Accurate control method for power output of energy storage power station

Abstract

The invention provides an accurate control method for power output of an energy storage power station, which belongs to the technical field of energy storage power stations, and comprises the following steps of: establishing a multi-dimensional monitoring data vector by collecting real-time operation data of the energy storage power station, and constructing a power output stable reference matrix and a power fluctuation characteristic matrix as control references; a virtual inertia control algorithm is used to calculate power deviation, an improved chessboard coverage algorithm is used to carry out abnormal mode identification, power fluctuation spectrum characteristics are analyzed through fast Fourier transform to establish an abnormal fluctuation identification vector, and a future power output prediction matrix is generated based on a power trend prediction function containing a neural network item. And the charge-discharge rate and the power factor of the energy storage system are adaptively adjusted according to the abnormal detection result and the prediction information, millisecond-level power output control is implemented to complete power fluctuation suppression, and the technical problem that the power output fluctuation of the energy storage power station is difficult to accurately control in real time is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of energy storage power stations, and in particular relates to a precise control method for power output of an energy storage power station. BACKGROUND

[0002] As an important frequency modulation and peak shaving device in the power system, the traditional power output control of the energy storage power station mainly relies on a voltage and current double-loop control strategy based on a PI controller, power regulation is achieved by monitoring the terminal voltage and output current of the energy storage unit, and a filter is used to suppress and process power fluctuations. This control method is widely used in the commercial application of the energy storage power station. However, the traditional control method has a slow response speed when facing rapid changes in the grid frequency and sudden changes in the load, and the power output regulation accuracy is not high enough, especially in the millisecond time scale, it is difficult to achieve precise control, the fixed time constant of the filter processing leads to poor suppression effect on different types of power fluctuations, and there is a lack of accurate prediction ability for the power output trend. In the current application environment of large-scale access of the energy storage power station to the power grid, due to the complex and diverse changes in the grid load and the constantly changing frequency fluctuation mode, the traditional fixed parameter control strategy is difficult to adapt to the dynamically changing operating conditions, resulting in large fluctuations in the power output of the energy storage power station, affecting the stability of the power grid and the operating efficiency of the energy storage system, that is, there is a technical problem of difficulty in precise and real-time control of the power output fluctuation of the energy storage power station in the prior art. SUMMARY

[0003] Therefore, the application provides a precise control method for power output of an energy storage power station, which can solve the technical problem of difficulty in precise and real-time control of the power output fluctuation of the energy storage power station in the prior art.

[0004] The application is implemented in the following manner: the application provides an accurate control method for power output of an energy storage power station, which collects real-time voltage, current, frequency, power factor and temperature data of the energy storage power station to establish a real-time monitoring data vector, simultaneously collects power grid frequency fluctuation data to establish a frequency fluctuation data set, constructs a power output stability reference matrix and a power fluctuation characteristic matrix based on historical operation data, calculates the deviation of current power output from the stability reference to establish a power deviation vector through a virtual inertia control algorithm, analyzes and processes the power deviation vector using an improved chessboard covering algorithm, starts a power recall control process when the power abnormality indication value output by the improved chessboard covering algorithm is greater than a stability threshold, analyzes power fluctuation frequency spectrum characteristics to establish a fluctuation frequency spectrum data set using fast Fourier transform, judges the fluctuation type according to the frequency spectrum characteristics and establishes an abnormal fluctuation identification vector, calculates power smoothing adjustment parameters based on an adaptive filtering algorithm and a feedforward compensation control, generates a power trend prediction matrix through a power trend prediction function containing a neural network term, adjusts the charge-discharge rate and power factor of the energy storage system according to the abnormal fluctuation identification vector and the power trend prediction matrix, implements millisecond-level power output control, and realizes power fluctuation suppression by adjusting the voltage and current output of the energy storage unit.

[0005] The real-time monitoring data vector is obtained by an electric energy quality monitoring device of the energy storage power station, and specifically includes voltage effective value, current effective value, frequency instantaneous value, power factor and surface temperature of the energy storage unit, and the sampling frequency is 1000 Hz.

[0006] The frequency fluctuation data set is collected by a power grid frequency monitoring system, and specifically records the change data of the power grid frequency in a 1-second time window, and the data accuracy is 0.01 Hz.

[0007] The power output stability reference matrix is established based on the historical operation data of the energy storage power station in the past 30 days, and specifically includes stable power output reference values under different load conditions, environmental temperatures and power grid states.

[0008] The power fluctuation characteristic matrix is established by analyzing the statistical characteristics of the power output data, and specifically includes amplitude, frequency, duration and periodicity characteristics of the power fluctuation.

[0009] The power deviation vector is calculated by the difference between the current power output value and the corresponding value of the power output stability reference matrix, and specifically includes active power deviation, reactive power deviation and power factor deviation.

[0010] The virtual inertia control algorithm is a control strategy simulating the inertia characteristics of a traditional synchronous generator, which improves the frequency stability of the power grid by introducing a virtual inertia response mechanism in the control loop of the energy storage system, and uses the frequency instantaneous value in the real-time monitoring data vector as the main input parameter, combined with the standard power value in the power output stability reference matrix and the fluctuation mode information in the power fluctuation characteristic matrix.

[0011] The improved chessboard coverage algorithm is used for analyzing the abnormal mode identification of the power deviation vector, and the input includes the power deviation vector, the power fluctuation characteristic matrix, the real-time monitoring data vector, the frequency fluctuation data set and the historical anomaly database, and the output is the power anomaly indication value.

[0012] The stable threshold is set according to 5% of the rated power of the energy storage power station, and when the power anomaly indication value exceeds the stable threshold, it is determined that adjustment is needed.

[0013] The power trend prediction function containing a neural network term is used to predict future power demand based on current power output data and power grid load change trend, and the input includes current power value, power change rate, power grid frequency, load change rate, environmental temperature and neural network term, and the output is power output prediction value within 3 seconds in the future.

[0014] The fluctuation spectrum data set is obtained by performing fast Fourier transform on the power output data, and the analysis frequency range is 0.1Hz to 50Hz.

[0015] The abnormal fluctuation mode includes three types of power output mutation, high-frequency oscillation and low-frequency swing, which are distinguished by frequency spectrum characteristics and amplitude characteristics.

[0016] The power callback control flow includes reducing the power output change rate, increasing the filter processing time constant and starting the coordinated control of the standby energy storage unit. The millisecond-level power output control refers to the control technology of completing power output adjustment within 1ms to 10ms, which adjusts the amplitude and phase of output voltage and current by pulse width modulation control of the inverter of the energy storage unit, realizes fast response and accurate control of power output.

[0017] The abnormal fluctuation identification vector is a binary vector, corresponding to three abnormal states of mutation, high-frequency oscillation and low-frequency swing, respectively, 1 indicating abnormality and 0 indicating normality. The power trend prediction matrix contains power output prediction values at every 100ms time point within 3 seconds in the future, which is used as a reference for feedforward control.

[0018] The application realizes accurate identification of power deviation and abnormal mode detection by establishing a multidimensional data system containing a real-time monitoring data vector, a power output stability benchmark matrix and a power fluctuation characteristic matrix, combining a virtual inertia control algorithm and an improved chessboard covering algorithm, using fast Fourier transform to analyze power fluctuation spectrum characteristics and establishing an abnormal fluctuation identification vector, and generating a power output prediction within the next 3 seconds through a power trend prediction function containing a neural network term, solving the technical defects of slow response speed and insufficient prediction ability of traditional control methods. The application dynamically adjusts filtering parameters through an adaptive filtering algorithm, automatically adjusts the charge-discharge rate and power smoothing processing strength according to the abnormal fluctuation identification vector, realizes millisecond-level power output control, overcomes the limitations of fixed parameter control strategies in traditional methods that are difficult to adapt to dynamic working conditions, and significantly improves the stability and control accuracy of the power output of the energy storage power station. In summary, the application solves the technical problem of the background art that the power output fluctuation of the energy storage power station is difficult to accurately and real-time control. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of the method of the application.

[0020] Figure 2 A neural network architecture diagram related to the application based on a multi-head attention mechanism.

[0021] Figure 3 A power output spectrum analysis diagram in Example 2.

[0022] Figure 4 A power trend prediction curve diagram in Example 2. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0024] As shown in Figure 1 , a flowchart of an accurate control method for power output of an energy storage power station provided by the application, the method includes the following steps: S01, collect real-time voltage, current, frequency, power factor and temperature data of the energy storage power station, establish a real-time monitoring data vector, and collect power grid frequency fluctuation data to establish a frequency fluctuation data set; S02, construct a power output stability benchmark matrix based on historical operation data, the power output stability benchmark matrix contains standard power output values under different working conditions, and establish a power fluctuation characteristic matrix for identifying the fluctuation mode of the power output; S03, calculate the deviation of the current power output from the stable reference through the virtual inertia control algorithm, establish a power deviation vector, and analyze and process the power deviation vector using an improved chessboard covering algorithm. When the power abnormality indication value output by the improved chessboard covering algorithm is greater than the stability threshold, start the power callback control process; S04, analyze the power fluctuation frequency spectrum characteristics using fast Fourier transform, establish a fluctuation frequency spectrum dataset, and determine the fluctuation type according to the frequency spectrum characteristics. When an abnormal fluctuation pattern is detected, an abnormal fluctuation identification vector is established; S05, based on the adaptive filtering algorithm and the feedforward compensation control, calculate the power smoothing adjustment parameters, generate a power trend prediction matrix through a power trend prediction function containing a neural network term, and the power trend prediction matrix is used to predict the power output change in the next 3 seconds; S06, according to the abnormal fluctuation identification vector and the power trend prediction matrix, adjust the charge-discharge rate and power factor of the energy storage system. When there is an abnormal identification in the abnormal fluctuation identification vector, the operator needs to adjust the charge-discharge rate to 80% of the normal value and increase the power smoothing processing strength; S07, implement millisecond-level power output control, adjust the voltage and current output of the energy storage unit to achieve power fluctuation suppression, and update the power output stability reference matrix after the power output is stable to complete a round of precise control cycle.

[0025] Among them, the real-time monitoring data vector is obtained through the power quality monitoring device of the energy storage power station, including voltage effective value, current effective value, frequency instantaneous value, power factor and energy storage unit surface temperature, and the sampling frequency is 1000Hz.

[0026] Among them, the frequency fluctuation dataset is collected through the power grid frequency monitoring system, recording the change data of the power grid frequency in a 1-second time window, and the data accuracy is 0.01Hz.

[0027] Among them, the power output stability reference matrix is established based on the historical operation data of the energy storage power station in the past 30 days, including stable power output reference values under different load conditions, environmental temperature and power grid state.

[0028] Among them, the power fluctuation feature matrix is established by analyzing the statistical characteristics of the power output data, including the amplitude, frequency, duration and periodicity of the power fluctuation.

[0029] Among them, the power deviation vector is calculated by the difference between the current power output value and the corresponding value of the power output stability reference matrix, including active power deviation, reactive power deviation and power factor deviation.

[0030] The virtual inertia control algorithm is a control strategy for simulating the inertia characteristics of a traditional synchronous generator, and the frequency stability of the power grid is improved by introducing a virtual inertia response mechanism in the control link of the energy storage system. The virtual inertia control algorithm uses the frequency instantaneous value in the real-time monitoring data vector as the main input parameter, combines the standard power value in the power output stability reference matrix and the fluctuation mode information in the power fluctuation characteristic matrix, and calculates the inertia response power that the energy storage system should provide through mathematical modeling. The core principle of the algorithm is that when the grid frequency changes, the energy storage system can quickly respond and provide or absorb power to maintain frequency stability. Compared with the traditional frequency regulation method, the virtual inertia control algorithm has the advantages of fast response speed, high regulation accuracy and strong adaptability. The algorithm dynamically adjusts the power output strategy of the energy storage system by monitoring the change rate and amplitude of the grid frequency. When the frequency decreases, the algorithm instructs the energy storage system to release power to support the grid frequency. When the frequency rises, the algorithm instructs the energy storage system to absorb excess power to stabilize the grid frequency. The matching method of the algorithm and the scheme of the application is that the input parameters of the algorithm are completely derived from the real-time monitoring data vector established in step one and the power output stability reference matrix constructed in step two, and the output result of the algorithm is directly used to generate the power deviation vector in step three, providing accurate guidance information for subsequent power regulation control.

[0031] The improved chessboard covering algorithm is used for analyzing the abnormal mode recognition of the power deviation vector, and the input includes the power deviation vector, the power fluctuation characteristic matrix, the real-time monitoring data vector, the frequency fluctuation data set and the historical abnormal database. The output is a power abnormality indication value.

[0032] The stable threshold is set according to 5% of the rated power of the energy storage power station. When the power abnormality indication value exceeds the stable threshold, it is determined that adjustment is needed.

[0033] The power callback control process includes reducing the power output change rate, increasing the filter processing time constant and starting the coordinated control of the standby energy storage unit.

[0034] The fluctuation spectrum data set is obtained by performing fast Fourier transform on the power output data, and the analysis frequency range is 0.1Hz to 50Hz.

[0035] The abnormal fluctuation mode includes three types of power output mutation, high-frequency oscillation and low-frequency swing, which are distinguished by spectral characteristics and amplitude characteristics.

[0036] The abnormal fluctuation identification vector is a binary vector, corresponding to three abnormal states of mutation, high-frequency oscillation and low-frequency swing, respectively. 1 represents an abnormality, and 0 represents normal.

[0037] The power trend prediction function, which includes a neural network term, is used to predict future power demand based on current power output data and grid load change trends. The inputs include the current power value, power change rate, grid frequency, load change rate, ambient temperature, and neural network term. The output is the predicted power output value for the next 3 seconds.

[0038] The power trend prediction matrix contains the predicted power output values ​​at 100 milliseconds per 3 seconds, which are used as a reference for feedforward control.

[0039] The charge / discharge rate is adjusted by controlling the current output amplitude of the energy storage unit. Under normal operation, it is 1C. When the power abnormality indication value is within the range of the stable threshold to twice the stable threshold, the operator needs to adjust it to 0.8C. When the power abnormality indication value exceeds twice the stable threshold, the operator needs to adjust it to 0.6C and activate the emergency power limiting mode.

[0040] The power smoothing intensity is achieved by adjusting the cutoff frequency of the adaptive filter. Under normal circumstances, the cutoff frequency is 10Hz. When there is one abnormal indicator in the abnormal fluctuation indicator vector, the cutoff frequency drops to 5Hz. When there are two or more abnormal indicators in the abnormal fluctuation indicator vector, the operator needs to adjust the cutoff frequency to 2Hz.

[0041] Among them, millisecond-level power output control refers to the control technology that completes power output regulation within a time range of 1 millisecond to 10 milliseconds. It uses a high-speed digital signal processor to perform pulse width modulation control on the inverter of the energy storage unit, and adjusts the amplitude and phase of the output voltage and current to achieve fast response and precise control of power output.

[0042] Among them, the improved chessboard covering algorithm is an optimization algorithm based on the traditional chessboard covering problem. It maps each element of the power deviation vector to chessboard grid points, and identifies abnormal power output patterns by analyzing the connection patterns and coverage paths between grid points. The algorithm adopts a recursive divide-and-conquer strategy and a dynamic weight adjustment mechanism to improve the anomaly detection accuracy.

[0043] The historical anomaly database is established through the energy storage power station operation monitoring system, which records the power deviation characteristics, occurrence time, duration and processing results of all abnormal events in the past 90 days.

[0044] Among them, the emergency power limiting mode is a protective control mode that is activated when the energy storage system experiences severe power fluctuations. It limits the power output to below 60% of the rated power, prioritizing grid stability.

[0045] like Figure 2As shown, the specific structure of the neural network term is a fast response neural network architecture based on a multi-head attention mechanism. The network includes an input layer, multiple hidden layers, and an output layer. The input layer receives five parameters: current power value, power change rate, grid frequency, load change rate, and ambient temperature. The hidden layers employ residual connections and layer normalization techniques to improve training stability and convergence speed. The number of attention heads in the multi-head attention mechanism is dynamically adjusted based on the fluctuation amplitude of the instantaneous frequency value in the real-time monitoring data vector and the fluctuation frequency in the power fluctuation feature matrix. When the fluctuation amplitude of the instantaneous frequency value is less than 0.1 Hz and the dominant fluctuation frequency in the power fluctuation feature matrix is ​​less than 5 Hz, the network is activated. The number of attention heads is set to 4 at time z. When the instantaneous frequency fluctuation is within the range of 0.1Hz to 0.3Hz or the dominant fluctuation frequency in the power fluctuation feature matrix is ​​within the range of 5Hz to 15Hz, the number of attention heads is adjusted to 8. When the instantaneous frequency fluctuation exceeds 0.3Hz or the dominant fluctuation frequency in the power fluctuation feature matrix exceeds 15Hz, the number of attention heads is increased to 16. Each attention head is specifically designed to handle power change patterns at different time scales and frequency ranges. The output layer generates power trend compensation values ​​through fully connected layers and activation functions. The entire neural network adopts an end-to-end training method to optimize compensation accuracy and response speed.

[0046] The steps for establishing the training dataset for the neural network term model specifically include: extracting power output data, grid frequency data, load change data, and ambient temperature data from the historical operation records of the energy storage power station over the past 180 days; organizing the data into a continuous data stream according to the time series; performing preprocessing on the data, including outlier detection and smoothing filtering; dividing the processed data into 5-minute time windows, with each time window containing 300 data points; extracting the data from the first 180 seconds of each time window as input features; calculating the basic predicted value using traditional power prediction methods; using the difference between the actual power output value and the basic predicted value as the power trend compensation value target that the neural network term needs to learn; generating training samples using a sliding window method; and finally establishing a complete training dataset containing more than 50,000 training samples. The training dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio.

[0047] The specific steps of training the neural network model include iteratively updating the neural network parameters using an adaptive learning rate optimization algorithm, with an initial learning rate set to 0.001. The learning rate is dynamically adjusted using a cosine annealing scheduling strategy. During training, the mean squared error loss function is used to measure the difference between the predicted power trend compensation value and the actual compensation value. At the same time, a regularization term is introduced to prevent overfitting. The training adopts the batch gradient descent method with a batch size of 64 and 200 training rounds. After each training round, the model performance is evaluated using a validation set. Training is stopped early when the validation set loss no longer decreases for 10 consecutive rounds. After training, the compensation accuracy of the model is finally evaluated using a test set. The model parameters with the best performance on the validation set are saved during training as the final pre-trained model. The pre-trained model can be directly loaded and used in practical applications without retraining.

[0048] The specific implementation methods of the above steps are described in detail below.

[0049] The specific implementation of step S01 involves real-time acquisition of multi-dimensional operational data through the power quality monitoring device of the energy storage power station. First, a high-precision data acquisition module is activated. This module, based on the principle of synchronous sampling, simultaneously acquires five key parameters: effective voltage, effective current, instantaneous frequency, power factor, and surface temperature of the energy storage unit. The sampling frequency is set to 1000Hz to ensure the timeliness and completeness of the data. During data acquisition, a Kalman filter algorithm is used to suppress noise and smooth the raw signal. The filter parameters are dynamically adjusted according to the signal characteristics, and the noise variance threshold is set to 0.01. Simultaneously, a frequency fluctuation dataset is established through a power grid frequency monitoring system. This system uses phase-locked loop technology to accurately track changes in the power grid frequency, recording frequency fluctuation data within a 1-second time window with a data accuracy of 0.01Hz. All acquired data, after data verification and outlier detection, is combined to form a real-time monitoring data vector, providing a reliable data foundation for subsequent power control.

[0050] The specific implementation of step S02 is to construct a multi-level power control benchmark system based on historical big data analysis. The system analyzes the historical operating records of the energy storage power station over the past 30 days using data mining technology, extracting stable power output characteristics under different load conditions, ambient temperatures, and grid conditions. A clustering analysis algorithm is used to divide the operating conditions into several typical categories, each corresponding to a certain standard power output value. Temperature segmentation thresholds are set at 10°C intervals, and load grading thresholds are set at 20% intervals of rated power. A power fluctuation feature matrix is ​​established using statistical methods, analyzing the mean, variance, skewness, and kurtosis of the power output data to identify the amplitude, frequency, duration, and periodicity characteristics of power fluctuations. Principal component analysis is used in the matrix construction process to reduce data dimensionality, retaining the main components with a cumulative contribution rate of 95%, providing standardized feature templates for power anomaly identification.

[0051] The specific implementation of step S03 involves power state assessment through intelligent deviation analysis and anomaly detection. The virtual inertia control algorithm uses the instantaneous frequency value as the main input parameter, combined with the standard power value in the power output stability reference matrix, to calculate the inertia response power of the energy storage system through differential equation solving. The algorithm dynamically adjusts the power output strategy based on the grid frequency change rate and amplitude, initiating a fast response mechanism when the frequency deviation exceeds 0.05Hz. The power deviation vector is obtained through vector operations and includes three components: active power deviation, reactive power deviation, and power factor deviation. The improved chessboard covering algorithm maps each element of the power deviation vector to chessboard grid points, employs a recursive divide-and-conquer strategy to analyze the connection patterns between grid points, and identifies abnormal patterns through a dynamic weight adjustment mechanism. The power anomaly indication value output by the algorithm is compared with a stability threshold, which is set at 5% of the rated power of the energy storage power station. If this threshold is exceeded, a power callback control process needs to be initiated.

[0052] The specific implementation of step S04 involves identifying the spectral characteristics and abnormal patterns of power fluctuations through frequency domain analysis. The system employs a Fast Fourier Transform (FFT) algorithm to perform frequency domain conversion on the power output time-series data, with the analysis frequency range set from 0.1Hz to 50Hz, covering the main power fluctuation frequency bands of the energy storage system. During the conversion process, a Hamming window function is used to reduce spectral leakage; the window length is selected based on signal characteristics, typically set to 1024 sampling points. Three abnormal fluctuation patterns are identified based on spectral amplitude and frequency distribution characteristics: sudden power output changes correspond to high amplitude and wide spectral characteristics; high-frequency oscillations manifest as energy concentration in certain frequency bands; and low-frequency oscillations show periodic energy distribution in low-frequency bands. Abnormal pattern discrimination uses a pattern recognition algorithm, determining the fluctuation type by comparing it with a standard spectral template. The abnormal fluctuation identification vector uses binary encoding, corresponding to three abnormal states: 1 indicates an anomaly has been detected, and 0 indicates a normal state, providing accurate state information for subsequent control strategies.

[0053] The specific implementation of step S05 involves optimizing power smoothing control parameters through intelligent prediction and adaptive adjustment. The adaptive filtering algorithm dynamically adjusts the filtering parameters based on power fluctuation characteristics, optimizes the filter coefficients using the minimum mean square error criterion, and adjusts the filter order between 8 and 32 depending on the complexity of the fluctuation. Feedforward compensation control establishes a compensation model based on the system transfer function, optimizing the system response through zero-pole configuration. The power trend prediction function, which includes neural network terms, employs a fast-response neural network architecture with a multi-head attention mechanism. The network input layer receives five parameters: current power value, power change rate, grid frequency, load change rate, and ambient temperature. The hidden layer uses residual connections and layer normalization techniques to improve training stability. The number of attention heads is dynamically adjusted according to the frequency fluctuation amplitude: 4 heads for fluctuations less than 0.1Hz, 8 heads for fluctuations between 0.1Hz and 0.3Hz, and 16 heads for fluctuations exceeding 0.3Hz. The network optimizes prediction accuracy through end-to-end training, generating a power trend prediction matrix containing predicted power output values ​​every 100 milliseconds within the next 3 seconds.

[0054] The specific implementation of step S06 is to achieve adaptive adjustment of energy storage system parameters based on abnormal states and prediction information. The system dynamically adjusts the charge / discharge rate and power smoothing intensity of the energy storage units according to the analysis results of the abnormal fluctuation indicator vector and power trend prediction matrix. Charge / discharge rate adjustment is achieved by controlling the current output amplitude of the energy storage units. During normal operation, it maintains a 1C rate. When the power anomaly indication value is within the range of the stable threshold to twice the stable threshold, it is adjusted to 0.8C. When it exceeds twice the stable threshold, it drops to 0.6C and an emergency power limiting mode is activated. Power smoothing intensity is controlled by adjusting the cutoff frequency of the adaptive filter. Under normal circumstances, the cutoff frequency is set to 10Hz. When there is one abnormal indicator in the abnormal fluctuation indicator vector, it drops to 5Hz. When there are two or more abnormal indicators, it is adjusted to 2Hz. The parameter adjustment process adopts a gradual adjustment strategy to avoid the impact of parameter abrupt changes on system stability. The adjustment rate is determined according to the severity of the anomaly, ensuring that the energy storage system maintains stable operation during parameter adjustment.

[0055] The specific implementation of step S07 involves achieving precise power output adjustment and system parameter updates through millisecond-level high-speed control technology. Millisecond-level power output control employs a high-speed digital signal processor to perform pulse-width modulation control on the inverter of the energy storage unit, with the control cycle set within the range of 1 to 10 milliseconds. Rapid power output response is achieved by precisely adjusting the amplitude and phase of the output voltage and current, with voltage regulation accuracy reaching 0.1% of the rated value and current regulation accuracy reaching 0.2% of the rated value. Power fluctuation suppression employs a multi-loop control strategy: the inner loop is a current control loop with a response time of approximately 1 millisecond, and the outer loop is a power control loop with a response time of approximately 5 milliseconds. Once the power output reaches a stable state, the system automatically updates the power output stability reference matrix, using a sliding window averaging method to update the standard power output value, with the window length set to the most recent 100 stable operating cycles. The completion of the reference matrix update marks the end of one precise control cycle, and the system continues into the next monitoring and control cycle, ensuring the continuous stability and precise control of the energy storage power station's power output.

[0056] It should be noted that the present invention adopts the following three key ideas.

[0057] The first is a collaborative anomaly detection technology combining virtual inertia control and an improved chessboard coverage algorithm. This technology integrates virtual inertia control and an improved chessboard coverage algorithm to achieve intelligent identification and rapid response to power anomalies in energy storage systems. The virtual inertia control algorithm simulates the inertial characteristics of a traditional synchronous generator, enabling the energy storage system to respond quickly to changes in grid frequency, offering advantages in speed and accuracy compared to traditional frequency regulation methods. The improved chessboard coverage algorithm maps power deviation data to a spatial geometric problem. Through a recursive divide-and-conquer strategy and a dynamic weight adjustment mechanism, it effectively identifies complex power anomaly patterns, overcoming the limitations of traditional threshold-based methods in recognizing complex anomaly patterns.

[0058] The second technique is the power trend prediction technology using a multi-head attention mechanism neural network. This technology employs a neural network architecture based on a multi-head attention mechanism to achieve accurate prediction of power trends. The multi-head attention mechanism can simultaneously focus on power change patterns across different time scales and frequency ranges, exhibiting stronger nonlinear modeling capabilities and temporal feature extraction capabilities compared to traditional linear prediction methods and simple neural networks. The dynamic adjustment mechanism for the number of attention heads in the network adaptively configures the model based on real-time frequency fluctuation amplitude and power fluctuation characteristics, enabling the prediction model to adapt to power change patterns under different operating conditions, significantly improving prediction accuracy and response speed.

[0059] Thirdly, there is the power fluctuation suppression technology that combines frequency domain analysis with time domain control. This technology integrates Fast Fourier Transform (FFT) frequency domain analysis with millisecond-level time domain control, enabling comprehensive identification and precise suppression of power fluctuations. Frequency domain analysis can accurately identify the spectral characteristics and abnormal patterns of power fluctuations, providing a scientific basis for control strategies. Compared with traditional time domain analysis methods, it has a stronger ability to identify fluctuation characteristics. Millisecond-level time domain control uses a high-speed digital signal processor to achieve rapid adjustment of power output. Its response time is much faster than traditional control methods, effectively suppressing power fluctuations across various frequency ranges and ensuring high stability of the energy storage system's output power.

[0060] The aforementioned key technological approaches, when organically combined, form a complete and precise power output control system for energy storage power stations. Virtual inertia control and an improved chessboard coverage algorithm provide an intelligent foundation for anomaly detection, while a multi-head attention mechanism neural network provides accurate trend prediction support for control decisions. The frequency- and time-domain combined control technology enables precise execution of fluctuation suppression. The synergistic effect of these three technologies achieves full-chain optimization from anomaly detection and trend prediction to control execution. Compared to existing single control methods, this approach offers higher control accuracy, faster response speed, and stronger adaptability, significantly improving the power output stability and control reliability of energy storage power stations in complex grid environments.

[0061] First, this invention addresses the technical problem of low power prediction accuracy in energy storage power stations. Traditional power prediction methods primarily rely on statistical analysis of historical data and simple linear regression models, which struggle to accurately capture the nonlinear characteristics of grid load changes and complex power fluctuation patterns, resulting in typically low prediction accuracy and limited prediction time windows. This invention constructs a power trend prediction function incorporating neural network terms and employs a deep learning architecture based on a multi-head attention mechanism. This architecture can simultaneously process multiple input parameters such as current power value, power change rate, grid frequency, load change rate, and ambient temperature. Residual connections and layer normalization techniques enhance the model's training stability. The multi-head attention mechanism dynamically adjusts the number of attention heads based on the fluctuation characteristics of real-time monitoring data, specifically addressing power change patterns across different time scales and frequency ranges, significantly improving the accuracy and timeliness of power prediction. For regional power grids, power conditions tend to be relatively stable over most of the time, so some power prediction can be made based on historical values. Although the accuracy is not high enough, it can be improved by adding a compensation term. In this scheme, a mathematical function is used to calculate a preliminary power prediction value with relatively low accuracy. Then, the neural network only processes the compensation term instead of the entire power prediction, which greatly improves the overall prediction efficiency and meets the millisecond-level control requirements.

[0062] Secondly, this invention solves the technical problem of delayed response in anomaly detection at energy storage power stations. Traditional anomaly detection methods typically employ fixed threshold judgments or simple statistical analysis methods, which have limited ability to identify complex power fluctuation patterns, low detection accuracy, and are prone to false alarms and missed alarms. When anomalies occur, they often cannot be detected and handled in a timely manner, leading to increased operational risks for energy storage systems. This invention maps each element of the power deviation vector to a chessboard grid point through an improved chessboard coverage algorithm. By analyzing the connection patterns and coverage paths between grid points, it identifies abnormal power output patterns. The algorithm employs a recursive divide-and-conquer strategy and a dynamic weight adjustment mechanism, combined with pattern matching based on anomaly event features recorded in a historical anomaly database. It can accurately identify various abnormal states such as sudden power output changes, high-frequency oscillations, and low-frequency swings, achieving millisecond-level anomaly detection response and providing a reliable guarantee for the safe and stable operation of energy storage systems.

[0063] Specifically, the principle of this invention is as follows: The fundamental principle that enables this invention to solve the problem of precise real-time control of power output fluctuations in energy storage power stations lies in the construction of a multi-layered power control system and an intelligent anomaly detection and prediction mechanism. First, by establishing a complete data foundation through real-time monitoring data vectors, a power output stability benchmark matrix, and a power fluctuation characteristic matrix, comprehensive state information is provided for precise control. The virtual inertia control algorithm simulates the inertial characteristics of a traditional synchronous generator, enabling the energy storage system to quickly respond to changes in grid frequency. The improved chessboard coverage algorithm maps the power deviation vector to chessboard grid points and analyzes the connection patterns, achieving accurate identification of power anomaly modes. Compared with traditional threshold detection methods, it has higher detection accuracy and stronger adaptability. Second, the Fast Fourier Transform analyzes the spectral characteristics of power fluctuations, accurately identifying three abnormal fluctuation modes: sudden changes, high-frequency oscillations, and low-frequency swings. This provides a scientific basis for subsequent control strategy selection. The power trend prediction function, which includes a neural network term, adopts a deep learning architecture with a multi-head attention mechanism, capable of handling power change modes at different time scales and frequency ranges. Compared with traditional linear prediction methods, it has stronger nonlinear fitting capabilities and more accurate prediction precision. Finally, the adaptive filtering algorithm dynamically adjusts the filtering parameters based on the abnormal fluctuation identification vector to achieve targeted suppression of different types of power fluctuations. The millisecond-level power output control uses a high-speed digital signal processor to perform pulse width modulation control on the inverter, which can complete power regulation in a very short time. The entire control system forms a complete closed loop from data acquisition, anomaly detection, trend prediction to precise control, ensuring high-precision real-time control of the power output of the energy storage power station.

[0064] 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.

[0065] The specific implementation of step S01 is to collect multi-dimensional operational data in real time through the power quality monitoring device of the energy storage power station, and establish a real-time monitoring data vector. Specifically, it is expressed as follows: ; In the formula, For real-time monitoring of data vectors; This is the effective value of the voltage, in volts (V). This is the effective value of the current, in amperes (A). This is the instantaneous frequency value, in Hz. Power factor; The surface temperature of the energy storage unit is given in °C. A frequency fluctuation dataset is also established. Specifically, it is expressed as follows: ; In the formula, This is a frequency fluctuation dataset; For the first The power grid frequency value at a given moment, in Hz; For the first Each sampling time, in seconds; The starting time is in seconds (s). The sampling interval is set to 1ms; The number of sampling points within a 1-second time window. ; This is the index for the sampling point number.

[0066] The parameter acquisition method is as follows: and The effective value is obtained through the power quality monitoring device's effective value calculation module, with a sampling frequency of 1000Hz. The frequency was obtained through real-time tracking using phase-locked loop technology, with a measurement accuracy of 0.01Hz. The power factor is obtained through a power factor measurement circuit, with a measurement range of 0 to 1. The temperature is measured by an infrared temperature sensor with an accuracy of ±0.1℃ and a measurement range of -20℃ to 80℃. The sampling time interval; This represents the total number of sampling points.

[0067] The specific implementation of step S02 is to construct a power output stability reference matrix based on historical operating data. Specifically, it is expressed as follows: ; In the formula, For the first Under what load conditions is the first Standard power output value under ambient temperature, in kW; The number of load condition levels, ; For temperature grade number, Power fluctuation characteristic matrix Specifically, it is expressed as follows: ; In the formula, This represents the average power fluctuation. For power fluctuation variance; Power fluctuation skewness; This refers to the kurtosis of power fluctuations.

[0068] The parameter acquisition method is as follows: The average power value under each operating condition was obtained by analyzing historical data from the past 30 days and using a clustering analysis algorithm. , , , The statistical characteristics of historical power data were obtained by calculating the statistical characteristics of the historical power data using statistical methods. The total number of load condition levels; This represents the total number of temperature classifications. For load condition indexes, the range is 1 to ; For temperature condition index, ranging from 1 to .

[0069] The specific implementation of step S03 is to calculate the power deviation vector using a virtual inertia control algorithm. Specifically, it is expressed as follows: ; In the formula, Actual active power, in kW; For reference active power, the unit is kW; Actual reactive power, in kvar; Reference reactive power, unit: kvar; This is the actual power factor; The power factor is used as a reference. The power calculation formula for the virtual inertia control algorithm is as follows: ; In the formula, The virtual inertial response power is expressed in kW. This is the virtual inertia constant, with a default value of 5s; The frequency change rate is expressed in Hz / s. This is the virtual damping coefficient, with a default value of 2. Frequency deviation, in Hz. Power anomaly indication value output by the improved checkerboard coverage algorithm. The calculation formula is as follows: ; In the formula, For the first The weighting coefficients of each deviation component, , , .

[0070] The parameter acquisition method is as follows: , , Acquired through real-time monitoring of data vectors; , , Obtained by looking up the power output stability reference matrix; It is obtained by numerical differentiation of the instantaneous frequency value; For the deviation component index, the range is 1 to 3; Here are the weighting coefficients, where Corresponding active power weight, Corresponding reactive power weight, Corresponding power factor weight; The first term represents the power deviation vector. Each component.

[0071] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0072] The specific implementation of step S05 involves calculating the power smoothing adjustment parameters and the power trend prediction matrix using an adaptive filtering algorithm. Specifically, it is expressed as follows: ; In the formula, For prediction Power output value at any given time, in kW; The prediction time step ranges from 0.1s to 3.0s, with a step size of 0.1s. A power trend prediction function incorporating neural network terms. Specifically, it is expressed as follows: ; In the formula, Based on the prediction function; Current power value, in kW; The power change rate is expressed in kW / s. The frequency is the power grid frequency, measured in Hz. Load change rate, in kW / s; Ambient temperature, in °C; This is a neural network compensation term.

[0073] The parameter acquisition method is as follows: Obtained through real-time monitoring data vector calculation; It is obtained by numerical differentiation calculation of power time series data; Obtained from frequency fluctuation datasets; Obtained through the power grid load monitoring system; Data is obtained through an ambient temperature sensor; The time step is for prediction, and the unit is seconds (s). This represents the compensation power value output by the neural network, in kW.

[0074] The specific implementation of step S06 is based on the abnormal fluctuation identification vector. The energy storage system parameters are adjusted using the power trend prediction matrix, and the charge / discharge rate adjustment formula is as follows: ; In the formula, This refers to the charge / discharge rate; To stabilize the threshold, it is set to 5% of the rated power. Power smoothing intensity is adjusted via the cutoff frequency. The calculation formula is as follows: ; In the formula, This represents the number of non-zero elements in the abnormal fluctuation identifier vector; This is the cutoff frequency of the adaptive filter, in Hz.

[0075] The parameter acquisition method is as follows: Obtained through step S03; Obtained through spectral analysis in step S04; Preset according to the rated power of the energy storage power station.

[0076] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0077] It should be explained that the power calculation formula of the virtual inertia control algorithm is based on the power system frequency response theory. This formula decomposes the power response of the energy storage system into two parts: the inertia response term and the damping response term.

[0078] The formula for calculating the inertia response term is: ; The formula for calculating the damping response term is: ; Inertia response term The damping response term reflects the system's response to the rate of frequency change and provides instantaneous power support when the frequency changes rapidly. The formula reflects the steady-state response of the system to frequency deviation, providing stable power regulation even when frequency deviation persists. Compared to traditional proportional-integral control methods, this formula better simulates the dynamic characteristics of synchronous generators, improving the frequency response performance and grid stability of energy storage systems. The power deviation vector calculation formula integrates multi-dimensional power deviation information into a unified mathematical expression through vector operations. This vector includes three components: active power deviation, reactive power deviation, and power factor deviation, comprehensively reflecting the operating state deviation of the energy storage system. Compared to a single power deviation index, this vector provides richer state information, offering a more accurate data foundation for anomaly detection and control decisions. The improved chessboard coverage algorithm's power anomaly indication calculation formula uses a weighted Euclidean distance mathematical form.

[0079] The formula for calculating the weighted Euclidean distance is: ; By setting different weighting coefficients for different deviation components, the importance of active power deviation can be highlighted, while also taking into account the influence of reactive power and power factor. Compared with the traditional simple threshold discrimination method, this formula can more accurately quantify the degree of power anomaly, improving the accuracy and reliability of anomaly detection. The power trend prediction matrix, in the form of a time series matrix, arranges the predicted power values ​​for the next 3 seconds in chronological order. This matrix provides detailed reference information for feedforward control. Compared with the traditional single-point prediction method, this matrix can provide more complete time-domain prediction information, enabling the control system to prepare in advance for power changes and significantly improving the predictability and accuracy of control. The power trend prediction function incorporating neural network terms combines the traditional basic prediction function with neural network compensation terms.

[0080] The formula for calculating the basic prediction function is: ; In the formula, to The linear coefficients of the basic prediction function are obtained by fitting historical data. The range is from 0.8 to 1.2. The range is 0.1 to 0.5 s. The range is -10 to 10 kW / Hz. The range is from 0.5 to 1.5. The range is -2 to 2 kW / ℃. The basic prediction function provides basic linear prediction capability, while the neural network compensation term provides nonlinear modeling and complex pattern recognition capabilities. Compared with traditional linear prediction methods, this function can better handle the nonlinear characteristics and complex patterns of power changes in energy storage systems, significantly improving prediction accuracy and adaptability. The charge / discharge rate adjustment formula adopts a piecewise function form, providing different adjustment strategies according to different degrees of power anomaly indication values. This formula can achieve adaptive parameter adjustment, and compared with fixed parameter control methods, it can better adapt to different operating conditions and anomaly degrees, improving the system's operational stability and safety. The power smoothing intensity adjustment formula achieves dynamic control of the power smoothing degree by adjusting the filter cutoff frequency. This formula determines the filtering intensity based on the number of anomaly types in the anomaly fluctuation identifier vector. Compared with fixed filter parameter methods, it can dynamically adjust the filtering effect according to the actual fluctuation situation, ensuring power smoothing while avoiding response delay caused by over-filtering.

[0081] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team needs to implement a precise power output control system for a 50MW lithium battery energy storage power station. This power station is connected to a 110kV power grid and undertakes grid frequency regulation and power smoothing tasks. The power station is equipped with 400 energy storage units, each with a rated power of 125kW, using lithium iron phosphate battery technology, with a total system capacity of 200MWh.

[0082] The technical team first deployed a real-time data acquisition system, which synchronously collects RMS voltage, RMS current, instantaneous frequency, power factor, and energy storage unit surface temperature at a frequency of 1000Hz using power quality monitoring devices. The voltage monitoring range is 380V to 420V, the current monitoring range is 0 to 200A, the frequency monitoring accuracy is 0.01Hz, the power factor monitoring range is 0.8 to 1.0, and the temperature monitoring range is -10℃ to 60℃. The noise variance threshold of the Kalman filter algorithm is set to 0.01 to ensure effective noise suppression of the original signal. The grid frequency monitoring system uses phase-locked loop (PLL) technology to record frequency fluctuation data within a 1-second time window. During normal operation, the grid frequency is maintained within the range of 49.95Hz to 50.05Hz.

[0083] Based on 720 hours of historical operating data from the energy storage power station over the past 30 days, the technical team constructed a power output stability benchmark matrix. Through cluster analysis, the operating conditions were divided into 15 typical categories, with ambient temperature segmented in 10°C intervals and load segmented in 20% intervals of rated power. As shown in Table 1, the standard power output values ​​under different operating conditions exhibit clear and regular variations.

[0084] Table 1 Standard power output values ​​under typical operating conditions

[0085] The power fluctuation characteristic matrix was established using statistical methods, and principal component analysis retained components with a cumulative contribution rate of 95%. The fluctuation amplitude characteristic range was 0.5MW to 5.0MW, the frequency characteristic range was 0.1Hz to 15Hz, the duration characteristic range was 1 second to 300 seconds, and the periodicity characteristic identified cyclic patterns ranging from 0.5 hours to 6 hours.

[0086] The virtual inertia control algorithm uses the instantaneous frequency value of 49.98Hz as the input parameter and combines it with the standard power value of 48.5MW to calculate the inertia response power demand. When the grid frequency drops by 0.05Hz to 49.93Hz, the algorithm instructs the energy storage system to release 2.5MW of power to support the grid frequency. The power deviation vector contains three components: active power deviation of 1.2MW, reactive power deviation of 0.3MVar, and power factor deviation of 0.02. The improved chessboard covering algorithm maps these deviation values ​​to 64×64 grid points, uses a recursive divide-and-conquer strategy to analyze the connection mode, and sets the initial weight of the dynamic weight adjustment mechanism to 1.0 with an adjustment step size of 0.1. The algorithm outputs a power anomaly indication value of 2.8MW, exceeding the stability threshold of 2.5MW, and initiates the power callback control process.

[0087] like Figure 3 As shown, the spectral characteristics of the power output data were analyzed using Fast Fourier Transform (FFT) with a 1024-point Hamming window, covering a frequency range of 0.1 Hz to 50 Hz. Spectral analysis identified a high-frequency oscillation mode, with the main energy concentrated in the 12 Hz to 18 Hz band, reaching an amplitude of 1.5 MW. The abnormal fluctuation flag vector was set to [0,1,0], indicating the detection of high-frequency oscillation anomalies.

[0088] The multi-head attention mechanism neural network receives the current power value of 48.2 MW, power change rate of 0.8 MW / s, grid frequency of 49.98 Hz, load change rate of 0.3 MW / s, and ambient temperature of 25°C as input parameters. Since the frequency fluctuation amplitude of 0.02 Hz is less than 0.1 Hz and the dominant fluctuation frequency of 12 Hz falls within the range of 5 Hz to 15 Hz, the number of attention heads is set to 8. The network uses a 4-layer hidden layer structure, with 256 neurons per layer. The residual connection and layer normalization parameters are set to 0.1 and 0.1, respectively. .like Figure 4 As shown, the power trend prediction matrix contains predicted power output values ​​for 30 time points within the next 3 seconds, with a prediction accuracy of 0.2MW.

[0089] Table 2 Neural Network Training Parameter Configuration

[0090] Based on the abnormal fluctuation indicator vector [0,1,0] and the power trend prediction results, the technical team adjusted the energy storage system parameters. Since the power anomaly indication value of 2.8MW falls within the range of the stability threshold of 2.5MW to twice the stability threshold of 5.0MW, the charge / discharge rate was adjusted from the normal 1C to 0.8C, corresponding to a decrease in current output from 200A to 160A. One anomaly indicator was found in the abnormal fluctuation indicator vector, so the adaptive filter cutoff frequency was reduced from the normal 10Hz to 5Hz, and the filter order was set to 16th. A gradual adjustment strategy was adopted, with an adjustment time constant set to 2 seconds to avoid the impact of sudden changes on system stability.

[0091] Millisecond-level power output control employs a dual-loop control strategy, with the high-speed digital signal processor's control cycle set to 5 milliseconds. The inner loop current control has a response time of 1 millisecond and a control accuracy of 0.2%, while the outer loop power control has a response time of 5 milliseconds and a control accuracy of 0.1%. Pulse width modulation (PWM) technology precisely adjusts the inverter output, with a voltage regulation range of 380V to 420V and a phase regulation accuracy of 0.1°. Power fluctuations are suppressed from an initial ±1.5MW to ±0.3MW, achieving a fluctuation suppression rate of 80%.

[0092] After 100 cycles of stable power output operation, the system automatically updates the power output stability reference matrix. A sliding window averaging method is used, with a window length of 100 cycles and an update weight of 0.95. The new standard power output value is updated from 48.5MW to 48.3MW, and the standard reactive power is updated from 9.2MVar to 9.0MVar. After the reference matrix update is complete, the system enters the next monitoring and control cycle, with a cycle period set to 10 seconds.

[0093] Throughout the control process, the technical team monitored the system's key performance indicators. Power output response time was reduced from 50 milliseconds using traditional methods to 8 milliseconds, frequency regulation accuracy improved from ±0.05Hz to ±0.02Hz, and power fluctuation suppression increased from 60% to 80%. The energy storage unit temperature was controlled below 35℃, battery charge / discharge efficiency remained above 95%, and the overall system availability reached 99.8%.

[0094] Compared to traditional energy storage power control methods, this invention simulates the inertial characteristics of a synchronous generator through a virtual inertia control algorithm, enabling the energy storage system to respond quickly to grid frequency changes and provide more stable grid support. The improved checkerboard covering algorithm maps complex power deviation patterns to a geometric structure, identifying abnormal patterns through grid connection analysis, achieving higher detection accuracy and lower false alarm rate compared to traditional threshold judgment methods. The multi-head attention mechanism neural network can simultaneously handle power change patterns across different time scales and frequency ranges, exhibiting stronger adaptability and prediction accuracy compared to single prediction models. Millisecond-level power output control achieves hierarchical regulation of current and power through a dual-loop control strategy, offering faster response speed and higher control accuracy compared to traditional single-loop control, fundamentally solving the power fluctuation problem of energy storage systems.

[0095] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0096] Table 3. Variable Explanation Table

[0097] 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 method for precise control of power output of an energy storage power station, characterized in that, Real-time voltage, current, frequency, power factor, and temperature data of the energy storage power station are collected to establish a real-time monitoring data vector. Simultaneously, grid frequency fluctuation data are collected to establish a frequency fluctuation dataset. Based on historical operating data, a power output stability benchmark matrix and a power fluctuation characteristic matrix are constructed. A power deviation vector is established by calculating the deviation between the current power output and the stability benchmark using a virtual inertia control algorithm. An improved checkerboard covering algorithm is used to analyze and process the power deviation vector. When the power anomaly indication value output by the improved checkerboard covering algorithm exceeds the stability threshold, a power callback control process is initiated. A fluctuation spectrum dataset is established by analyzing the power fluctuation spectrum characteristics using Fast Fourier Transform. The fluctuation type is determined based on the spectrum characteristics, and an abnormal fluctuation identification vector is established. Power smoothing adjustment parameters are calculated based on an adaptive filtering algorithm and feedforward compensation control. A power trend prediction matrix is ​​generated using a power trend prediction function containing a neural network term. The charge / discharge rate and power factor of the energy storage system are adjusted according to the abnormal fluctuation identification vector and the power trend prediction matrix. Millisecond-level power output control is implemented, and power fluctuation suppression is achieved by adjusting the voltage and current output of the energy storage units.

2. The precise control method for power output of an energy storage power station according to claim 1, characterized in that, The real-time monitoring data vector is obtained through the power quality monitoring device of the energy storage power station, specifically including the effective value of voltage, the effective value of current, the instantaneous value of frequency, the power factor and the surface temperature of the energy storage unit, with a sampling frequency of 1000Hz.

3. The precise control method for power output of an energy storage power station according to claim 2, characterized in that, The frequency fluctuation dataset is collected through a power grid frequency monitoring system. Specifically, it records the changes in power grid frequency within a 1-second time window, with a data accuracy of 0.01Hz.

4. The precise control method for power output of an energy storage power station according to claim 3, characterized in that, The power output stability benchmark matrix is ​​established based on the historical operating data of the energy storage power station over the past 30 days, specifically including stable power output reference values ​​under different load conditions, ambient temperatures, and grid conditions.

5. The precise control method for power output of an energy storage power station according to claim 4, characterized in that, The power fluctuation feature matrix is ​​established by analyzing the statistical characteristics of power output data, specifically including the amplitude, frequency, duration, and periodicity of power fluctuations.

6. The precise control method for power output of an energy storage power station according to claim 5, characterized in that, The power deviation vector is calculated by the difference between the current power output value and the corresponding value of the power output stability reference matrix, specifically including active power deviation, reactive power deviation and power factor deviation.

7. The precise control method for power output of an energy storage power station according to claim 6, characterized in that, The virtual inertia control algorithm is a control strategy that simulates the inertial characteristics of a traditional synchronous generator. It improves the frequency stability of the power grid by introducing a virtual inertia response mechanism into the control loop of the energy storage system. It uses the instantaneous frequency value in the real-time monitoring data vector as the main input parameter, and combines the standard power value in the power output stability reference matrix and the fluctuation mode information in the power fluctuation characteristic matrix.

8. The precise control method for power output of an energy storage power station according to claim 7, characterized in that, The improved chessboard coverage algorithm is used to analyze the abnormal pattern recognition of the power deviation vector. The input includes the power deviation vector, the power fluctuation feature matrix, the real-time monitoring data vector, the frequency fluctuation dataset, and the historical anomaly database. The output is the power anomaly indication value.

9. The precise control method for power output of an energy storage power station according to claim 8, characterized in that, The stability threshold is set at 5% of the rated power of the energy storage power station. When the power anomaly indication value exceeds the stability threshold, it is determined that adjustment is required.

10. The precise control method for power output of an energy storage power station according to claim 9, characterized in that, The power trend prediction function containing neural network terms is used to predict future power demand based on current power output data and grid load change trends. The inputs include current power value, power change rate, grid frequency, load change rate, ambient temperature, and neural network terms. The output is the predicted power output value within the next 3 seconds.

Images