A collaborative control method and system for energy storage converters

Abstract

This invention relates to energy storage technology and discloses a collaborative control method and system for an energy storage converter. The collaborative control method for the energy storage converter includes the following steps: acquiring multi-time-period electrical parameter measurements of the target energy storage converter's grid connection point through a data acquisition interface; dividing the electrical parameter measurements into a training dataset and a target dataset according to a program-defined partitioning rule; calling a covariance calculation module to calculate an autocovariance matrix based on the training dataset and a cross-covariance vector based on the training dataset and the target dataset; generating predicted values ​​of the target electrical parameters based on the autocovariance matrix and the cross-covariance vector using a prediction algorithm module; subtracting the predicted values ​​from the actual measured values ​​of the target dataset using a compensation processing module to obtain compensated electrical parameter commands; and controlling the energy storage converter to switch operating modes according to the compensated electrical parameter commands.

Description

Technical Field

[0001] This application relates to energy storage technology, specifically to a collaborative control technology for an energy storage converter, and particularly to a collaborative control method and system for an energy storage converter. Background Technology

[0002] With the increasing penetration of renewable energy and the growing complexity of power systems, energy storage technology has received widespread attention as a key means to balance supply and demand and improve grid stability. As the interface device between the energy storage system and the grid, the energy storage converter's control performance directly affects the system's power quality, operating efficiency, and security. Traditional energy storage converter control methods mainly include PID control and droop control based on fixed parameters. These methods perform well under steady-state conditions, but often suffer from problems such as response lag, poor anti-interference capability, and inflexible mode switching under complex scenarios such as grid transients, nonlinear load access, or distributed energy fluctuations.

[0003] For example, during grid voltage dips or frequency surges, traditional control methods struggle to quickly distinguish between disturbances and normal fluctuations, leading to oscillations in control commands and even system instability. Furthermore, existing methods handle grid quiescent states (such as light loads or the initial stages of islanding) rather crudely, easily misjudging due to measurement noise, causing unnecessary mode switching or power output, impacting equipment lifespan and potentially posing safety hazards. On the other hand, the lack of an effective coordination mechanism when multiple energy storage converters operate in parallel can easily lead to uneven power distribution and increased circulating current. Summary of the Invention

[0004] In view of this, the present invention provides a collaborative control method and system for energy storage converters to solve the technical problems mentioned in the background art.

[0005] The specific content includes:

[0006] A collaborative control method for an energy storage converter includes the following steps:

[0007] The electrical parameter measurement values ​​of the target energy storage converter grid connection point are obtained through the data acquisition interface;

[0008] The electrical parameter measurements are divided into a training dataset and a target dataset according to the division rules set by the program.

[0009] The covariance calculation module is invoked to calculate the autocovariance matrix based on the training dataset, and to calculate the cross-covariance vector based on the training dataset and the target dataset.

[0010] The prediction algorithm module generates predicted values ​​of the target electrical parameters based on the autocovariance matrix and the cross-covariance vector.

[0011] The compensation processing module subtracts the predicted value from the actual measured value of the target dataset to obtain the compensated electrical parameter command.

[0012] Based on the compensated electrical parameter instructions, the energy storage converter is controlled to switch operating modes, and charging and discharging power instructions are sent to the energy storage battery.

[0013] The compensation processing module outputs predefined electrical parameter commands to suppress power grid interference when the measured electrical parameter values ​​are in a silent state.

[0014] The compensation processing module includes a logic control unit, which is configured to output a predefined logic state to the control interface in a grid silent state, and the control interface receives a predefined electrical parameter instruction when the energy detection output is logic low and a predefined time is reached.

[0015] The logic control unit includes a time delay unit, which is used to trigger the output of a predefined command after the power grid remains in a silent state for a predefined time.

[0016] Furthermore, the division rule is a time window division strategy preset by the program, or an adaptive division strategy dynamically adjusted based on the characteristics of the power grid state.

[0017] Furthermore, the prediction algorithm module is a linear predictor, and its weight coefficients are obtained by multiplying the inverse of the self-covariance matrix with the cross-covariance vector.

[0018] Furthermore, the compensation processing module performs the following operations:

[0019] Input the predicted value and the actual measured value into the subtractor interface;

[0020] The compensated output signal serves as the source of control commands;

[0021] When the compensation processing module detects that the power grid is in a silent state, it outputs a predefined logical state to replace the actual measured value.

[0022] Furthermore, the operating modes include grid-following mode and grid-building mode, and the mode switching logic is automatically determined and executed by the program based on the compensated electrical parameter instructions;

[0023] The mode switching logic forces the system to enter grid-following mode in a grid-quiet state to suppress interference.

[0024] The logic control unit outputs a constant logic state in a silent state to avoid unnecessary mode switching.

[0025] Furthermore, before calculating the autocovariance matrix and crosscovariance vector, the training dataset is preprocessed, including mean reduction and normalization operations to eliminate DC components and unify the data scale.

[0026] The preprocessing also includes low-pass filtering to remove high-frequency noise.

[0027] The present invention also provides a cooperative control system for an energy storage converter, comprising:

[0028] The data acquisition unit is used to acquire electrical parameters of the grid connection point in real time;

[0029] The dataset partitioning unit is configured to divide the data into a training dataset and a target dataset according to the program's preset rules;

[0030] The covariance processing unit is used to calculate the autocovariance matrix and the cross-covariance vector;

[0031] The prediction and compensation unit is used to generate predicted values ​​and perform compensation processing.

[0032] The control decision unit is used to select the operating mode and generate power commands based on the compensation results.

[0033] The prediction and compensation unit includes a state detection module, which is used to output a predefined control command when a grid quiescent state is detected.

[0034] The state detection module includes a logic control circuit for outputting a predefined logic state in a silent state.

[0035] Furthermore, the prediction and compensation unit includes:

[0036] Pre-trained linear prediction model;

[0037] The real-time compensation interface is used to output the compensated control commands.

[0038] The time delay unit is used to trigger the output of a predefined command after the power grid has been in a state of continuous silence for a predefined time.

[0039] The logic control circuit includes AND gate or NAND gate structures.

[0040] Furthermore, the data acquisition unit includes a voltage transformer, a current transformer, and a phase-locked loop circuit, and its sampling accuracy and dynamic response speed meet the requirements for capturing transient processes of the power grid.

[0041] The data acquisition unit also includes a comparator and a low-pass filter for generating energy detection output.

[0042] Furthermore, the predefined time of the time delay unit can be remotely configured to avoid the silent state processing logic being mistakenly triggered due to short-term power grid disturbances.

[0043] The time delay unit is implemented using a counter or an RC circuit.

[0044] This invention employs a prediction and compensation mechanism based on covariance calculation. By calculating the autocovariance matrix of the training dataset and the cross-covariance vector with the target dataset, a linear prediction model is constructed to generate predicted values ​​for electrical parameters. Furthermore, compensation processing is used to eliminate grid disturbance components. The core of this mechanism lies in utilizing the statistical characteristics of historical data to predict future trends, thereby achieving proactive intervention in the grid's state.

[0045] By extracting the intrinsic correlation of signals through covariance calculation, normal fluctuations and abnormal disturbances can be effectively distinguished, thereby generating smoother and more accurate control commands. For example, when the grid voltage suddenly drops or the frequency fluctuates, the prediction algorithm can identify the trend in advance and quickly offset the disturbance component through compensation processing, enabling the converter to adjust its power output within milliseconds and avoid grid instability.

[0046] Although the linear predictor used has a simple structure, it can effectively capture the local features of the signal by dynamically adjusting the prediction order and weight coefficients, balancing computational efficiency and prediction accuracy. In addition, the recursive update mechanism of the covariance matrix ensures that the model can adaptively adjust to changes in the power grid state, avoiding the performance degradation of fixed-parameter models when operating conditions change.

[0047] Through compensation processing, noise and DC components in the actual measurements are effectively suppressed, and the control commands are closer to the actual grid requirements. In silent mode, the system automatically switches to predefined commands, avoiding frequent mode switching or power oscillations caused by minor fluctuations in measurements.

[0048] Covariance calculation can be used not only for prediction, but also for identifying grid resonance modes and evaluating power quality indicators.

[0049] Accurate identification and robust response to grid quiescent states are achieved through a logic control unit and a time delay unit. A quiescent state refers to a situation where electrical parameter measurements are within an extremely low fluctuation range, often occurring during periods of light grid load or the initial stages of islanding. In a quiescent state, by setting energy detection thresholds and a time delay unit, predefined command outputs are only triggered after the quiescent state has lasted for a certain period, avoiding the impact of short-term interference. For example, the time delay unit can be configured from 100ms to 10 seconds, effectively filtering out instantaneous fluctuations caused by lightning or switching operations. In a quiescent state, a zero-power command or rated voltage command is forcibly output, causing the converter to enter grid-connected mode or standby mode, complying with grid connection specifications and avoiding unplanned islanding operation. Simultaneously, the logic control unit achieves seamless command switching through a multiplexer, ensuring a smooth transition process. Predefined commands can be dynamically adjusted according to the requirements of the upper-level energy management system, supporting multiple control strategies. For example, in microgrid applications, a grid-connected mode can be switched to support local loads in a quiescent state; in distribution scenarios, a zero-power command can be selected to reduce losses.

[0050] This invention employs a pre-defined time window partitioning strategy or an adaptive partitioning strategy dynamically adjusted based on grid state characteristics to divide electrical parameter data into training and target datasets. The adaptive partitioning rule enhances adaptability to dynamic changes in the grid. While fixed time window strategies perform well when the grid is stable, they are prone to prediction errors due to data lag during transient processes such as load surges and fault recovery. The adaptive strategy dynamically adjusts the window length and position by analyzing the fluctuation characteristics of electrical parameters in real time, ensuring that the training data always reflects the current grid state. For example, when increased voltage fluctuations are detected, the system automatically shortens the training window to improve response speed; after the grid returns to stability, the window is extended to improve prediction accuracy. The adaptive partitioning strategy can adjust the computational load according to grid complexity, avoiding unnecessary complex calculations under simple operating conditions. For example, during light loads at night, the system can use a longer training window and a lower prediction order to reduce processor load; while during peak daytime loads, high-frequency updates and complex models are used to cope with rapid changes.

[0051] The system automatically switches between grid-following mode and grid-building mode based on compensated electrical parameter commands. The beneficial effects of this logic include: grid-following mode is suitable for stable grid conditions, with the converter tracking grid parameters; grid-building mode proactively establishes voltage and frequency support when the grid is weak or faulty. This invention achieves precise judgment of mode switching by setting a multi-dimensional stable operating range. For example, when the voltage deviation exceeds ±10% or the frequency change rate exceeds 1Hz / s, the system automatically switches to grid-building mode, effectively preventing grid collapse. The mode switching process employs a seamless transition strategy, with the grid-building control loop synchronizing its internal state in advance to ensure a shock-free switching moment. This soft switching mechanism avoids voltage spikes or current overshoots caused by traditional hard switching, protecting the converter and grid equipment. In silent mode, the system forcibly enters grid-following mode and outputs predefined commands, avoiding unplanned islanding operation. Simultaneously, the logic control unit blocks the transmission of abnormal measurement values, preventing mode oscillation. In parallel systems, each converter independently executes the mode switching logic and achieves synchronized action through communication, avoiding system oscillation or circulating current caused by inconsistent individual behavior. Attached Figure Description

[0052] Figure 1 A flowchart of the method provided by the present invention;

[0053] Figure 2 This is a schematic diagram of the system framework principle provided by the present invention. Detailed Implementation

[0054] Reference Figure 1 and Figure 2This invention provides a collaborative control method for an energy storage converter, comprising the following steps: acquiring multi-time period electrical parameter measurement values ​​of the target energy storage converter grid connection point through a data acquisition interface; dividing the electrical parameter measurement values ​​into a training dataset and a target dataset according to a partitioning rule set by the program; calling a covariance calculation module to calculate an autocovariance matrix based on the training dataset, and calculating a cross-covariance vector based on the training dataset and the target dataset; generating predicted values ​​of the target electrical parameters according to the autocovariance matrix and the cross-covariance vector through a prediction algorithm module; and subtracting the predicted values ​​from the actual measurement values ​​of the target dataset through a compensation processing module to obtain compensated electrical parameter commands. Based on the compensated electrical parameter commands, the energy storage converter is controlled to switch operating modes and charge / discharge power commands are sent to the energy storage battery. The compensation processing module outputs predefined electrical parameter commands to suppress grid interference when the measured electrical parameter values ​​are in a silent state. The compensation processing module includes a logic control unit configured to output a predefined logic state to the control interface in a grid silent state, and the control interface receives and outputs the predefined electrical parameter commands when the energy detection output is logic low and a predefined time has elapsed. The logic control unit includes a time delay unit for triggering the output of the predefined commands after the grid silent state has lasted for a predefined time.

[0055] In some embodiments, the measured values ​​include key electrical parameters such as voltage, frequency, and phase angle, which are continuously acquired at a fixed sampling frequency to form time-series data. The data acquisition interface typically consists of voltage transformers, current transformers, and phase-locked loop circuits to ensure high accuracy and dynamic response speed to capture transient processes in the power grid. The partitioning rules can be a preset time window strategy, for example, using the first N sampling points as training data and the subsequent M points as target data, or an adaptive strategy dynamically adjusted based on power grid state characteristics, for example, automatically adjusting the window size according to the degree of power grid fluctuations. The training dataset is used for model training, and the target dataset is used for actual prediction and compensation.

[0056] The autocovariance matrix reflects the correlation within the training data, while the cross-covariance vector characterizes the correlation between the training and target data. During computation, low-pass filtering is typically used for data preprocessing to eliminate DC components, standardize data scale, and remove high-frequency noise. The prediction algorithm module is a linear predictor, and its weighting coefficients are obtained by multiplying the inverse of the autocovariance matrix by the cross-covariance vector. The prediction order is dynamically adjusted based on the length of the training dataset and the characteristics of the power grid background noise to balance prediction accuracy and computational efficiency. A silent state refers to a state where electrical parameter measurements are between positive and negative energy detection thresholds (i.e., minimal fluctuation). When the logic control unit detects a silent state, it outputs a predefined logic state (such as constant rated voltage or zero power command) to the control interface, replacing the actual measured value. A time delay unit is used to trigger a predefined command output after the silent state has lasted for a predefined time, avoiding false triggering due to short-term disturbances. The mode switching logic is automatically determined and executed by the program: when the fluctuation of the compensated command exceeds the stable operating range of the grid-connected mode, it switches to the grid-building mode to actively stabilize the power grid; in the silent state, it forces entry into the grid-connected mode to avoid unnecessary switching.

[0057] In some embodiments, the partitioning rule is a pre-defined time window partitioning strategy or an adaptive partitioning strategy dynamically adjusted based on power grid state characteristics. The pre-defined time window partitioning strategy is a fixed rule, such as dividing continuously collected electrical parameter data into equal-length training and target segments in chronological order. The training segment is used to calculate the autocovariance matrix and train the prediction model, while the target segment is used to generate predicted values ​​and perform compensation. This strategy is simple and efficient, suitable for scenarios with stable power grid conditions. The window size is typically preset based on the sampling frequency and historical data characteristics; for example, selecting the first 100 sampling points as training data and the last 20 points as target data.

[0058] Adaptive partitioning strategies based on dynamic adjustment of grid state characteristics are more flexible and can adapt to the time-varying characteristics of the grid. This strategy dynamically adjusts the length or position of the training and target windows by monitoring the fluctuation characteristics of electrical parameters in real time (such as variance, spectral characteristics, or abrupt change detection). For example, when an increase in grid disturbance is detected, the training window is automatically shortened to improve model response speed; when the grid is stable, the training window is extended to improve prediction accuracy. Adaptive strategies are usually combined with sliding window or recursive update mechanisms to ensure that the model always runs based on the latest data. Both strategies are implemented through program algorithms, with the core being the balance between computational complexity and prediction performance. Fixed strategies have a lower computational burden but poorer adaptability; adaptive strategies have a higher computational load but can better cope with dynamic changes in the grid. In practical applications, these two strategies can be selected or used in combination depending on specific needs.

[0059] In some embodiments, the prediction algorithm module is a linear predictor, whose weight coefficients are obtained by multiplying the inverse of the autocovariance matrix by the cross-covariance vector. The linear predictor uses historical data to predict future values. Its mathematical model can be simplified to: the weight coefficient vector is equal to the product of the inverse of the autocovariance matrix and the cross-covariance vector. The linear predictor is typically a finite impulse response filter, and its output is a weighted sum of historical inputs. The prediction order is dynamically adjusted according to the length of the training dataset and the characteristics of the power grid background noise: it has good effects on quasi-periodic signals such as power grid voltage and frequency, effectively extracting signal trends and suppressing noise. The calculation of the weight coefficients must ensure that the autocovariance matrix is ​​invertible; this is usually avoided by regularization or recursive updates.

[0060] In some embodiments, the compensation processing module performs the following operations: inputting the predicted value and the actual measured value into the subtractor interface; outputting the compensated signal as the source of control commands; wherein, when the compensation processing module detects that the power grid is in a silent state, it outputs a predefined logic state to replace the actual measured value. The predicted value and the actual measured value are input into the subtractor interface, and a subtraction operation is performed to obtain the compensation signal. The compensation signal is the difference between the actual value and the predicted value, representing the disturbance component after removing the trend, and can serve as a purer source of control commands. When the compensation processing module detects that the power grid is in a silent state, it outputs a predefined logic state to replace the actual measured value. The predefined logic state can be a constant rated voltage, rated frequency command, or zero power command, the purpose of which is to stop the converter from power exchange and avoid introducing additional disturbances to the power grid. When the energy detection output (representing signal energy) is logic low for a predefined time, a command switch is triggered. The time delay unit is used to avoid short-term noise false triggering and ensure the reliability of the silent state determination.

[0061] In some embodiments, the operating modes include grid-following mode and grid-connecting mode. The mode switching logic is automatically determined and executed by the program based on compensated electrical parameter commands. Specifically, the mode switching logic forces the system into grid-following mode to suppress interference during grid quiescent conditions. The logic control unit outputs a constant logic state during quiescent conditions to avoid unnecessary mode switching. In grid-following mode, the converter tracks the grid voltage and frequency; in grid-connecting mode, the converter actively establishes and stabilizes grid parameters. The mode switching logic is automatically determined by the program based on compensated electrical parameter commands. When command fluctuations exceed the stable operating range of grid-following mode, for example, a voltage deviation exceeding ±10% or a frequency deviation exceeding ±0.5Hz, the system switches to grid-connecting mode to enhance grid stability. The stable operating range is comprehensively set based on the converter's rated capacity and the grid's allowable fluctuation range. During grid quiescent conditions, the logic control unit outputs a constant logic state, forcing the system into grid-following mode and suppressing mode switching, avoiding frequent switching due to minor fluctuations. This improves system reliability and reduces switching losses.

[0062] In some embodiments, the measured electrical parameters include at least one of voltage, frequency, and phase angle, and the multi-time period measured electrical parameters are time-series data continuously collected at a fixed sampling frequency; the determination of the silent state is based on the measured electrical parameters falling between positive and negative energy detection thresholds. The energy detection threshold is typically set as a percentage of the rated value, such as ±0.5% of the rated voltage. When the fluctuation of the measured value in multiple consecutive cycles is less than the threshold, it is determined to be a silent state.

[0063] In some embodiments, the training dataset is preprocessed before calculating the autocovariance matrix and crosscovariance vector. The preprocessing includes mean reduction and normalization operations to eliminate DC components and unify the data scale. The preprocessing also includes low-pass filtering to remove high-frequency noise.

[0064] In some embodiments, the linear predictor is a finite impulse response filter (FIR). Its prediction order is dynamically adjusted based on the length of the training dataset and the characteristics of the power grid background noise to achieve optimal prediction accuracy and computational efficiency. The process of dynamically adjusting the prediction order specifically includes: setting an initial maximum feasible order Pmax based on the length of the training dataset, typically set to 1 / 5 to 1 / 3 of the dataset length, to ensure sufficient data samples to reliably estimate the covariance matrix and avoid overfitting; introducing the Akaike Information Criterion as a trade-off between the performance and complexity of the prediction model. Its calculation considers not only the sum of squared prediction residuals but also includes a penalty term proportional to the model order to prevent the inclusion of power grid background noise in the model due to an excessively high order, which would decrease prediction accuracy in practical applications; the characteristics of the power grid background noise are obtained by analyzing the power spectral density of the training dataset. When the noise level is high or the spectrum is wide, the algorithm automatically tends to select a lower prediction order to enhance the model's performance. Robustness: The weight coefficients of the linear predictor are solved using the Levinson-Durbin recursive algorithm instead of direct matrix inversion. This algorithm is particularly suitable for autocovariance matrices, which can significantly reduce computational complexity, reduce the computational burden on the processor, and make real-time adjustment of the prediction order feasible. The system pre-sets an optimal order search range, for example, from 3 to Pmax. By iteratively calculating data values ​​at different orders, the system finally selects the order that minimizes the criterion function value as the optimal prediction order for the current control cycle. This dynamic adjustment process is executed once every one or every N training dataset update cycles to ensure that the prediction algorithm can adapt to changes in the power grid operating state, such as load switching, distributed energy output fluctuations, and system recovery after a fault, and always maintain optimal prediction accuracy and computational efficiency.

[0065] In some embodiments, the criteria for determining the silent state are: the fluctuation amplitude of the electrical parameter measurement values ​​for multiple consecutive sampling periods is less than a preset percentage threshold of its rated value, and the duration exceeds a preset silent time window; the silent time window is implemented by a time delay unit. The criteria for determining the silent state include: the fluctuation amplitude is calculated using a sliding window mechanism, calculating the standard deviation σ or peak-to-peak value of the electrical parameter measurement values ​​(taking voltage as an example) within M consecutive sampling periods of the silent time window, and comparing this fluctuation amplitude statistic with a dynamic threshold; the preset percentage threshold is not a fixed value, but is adaptively adjusted according to the historical operating data of the grid connection point, the rated voltage level of the grid, and the current operating mode. For example, in a weak grid environment, the threshold can be appropriately relaxed to avoid overly sensitive misjudgments; the predefined time T1 of the silent time window is implemented by the time delay unit, which consists of a configurable software counter or hardware timer, and its timing start condition is that the energy detection output first becomes logic low; only when the energy detection output remains logic low for the entire T1 time period. Furthermore, the silent state determination is only finalized when the fluctuation amplitude remains below the threshold, effectively avoiding false triggering caused by instantaneous power grid disturbances. The parameters of the time delay unit can be remotely modified online to adapt to the stability requirements of different power grid environments. In addition, the determination logic also incorporates hysteresis logic, where the threshold for recovering from the silent state to the normal state is slightly higher than the threshold for entering the silent state, to prevent the system from frequently switching near the state critical point. The energy detection output is generated by a dedicated hardware circuit, including a high-speed comparator (used to compare the sampled signal with a positive and negative energy detection threshold V+ and V-) and a subsequent low-pass filter circuit, the output of which is a binary logic signal. The preset percentage threshold of the rated value and the T1 time together constitute a highly reliable AND judgment condition, ensuring the accuracy and anti-interference capability of the silent state determination.

[0066] In some embodiments, the predefined logic state is a constant rated voltage and frequency command, or a zero-power command, used to stop the energy storage converter from exchanging active and reactive power with the grid; the logic control unit outputs the predefined logic state to the receiving data output interface in a silent state. The selection strategy and output mechanism of the predefined logic state include: the specific content of the predefined logic state is not static, but is dynamically configured by the upper-level energy management system based on the global grid status, the SOC level of the energy storage power station, and the current operating strategy; the strategy of selecting a constant rated voltage and frequency command is suitable for scenarios where the energy storage converter needs to transform into a grid-building source during grid silence, providing voltage and frequency support for a potentially isolated local network, in which case the converter is equivalent to a voltage source; while the strategy of selecting a zero-power command is suitable for scenarios where the grid disappears completely or the energy storage system needs to be completely offline, in which case the converter will stop all active and reactive power exchange and enter a standby or hibernation state to ensure equipment safety; the logic control unit internally includes a multiplexer. The controller's control terminal is driven by a silent state flag signal. In normal state, it selects the compensated electrical parameter command; in silent state, it switches to selecting a predefined command source. The predefined command source can be a built-in constant register or a setpoint register received from the upper-level system through the communication unit. The "output to receive data output interface" means that the logic control unit has completed the assignment and formatting of the final control command to make it conform to the interface protocol requirements of the lower-level power drive unit. In addition, when outputting the predefined logic state, the system will simultaneously lock out the conventional control algorithm based on predictive compensation and may initiate a gentle command transition process to avoid voltage or current surges on the converter output side caused by abrupt changes in the control command, ensuring the smoothness and stability of the mode switching process.

[0067] In some embodiments, the mode switching logic specifically involves: when the fluctuation of the compensated electrical parameter command exceeds the stable operating range of the grid-connected mode, controlling the energy storage converter to switch from the grid-connected mode to the grid-connected mode to actively establish and stabilize the grid voltage and frequency; the stable operating range of the grid-connected mode is set by the program, and its boundary value is determined comprehensively based on the rated capacity of the energy storage converter and the allowable fluctuation range of the grid; the logic control unit blocks the transmission of actual measured values ​​in a silent state.

[0068] The criteria and execution process of the mode switching logic include: the stable operating range of the grid-following mode is a multi-dimensional dynamic domain, whose boundary values ​​include not only the upper and lower limits of voltage and frequency, but also transient indicators such as voltage change rate, frequency change rate, and phase jump. These boundary values ​​together constitute a stable operating state; the comprehensive determination process needs to consider factors such as the overload capacity of the energy storage converter inverter unit, thermal design margin, power quality requirements stipulated by the power grid regulations, and short-circuit capacity at the connection point; when the compensated electrical parameter command continuously exceeds the state range for more than a preset delay, the control decision unit issues a command to switch from the grid-following mode to the grid-connecting mode; the switching process is not a simple and crude hard switch, but includes a seamless transition strategy: the grid-connecting control loop is advanced. The system calculates and gradually synchronizes its internal state, achieving a smooth handover of control output at the moment of switching, avoiding impact on the filter and the power grid. The forced entry into grid-following mode, which seems contradictory in the silent state, is actually a safety strategy. When the power grid is truly silent, blindly entering grid-following mode and supplying power to lines that may have been de-energized poses a safety risk of islanding operation, which does not comply with grid connection specifications. Therefore, in this state, forcibly maintaining grid-following mode and outputting zero-power commands, waiting for the power grid to recover or waiting for a clear islanding operation command from the superior system, is a safer choice. The logic control unit ensures that the commands received by the mode switching logic unit are stable predefined values ​​by blocking the transmission of actual measured values, thereby avoiding mode oscillation decisions that may be caused by fluctuations in its input signals.

[0069] This invention also provides a collaborative control system for an energy storage converter, comprising: a data acquisition unit for acquiring electrical parameters at the grid connection point in real time; a dataset partitioning unit configured to partition the data into a training dataset and a target dataset according to preset program rules; a covariance processing unit for calculating the autocovariance matrix and the cross-covariance vector; a prediction and compensation unit for generating predicted values ​​and performing compensation processing; and a control decision unit for selecting an operating mode and generating power commands based on the compensation results. The prediction and compensation unit includes a state detection module for outputting predefined control commands when a grid quiescent state is detected. The state detection module includes a logic control circuit for outputting predefined logic states in the quiescent state.

[0070] The state detection module is a mixed-signal circuit board integrating analog signal conditioning, digital logic processing, and a communication interface; its input comes directly from the high-speed ADC sampling signal of the data acquisition unit or a pre-processed electrical parameter data stream; the logic control circuit is a finite state machine implemented by an FPGA, whose states include normal monitoring, silent suspicion, silent confirmation, and silent processing; the conditions for state transitions are jointly determined by the logic level output of the energy detection, the timeout signal of the time delay unit, and the configurable threshold comparison result; the output of the predefined control command is converted into an analog signal by a high-precision DAC circuit. The command can be sent to the control decision unit via a digital bus in the form of data frames. The module also has a self-diagnostic function, which can monitor the health status of its internal reference voltage source, comparator and logic circuit, and send an alarm signal to the host computer when a fault occurs, while forcibly switching the output to a safe default state. All key parameters of the module, such as energy detection threshold (V+, V-), silent confirmation time (T1), hysteresis width, predefined command values, etc., are stored in non-volatile memory and can be remotely read and written through the communication unit, so that the system can flexibly adapt to the needs of different application scenarios without changing the hardware.

[0071] In some embodiments, the prediction and compensation unit includes: a pre-trained linear prediction model; a real-time compensation interface for outputting compensated control commands; and a time delay unit for triggering the output of predefined commands after the power grid remains silent for a predefined time. The logic control circuit includes AND gates or NAND gates. The pre-trained linear prediction model is an FIR digital filter whose weight coefficients can be updated online. Its coefficients are stored in a dual-port RAM, allowing the covariance processing unit to calculate and update a new set of coefficients in the background while the prediction algorithm in the foreground continues to use the old coefficients for real-time prediction. An atomic switch is performed after the new coefficients are calculated, thus achieving seamless updating of the prediction model. The real-time compensation interface is a high-precision digital subtractor that subtracts the predicted value from the measured value of the original target dataset, outputting the compensated control command. The time delay unit in the FPGA is typically implemented by a clock-driven decrementing counter, whose initial value is determined by a configured silence time window. The counter starts counting when the energy detection signal goes low. If the signal returns to a high level before counting to zero, the counter is reset. If the count reaches zero, a timeout interrupt signal is generated, triggering the state machine to enter a silent confirmation state. The AND or NAND gate structure in the logic control circuit is part of the output logic of the finite state machine and is used to generate the final control command selection signal. The entire prediction and compensation unit is designed as a high-throughput pipeline structure. Data acquisition, covariance calculation, matrix operation, prediction, compensation, and logical judgment are performed in parallel over multiple clock cycles to ensure that all processing is completed before the next sampling point arrives, meeting the stringent timing requirements of real-time control.

[0072] In some embodiments, the data acquisition unit includes a voltage transformer, a current transformer, and a phase-locked loop circuit, whose sampling accuracy and dynamic response speed meet the requirements for capturing transient processes of the power grid; the data acquisition unit also includes a comparator and a low-pass filter for generating energy detection output. Voltage and current transformers employ high-linearity, wide-bandwidth Rogowski coils or optical sensors to ensure accurate transmission of primary-side signals even during grid distortion or high-frequency oscillations. Subsequent signal conditioning circuitry includes an anti-aliasing filter, a programmable gain amplifier, and a synchronous sampling ADC array to guarantee simultaneous multi-channel sampling. The phase-locked loop (PLL) circuit uses an enhanced software PLL based on a second- or third-order adaptive filter, with adjustable bandwidth. It employs a narrow bandwidth to filter noise when the grid is stable and automatically widens the bandwidth for rapid tracking when grid disturbances occur. The comparator is a high-speed hysteresis comparator, with its positive and negative thresholds (V+, V-) set by a high-precision digital-to-analog converter, enabling programmable control of the energy detection threshold. The low-pass filter smooths the digital pulse signal output by the comparator; its extremely low cutoff frequency filters out glitches and generates a stable energy detection flag signal. The group delay of the entire data acquisition link is precisely calibrated and compensated to ensure accurate time stamps for all electrical parameter measurements, providing a high-precision spatiotemporal consistency data foundation for subsequent collaborative control algorithms.

[0073] In some embodiments, the covariance processing unit is implemented by a digital signal processor or a field-programmable gate array (FPGA) to perform matrix inversion and vector multiplication operations at high speed. The covariance processing unit further includes a noise filtering function. The covariance processing unit utilizes a Virtex series digital signal processor with a floating-point unit to specifically handle the calculation of the covariance matrix and vectors. To improve computational efficiency, the algorithm is optimized for the symmetric structure of the covariance matrix, employing a recursive matrix inversion algorithm to avoid the computationally intensive and numerically unstable direct Gaussian elimination method. The noise filtering function is implemented in the preprocessing stage, including wavelet threshold denoising or Wiener filtering of the training dataset to suppress the influence of measurement noise before calculating the covariance, thereby improving the signal-to-noise ratio of subsequent predictions. Data is transferred between the covariance processing unit and the prediction algorithm module via a high-speed DMA channel, and the calculated optimal weight coefficients are directly written into the coefficient memory of the prediction FIR filter. This unit also monitors the condition number of the matrix; if the matrix is ​​found to be close to singular, a small regularization parameter is automatically injected to ensure the numerical stability of the solution and prevent control command divergence due to data anomalies.

[0074] In some embodiments, the system further includes a communication unit for receiving scheduling instructions from the upper-level energy management system and uploading the operating mode, power instructions, and grid status information; the communication unit supports remote configuration of logic control parameters. The communication unit adopts a multi-mode design and supports at least industry-standard protocols such as Ethernet and fiber optic serial communication to ensure reliable interconnection with upper-level energy management systems from different manufacturers; its functionality goes far beyond simple instruction reception and data uploading, implementing a complete configuration and management interface: remotely configured parameters cover logic control parameters (such as silence determination threshold and time delay), prediction algorithm parameters (such as prediction order and training dataset length), control mode parameters (such as grid-connected / grid-connected switching threshold), and protection settings.

[0075] In some embodiments, the scheduling instructions of the upper-level energy management system are used to override the locally generated charging and discharging power instructions, or to modify the partitioning rules, the parameters of the prediction algorithm module, and the predefined electrical parameter instructions; the predefined electrical parameter instructions include the logical state types output in the silent state. The dispatch instructions from the upper-level system can be divided into two categories for local control: one is setpoint instructions (such as charging and discharging power instructions and voltage setpoints), and the other is parameter configuration instructions. For setpoint instructions, the system is designed with a multi-source instruction arbitration logic: under normal circumstances, instructions generated by the local collaborative control algorithm have the highest priority to achieve rapid grid support; when the upper-level system issues an instruction and its priority flag is high, such as a dispatch plan or an emergency frequency regulation request, the upper-level system instruction will override the local instruction; flexible switching can also be performed using weighted averaging or selectors; for parameter configuration instructions, the upper-level system can remotely modify the division rules, adjust the parameters of the prediction algorithm module, and even redefine the logic state type output in the silent state, so that the entire collaborative control strategy can be dynamically optimized according to the long-term evolution of the grid without replacing hardware; all instructions from the upper-level system undergo validity verification and range checks to prevent illegal values ​​from entering the control system.

[0076] In some embodiments, the predefined time of the time delay unit can be remotely configured to avoid the silent state processing logic being mistakenly triggered due to short-term power grid disturbances; the time delay unit is implemented through a counter or RC circuit. The predefined time of the time delay unit is stored in a read / write register of non-volatile memory, and its value is usually in milliseconds or seconds, ranging from 100ms to 10 seconds or even longer depending on the application scenario. Remote configuration is completed by receiving configuration frames through the communication unit. The new delay value takes effect immediately after verification or at the start of the next silent event. The counter is implemented based on the system's high-precision clock, and its count value is calculated based on the clock frequency and the set T1 time. In the analog RC circuit implementation scheme, the time constant is changed by switching resistors with different resistance values, thereby achieving digital controllability of the delay value. In order to further enhance the anti-interference capability and avoid short-term disturbances from falsely triggering the silent state, a short-time pulse interference suppression circuit or digital filter can be added before the time delay unit. This filter has an extremely short delay (e.g., <1ms) and can filter out microsecond to millisecond-level spike pulses, ensuring that only a continuous voltage disappearance or drop will trigger the subsequent delay counting logic.

[0077] In some embodiments, the control decision unit further includes a protection submodule, which triggers an alarm and executes a safety shutdown procedure when it detects that the compensated electrical parameter command is continuously abnormal or the mode switching is frequent; the protection submodule is linked with the logic control unit and does not trigger abnormal alarms in the silent state. The protection submodule is an independent multi-level protection system. Its inputs include compensated electrical parameter commands, mode switching signals, power device temperatures, DC-side voltage, etc. Its judgment of continuous anomalies is based on multiple dimensions: first, the voltage command exceeds the safety hard limit; second, it is implemented through multiple different delay timers, with short delays triggering warnings and long delays triggering faults; and third, a continuous monotonically increasing command indicates potential loss of control. When frequent mode switching between grid connection and grid connection is detected, the protection submodule determines that the system may be in a critical unstable state, thereby triggering an alarm and executing a safety shutdown procedure. This procedure does not immediately cut off power, but first attempts to switch to a fixed, conservative control mode. If stability is still not achieved, the power is smoothly reduced to zero according to a preset slope. The linkage with the logic control unit is crucial: in the silent state, the commands received by the protection submodule are stable predefined values, so it automatically relaxes its anomaly detection criteria or temporarily blocks certain alarms based on command fluctuations, preventing unnecessary false alarms and shutdowns under the special condition of grid silence, thereby improving the system's availability and reliability. The protection system will only be triggered unconditionally when an absolute anomaly, such as a hardware circuit fault, is detected.

[0078] In some embodiments, in a parallel system comprising multiple energy storage converters, each converter independently executes coordinated control based on its own grid connection point electrical parameters and coordinates power distribution through communication. The logic control units of each converter uniformly output predefined logic states in a silent state to avoid system oscillation. In a parallel system comprising multiple energy storage converters, independent execution means that each converter controller independently collects its grid connection point electrical parameters and independently runs a complete coordinated control system, thereby achieving distributed decision-making, independent of a single central controller, and improving system redundancy and reliability. However, to achieve power balance and avoid circulating current among parallel units, they need to coordinate communication through a high-speed communication network, exchanging their respective operating status, available capacity, and local command information. A typical coordination strategy is to use a distributed consensus algorithm based on average power, where each converter iteratively adjusts its own power output reference value based on information from neighboring nodes, ultimately achieving total power indexing. The power supply is allocated proportionally among the units. For the silent state, in addition to receiving local status detection results, the logic control unit of each converter also receives a system-level silent state flag via the communication bus. This flag is generated by a master node or through a voting mechanism. Once the system-level silent state is established, the logic control units of all converters will uniformly and synchronously switch to the predefined output logic state. This synchronization is crucial, ensuring that all converters behave consistently during grid silence, either simultaneously switching to grid-connected mode to support the islanded grid or simultaneously switching to standby mode, completely avoiding system oscillations, circulating currents, or even equipment damage caused by misjudgments or asynchronous actions of individual units. Communication latency is taken into account, and the time difference between nodes is minimized as much as possible through a time synchronization protocol.

[0079] It should also be noted that this application can also be implemented in the following ways.

[0080] In some embodiments, the output of the logic control unit is connected to the mode switching logic input, such that when the logic control unit detects a grid quiescent state, the predefined logic state it outputs directly triggers the mode switching logic, forcing the energy storage converter into grid-connected mode to suppress grid interference. The mode switching logic executes the mode switch immediately after the logic control unit outputs the predefined logic state. In a grid quiescent state, such as the initial stage of islanding or after a line disconnection, the electrical parameters (voltage, frequency) at the grid connection point may exhibit extremely low, near-zero noise signals, rather than being completely zero. In this case, relying on traditional mode switching logic based on the amplitude or rate of change of compensated electrical parameters, for example, determining whether the voltage is below 80% of the rated value, may cause the decision-maker to oscillate repeatedly at the critical point due to random fluctuations in the noise signal: sometimes determining a grid fault and entering grid connection mode, and sometimes determining that the grid has returned to normal and returning to grid-connected mode. Such frequent mode switching can generate enormous electrothermal stress on the converter's power devices and even cause output current surges, endangering equipment safety. More importantly, during a true grid outage and quiet period, blindly entering grid connection mode and supplying power to lines that have been de-energized will create unplanned islands, which seriously violate grid connection standards. Therefore, there must be an absolutely reliable mechanism that goes beyond conventional parameter criteria to force the system into a safe state (i.e., grid connection mode).

[0081] In the above, direct triggering means that the logic control unit's output signal in a predefined logic state is given higher priority and directly accesses the decision core of the mode switching logic. At the hardware level, this is expressed as a high or low level signal on the FPGA. When the logic control unit, based on its internal state machine (e.g., after a state transition from normal monitoring -> silent suspicion -> silent confirmation), finally determines that the silent state is established, this pin outputs a specific logic level; for example, a high level represents forced following.

[0082] The software interrupt service routine or hardware logic circuit of the mode switching logic monitors the state of this pin in real time. Once the forced grid connection signal is detected to be valid, the mode switching logic will immediately interrupt its current complex algorithm judgment process based on compensation instructions and unconditionally execute a forced switching instruction to set the converter's control loop structure, reference value source, and other internal parameters to grid connection mode. The delay of this trigger path is extremely short, much faster than the judgment cycle based on software algorithms, thus ensuring that the system's response to the silent state is near instantaneous.

[0083] In some embodiments, the time delay unit of the logic control unit is configured in conjunction with the mode switching logic, such that when the grid quiescent state lasts for a predefined time, the time delay unit simultaneously triggers a predefined electrical parameter command output and a mode switching command, and the mode switching logic forces the energy storage converter to enter the grid-connected mode according to the command.

[0084] In the above, the time delay unit is used to start timing when the energy detection output of the data acquisition unit first goes low, indicating that the signal energy is below the threshold. However, at this time, the system is only in a silent suspicion state. Only when the energy detection signal remains low for the entire predefined time, and the time delay unit expires, will a timeout event be generated. This timeout event is the trigger for the entire coordinated action.

[0085] The key is simultaneous triggering. In hardware design, this timeout signal enables two parallel processes simultaneously: Process 1: Instruction switching; triggering a multiplexer to switch its output from the compensated electrical parameter instruction channel to a predefined electrical parameter instruction channel, such as zero-power instruction or rated voltage instruction. Process 2: Mode switching; directly sending the aforementioned forced grid connection signal to the mode switching logic. This simultaneity is guaranteed by sharing the same timeout signal and synchronous hardware logic (e.g., executing on the same clock edge of the FPGA). It ensures that at the same moment a new control instruction takes effect, the converter's operating mode has also switched to the grid connection mode, which can correctly interpret and execute the instruction.

[0086] The predefined time T1 acts as a filter, effectively distinguishing between continuous grid silence and brief grid disturbances. The latter typically lasts much shorter than T1, thus avoiding triggering subsequent actions and preventing malfunctions. If command switching and mode switching are not synchronized—for example, if the command switches to zero power first, but the mode remains in grid-connected mode—the grid-connected mode may attempt to interpret the zero-power command as an invalid value or produce unpredictable behavior. Conversely, if the mode switches first, but the command is still a compensated fluctuation value, the grid-connected mode may attempt to track these noises, leading to unstable output. The "simultaneous triggering" in this claim ensures a high degree of intrinsic consistency between the command system and the control mode, avoiding logical conflicts between subsystems. The predefined time T1 is a remotely configurable key parameter, allowing maintenance personnel to adjust it according to the specific characteristics of the grid and safety procedures. In grids with high stability requirements, a longer T1 can be set to maximize disturbance immunity; in scenarios requiring rapid islanding detection, a shorter T1 can be set.

[0087] In some embodiments, the state detection module of the prediction and compensation unit is communicatively connected to the control decision unit. The state detection module includes a logic control circuit. When a grid quiescent state is detected, the logic control circuit outputs a predefined logic state to the control decision unit, directly controlling the switching of the operating mode and forcing the energy storage converter to enter the grid-connected mode.

[0088] In the above description, the prediction and compensation unit is a complex core processing unit. Separating the state detection module from this unit means that the system delegates the critical task of grid state identification to a dedicated hardware or firmware module. This module integrates analog signal conditioning, digital logic processing, and timing functions. The control decision unit, on the other hand, focuses more on generating the final power device drive signals based on predetermined control laws and mode switching strategies. The state detection module can be optimized for specific needs of silent state identification, for example, by using high-speed comparators, configurable hysteresis circuits, and high-precision timers, resulting in performance and reliability far exceeding those achieved through software simulation on a general-purpose processor. The control decision unit does not need to continuously poll the original electrical parameters to determine whether it is silent; it only needs to wait for explicit digital instructions from the state detection module, thus concentrating valuable computing resources on complex power control algorithms. The state detection module is typically implemented using programmable logic devices such as FPGAs or CPLDs, with response times in the nanosecond to microsecond range, enabling extremely fast state determination and signal generation, ensuring rapid system awareness of silent states.

[0089] Communication connections refer to high-speed, reliable data exchange channels between boards or chips within a remote system. Specific implementation methods include: Parallel digital bus: The status detection module encodes a "predefined logical state" into a multi-bit parallel signal (e.g., a 2-bit signal can represent "00 - Normal", "01 - Forced Grid Connection - Zero Power", "10 - Forced Grid Connection - Rated Voltage"), which is directly sent to the I / O port of the control decision unit via physical connections. High-speed serial bus: Such as SPI or I2C, the status detection module acts as a slave device, actively interrupting the control decision unit when the state changes, and transmitting status information through data frames. This method requires fewer connections, but the protocol is slightly more complex. Shared memory: Within the FPGA, the logic control circuit of the status detection module writes the status flag into a memory map shared by the processor core of the control decision unit. The state is stored in the register. The processor acquires this state through periodic readings or interrupt triggering. When the logic control circuit determines that a silent state is established, the predefined logic state it issues is a command with the highest priority. The control decision unit must treat this as an absolute instruction, unconditionally interrupting the current task and forcibly switching to grid-connected mode. This modular design makes the entire control system more deterministic and secure in the face of power grid anomalies. The state detection module acts like a sentinel, specifically responsible for monitoring a specific dangerous situation. Once a danger is detected, it reports directly to the control decision unit via a dedicated alarm line without going through complex software algorithms, and the commander immediately issues a mandatory action command.

[0090] Furthermore, this design facilitates system self-diagnosis and fail-safe operation. For example, the status detection module can have a built-in self-test circuit to periodically check the accuracy of its comparator reference voltage and the normality of its logic circuits. Once a fault is detected, it can send a fault signal to the control decision unit, allowing the system to switch to a backup, conservative control strategy. This prevents the entire system from malfunctioning or losing its protective functions due to the failure of the detection module.

[0091] In some embodiments, the configurable predefined time of the time delay unit is synchronized with the silent state detection time of the state detection module, and the output of the time delay unit is used to trigger the output of predefined control commands and mode switching commands simultaneously, ensuring that the control decision unit consistently executes the network following mode in the silent state.

[0092] In the above, synchronization refers to the alignment and coordination of logical timing. The time base used by the entire determination process of the state detection module (from the start of timing when the energy detection signal goes low, to the time delay unit timeout, and finally the generation of the silent state flag) is unique and centrally configurable. All subsequent operations in the system that depend on the silent state determination have a waiting time based on the same T1 parameter, assumed to be 2 seconds. Inconsistencies such as a 2-second wait for instruction switching and a 1.5-second wait for mode switching are not allowed. This synchronization is achieved through system design: the "timeout" signal generated by the time delay unit is the unique and global trigger source. When the timeout signal is valid, the processing logic inside the control decision unit will receive this event simultaneously. Subsequently, the task responsible for instruction management will switch the instruction source, while the task responsible for mode management will perform mode switching. Because they respond to the same trigger event, and the design ensures that the processing delay is minimal and controllable, from the outside of the system, the converter enters the grid-connected mode and begins executing predefined instructions almost simultaneously. Designing the predefined time T1 of the time delay unit as configurable and emphasizing its synchronization with the entire silent detection time greatly enhances the system's adaptability and intelligence. Different grid application scenarios have different requirements for islanding detection time. For example, some grid connection standards require islanding to be detected and power supply to be stopped within 2 seconds. Through a remote communication interface, maintenance personnel can easily set T1 to a value conforming to local specifications without modifying hardware or firmware code. In more advanced implementations, T1 may not even be a fixed value, but rather dynamically adjusted by the upper-level energy management system based on the real-time operating status of the grid. For example, during periods when the grid is very vulnerable, T1 can be appropriately shortened to allow for faster implementation of safety measures; during periods when the grid is stable, T1 can be extended to further improve disturbance rejection capabilities. This dynamic adjustment requires the implementation of the time delay unit to accept online parameter updates and ensure the integrity of timing logic during the update process.

[0093] In some embodiments, the update of the weight coefficients of the linear predictor of the prediction algorithm module is controlled by the state of the logic control unit. When the logic control unit detects a grid quiescent state, it freezes the update of the weight coefficients until the quiescent state is lifted, so as to maintain prediction stability.

[0094] In the above, the prediction algorithm operates based on the statistical characteristics of historical data. Its basic assumption is that historical data can reflect future trends to some extent. This assumption holds true when the power grid is operating normally and the data has a certain degree of continuity and correlation. However, in a quiescent state, the electrical parameters measured at the grid connection point are no longer valid power grid state information, but may be measurement noise, ground potential drift, or extremely small induced voltages. If, under these circumstances, the prediction algorithm continues to recursively update its weight coefficients based on this invalid data, it will lead to serious consequences: the algorithm will learn the characteristics of the noise, causing the weight coefficients to deviate from normal values, becoming meaningless or even harmful. When the power grid recovers, a contaminated prediction model cannot immediately make accurate predictions and requires a period of reconvergence, during which control performance will degrade.

[0095] In the above scenario, the logic control unit, acting as the system's state sensor, is the first to detect the establishment of a grid quiescent state. Besides sending instructions to the mode switching logic, it simultaneously sends a freeze signal to the prediction algorithm module (specifically, the covariance processing unit responsible for updating the weight coefficients). Upon receiving the freeze signal, the covariance processing unit immediately suspends its internal recursive update algorithm for the covariance matrix. It stops collecting new training data, maintains the current set of weight coefficients unchanged, and ceases new matrix inversions and coefficient calculations. During the quiescent period, the prediction algorithm module continues to use the frozen set of coefficients for prediction. When the logic control unit detects that the grid has recovered, it removes the freeze signal. The covariance processing unit then resumes collecting new, valid grid data from where it left off, gradually updating the weight coefficients to allow the prediction model to readjust to the normal grid state.

[0096] In some embodiments, the data acquisition unit includes a comparator and a low-pass filter to generate an energy detection output. The energy detection output is connected to the logic control circuit of the prediction and compensation unit. When the energy detection output is low for a predetermined time, the logic control circuit outputs a predetermined logic state, triggering the control decision unit to switch to network-following mode.

[0097] In the above, the data acquisition unit includes a voltage / current transformer, an ADC, and a comparator and a low-pass filter for generating the energy detection output. The comparator is a high-speed operational amplifier that compares its input electrical parameter signal with a preset positive and negative energy detection threshold (V+, V-). The comparator outputs a low level when the input signal amplitude remains between these two thresholds; otherwise, it outputs a high level. The comparator's output is a digital pulse signal containing numerous glitches. A low-pass filter with an extremely low cutoff frequency is used to smooth this signal. Its function is to filter out short-lived noise pulses. Only when a low-level signal persists for a sufficiently long time will the filter's output stably become logic low. This "filtered" signal is the actual energy detection output.

[0098] This path constitutes a two-stage filtered logical judgment: The first stage is hardware filtering: a comparator + low-pass filter, which determines whether there is a low-energy signal lasting for a sufficiently long time. This filters out transient interference in the microsecond to millisecond range. The second stage is digital logic / software delay: a time delay unit (T1) in the logic control circuit, which determines whether the low-energy signal has lasted for a predefined time. This filters out short-term disturbances in the second range. Only when both of these conditions are met simultaneously, i.e., (energy detection output is logic low) and (duration > T1), is the silent state finally confirmed. This double-protection mechanism greatly improves the anti-interference capability and reliability of the judgment.

[0099] The energy detection output, as the source signal for the entire silent state determination, is of paramount importance in terms of quality and reliability. It is directly connected to the logic control circuit of the prediction and compensation unit. The state machine inside the logic control circuit continuously monitors the level of this pin. Only when the complete judgment logic chain described above is true does the logic control circuit finally output the crucial predefined logic state. This state signal propagates upwards, triggering a series of coordinated actions as described earlier: instruction switching, mode switching, and even prediction coefficient freezing.

Claims

1. A method of coordinated control of energy storage converters, characterized by, The method comprises the following steps: obtaining multi-period electrical parameter measurement values of a target energy storage converter grid-connected point through a data acquisition interface; dividing the electrical parameter measurement values into a training data set and a target data set through a program-set division rule; calling a covariance calculation module, calculating a self-covariance matrix based on the training data set, and calculating a cross-covariance vector based on the training data set and the target data set; generating a predicted value of a target electrical parameter through a prediction algorithm module according to the self-covariance matrix and the cross-covariance vector; subtracting the predicted value from an actual measurement value of the target data set through a compensation processing module to obtain a compensated electrical parameter instruction; controlling the energy storage converter to switch operation modes and issuing a charge-discharge power instruction to the energy storage battery according to the compensated electrical parameter instruction; wherein the compensation processing module outputs a predefined electrical parameter instruction to suppress grid interference when the electrical parameter measurement value is in a silent state; the compensation processing module comprises a logic control unit configured to output a predefined logic state to a control interface in a grid silent state, and the control interface outputs the predefined electrical parameter instruction when the energy detection output is logic low and the predefined time is reached; the logic control unit comprises a time delay unit for triggering the output of the predefined instruction after the grid silent state lasts for a predefined time; the operation modes include a grid-following mode and a grid-forming mode, and mode switching logic is automatically executed according to the compensated electrical parameter instruction by a program; wherein the mode switching logic is forced to enter the grid-following mode to suppress interference in a grid silent state; the logic control unit outputs a constant logic state in a silent state to avoid unnecessary mode switching; the determination basis of the silent state is that the fluctuation amplitude of the electrical parameter measurement value in a plurality of consecutive sampling periods is less than a preset percentage threshold of the rated value, and the duration exceeds a preset silent time window; in the silent state, the system is forced to enter the grid-following mode and output a predefined instruction, avoiding unplanned island operation, and the predefined instruction can be dynamically adjusted according to the requirements of the upper energy management system; the division rule is a program-predefined time window division strategy or an adaptive division strategy dynamically adjusted based on grid state characteristics; the program-predefined time window division strategy is a fixed rule that divides the continuously collected electrical parameter data into equal-length training data sets and target data sets in chronological order; the adaptive division strategy dynamically adjusts the length of the training and target windows by monitoring the fluctuation characteristics of the electrical parameter in real time, automatically shortens the training window to improve the response speed of the model when the grid disturbance increases, and prolongs the training window to improve the prediction accuracy when the grid is stable; the prediction algorithm module is a linear predictor, and the weight coefficients are obtained by multiplying the inverse matrix of the self-covariance matrix and the cross-covariance vector; the prediction order is dynamically adjusted according to the length of the training data set and the background noise characteristics of the grid to balance the prediction accuracy and computational efficiency.

2. The method of coordinated control of energy storage converters according to claim 1, characterized in that, The compensation processing module performs the following operations: The predicted value and the actual measured value are input into a subtractor interface; The compensated signal is output as the source of the control instruction; Wherein, the compensation processing module detects when the power grid is in a silent state, and outputs a predefined logic state to replace the actual measured value.

3. The method of coordinated control of energy storage converters according to claim 1, characterized in that, Before calculating the autocovariance matrix and the cross-covariance vector, the training data set is preprocessed, which includes mean removal and normalization operations to eliminate DC components and unify the data scale; The preprocessing also includes low-pass filtering to remove high-frequency noise.

4. A coordinated control system for an energy storage converter, characterized by, It is used to realize the collaborative control method of the energy storage converter of any one of claims 1-3, comprising: A data acquisition unit for real-time acquisition of grid-connected point electrical parameters; A data set division unit configured to divide the data into a training data set and a target data set according to a program preset rule; A covariance processing unit for calculating the autocovariance matrix and the cross-covariance vector; A prediction and compensation unit for generating a predicted value and performing compensation processing; A control decision unit for selecting an operating mode according to the compensation result and generating a power instruction; Wherein, the prediction and compensation unit includes a state detection module for outputting a predefined control instruction when the grid is in a silent state; The state detection module includes a logic control circuit for outputting a predefined logic state in a silent state; The prediction and compensation unit includes: A pre-trained linear prediction model; A real-time compensation interface for outputting a compensated control instruction; A time delay unit for triggering a predefined instruction output after the grid state remains silent for a predefined time; The logic control circuit includes an AND gate or an NAND gate structure.

5. The coordinated control system of an energy storage converter according to claim 4, wherein, The data acquisition unit includes a voltage transformer, a current transformer, and a phase-locked loop circuit, and its sampling accuracy and dynamic response speed meet the requirements of capturing the transient process of the grid; The data acquisition unit also includes a comparator and a low-pass filter for generating an energy detection output.

6. The coordinated control system of an energy storage converter according to claim 5, wherein, The predefined time of the time delay unit can be remotely configured to avoid false triggering of the silent state processing logic due to short-term disturbances in the grid; The time delay unit is implemented by a counter or an RC circuit.

Images