Battery car charging comprehensive management system based on edge computing

Abstract

The present application relates to the technical field of battery car charging management, in particular to a battery car charging comprehensive management system based on edge computing, which collects battery parameters of battery cars and state parameters of charging piles in real time through a battery car end collection module and a charging pile end collection module, and performs local processing through an edge computing unit, thereby improving the real-time performance of data processing and enabling rapid response to abnormal conditions. Through deep learning analysis of charging behavior, abnormal conditions are judged, and data is distributed to a load balancing unit or a risk assessment unit. The load balancing unit optimizes charging load distribution according to power grid load, battery car demand and charging pile state by using a reinforcement learning model, thereby avoiding power grid overload. The risk assessment unit establishes a model through historical data, monitors risks in real time and issues warnings, and at the same time, through edge computing, a classification model is deployed locally, thereby reducing dependence on the cloud, improving computing efficiency and management efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery car charging management, and particularly relates to a battery car charging comprehensive management system based on edge computing. BACKGROUND

[0002] Battery cars have become one of the important tools for short-distance travel in cities. As a high-efficiency vehicle parking solution, multi-level garages have been widely used. Multi-level garages store vehicles through multi-layer mechanical structures, effectively save ground space, and greatly alleviate the parking pressure in cities. However, with the increase in the number of battery car users, how to reasonably, safely and efficiently charge battery cars in multi-level garages has become a problem to be solved.

[0003] Traditional battery car charging methods are mostly concentrated in flat parking lots or independent charging stations. The charging pile layout in these scenarios is relatively simple, and the charging load management and safety problems are also relatively single. However, in the multi-level garage scenario, battery car charging involves multi-layer space charging safety management. Due to the high density of vehicles in multi-level garages, the power load of battery car charging piles fluctuates greatly, and the impact on the power grid load is more complex when battery car centralized charging behavior occurs. If the power load cannot be reasonably allocated or abnormal conditions cannot be found in time, it is easy to cause power overload, equipment damage and even fire safety hazards. SUMMARY

[0004] To solve the above problems, the present application provides a battery car charging comprehensive management system based on edge computing, which realizes intelligent management of large-scale battery car charging by combining edge computing and machine learning models and reduces the server computing power cost.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is:

[0006] A battery car charging comprehensive management system based on edge computing, comprising a battery car end acquisition module, a charging pile end acquisition module and an intelligent management module, the battery car end acquisition module and the charging pile end acquisition module are connected with the intelligent management module respectively;

[0007] The battery car end acquisition module is used for acquiring battery parameters in the battery car and sending them to an edge computing unit, including battery model, capacity, temperature, battery voltage and charging rate;

[0008] The charging pile end acquisition module is used for acquiring the state parameters of the charging pile load and each port of the charging pile that can be supported by the power grid, and sending them to the edge computing unit, the state parameters of the port including output voltage, output current and load power;

[0009] The intelligent management module comprises an edge computing unit, a load balancing unit and a risk assessment unit, the edge computing unit being in communication connection with the load balancing unit and the risk assessment unit respectively;

[0010] The edge computing unit is configured to determine whether a current charging behavior is abnormal and determine an abnormal type according to a power grid supportable charging pile load, battery parameters and state parameters of each port through deep learning, and distribute and send the power grid supportable charging pile load, the battery parameters and the state parameters of each port to the load balancing unit or the risk assessment unit according to the abnormal type;

[0011] The load balancing unit is configured to perform load adjustment according to a reinforcement learning model based on the power grid supportable charging pile load, the battery parameters and the state parameters of each port;

[0012] The risk assessment unit is configured to construct a risk assessment model according to the abnormal type, the battery parameters and the state parameters of each port in history, and output a risk level and perform an alarm prompt through the risk assessment model according to real-time abnormal type, battery parameters and state parameters of each port.

[0013] Further, the edge computing unit comprises a battery control subunit and a charging pile control subunit, the battery control subunit and the charging pile control subunit being in communication connection;

[0014] The battery control subunit is configured to receive battery parameters for standardization and send the battery parameters to the charging pile control subunit;

[0015] The charging pile control subunit is configured to perform the following steps:

[0016] The power grid supportable charging pile load, the battery parameters and the state parameters of the port are taken as inputs of a feedforward neural network, and an abnormal type is output, the abnormal type comprising a load abnormality and a charging abnormality;

[0017] The battery parameters and the state parameters of each port are sent to the load balancing unit or the risk assessment unit according to the abnormal type.

[0018] Further, the training steps of the feedforward neural network comprise:

[0019] The power grid supportable charging pile load, the battery parameters and the state parameters of the port are normalized, and an abnormal type is labeled;

[0020] Each parameter in the normalized power grid supportable charging pile load, the battery parameters and the state parameters of the port is split to construct an input neuron;

[0021] The features of the input neurons are calculated and extracted layer by layer through the hidden layers in the feedforward neural network, and the high-order features of the abnormal patterns are extracted and output through the output layer by using the activation function for nonlinear feature mapping;

[0022] The output result of the output layer is used for loss function calculation based on the labeled abnormal type, and the network weight is adjusted by using the back propagation algorithm.

[0023] Further, the sending of the battery parameters and the state parameters of each port to the load balancing unit or the risk assessment unit according to the abnormal type comprises:

[0024] If the abnormal type is a load abnormality, the grid supportable charging pile load, the battery parameters and the state parameters of each port are sent to the load balancing unit;

[0025] If the abnormal type is a charging abnormality, the battery parameters and the state parameters of each port are sent to the risk assessment unit.

[0026] Further, the load balancing unit is used to perform the following steps:

[0027] A preset ratio of the grid supportable charging pile load to the charging pile disposable load is set, and the real-time charging pile disposable load is generated according to the real-time grid supportable charging pile load;

[0028] The current total load of the charging pile is calculated according to the state parameters of the ports, and the reinforcement learning model is constructed by taking the load adjustment of the ports as the action and taking the current total load of the charging pile, the real-time charging pile disposable load and the battery parameters of each port as the state.

[0029] Further, the formula of the reinforcement learning model is as follows:

[0030]

[0031] Wherein, s is the state, including the battery parameters and the state parameters of the charging pile ports; a is the action, including the load adjustment parameters of all ports; Q(s, a) is the expected return that can be obtained after taking action a in state s; r is the immediate reward value after executing the adjustment action; γ is the discount factor; is the executed action; s' is the next state after executing the action a; a' is the optimal adjustment action; max a′ Q(s', a') is the maximum expected return that can be obtained by selecting the optimal adjustment action a' in the next state s'.

[0032] Further, the formula of the immediate reward value r is as follows:

[0033] r=-α·|L current -L optimal |-β·Eloss ;

[0034] wherein, L current is the current load of the charging pile; L optimal is the real-time available load of the charging pile; E loss is the energy loss due to load imbalance or power loss; and alpha and beta are the weight coefficients of load balance degree and energy loss, respectively.

[0035] Further, the construction of the risk assessment model according to the abnormal types in the history, the battery parameters and the state parameters of each port comprises the following steps:

[0036] Based on the abnormal data in the history, a historical data set containing abnormal types, battery parameters and state parameters of each port is collected, and a corresponding risk level is labeled for each record in the historical data set;

[0037] The historical data set with completed risk level labeling is divided into a training set and a test set;

[0038] Based on the decision tree algorithm, the parameters in the training set are recursively binary segmented, the risk level is output through the leaf node, and the risk assessment model is obtained;

[0039] The risk assessment model is verified through the test set.

[0040] Further, the risk assessment unit and the load balancing unit are both deployed in the cloud server.

[0041] The beneficial effects of the present application are that the present application collects the battery parameters of the electric vehicle and the state parameters of the charging pile in real time through the electric vehicle end collection module and the charging pile end collection module. These data are first processed locally through the edge computing unit, thereby reducing the cloud computing power pressure and being able to quickly respond to abnormal conditions in the charging process. The electric vehicle charging behavior is intelligently analyzed through the deep learning algorithm to determine whether an abnormality occurs in the charging process, and the data is distributed to the load balancing unit or the risk assessment unit according to the abnormal type. Through the load balancing unit, the available load of the power grid, the charging demand of the electric vehicle and the state of the charging pile are used to optimize the charging load distribution using a reinforcement learning model to avoid power grid overload during peak hours. At the same time, the risk assessment unit establishes a risk assessment model through the training of historical abnormal data, can monitor potential risks in the charging state in real time, and issue an alarm prompt according to the risk level to prevent accidents in advance. At the same time, the system uses edge computing to deploy the classification model locally for data classification processing, reducing the dependence on cloud server computing power and effectively improving the computing management efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0042] Fig. 1 is a structural schematic diagram of an electric vehicle charging comprehensive management system based on edge computing in the present application.

[0043] Fig. 2 is the training step flow chart of the feedforward neural network in the present application. DETAILED DESCRIPTION

[0044] Referring to Figs. 1-2 The application relates to an edge-computing-based comprehensive management system for battery car charging, which comprises a battery car end acquisition module, a charging pile end acquisition module and an intelligent management module.

[0045] The battery car end acquisition module is used for acquiring battery parameters in the battery car and sending the battery parameters to an edge computing unit, and the battery parameters include a battery model, a capacity, a temperature, a battery voltage and a charging rate.

[0046] The charging pile end acquisition module is used for acquiring state parameters of a charging pile load supported by a power grid and each port of the charging pile, and the state parameters of the port include an output voltage, an output current and a load power, and the state parameters are sent to the edge computing unit.

[0047] The intelligent management module comprises an edge computing unit, a load balancing unit and a risk assessment unit, and the edge computing unit is in communication connection with the load balancing unit and the risk assessment unit.

[0048] The edge computing unit is used for determining whether a current charging behavior is abnormal and determining an abnormal type according to the power grid supported charging pile load, the battery parameters and the state parameters of each port through deep learning, and the power grid supported charging pile load, the battery parameters and the state parameters of each port are distributed and sent to the load balancing unit or the risk assessment unit according to the abnormal type.

[0049] The load balancing unit is used for performing load adjustment according to a reinforcement learning model based on the power grid supported charging pile load, the battery parameters and the state parameters of each port.

[0050] The risk assessment unit is used for constructing a risk assessment model according to the abnormal type, the battery parameters and the state parameters of each port in history, and outputting a risk level and giving an alarm prompt through the risk assessment model according to real-time abnormal type, battery parameters and state parameters of each port.

[0051] It should be noted that the electric vehicle end collection module as the front-end data collection device of the system mainly monitors the key parameters of the battery in real time through a series of sensors. The collected parameters include the model, capacity, temperature, battery voltage and charging rate of the battery. Each parameter has a dedicated sensor responsible for collection, which can include a voltage sensor (such as a Hall effect sensor) for monitoring voltage changes to ensure that voltage fluctuations are within a safe range; a temperature sensor (such as an NTC thermistor) responsible for collecting battery temperature to avoid overheating problems caused by excessive temperature; a current sensor for detecting changes in charging rate to evaluate charging efficiency and safety. All collected data is processed and filtered by an embedded collection controller (such as an STM32 microcontroller) to eliminate noise data, and transmitted to the edge computing unit through a wireless communication module (such as ZigBee, LoRa, BLE, etc. Low-power communication protocols). The charging pile end collection module is used to collect the state information of the charging pile and the power grid in real time to ensure that the charging process is within the load range that the power grid can withstand. Each charging pile port is equipped with current, voltage sensors to collect the output voltage, output current and load power of each port in real time. The voltage sensor can accurately detect the output voltage to ensure its stability; the current sensor monitors the transient changes of the output current to detect whether short circuit or overload occurs; the power sensor is used to calculate the total power consumption of the charging pile to evaluate the load capacity of the power grid. The data of these sensors is processed in real time by the embedded processing unit (such as FPGA or DSP) built-in the charging pile, and the results are transmitted to the edge computing unit of the system as the basis for charging strategy adjustment. The intelligent management module is the core control unit of the entire system, which includes the edge computing unit, the load balancing unit and the risk assessment unit. The design of this module is based on an embedded computing platform, such as the NVIDIA Jetson series or the Intel Movidius module, which can process a large amount of data locally in real time, reduce dependence on remote cloud computing, and improve response speed and data privacy. In the intelligent management module, the edge computing unit is responsible for receiving and integrating data from the electric vehicle end and the charging pile end. Through deep learning algorithms (such as convolutional neural networks), this unit analyzes the state data of the battery and the load data of the charging pile in real time to detect abnormal behavior during charging. The deep learning model is trained in advance on a cloud server database containing a large amount of battery and charging data, which can accurately identify different types of abnormalities (such as battery temperature abnormalities, high charging rate, voltage fluctuations, etc.), and then deploy the trained model locally on the edge computing device. When abnormal behavior is detected, the edge computing unit labels and classifies the abnormal information, and sends the data to the load balancing unit or the risk assessment unit for further processing according to the specific type of the anomaly. The load balancing unit realizes intelligent allocation of charging power based on a reinforcement learning model to optimize the load balancing between the power grid and the charging pile.The reinforcement learning model (such as deep Q-learning algorithm DQN or policy gradient algorithm) learns how to optimally allocate charging power under different charging conditions through training on historical charging data. The core task of this unit is to dynamically adjust the charging output of each port according to real-time battery parameters and the state of the charging pile. For example, when detecting that the load power of a certain port is too high, the system will automatically reduce the output power of that port and redistribute the excess power to ports with lower loads, thereby avoiding local overload of the charging pile. The risk assessment unit is deployed through the cloud server, triggers the service support of the edge computing unit through cloud instructions, and uses machine learning algorithms to make real-time predictions and assessments of potential risks during the charging process. This unit builds a risk assessment model based on abnormal types, battery parameters, and charging pile state parameters in historical data. The risk assessment model is based on random forest or gradient boosting decision tree (GBDT) algorithms, which can extract abnormal patterns from a large amount of historical data and perform classification and prediction. During real-time monitoring, when the system detects new abnormal signals, the risk assessment unit inputs the current data into the model, outputs the corresponding risk level, and triggers the alarm mechanism. For example, when the system determines that an abnormal behavior may cause the battery to overheat or the charging pile to fail, it will immediately issue a high-risk warning and display detailed risk information on the charging pile LED screen, while notifying relevant maintenance personnel to take emergency measures through a buzzer or remote communication module (such as 4G / 5G network).

[0052] Further, the edge computing unit comprises a battery control subunit and a charging pile control subunit, and the battery control subunit and the charging pile control subunit are in communication connection;

[0053] The battery control subunit is used to receive battery parameters for standardization and send them to the charging pile control subunit;

[0054] The charging pile control subunit is used to perform the following steps:

[0055] The state parameters of the grid, the battery, and the ports are used as the input of the feedforward neural network, and the abnormal type is output, including load abnormality and charging abnormality;

[0056] According to the abnormal type, the battery parameters and the state parameters of each port are sent to the load balancing unit or the risk assessment unit.

[0057] In some embodiments, the battery control subunit is mainly responsible for receiving and preprocessing battery parameters. Battery parameters include battery model, capacity, temperature, battery voltage, and charging rate, etc. These parameters are collected in real time through sensors embedded in the battery car, and the data is transmitted to the battery control subunit through a low-latency communication interface such as CAN bus or RS485 interface. The battery control subunit processes these collected battery parameters through a built-in microcontroller such as ARM Cortex-M series. The purpose of standardization is to convert various parameters into a unified scale so that subsequent algorithms can be accurately analyzed. For example, the battery temperature may be collected in Celsius, while the battery voltage is in volts. Through standardization, these data are mapped to the same numerical range, reducing the influence between different dimensions and improving data consistency. After standardization, the battery parameters are sent to the charging pile control subunit through a real-time communication protocol such as MQTT or HTTP RESTful interface for further anomaly detection and analysis. After receiving the standardized battery parameters from the battery control subunit, the charging pile control subunit combines the information of the charging load supported by the power grid and the state parameters of the charging pile port to execute intelligent algorithms for anomaly detection. This process uses a feedforward neural network (FNN) for real-time analysis. Feedforward neural network is a basic deep learning model suitable for processing nonlinear relationships between input and output. After detecting and classifying the type of anomaly, the charging pile control subunit sends relevant parameter data to specific subsequent modules for processing according to the different types of anomalies. If the system detects a "load anomaly", the control subunit sends battery parameters and port state parameters to the load balancing unit. The load balancing unit adjusts the power output of the charging pile through a reinforcement learning algorithm to avoid power grid overload and ensure charging safety. If a "charging anomaly" is detected, the relevant data will be transmitted to the risk assessment unit. The risk assessment unit constructs a risk assessment model based on abnormal parameters to assess the risk level in real time during the battery charging process and issue an alarm prompt if necessary.

[0058] Further, the training step of the feedforward neural network comprises:

[0059] normalizing the power grid supportable charging pile load, battery parameters, and port state parameters, and labeling the anomaly type;

[0060] splitting and constructing input neurons for each parameter in the normalized power grid supportable charging pile load, battery parameters, and port state parameters;

[0061] The features of the input neurons are calculated and extracted layer by layer through the hidden layers in the feedforward neural network, and the activation function is used for nonlinear feature mapping, high-order features of the abnormal patterns are extracted, and the results are output through the output layer integration;

[0062] The output results of the output layer are used for loss function calculation based on the labeled abnormal types, and the network weights are adjusted using the back propagation algorithm.

[0063] In some embodiments, data preprocessing is performed first, in which the grid normalizes the load of the charging pile, various parameters of the battery (such as temperature, voltage, capacity), and real-time state parameters of the charging pile port (such as output current, voltage, etc.). The goal of normalization is to scale the data to a unified range (usually between 0 and 1), thereby eliminating the dimensional differences between different data and ensuring that subsequent algorithms are not affected by the size of the feature values. For example, battery voltage and battery temperature may differ greatly in data range, but after normalization, they are converted to the same scale, facilitating unified analysis by the model. At this time, the normalized data is also labeled with the type of anomaly as a reference target for training the model. After data preparation is complete, the step of constructing input neurons begins. The input layer of the neural network is composed of multiple input neurons, each corresponding to a normalized data feature. The input layer is responsible for receiving all standardized parameter data, such as battery temperature, battery voltage, charging load supported by the grid, and state information of the charging pile port. Through this structural design, the input layer can provide comprehensive operational state data, covering key factors that affect the charging process. After the input data enters the hidden layer, the feature extraction and nonlinear mapping process begins. The hidden layer is composed of multiple neurons, each connected to the neurons of the previous layer through weights. Each neuron in the hidden layer performs weighted summation on the input features and passes them through an activation function to capture complex nonlinear relationships in the data. For example, when detecting a sudden increase in battery temperature, the hidden layer neurons will respond to this abnormal change and reflect it in feature extraction. The choice of activation function, such as ReLU (Rectified Linear Unit), enhances the model's ability to recognize complex features, ensuring the system's sensitivity to changes in battery and charging pile state. Under the step-by-step mapping of multiple hidden layers, the model gradually extracts high-order anomaly pattern features that can reflect hidden abnormal behaviors in the input data. After feature extraction is complete, the data enters the output layer for final integration. The output layer, based on the high-order features extracted by the hidden layer, determines whether the data is abnormal and classifies it as "load anomaly" or "charging anomaly." This classification is based on the probability results of the output layer neurons, and the model finally outputs the likelihood of each type of anomaly. For example, if the system detects that the grid load abnormally increases during a charging process and the battery temperature rapidly increases, the output layer will tend to label this situation as "charging anomaly." In this way, the system can effectively classify different types of anomalies. To improve the prediction accuracy of the network, the model also performs loss calculation and weight adjustment. The loss function is used to measure the difference between the predicted results and the actual labeled values. In this process, the cross-entropy loss function is often used, which can quantify the error in classification. For example, if the network incorrectly classifies "load anomaly" as "charging anomaly," the loss function will give a specific loss value based on the degree of error.The system adjusts the weights and biases of the neural network according to the loss value through the backpropagation algorithm. The process of backpropagation involves gradient calculation for each weight, and the weights are updated according to the gradient information, so that the model is closer to the correct result in the next iteration.

[0064] Further, the sending of the battery parameters and the state parameters of each port to the load balancing unit or the risk assessment unit according to the abnormal type comprises:

[0065] If the abnormal type is a load abnormality, the grid supportable charging pile load, the battery parameters and the state parameters of each port are sent to the load balancing unit;

[0066] If the abnormal type is a charging abnormality, the battery parameters and the state parameters of each port are sent to the risk assessment unit.

[0067] In some embodiments, load abnormalities generally involve load issues of the power grid or charging piles, such as power grid load exceeding limits, excessive current, or voltage fluctuation abnormalities, etc.; charging abnormalities are related to the state of the battery or the supply current voltage, such as excessive battery temperature, voltage abnormalities, or excessive charging rate, etc. After determining the type of abnormality, the system takes corresponding measures according to different types of abnormalities, and transmits data to the appropriate module for further processing. When the type of abnormality is a load abnormality, the system will focus on parameters related to the power grid and charging piles. In this case, the system needs to analyze the load of the charging piles supported by the power grid, the battery parameters, and the state parameters of each charging port in detail to adjust the load balancing of the charging piles. To achieve this goal, the system sends these parameter data to the load balancing unit. Specifically, the load capacity of the power grid, the current voltage and temperature of the battery, and the real-time output current and voltage of each port of the charging pile are transmitted as input data to the load balancing unit. This unit will dynamically adjust the power distribution of the charging pile based on the real-time data received using a reinforcement learning algorithm. The reinforcement learning algorithm finds the optimal charging strategy based on the historical load and current state of the power grid and the ports through continuous iteration, thereby avoiding overloading of the power grid. For example, when the system detects that the current of a charging port is too high, the load balancing unit will reduce the charging power of that port and distribute more power to other ports with lower load to ensure the balance of the overall load. When the type of abnormality is a charging abnormality, the system will focus on the analysis of the battery state. This type of abnormality usually means that there may be potential risks in the battery itself, such as excessive temperature, excessive charging current, or abnormal voltage fluctuations, etc. For this purpose, the system sends detailed battery parameters and port state parameters to the risk assessment unit. The risk assessment unit uses a machine learning-based risk prediction model internally. This model can identify potential risk factors in battery parameters through analysis and learning of historical data. For example, if the system detects that the temperature of the battery is rapidly rising in a short period of time, or the battery voltage is fluctuating unstably, the risk assessment unit will combine similar data from the past to determine whether the current charging behavior may cause safety problems. The unit will comprehensively evaluate the received battery parameters and port states to generate a risk level and output specific risk prompts. For example, if the risk assessment model determines that the current charging has a high risk, the system will immediately trigger an alarm to notify users or maintenance personnel to take appropriate measures, such as pausing charging or reducing the charging rate, to ensure safety.

[0068] Further, the load balancing unit is configured to perform the following steps:

[0069] a preset ratio of the power grid supportable charging pile load to the charging pile disposable load, and generating the real-time charging pile disposable load according to the real-time power grid supportable charging pile load;

[0070] The current total load of the charging pile is calculated according to the state parameters of the ports, and a reinforcement learning model is constructed with the current total load of the charging pile, real-time distributable load of the charging pile and battery parameters of each port as states and load adjustment of the ports as actions, and load adjustment is output through the reinforcement learning model.

[0071] In some embodiments, first, the system needs to preset the ratio of the grid supportable charging pile load to the charging pile disposable load. The grid supportable charging pile load represents the maximum load capacity that the grid can provide to the charging pile under the current state, while the charging pile disposable load indicates the load range that the charging pile can flexibly adjust under this load capacity. To ensure that the charging pile does not overload, the system usually sets a safety factor according to the grid stability to represent the proportional relationship between the charging pile disposable load and the grid supportable charging pile load. For example, if the safety factor is set to 0.8, it means that the disposable load of the charging pile is 80% of the maximum support load of the grid. In actual operation, when the grid state changes, the system generates the real-time disposable load of the charging pile according to this safety factor. After obtaining the real-time charging pile disposable load, the system calculates the current total load of the charging pile according to the state parameters of each port. The state parameters include the output voltage, output current, battery temperature, and battery voltage of each port. These parameters are monitored in real time by high-precision sensors on the port and sent to the load balancing unit for calculation. Through these data, the system can accurately judge the load status of each port at the moment, and thus obtain the total load of the charging pile. This total load represents the power consumed by all ports together, which is compared with the real-time disposable load of the charging pile as the basis for the next decision. The decision of load balancing relies on the construction of a reinforcement learning model. The reinforcement learning model represents the state of the charging pile as a state space, including the current total load, real-time disposable load, and battery parameters of each port. Each state in the state space represents the running status of the charging pile at a certain moment. Based on this state, the system aims to find the optimal load adjustment strategy to maintain the stability of the grid and the charging pile. In the reinforcement learning model, the load adjustment of each port (such as adjusting the charging current or voltage) is regarded as the action of the model. The action space includes increasing or decreasing the charging power of the port, and through these actions, the system can adjust the load distribution of each port. The training process of the reinforcement learning model aims to minimize the instability of the load and maximize the efficiency of charging. For this purpose, the model uses a reward function to measure the effect of each action. The design of the reward function is the key to the whole model, which evaluates the good or bad of each action according to the state change of the port, the stability of the grid, and the charging efficiency. For example, if the adjusted port load stabilizes the total load of the charging pile within the disposable range and optimizes the charging efficiency of the battery, the model will give a high reward; conversely, if the adjusted load exceeds the disposable range or causes fluctuations in the grid, it will be punished. The specific expression of the reward function may involve the weighted combination of multiple factors, such as charging time, battery health status, and grid stability. The model adjusts the load strategy of the port through repeated experiments and experience accumulation to find the optimal load adjustment scheme.In this process, the reinforcement learning algorithm uses algorithms such as Q-learning or Deep Q Network (DQN), and uses experience replay and Q-value update mechanism to optimize the model. Each load adjustment action updates the Q-value to represent the state-action pair in the long-term benefits. As the training goes deeper, the model gradually converges and can output the optimal load adjustment scheme under different grid and charging pile states.

[0072] Further, the reinforcement learning model is formulated as follows:

[0073]

[0074] where s is the state, including battery parameters and charging pile port state parameters; a is the action, including all port load adjustment parameters; Q(s, a) is the expected return that can be obtained after taking action a in state s; r is the immediate reward value after executing the adjustment action; γ is the discount factor; is the execution action; s' is the next state after executing action a; a' is the optimal adjustment action; max a′ Q(s', a') is the maximum expected return that can be obtained by selecting the optimal adjustment action a' in the next state s'.

[0075] Specifically, the state s represents all the environmental information of the charging system at a specific time point, which includes battery parameters and charging pile port state parameters. Battery parameters can include battery temperature, voltage, and state of charge, etc., while charging pile port state parameters cover output voltage, output current, and load situation of each port. The combination of all these parameters constitutes the current state of the system, describing the charging status at the current time. The expected return Q(s, a) represents the cumulative reward the system can achieve in the long term after choosing action a in state s. This value reflects the future impact of the current action on charging efficiency and grid stability. The expected return is the core goal of the reinforcement learning model, and the system continuously adjusts the Q value through multiple trials and training to find the best adjustment strategy in different states. The immediate reward value r is the feedback the system receives immediately after performing an action in a certain state. The reward value is usually based on the effectiveness of the current charging behavior, for example, when the adjusted charging power reduces the battery temperature or improves the charging efficiency, the system will receive a positive reward; on the contrary, if the adjustment leads to grid fluctuations or abnormal battery temperature rise, the system will be punished. The design of the reward value needs to consider the charging safety and efficiency to guide the system to learn the best adjustment strategy. The discount factor γ measures the value of future rewards. The discount factor is between 0 and 1, balancing the weight of immediate rewards and long-term returns. A larger value of γ indicates that the system pays more attention to long-term optimization, while a smaller value of γ emphasizes current immediate income. For the charging management system, appropriate setting of γ can ensure that the system optimizes the current charging behavior while considering the possible future impact. The next state s' is the new state the system enters after performing action a. This new state depends on the impact of the action on the current charging system, for example, after adjusting the output current of a certain port, the battery temperature may change and the load distribution of the charging pile will be adjusted. The system records this change and uses it as the basis for the next decision. The optimal adjustment action a' is the best adjustment strategy the system can take in the next state s'. By searching and evaluating all possible actions, the system will choose the optimal action that maximizes the expected return. The process of selecting the optimal action is based on the existing Q value calculation, and the system will continuously update and optimize these values to ensure the rationality of each decision. The maximum expected return is the highest return the system can achieve after choosing the optimal action a' in state s'. This parameter is used to update the Q value, helping the system evaluate the long-term effect of each action. By continuously calculating the maximum expected return, the system can gradually identify the optimal load adjustment strategy.

[0076] Further, the formula of the immediate reward value r is as follows:

[0077] r = -a · |L current -L optimal -β · E loss ;

[0078] wherein, L current is the current load of the charging pile; L optimal is the real-time available load of the power grid; E loss is the energy loss due to load imbalance or power loss; and a and b are the weight coefficients of load balance and energy loss, respectively.

[0079] Specifically, L optimal is the optimal load value calculated according to the available load of the power grid and the charging demand of the electric vehicle, ensuring that the impact of excessive load on the power grid is avoided while meeting the charging efficiency of the electric vehicle. The parameter E loss represents the energy loss due to load imbalance or power loss during the charging process. The two weight coefficients a and b correspond to the importance of load balance and energy loss, respectively, and the specific values depend on the optimization goal of the system and the characteristics of the power grid. In the algorithm implementation, the system first collects the battery parameters of the electric vehicle and the state parameters of the charging pile port through sensors, and calculates the current load L current in real time. The system simultaneously analyzes the available load data of the power grid, combines the charging demand of the electric vehicle, and dynamically generates the optimal load value L optimal . At each time step, the system calculates the immediate reward value r based on the above formula to evaluate the pros and cons of the current charging adjustment strategy. If L current is significantly different from L optinal , the immediate reward value will be significantly reduced, indicating that there is a large imbalance in the current load configuration. At the same time, if E loss is high, the impact of energy loss on the immediate reward value will also increase, and by increasing the penalty weight b of energy loss, the system can more effectively control power distribution and reduce energy waste during the charging process. In each charging load adjustment process, the system uses the current immediate reward value as feedback information for the reinforcement learning model to update the strategy. Specifically, the reinforcement learning model continuously adjusts the output power by maximizing the cumulative reward to optimize the load adjustment scheme. In the training phase, the system uses historical data to build an experience pool and trains the model based on the immediate reward value using Q-learning or DQN (Deep Q Network) algorithm. The model gradually learns how to choose the optimal load adjustment action under different charging environments and power grid load conditions, so as to make better decisions in future real-time charging scenarios.

[0080] Further, the construction of the risk assessment model according to the abnormal types, battery parameters, and state parameters of each port in the history includes the following steps:

[0081] Based on the abnormal data in the history, a historical data set containing abnormal types, battery parameters, and state parameters of each port is collected, and each record is labeled with the corresponding risk level.

[0082] dividing the historical dataset with completed risk level annotations into a training set and a test set;

[0083] based on a decision tree algorithm, recursively performing binary segmentation on the parameters in the training set, outputting risk levels through leaf nodes, and obtaining a risk assessment model;

[0084] verifying the risk assessment model through the test set.

[0085] Specifically, first, the system forms a dataset containing abnormal types, battery parameters (such as battery model, capacity, temperature, battery voltage, charging rate, etc.), and state parameters of each port (such as output voltage, output current, load power) through the collection of historical data. To ensure the accuracy and effectiveness of model training, each historical record is annotated with the corresponding risk level, and the risk level is set according to the specific abnormal type and its potential impact on the safety of the charging system. Through this annotation, the historical dataset becomes a complete dataset that can be used for supervised learning. In the data processing stage, the system divides the annotated historical dataset into a training set and a test set. The training set is used to build the model, and the test set is used for subsequent verification and evaluation. Such division can ensure that the model still has good prediction ability and generalization ability when facing unseen data. Then, the system uses a decision tree algorithm to process the data in the training set. The decision tree algorithm recursively performs binary segmentation on each feature when building the model. Specifically, the algorithm selects a feature at each node, and divides the dataset into two parts using a certain threshold value of the feature. This segmentation process selects the optimal segmentation point according to indicators such as information gain, entropy, or Gini coefficient. In this way, through recursive segmentation, a tree structure is finally generated, where each leaf node corresponds to a specific risk level output. These leaf nodes reflect the combination of battery parameters and port state parameters under different abnormal conditions, providing a basis for risk assessment during the charging process. After the model is built, to verify its accuracy, the system uses the previously divided test set to verify the risk assessment model. The data in the test set did not participate in the training of the model. By inputting the data in the test set into the built model, the system can predict the risk level and compare it with the actual annotated risk level. Through this verification process, the accuracy and precision of the model can be evaluated, and whether the model is overfitting or underfitting can be found. If the verification result shows that the model performs poorly on certain specific abnormal types, the system will further adjust the parameters of the model, such as changing the maximum depth of the decision tree, adjusting the segmentation standard, etc., to improve the prediction performance of the model in different risk scenarios. Finally, after multiple verifications and optimizations, the risk assessment model can effectively deal with various abnormal situations during the charging process, provide accurate risk warnings, and improve the safety and stability of the charging system.

[0086] The above embodiments are only used to describe the preferred embodiments of the present application, and are not used to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements made by those skilled in the art to the technical solutions of the present application shall fall within the protection scope of the claims of the present application.

Claims

1. A comprehensive electric vehicle charging management system based on edge computing, characterized in that, It includes an electric vehicle-side data acquisition module, a charging pile-side data acquisition module, and an intelligent management module, wherein the electric vehicle-side data acquisition module and the charging pile-side data acquisition module are respectively connected to the intelligent management module; The electric vehicle-side data acquisition module is used to collect battery parameters in the electric vehicle and send them to the edge computing unit, including battery model, capacity, temperature, battery voltage and charging rate; The charging pile terminal acquisition module is used to acquire the status parameters of the charging pile load that the power grid can support and the status parameters of each port of the charging pile. The status parameters of the port include output voltage, output current and load power, and are sent to the edge computing unit. The intelligent management module includes an edge computing unit, a load balancing unit, and a risk assessment unit, and the edge computing unit is communicatively connected to the load balancing unit and the risk assessment unit, respectively. The edge computing unit is used to determine whether the current charging behavior is abnormal and to determine the type of abnormality based on the grid-supportable charging pile load, battery parameters and the status parameters of each port through deep learning. Based on the type of abnormality, the grid-supportable charging pile load, battery parameters and the status parameters of each port are allocated and sent to the load balancing unit or the risk assessment unit. The load balancing unit is used to adjust the load based on the power grid's support for charging pile loads, battery parameters, and the status parameters of each port, according to a reinforcement learning model. The risk assessment unit is used to construct a risk assessment model based on historical anomaly types, battery parameters, and status parameters of each port. The risk assessment model outputs the risk level and provides alarm prompts based on real-time anomaly types, battery parameters, and status parameters of each port. The load balancing unit is used to perform the following steps: The ratio of the load that the power grid can support to the available load of the charging pile is preset, and the available load of the charging pile is generated in real time based on the load that the power grid can support. The current total load of the charging pile is calculated based on the state parameters of the port. The current total load of the charging pile, the real-time available load of the charging pile, and the battery parameters of each port are used as the state. The load adjustment of the port is used as the action to build a reinforcement learning model. The load adjustment is output through the reinforcement learning model. The formula for the reinforcement learning model is as follows: ； in, This refers to the status, including battery parameters and charging station port status parameters; For actions, including load regulation parameters for all ports; In the state Take action below The expected return that can be obtained afterward; The immediate reward value after performing the adjustment action; Discount factor; Action to be performed; To perform the action The next state after; This is the optimal adjustment action; For the next state Next, select the optimal adjustment action. The maximum expected return that can be obtained; The instant reward value The formula is as follows: ； in, The current load of the charging station; The available load for real-time charging piles; This refers to energy loss caused by load imbalance or power loss. and These are the weighting coefficients for load balance and energy loss, respectively.

2. The electric vehicle charging integrated management system based on edge computing according to claim 1, characterized in that, The edge computing unit includes a battery central control subunit and a charging pile central control subunit, which are communicatively connected. The battery central control subunit is used to receive battery parameters, standardize them, and send them to the charging pile central control subunit; The central control subunit of the charging pile is used to perform the following steps: The power grid supports the charging pile load, battery parameters, and port status parameters as inputs to the feedforward neural network, and outputs anomaly types, including load anomalies and charging anomalies. Based on the anomaly type, the battery parameters and the status parameters of each port are sent to the load balancing unit or the risk assessment unit.

3. The electric vehicle charging integrated management system based on edge computing according to claim 2, characterized in that, The training steps of the feedforward neural network include: The load, battery parameters, and port status parameters of the charging piles supported by the power grid are normalized, and the anomaly types are labeled. The normalized power grid can support the decomposition of each parameter among the charging pile load, battery parameters, and port state parameters to construct input neurons; The features of the input neurons are calculated and extracted layer by layer through the hidden layers in the feedforward neural network. The activation function is used to perform non-linear feature mapping, extract the high-order features of the abnormal pattern, and integrate them through the output layer to output the result. The loss function is calculated based on the output of the output layer and the labeled anomaly types, and the network weights are adjusted using the backpropagation algorithm.

4. The electric vehicle charging integrated management system based on edge computing according to claim 2, characterized in that, The step of sending battery parameters and status parameters of each port to the load balancing unit or the risk assessment unit according to the anomaly type includes: If the anomaly type is load anomaly, the grid-supported charging pile load, battery parameters, and status parameters of each port will be sent to the load balancing unit. If the anomaly type is charging anomaly, the battery parameters and the status parameters of each port will be sent to the risk assessment unit.

5. The electric vehicle charging integrated management system based on edge computing according to claim 1, characterized in that, The risk assessment model constructed based on historical anomaly types, battery parameters, and status parameters of each port includes the following steps: Based on historical anomaly data, a historical dataset containing anomaly types, battery parameters, and status parameters of each port is collected, and each record is labeled with its corresponding risk level. The historical dataset with completed risk level labeling is divided into a training set and a test set; Based on the decision tree algorithm, the parameters in the training set are recursively split into binary categories, and the risk level is output through the leaf nodes to obtain the risk assessment model. The risk assessment model was validated using a test set.

6. The electric vehicle charging integrated management system based on edge computing according to claim 1, characterized in that, Both the risk assessment unit and the load balancing unit are deployed on cloud servers.

Images