Intelligent energy management system of multi-charging pile micro-grid combining photovoltaic and energy storage

Abstract

The application relates to the technical field of energy management systems, and discloses a multi-charging-pile micro-grid intelligent energy management system integrating photovoltaics and energy storage, which comprises a cloud layer central server, an edge layer regional server and a plurality of intelligent charging piles in a terminal layer; wherein the cloud layer central server is provided with a counter-environment simulator and a federal meta-learning aggregator; the counter-environment simulator is constructed based on a generative adversarial network and is used for learning desensitization feature distribution uploaded by each intelligent charging pile and generating multi-modal virtual scene data conforming to real statistical rules. The multi-charging-pile micro-grid intelligent energy management system integrating photovoltaics and energy storage realizes whole-process protection of terminal data privacy, simultaneously adapts to data sparse working conditions of newly put-in charging piles, completes fine adjustment of a lightweight model on the terminal side, finally effectively improves the overall energy utilization efficiency of the multi-charging-pile micro-grid, and provides a more optimal technical scheme for intelligent energy management in the field.

Description

Technical Field

[0001] This invention relates to the field of energy management system technology, and in particular to a smart energy management system for multi-charging pile microgrids that integrates photovoltaics and energy storage. Background Technology

[0002] In the current field of smart energy management for microgrids integrating photovoltaics and energy storage, mainstream technical solutions mostly adopt cloud-based centralized model training or traditional federated learning modes to achieve charging pile load prediction and photovoltaic energy storage collaborative scheduling. That is, a globally unified energy management prediction model is trained by a cloud server, and then the model is distributed to each terminal charging pile. The terminal uses this model to complete charging load prediction, photovoltaic output adaptation, and energy storage battery charging and discharging strategy formulation. Some technical solutions introduce adversarial generative networks to generate virtual scene data to assist cloud model training, and also collect terminal charging pile operation data and upload it to the cloud for global model iterative optimization. However, the virtual data generation of such solutions is mostly based on full feature learning, and the terminal data upload is mostly simple processing, without dedicated lightweight design and privacy protection optimization for charging pile scenarios.

[0003] In practical applications, the risk of data privacy leakage is high. Existing solutions often upload raw terminal operating data directly or after simple processing, which can easily leak private information such as user charging details, accurate operating conditions of charging piles, and regional charging load patterns. An effective privacy protection mechanism has not been formed at the source. The problem of data sparsity of newly deployed charging piles is difficult to solve. The global model in the cloud lacks adaptability to local operating conditions. Terminals need to rely on massive amounts of operating data to complete model optimization, and cannot quickly generate personalized prediction models that fit local operating conditions. In addition, the model and terminal hardware adaptability is poor. Traditional prediction models have complex structures and large computational loads. The embedded CPUs and lightweight computing modules of ordinary charging piles cannot support local model fine-tuning. Adding extra hardware will increase deployment costs and cannot be dynamically adjusted according to terminal operating conditions, resulting in a continuous decline in adaptability. The insufficient adaptability of the model to local operating conditions leads to large load prediction errors, which in turn causes unreasonable energy allocation within the microgrid and low energy utilization efficiency. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing smart energy management scheme for microgrids with multiple charging piles that integrate photovoltaic and energy storage has the disadvantages of sparse data of newly commissioned charging piles, which makes it difficult to quickly adapt to local operating conditions, and low scheduling and matching accuracy of energy storage charging and discharging with charging load. To address this, we propose a smart energy management system for microgrids with multiple charging piles that integrate photovoltaic and energy storage.

[0005] To achieve the above objectives, this application adopts the following technical solution: a smart energy management system for a multi-charging pile microgrid integrating photovoltaic and energy storage, comprising a cloud-based central server, edge-layer regional servers, and multiple smart charging piles at the terminal layer; wherein, the cloud-based central server is deployed with an adversarial environment simulator and a federated meta-learning aggregator; the adversarial environment simulator is constructed based on a generative adversarial network and is used to learn the distribution of desensitized features uploaded by each of the smart charging piles, and generate multimodal virtual scene data that conforms to real statistical laws; the federated meta-learning aggregator is connected to the adversarial environment simulator and is used to receive the virtual scene data for training, and generate model initialization parameters that can be quickly fine-tuned based on a small amount of data. The system includes a lightweight local prediction model within the smart charging pile. This model receives the initialization parameter set to initialize the model and performs minor adjustments using locally collected operational data. This generates a personalized prediction model adapted to local operating conditions, with all data processing for the adjustment process completed locally. The system also includes an asynchronous federated feedback module across the cloud and terminal layers. This module encrypts and uploads the locally adjusted performance gain indicators and abstract feature distribution parameters of each smart charging pile to the central cloud server. This drives the adversarial environment simulator to update the feature distribution, the federated meta-learning aggregator to optimize the parameter set generation logic, and the optimized model initialization parameter set to be redistributed to the smart charging pile.

[0006] Preferably, the adversarial environment simulator performs desensitization processing on the data received from the terminal charging pile. The desensitization processing includes data normalization, feature abstraction and extraction, and privacy masking to obtain feature distribution parameters that characterize the operating rules of the charging pile.

[0007] Preferably, the adversarial environment simulator includes a generator and a discriminator built based on a generative adversarial network; the discriminator is configured to distinguish between the real desensitized feature distribution and the feature distribution generated by the generator; the generator is configured to adjust its generation logic according to the discrimination result of the discriminator until the feature distribution output by the generator meets a preset standard of statistical distribution consistency.

[0008] Preferably, the adversarial environment simulator includes a generator and a discriminator built based on a generative adversarial network; the discriminator is configured to distinguish between the real desensitized feature distribution and the simulated feature distribution generated by the generator; the generator is configured to dynamically adjust its generation logic according to the discrimination result of the discriminator until the simulated feature distribution output by the generator is highly fitted to the real desensitized feature distribution, reaching a preset standard of statistical distribution consistency.

[0009] Preferably, the federated meta-learning aggregator preprocesses the received virtual scene data, splitting the data into multiple sample sets based on a single virtual charging pile, and performing lightweight screening and standardization on each sample set.

[0010] Preferably, during the training process of the federated meta-learning aggregator based on the sample set, if the training result does not reach the threshold, a feedback signal is sent to the adversarial environment simulator to trigger the generation of supplementary virtual scene data of the corresponding type.

[0011] Preferably, the lightweight local prediction model receives the model initialization parameter set as a basis and only performs a limited number of iterative adjustments on the parameters related to the local operating conditions within the model to complete the lightweight fine-tuning.

[0012] Preferably, the local operating data on which the lightweight local prediction model is fine-tuned includes charging load data, photovoltaic output data, energy storage status data, and local environmental data.

[0013] Preferably, the asynchronous federated feedback module encrypts the extracted performance gain indicators and highly abstract feature distribution parameters respectively, and transmits them asynchronously through an encrypted communication channel; the edge layer regional server is configured to provide data caching and relay support during the data transmission process, and triggers a retransmission mechanism when a transmission interruption is detected.

[0014] Preferably, the asynchronous federated feedback module classifies and summarizes the decrypted feedback data on the central server side of the cloud layer to form a feature distribution dataset for updating the adversarial environment simulator and a model performance analysis dataset for optimizing the federated meta-learning aggregator.

[0015] The technical effects and advantages of this invention are as follows: By realizing full-process protection of terminal data privacy and adapting to the data sparse working conditions of newly deployed charging piles, this invention completes the fine-tuning of the lightweight model on the terminal side, allowing the model to be dynamically optimized according to the actual working conditions of the terminal. This significantly improves the model's adaptability to local working conditions and the scheduling and matching accuracy of photovoltaic energy storage and charging load, reduces the terminal hardware deployment cost, and ultimately effectively improves the overall energy utilization efficiency of multi-charging pile microgrids, providing a better technical solution for smart energy management in this field. Attached Figure Description

[0016] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0017] Figure 1 This is a flowchart of the cloud-based training and distribution process of the present invention; Figure 2This is a flowchart of the terminal fine-tuning and feedback evolution process of the present invention. Detailed Implementation

[0018] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0019] Reference Figure 1-2 As shown, this invention provides a technical solution: a smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage. This system includes an adversarial environment simulator deployed on a cloud-based central server. This simulator is a simulation training module built upon a Generative Adversarial Network (GAN), whose core function is to generate massive amounts of multimodal virtual challenging scenario data that conform to real statistical laws, providing diverse sample support for the training of the federated meta-learning aggregator. The adversarial environment simulator only learns the feature distribution of anonymized data uploaded by charging piles in each region, without acquiring any original operational data, thus avoiding privacy leakage risks from the source.

[0020] After the adversarial environment simulator starts, it receives anonymized data feature distribution parameters uploaded by charging piles in various regions through an upgraded asynchronous federated feedback module. The anonymization process employs data normalization, feature abstraction and extraction, and privacy masking techniques to remove private information such as user charging details, precise charging pile locations, and raw real-time grid dispatch data. Only core features of charging pile operation patterns are retained, including peak and off-peak charging load distribution characteristics, the correlation between traffic flow and charging demand, the impact of weather factors on charging power, charging load fluctuation characteristics during large-scale events, and the matching characteristics between photovoltaic output and charging load. The adversarial environment simulator then summarizes, removes redundancy, and standardizes the received anonymized feature parameters from multiple regions to form a unified feature distribution dataset. The calculation formula is as follows: ;in The normalized feature data takes values ​​in the range [0,1]. The original feature data to be normalized; This is the global minimum value under this feature dimension; This represents the global maximum value within this feature dimension; it provides standardized input for subsequent feature learning and ensures that the data processing meets data compliance requirements. When all data in that feature dimension are completely consistent and without any fluctuations, the denominator of the formula is 0, which has no mathematical meaning and cannot be calculated. For invalid features that are consistent across the entire dataset, they should be directly removed during the preprocessing stage and not included in the normalization calculation process.

[0021] The adversarial environment simulator relies on the generator and discriminator of a GAN to perform deep learning of feature distributions. The two work together through multiple rounds of iterative training to achieve optimization. The discriminator learns and distinguishes between the real de-identified feature distribution and the feature distribution simulated by the generator, continuously improving the recognition accuracy of the real feature distribution through iteration. The generator, on the other hand, uses the discriminator's recognition results as feedback signals to dynamically adjust its generation logic, gradually approximating the actual operating feature distribution of the charging pile. Through iterative optimization of the loss function, a high degree of fit is achieved between the generator's output feature distribution and the real de-identified feature distribution. The calculation formula is as follows: Discriminator loss function: Generator loss function: ;in This represents the loss value of the discriminator. The smaller the value, the higher the accuracy of the discriminator in recognizing real data and generating features. This represents the loss value of the generator; the smaller the value, the closer the feature distribution generated by the generator is to the true distribution. For the discriminator to analyze the true desensitized feature distribution data The probability of determination, with a value range of [0,1]; For generators based on random noise The generated simulated feature distribution data; For mathematical expectation; The probability distribution represents the true distribution of desensitized features; random noise The probability distribution is determined; ultimately, the generator output feature distribution is highly fitted to the real desensitized feature distribution. When the discriminator completely fails and outputs a fixed value of 0.5 for all samples, the loss function has no optimization gradient, and the training cannot converge. This is a common degradation scenario in GAN training. The training termination condition is constrained by the cosine similarity fit threshold to avoid invalid training scenarios where the discriminator degenerates.

[0022] A quantitative algorithm for determining the goodness of fit between generated feature distributions and true feature distributions in an adversarial environment simulator is presented, and the calculation formula is as follows: ;when When the generated feature distribution is determined to be highly fitted to the real feature distribution, the generator reaches a stable convergence state. The cosine similarity value ranges from -1 to 1. The closer the value is to 1, the higher the similarity between the two sets of feature distributions. The feature vector is the true desensitization feature distribution, with dimension . , The number of core operational features of charging piles; The feature vector of the simulated feature distribution output by the generator; and for and The Feature values ​​in each dimension; and For feature vectors and The modulus length; when and All vectors are non-zero. Invalid all-zero feature vectors must be removed during the preprocessing stage and will not be included in the fit determination process.

[0023] The deep learning process for feature distribution does not rely on massive amounts of real-world operational data. It can reproduce the operating characteristics of charging piles under different regions and conditions simply by learning the desensitized feature distribution, thus adapting to the training needs of scenarios with sparse data from newly commissioned charging piles. Once the generator's output feature distribution reaches a stable convergence state, the adversarial environment simulator, combined with the actual operating scenario of the charging pile microgrid, initiates the virtual data generation process. The generated virtual data is precisely adapted to the training requirements of the federated meta-learning aggregator and can be directly used as training samples. The generated virtual data is in a multi-dimensional fusion format; each set of data includes scenario parameters, photovoltaic output data, charging load data, and energy storage adaptation data. The data volume can be flexibly adjusted according to training needs. Based on scenario type, the data is divided into three categories: The first category is virtual traffic flow pattern data, which covers the number of charging piles connected, charging duration, and charging power distribution data under off-peak, peak, and sudden congestion conditions, and is used to simulate the impact of different traffic scenarios on charging load; The second category is virtual weather change sequence data, which includes photovoltaic output simulation data and corresponding charging pile charging load data under extreme weather conditions such as sudden changes in sunlight, rainstorms, high temperatures, and cold waves, and is used to reproduce the interference of weather factors on the coordinated operation of photovoltaic-energy storage systems and charging piles; The third category is virtual large-scale event operation data, which simulates the sudden increase in charging demand and load peak shift under scenarios such as sports events, large-scale exhibitions, and holiday gatherings, and covers charging load fluctuation data corresponding to different event scales and durations.

[0024] After classifying and labeling the massive amounts of virtual data generated by the adversarial environment simulator, it directly outputs the data to the federated meta-learning aggregator deployed on the same server, achieving seamless integration with the next module. The adversarial environment simulator continuously receives training feedback signals indirectly transmitted by the federated meta-learning aggregator through an upgraded asynchronous federated feedback module. When it detects insufficient model adaptability or poor training performance for a certain scenario during training, the adversarial environment simulator adjusts the generation ratio of virtual data accordingly, supplements virtual samples for the corresponding scenario, and further optimizes the training effect of the federated meta-learning aggregator.

[0025] The Federated Meta-Learning Aggregator is deployed on a central cloud server and works in conjunction with an adversarial environment simulator. Its core training objective is to generate a set of initialization parameters for the model. The output of this initialization parameter set is characterized by its ability to quickly fine-tune with limited local data and can be distributed to charging stations to initialize a lightweight local prediction model. The Federated Meta-Learning Aggregator receives massive amounts of multimodal virtual challenge scenario data from the adversarial environment simulator after classification and labeling, and performs targeted preprocessing for the training objective. First, it splits the virtual data according to the operational dimension of a single charging station, forming a single virtual station sample set. Then, it performs lightweight filtering on each sample set, retaining only a small amount of feature data representing the core operational rules of the charging station, ensuring that the effective data volume of the single virtual station sample set matches the actual local data volume of newly deployed charging stations to simulate data sparsity at the terminal. Finally, it normalizes and standardizes all sample sets, unifying the data format and feature dimensions to ensure the standardization of training input.

[0026] The federated meta-learning aggregator trains models based on a hierarchical training architecture of federated meta-learning. It uses each single virtual peg sample set as the training object, inputs a small amount of selected virtual feature data into the built-in basic model framework, and drives the framework to complete rapid optimization training. At the same time, it extracts and aggregates the optimization process, parameter adjustment logic, and correlation between features and model output in all single virtual peg scenarios to continuously optimize its own parameter generation logic.

[0027] The algorithm for the federated meta-learning aggregator is calculated using the following formula: ; ;when When the denominator of the aggregation formula is 0, it has no mathematical meaning; when With all initial parameters set to zero, the model has no initial feature extraction capability, fine-tuning becomes completely ineffective, and the aggregation formula is only applicable under certain conditions. The number of training samples should be no less than one set; The valid range of values ​​is The value of 0 is prohibited. The standard engineering value is 0.8-1.0 to adapt to the lightweight fine-tuning requirements of the terminal.

[0028] in, These are the global model parameters aggregated by federated meta-learning; This represents the total number of single virtual stake sample sets. No. A single virtual pile sample set was obtained The model parameters were fine-tuned through a small number of iterations. ; Initialize the parameter set for the final generated model; For parameter adaptation coefficients ; Element-wise multiplication enables lightweight scaling of global parameters to adapt to the computing power of terminal hardware. The aggregator sets clear performance verification metrics for each optimization training, including the prediction accuracy after model optimization, the number of optimization iterations, and the magnitude of parameter adjustments. If the optimization result in a certain virtual charging pile scenario does not reach a preset threshold, the aggregator immediately generates training feedback information for that scenario and transmits it to the adversarial environment simulator, driving it to supplement the corresponding virtual samples. Then, a second training is carried out based on the supplemented samples until fast and high-precision model optimization based on a small amount of virtual data is achieved. When the training effect reaches the preset global threshold, the aggregator, based on its learned multi-scenario optimization capabilities, extracts the common parameter logic of model optimization and generates a model initialization parameter set. At the same time, the parameter set is lightweighted and standardized to compress the parameter data volume, ensuring that the parameter set is adapted to the hardware computing and data receiving capabilities of the terminal charging pile.

[0029] The federated meta-learning aggregator distributes the processed model initialization parameter set to all smart charging piles at the terminal layer through an encrypted channel. This parameter set serves as the basic initialization parameter for a lightweight local prediction model, completing the cloud initialization of the model. At the same time, the parameter set distribution record and basic adaptation information are synchronized to the edge layer server for backup, providing data support for subsequent parameter set optimization and re-distribution. In addition, the basic version of the parameter set is retained to prepare for iterative optimization of the parameter set based on terminal feedback information.

[0030] The lightweight local prediction model is deployed within each smart charging pile at the terminal layer. Designed with minimal computational overhead and high speed, it is fully compatible with the hardware resources of ordinary charging piles. Its core function is to initialize itself using model initialization parameter sets distributed from the cloud, and then rapidly fine-tune it within the terminal using a small amount of new local operating data from the charging pile. This generates a highly customized prediction model adapted to local conditions, providing data support for charging pile load forecasting, photovoltaic output adaptation, and energy storage charging and discharging scheduling. Furthermore, the entire fine-tuning process ensures that data remains locally available, guaranteeing data privacy and security from the terminal side.

[0031] The lightweight local prediction model has a built-in parameter receiving and parsing unit. It receives the model initialization parameter set from the federated meta-learning aggregator on the cloud central server through an encrypted channel. During the receiving process, the parameter set is first checked for integrity and decrypted. If parameter set transmission loss or format errors are detected, a retransmission request is immediately initiated to the cloud. After successful verification, the parameter set is directly loaded into its own lightweight basic framework, completing the initial model configuration. The initialized model has basic prediction capabilities and can adapt to the general operating conditions of charging piles. At the same time, the model retains the basic data of the initialization parameter set as a benchmark reference for subsequent local fine-tuning.

[0032] The lightweight local prediction model initiates local data acquisition, connecting to the charging pile's own sensing and monitoring modules and the photovoltaic energy storage monitoring unit. It collects a small amount of new operational data from the past 3 to 7 days after the charging pile's commissioning. Core data types include: charging load data such as real-time charging power, charging duration, and number of connected vehicles; photovoltaic operation data such as real-time output and light intensity of the photovoltaic modules; energy storage data such as the charging and discharging status and remaining capacity of the energy storage battery; and local environmental data such as real-time temperature and precipitation in the charging pile's area. After acquisition, the data undergoes lightweight preprocessing within the terminal, removing outliers and missing values, retaining only valid data representing the core patterns of local operating conditions. Simultaneously, the local data is standardized and matched according to the feature dimensions of the cloud parameter set to ensure consistency between the local data and the model input dimensions, providing highly adaptable samples for subsequent rapid fine-tuning. All acquisition and preprocessing actions are completed within the terminal, with no data transmitted externally. The model then performs rapid fine-tuning calculations within the terminal based on the preprocessed, limited amount of valid local data. This process is specifically optimized for lightweight model frameworks. Model adaptation is completed through only a few iterations and local parameter adjustments with low computational cost. It is fully compatible with the embedded CPU or lightweight computing power module of ordinary charging piles, without the need for additional hardware.

[0033] The lightweight local prediction model terminal fine-tuning algorithm enables rapid parameter updates based on limited local data. The calculation formula is as follows: ; For the model through the first Local parameters after fine-tuning; For the model through the first Local parameters after fine-tuning; The learning rate; For loss function For parameters The gradient; This is a small amount of operational characteristic data of the charging pile after local preprocessing. Predictive labels for local operating data of charging piles.

[0034] when When the parameters are not updated, fine-tuning becomes completely ineffective; when When the value is too large, the parameter update range is too large, the model diverges, and it cannot converge; when combine Since the set is empty, the loss function and gradient cannot be calculated, and the formula is meaningless.

[0035] Fine-tuning is based on the initial parameter set, making only minor adjustments to core parameters such as feature weights and prediction thresholds that are strongly correlated with local operating conditions, without changing the overall model framework. All calculations and parameter adjustments are completed locally on the charging pile terminal. The original local operating data and parameter change data during the fine-tuning process are all stored in the terminal's local cache, ensuring that data does not leave the charging pile throughout the entire process.

[0036] After local fine-tuning is completed, the model immediately starts a self-verification mechanism, using the latest local operating data as a verification sample to verify the prediction accuracy of the fine-tuned personalized prediction model: if the prediction error is within the preset threshold, the model is determined to be successfully adapted; if the threshold is not reached, 1 to 3 supplementary step-by-step fine-tunings are performed based on local data until the prediction requirements of local working conditions are met.

[0037] The personalized prediction model, after passing verification, has been officially put into terminal application, providing core data support for the management of charging pile photovoltaic energy storage microgrids. Specifically, it achieves the following: accurately predicting the charging load demand of charging piles in future periods, providing an adaptation basis for photovoltaic module output scheduling, and achieving maximum matching between photovoltaic output and charging load; and optimizing the charging and discharging strategy of energy storage batteries based on load forecasts and photovoltaic output data, controlling energy storage charging when photovoltaic output is sufficient and charging load is low, and controlling energy storage discharging when photovoltaic output is insufficient and charging load is high, thereby achieving efficient utilization of energy within the microgrid.

[0038] After the personalized prediction model is put into operation, it continuously receives the latest local operating data and dynamically makes minor adjustments to the model to ensure that it always adapts to the dynamic changes in local operating conditions. Simultaneously, it retains real-time performance metrics data after model fine-tuning, including core performance gain indicators such as prediction accuracy, number of fine-tuning iterations, and parameter adjustment magnitude. Furthermore, it abstracts and characterizes the local operating data to extract highly abstract feature distribution parameters that represent local operating conditions. These performance gain indicators and abstract feature distribution parameters serve as core feedback data from the terminal, providing data support for subsequent uploading to the cloud via the upgraded asynchronous federated feedback module.

[0039] The upgraded asynchronous federated feedback module is a two-way data interaction and optimization module set up across the cloud layer and the terminal layer. Its core function is to build an encrypted feedback channel between the terminal charging pile and the central server in the cloud, realize the secure uploading of terminal feedback data and the effective distribution of the model initialization parameter group after cloud optimization. Through terminal data feedback, it drives the synchronous optimization of the cloud adversarial environment simulator and the federated meta-learning aggregator, realizing the self-evolution of the entire system.

[0040] Once the lightweight local prediction model of the terminal charging pile completes local fine-tuning and passes accuracy verification, the upgraded asynchronous federated feedback module initiates the feedback data extraction process within the charging pile terminal: First, it accurately extracts the performance gain indicators after model fine-tuning, including the model loss function decline rate, prediction accuracy improvement rate, fine-tuning iteration steps, parameter adjustment range, and prediction error fluctuation range; Second, it extracts highly abstract feature distribution parameters from the pre-processed operating data of the terminal, retaining only feature information that characterizes the core laws of local operating conditions, such as the peak and valley distribution characteristics of local charging load, and the matching characteristics of photovoltaic output and charging load, while removing all original operating data and privacy-related information, and ensuring that the extraction dimension is consistent with the feature learning input dimension of the adversarial environment simulator.

[0041] The upgraded asynchronous federated feedback module standardizes and removes redundancy from the extracted performance gain metrics and abstract feature distribution parameters, forming a unified format terminal feedback dataset to prepare for subsequent encryption and transmission. The upgraded asynchronous federated feedback module then employs a high-strength encryption algorithm for layered encryption of the preprocessed terminal feedback dataset: the performance gain metrics and highly abstract feature distribution parameters are independently encrypted using the AES-256 encryption algorithm, generating unique encrypted ciphertext; simultaneously, a CRC32 checksum and a unique terminal identifier are added to each encrypted ciphertext, and the integrity of the encrypted ciphertext is verified locally. If data loss or format errors are detected, the data is immediately re-extracted and re-encrypted. After encryption, data transmission is carried out using a dedicated end-to-cloud encryption communication channel built by the module. This channel uses national cryptographic algorithms for channel encryption and is coupled with anti-interference differential lines to ensure transmission stability; at the same time, the module adopts an asynchronous transmission mechanism, dynamically allocating transmission time slots based on the model fine-tuning completion time of each terminal charging pile and network status, avoiding channel congestion and data packet loss caused by multiple terminals uploading simultaneously. The edge layer server temporarily caches and relays data during transmission; if a data transmission interruption is detected, the module immediately triggers a retransmission mechanism until the data is successfully uploaded to the central cloud server.

[0042] The upgraded asynchronous federated feedback module allows for flexible settings of data upload frequency based on system operational needs. It supports both immediate reporting after terminal model fine-tuning and batch reporting at preset intervals, balancing real-time data transmission with efficiency. Upon receiving encrypted data from each terminal, the upgraded asynchronous federated feedback module on the cloud central server first decrypts the data using a dedicated decryption algorithm. Then, it performs a secondary integrity and validity check on the decrypted feedback data using a CRC32 checksum, eliminating invalid and abnormal data caused by transmission or encryption anomalies. After verification, the upgraded asynchronous federated feedback module categorizes and extracts valid feedback data according to data type and regional aggregation: highly abstract feature distribution parameters are categorized and integrated according to the charging pile's location and application scenario to form an updated multi-region feature distribution dataset that matches the feature learning input format of the adversarial environment simulator; performance gain indicators are categorized and statistically analyzed according to charging pile type, fine-tuning scenario, and model adaptation effect to form a standardized model performance analysis dataset, providing quantitative basis for parameter optimization of the federated meta-learning aggregator. Meanwhile, the upgraded asynchronous federated feedback module binds and retains all valid feedback data with the terminal's unique identifier, forming an edge-cloud feedback data log, which provides data support for tracking the effects of subsequent system optimizations.

[0043] The upgraded asynchronous federated feedback module delivers the two types of datasets, after classification and aggregation, to the adversarial environment simulator and the federated meta-learning aggregator, respectively, driving the two modules to perform precise synchronous optimization: The updated multi-region feature distribution dataset is delivered to the adversarial environment simulator, which integrates this data into the original feature distribution dataset, re-performs feature distribution learning, adjusts the generator's generation logic and the virtual data generation ratio, and supplements corresponding virtual samples for scarce scenarios or new operating conditions reported by the terminal, making the subsequently generated virtual challenge scenario data more closely match the actual operating conditions of the terminal; The model performance analysis dataset is delivered to the federated meta-learning aggregator, which analyzes the adaptation shortcomings of the existing model's initialization parameter set based on performance gain indicators under different scenarios and charging pile types, adjusts its own training weights and parameter generation logic, and performs targeted optimization of the initialization parameter set, improving the parameter set's cross-scenario and cross-terminal adaptability.

[0044] The two cloud-based modules conduct a new round of virtual sample generation and model training based on the optimized logic and parameters. The generated better model initialization parameter set is redistributed to all terminal charging piles through the encrypted channel of the upgraded asynchronous federated feedback module. The terminal charging piles can reinitialize or fine-tune the lightweight local prediction model based on the new initialization parameter set to achieve iterative upgrade of the terminal model.

[0045] Working principle: When the system is running, the adversarial environment simulator deployed on the central cloud server first receives the de-identified data feature distribution parameters uploaded by the upgraded asynchronous federated feedback module of each terminal charging pile. Through iterative collaborative learning between the GAN generator and discriminator, the real operating characteristics and rules of the charging pile are reproduced, thereby generating massive multimodal virtual challenge scenario data that conforms to real statistical rules. This provides diverse training samples for the federated meta-learning aggregator. No original operating data is acquired throughout the process, thus avoiding privacy leakage from the source.

[0046] Subsequently, the cloud-based federated meta-learning aggregator uses this virtual data as a training sample, splits and lightweight filters the data according to the single virtual pile dimension, simulates the sparse working condition of new terminal pile data to carry out federated meta-learning training, and learns the core ability to quickly optimize the model based on a small amount of data. Finally, it generates a generalized model initialization parameter set with cross-scenario and cross-terminal adaptability, which is encrypted and uniformly distributed to all terminal charging piles through a dedicated channel to complete the cloud initialization of the lightweight local prediction model in each pile.

[0047] After the lightweight local prediction model of each smart charging pile at the terminal layer is initialized, it connects to the local sensing and photovoltaic energy storage monitoring unit to collect a small amount of new operating data from the past 3-7 days. Data preprocessing and lightweight fine-tuning of the model are completed within the pile. Only a small number of iterations are used to locally correct the core parameters of the model. The entire process involves no data leaving the local area and no external data interaction. Finally, a personalized prediction model that is highly adapted to local operating conditions is generated. This model directly provides core data support for the charging load prediction of charging piles, photovoltaic output adaptation, and intelligent scheduling of energy storage charging and discharging, realizing the fine allocation of energy within the microgrid.

[0048] Next, the upgraded asynchronous federated feedback module acts as the bidirectional interaction hub between the edge and the cloud. It extracts the performance gain index after fine-tuning the personalized prediction model and the highly abstract feature distribution parameters representing the local operating conditions from the terminal charging pile. After encrypting the two types of data, they are uploaded to the cloud asynchronously through the dedicated encrypted channel between the edge and the cloud. After receiving and decrypting the verification data, the central server in the cloud sends the abstract feature distribution parameters to the adversarial environment simulator, driving it to update the feature learning logic and adjust the virtual data generation ratio so that the subsequently generated virtual data is more in line with the actual operating conditions of the terminal. At the same time, the performance gain index is sent to the federated meta-learning aggregator, driving it to optimize the training weights and parameter generation logic, and improve the terminal adaptability of the model initialization parameter set.

[0049] After the cloud-based simultaneous optimization of the two core modules, the adversarial environment simulator and the federated meta-learning aggregator immediately begin a new round of virtual data generation and model training. This generates a better set of model initialization parameters and encrypts and distributes them to each terminal. Based on the new parameter set, the terminal charging piles re-initialize or fine-tune the lightweight local prediction model. This effectively solves industry pain points such as data sparsity, poor model adaptability, and data privacy leakage in newly deployed charging piles. It ensures that the terminal charging piles always have high-precision local operating condition prediction capabilities, continuously optimizes the scheduling and matching accuracy of photovoltaic energy storage and charging load, and ultimately realizes intelligent and refined energy management of multi-charging pile microgrids that integrate photovoltaic and energy storage.

[0050] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage, characterized in that: The system comprises a central cloud server, edge layer regional servers, and multiple smart charging piles at the terminal layer. The central cloud server is equipped with an adversarial environment simulator and a federated meta-learning aggregator. The adversarial environment simulator, built upon a generative adversarial network, learns the distribution of de-identified features uploaded by each smart charging pile and generates multimodal virtual scene data conforming to real statistical laws. The federated meta-learning aggregator, connected to the adversarial environment simulator, receives the virtual scene data for training and generates a set of model initialization parameters that can be quickly fine-tuned based on a small amount of data. Each smart charging pile is equipped with a lightweight local prediction module. The system is configured to receive the model initialization parameter set to complete model initialization, and to perform minor fine-tuning within the charging pile using locally collected operational data to generate a personalized prediction model adapted to local operating conditions. The data processing of the fine-tuning process is completed entirely locally. The system also includes an asynchronous federated feedback module across the cloud and terminal layers, which is used to encrypt and upload the performance gain indicators and abstract feature distribution parameters of each smart charging pile after local fine-tuning to the central server in the cloud layer. This drives the adversarial environment simulator to update the feature distribution, the federated meta-learning aggregator to optimize the parameter set generation logic, and redistributes the optimized model initialization parameter set to the smart charging pile.

2. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage as described in claim 1, characterized in that: The adversarial environment simulator performs desensitization processing on the data received from the terminal charging pile. The desensitization processing includes data normalization, feature abstraction and extraction, and privacy masking to obtain feature distribution parameters that characterize the operating rules of the charging pile.

3. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage as described in claim 2, characterized in that: The adversarial environment simulator includes a generator and a discriminator built based on a generative adversarial network; the discriminator is configured to distinguish between the real desensitized feature distribution and the feature distribution generated by the generator; the generator is configured to adjust its generation logic according to the discrimination result of the discriminator until the feature distribution output by the generator meets the preset standard of statistical distribution consistency.

4. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 3, characterized in that: The adversarial environment simulator includes a generator and a discriminator built on a generative adversarial network; the discriminator is configured to distinguish between the real desensitized feature distribution and the simulated feature distribution generated by the generator; the generator is configured to dynamically adjust its generation logic according to the discrimination result of the discriminator until the simulated feature distribution output by the generator is highly fitted to the real desensitized feature distribution, reaching a preset standard of statistical distribution consistency.

5. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 1, characterized in that: The federated meta-learning aggregator preprocesses the received virtual scene data, splitting the data into multiple sample sets based on a single virtual charging pile, and then performing lightweight filtering and standardization on each sample set.

6. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 5, characterized in that: If the training result of the federated meta-learning aggregator based on the sample set does not reach the threshold during training, it sends a feedback signal to the adversarial environment simulator to trigger the generation of supplementary virtual scene data of the corresponding type.

7. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 1, characterized in that: The lightweight local prediction model receives the model initialization parameter set as a basis and only performs a limited number of iterative adjustments on the parameters related to the local operating conditions within the model to complete the lightweight fine-tuning.

8. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 1, characterized in that: The local operating data on which the lightweight local prediction model is fine-tuned includes charging load data, photovoltaic output data, energy storage status data, and local environmental data.

9. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 1, characterized in that: The asynchronous federated feedback module encrypts the extracted performance gain indicators and highly abstract feature distribution parameters respectively, and transmits them asynchronously through an encrypted communication channel; the edge layer regional server is configured to provide data caching and relay support during data transmission, and triggers a retransmission mechanism when a transmission interruption is detected.

10. The smart energy management system for multi-charging pile microgrids integrating photovoltaics and energy storage according to claim 9, characterized in that: The asynchronous federated feedback module classifies and summarizes the decrypted feedback data on the central server side of the cloud, forming a feature distribution dataset for updating the adversarial environment simulator and a model performance analysis dataset for optimizing the federated meta-learning aggregator.

Images