Charging pile system management method based on dynamic load balancing

Abstract

The invention discloses a charging pile system management method based on dynamic load balancing, and particularly relates to the technical field of charging pile management, and the method comprises the steps: collecting time domain scheduling data of a charging pile side, a power grid side and a distributed renewable energy output device side, carrying out the fusion filtering, and obtaining a station-level state vector; generating a station-level prediction sequence of the charging pile in the next control period based on the station-level state vector and a rolling prediction model; an inner-layer controller reads the station-level state vector and the station-level prediction sequence, establishes a power prediction optimization model and solves the model, and outputs an expected station-level power trajectory and station-level power redundancy; a result output by the inner-layer controller is received, and a station-level power envelope curve and an energy storage charging and discharging set value are comprehensively solved and transmitted; the slope upper limit of the station-level power envelope curve is contracted based on the oscillation amplitude, the new slope upper limit is substituted into the next solution, the subsequent power change rate is limited, and the problem of unstable power exchange between the charging station and the superior power grid is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of charging pile management. More specifically, the present invention relates to a management method for a charging pile system based on dynamic load balancing. Background Art

[0002] With the parallel deployment of high-power DC fast charging technology and distributed renewable energy output devices, charging infrastructure has gradually evolved into "active load" nodes in the power system; multiple charging stations are both coupled to the public power grid through a three-phase low-voltage bus, and are also equipped with distributed renewable energy output devices such as rooftop photovoltaic, energy storage batteries, and power converters; the output of renewable energy devices and grid dispatching instructions are unstable, increasing the difficulty of dynamic load balancing of the charging pile system, specifically reflected in:

[0003] The externally given power boundary and the random fluctuation of renewable power cause the site load to exhibit fast, bidirectional, and unpredictable jump characteristics, and its change time constant is significantly mismatched with the vehicle battery charging time constant and the action time constant of the distribution network voltage regulating device, thus introducing an operating scenario of multi-time scale coupling; the exposed technical problem is: how to ensure the steady-state safety of the power exchange between the charging station and the superior power grid while meeting the instantaneous power demand of the vehicle, that is, to avoid rapid voltage drop or overshoot of the bus, frequency deviation exceeding the limit, and deterioration of power quality. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a management method for a charging pile system based on dynamic load balancing to solve the problems proposed in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solution: A management method for a charging pile system based on dynamic load balancing, the charging pile system is deployed in parallel with a distributed renewable energy output device, including the following steps:

[0006] Step 1: Multi-time domain state acquisition and fusion prediction, including: collecting first-time domain operation data on the charging pile side, second-time domain scheduling data on the grid side, and third-time domain scheduling data on the distributed renewable energy output device side, performing fusion filtering to obtain a station-level state vector, and generating a station-level prediction sequence for the charging pile in the next control cycle based on the station-level state vector and a rolling prediction model;

[0007] Step 2: Inner-layer prediction and balancing control, including: the inner-layer controller reads the station-level state vector and the station-level prediction sequence, establishes a power prediction optimization model and solves it, and outputs an expected station-level power trajectory and a station-level power redundancy;

[0008] Step 3: Outer grid collaborative envelope generation, including: receiving the results output by the inner layer controller, comprehensively solving the station-level power envelope curve and the energy storage charge and discharge set value, and transmitting them;

[0009] Step 4: Update control parameters, including: adjusting the smoothing weight factor, the power damping dead zone boundary, and the slope upper limit, and synchronously writing them into the status database; based on the oscillation amplitude, shrink the slope upper limit of the station-level power envelope curve, substitute the new slope upper limit into the next solution, and limit the subsequent power change rate.

[0010] Preferably, the first time-domain operation data is collected by sensors deployed on each charging pile, and at least includes instantaneous current, instantaneous voltage, vehicle battery power, and in-pile temperature; the second time-domain scheduling data is received through a communication link with an external power grid management system, and at least includes grid scheduling instructions, real-time electricity price signals, common bus voltage and frequency; the third time-domain scheduling data includes environmental parameters affecting renewable energy output and energy storage status; the station-level status vector refers to a multi-dimensional vector constructed based on the fusion filtering result, including the total power of the charging piles, the power distribution of each charging pile, the bus voltage frequency, prediction error, renewable energy output status, and energy storage status, and is used as the input of the station-level prediction sequence of the rolling prediction model.

[0011] Preferably, the rolling prediction model includes a renewable energy output prediction channel, a station-level power prediction channel, and a grid-side voltage prediction channel. The renewable energy output prediction channel is used to predict the renewable energy output status and energy storage status in the next control cycle (the next 1-5 minutes); the station-level power prediction channel is used to predict the total power of the charging piles in the next control cycle; the grid-side voltage prediction channel is used to predict the bus voltage frequency in the next control cycle. The operation process of the rolling prediction model includes:

[0012] State space modeling, taking the observed initial station-level state vector as the input, and respectively establishing linear time-varying state space equations for renewable energy predicted output, station-level predicted power, and bus voltage frequency;

[0013] Explanation: The predicted value of photovoltaic output is generated by short-term trend fitting of irradiance historical data, irradiance sensor data, and weather API interface; the bus voltage change rate has a linear relationship with the difference between the grid scheduling instruction power and the total power of the charging piles, and the proportional coefficient is determined by historical data regression analysis; renewable energy is affected by the integration of charge and discharge power and is restricted by the upper and lower limits of physical capacity;

[0014] Use recursive least squares or extended Kalman filter to update the state transition matrix and input gain matrix;

[0015] Multi-step look-ahead deduction, after identification at each control step, expand forward to the prediction time domain and output the station-level prediction sequence;

[0016] Confidence interval evaluation, calculating the noise covariance through Bayesian estimation, and generating the upper and lower confidence bounds corresponding to the renewable energy output prediction channel, the substation-level power prediction channel, and the grid-side voltage prediction channel.

[0017] Preferably, the inner-layer controller is used to quickly control the power of charging piles within the local area network of the charging pile group, and has a refresh cycle of milliseconds; the power prediction optimization model adopts a rolling optimization method and takes into account the vehicle charging demand and hardware safety constraints;

[0018] Explanation: The power prediction optimization model takes the charging pile power as the decision variable, takes the vehicle charging demand and hardware safety constraints as the boundaries, and uses a rolling optimization method based on the substation-level prediction sequence of the next control cycle to minimize the power fluctuation and the imbalance between piles, realizing the load balancing of charging piles, and continuously learning and optimizing the parameters of the power prediction optimization model according to historical charging data and real-time feedback to improve the accuracy and adaptability of load distribution; the charging pile power sequence is the power setting vector generated by the inner-layer controller for each charging pile within the prediction window and is sent to the execution layer in time series; the substation-level power redundancy is the difference between the total available power upper limit allowed by the charging pile hardware and the total output power after inner-layer optimization; the expected substation-level power trajectory is the curve of the substation-level power changing with time obtained by the inner-layer controller summarizing the charging pile power sequences.

[0019] Preferably, the prediction step of the inner-layer prediction model is equal to the refresh cycle multiplied by the number of optimization steps (the millisecond-level control cycle of the inner-layer controller, with a typical value of 1 - 10 milliseconds); the hardware safety constraints include that the current of a single pile does not exceed the hardware limit and the temperature inside the pile is lower than the safety threshold; the number of optimization steps is determined according to the balance between the system computing power and the prediction accuracy requirements.

[0020] Preferably, the outer-layer controller is used to perform the slow scheduling of the charging pile power, and the refresh cycle is at the second level (1 - 10 seconds); when the error between the actual power and the envelope curve alternates between positive and negative for 3 consecutive cycles and the absolute value increases, a warning state from the stable state to the oscillation warning state is triggered, and the control strategy is adjusted according to the situation.

[0021] Explanation: The slow scheduling of the outer-layer controller includes the following state transition rules:

[0022] When the error between the actual power and the envelope curve alternates between positive and negative for 3 consecutive cycles and the absolute value increases, a warning state from the stable state to the oscillation warning state is triggered;

[0023] If the warning state lasts for more than 2 outer-layer control cycles, a strategy of increasing the smoothing weight factor is started, and the update of the outer-layer envelope slope is frozen, triggering an intervention state from the oscillation warning state to the damping intervention state;

[0024] When the absolute value of the error continuously drops to the preset range, gradually release the slope constraint and reset the weight factor to trigger the recovery state of the control parameter recovery.

[0025] Preferably, the inner layer controller feeds back the inter-pile imbalance (standard deviation of each pile's power) after power distribution to the outer layer controller. The inter-pile imbalance is the standard deviation of the power distribution of each charging pile. The outer layer controller dynamically adjusts the distribution strategy of the station-level power redundancy based on the inter-pile imbalance. If the inter-pile imbalance rises, the outer layer controller increases the station-level power redundancy and relaxes the limit of the slope upper limit; if the inter-pile imbalance drops, the outer layer controller reduces the station-level power redundancy and tightens the limit of the slope upper limit.

[0026] Preferably, in the off-grid microgrid scenario, the outer layer controller constructs a closed-loop correlation between the power ramp rate limit and the inertia support ability through equivalent inertia perception, dynamically adjusting the power envelope slope upper limit, and priority compensation of energy storage.

[0027] When the charging station operates in the off-grid microgrid mode, the outer layer controller performs the following coordinated control steps:

[0028] Calculate the equivalent moment of inertia: Assume the number of virtual synchronous generators is n, and use t to represent the index number; Denote the output power of the t-th virtual synchronous generator as P gen,t ; Denote the mechanical time constant of the t-th virtual synchronous generator as T m,t , which reflects the inertia response characteristic. Denote the real-time charge and discharge power of the energy storage device as P ess ; Denote the virtual inertia time constant of the energy storage device as T ess , preset through the configuration interface; Denote the rated angular frequency as ω0; Denote the maximum allowable frequency deviation as Δf max ; Through the formula Calculate the equivalent moment of inertia of the microgrid. The equivalent moment of inertia of the microgrid is used to characterize the ability of the charging pile system to resist frequency fluctuations;

[0029] Adjust the power envelope slope upper limit: According to the equivalent moment of inertia J eq Adaptive adjust the slope upper limit S max of the station-level power envelope curve to ensure that the power change rate matches the system inertia. The adjustment formula is: where S base represents the preset slope upper limit, l represents the slope gain coefficient, obtained through training with historical data; J min , J max represent the allowable minimum and maximum equivalent moments of inertia respectively.

[0030] Preferably, the specific implementation process of step four includes the following steps:

[0031] Collect the actual station-level power after the inner-layer controller executes in real time for each control cycle, and synchronously receive the station-level power envelope curve set by the outer-layer controller;

[0032] For each control cycle, judge the change trend of the error sequence according to the error between the actual station-level power and the power envelope curve; if the error reverses direction and the amplitude increase ratio exceeds a preset value within two adjacent control cycles, it is determined that there is an oscillation trend;

[0033] When an oscillation trend is detected, increase the smoothing weight factor of the inner-layer controller to enhance the time continuity of power scheduling, set the power damping dead zone boundary centered on the current power mean value, and dynamically shrink the slope upper limit of the station-level power envelope curve; make the error fall within the power damping dead zone boundary range, suppress the high-frequency correction of the inner and outer layer control commands, and prevent loop high-frequency coupling. Loop high-frequency coupling refers to the phase resonance phenomenon caused by the bandwidth overlap of the inner and outer layer control loops;

[0034] When the inner-layer controller and the outer-layer controller quickly correct the power command at the same time and lack damping, and the error sequence shows high-frequency reverse flipping and gradually amplifies, it is regarded as high-frequency coupling oscillation); the smoothing weight factor is the continuity penalty coefficient in the objective function of the inner-layer controller;

[0035] The updated control parameters include the smoothing weight factor, the power damping dead zone boundary, and the outer envelope slope upper limit.

[0036] Preferably, the power prediction optimization model includes a capacity drift self-calibration compensation step: continuously collect the energy storage current, terminal voltage, temperature, and internal resistance within a sliding time window, suppress the noise of the energy storage current and terminal voltage data through ampere-hour integration combined with Kalman filtering to obtain the cumulative charge and discharge amount, and ensure the accuracy of capacity estimation; input the cumulative charge and discharge amount and the nominal capacity into the recursive least squares algorithm to online estimate the effective capacity, evaluate the battery health state based on the increase in internal resistance and cycle life loss, and predict the future capacity drift and health degradation trend of the battery based on historical data; use the difference between the nominal capacity and the effective capacity to represent the capacity drift amount, and update the state and charge mapping function according to the capacity drift amount; use the updated charge mapping function to regenerate the power difference correction value, and adjust the charge and discharge power of the energy storage energy conversion device according to the power difference correction value to complete the compensation; dynamically adjust the control strategy according to the evaluated battery health state and the predicted future capacity drift and health degradation trend, such as limiting the maximum charging current to slow down aging, or redistributing the energy storage tasks among battery modules to balance the health state; integrate the battery health state into the dynamic load balancing algorithm, preferentially allocate high-load tasks to the charging piles with better health states, and adjust the charge and discharge power of the energy storage energy conversion device according to the real-time capacity and health state of the battery.

[0037] Technical effects and advantages of the present invention:

[0038] (1) Through the hierarchical cooperation mechanism of the inner and outer layer controllers (millisecond-level fast power distribution in the inner layer and second-level grid coordinated scheduling in the outer layer), the present invention effectively solves the problem of time constant mismatch between the charging station and the distribution network voltage regulating device; the inner layer predictive equilibrium control quickly responds to vehicle demands based on the power trajectory of rolling optimization, and the outer layer grid coordination envelope ensures the steady-state safety of power exchange on the grid side through slope limitation, enabling the dynamic decoupling of the minute-level battery charging process and the second-level grid voltage regulation demand; by adopting the power damping dead zone boundary and slope upper limit dynamic contraction strategy, when it is detected that the error sequence alternates between positive and negative and the amplitude increases, a phase compensation barrier is constructed by increasing the smoothing weight factor and shrinking the slope upper limit; the problem of low-frequency oscillation generated by the voltage regulating device or inverter of the site and the distribution network is solved.

[0039] (2) Based on the three-channel fusion mechanism of the rolling prediction model, the present invention aligns the irradiance sensor data and the short-term prediction data of the weather API spatiotemporally through the Kalman gain matrix. Compared with the traditional single prediction channel, the prediction error of renewable energy output is reduced. The power envelope band generated through confidence interval evaluation provides a ±5% dynamic safety margin for the outer layer control, avoiding power command overshoot caused by prediction deviation, and solving the problem that the dynamic equilibrium algorithm makes decisions based on lagged information through the unified time series alignment and prediction fusion mechanism. Description of the Drawings

[0040] Figure 1 It is a flowchart of the management method of the charging pile system based on dynamic load balancing of the present invention.

[0041] Figure 2 It is a flowchart of the control parameter update of the present invention. Detailed Embodiments

[0042] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully communicated to those skilled in the art.

[0043] At the same time, it should be understood that, for the sake of description, the sizes of the various parts shown in the drawings are not drawn in actual proportional relationships.

[0044] The description of at least one exemplary embodiment hereinafter is actually only illustrative and in no way limits the application or use of the present application.

[0045] Techniques, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, such techniques, methods, and devices shall be regarded as part of the specification.

[0046] Example 1, refer to Figure 1 The flowchart of the management method for a charging pile system based on dynamic load balancing. The present invention provides a Figure 1 management method for a charging pile system based on dynamic load balancing as shown in the figure. The charging pile system is deployed in parallel with a distributed renewable energy output device, and includes the following steps:

[0047] Step 1: Multi-time domain state acquisition and fusion prediction, including: acquiring the first time domain operation data on the charging pile side, the second time domain scheduling data on the grid side, and the third time domain scheduling data on the distributed renewable energy output device side, performing fusion filtering to obtain a station-level state vector, and generating a station-level prediction sequence for the charging pile in the next control period based on the station-level state vector and a rolling prediction model;

[0048] Step 2: Inner layer prediction and balancing control, including: the inner layer controller reads the station-level state vector and the station-level prediction sequence, establishes and solves a power prediction optimization model, and outputs an expected station-level power trajectory and a station-level power redundancy;

[0049] Explanation: The power prediction optimization model is an optimal scheduling model with MPC or rolling horizon optimization as the core. Through the station-level state vector and the station-level prediction sequence, it continuously iteratively solves the optimal power distribution; the objective function of the power prediction optimization model includes: vehicle charging demand constraint and power smoothing penalty term. The vehicle charging demand constraint refers to dynamically allocating power based on the remaining battery power of the vehicle and the user priority weight; the power smoothing penalty term is used to suppress the power jump between adjacent control steps;

[0050] Step 3: Outer layer grid coordination envelope generation, including: receiving the results output by the inner layer controller, comprehensively solving the station-level power envelope curve, the energy storage charge and discharge setting value, and transmitting them;

[0051] Explanation: The updated control parameters are synchronously written into the state database for calling in the next round of multi-time domain state acquisition and fusion prediction to achieve stable operation of cross-time domain coordinated control; the updated control parameters include new parameters such as a smoothing weight factor, a power damping dead zone boundary, and a slope upper limit obtained through adaptive adjustment;

[0052] Explanation: The smoothing weight factor refers to the continuity penalty coefficient of the power smoothing penalty term in the objective function of the power prediction optimization model, which is used to suppress the jump of power commands in adjacent control steps. The larger the value, the smoother the output power. The power damping dead zone boundary refers to the power tolerance band set centered on the current total power mean value. Within the tolerance band, the command differences between the inner-layer controller and the outer-layer controller are temporarily ignored to prevent the superposition of high-frequency micro-amplitude commands. The upper limit of the slope of the station-level power envelope curve refers to the maximum allowable slope of the station-level power envelope curve on the time axis, which is used to limit the total power change rate and prevent the outer-layer scheduling from generating too fast power steps.

[0053] Step 4: Update the control parameters including the smoothing weight factor, the power damping dead zone boundary, and the upper limit of the slope, and synchronously write them into the status database.

[0054] Explanation of Step 4: Based on the upper limit of the slope of the station-level power envelope curve (the upper limit of the power change rate) that shrinks with the oscillation amplitude, the new upper limit of the slope is substituted into the next solution to limit the subsequent power change rate. According to the detected power oscillation trend, the power change rate constraint is dynamically adjusted to preferentially suppress the conflict between the inner-layer and outer-layer control commands and ensure the stability of the power exchange between the charging station and the superior power grid.

[0055] In the embodiments of the present invention, it needs to be explained that the first time-domain operation data is collected by sensors deployed on each charging pile and at least includes instantaneous current, instantaneous voltage, vehicle battery power, and in-pile temperature; the second time-domain scheduling data is received through the communication link with the external power grid management system and at least includes power grid scheduling commands, real-time electricity price signals, common bus voltage and frequency; the third time-domain scheduling data includes environmental parameters (such as irradiance, wind speed, temperature) affecting the output of renewable energy and the energy storage state; the station-level state vector refers to a multi-dimensional vector constructed based on the fusion filtering result, including the total power of the charging piles, the power distribution of each charging pile, the bus voltage frequency, the prediction error, the output state of renewable energy, and the energy storage state, which is used as the input of the station-level prediction sequence of the rolling prediction model.

[0056] Explanation: The predicted value of the photovoltaic output is generated by short-term trend fitting of environmental parameters such as historical irradiance data, irradiance sensor data, temperature, cloud cover, and the weather API interface. When the API data is unavailable, historical data interpolation or the default model is used; the environmental parameters are spatially and temporally aligned and processed through weighted fusion or machine learning models (such as support vector machines or neural networks) to improve the prediction accuracy; for example, the temperature data is collected in real time by sensors, and the cloud cover data is obtained from the weather API. After fusion, a short-term output prediction of 1 - 5 minutes is generated.

[0057] In the embodiments of the present invention, it should be explained that the rolling prediction model includes a renewable energy output prediction channel, a station-level power prediction channel, and a grid-side voltage prediction channel. The renewable energy output prediction channel is used to predict the renewable energy output state and energy storage state in the next control cycle (the next 1-5 minutes); the station-level power prediction channel is used to predict the total power of the charging piles in the next control cycle; the grid-side voltage prediction channel is used to predict the bus voltage frequency in the next control cycle. The operation process of the rolling prediction model includes:

[0058] State space modeling, using the observed initial station-level state vector as the input, and respectively establishing linear time-varying state space equations for renewable energy predicted output, station-level predicted power, and bus voltage frequency;

[0059] Explanation: The predicted value of photovoltaic output is generated by short-term trend fitting of irradiance historical data, irradiance sensor data, and weather API interface; the bus voltage change rate has a linear relationship with the difference between the grid dispatching instruction power and the total power of the charging piles, and the proportional coefficient is determined by historical data regression analysis; renewable energy is affected by the integration of charge and discharge power and is restricted by the upper and lower limits of physical capacity;

[0060] Using recursive least squares or extended Kalman filter to update the state transition matrix and input gain matrix;

[0061] Multi-step forward deduction, after identification at each control step, expand forward to the prediction time domain and output the station-level prediction sequence;

[0062] Confidence interval evaluation, calculating the noise covariance through Bayesian estimation, and generating the upper and lower confidence boundaries corresponding to the renewable energy output prediction channel, the station-level power prediction channel, and the grid-side voltage prediction channel.

[0063] Explanation: The rolling prediction model includes three channels for renewable energy / power / voltage, and uses recursive least squares or extended Kalman filter to update the matrix; the inner and outer layer controllers dynamically adjust the redundancy and slope upper limit based on the imbalance between piles; when an oscillation trend is detected, the smoothing weight factor is increased and the power damping dead zone boundary is set; the parts of the three channels for renewable energy / power / voltage in the rolling prediction model that are not explained in detail are prior art and can be implemented based on conventional technical means, and the present invention does not make specific limitations on this.

[0064] In the embodiments of the present invention, it should be explained that the inner layer controller is used to quickly control the power of the charging piles in the charging pile group local area network, and has a millisecond-level refresh cycle; the power prediction optimization model adopts a rolling optimization method and considers vehicle charging requirements and hardware safety constraints;

[0065] It should be explained that the power prediction optimization model takes the charging pile power as the decision variable, takes the vehicle charging demand and hardware safety constraints as the boundaries, and uses a rolling optimization method based on the station-level prediction sequence in the next control period to minimize power fluctuations and imbalance between piles, so as to achieve load balancing of the charging piles. It continuously learns and optimizes the parameters of the power prediction optimization model according to historical charging data and real-time feedback to improve the accuracy and adaptability of load distribution; the charging pile power sequence is the power setting vector generated by the inner-layer controller for each charging pile within the prediction window and is sent to the execution layer in sequence; the station-level power redundancy is the difference between the total available power upper limit allowed by the charging pile hardware and the total output power after inner-layer optimization; the expected station-level power trajectory is the curve of the station-level power changing with time obtained by the inner-layer controller summarizing the charging pile power sequences.

[0066] In the embodiments of the present invention, it should be explained that the prediction step of the inner-layer prediction model is equal to the refresh period multiplied by the number of optimization steps, and the number of optimization steps is determined according to the balance of system computing power and prediction accuracy requirements; for example, the number of optimization steps is 3 - 10 steps and is adjusted according to the system computing power; the hardware safety constraints include that the current of a single pile does not exceed the hardware limit and the temperature inside the pile is lower than the safety threshold.

[0067] In a possible embodiment, the method for obtaining the number of optimization steps is as follows:

[0068] When the inner-layer controller starts a new control period, it collects the total power demand of the charging piles, the power grid dispatching instruction, the output of renewable energy, and the energy storage state, synchronizes and aligns the data through timestamps, and classifies the operating scenarios into high-load, medium-load, or low-load modes by applying a decision tree classifier. At the same time, it identifies abnormal frequency deviations, generates scenario state labels and abnormal feature records, and stores them in the shared state database to support scenario-based computing resource allocation;

[0069] After receiving the scenario state labels and abnormal feature records, the inner-layer controller measures the CPU usage rate, available memory, and communication delay, calculates the maximum number of optimization iterations per second using a resource allocation model, generates computing power indicators and resource constraint alarms, and stores them in the shared state database to constrain the range of the number of optimization steps;

[0070] After obtaining the scenario state labels, abnormal feature records, and computing power indicators, the prediction model analyzes the historical prediction error data, estimates the prediction accuracy at different numbers of steps through a Bayesian inference model, selects a set of candidate steps that meet the accuracy threshold, and stores the set of candidate steps and the accuracy anomaly flag in the shared state database to support the final step optimization;

[0071] After obtaining the candidate step set and the precision anomaly flag, the inner-layer controller tests each step in the set, uses an online learning algorithm to minimize the power fluctuation error between the actual power trajectory and the predicted power trajectory to optimize the final steps, selects the step with the minimum error, verifies the stability by monitoring the convergence of the power trajectory, and stores the final optimized steps and the verification results in the shared state database to ensure the stability of power distribution.

[0072] Explanation: The prediction accuracy formula calculates the prediction accuracy at each step in the following way: Divide the historical prediction error data by the total system power input (obtained by adding the total power demand of the charging pile and the output of renewable energy), take the absolute value, then apply an exponential decay function, multiply by the scenario error amplification factor determined by historical data regression analysis based on the scenario state label, integrate the resulting value over the time axis, divide the integral value by the integral of the exponential decay function containing only the scenario error amplification factor over the time axis to generate the prediction accuracy value, with a value range of 0 to 1, and store it in the shared state database to quantify the accuracy of the prediction model and support the selection of the candidate step set.

[0073] In a possible embodiment, the prediction accuracy is calculated by the following formula:

[0074]

[0075] where, E history (t, N) represents the historical prediction error, which is obtained by extracting the difference between the predicted power and the actual power in the past control cycle from the shared state database, with the unit of kilowatt, and is calculated by the inner-layer controller based on sensor data and the output of the prediction model; P total (t) represents the total system power input, which is obtained by adding the total power demand of the charging pile and the output of renewable energy after time stamp synchronization alignment, with the unit of kilowatt, and is collected in real time from the charging pile sensor and the renewable energy sensor; F scenario (L scenario ) represents the scenario error amplification factor, which is obtained by historical data regression analysis based on the scenario state label (high load, medium load, low load or off-grid), dimensionless, and is stored in the shared state database, N represents the optimized steps, which are obtained by iteratively testing within the candidate step set by the prediction model to determine the steps that meet the accuracy requirements, dimensionless, and are dynamically generated by the prediction model; A predict (N) represents the prediction accuracy at the step, which is integrated after exponential decay and scenario adjustment based on the ratio of the historical prediction error and the total system power input, dimensionless, with a value range of [0, 1], and is stored in the shared state database.

[0076] For ease of understanding, in the embodiments of the present invention, the scenario error amplification factor is generated in the following manner: Extract the operation data of the past 30 days from the shared state database, including scenario status tags (high load, medium load, low load, or off-grid), total charging pile power demand, renewable energy output, and frequency deviation records, and construct a multi-dimensional data set containing scenario features and prediction errors; Use a support vector regression model to train the data set, with the scenario status tag as the input and the mean value of the ratio of the prediction error to the total power input as the output, to generate a scenario error amplification factor with a value of 1.5 in the high-load scenario, a value of 1.0 in the medium-load scenario, a value of 0.8 in the low-load scenario, and a value of 2.0 in the off-grid scenario, and store it in the shared state database, which is updated every 24 hours. The abnormal fluctuation amplification factor is generated in the following manner: Extract the abnormal feature records containing frequency deviation, power mutation, and communication delay from the shared state database, and construct an abnormal data set; Use a random forest regression model, with the abnormal feature records as the input and the peak value of the power fluctuation error as the output, to generate an abnormal fluctuation amplification factor with a value of 2.5 when the frequency deviation is greater than 0.5 Hz, a value of 2.0 when the power mutation is greater than 10% of the rated power, and a value of 1.5 when the communication delay is greater than 50 ms, and store it in the shared state database, which is updated every 12 hours.

[0077] Explanation: The power fluctuation error formula calculates the power fluctuation error at the following number of steps: Take the absolute value of the difference between the actual power trajectory and the predicted power trajectory, divide it by the system constraint power (obtained by adding the grid dispatch instruction and the energy storage state multiplied by the energy storage capacity factor), square it, and then multiply it by the abnormal fluctuation amplification factor determined through historical data analysis based on the abnormal feature records. The resulting value is integrated on the time axis to generate a power fluctuation error value in units of kilowatt square seconds, which is stored in the shared state database to evaluate the power distribution stability and optimize the selection of the final number of steps.

[0078] In a possible embodiment, the power fluctuation error is calculated by the following formula:

[0079]

[0080] Among them, P actual (t) represents the actual power trajectory, which is obtained by real-time collecting the total station-level power output through a charging pile sensor, in units of kilowatts, and is stored in the shared state database for use by the inner-layer controller; P predict (t,N) represents the predicted power trajectory, which is calculated by the inner-layer prediction model based on the number of steps N and the station-level state vector, in units of kilowatts, and is extracted from the shared state database; P constraint (t) represents the system constraint power, which is obtained by adding the grid dispatch instruction and the energy storage state multiplied by the energy storage capacity factor after time stamp synchronization alignment, in units of kilowatts, and is collected from the grid communication interface and the energy storage monitoring device; F anomaly(A anomaly ) represents the abnormal fluctuation amplification factor, which is obtained by determining through historical data analysis based on abnormal feature records (such as frequency deviation), has no unit, and is stored in the shared state database; E fluctuation (N) represents the power fluctuation error, which is obtained by normalizing the square of the difference between the actual and predicted power trajectories, multiplying by the abnormal fluctuation amplification factor, and integrating. The unit is kilowatt-square seconds and it is stored in the shared state database.

[0081] To increase the robustness of obtaining the optimization steps, during the process of obtaining the optimization steps, to cope with abnormal situations such as sensor failures and communication interruptions, the inner-layer controller executes the following logic: When it is detected that the difference between the charging pile sensor readings (such as instantaneous current or voltage) and the previous cycle exceeds 20% of the rated value, or the readings are missing for 3 consecutive cycles, it is determined as a sensor failure, and the standby historical data mode is triggered. The average power demand and renewable energy output in the shared state database for the most recent 24 hours are used to replace the real-time data to generate temporary state tags and abnormal feature records; When the communication delay exceeds 100 milliseconds or the grid dispatch instructions are not received for 2 consecutive cycles, it is determined as a communication interruption, and the inner-layer controller switches to the local prediction mode. Based on the most recent valid station-level state vector and prediction sequence, the optimization steps are fixed at 3, and the online learning optimization is paused until the communication is restored; The above abnormal handling results are stored in the shared state database, and an abnormal event report is generated after the normal operation is restored and pushed to the outer-layer controller to adjust the subsequent power envelope curve.

[0082] In the embodiments of the present invention, it should be explained that the outer-layer controller is used to perform slow scheduling of the charging pile power, and the refresh period is at the second level (1 - 10 seconds); When the error between the actual power and the envelope curve alternates between positive and negative for 3 consecutive cycles and the absolute value increases, a warning state from the stable state to the oscillation warning state is triggered, and the control strategy is adjusted according to the situation.

[0083] Explanation, the slow scheduling of the outer-layer controller includes the following state transition rules:

[0084] When the error between the actual power and the envelope curve alternates between positive and negative for 3 consecutive cycles and the absolute value increases, a warning state from the stable state to the oscillation warning state is triggered;

[0085] If the warning state lasts for more than 2 outer-layer control cycles, the smooth weight factor increment strategy is started, and the update of the outer-layer envelope slope is frozen, triggering an intervention state from the oscillation warning state to the damping intervention state;

[0086] When the absolute value of the error continuously drops to the preset range, the slope constraint is gradually released and the weight factor is reset, triggering a recovery state of control parameter recovery.

[0087] In the embodiments of the present invention, it should be explained that the inner layer controller feeds back the inter-pile imbalance (standard deviation of each pile's power) after power distribution to the outer layer controller. The inter-pile imbalance is the standard deviation of the power distribution of each charging pile. The outer layer controller dynamically adjusts the distribution strategy of the station-level power redundancy based on the inter-pile imbalance. If the inter-pile imbalance rises, the outer layer controller increases the station-level power redundancy and relaxes the limit of the slope upper limit; if the inter-pile imbalance drops, the outer layer controller reduces the station-level power redundancy and tightens the limit of the slope upper limit.

[0088] Background description: In the off-grid microgrid scenario, the charging station faces the severe challenge of frequency instability when the charging load suddenly changes due to the loss of the inertia and power support of the main power grid. The traditional control strategy does not dynamically match the equivalent inertia (such as the rotational inertia simulated by the virtual synchronous machine) with the power change rate. When the high-power charging demand suddenly increases, the microgrid cannot buffer the power impact due to insufficient inertia, resulting in frequency drop exceeding the limit or even protection tripping. At the same time, the fixed threshold power ramp limit and the decoupled energy storage response mechanism are difficult to balance the charging efficiency and system stability. Therefore, the embodiments of the present invention construct a closed-loop association between the power ramp rate limit and the system inertia support ability through a cooperative mechanism of equivalent inertia perception, dynamic constraint of the power envelope slope, and priority compensation of energy storage: calculate the equivalent inertia in real time and dynamically adjust the upper limit of the power envelope slope to make the power change rate strictly adapt to the inertia level of the microgrid. When the frequency deviation is detected to exceed the limit, trigger the energy storage millisecond-level power compensation first and freeze the slope adjustment to form a stable control loop across time scales, ultimately ensuring the charging efficiency while avoiding the risk of frequency collapse caused by power mutation.

[0089] In the embodiments of the present invention, it should be explained that in the off-grid microgrid scenario, the outer layer controller constructs a closed-loop association between the power ramp rate limit and the inertia support ability through equivalent inertia perception, dynamic adjustment of the upper limit of the power envelope slope, and priority compensation of energy storage;

[0090] When the charging station operates in the off-grid microgrid mode, the outer layer controller executes the following cooperative control steps:

[0091] Calculate the equivalent rotational inertia: Assume the number of virtual synchronous generators is n, and use t to represent the index number; Denote the output power of the t-th virtual synchronous generator as P gen,t ; Denote the mechanical time constant of the t-th virtual synchronous generator as T m,t , which reflects the inertia response characteristic. Denote the real-time charge and discharge power of the energy storage device as P ess ; Denote the virtual inertia time constant of the energy storage device as T ess , which is preset through the configuration interface; Denote the rated angular frequency as ω0; Denote the maximum allowable frequency deviation as Δf max ; Through the formula The equivalent moment of inertia of the microgrid is calculated and used to characterize the ability of the charging pile system to resist frequency fluctuations;

[0092] Adjust the upper limit of the power envelope slope: according to the equivalent moment of inertia J eq Adaptively adjust the slope upper limit S of the station-level power envelope curve max , to ensure that the power change rate matches the system inertia, the adjustment formula is: Among them S base represents the preset slope upper limit, k represents the slope gain coefficient, which is obtained through historical data training; J min , J max They represent the minimum and maximum allowable equivalent moments of inertia respectively.

[0093] Furthermore, when the dynamic offset of the microgrid frequency exceeds the preset stability threshold and continues to exceed the limit, the transient power compensation mechanism of the energy storage device is triggered and the slope constraint parameter of the power envelope curve is locked. A topology optimization alarm of insufficient inertial support capacity is simultaneously generated, and a virtual synchronous unit expansion configuration recommendation is pushed.

[0094] It is necessary to explain in the embodiments of the present invention that Figure 2 The control parameter update flow chart of FIG. 1 includes the following steps:

[0095] Real-time collection of the actual station-level power after the inner controller executes each control cycle, and synchronous reception of the station-level power envelope curve set by the outer controller;

[0096] For each control cycle, the error sequence trend is determined based on the error between the actual station-level power and the power envelope curve. If the error reverses between two consecutive control cycles and the amplitude increase ratio exceeds the preset value, it is determined to be an oscillation trend.

[0097] When an oscillation trend is detected, the smoothing weight factor of the inner controller is increased to enhance the temporal continuity of power scheduling, a power damping dead zone boundary centered on the current power mean is set, and the slope upper limit of the station-level power envelope curve is dynamically reduced;

[0098] The error is kept within the power damping dead zone boundary, suppressing high-frequency corrections to the inner and outer control commands and preventing high-frequency coupling of the loops; high-frequency coupling refers to the phase resonance phenomenon caused by the bandwidth overlap of the inner and outer control loops.

[0099] Explanation: When the inner controller and the outer controller simultaneously and rapidly correct the power instructions and lack damping, the error sequence undergoes high-frequency reverse flipping and gradually amplifies, which is considered to be a high-frequency coupled oscillation.

[0100] In a possible embodiment, for each control period, according to the error between the actual station-level power and the power envelope curve, the change trend of the error sequence is judged; the station-level power error sequence is processed by fast Fourier transform (FFT) to extract the main frequency component in the range of 0.1 Hz to 1 Hz; if the amplitude of the main frequency component exceeds a preset threshold (for example, 5% of the rated power) and remains consistent in two consecutive control periods, it is determined that there is an oscillation trend; according to the detected oscillation frequency, the proportional-integral (PI) controller is used to dynamically adjust the smoothing weight factor to enhance the time continuity of power scheduling; the specific implementation process is as follows:

[0101] In each control period, according to the station-level power error sequence recorded in the current period, Fourier transform is performed to obtain spectrum information;

[0102] If there are frequency components in the spectrum that are within a preset frequency range and the amplitude exceeds a set threshold, and this condition is satisfied in two consecutive periods, it is judged that there is periodic oscillation; if there is no periodic oscillation, the error sequence is processed using a band-pass filter;

[0103] If the filtered output continuously shows high-amplitude oscillations exceeding the energy threshold within a set frequency range and the duration exceeds a preset time threshold, it is judged that there is an oscillation trend;

[0104] If any of the above conditions is satisfied, it is determined that the oscillation trend is established.

[0105] In a possible embodiment, the inner-layer controller and the outer-layer controller are connected through a local area network to achieve data exchange and instruction issuance; the output data of the inner-layer controller and the outer-layer controller are received through a communication network, and the updated control parameters are written into the shared state database; the state database is a shared storage device for access by the inner-layer controller and the outer-layer controller, and is used to store and update control parameters and prediction sequences.

[0106] Summary: Embodiment 1 of the present invention provides a management method for a charging pile system based on dynamic load balancing. Through a hierarchical collaborative control architecture (inner-layer millisecond-level fast power distribution + outer-layer second-level grid collaborative scheduling) and a multi-time-domain prediction fusion mechanism, the power jump and steady-state safety problems caused by the multi-time-scale coupling between the charging station and the grid are solved; the specific implementation includes:

[0107] Integrate the real-time data of the charging pile (current / voltage / temperature), the grid scheduling instructions (electricity price / bus voltage), and the renewable energy output parameters (irradiance / storage SOC), construct a multi-dimensional state vector (such as total power distribution, prediction error, etc.), and eliminate sensor noise through Kalman filtering;

[0108] Adopt a three-channel prediction model (renewable energy output, charging pile power, bus voltage), generate power / voltage prediction sequences for the next 1-5 minutes based on linear time-varying state space equations (the coefficients are updated by recursive least squares), and calculate the Bayesian confidence boundary (±5% error band);

[0109] Trigger the upper limit contraction of the slope and the increase of the smoothing weight factor according to the oscillation trend detection (the error alternates between positive and negative and the amplitude increases), and suppress the conflict of inner and outer layer control commands through the power damping dead zone;

[0110] In the microgrid mode, dynamically adjust the upper limit of the power envelope slope based on the equivalent moment of inertia formula to ensure that the power ramp rate matches the system inertia.

[0111] Compared with the traditional fixed-step method, the optimized step acquisition scheme of this method significantly improves the prediction accuracy and power distribution stability in actual tests. In the 30-day operation test of the charging pile system, the average prediction error of the fixed-step method (the number of steps is set to 5) is 8.2%, and the mean value of the power fluctuation error is 12.5 kW·s²; while this method optimizes through dynamic scenario classification and online learning, the average prediction error is reduced to 4.7%, and the mean value of the power fluctuation error is reduced to 7.3 kW·s², improving the performance by 42.7% and 41.6% respectively. In the off-grid microgrid scenario, in the abnormal situation where the frequency deviation is greater than 0.5 Hz, the power distribution stable time of this method is shortened from 1.2 s of the fixed-step method to 0.6 s, improving the response speed by 50%.

[0112] Embodiment 2. The difference between the embodiment of the present invention and Embodiment 1 is that the power prediction optimization model includes a capacity drift self-calibration compensation step: continuously collect the energy storage current, terminal voltage, temperature, and internal resistance within a sliding time window, suppress the noise of the energy storage current and terminal voltage data through ampere-hour integration combined with Kalman filtering to obtain the cumulative charge and discharge amount, and ensure the accuracy of capacity estimation; input the cumulative charge and discharge amount and the nominal capacity into the recursive least squares algorithm to online estimate the effective capacity, evaluate the battery health state based on the increase in internal resistance and cycle life loss, and predict the future capacity drift and health degradation trend of the battery based on historical data; represent the capacity drift amount by the difference between the nominal capacity and the effective capacity, and update the state and charge mapping function according to the capacity drift amount; use the updated charge mapping function to regenerate the power difference correction value, and adjust the charge and discharge power of the energy storage energy conversion device according to the power difference correction value to complete the compensation; dynamically adjust the control strategy according to the evaluated battery health state and the predicted future capacity drift and health degradation trend, such as limiting the maximum charging current to slow down aging, or redistributing the energy storage tasks among battery modules to balance the health state; incorporate the battery health state into the dynamic load balancing algorithm, preferentially allocate high-load tasks to the charging piles with better health states, and adjust the charge and discharge power of the energy storage energy conversion device according to the real-time capacity and health state of the battery.

[0113] In a possible embodiment, after constructing the historical record of the oscillation event, when a similar oscillation is detected subsequently, adjust the parameters according to the historical data. The specific implementation process is as follows:

[0114] After each oscillation is detected, record the characteristic information corresponding to the oscillation, including the oscillation frequency, the number of continuous cycles at that time, as well as the smoothing weight factor and the power damping dead zone boundary value used;

[0115] Construct the above oscillation events into a historical record to form an oscillation event database;

[0116] When a new oscillation is detected in a subsequent control cycle, compare the similarity with the records in the historical database according to the current oscillation frequency and duration;

[0117] Select the parameter values corresponding to the record in the historical record that is most similar to the current oscillation characteristics as the control parameter setting for the current cycle;

[0118] Update the smoothing weight factor and the power damping dead zone boundary value to the corresponding optimal setting values in the historical record respectively.

[0119] In a possible embodiment, before detecting the oscillation trend, perform an abnormal identification and classification response process to avoid the adaptive parameter adjustment operation. The specific implementation process is as follows:

[0120] During the current control cycle, respectively obtain the residual value between the model prediction result and the actual power measurement value, the deviation value between the bus voltage and the target voltage, and the difference in consecutive cycle readings of the key sensors; if the total residual exceeds the preset residual judgment threshold, it is determined that the model prediction is abnormal;

[0121] If the bus voltage deviation value exceeds the preset voltage stability threshold, it is determined that there is an external disturbance; if the difference between the measured value of the key sensor and the previous cycle value exceeds the set mutation amplitude threshold, it is determined that the sensor is faulty;

[0122] When any one of the above three items meets the abnormal determination condition, skip the adaptive parameter adjustment operation of the current cycle, send an alarm signal, and switch to the backup control strategy to ensure stable operation;

[0123] When none of the above three criteria are triggered, perform the parameter adaptive adjustment process.

[0124] In a possible embodiment, the upper limit of the slope is dynamically adjusted according to the voltage harmonic distortion rate monitored in real time. When it is detected that the harmonic distortion rate exceeds the standard limit, the upper limit of the slope is automatically tightened; analyze the power error spectrum. If the main frequency exceeds the bandwidth of the outer layer controller, it is determined that the oscillation trend is caused by the overlap of the inner and outer layer control bandwidths, generate an emergency power envelope curve, and preferentially call the energy storage compensation instruction; the operation process of the energy storage compensation instruction includes:

[0125] Compare the station-level predicted power with the actual station-level power in each control cycle; if the absolute value of the power gap exceeds the confidence interval, trigger the compensation process;

[0126] The difference between the station-level predicted power and the actual station-level power represents the power gap; if the power gap is positive, control the energy storage conversion device to discharge immediately to make up the gap; if the power gap is negative, instruct the energy storage conversion device to absorb the excess power;

[0127] Analyze the mode and frequency of the predicted power gap in real time, and use the fault diagnosis model to judge whether there is a sensor fault or an abnormal event (such as several charging piles going offline suddenly); if an abnormality is detected, automatically adjust the compensation strategy, preferentially ensure the key load or take other preset measures; input the predicted power gap and the real-time power gap of the charging piles, and the output is the fault diagnosis result and the adjusted compensation strategy. Use the support vector machine to analyze the data to identify sensor or charging pile abnormalities and adjust the instructions according to the preset strategy;

[0128] Increase the weight of the latest measurement data in the state space model or switch to the backup prediction model to quickly correct the subsequent prediction results;

[0129] Write the error correction amount and the updated model parameters into the state database for the next round of power prediction optimization model to call.

[0130] In a possible embodiment, the problem that a single control strategy cannot adapt to multiple scenarios is solved through dynamic mode switching logic; according to the urgency of the power grid dispatching instruction (such as peak shaving demand, harmonic suppression instruction) and the real-time load rate of the charging pile group, the cooperation mode of the inner layer controller and the outer layer controller is dynamically selected, including the following modes:

[0131] Mode A (steady-state cooperation): The outer layer controller dominates the generation of the station-level power envelope curve, and the inner layer controller distributes according to the conventional rolling optimization;

[0132] Mode B (emergency response): When it is detected that the power grid instruction contains harmonic suppression requirements, it switches to the joint calculation mode of the inner and outer layers, and the inner layer controller feeds back the power distribution result to the outer layer envelope slope constraint in real time;

[0133] Mode C (fault fallback): When it is detected that the communication delay exceeds the threshold, the inner layer local cache data is enabled to generate a power sequence, and the outer layer controller only monitors without intervention.

[0134] Summary: Based on Embodiment 1, this embodiment adds a capacity drift self-calibration compensation and a multi-modal oscillation suppression mechanism, and focuses on solving the power distribution deviation and control loop resonance problems caused by battery aging.

[0135] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A management method for a charging pile system based on dynamic load balancing, where the charging pile system is deployed in parallel with a distributed renewable energy output device, characterized in that It includes the following steps: Step 1: Multi-time domain state acquisition and fusion prediction, including: collecting the first time domain operation data on the charging pile side, the second time domain scheduling data on the grid side, and the third time domain scheduling data on the distributed renewable energy output device side, performing fusion filtering to obtain the station-level state vector, and generating the station-level prediction sequence of the charging pile in the next control period based on the station-level state vector and the rolling prediction model; Step 2: Inner layer prediction and equilibrium control, including: the inner layer controller reads the station-level state vector and the station-level prediction sequence, establishes and solves the power prediction optimization model, and outputs the expected station-level power trajectory and the station-level power redundancy; Step 3: Outer layer grid collaborative envelope generation, including: receiving the result output by the inner layer controller, comprehensively solving the station-level power envelope curve, the energy storage charge and discharge set value and transmitting them; the upper limit of the slope of the station-level power envelope curve is used to prevent excessive power steps in the outer layer scheduling; Step 4: Update the control parameters including the smoothing weight factor, the power damping dead zone boundary, and the slope upper limit, and synchronously write them into the state database; based on the oscillation amplitude, shrink the slope upper limit of the station-level power envelope curve, and substitute the new slope upper limit into the next solution; the smoothing weight factor refers to the continuity penalty coefficient of the power smoothing penalty term in the objective function of the power prediction optimization model, which is used to suppress the jump of the power command between adjacent control steps, and the power damping dead zone boundary refers to the power tolerance band set centered on the current total power mean value to prevent the superposition of high-frequency micro-amplitude commands.

2. The management method of a charging pile system based on dynamic load balancing according to claim 1, wherein, The first time domain operation data is collected by sensors deployed on each charging pile, and at least includes instantaneous current, instantaneous voltage, vehicle battery power and in-pile temperature; the second time domain scheduling data is received through the communication link with the external grid management system, and at least includes grid scheduling instructions, real-time electricity price signals, common bus voltage and frequency; the third time domain scheduling data includes environmental parameters affecting renewable energy output and energy storage status; The station-level state vector refers to a multi-dimensional vector constructed based on the fusion filtering result, including the total power of the charging pile, the power distribution of each charging pile, the bus voltage frequency, the prediction error, the renewable energy output status and the energy storage status, and is used as the input of the station-level prediction sequence of the rolling prediction model.

3. A management method for a charging pile system based on dynamic load balancing according to claim 2, characterized in that, The rolling prediction model includes a renewable energy output prediction channel, a station-level power prediction channel and a grid-side voltage prediction channel. The renewable energy output prediction channel is used to predict the renewable energy output status and the energy storage status in the next control period; the station-level power prediction channel is used to predict the total power of the charging pile in the next control period; The grid-side voltage prediction channel is used to predict the bus voltage frequency in the next control period. The operation process of the rolling prediction model includes: State-space modeling, with the observed initial substation-level state vector as the input, respectively establish linear time-varying state-space equations for renewable energy predicted output, substation-level predicted power, and bus voltage frequency; the predicted value of photovoltaic output is generated by short-term trend fitting of irradiance historical data, irradiance sensor data, and weather API interface; the rate of change of bus voltage has a linear relationship with the difference between the grid dispatching command power and the total power of charging piles, and the proportional coefficient is determined by regression analysis of historical data; renewable energy is affected by the integration of charging and discharging power and is constrained by the upper and lower limits of physical capacity. Use recursive least squares or extended Kalman filter to update the state transition matrix and input gain matrix. Multi-step look-ahead deduction, after identification at each control step, expand forward to the prediction time domain to output the substation-level prediction sequence. Confidence interval evaluation, calculate the noise covariance through Bayesian estimation, and generate the upper and lower confidence bounds corresponding to the renewable energy output prediction channel, substation-level power prediction channel, and grid-side voltage prediction channel.

4. A management method for a charging pile system based on dynamic load balancing according to claim 1, characterized in that, The inner-layer controller is used to quickly control the power of charging piles within the charging pile group local area network. The power prediction optimization model takes the charging pile power as the decision variable, with vehicle charging demand and hardware safety constraints as the boundaries, and uses a rolling optimization method based on the substation-level prediction sequence of the next control cycle to minimize power fluctuations and imbalance between piles, realizing the load balancing of charging piles, and continuously learning and optimizing the parameters of the power prediction optimization model according to historical charging data and real-time feedback; the charging pile power sequence is the power setting vector generated by the inner-layer controller for each charging pile within the prediction window and is sent to the execution layer in sequence; the substation-level power redundancy is the difference between the upper limit of the total available power allowed by the charging pile hardware and the total output power after inner-layer optimization; the expected substation-level power trajectory is the curve of the substation-level power changing with time obtained by the inner-layer controller summarizing the charging pile power sequences.

5. A management method for a charging pile system based on dynamic load balancing according to claim 4, characterized in that The prediction step length of the inner-layer prediction model is equal to the refresh period multiplied by the number of optimization steps; the hardware safety constraints include that the current of a single pile does not exceed the hardware limit and the temperature inside the pile is lower than the safety threshold.

6. A management method for a charging pile system based on dynamic load balancing according to claim 4, characterized in that, The outer-layer controller is used to perform slow scheduling of the charging pile power, and the refresh period is at the second level; when the error between the actual power and the envelope curve alternates between positive and negative for 3 consecutive cycles and the absolute value increases, trigger the oscillation warning state.

7. A management method for a charging pile system based on dynamic load balancing according to claim 1, characterized in that, The inner-layer controller feeds back the imbalance degree between piles after power distribution to the outer-layer controller. The imbalance degree between piles is the standard deviation of the power distribution of each charging pile. The outer-layer controller dynamically adjusts the distribution strategy of the substation-level power redundancy based on the imbalance degree between piles. If the imbalance degree between piles rises, increase the substation-level power redundancy and relax the limit of the slope upper limit; If the imbalance degree between piles drops, reduce the substation-level power redundancy and tighten the limit of the slope upper limit.

8. A method for managing a charging pile system based on dynamic load balancing according to claim 1, characterized in that When the charging station operates in the off-grid microgrid mode, the outer-layer controller performs the following coordinated control steps: Calculating the equivalent moment of inertia: Assume the number of virtual synchronous generators is n, and use t to represent the index number; Denote the output power of the t-th virtual synchronous generator as P gen,t ; Denote the mechanical time constant of the t-th virtual synchronous generator as T m,t , which reflects the inertia response characteristic. Denote the real-time charge and discharge power of the energy storage device as P ess ; Denote the virtual inertia time constant of the energy storage device as T ess , which is preset through the configuration interface; Denote the rated angular frequency as ω0; Denote the maximum allowable frequency deviation as Δf max ; Calculate the equivalent moment of inertia of the microgrid through the formula The equivalent moment of inertia of the microgrid is calculated, and the equivalent moment of inertia of the microgrid is used to characterize the ability of the charging pile system to resist frequency fluctuations; Adjust the upper limit of the power envelope slope: According to the equivalent moment of inertia J eq Adaptively adjust the upper limit S of the slope of the station-level power envelope curve max , ensuring that the power change rate matches the system inertia. The adjustment formula is: Where S base represents a preset slope upper limit, and k represents a slope gain coefficient, which is obtained by training with historical data; J min and J max represent the minimum and maximum allowable equivalent moments of inertia, respectively.

9. A method for managing a charging pile system based on dynamic load balancing according to claim 8, characterized in that, The specific implementation process of Step 4 includes the following steps: Real-time collect the actual substation-level power after the execution of the inner-layer controller in each control cycle, and synchronously receive the substation-level power envelope curve set by the outer-layer controller; For each control cycle, judge the change trend of the error sequence according to the error between the actual substation-level power and the power envelope curve. If the error reverses direction between two adjacent control cycles and the amplitude increase ratio exceeds a preset value, it is determined that there is an oscillation trend; When an oscillation trend is detected, increase the smoothing weight factor of the inner-layer controller to enhance the time continuity of power scheduling, set the power damping dead-zone boundary centered on the current power mean, and dynamically shrink the slope upper limit of the station-level power envelope curve; Keep the error within the power damping dead-zone boundary to suppress high-frequency corrections to the inner- and outer-layer control commands and prevent high-frequency loop coupling; high-frequency loop coupling refers to the phase resonance phenomenon caused by the bandwidth overlap of the inner- and outer-layer control loops.

10. A management method for a charging pile system based on dynamic load balancing according to claim 9, characterized in that, The power prediction optimization model includes a capacity drift self-calibration compensation step: continuously collect the energy storage current, terminal voltage, temperature, and internal resistance within a sliding time window, suppress the noise of the energy storage current and terminal voltage data through ampere-hour integration combined with Kalman filtering to obtain the cumulative charge and discharge amount; Input the cumulative charge and discharge amount and the nominal capacity into the recursive least squares algorithm to online estimate the effective capacity, evaluate the battery health state based on the internal resistance increase and cycle life loss, and predict the future capacity drift and health degradation trend of the battery based on historical data; Use the difference between the nominal capacity and the effective capacity to represent the capacity drift amount, and update the state and charge mapping function according to the capacity drift amount; Use the updated charge mapping function to regenerate the power difference correction value, and adjust the charge and discharge power of the energy storage energy conversion device according to the power difference correction value to complete the compensation; Dynamically adjust the control strategy according to the evaluated battery health state and the predicted future capacity drift and health degradation trend; Integrate the battery health state into the dynamic load balancing algorithm, preferentially allocate high-load tasks to the charging piles with better health states, and adjust the charge and discharge power of the energy storage energy conversion device according to the real-time capacity and health state of the battery.

Images