accurately calculate the current thermal storage state. Heat release power monitoring is achieved using a high-precision power transmitter, with a measurement accuracy of 0.2%. Grid parameter monitoring includes frequency, phase angle, voltage amplitude, power factor, and harmonic content, using a power quality analyzer for millisecond-level data acquisition. The monitoring system transmits data via a fiber optic communication network, ensuring data real-time performance and reliability.

[0068] In the design of the adaptive phase-locked loop (PLL) controller, the technical team adopted a second-order generalized integrator structure as the core architecture. This controller can achieve rapid synchronization tracking within a grid frequency fluctuation range of ±0.3Hz, with a response time of less than 20ms. The feedforward compensation algorithm predicts the phase change trend by analyzing the differential signal of the grid voltage, effectively reducing the lag time of phase tracking. The fuzzy control algorithm dynamically adjusts the proportional-integral parameters according to the magnitude of the frequency deviation and phase error, controlling the frequency deviation within ±0.05Hz and the phase error within ±2°. The synchronization control signal output by the PLL is a standard sine wave with stable amplitude, providing a precise synchronization reference for subsequent power control.

[0069] The thermal energy storage power prediction model is built based on the heat transfer characteristics of the storage medium and historical operating data. The technical team collected six consecutive months of operating data and established a prediction model incorporating multiple variables such as temperature distribution, mass flow rate, and ambient temperature. The model employs a numerical solution of the heat transfer differential equation of the storage medium, combined with machine learning algorithms to improve prediction accuracy. In actual operation, when the average temperature of the storage medium is 485℃ and the mass flow rate is 125kg / s, the predicted power output is 12.3MW, with an error of less than 3% compared to the actual output power. The prediction model can provide power output curves for the next two hours, offering forward-looking guidance for power control strategies.

[0070] like Figure 2 As shown, the sliding mode variable structure power controller is designed using an exponential reaching law method. The controller dynamically adjusts control parameters to eliminate power output fluctuations based on the deviation between predicted power data and actual heat release power data. The adaptive reaching speed adjustment mechanism designed by the technical team can automatically adjust the reaching speed according to the magnitude of the power error; a fast reaching mode is used when the power error is greater than 5%, and a fine adjustment mode is used when the error is less than 2%. Power smoothing filtering technology, combined with a moving average algorithm and Kalman filtering, can effectively suppress power fluctuations caused by uneven flow of the heat storage medium. In actual operation, the power fluctuation amplitude is controlled within 3.2% of the rated power, and the response time is 85ms, meeting the grid's technical requirements for distributed power source integration.

[0071] The thermal energy storage grid-connected optimization control model adopts a multi-layer feedforward neural network structure, comprising one input layer, three hidden layers, and one output layer. The input layer receives 12 grid parameters and 8 thermal energy storage system state parameters. The hidden layers contain 64, 128, and 64 neurons respectively. The output layer outputs 6 control commands through a sigmoid activation function. The network has a total of 15,247 parameters, and batch normalization is used to improve training stability. The training dataset established by the technical team includes 1,000 typical operating scenarios, covering various conditions such as grid frequency fluctuations, load abrupt changes, and thermal energy storage temperature variations. Each scenario contains 24 hours of continuous operating data, with a data sampling interval of 1 second, forming a complete training set of 86,400,000 sample points. Table 3 shows the typical values ​​of system operating parameters under different operating conditions.

[0072] Table 3 System operating parameters under typical operating conditions

[0073] The model was trained using the Adam optimizer for parameter updates, with an initial learning rate of 0.001 and a batch size of 256. After 500 training iterations, the model achieved an accuracy of 95.8% on the test set, and the loss function converged to 0.0023. A progressive network growth mechanism started with a simple structure containing 50 neurons in a single hidden layer, gradually increasing to a final 3-layer hidden layer structure based on the complexity of the training data. A second-order optimization framework based on meta-gradients utilized the L-BFGS algorithm to approximate the second derivative information, significantly improving the accuracy and convergence speed of parameter updates by introducing curvature information.

[0074] The power factor compensation control strategy was designed to address the complex load characteristics of the industrial park. The technical team configured six 400... The capacitor banks utilize an intelligent switching controller for automatic switching. When the power factor is detected to be below 0.9, the system automatically calculates the required compensation capacity and switches on the corresponding capacitor banks. The switching action time is less than 50ms, enabling rapid response to load changes. In actual operation, when the power factor drops to 0.87 due to the start-up of a large motor in the park, the compensation system completes the switching action within 35ms, raising the power factor to 0.94. The capacitor banks employ a group switching method to avoid overcompensation.

[0075] The comprehensive power quality management system includes multi-stage active filtering units. When the detected harmonic content exceeds 5%, the system automatically activates a three-stage filtering process: the fundamental frequency filter unit handles the fundamental harmonics, the low-order harmonic filter unit filters the 3rd, 5th, and 7th harmonics, and the high-order harmonic filter unit handles harmonics above the 11th order. Under conditions where numerous frequency converters are used in the industrial park, the 5th harmonic content reaches 6.8%, and the filtering system, after activation, controls the total harmonic distortion rate to within 4.2%. The imbalance compensation algorithm, based on the symmetrical component method, calculates the negative sequence current in real time, and through compensation control, keeps the three-phase voltage imbalance to within 2%.

[0076] like Figure 3 As shown, the system operation status evaluation mechanism establishes a complete performance monitoring system. The comprehensive evaluation algorithm designed by the technical team calculates the system operation score based on synchronization control signals, stable power control signals, power factor correction control signals, and power quality control signals. The synchronization accuracy error is calculated by comparing the grid phase angle and the system phase angle in real time, with an error of 0.05% during normal operation. The power fluctuation coefficient is calculated by statistically analyzing the standard deviation of power output, with a coefficient of 2.8% during stable operation. The power quality index comprehensively considers factors such as harmonic content, voltage deviation, and frequency stability, and the index is 94 points under normal operating conditions.

[0077] The backup control strategy includes two operating modes. The reduced-power operating mode is suitable for situations where grid frequency fluctuations are significant but still within acceptable limits. The system adjusts its output power to 50% of the rated power by reducing the flow rate of the heat storage medium from 125 kg / s to 65 kg / s, while simultaneously adjusting the effective heat exchange area of ​​the heat exchanger to achieve power control. The off-grid operating mode is activated during grid faults. The system disconnects its electrical connection to the grid and independently supplies power to critical loads in the industrial park. Islanding detection technology combines active frequency disturbance and passive detection methods, with a detection time of less than 100 ms. When the grid returns to normal, the system confirms grid parameter stability through synchronous detection and automatically executes the reconnection procedure.

[0078] During six consecutive months of operation, the thermal energy storage system demonstrated excellent grid connection performance. The system maintained stable grid connection for 99.2% of the time, achieving design accuracy in power output, and all power quality indicators met national standards. In handling grid disturbances, the system successfully responded to 34 grid frequency fluctuation events and 8 voltage dip events, achieving rapid recovery and stable operation in all cases. The temperature control accuracy of the thermal storage medium reached ±3℃, the average power response time was 78ms, and the synchronization accuracy error remained stable within 0.04%.

[0079] This invention represents a significant technological advancement over traditional thermal energy storage grid-connected control methods. Traditional control methods typically employ fixed-parameter controller designs, which struggle to adapt to complex and ever-changing grid environments and the nonlinear characteristics of thermal energy storage systems. This invention, by introducing an adaptive phase-locked loop controller and fuzzy control algorithms, achieves dynamic adjustment of control parameters, significantly improving the system's adaptability to grid disturbances. The application of a sliding mode variable structure power controller overcomes the limitations of traditional PI controllers in handling nonlinear systems, achieving more precise power point tracking control through a reaching law method. The neural network-based optimization control model, compared to traditional empirical formula control methods, can handle more complex multi-parameter coupling relationships, achieving multi-objective coordinated optimization. The application of an incremental network growth mechanism and a meta-gradient optimization framework endows the control system with self-learning and self-optimization capabilities, enabling dynamic adjustment of control strategies based on the actual operating environment, avoiding the insufficient adaptability of traditional fixed control strategies when facing new operating conditions. The integrated application of comprehensive power quality management technologies, compared to traditional single-filter methods, can more comprehensively address harmonic pollution and voltage imbalance issues, improving the overall power quality level of the system.

[0080] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for synchronizing a thermal energy storage system with a power grid, characterized in that, Establish a multi-parameter real-time monitoring network for the thermal energy storage system, collecting operating parameters such as thermal energy storage medium temperature distribution, thermal storage capacity, heat release power, grid frequency, phase angle, voltage amplitude, power factor, and harmonic content. Construct an adaptive phase-locked loop (PLL) controller, combining a feedforward compensation algorithm to detect changes in grid frequency and phase angle data, and use a fuzzy control algorithm to dynamically adjust the PLL parameters based on the grid frequency and phase angle data. Establish a thermal energy storage power prediction model to predict the power output curve for future periods. Design a sliding mode variable structure power controller to eliminate power output fluctuations based on predicted power and heat release power data. Employ a thermal energy storage grid-connected optimization control model to output optimized control commands. Establish a power factor compensation control strategy; when the power factor is below 0.9, improve power utilization efficiency by adjusting the power factor correction circuit parameters. Implement comprehensive power quality management; when the harmonic content exceeds 5%, activate a multi-stage active filter system. Establish a system operation status assessment mechanism and a backup control strategy. When the synchronization accuracy error is ≥0.1%, the power fluctuation coefficient is ≥5%, or the power quality index is <90 points, the system will automatically switch to the backup control strategy.

2. The method of claim 1, wherein, The steps for establishing a multi-parameter real-time monitoring network for the thermal energy storage system specifically involve using a distributed sensor network to achieve millisecond-level data updates, thereby obtaining data on the thermal energy storage medium temperature distribution, thermal storage capacity, heat release power, grid frequency, phase angle, voltage amplitude, power factor, and harmonic content.

3. The method of claim 2, wherein, The adaptive phase-locked loop controller specifically adopts a second-order generalized integrator structure to achieve fast synchronous tracking. The feedforward compensation algorithm predicts the phase change trend through the differential signal of the grid voltage. The fuzzy control algorithm dynamically adjusts the proportional-integral parameters according to the frequency deviation and phase error.

4. The method of claim 3, wherein, The thermal energy storage power prediction model is specifically based on the variation law of temperature distribution data and thermal storage capacity data of the thermal energy storage medium and the thermodynamic characteristics of the thermal storage medium. It is based on the heat transfer differential equation of the thermal storage medium, where the predicted power is the product of the correction coefficient, the temperature difference, and the mass flow rate, multiplied by the rated power.

5. The method of claim 4, wherein, The sliding mode variable structure power controller specifically employs a reaching law method and an exponential reaching law. The reaching speed is adaptively adjusted according to the power error amplitude. Combined with power smoothing filtering technology, the power fluctuation amplitude is controlled within a set threshold range. The power smoothing filtering is implemented by combining a moving average algorithm with a Kalman filter.

6. The method of claim 5, wherein, The aforementioned thermal energy storage grid-connected optimization control model is specifically a multi-layer feedforward neural network structure, including an input layer, three hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer outputs control commands through the Sigmoid function. The inputs are grid frequency data, phase angle data, and stable power control signals, and the output is optimized control commands.

7. The method of claim 6, wherein, The power factor compensation control strategy specifically maintains the existing circuit configuration without change when the power factor data is in the range [0.9, 1.0]. The power factor correction circuit parameters include the capacitor bank capacity and the timing of the switching operation. The power factor correction control signal is used to adjust the switching state of the capacitor bank.

8. The method of claim 7, wherein, The comprehensive power quality management specifically involves adopting a single-stage active filtering scheme if the harmonic content data is controlled within the range (0, 5%). This scheme uses an imbalance compensation algorithm to adjust the three-phase voltage balance and eliminate harmonic interference. The multi-stage active filtering system includes a fundamental wave filtering unit, a low-order harmonic filtering unit, and a high-order harmonic filtering unit. The single-stage active filtering scheme only uses the fundamental wave filtering unit.

9. The method of claim 8, wherein, The system operation status evaluation mechanism specifically evaluates the system based on the synchronization control signal, stable power control signal, power factor correction control signal, and power quality control signal, and calculates the synchronization accuracy error, power fluctuation coefficient, and power quality index. The imbalance compensation algorithm calculates the negative sequence current based on the symmetrical component method and performs real-time compensation.

10. The method according to claim 9, characterized in that, The off-grid operation mode disconnects the electrical connection with the power grid, and the thermal energy storage system independently supplies power to the local load. Islanding detection technology is used to monitor the grid status, and the system automatically reconnects to the grid when the grid returns to normal.

Images