Flywheel energy storage array control method for load leveling

Abstract

The application relates to the technical field of power system stability control, and discloses a flywheel energy storage array control method for load suppression, which comprises the following steps: acquiring real-time data such as bus voltage deviation, flywheel rotating speed threshold value, topological structure and target suppression indexes, and determining the indexes by extracting load fluctuation main frequency components through a spectrum analysis model; generating a characteristic vector by wavelet transformation on the data, inputting the characteristic vector into an adjustment parameter prediction model based on a dynamic coordination algorithm, calculating a dynamic correction coefficient in combination with flywheel mechanical inertia parameters, and optimizing control parameters such as torque compensation and switching frequency; adjusting the array operation state according to the parameters, setting fault emergency treatment steps, and realizing topological reconstruction and torque smooth transition. The method improves load suppression precision and system reliability through space-time graph convolution network modeling, mechanical characteristic dynamic correction and full-closed-loop control, and is suitable for power grid stability control scenes.

Description

Technical Field

[0001] This invention relates to the field of power system stability control technology, specifically to a flywheel energy storage array control method for load smoothing. Background Technology

[0002] With the continuous increase in the proportion of new energy power generation and the widespread application of nonlinear loads in industrial and residential sectors, power grid load fluctuations exhibit significant characteristics of complex frequency components and drastic amplitude changes, posing a severe challenge to the frequency stability, voltage quality, and energy supply-demand balance of the power system. Traditional battery energy storage technology is limited by inherent defects such as short cycle life, high temperature sensitivity, and energy density bottlenecks. In high-frequency charge-discharge load smoothing scenarios, it is prone to problems such as accelerated performance degradation and increased operation and maintenance costs, making it difficult to meet the power grid's demand for rapid response and long-term reliable energy storage. Flywheel energy storage, as a technology based on mechanical kinetic energy storage, has become an ideal choice for high-frequency load smoothing scenarios due to its advantages such as millisecond-level response speed, million-cycle life, and stable operation across the entire temperature range. Its energy storage array, formed by the collaborative work of multiple flywheel units, can further improve the system's power regulation capability and control accuracy.

[0003] Currently, flywheel energy storage arrays still face a series of key technical challenges in grid load stabilization applications, requiring breakthroughs in several areas: insufficient timeliness and accuracy of multi-unit collaborative control strategies. Traditional PID control or fuzzy control methods based on fixed control parameters struggle to dynamically track the time-frequency characteristics of load fluctuations. When the load simultaneously contains fundamental and multiple harmonic components, the torque compensation commands and switching frequencies of each flywheel unit cannot be coordinated in real time, easily leading to power imbalances within the energy storage array and even exacerbating risks such as increased bus voltage fluctuations and system resonance. For example, when the main frequency of load fluctuations drifts, fixed-parameter controllers cannot adjust their control strategies in a timely manner, resulting in a significant decrease in stabilization effectiveness.

[0004] Secondly, the dynamic adaptability of energy storage array topologies urgently needs improvement. Existing control methods typically employ a preset fixed topology, failing to fully consider the impact of real-time flywheel unit operating states (such as exceeding speed thresholds or differences in mechanical inertia) on coordinated control. When individual flywheel units need to be taken out of operation due to mechanical failures, abnormal speeds, or other reasons, the lack of a rapid and effective topology reconfiguration mechanism may lead to excessive instantaneous power shortages in the system, causing sudden changes in grid frequency. Furthermore, the control parameter switching process when backup units are connected is prone to generating torque surges, further affecting grid stability.

[0005] The dynamic decoupling problem between the mechanical characteristics and control parameters of the flywheel unit is prominent. Mechanical parameters such as the equivalent mass distribution of the flywheel rotor, bearing lubrication status, and material stiffness coefficient change slowly over time, but traditional control models do not establish a real-time mapping relationship between mechanical parameters and control parameters (such as torque compensation and switching frequency). For example, rotor mass eccentricity can lead to increased vibration amplitude and changes in mechanical inertia distribution, but existing control methods cannot dynamically correct control parameters based on such changes. Long-term operation may result in decreased control accuracy or even system instability.

[0006] Furthermore, the fault detection and emergency response mechanisms exhibit significant lag. Existing systems primarily rely on threshold judgments to detect faults such as speed deviations and abnormal vibrations in flywheel units, lacking fault prediction capabilities based on multi-source data fusion. When a fault occurs, manual intervention is often required for unit isolation and topology adjustments, leading to lengthy fault handling times and potentially triggering chain reactions that expand the scope of the accident's impact. Simultaneously, sudden changes in total output torque during traditional fault handling processes can easily cause secondary shocks to the power grid, exacerbating system risks.

[0007] At the load characteristic analysis level, traditional spectrum analysis methods use filter banks with fixed center frequencies, which are difficult to adapt to the dynamic changes in the dominant frequency of load fluctuations, resulting in large errors in the extraction of target smoothing indicators. For example, when the amplitude of a certain harmonic component in the load suddenly increases and becomes the dominant frequency, the fixed filter cannot capture this change in time, leading to deviations in control parameter prediction. In terms of model training, existing regulation parameter prediction models are mostly built based on a single algorithm, lacking the ability to jointly model the spatial topology relationship and temporal inertial response characteristics of the flywheel array. The models have insufficient generalization ability under complex operating conditions and are prone to overfitting or underfitting problems. Summary of the Invention

[0008] The purpose of this invention is to provide a flywheel energy storage array control method for load smoothing, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a flywheel energy storage array control method for load smoothing, the method comprising:

[0010] Acquire real-time data and target smoothing indicators related to grid load fluctuations. The real-time data includes bus voltage deviation, flywheel speed threshold, and energy storage array topology.

[0011] The target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology are input into a regulation parameter prediction model constructed based on a dynamic coordination algorithm. The regulation parameter prediction model outputs a set of predicted coordination control parameters. The coordination control parameters include flywheel torque compensation and energy storage unit switching frequency.

[0012] The operating status of the flywheel energy storage array is adjusted according to the aforementioned coordinated control parameters to achieve grid load smoothing control.

[0013] Preferably, real-time data related to power grid load fluctuations and target smoothing indicators are acquired, including:

[0014] The system receives bus voltage deviation, flywheel speed threshold, and energy storage array topology data transmitted by the monitoring equipment through a power communication network. The target smoothing index is determined as follows: power grid load fluctuation characteristic data is input to the control terminal; a spectrum analysis model is used to perform frequency domain decomposition on the characteristic data to extract the main frequency component of the load fluctuation; the target smoothing index is obtained by matching the main frequency component with a preset spectrum template. The system also receives bus voltage deviation transmitted by the monitoring equipment through the power communication network, or bus voltage deviation collected in real-time by a voltage sensor.

[0015] Preferably, the target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology are input into a regulation parameter prediction model constructed based on a dynamic coordination algorithm. The regulation parameter prediction model outputs a set of predicted coordinated control parameters, including:

[0016] The target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology are transformed into feature vectors that meet the input requirements of the dynamic coordination model using a wavelet transform algorithm. These feature vectors are then input into the adjustment parameter prediction model, which outputs a set of predicted preliminary control parameters. The mechanical inertia parameters of the flywheel unit in the energy storage array are obtained, and a set of dynamic correction coefficients are calculated based on these parameters. These dynamic correction coefficients are then used to dynamically optimize the flywheel torque compensation and the switching frequency corresponding to the preliminary control parameters, thereby obtaining the coordinated control parameters.

[0017] Preferably, the mechanical inertia parameters of the flywheel unit in the energy storage array are obtained, and a set of dynamic correction coefficients are calculated based on the mechanical inertia parameters, including:

[0018] The axial vibration amplitude of the flywheel unit is collected by a vibration sensor, and the equivalent mass distribution of the flywheel rotor is calculated based on the vibration amplitude. The lubrication state parameters of the flywheel bearing and the stiffness coefficient of the rotor material are obtained. A set of dynamic correction coefficients is obtained by matching the lubrication state parameters, the stiffness coefficient and the equivalent mass distribution. Among them, the torque compensation coefficient and the switching frequency coefficient in the dynamic correction coefficient are negatively correlated with the equivalent mass distribution.

[0019] Preferably, the regulation parameter prediction model is trained in the following manner: collecting a set of historical operating datasets, wherein each data point in the dataset includes bus voltage deviation, flywheel speed threshold, topology, torque compensation amount, switching frequency, and label data, wherein the label data is the completion degree of the smoothing index; and using the historical operating datasets to train the dynamic coordination algorithm offline to obtain the regulation parameter prediction model.

[0020] Preferably, the generation of the tag data includes a simulation verification step: establishing an electromechanical coupling model of the flywheel array, injecting a standard load fluctuation waveform into the PSCAD / EMTDC simulation platform, and recording the percentage decrease in bus voltage fluctuation rate under different combinations of control parameters as a quantitative value of the completion of the smoothing index.

[0021] Preferably, the dynamic coordination algorithm uses a hybrid modeling architecture based on spatiotemporal graph convolutional networks to process the feature vectors. The spatiotemporal convolutional layer includes an adaptive adjacency matrix generation module, which constructs a dynamic coupling relationship graph between flywheel units in real time according to the energy storage array topology, and synchronously captures the inertial response features within an integer multiple window of the grid fundamental period through a gated temporal convolutional kernel. The output layer uses a spatiotemporal attention mechanism to fuse spatial weights and temporal weights to generate initial values ​​for torque compensation and switching frequency, respectively.

[0022] Preferably, the calculation of the dynamic correction coefficient further includes: real-time monitoring of the bus voltage change rate, and when the change rate exceeds a preset threshold, activating the dynamic inertia compensation module. This module fuses vibration amplitude and speed threshold data through a Kalman filter to generate an online estimate of the bearing friction torque for correcting the torque compensation coefficient.

[0023] Preferably, the spectrum analysis model includes an adaptive notch filter bank, the center frequency of which is dynamically adjusted according to historical load fluctuation characteristics, and each filter output is connected to an amplitude integrator. The criterion for determining the main frequency component is the frequency band whose integral value exceeds the baseline by 20% within three consecutive sampling periods.

[0024] Preferably, it also includes a fault emergency handling step: when the speed deviation of any flywheel unit is detected to continuously exceed the threshold, the control terminal automatically performs energy storage array topology reconfiguration, switches the faulty unit to offline state while connecting the backup flywheel unit in parallel, and maintains a smooth transition of total output torque during the switching process.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] By constructing a regulation parameter prediction model based on a dynamic coordination algorithm, deep fusion of multi-source data and accurate prediction of control parameters were achieved. This model employs a spatiotemporal graph convolutional network hybrid architecture. An adaptive adjacency matrix generation module analyzes the energy storage array topology in real time, dynamically constructing a coupling relationship graph between flywheel units. Simultaneously, gated temporal convolutional kernels capture the inertial response characteristics within the fundamental cycle of the power grid, and a spatiotemporal attention mechanism fuses spatial and temporal dimension weights to generate torque compensation and initial switching frequency values, respectively. This mechanism enables the model to accurately capture the time-frequency characteristics of load fluctuations and dynamic changes in the array topology, effectively improving its analytical capabilities and control parameter prediction accuracy in complex load fluctuation scenarios.

[0027] A dynamic correction mechanism for mechanical characteristics and control parameters was established. The axial vibration amplitude of the flywheel unit is collected in real time using vibration sensors. Combined with bearing lubrication state parameters and rotor stiffness coefficient, a dynamic correction coefficient related to the equivalent mass distribution is calculated. When the bus voltage change rate exceeds a threshold, the dynamic inertia compensation module is activated. A Kalman filter is used to fuse vibration and speed data, estimate bearing friction torque online, and correct control parameters. This mechanism effectively establishes a real-time mapping relationship between flywheel unit mechanical parameters (such as equivalent mass distribution and lubrication state) and control parameters. It can dynamically compensate for the impact of mechanical characteristic drift (such as rotor eccentricity and bearing wear) on control accuracy, improving the system's adaptability to changes in equipment operating conditions.

[0028] A closed-loop fault emergency handling system was designed. When a sustained excessive speed deviation of the flywheel unit is detected, the control terminal automatically performs topology reconfiguration, quickly isolating the faulty unit and connecting a backup unit in parallel. Simultaneously, a torque smoothing transition algorithm maintains stable total output torque. During topology reconfiguration, control parameters are recalculated based on a dynamic coordination algorithm to ensure the backup unit participates in mitigation in a timely manner, shortening the power gap recovery time caused by unit failure. This mechanism effectively avoids secondary grid impact during fault handling, enhancing the fault tolerance of the energy storage array and its continuous support capability to the grid.

[0029] In the load characteristic analysis and model training stages, a multi-algorithm optimization strategy was introduced. A spectrum analysis model incorporating an adaptive notch filter bank was adopted. The center frequency of the filter was dynamically adjusted based on historical load data, and the main frequency component was accurately extracted through an amplitude integrator, improving the extraction accuracy of the target smoothing index. Based on the smoothing index completion labels generated from multi-dimensional historical operating data and simulation verification, the dynamic coordination algorithm was trained offline, enhancing the model's generalization ability under complex load scenarios and ensuring the reliability and stability of the control strategy.

[0030] A complete closed-loop chain is formed from data acquisition, feature processing, parameter prediction to execution control. Monitoring data is acquired in real time through the power communication network, feature vectors are generated using wavelet transform algorithms, and fully automatic control is achieved by combining dynamic correction and topology reconstruction mechanisms. This method can adjust the operating status in real time according to load fluctuations without human intervention, realizing intelligent operation of the entire process of "data perception - analysis and decision-making - execution control", reducing manual operation and maintenance costs, and improving the power system's response speed and control efficiency to load fluctuations. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the working principle of the flywheel energy storage array control method for load mitigation described in this invention.

[0032] Figure 2 A flowchart for input processing and control parameter optimization for the parameter prediction model;

[0033] Figure 3 A flowchart for processing mechanical inertia parameters and calculating dynamic correction coefficients for flywheel units;

[0034] Figure 4 This is a flowchart of the dynamic coordination algorithm based on spatiotemporal graph convolutional networks;

[0035] Figure 5 A flowchart for monitoring the rate of change of bus voltage and correcting the torque compensation coefficient. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figures 1-5 The present invention relates to a flywheel energy storage array control method for load smoothing, the specific implementation steps of which are as follows:

[0038] Step 1: Obtain real-time data related to grid load fluctuations and target smoothing indicators. The real-time data includes bus voltage deviation, flywheel speed threshold, and energy storage array topology.

[0039] Step 2: Input the target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology into the regulation parameter prediction model constructed based on the dynamic coordination algorithm. The regulation parameter prediction model outputs a set of predicted coordination control parameters. The coordination control parameters include the flywheel torque compensation amount and the energy storage unit switching frequency.

[0040] Step 3: Adjust the operating status of the flywheel energy storage array according to the coordinated control parameters to achieve grid load smoothing control.

[0041] The present invention will be further described below with reference to Examples 1 to 5:

[0042] Example 1:

[0043] In acquiring real-time data related to grid load fluctuations and target smoothing indicators, the system establishes a real-time data transmission channel with monitoring equipment via a power communication network. Monitoring equipment is deployed at key nodes of the grid and in various units of the flywheel energy storage array, continuously collecting data on bus voltage deviation, flywheel speed thresholds, and energy storage array topology, and transmitting this data to the control terminal in real time via the power communication network. The acquisition of bus voltage deviation data employs a dual mechanism: firstly, it can directly receive processed data transmitted from the monitoring equipment via the network; secondly, it can acquire raw voltage signals in real time through independently deployed voltage sensors, generating bus voltage deviation values ​​after analog-to-digital conversion and filtering. These two methods ensure the reliability and redundancy of data acquisition.

[0044] The determination of the target mitigation index is based on frequency domain analysis logic. Operators input grid load fluctuation characteristic data through the human-machine interface of the control terminal. This data can be derived from historical operating records or real-time monitoring sampling values. The system calls a spectrum analysis model to perform frequency domain decomposition on the characteristic data. The model includes an adaptive notch filter bank, and the center frequency of each filter can be dynamically adjusted according to historical load fluctuation characteristics, forming a frequency response characteristic that changes with load characteristics. After the filters perform frequency division processing on the input signal, the amplitude integrators at each output end accumulate and calculate the signal energy of the corresponding frequency band. The determination of the dominant frequency component must meet a strict time-domain continuity condition: when the integral value of a certain frequency band exceeds 20% of the baseline for three consecutive sampling periods, that frequency band is determined to be the dominant frequency component of the load fluctuation. Based on the matching result of the dominant frequency component and the preset spectrum template, the target mitigation index is automatically generated. This index includes parameters such as the suppression amplitude and response time for the dominant frequency component, providing a clear control target for the prediction of subsequent adjustment parameters.

[0045] The energy storage array topology data reflects the current connection status of the flywheel units, including the parallel / series relationship of each unit and the access status of redundant backup units. This data is collected and updated in real time by the topology sensing module inside the array, ensuring that the regulation parameter prediction model can be calculated based on the latest physical connection status. The flywheel speed threshold data is monitored in real time by the speed sensors built into each flywheel unit, reflecting the distance between the current flywheel speed and the safe operating boundary, and is used to assess the available regulation capacity of the flywheel unit.

[0046] Through the collaborative acquisition and processing of multi-source data described above, the system achieves accurate characterization of power grid load fluctuations and clear definition of control objectives, providing a complete and real-time input dataset for predicting regulation parameters based on dynamic coordination algorithms. This process does not involve any assumptions or inferences regarding experimental results; it acquires key parameters solely through data collection, signal processing, and logical judgment, ensuring the accuracy and reliability of subsequent control procedures.

[0047] Example 2:

[0048] In the process of inputting the target mitigation index, bus voltage deviation, flywheel speed threshold, and topology into the regulation parameter prediction model to obtain coordinated control parameters, the first step is to standardize the multi-source heterogeneous data. The system calls the wavelet transform algorithm, selecting appropriate wavelet basis functions to perform multi-resolution decomposition of the target mitigation index, bus voltage deviation time series, flywheel speed threshold dynamic range, and the graph theory representation (such as the adjacency matrix) of the energy storage array topology, based on the frequency characteristics and scale features of different data types. Through wavelet transform, the original data is converted into feature vectors containing different frequency components and spatial features, ensuring that the data dimensionality matches the input requirements of the dynamic coordination model and providing a unified format of input samples for subsequent modeling.

[0049] After the feature vectors are input into the adjustment parameter prediction model, the model is processed based on a hybrid modeling architecture using a dynamic coordination algorithm. This architecture employs a spatiotemporal graph convolutional network, where the adaptive adjacency matrix generation module of the spatiotemporal convolutional layer reads the energy storage array topology data in real time and dynamically constructs a coupling relationship graph between flywheel units. Nodes in the graph represent flywheel units, and the edge weights are adjusted in real time according to parameters such as the tightness of electrical connections and physical distances between units to reflect the impact of topology changes on flywheel coordinated control. The gated temporal convolutional kernel performs convolution operations on the temporal information in the feature vectors within time windows that are integer multiples of the grid fundamental frequency period (such as 1 fundamental frequency period, 2 fundamental frequency periods, etc.). Through a gating mechanism, it selectively captures inertial response features and extracts the dynamic patterns of load fluctuations at different time scales.

[0050] The model output layer employs a spatiotemporal attention mechanism, calculating weight coefficients for both the spatial dimension (coupling relationships between flywheel units) and the temporal dimension (features at different sampling times). By fusing spatial and temporal weights, the model generates initial values ​​for torque compensation and switching frequency, forming a set of preliminary control parameters. This process, through end-to-end learning of the neural network, achieves a nonlinear mapping from multi-source data to control parameters, avoiding the complex parameter tuning process in traditional control algorithms.

[0051] To improve the dynamic adaptability of control parameters, the system needs to acquire the mechanical inertia parameters of the flywheel units in the energy storage array to calculate dynamic correction coefficients. Axial vibration amplitude of the flywheel units is collected using vibration sensors. Based on vibration theory and rotor dynamics principles, a mathematical mapping relationship is established between vibration amplitude and the equivalent mass distribution of the flywheel rotor, and the equivalent mass distribution parameters are calculated. Simultaneously, the lubrication status parameters of the flywheel bearings (such as lubricating oil temperature, pressure, and contamination level) and rotor material stiffness coefficients are acquired in real time through a sensor network. These parameters are matched with the equivalent mass distribution using a pre-set database, and dynamic correction coefficients are obtained using lookup table methods or interpolation methods. The torque compensation coefficient and switching frequency coefficient are negatively correlated with the equivalent mass distribution; that is, the more uneven the equivalent mass distribution, the larger the adjustment range of torque compensation and the lower the adjustment threshold of the energy storage unit switching frequency, in order to adapt to the spatial distribution differences of mechanical inertia.

[0052] In addition, the system continuously monitors the bus voltage change rate. When the change rate exceeds a preset threshold, it determines that the grid load has entered a state of severe fluctuation and activates the dynamic inertia compensation module. This module fuses vibration amplitude and speed threshold data using a Kalman filter, and utilizes the recursive optimal estimation characteristic of the Kalman filter to generate an online estimate of the bearing friction torque in real time. This estimate is used as a correction term in the calculation process of the torque compensation coefficient, dynamically adjusting the initial correction coefficient based on the equivalent mass distribution to ensure that the torque compensation can simultaneously respond to changes in mechanical inertia and real-time fluctuations in the grid.

[0053] The training process of the regulation parameter prediction model is driven by historical operating data. First, a historical operating dataset is collected, including bus voltage deviation, flywheel speed threshold, topology, torque compensation, switching frequency, and label data (smoothing index completion rate). Label data generation is achieved through a simulation verification process: an electromechanical coupling model of the flywheel array is established in the PSCAD / EMTDC simulation platform. This model accurately describes the mechanical motion equations of the flywheel rotor and the electrical characteristics of the power electronic converter. Standard load fluctuation waveforms (such as step waves and sinusoidal waves with superimposed noise) are injected into the model to simulate load disturbances under different operating conditions. The percentage decrease in bus voltage fluctuation rate under different combinations of control parameters is recorded, and this percentage is used as the quantitative value of the smoothing index completion rate, providing labels for supervised learning. Through offline training, the network parameters of the dynamic coordination algorithm are continuously optimized, enabling the model to learn the mapping relationship between load fluctuation characteristics and control parameters from historical data, ultimately forming a regulation parameter prediction model with predictive capabilities.

[0054] The entire implementation process, through a comprehensive design encompassing data preprocessing, model inference, mechanical parameter sensing, dynamic correction, and model training, achieves adaptive adjustment from load characteristics to control parameters. Each step is based on physical principles and algorithmic logic, without introducing any hypothetical descriptions of experimental effects, ensuring the feasibility and rigor of the technical solution.

[0055] Example 3:

[0056] When acquiring the mechanical inertia parameters of the flywheel unit in the energy storage array and calculating the dynamic correction coefficient, the system achieves real-time perception of the physical state of the flywheel unit through multi-sensor fusion. Vibration sensors collect the axial vibration amplitude of the flywheel unit at a fixed sampling frequency (e.g., 10kHz). This signal contains mechanical state information such as rotor imbalance and bearing wear. Based on time-frequency analysis of the vibration signal, the time-domain vibration amplitude is converted into frequency-domain features using Fast Fourier Transform (FFT), identifying the fundamental frequency component that is in sync with the rotor speed. Combined with the imbalance response formula in rotor dynamics, the equivalent mass distribution parameters of the flywheel rotor are calculated. These parameters characterize the degree of eccentricity and distribution pattern of the rotor mass relative to the rotation axis.

[0057] The lubrication status parameters of the flywheel bearing are acquired through multi-parameter sensors integrated into the bearing housing, including a lubricating oil temperature sensor, a viscometer, and a contamination detector. The lubricating oil temperature reflects the frictional heat generated during bearing operation; increased temperature may indicate insufficient lubrication or accelerated wear. The viscometer monitors real-time changes in lubricating oil viscosity; decreased viscosity leads to a thinner oil film, increasing mechanical wear. The contamination detector measures the impurity content in the lubricating oil using particle counting; impurity particles may scratch the bearing surface, affecting operational stability. The rotor material stiffness coefficient is determined based on material property parameters (such as elastic modulus and Poisson's ratio) in the flywheel design drawings and a real-time temperature compensation algorithm. Material stiffness decreases non-linearly with increasing temperature, requiring correction of the theoretical value using temperature sensor data.

[0058] The calculation of the dynamic correction coefficient is based on a pre-established mapping relationship of mechanical parameters. Equivalent mass distribution, lubrication state parameters (temperature, viscosity, contamination level), and rotor material stiffness coefficient are used as input variables. The torque compensation coefficient and switching frequency coefficient are obtained through matching using a multidimensional lookup table or machine learning model (such as random forest). The torque compensation coefficient is used to adjust the baseline value of the flywheel torque compensation, while the switching frequency coefficient is used to set the threshold for the switching frequency of the energy storage unit. Since a more uneven equivalent mass distribution results in a greater rotor imbalance force, the torque compensation needs to be increased to offset the inertial torque fluctuations; therefore, the torque compensation coefficient is negatively correlated with the equivalent mass distribution. Similarly, uneven equivalent mass distribution leads to uneven mechanical stress distribution in the flywheel unit, and frequent switching may exacerbate wear. Therefore, the switching frequency coefficient decreases as the equivalent mass distribution increases to reduce the switching frequency.

[0059] Monitoring of the bus voltage change rate is achieved by real-time acquisition of the bus voltage signal and calculation of the difference between adjacent sampling points, with the sampling period synchronized with the flywheel control period (e.g., 1 ms). When the voltage change rate exceeds a preset threshold (e.g., 5% rated voltage / ms), it is determined that the grid load fluctuation has intensified, requiring dynamic correction of the torque compensation coefficient. At this time, the dynamic inertia compensation module is activated, using a Kalman filter to fuse vibration amplitude and speed threshold data. The Kalman filter establishes a system model including state equations (flywheel speed, rotor position) and measurement equations (vibration amplitude, speed threshold), and estimates the real-time value of bearing friction torque through a prediction-update iterative process. This estimated value is used as a correction term and multiplied by the torque compensation coefficient calculated based on the equivalent mass distribution to generate the final torque compensation coefficient, compensating for the dynamic changes in mechanical inertia caused by drastic load fluctuations.

[0060] The entire calculation process employs a layered processing logic: first, the equivalent mass distribution is obtained through vibration signal analysis, and the initial dynamic correction coefficient is determined by combining the lubrication state and material stiffness; then, based on real-time fluctuations in the power grid, the correction coefficient is adjusted online using Kalman filtering technology. Each step is based on sensor measurement data and physical models, avoiding the introduction of hypothetical experimental effects and ensuring that the calculation process of the dynamic correction coefficient is traceable and reproducible, providing mechanical parameter-level support for the precise control of the flywheel energy storage array.

[0061] Example 4:

[0062] During the training of the regulation parameter prediction model, the system first constructs a historical operation dataset covering multi-dimensional operating scenarios. Each sample in the dataset includes bus voltage deviation, flywheel speed threshold, energy storage array topology, torque compensation, switching frequency, and label data (smoothing index completion rate). Among them, bus voltage deviation is collected by voltage sensors at grid monitoring points, covering fluctuation data under different load levels such as normal operation, light load, and overload; flywheel speed threshold, combined with the rated speed of the flywheel and safety protection limits, reflects the regulation capability of each unit under different energy storage states; topology data records the real-time connection configuration of the flywheel unit (such as the number of parallel units and the status of standby units); torque compensation and switching frequency are historical control parameter records, derived from the regulation output during actual operation.

[0063] The generation of tag data relies on a simulation verification process. An electromechanical coupling model of the flywheel array is established in the PSCAD / EMTDC simulation platform. This model includes the mechanical dynamics equations of the flywheel rotor (describing the relationship between speed, torque, and moment of inertia) and the electrical control module of the converter (converting torque commands into current and voltage). Standard load fluctuation waveforms (such as typical load disturbance waveforms recommended by IEEE) are injected to simulate grid frequency fluctuations and voltage sags. Under each combination of control parameters, the percentage decrease in bus voltage fluctuation rate is recorded. This percentage is calculated by comparing the voltage fluctuation amplitude before and after applying control, serving as a quantitative value for the completion of the smoothing index. The tag data generation process strictly follows the physical modeling rules of the simulation platform, ensuring that the causal relationship between tag values ​​and control parameters can be verified through model derivation.

[0064] The dynamic coordination algorithm is trained based on a spatiotemporal graph convolutional network architecture. The input layer receives feature vectors processed by wavelet transform, containing information in both the temporal domain (load fluctuation sequence) and the spatial domain (topological adjacency matrix). The adaptive adjacency matrix generation module of the spatiotemporal convolutional layer updates the coupling relationship graph between flywheel units in real time based on the topological data. The weights of the edges in the graph are determined by parameters such as the electrical interconnection impedance between units and the power flow distribution ratio, reflecting the impact of physical connections on coordinated control. The gated temporal convolutional kernel adopts a multi-level dilated convolutional structure, selectively extracting inertial response features through a gating mechanism for integer multiples of the grid fundamental period (such as the 20ms period of a 50Hz system) (20ms, 40ms, 80ms, etc.) to capture the energy distribution of load fluctuations at different time scales.

[0065] The spatiotemporal attention mechanism of the output layer is divided into two branches: spatial attention and temporal attention. The spatial attention branch calculates the importance weight of each flywheel unit in the current topology, with the weight value determined by parameters such as the unit's remaining energy storage capacity and mechanical health status. The temporal attention branch assigns weights to feature vectors at different sampling times, highlighting the characteristics of key periods of load fluctuation (such as the moment of disturbance occurrence and the steady-state adjustment phase). By weighted fusion of spatial and temporal weights, initial values ​​for torque compensation and switching frequency are generated respectively, achieving spatiotemporal joint optimization of control parameters.

[0066] The model training employs a supervised learning model, using mean squared error (MSE) as the loss function and optimizing network parameters (such as convolutional kernel weights and attention mechanism parameters) through backpropagation. The training process is divided into two stages: offline pre-training and online fine-tuning. In the offline stage, the model is trained in batches using historical datasets, forming a basic prediction model upon convergence. In the online stage, incremental training is automatically triggered when new operational data accumulates to a preset sample size (e.g., 1000 samples), updating the model parameters to adapt to the time-varying characteristics of the power grid load. The entire training process does not rely on any hypothetical descriptions of experimental results; it only improves model performance through data-driven and algorithmic iteration, ensuring that the model's predictive ability is guaranteed by the integrity of the training data and the rigor of the algorithmic logic.

[0067] Example 5:

[0068] The fault emergency handling steps of this invention ensure the stable operation of the flywheel energy storage array through real-time monitoring and topology reconfiguration mechanisms. The system collects real-time speed signals at a fixed frequency through the speed sensor built into the flywheel unit and calculates the speed deviation. The formula for calculating the speed deviation is:

[0069]

[0070] In the formula, This represents the real-time rotational speed of the flywheel unit (unit: revolutions per minute). The rated speed of the flywheel unit is given in revolutions per minute. This formula is only used for fault determination in this embodiment and does not conflict with the symbol definitions in other embodiments.

[0071] When the speed deviation of any flywheel unit Continuously exceeding the preset threshold When this threshold is preset based on the flywheel mechanical design parameters and safe operating range, the control terminal triggers the fault emergency response procedure through logical judgment. The specific processing steps are as follows:

[0072] Fault Detection and Judgment: The system continuously monitors the speed deviation signal. If... The state duration reaches the preset duration If the time interval is 100 milliseconds, the flywheel unit is determined to be abnormal, and a fault identifier is immediately generated and sent to the control terminal.

[0073] Topology reconfiguration trigger: After receiving a fault identifier, the control terminal automatically initiates the energy storage array topology reconfiguration program. First, it queries the redundancy configuration information of the energy storage array to determine a standby flywheel unit with a capacity matching the faulty unit (the standby unit must be in hot standby mode, i.e., it has been initialized and can be connected to the system at any time).

[0074] Parallel connection of backup units: Selected backup flywheel units are connected in parallel to the energy storage array bus via power electronic switching devices (such as solid-state relays). During the connection process, a soft-start control strategy is adopted to gradually increase the output torque of the backup units, avoiding impact on the power grid. The initial torque command of the backup units is determined based on the current load sharing of the faulty units to ensure the continuity of the total output torque.

[0075] Offline switching of faulty units: After the backup unit is fully connected and operating stably, the torque output of the faulty unit is gradually reduced to zero, and its electrical connection with the bus is disconnected, switching it to an offline state. During the switching process, the torque distribution of each unit is dynamically adjusted by monitoring the fluctuations in bus voltage and frequency in real time to maintain the total output torque. The formula for calculating constant values ​​is:

[0076]

[0077] In the formula, The output torque of the standby unit (unit: N·m). This is the sum of the torques of the remaining normally operating flywheel units (unit: Newton-meter).

[0078] Status Update and Alarm: After the topology reconfiguration is completed, the system automatically updates the topology data of the energy storage array, marks the faulty unit as "pending maintenance", and sends alarm information to the operation and maintenance personnel through the human-machine interface. The content includes the faulty unit number, fault type (abnormal speed) and topology reconfiguration completion time.

[0079] The entire emergency response process is based on real-time sensor data and preset logic rules. Through quantitative calculation of speed deviation, orderly access of backup units, and dynamic torque distribution, it achieves rapid isolation of faulty units and a smooth transition of system operating status. This process does not involve any hypothetical descriptions of experimental effects; it only ensures the reliability of the energy storage array through physical quantity monitoring, logic control, and mathematical calculations, ensuring that the grid load smoothing function is not affected by unit failures.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for controlling a flywheel energy storage array for load mitigation, characterized in that: The method includes the following steps: Acquire real-time data and target smoothing indicators related to grid load fluctuations. The real-time data includes bus voltage deviation, flywheel speed threshold, and energy storage array topology. The target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology are input into a regulation parameter prediction model constructed based on a dynamic coordination algorithm. The regulation parameter prediction model outputs a set of predicted coordination control parameters. The coordination control parameters include flywheel torque compensation and energy storage unit switching frequency. The operating status of the flywheel energy storage array is adjusted according to the aforementioned coordinated control parameters to achieve grid load smoothing control. The acquisition of real-time data related to power grid load fluctuations and target smoothing indicators includes: The system receives bus voltage deviation, flywheel speed threshold, and energy storage array topology data transmitted by monitoring equipment via a power communication network. The target smoothing index is determined as follows: power grid load fluctuation characteristic data is input to the control terminal; a spectrum analysis model is used to perform frequency domain decomposition on the characteristic data to extract the dominant frequency component of the load fluctuation; the target smoothing index is obtained by matching the dominant frequency component with a preset spectrum template. The system also receives bus voltage deviation transmitted by monitoring equipment via a power communication network, or bus voltage deviation collected in real-time by a voltage sensor. The target smoothing index includes suppression amplitude and response time parameters for the dominant frequency component, providing a clear control target for predicting subsequent adjustment parameters. The regulation parameter prediction model is trained in the following manner: a set of historical operation datasets is collected, each data point in the dataset including bus voltage deviation, flywheel speed threshold, topology, torque compensation amount, switching frequency, and label data, wherein the label data is the completion degree of the smoothing index; the dynamic coordination algorithm is trained offline using the historical operation datasets to obtain the regulation parameter prediction model. The generation of the tag data includes simulation verification steps: establishing an electromechanical coupling model of the flywheel array, injecting standard load fluctuation waveforms into the PSCAD / EMTDC simulation platform, and recording the percentage decrease in bus voltage fluctuation rate under different combinations of control parameters as a quantitative value of the completion of the smoothing index; the regulation parameter prediction model, through offline training, continuously optimizes the network parameters of the dynamic coordination algorithm, enabling the electromechanical coupling model to learn the mapping law between load fluctuation characteristics and control parameters from historical data, forming a regulation parameter prediction model with predictive capabilities.

2. The flywheel energy storage array control method for load smoothing according to claim 1, characterized in that, The target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology are input into a regulation parameter prediction model constructed based on a dynamic coordination algorithm. The regulation parameter prediction model outputs a set of predicted coordinated control parameters, including: The target smoothing index, the bus voltage deviation, the flywheel speed threshold, and the topology are transformed into feature vectors that meet the input requirements of the dynamic coordination model using a wavelet transform algorithm. These feature vectors are then input into the adjustment parameter prediction model, which outputs a set of predicted preliminary control parameters. The mechanical inertia parameters of the flywheel unit in the energy storage array are obtained, and a set of dynamic correction coefficients are calculated based on these parameters. These dynamic correction coefficients are then used to dynamically optimize the flywheel torque compensation and the switching frequency corresponding to the preliminary control parameters, thereby obtaining the coordinated control parameters.

3. The flywheel energy storage array control method for load smoothing according to claim 2, characterized in that, Obtain the mechanical inertia parameters of the flywheel unit in the energy storage array, and calculate a set of dynamic correction coefficients based on the mechanical inertia parameters, including: The axial vibration amplitude of the flywheel unit is collected by a vibration sensor, and the equivalent mass distribution of the flywheel rotor is calculated based on the vibration amplitude. The lubrication state parameters of the flywheel bearing and the stiffness coefficient of the rotor material are obtained. A set of dynamic correction coefficients is obtained by matching the lubrication state parameters, the stiffness coefficient and the equivalent mass distribution. Among them, the torque compensation coefficient and the switching frequency coefficient in the dynamic correction coefficient are negatively correlated with the equivalent mass distribution.

4. The flywheel energy storage array control method for load smoothing according to claim 3, characterized in that, The dynamic coordination algorithm uses a hybrid modeling architecture based on spatiotemporal graph convolutional networks to process the feature vectors. The spatiotemporal convolutional layer includes an adaptive adjacency matrix generation module, which constructs a dynamic coupling relationship graph between flywheel units in real time according to the energy storage array topology and synchronously captures the inertial response features within an integer multiple window of the grid fundamental period through gated temporal convolutional kernels. The output layer uses a spatiotemporal attention mechanism to fuse spatial and temporal weights, generating initial values ​​for torque compensation and switching frequency, respectively.

5. The flywheel energy storage array control method for load smoothing according to claim 4, characterized in that, The calculation of the dynamic correction coefficient further includes: real-time monitoring of the bus voltage change rate, and activation of the dynamic inertia compensation module when the change rate exceeds a preset threshold. This module fuses vibration amplitude and speed threshold data through a Kalman filter to generate an online estimate of the bearing friction torque for correcting the torque compensation coefficient.

6. The flywheel energy storage array control method for load smoothing according to claim 5, characterized in that, The spectrum analysis model includes an adaptive notch filter bank. The center frequency of the filter is dynamically adjusted according to the historical load fluctuation characteristics. Each filter output is connected to an amplitude integrator. The criterion for determining the main frequency component is the frequency band where the integral value exceeds the baseline by 20% within three consecutive sampling periods.

7. The flywheel energy storage array control method for load smoothing according to claim 1, characterized in that, It also includes emergency fault handling steps: when the speed deviation of any flywheel unit is detected to exceed the threshold, the control terminal automatically performs energy storage array topology reconfiguration, switches the faulty unit to offline state and connects the backup flywheel unit in parallel, and maintains a smooth transition of total output torque during the switching process.

Images