A charging module energy management method employing predictive control strategy

Abstract

This invention discloses a charging module energy management method employing a predictive control strategy, relating to the field of charging management technology. The method includes: Step 100, synchronously collecting local data from submodules and global data from the power grid; preprocessing the local data from submodules and the global data from the power grid to obtain standardized submodule data and standardized global data, respectively; Step 200, based on the standardized submodule data, obtaining the local prediction results of each submodule through a local prediction model, constructing a submodule correlation matrix, calculating the prediction deviation difference between related submodules using the submodule correlation matrix, and collaboratively correcting local prediction results exceeding a preset deviation threshold to obtain corrected prediction results. This invention can improve the accuracy of energy management in charging systems, optimize submodule load allocation, reduce power grid impact, reduce submodule operational failures, and achieve high efficiency in global energy utilization and stable system operation.

Description

Technical Field

[0001] This invention relates to the field of charging management technology, specifically to a charging module energy management method employing a predictive control strategy. Background Technology

[0002] In charging systems composed of multiple sub-modules, predictive control strategies have become the mainstream technology for energy management. By predicting and analyzing charging demand, module status, and grid constraints, global energy allocation and scheduling optimization can be achieved. This type of technology is widely used in large-scale charging scenarios such as charging pile clusters and integrated photovoltaic-storage-charging power stations, and is the core support for improving the operating efficiency of charging systems.

[0003] Current mainstream solutions mostly adopt an architecture of independent prediction by sub-modules and centralized decision-making in the cloud. Each sub-module completes local state prediction based on its own data and uploads the results to the cloud to generate scheduling instructions. In this mode, the prediction of a single sub-module is susceptible to local interference such as sensor errors and short-term load fluctuations, which can cause deviations. The independent deviations of multiple sub-modules can accumulate through superposition and amplification effects, causing the cloud scheduling instructions to become disconnected from actual needs, resulting in an imbalance in global energy allocation and failing to achieve accurate and efficient energy management. Summary of the Invention

[0004] The purpose of this invention is to provide a charging module energy management method employing a predictive control strategy to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a charging module energy management method employing a predictive control strategy, comprising the following steps:

[0006] Step 100, Data Acquisition and Preprocessing: Simultaneously acquire local data of the submodule and global data of the power grid, and preprocess the local data of the submodule and the global data of the power grid respectively to obtain standardized submodule data and standardized global data;

[0007] Step 200, Deviation Cooperative Correction: Based on the standardized sub-module data, the local prediction results of each sub-module are obtained through the sub-module local prediction model, a sub-module correlation matrix is ​​constructed, the prediction deviation difference between the related sub-modules is calculated in combination with the sub-module correlation matrix, and the local prediction results that exceed the preset deviation threshold are cooperatively corrected to obtain the corrected prediction results.

[0008] Step 300, Global Prediction Calibration: Construct a global optimization objective function and constraints based on the standardized global data, substitute the corrected prediction result into the global optimization objective function, solve the calibrated prediction result through the optimization algorithm, and calculate the calibration deviation between the calibrated prediction result and the corrected prediction result.

[0009] Step 400, Energy scheduling execution: Generate energy scheduling instructions based on the calibrated prediction results, and send the energy scheduling instructions to each submodule for execution;

[0010] Step 500, Feedback Iterative Optimization: Collect the actual operating data of each submodule after executing the energy scheduling command, and update the parameters of the submodule local prediction model, submodule correlation matrix and optimization algorithm by combining the actual operating data and the calibration deviation.

[0011] Preferably, step 100 includes:

[0012] Step 110, Multi-dimensional data acquisition: The sub-module collects its own real-time operation data, historical charging data and local environmental data through built-in sensors to form the sub-module local data; the cloud controller collects the real-time power supply capacity, peak and valley electricity price periods, power grid load warning and the system's overall sub-module operation status summary and total charging demand prediction value to form the power grid global data;

[0013] Step 120, Data Preprocessing: The submodule performs filtering and normalization processing on the local data of the submodule in sequence to eliminate sensor noise and unify data dimensions; the cloud controller removes outliers from the global power grid data and deletes abnormal power data caused by submodule failures.

[0014] Step 130, Data Timing Synchronization: The acquisition time of the local data of the submodule and the global data of the power grid are marked by timestamps to ensure that the standardized submodule data and the standardized global data of all submodules are consistent in timing.

[0015] Preferably, step 200 includes:

[0016] Step 210, Submodule Local Prediction: Each submodule calls a preset lightweight long short-term memory model, takes the standardized submodule data as input, predicts the load demand and operating efficiency within a preset time period, and outputs the local prediction results, which include the predicted values ​​P1, P2...P for each submodule. n , where n is the number of submodules;

[0017] Step 220, Sub-module association matrix construction: The edge nodes construct the sub-module association matrix based on the physical location and historical collaborative data of the sub-modules. The correlation between each sub-module is quantified by association weight in the sub-module association matrix. The association weight of adjacent sub-modules is higher than that of non-adjacent sub-modules.

[0018] Step 230, Calculation of prediction deviation difference: Based on the sub-module association matrix, determine the associated sub-modules of each sub-module, and calculate the prediction deviation difference ΔP between the target sub-module and each associated sub-module. ij The prediction deviation difference ΔP ij The calculation formula is ΔP ij =|P i -P j |, where P i P is the local predicted value of target submodule i. j This represents the local predicted value of the associated submodule j;

[0019] Step 240, Collaborative Correction Triggering and Execution: Set the preset deviation threshold ΔP0, when the prediction deviation difference ΔP ij When the preset deviation threshold ΔP0 is reached, collaborative correction is triggered; the local prediction results are weighted and corrected using the historical prediction accuracy of the target submodule and related submodules as weights, to obtain the corrected prediction results P1', P2'...P n The formula for calculating the weighted correction is P. i '=(w i ×P i +Σw j ×P j ) / (w i +Σw j ), where w i w represents the historical prediction accuracy weights for target submodule i. j The historical prediction accuracy weight of the associated submodule j is used; the higher the historical prediction accuracy, the greater the corresponding weight.

[0020] Preferably, in step 220, the historical collaborative data is the load fluctuation trend consistency data of each sub-module within the same time period, and the association weight is determined by statistically analyzing the collaborative operation data of the sub-modules in the past preset period, which is 30 days.

[0021] Preferably, in step 210, the preset time period is 15 minutes, the input dimension of the lightweight long short-term memory model is 8, and the parameters corresponding to the input dimension include the load rate, temperature, ambient temperature and current power of the five historical time points.

[0022] Preferably, step 300 includes:

[0023] Step 310, Construction of global optimization objectives and constraints: The global optimization objective function is constructed with the objectives of maximizing global energy utilization, minimizing grid impact, and balancing sub-module load. The constraints include the maximum power supply of the grid, the maximum output power of the sub-module, and energy usage restrictions during peak and off-peak electricity price periods.

[0024] Step 320, Optimization algorithm solution: The corrected prediction result is used as the input of the global optimization objective function, and the improved particle swarm optimization algorithm is called to solve the optimal solution of the global optimization objective function. The improved particle swarm optimization algorithm shortens the calculation time by introducing a convergence acceleration factor to obtain the calibrated prediction result.

[0025] Step 330, Calibration Deviation Calculation and Feedback: Calculate the difference between the calibrated prediction result and the corrected prediction result of the corresponding submodule to obtain the calibration deviation, and feed the calibration deviation back to the corresponding submodule.

[0026] Preferably, in step 320, the convergence acceleration factor is 0.7, the maximum power supply of the power grid is the rated power supply of the charging system connected to the power grid, and the maximum output power of the submodule is the rated output power of each submodule.

[0027] Preferably, step 400 includes:

[0028] Step 410, Energy dispatch instruction generation: The energy dispatch instruction includes submodule output power allocation value, charging timing adjustment instruction and operation status warning instruction. The submodule output power allocation value is determined based on the calibrated prediction result. The charging timing adjustment instruction is used to transfer the charging demand of some submodules to the off-peak period of the power grid load. The operation status warning instruction is used to give advance notice of the risk of temperature rise in the submodule.

[0029] Step 420, Instruction Issuance and Execution: The energy scheduling instruction is issued to the controller of each submodule through the 5G industrial module. The execution delay of the control instruction does not exceed 50ms. Each submodule controller adjusts its own operating parameters according to the energy scheduling instruction.

[0030] Preferably, step 500 includes:

[0031] Step 510, actual operation data acquisition: The submodule acquires the actual output power, load rate and temperature data after executing the energy scheduling command, forms the actual operation data and uploads it to the edge node and cloud controller;

[0032] Step 520, Model and Parameter Update: The edge node calculates the error rate between the actual value of the submodule and the corrected prediction result based on the actual running data, and updates the association weights of the submodule association matrix and the preset deviation threshold; the cloud controller calculates the error rate between the actual value of the submodule and the calibrated prediction result based on the actual running data, and optimizes the parameters of the global optimization objective function and the convergence factor of the improved particle swarm optimization algorithm through the gradient descent algorithm; the submodule updates its own lightweight long short-term memory model weight parameters based on the calibration deviation.

[0033] Preferably, in step 520, the feedback iteration optimization period is consistent with the preset time period in step 210, which is 15 minutes. After each iteration update, steps 100 to 500 are re-executed to achieve dynamic optimization of the full-link model parameters.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] By synchronously collecting and preprocessing local data from submodules and global data from the power grid, and combining standardized data with local prediction models of submodules to obtain local prediction results, a submodule correlation matrix is ​​constructed to correct deviations. The prediction results are then calibrated using a global optimization objective function and constraints. Based on the calibration results, scheduling instructions are generated and executed, and the model parameters are iteratively optimized through feedback from actual operating data. This achieves the goals of improving the accuracy of energy management in the charging system, optimizing submodule load allocation, reducing grid impact, reducing submodule malfunctions, and realizing high efficiency in global energy utilization and stable system operation. Attached Figure Description

[0036] Figure 1 The main flowchart of a charging module energy management method using a predictive control strategy is provided in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please see Figure 1 This invention provides a charging module energy management method employing a predictive control strategy. The method is applied to energy allocation, load scheduling, deviation correction, and global optimization in multi-module collaborative scenarios. It integrates data acquisition and preprocessing, predictive model construction, multi-objective optimization algorithms, and feedback iteration techniques to achieve efficient and coordinated operation of the charging system and the power grid. The method includes:

[0039] Step 100, Data Acquisition and Preprocessing: Simultaneously acquire local data of the submodule and global data of the power grid, and preprocess the local data of the submodule and the global data of the power grid respectively to obtain standardized submodule data and standardized global data.

[0040] Specifically, step 100 includes:

[0041] Step 110, Multi-dimensional data acquisition: The sub-module collects its own real-time operation data, historical charging data and local environmental data through built-in sensors to form the sub-module local data; the cloud controller collects the real-time power supply capacity, peak and valley electricity price periods, power grid load warning and the system's overall sub-module operation status summary and total charging demand prediction value to form the power grid global data;

[0042] Step 120, Data Preprocessing: The submodule performs filtering and normalization processing on the local data of the submodule in sequence to eliminate sensor noise and unify data dimensions; the cloud controller removes outliers from the global power grid data and deletes abnormal power data caused by submodule failures.

[0043] Step 130, Data Timing Synchronization: The acquisition time of the local data of the submodule and the global data of the power grid are marked by timestamps to ensure that the standardized submodule data and the standardized global data of all submodules are consistent in timing.

[0044] It should be noted that in this embodiment, the local data of the submodule is the basic data supporting subsequent local predictions. Real-time operating data directly reflects the current working status of the submodule, historical charging data is used to explore load change patterns, and local environmental data is a key factor affecting charging efficiency and module lifespan. The combination of these three can improve the completeness of the data dimensions. The peak and valley electricity price data in the global power grid data is the core basis for cost-optimized scheduling, and the power grid load early warning data can avoid the risk of power grid overload in advance. The filtering process in the data preprocessing stage adopts the Kalman filter algorithm. This algorithm effectively filters out high-frequency noise generated by electromagnetic interference from sensors by establishing a dynamic system model and combining measured values ​​and predicted values ​​to iteratively correct the data. The normalization process converts data of different dimensions to the [0,1] interval to avoid the training deviation of the prediction model due to excessive differences in parameters such as power (kW level) and temperature (°C level). Data timing synchronization is a prerequisite for ensuring the coordination of subsequent "local prediction-global calibration". If there is a timing misalignment in the data of each sub-module, the deviation calculation of the related sub-modules will become meaningless, thereby affecting the accuracy of the entire energy management decision. By synchronizing the timestamps, the time error of each data can be controlled within 10ms, ensuring the timing consistency of the data.

[0045] Additionally, it should be noted that the sensors at the sub-module end can adopt a multi-sensor fusion scheme, that is, each monitoring parameter is configured with at least two different types of sensors. For example, temperature monitoring uses both thermocouples and infrared sensors. Data cross-validation further improves the accuracy of real-time operation data. In another possible implementation, the cloud controller can use the 3σ criterion to remove outliers, that is, when the data deviates from the mean by more than 3 times the standard deviation, it is judged as an outlier. Compared with the traditional threshold method, it can better adapt to the dynamic fluctuation characteristics of charging data.

[0046] For example, in one feasible implementation, a charging pile cluster comprising 20 sub-modules has each sub-module connected via an RS485 bus to built-in current sensors, voltage sensors, thermocouple temperature sensors, and humidity sensors. It collects real-time output power, input voltage, module temperature, and ambient humidity data every 10Hz, while simultaneously accessing locally stored historical data such as charging duration and charging power variation curves from the past 24 hours to form local sub-module data. The cloud obtains real-time power supply capacity of 500kW, peak-valley electricity price periods (10:00-15:00 peak, 23:00-7:00 valley), grid load warning level of "normal," and real-time status summary data from the 20 sub-modules through the power system interface to form global grid data. The submodules use Kalman filtering to process the power data, filtering out noise caused by voltage fluctuations during charging. Then, min-max normalization is used to convert all data to the [0,1] range. The cloud uses the 3σ criterion to remove the 150kW abnormal power data caused by a temporary fault in submodule 3. Finally, the Network Time Protocol (NTP) is used to add precise timestamps to all data to ensure that the data of the 20 submodules are completely synchronized with the cloud data in time, providing a reliable data foundation for subsequent collaborative correction of deviations.

[0047] Step 200, Deviation Co-correction: Based on the standardized sub-module data, the local prediction results of each sub-module are obtained through the sub-module local prediction model, a sub-module correlation matrix is ​​constructed, the prediction deviation difference between the related sub-modules is calculated in combination with the sub-module correlation matrix, and the local prediction results that exceed the preset deviation threshold are co-corrected to obtain the corrected prediction results.

[0048] Specifically, step 200 includes:

[0049] Step 210, Submodule Local Prediction: Each submodule calls a preset lightweight long short-term memory model, takes the standardized submodule data as input, predicts the load demand and operating efficiency within a preset time period, and outputs the local prediction results, which include the predicted values ​​P1, P2...P for each submodule. n , where n is the number of submodules;

[0050] Step 220, Sub-module association matrix construction: The edge nodes construct the sub-module association matrix based on the physical location and historical collaborative data of the sub-modules. The correlation between each sub-module is quantified by association weight in the sub-module association matrix. The association weight of adjacent sub-modules is higher than that of non-adjacent sub-modules.

[0051] Step 230, Calculation of prediction deviation difference: Based on the sub-module association matrix, determine the associated sub-modules of each sub-module, and calculate the prediction deviation difference ΔP between the target sub-module and each associated sub-module. ij The prediction deviation difference ΔP ij The calculation formula is ΔP ij =|P i -P j |, where P i P is the local predicted value of target submodule i. j This represents the local predicted value of the associated submodule j;

[0052] Step 240, Collaborative Correction Triggering and Execution: Set the preset deviation threshold ΔP0, when the prediction deviation difference ΔP ij When the preset deviation threshold ΔP0 is reached, collaborative correction is triggered; the local prediction results are weighted and corrected using the historical prediction accuracy of the target submodule and related submodules as weights, to obtain the corrected prediction results P1', P2'...P n The formula for calculating the weighted correction is P. i '=(w i ×P i +Σw j ×P j ) / (w i +Σw j ), where w i w represents the historical prediction accuracy weights for target submodule i. j The historical prediction accuracy weight of the associated submodule j is used; the higher the historical prediction accuracy, the greater the corresponding weight.

[0053] It should be noted that, in this embodiment, the lightweight long short-term memory model is an improved model designed for the limited edge computing power of sub-modules. Compared with the traditional long short-term memory model, it achieves lightweighting by simplifying the number of network layers (reducing from 3 layers to 2 layers) and reducing the number of hidden layer neurons (reducing from 128 to 64). While ensuring prediction accuracy, it controls the inference time to within 100ms, meeting the real-time prediction requirements of sub-modules. The sub-module correlation matrix is ​​the core carrier for realizing collaborative correction of deviations. Its essence is to quantify the correlation between sub-modules in matrix form. Sub-modules with physically adjacent locations are more correlated in terms of load change trends due to the influence of charging demand and grid branch load in the same area. Therefore, they are given higher correlation weights. This design idea is based on the principle of "spatial correlation", which can avoid mutual interference of prediction deviations of unrelated sub-modules. The historical prediction accuracy weights are calculated based on the error rate between the predicted and actual values ​​over the past 7 days. Error rate = |predicted value - actual value| / actual value × 100%. The weights are negatively correlated with the error rate; that is, the lower the error rate, the greater the weight. This design ensures that submodules with higher prediction accuracy play a greater role in collaborative correction, improving the reliability of the correction results. Through the deviation collaborative correction in step 200, the prediction deviation of a single submodule can be reduced from ±8% to ±3%, suppressing the accumulation of local deviations at the source and laying the foundation for subsequent global calibration.

[0054] The training process of the lightweight long short-term memory model includes: using standardized sub-module data from the past 90 days as the training set (divided into training, validation, and test sets in a 7:2:1 ratio), and using mean squared error (MSE) as the loss function. The loss function formula is as follows: , where m is the number of training samples; the optimizer is the Adam adaptive moment estimator, the learning rate is set to 0.001, the batch size is set to 32, the number of training iterations is set to 200, and the early stopping mechanism is triggered when the validation set loss does not decrease for 10 consecutive iterations to avoid overfitting.

[0055] Model structure details: The input layer has a dimension of 8, the hidden layer has 2 layers (64 neurons per layer), the activation function is the ReLU function, the output layer has a dimension of 2 (corresponding to load requirements and running efficiency respectively), the dropout layer (dropout rate of 0.2) is used between layers to suppress overfitting, and the output layer uses a linear activation function to ensure that the predicted values ​​conform to the actual physical quantity range.

[0056] Additionally, it should be noted that the input dimensions of the lightweight long short-term memory model can be dynamically adjusted according to the sub-module type. For the fast charging sub-module, two input parameters, "battery type" and "charging mode," are added to improve the predictive adaptability for fast charging scenarios. In another possible implementation, the preset deviation threshold ΔP0 can adopt a dynamic threshold mechanism, which is adjusted according to the grid load status. When the grid peak load is tight, ΔP0 is reduced from 5% to 3% to improve the sensitivity of deviation correction and avoid grid overload due to prediction deviation.

[0057] For example, in one feasible implementation, each of the 20 sub-modules invokes a lightweight long short-term memory model, using standardized historical load rates, temperatures, ambient temperatures, and current power as inputs, to predict the load demand for the next 15 minutes, obtaining local predicted values ​​P1=85kW, P2=78kW...P 20 =82kW. A 20×20 submodule association matrix is ​​constructed at the edge nodes. The association weight for adjacent submodules (e.g., submodule 1 and submodule 2, submodule 1 and submodule 5) is set to 0.8, and the association weight for non-adjacent submodules is set to 0.2. This weight is determined based on historical collaboration data from the past 30 days. The load fluctuation trend consistency of adjacent submodules reaches 85%, significantly higher than the 50% for non-adjacent submodules. A preset deviation threshold ΔP0 = 5% is set, and the deviation difference ΔP between submodule 1 and its associated submodule 2 is calculated. 12 =|85-78|=7kW. The average actual load demand of submodule 1 is 80kW. The deviation ratio corresponding to 7kW is 8.75%>5%, triggering collaborative correction. Checking the historical prediction accuracy, the error rate of submodule 1 is 3% (weight w1=0.97), and the error rate of submodule 2 is 2% (weight w2=0.98). Therefore, the corrected prediction value of submodule 1, P1'=(0.97×85 + 0.98×78) / (0.97+0.98)= (82.45+76.44) / 1.95=158.89 / 1.95≈81.48kW. After correction, the prediction deviation of submodule 1 is reduced to 2.98%, effectively suppressing local deviations and providing an accurate prediction basis for global energy scheduling.

[0058] In step 220, the historical collaborative data is the load fluctuation trend consistency data of each sub-module within the same time period, and the association weight is determined by statistically analyzing the collaborative operation data of the sub-modules in the past preset period, which is 30 days.

[0059] It should be noted that in this embodiment, the load fluctuation trend consistency data is the core indicator for measuring the correlation strength between sub-modules. It is obtained by calculating the Pearson correlation coefficient of the load change curves of each sub-module within the same time period. The correlation coefficient ranges from -1 to 1. The closer the value is to 1, the more consistent the load fluctuation trend and the stronger the correlation between sub-modules. The previously preset period of 30 days is the result of comprehensively considering data validity and computational cost: if the period is too short (e.g., 7 days), it is easily affected by short-term sudden factors (e.g., extreme weather, temporary events), leading to deviations in the correlation weight calculation; if the period is too long (e.g., 90 days), it will increase the data storage and computational burden and cannot reflect changes in the operating status of sub-modules in a timely manner (e.g., changes in load characteristics caused by module aging). Calculating the correlation weight using 30 days of historical collaborative data can ensure the accuracy of the weights while taking into account data processing efficiency and timeliness, enabling the sub-module correlation matrix to dynamically adapt to the operational changes of sub-modules. The core function of this step is to provide a scientific basis for screening related sub-modules for subsequent deviation difference calculations, avoid including prediction data of unrelated sub-modules in the correction scope, and ensure the pertinence and effectiveness of collaborative correction.

[0060] Additionally, it should be noted that the statistical time period of the historical collaborative data can be divided by hours, that is, the load fluctuation trend consistency data of 24 time periods in a day are calculated separately, and different correlation weights are set for different time periods. For example, the correlation weight between sub-modules during the morning peak (7:00-9:00) on weekdays is higher than that at night (0:00-6:00), which is more in line with the charging scenario characteristics of different time periods. In another possible implementation, when a sub-module is repaired or replaced, the previously preset cycle can be shortened to 15 days to speed up the update speed of the correlation weights, so that the correlation matrix can quickly adapt to the operating characteristics of the new module.

[0061] For example, in one feasible implementation, load data from 0:00 to 24:00 daily over the past 30 days is collected for 20 sub-modules. The data is divided into hourly sub-periods, and the Pearson correlation coefficient of the load curves for each sub-module within each sub-period is calculated. Taking the weekday morning peak of 7:00-8:00 as an example, the correlation coefficient between sub-module 1 and sub-module 2 (adjacent) is 0.92, and the correlation coefficient between sub-module 1 and sub-module 10 (non-adjacent) is 0.35. However, during the nighttime period of 0:00-1:00, the correlation coefficient between sub-module 1 and sub-module 2 drops to 0.65, and the correlation coefficient between sub-module 1 and sub-module 10 drops to 0.12. The association weight is calculated based on the correlation coefficient. The association weight = correlation coefficient / sum of correlation coefficients of all associated sub-modules. During the morning peak period, the association weight between sub-module 1 and sub-module 2 is 0.85, which drops to 0.72 at night. This time-differentiated association weight setting makes the sub-module association matrix more in line with the actual charging scenario, providing a scientific basis for subsequent accurate deviation correction and avoiding the problem of insufficient adaptability of fixed weights in different time periods.

[0062] In step 210, the preset time period is 15 minutes, the input dimension of the lightweight long short-term memory model is 8, and the parameters corresponding to the input dimension include the load rate, temperature, ambient temperature and current power of the five historical time points.

[0063] It should be noted that in this embodiment, the preset time period is set to 15 minutes, which is determined by combining the response requirements of the charging module's energy scheduling and the prediction accuracy: the power adjustment response time of the charging module is usually 5-10 minutes, and the 15-minute prediction time period can ensure that the scheduling command has sufficient preparation time for execution; at the same time, 15 minutes is within the short-term prediction range, and the changes in factors such as charging demand and ambient temperature are relatively stable within this period, which can ensure prediction accuracy. If the prediction time period is too long (such as 60 minutes), the prediction deviation will increase significantly due to the increase in uncertainties; if it is too short (such as 5 minutes), it will increase the prediction frequency and computational burden. The input dimension of the lightweight long short-term memory model is set to 8, which is the optimal dimension combination verified by a large number of experiments: the load rate data of the past 5 moments can capture the temporal characteristics of load changes, the temperature and ambient temperature data reflect the thermal state of the module operation, and the current power data provides a real-time benchmark. The parameters of the 8 dimensions not only cover the core factors affecting the load demand and operating efficiency of the submodule, but also avoid the problems of model overfitting and computational complexity caused by too many dimensions. This step, through precise parameter settings, enables the lightweight long short-term memory model to achieve efficient and accurate local predictions at the edge, providing reliable basic data for subsequent collaborative correction of biases.

[0064] Additionally, it should be noted that the preset time period can be dynamically adjusted according to the grid dispatching needs. When the grid issues a load adjustment warning, the preset time period can be shortened to 10 minutes to improve the response speed of prediction and dispatching. In another possible implementation, for the charging sub-module connected to the energy storage module, the "energy storage module SOC (State of Charge)" parameter can be added to the input dimension to make the prediction results more in line with the needs of the energy storage-charging collaborative operation scenario.

[0065] For example, in one feasible implementation, a charging submodule is currently at 10:00 AM. A lightweight long short-term memory model is invoked to predict the load demand and operating efficiency from 10:00 AM to 10:15 AM. The model input parameters are standardized data from five times: 9:59 AM, 9:58 AM, 9:57 AM, 9:56 AM, and 9:55 AM (82%, 80%, 78%, 79%, and 81%, respectively), module temperature (42°C, 41°C, 40°C, 40°C, and 41°C, respectively), ambient temperature (28°C), and current power (80kW), totaling eight dimensions. The model performs inference calculations through two hidden layers (64 neurons per layer), outputting a predicted load demand of 83kW and a predicted operating efficiency of 92% for the period from 10:00 AM to 10:15 AM. Subsequent actual operation verification showed that the actual load demand during this period was 82.5kW, with a prediction deviation of only 0.61%, which is far lower than the 5% deviation of the unoptimized model. This fully demonstrates the rationality of the input dimension and the preset time period setting, and provides high-quality local prediction results for subsequent collaborative correction of deviations.

[0066] Step 300, Global Prediction Calibration: Construct a global optimization objective function and constraints based on the standardized global data, substitute the corrected prediction result into the global optimization objective function, and obtain the calibrated prediction result through an optimization algorithm. At the same time, calculate the calibration deviation between the calibrated prediction result and the corrected prediction result.

[0067] Specifically, step 300 includes:

[0068] Step 310, Construction of global optimization objectives and constraints: The global optimization objective function is constructed with the objectives of maximizing global energy utilization, minimizing grid impact, and balancing sub-module load. The constraints include the maximum power supply of the grid, the maximum output power of the sub-module, and energy usage restrictions during peak and off-peak electricity price periods.

[0069] Step 320, Optimization algorithm solution: The corrected prediction result is used as the input of the global optimization objective function, and the improved particle swarm optimization algorithm is called to solve the optimal solution of the global optimization objective function. The improved particle swarm optimization algorithm shortens the calculation time by introducing a convergence acceleration factor to obtain the calibrated prediction result.

[0070] Step 330, Calibration Deviation Calculation and Feedback: Calculate the difference between the calibrated prediction result and the corrected prediction result of the corresponding submodule to obtain the calibration deviation, and feed the calibration deviation back to the corresponding submodule.

[0071] It should be noted that in this embodiment, the global optimization objective function is the core mathematical model for achieving multi-objective collaborative optimization. It integrates the three objectives into a single objective function through weighted summation. The weight coefficients are dynamically adjusted according to the grid operating status and operational needs. For example, during peak grid periods, the weight of "minimizing grid impact" is increased to 0.4, and during off-peak periods, the weight of "maximizing global energy utilization" is increased to 0.5, ensuring that the objective function aligns with actual needs. The improved particle swarm optimization algorithm addresses the shortcomings of traditional particle swarm optimization algorithms, such as slow convergence speed and susceptibility to local optima. By introducing a convergence acceleration factor to adjust the particle flight speed update formula, the algorithm maintains strong global search capabilities in the early stages of iteration and converges quickly to the optimal solution in the later stages, reducing the solution time from 500ms in the traditional algorithm to less than 200ms, meeting the real-time requirements of global scheduling. The calibration deviation is not only important data for feedback iteration but also reflects the matching degree between the local prediction of a submodule and the global constraints. If the calibration deviation of a certain submodule remains large, it indicates that its local prediction model may have an adaptability problem and requires focused optimization. The global prediction calibration in step 300 ensures that the energy dispatch command meets the prediction requirements of each submodule, as well as the grid constraints and global optimization objectives, thus avoiding energy waste caused by conflicts between local demands and global constraints.

[0072] The core improvement of the improved particle swarm optimization algorithm lies in embedding the convergence acceleration factor into the particle velocity update formula, specifically the update formula as follows: Where λ is the convergence acceleration factor (ranging from 0.5 to 0.9, and 0.7 in this embodiment), ω is the inertia weight (dynamically adjusted, initially 0.9, then linearly reduced to 0.4), c1 and c2 are learning factors (2.0 and 2.0 respectively), r1 and r2 are random numbers in the interval [0,1], and v i x is the particle's current velocity. i pbest represents the particle's current position. i gbest represents the optimal position for an individual particle, while gbest represents the globally optimal position for the particle swarm.

[0073] The particle fitness function directly corresponds to the global optimization objective function, and is defined as: α, β, and γ are target weighting coefficients (dynamically adjusted according to the grid status; peak segment β=0.4, α=0.3, γ=0.3; valley segment α=0.5, β=0.2, γ=0.3); the grid impact coefficient is the ratio of the actual total power to the grid's rated power supply (calculated as 1 if it exceeds 1); and the submodule load balancing coefficient is the ratio of the standard deviation of the load rate of each submodule to the average load rate.

[0074] Additionally, it should be noted that the global optimization objective function can be supplemented with a "minimizing charging costs" objective, which combines peak and off-peak electricity price data to calculate charging costs, making the optimization results more aligned with the operator's economic benefit needs. In another possible implementation, when the charging system is connected to distributed photovoltaics, a "photovoltaic output limit" parameter can be added to the constraints to ensure that global scheduling makes full use of photovoltaic energy and reduces dependence on the power grid.

[0075] For example, in one feasible implementation, the cloud controller obtains a total of 1650kW after correction of the prediction results of 20 sub-modules, while the maximum power supply of the grid is 1500kW. Load reduction and balanced distribution need to be achieved through global calibration. A global optimization objective function is constructed, and the weight coefficients of the grid peak period (10:00-15:00) are set as follows: grid impact minimization (0.4), global energy utilization maximization (0.3), and sub-module load balancing (0.3). The constraints are: total power ≤ 1500kW, power of a single sub-module ≤ 100kW, and peak power not exceeding 1.5 times that of the valley period. An improved particle swarm optimization algorithm (with a convergence acceleration factor of 0.7) was applied. Using the corrected prediction results as initial input, the optimal solution was obtained after 50 iterations. The total corrected prediction result was 1495kW, with the corrected prediction value of submodule 1 being 78kW (corrected prediction value 81.48kW, calibration deviation -3.48kW) and the corrected prediction value of submodule 2 being 75kW (corrected prediction value 76kW, calibration deviation -1kW). The power distribution of all submodules met the constraints, and the load rate difference was controlled within 5%. The calibration deviation of each submodule was fed back to the corresponding module to provide a basis for subsequent model parameter updates, ensuring that subsequent predictions better fit the global constraint requirements.

[0076] In step 320, the convergence acceleration factor is 0.7, the maximum power supply of the power grid is the rated power supply of the charging system connected to the power grid, and the maximum output power of the submodule is the rated output power of each submodule.

[0077] It should be noted that in this embodiment, the convergence acceleration factor is a core parameter of the improved particle swarm optimization algorithm. Its function is to adjust the step size of particle velocity updates, balancing the algorithm's global search capability and local convergence capability. Extensive experimental verification shows that when the convergence acceleration factor is 0.7, the algorithm converges to the optimal solution within 50 iterations, avoiding the problems of particle oscillation and failure to converge due to excessively large values ​​(e.g., 1.0), or excessively slow convergence speed due to excessively small values ​​(e.g., 0.4), thus achieving a balance between solution efficiency and accuracy. The maximum power supply of the power grid is the rated power supply of the charging system connected to the grid. This value is determined by the power grid company based on the capacity approval of the charging system and is an inviolable hard constraint. If the dispatch command exceeds this power, it will cause the grid branch to overload and trip. Therefore, using it as a core constraint ensures the safety of system operation. The maximum output power of the submodule is its rated output power, which is determined by the hardware specifications of the submodule. Exceeding this power will cause overheating and damage to the internal components of the module. Therefore, this constraint protects the safety of the submodule equipment. By clarifying the values ​​and basis of these parameters, the constraints of the global optimization solution are made clearer and the parameters are more accurate, ensuring the feasibility and security of the prediction results after calibration.

[0078] Additionally, it should be noted that the convergence acceleration factor can be adaptively adjusted, with a value of 0.9 in the early stages of algorithm iteration (enhancing global search) and a value reduced to 0.5 in the later stages (accelerating local convergence), further improving the algorithm's solution performance. In another possible implementation, when the power grid experiences an emergency power supply limitation, the maximum power supply of the power grid can be temporarily adjusted to 80% of the rated value to ensure the safe and stable operation of the power grid.

[0079] For example, in one feasible implementation, the rated power of a charging system connected to the power grid is 1500kW (i.e., the maximum power of the power grid), and the rated output power of each of the 20 sub-modules is 100kW (i.e., the maximum output power of the sub-modules). During global optimization, the convergence acceleration factor of the improved particle swarm optimization algorithm is initially set to 0.7, and the initial particle velocity is set based on the difference between the corrected prediction result and the constraint conditions. During iteration, the algorithm adjusts the particle flight direction and step size through the acceleration factor. By the 10th iteration, the particle swarm is close to the optimal solution region, and by the 45th iteration, the fitness values ​​(objective function values) of all particles tend to stabilize. The sum of the calibrated prediction results obtained is 1495kW, and the calibrated values ​​of individual sub-modules are all between 80-90kW, neither exceeding the maximum power of the power grid nor exceeding the rated output power of the sub-modules. If the traditional particle swarm optimization algorithm (without acceleration factor) is used, it requires 120 iterations to converge, and the total power of the solution fluctuates greatly. This fully demonstrates the rationality of the convergence acceleration factor of 0.7 and the advantages of the improved algorithm, ensuring that global calibration is efficient and safe.

[0080] Step 400, Energy scheduling execution: Generate energy scheduling instructions based on the calibrated prediction results, and send the energy scheduling instructions to each submodule for execution.

[0081] Specifically, step 400 includes:

[0082] Step 410, Energy dispatch instruction generation: The energy dispatch instruction includes submodule output power allocation value, charging timing adjustment instruction and operation status warning instruction. The submodule output power allocation value is determined based on the calibrated prediction result. The charging timing adjustment instruction is used to transfer the charging demand of some submodules to the off-peak period of the power grid load. The operation status warning instruction is used to give advance notice of the risk of temperature rise in the submodule.

[0083] Step 420, Instruction Issuance and Execution: The energy scheduling instruction is issued to the controller of each submodule through the 5G industrial module. The execution delay of the control instruction does not exceed 50ms. Each submodule controller adjusts its own operating parameters according to the energy scheduling instruction.

[0084] It should be noted that in this embodiment, the energy scheduling command is the core carrier connecting global decision-making and submodule execution. The three types of commands correspond to the three core requirements of "power allocation," "timing optimization," and "safety warning," forming a comprehensive scheduling system. The submodule output power allocation value is directly generated based on the calibrated prediction results, ensuring that power allocation accurately matches the global optimization target. The charging timing adjustment command is generated by combining peak-valley electricity prices and grid load prediction. By transferring some charging demand during peak periods to valley periods, charging costs can be reduced while smoothing grid load, reflecting the energy management concept of "peak shaving and valley filling." The operating status warning command is generated based on predicted submodule temperature changes, providing a 10-15 minute advance warning of temperature rise risk, reserving adjustment time for the submodule's thermal management system, and avoiding failures due to overheating. The 5G industrial module features low latency and high reliability, with a transmission rate exceeding 1Gbps and end-to-end latency controlled within 50ms, ensuring that scheduling commands can be quickly and accurately sent to each submodule, avoiding scheduling deviations due to command delays, and guaranteeing the real-time performance of energy management.

[0085] Additionally, it should be noted that the energy dispatch command can be supplemented with a "V2G (Vehicle to Grid) dispatch command". When the power grid experiences a load gap, the command instructs electric vehicles to discharge to the power grid through a submodule to participate in power grid frequency regulation and peak shaving. In another possible implementation, the command issuance can adopt a "5G + edge computing" dual-link backup mechanism. When the 5G link fails, it automatically switches to the local communication link of the edge node to ensure the reliability of the command issuance.

[0086] For example, in one feasible implementation, the cloud controller generates energy scheduling instructions based on the calibrated prediction results: the output power allocation value of submodule 1 is 78kW, submodule 2 is 75kW... and submodule 20 is 82kW; for the 5kW excess demand of submodule 3 (corrected prediction value is 90kW, calibrated value is 85kW), a charging timing adjustment instruction is generated to transfer this part of the demand to 23:00 (valley time in the power grid); based on the temperature prediction result of submodule 15 (it will rise to 48℃ in 15 minutes, close to the warning threshold of 50℃), an operation status warning instruction is generated, prompting to reduce the power by 2kW in advance. These instructions are sent to the controllers of each submodule through a 5G industrial module (using SA standalone networking mode). The test showed that the instruction transmission delay was 32ms, far below the upper limit of 50ms. After receiving the instruction, the controller of submodule 1 smoothly adjusted the output power from the current 80kW to 78kW through the PID (proportional-integral-derivative) adjustment algorithm, with no power fluctuation during the adjustment process; submodule 15 started the cooling fan in advance to accelerate, and the actual temperature was 46℃ after 15 minutes, successfully avoiding the risk of overheating, fully demonstrating the accuracy and effectiveness of the energy dispatch instruction.

[0087] Step 500, Feedback Iterative Optimization: Collect the actual operating data of each submodule after executing the energy scheduling command, and update the parameters of the submodule local prediction model, submodule correlation matrix and optimization algorithm by combining the actual operating data and the calibration deviation.

[0088] Specifically, step 500 includes:

[0089] Step 510, actual operation data acquisition: The submodule acquires the actual output power, load rate and temperature data after executing the energy scheduling command, forms the actual operation data and uploads it to the edge node and cloud controller;

[0090] Step 520, Model and Parameter Update: The edge node calculates the error rate between the actual value of the submodule and the corrected prediction result based on the actual running data, and updates the association weights of the submodule association matrix and the preset deviation threshold; the cloud controller calculates the error rate between the actual value of the submodule and the calibrated prediction result based on the actual running data, and optimizes the parameters of the global optimization objective function and the convergence factor of the improved particle swarm optimization algorithm through the gradient descent algorithm; the submodule updates its own lightweight long short-term memory model weight parameters based on the calibration deviation.

[0091] It should be noted that in this embodiment, step 500 is the core link in realizing the system's "self-evolution." By comparing the actual operating data with the prediction and calibration results, the entire link's model and parameters are optimized in reverse, solving the problem of decreased adaptability caused by fixed model parameters in traditional methods. The error rate is the core indicator for measuring the accuracy of prediction and calibration. It is calculated as: Error Rate = |Actual Value - Predicted / Calibrated Value| / Actual Value × 100%. The magnitude of the error rate directly reflects the rationality of the model and parameters. The edge nodes update the association weights of the sub-module association matrix, making the quantification of association relationships more in line with the actual operational changes of the sub-modules. For example, if the charging demand characteristics of a sub-module change due to the development of the surrounding commercial area, its association weights with adjacent modules will be automatically adjusted through error rate feedback. The cloud optimizes the global objective function parameters and algorithm convergence factors, ensuring that the global optimization is more in line with the dynamic needs of the power grid and the system. The sub-modules update the weights of the lightweight long short-term memory model, which can improve the accuracy and specificity of local predictions. Through this full-link feedback iteration, the energy management accuracy of the system will continue to improve with the increase of operating time, realizing a closed-loop cycle of "prediction-decision-execution-optimization".

[0092] Additionally, it should be noted that the model and parameter updates can adopt an "incremental update" mechanism, where parameter updates are only performed when the error rate exceeds a preset threshold (e.g., 5%), thus avoiding system fluctuations caused by frequent updates. In another possible implementation, the concept of "transfer learning" can be introduced, whereby when a new submodule is added, the model parameters of the new module are initialized using the update experience of the existing submodules, thereby shortening the adaptation cycle of the new module.

[0093] For example, in one feasible implementation, after 20 submodules execute energy dispatch instructions for 15 minutes, actual operating data is collected: the actual output power of submodule 1 is 77.5kW (corrected predicted value 81.48kW, error rate 5.13%; calibrated value 78kW, error rate 0.65%); the actual temperature of submodule 15 is 46℃ (predicted temperature 48℃, error rate 4.35%). The error rate of edge node computing submodule 1 exceeded the preset 3% threshold. Therefore, its association weight with submodule 2 was adjusted from 0.8 to 0.75, and the preset deviation threshold was fine-tuned from 5% to 4.8%. The calibration error rate of cloud computing submodule 1 was 0.65%. Using gradient descent, the weight of "submodule load balancing" in the global objective function was increased from 0.3 to 0.32, and the convergence factor of the improved particle swarm optimization algorithm was fine-tuned from 0.7 to 0.71. Based on the calibration deviation (-3.48kW), the hidden layer weights of the lightweight long short-term memory model in submodule 1 were updated using backpropagation, making the model's predicted output closer to the calibration requirements. After 10 such feedback iterations, the local prediction deviation of submodule 1 decreased to 2.1%, and the calibration deviation decreased to 0.3%. The energy management accuracy of the system was significantly improved, fully demonstrating the core value of feedback iterative optimization.

[0094] In step 520, the feedback iteration optimization period is the same as the preset time period in step 210, which is 15 minutes. After each iteration update, steps 100 to 500 are re-executed to achieve dynamic optimization of the full-link model parameters.

[0095] It should be noted that in this embodiment, the feedback iteration optimization cycle is set to 15 minutes, consistent with the preset time period of the submodule's local prediction, forming a synchronous cyclic rhythm of "prediction-calibration-execution-optimization". The core idea of ​​this design is to ensure that each optimization is based on the latest operating data, promptly correcting the prediction and calibration deviations of the previous cycle, and avoiding the accumulation of deviations. If the iteration cycle is longer than the preset time period (e.g., 30 minutes), the deviations of the previous cycle cannot be corrected in time, affecting the prediction accuracy of the next cycle; if it is shorter than the preset time period (e.g., 5 minutes), insufficient accumulation of actual operating data will lead to insufficient basis for optimization parameters, resulting in over-optimization. After each iteration update, steps 100 to 500 are re-executed, constructing a dynamic optimization closed loop across the entire chain. This allows the local prediction of the submodule, the deviation correction of the edge node, and the global calibration in the cloud to form a synergistic evolution, rather than isolated optimization, ensuring the continuous improvement of the overall energy management performance of the system. This cycle setting takes into account both the timeliness and reliability of optimization, and keeps the computational burden of the system within a reasonable range, achieving a balance between performance and efficiency.

[0096] Additionally, it should be noted that the feedback iteration optimization cycle can adopt a "dynamic cycle" mechanism. When the system error rate is less than 2%, the cycle can be extended to 20 minutes to reduce the computational burden; when the error rate is greater than 5%, the cycle can be shortened to 10 minutes to speed up the optimization. In another possible implementation, when the charging system is in a low-load operation period (such as at night), the iteration cycle can be extended to 30 minutes to save energy and computing power consumption while ensuring the optimization effect.

[0097] For example, in one feasible implementation, the charging system starts operating at 7:00 AM. At 7:00 AM, steps 100-400 are executed: the submodule predicts the load demand from 7:00 AM to 7:15 AM, and after deviation correction and global calibration, issues scheduling instructions. At 7:15 AM, step 500 is executed: actual operating data from 7:00 AM to 7:15 AM is collected, the error rate is calculated, and each model and parameter is updated, completing the first feedback iteration. At 7:15 AM, step 100 is immediately re-executed to collect real-time data from 7:15 AM, predict the load demand from 7:15 AM to 7:30 AM, and enter the next cycle. During the morning peak from 7:00 AM to 9:00 AM, the system error rate briefly rises to 4.2%, triggering a dynamic cycle mechanism that shortens the iteration cycle to 10 minutes. After three rapid iterations, the error rate drops to 1.8%, and then the 15-minute cycle is restored. Throughout the operation, the system used a closed-loop iteration with a 15-minute benchmark to gradually reduce the global prediction error from the initial 3.2% to below 1.5%, and increase the energy utilization rate from 90% to 98.2%, which fully demonstrates the rationality of the iteration cycle setting and achieves continuous optimization of system performance.

[0098] In this embodiment, by synchronously collecting and preprocessing local data from submodules and global data from the power grid, and combining standardized data with a local prediction model of the submodule to obtain local prediction results, a submodule correlation matrix is ​​constructed to correct deviations. The prediction results are then calibrated using a global optimization objective function and constraints. Based on the calibration results, scheduling instructions are generated and executed, and the model parameters are iteratively optimized through feedback from actual operating data. This achieves the effects of improving the accuracy of energy management in the charging system, optimizing submodule load allocation, reducing power grid impact, reducing submodule malfunctions, and realizing high efficiency in global energy utilization and stable system operation.

[0099] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0100] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A charging module energy management method employing a predictive control strategy, applied to a charging system comprising multiple sub-modules, characterized in that, Includes the following steps: Step 100, Data Acquisition and Preprocessing: Simultaneously acquire local data of the submodule and global data of the power grid, and preprocess the local data of the submodule and the global data of the power grid respectively to obtain standardized submodule data and standardized global data; Step 200, Deviation Cooperative Correction: Based on the standardized sub-module data, the local prediction results of each sub-module are obtained through the sub-module local prediction model, a sub-module correlation matrix is ​​constructed, the prediction deviation difference between the related sub-modules is calculated in combination with the sub-module correlation matrix, and the local prediction results that exceed the preset deviation threshold are cooperatively corrected to obtain the corrected prediction results. Step 300, Global Prediction Calibration: Construct a global optimization objective function and constraints based on the standardized global data, substitute the corrected prediction result into the global optimization objective function, solve the calibrated prediction result through the optimization algorithm, and calculate the calibration deviation between the calibrated prediction result and the corrected prediction result. Step 400, Energy scheduling execution: Generate energy scheduling instructions based on the calibrated prediction results, and send the energy scheduling instructions to each submodule for execution; Step 500, Feedback Iterative Optimization: Collect the actual operating data of each submodule after executing the energy scheduling command, and update the parameters of the submodule local prediction model, submodule correlation matrix and optimization algorithm by combining the actual operating data and the calibration deviation; Step 200 includes: Step 210, Submodule Local Prediction: Each submodule calls a preset lightweight long short-term memory model, takes the standardized submodule data as input, predicts the load demand and operating efficiency within a preset time period, and outputs the local prediction results, which include the predicted values ​​P1, P2...P for each submodule. n , where n is the number of submodules; Step 220, Sub-module association matrix construction: The edge nodes construct the sub-module association matrix based on the physical location and historical collaborative data of the sub-modules. The correlation between each sub-module is quantified by association weight in the sub-module association matrix. The association weight of adjacent sub-modules is higher than that of non-adjacent sub-modules. Step 230, Calculation of prediction deviation difference: Based on the sub-module association matrix, determine the associated sub-modules of each sub-module, and calculate the prediction deviation difference ΔP between the target sub-module and each associated sub-module. ij The prediction deviation difference ΔP ij The calculation formula is ΔP ij =|P i -P j |, where P i P is the local predicted value of target submodule i. j This represents the local predicted value of the associated submodule j; Step 240, Collaborative Correction Triggering and Execution: Set the preset deviation threshold ΔP0, when the prediction deviation difference ΔP ij When the preset deviation threshold ΔP0 is reached, collaborative correction is triggered; the local prediction results are weighted and corrected using the historical prediction accuracy of the target submodule and related submodules as weights, to obtain the corrected prediction results P1', P2'...P n The formula for calculating the weighted correction is P. i '=(w i ×P i +Σw j ×P j ) / (w i +Σw j ), where w i w represents the historical prediction accuracy weights for target submodule i. j The historical prediction accuracy weight of the associated submodule j is used; the higher the historical prediction accuracy, the greater the corresponding weight.

2. The method according to claim 1, characterized in that, Step 100 includes: Step 110, Multi-dimensional data acquisition: The sub-module collects its own real-time operation data, historical charging data and local environmental data through built-in sensors to form the sub-module local data; the cloud controller collects the real-time power supply capacity, peak and valley electricity price periods, power grid load warning and the system's overall sub-module operation status summary and total charging demand prediction value to form the power grid global data; Step 120, Data Preprocessing: The submodule performs filtering and normalization processing on the local data of the submodule in sequence to eliminate sensor noise and unify data dimensions; the cloud controller removes outliers from the global power grid data and deletes abnormal power data caused by submodule failures. Step 130, Data Timing Synchronization: The acquisition time of the local data of the submodule and the global data of the power grid are marked by timestamps to ensure that the standardized submodule data and the standardized global data of all submodules are consistent in timing.

3. The method according to claim 1, characterized in that, In step 220, the historical collaborative data is the load fluctuation trend consistency data of each sub-module within the same time period, and the correlation weight is determined by statistically analyzing the collaborative operation data of the sub-modules in the past preset period, which is 30 days.

4. The method according to claim 1, characterized in that, In step 210, the preset time period is 15 minutes, the input dimension of the lightweight long short-term memory model is 8, and the parameters corresponding to the input dimension include the load rate, temperature, ambient temperature and current power of the five historical time points.

5. The method according to claim 1, characterized in that, Step 300 includes: Step 310, Construction of global optimization objectives and constraints: The global optimization objective function is constructed with the objectives of maximizing global energy utilization, minimizing grid impact, and balancing sub-module load. The constraints include the maximum power supply of the grid, the maximum output power of the sub-module, and energy usage restrictions during peak and off-peak electricity price periods. Step 320, Optimization algorithm solution: The corrected prediction result is used as the input of the global optimization objective function, and the improved particle swarm optimization algorithm is called to solve the optimal solution of the global optimization objective function. The improved particle swarm optimization algorithm shortens the calculation time by introducing a convergence acceleration factor to obtain the calibrated prediction result. Step 330, Calibration Deviation Calculation and Feedback: Calculate the difference between the calibrated prediction result and the corrected prediction result of the corresponding submodule to obtain the calibration deviation, and feed the calibration deviation back to the corresponding submodule.

6. The method according to claim 5, characterized in that, In step 320, the convergence acceleration factor is 0.7, the maximum power supply of the power grid is the rated power supply of the charging system connected to the power grid, and the maximum output power of the submodule is the rated output power of each submodule.

7. The method according to claim 1, characterized in that, Step 400 includes: Step 410, Energy dispatch instruction generation: The energy dispatch instruction includes submodule output power allocation value, charging timing adjustment instruction and operation status warning instruction. The submodule output power allocation value is determined based on the calibrated prediction result. The charging timing adjustment instruction is used to transfer the charging demand of some submodules to the off-peak period of the power grid load. The operation status warning instruction is used to give advance notice of the risk of temperature rise in the submodule. Step 420, Instruction Issuance and Execution: The energy scheduling instruction is issued to the controller of each submodule through the 5G industrial module. The execution delay of the control instruction does not exceed 50ms. Each submodule controller adjusts its own operating parameters according to the energy scheduling instruction.

8. The method according to claim 5, characterized in that, Step 500 includes: Step 510, actual operation data acquisition: The submodule acquires the actual output power, load rate and temperature data after executing the energy scheduling command, forms the actual operation data and uploads it to the edge node and cloud controller; Step 520, Model and Parameter Update: The edge node calculates the error rate between the actual value of the submodule and the corrected prediction result based on the actual running data, and updates the association weights of the submodule association matrix and the preset deviation threshold; the cloud controller calculates the error rate between the actual value of the submodule and the calibrated prediction result based on the actual running data, and optimizes the parameters of the global optimization objective function and the convergence factor of the improved particle swarm optimization algorithm through the gradient descent algorithm; the submodule updates its own lightweight long short-term memory model weight parameters based on the calibration deviation.

9. The method according to claim 8, characterized in that, In step 520, the feedback iteration optimization period is consistent with the preset time period in step 210, which is 15 minutes. After each iteration update, steps 100 to 500 are re-executed to achieve dynamic optimization of the full-link model parameters.

Images