A Method and System for Predicting and Adaptively Controlling Voltage Drop in Charging Piles Based on Artificial Intelligence

By combining the cloud-based Transformer-LSTM hybrid prediction model and the edge prediction model, along with the electrical digital twin model and PPO decision model of the charging station, the system achieves early prediction and adaptive control of the voltage drop risk of charging piles. This solves the problem of difficulty in predicting and adapting to different scenarios in existing technologies, and improves the utilization rate of charging equipment and user experience.

CN123104451APending Publication Date: 2026-11-17CHINA CONSTR EIGHTH BUREAU FIRST DIGITAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611609992.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing charging pile voltage drop management methods are unable to predict the voltage drop risk at the AC input end in advance, and lack the ability to adapt to different times, different loads and different operating scenarios, resulting in charging interruptions and reduced equipment utilization.

Method used

An AI-based method for predicting and adaptively controlling voltage drop in charging piles is adopted. By combining a cloud-based Transformer-LSTM hybrid prediction model and an edge prediction model, multi-source feature correlation and time-series modeling are performed to form a priori voltage drop trend. The electrical digital twin model of the charging station and the PPO decision model are used to screen safe and executable control actions, including energy storage compensation actions.

Benefits of technology

It enables advance prediction and adaptive control of voltage drop risk at the AC input terminal of charging piles, improves the utilization rate of charging equipment and user charging experience, and ensures the safety and feasibility of control actions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104451A_ABST
    Figure CN123104451A_ABST
Patent Text Reader

Abstract

This invention provides an artificial intelligence-based method and system for predicting and adaptively controlling voltage drop in charging piles, belonging to the field of electric vehicle charging control technology. The method collects multi-source operational data from charging piles and charging stations, generates a voltage drop trend prior using a cloud-based Transformer-LSTM hybrid prediction model, and corrects this prior using an edge prediction model combined with real-time electrical data to obtain the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop. Based on the current operating state, continuous value candidate control actions are generated. The effective value state of the AC input voltage after each candidate action is pre-simulated using an electrical digital twin model of the charging station, and safe and executable candidate actions are selected. Finally, the PPO decision model determines and executes the final control action. The voltage drop prediction model and the PPO decision model are updated based on actual operational feedback. This invention can predict voltage drop before it occurs and perform electrical safety screening before the control action is executed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging control technology, and in particular to a method and system for predicting and adaptively controlling the voltage drop of charging piles based on artificial intelligence. Background Technology

[0002] With the continuous increase in the number of electric vehicles, charging infrastructure such as charging piles has been widely deployed. During the operation of charging stations, when multiple electric vehicles are charging simultaneously or the power grid load is heavy, the voltage at the AC input terminal of the charging pile may drop significantly. In severe cases, this can trigger the charging protection mechanism and cause charging interruption, thus affecting the user's charging experience and the utilization rate of charging equipment. Existing charging pile voltage drop management methods mostly adopt reactive control, that is, power adjustment or compensation is performed only after a voltage drop is detected, making it difficult to predict the risk before the voltage drop occurs. At the same time, some existing solutions mainly rely on limited data such as the real-time voltage of the charging pile itself, and do not make sufficient use of multi-source data such as weather, holidays, historical charging behavior, and the operating status of other charging piles in the station, resulting in limited predictive capabilities. In addition, existing control strategies usually use fixed control parameters, which are not adaptable to different times, loads, and operating scenarios, and are also difficult to coordinate the adjustment of multiple charging piles at the overall charging station level. Therefore, how to predict the voltage drop risk at the AC input terminal of the charging pile in advance and adjust the charging control strategy in a timely manner based on the actual operating status has become a technical problem that needs to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a charging pile voltage drop prediction and adaptive control method and system based on artificial intelligence, so as to realize the early prediction of the low voltage risk at the AC input terminal of the charging pile, and to pre-screen the voltage safety and equipment feasibility of candidate control actions before the actual execution of control actions.

[0004] This invention is achieved through the following measures: A method for predicting and adaptively controlling voltage drop in charging piles based on artificial intelligence, characterized by the following steps: S1 collects the effective value of AC input voltage, current, charging power, station load, operating status of other charging piles, historical charging records, and environmental and time information of the charging piles. The collected data is preprocessed to form a standardized multi-source feature dataset, edge real-time electrical data, and the current operating status of the charging station. S2, input the standardized multi-source feature dataset into the cloud Transformer-LSTM hybrid prediction model, establish global correlation between multi-source features based on Transformer, and perform temporal evolution modeling on the correlated features based on LSTM to form a priori cloud pressure drop trend within a future preset time window. S3, input the prior information of the cloud voltage drop trend, the real-time electrical data of the edge, and the current operating status of the charging station into the edge prediction model for short-time scale correction, and obtain the corrected voltage drop prediction result including the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop; S4. Based on the current operating status of the charging station and the current allowable power adjustment range of the charging pile, multiple candidate control actions with consecutive values ​​are formed. The candidate control actions include at least a charging power adjustment action. The correction voltage drop prediction result, the current operating status of the charging station, and each candidate control action are input into the electrical digital twin model of the charging station. The effective value of the AC input voltage of the charging pile after the execution of each candidate control action is predicted. A set of safe and executable candidate actions is formed by screening based on voltage safety conditions and charging pile operating constraints. When the charging station is equipped with controllable energy storage, the candidate control actions also include energy storage compensation actions. The screening also considers energy storage operating constraints. S5, input the set of safe executable candidate actions and the digital twin pre-simulation results corresponding to each safe executable candidate action into the PPO decision model, determine the final control action from the set of safe executable candidate actions, adjust the charging power of the charging pile according to the final control action, and control the controllable energy storage to perform compensation when the final control action includes an energy storage compensation action; S6, collect actual operation feedback data after the final control action is executed; for the operating range where active control to change the target charging power is not executed and energy storage compensation is not executed, form voltage drop prediction model update data based on the corresponding correction voltage drop prediction result and actual operating state; and form PPO decision model interactive update data based on the operating state before control, the actual executed action and the operating state after execution, so as to update the corresponding models respectively.

[0005] The invention also has the following specific features: The cloud-based Transformer-LSTM hybrid prediction model first establishes a global correlation between multi-source features through Transformer, and then performs temporal evolution modeling on the correlated features through LSTM; the edge prediction model uses the cloud-based voltage drop trend as prior information, and combines real-time electrical data at the edge with the current operating status of the charging station for short-time scale correction, outputting the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop.

[0006] The electrical digital twin model of the charging station is used to characterize the electrical characteristics of the power supply line from the low-voltage side of the station-level distribution transformer to the AC input terminal of each charging pile, the response relationship between the power of each charging pile and the effective value of the AC input terminal voltage, and the current operating status of each charging pile. The corrected voltage drop prediction result is used as the benchmark voltage drop risk state within a future preset time window, and each candidate control action is applied to obtain the trajectory of the predicted effective value of the AC input terminal voltage of the charging pile after the execution of each candidate control action.

[0007] The formation of the continuous value candidate control action includes: determining the actual power adjustment ratio range based on the power adjustment range configured by the system and the current allowed charging power range of the charging pile; selecting multiple candidate power adjustment ratios within the actual power adjustment ratio range; and determining the corresponding candidate target charging power based on the current charging power.

[0008] The voltage safety margin is determined based on the lowest predicted voltage effective value in the AC input terminal predicted voltage effective value trajectory corresponding to each candidate control action and the voltage safety threshold. Candidate control actions with a voltage safety margin greater than zero and that meet the charging pile operation constraints are retained in the set of safe executable candidate actions.

[0009] When a controllable energy storage system is configured in a charging station, the candidate control actions also include energy storage compensation state and energy storage compensation power. The energy storage compensation power is limited by the energy storage state of charge, available charging and discharging power, and remaining available energy. The electrical digital twin model of the charging station also simulates the impact of energy storage compensation on the effective value of the AC input voltage of the charging pile and filters it according to the energy storage operation constraints.

[0010] The PPO decision model only evaluates the policy of candidate control actions in the set of safe and executable candidate actions; when the set of safe and executable candidate actions contains only one candidate control action, it is directly determined as the final control action; when multiple candidate control actions have the same highest policy evaluation, the final control action is determined according to a fixed candidate order.

[0011] For operating ranges where no active control to change the target charging power is executed and no energy storage compensation is executed, the voltage drop prediction model updates data based on the actual voltage drop events, actual voltage drop amplitude, and actual voltage drop cumulative duration; for operating ranges where the active control or energy storage compensation is executed, a control intervention flag is added; and the PPO decision model interactively updates data based on the operating state before control, the actual executed action, the reward value, and the operating state after execution.

[0012] An artificial intelligence-based charging pile voltage drop prediction and adaptive control system includes: The data acquisition and preprocessing module is used to collect the effective value of the AC input voltage of the charging pile, current, charging power, station load, operating status of other charging piles, historical charging records, and environmental and time information, and to form a standardized multi-source feature dataset, edge real-time electrical data, and current charging station operating status. The cloud-based pressure drop trend prediction module is used to perform global correlation and temporal evolution modeling of the standardized multi-source feature dataset to form a priori cloud-based pressure drop trend. The edge real-time correction module is used to generate a corrected voltage drop prediction result based on the cloud voltage drop trend prior, edge real-time electrical data and the current charging station operating status, including the probability of voltage drop occurrence, the predicted voltage drop amplitude and the predicted cumulative duration of voltage drop. The candidate action formation and digital twin pre-simulation screening module is used to generate multiple consecutive candidate control actions based on the current operating status of the charging station and the current allowable power adjustment range of the charging pile. The module inputs the corrected voltage drop prediction result, the current operating status of the charging station, and each candidate control action into the electrical digital twin model of the charging station. It then predicts the effective value of the AC input voltage of the charging pile after each candidate control action is executed. Based on voltage safety conditions and charging pile operating constraints, it screens and forms a set of safe and executable candidate actions. When the charging station is equipped with controllable energy storage, the candidate control actions also include energy storage compensation actions, and the screening also considers energy storage operating constraints. The PPO decision and control execution module is used to input the set of safe executable candidate actions and the digital twin pre-simulation results corresponding to each safe executable candidate action into the PPO decision model, determine the final control action from the set of safe executable candidate actions, adjust the charging power of the charging pile according to the final control action, and control the controllable energy storage to perform compensation when the final control action includes an energy storage compensation action. The effect feedback and model update module is used to generate pressure drop prediction model update data and PPO decision model interactive update data based on the actual operation feedback data after the final control action is executed.

[0013] The cloud-based voltage drop trend prediction module is deployed in the cloud, and the edge real-time correction module is deployed on the charging station edge controller. The cloud transmits the cloud-based voltage drop trend prior to the edge controller. The edge controller combines the edge real-time electrical data and the current operating status of the charging station to form the corrected voltage drop prediction result, and sends control commands to the charging pile. When a controllable energy storage system is configured and the final control action includes an energy storage compensation action, the controller sends an energy storage control command to the energy storage control device.

[0014] The beneficial effects of this invention are as follows: Compared with the prior art, this invention combines long-term trend prediction in the cloud with real-time edge correction, enabling voltage drop prediction to simultaneously reflect the station-level load evolution trend and the current real-time electrical status; before the execution of control actions, the electrical digital twin model of the charging station is used to perform electrical pre-simulation of multiple candidate control actions, and actions that do not meet voltage safety conditions or equipment operation constraints are excluded, so that the PPO decision model can select strategies within the range of safe and executable candidate actions; the voltage drop prediction model and the PPO decision model are updated respectively through actual operation feedback, thereby forming a closed loop of prediction, pre-simulation screening, control execution and feedback update. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the Transformer-LSTM hybrid voltage drop prediction model in this invention.

[0017] Figure 3 This is a diagram illustrating the algorithmic framework for deep reinforcement learning PPO adaptive decision-making in this invention.

[0018] Figure 4 This is a flowchart illustrating the workflow of the system of the present invention in a typical application scenario. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to describe the solution.

[0021] Example 1 See Figures 1-4 A method for predicting and adaptively controlling voltage drop in charging piles based on artificial intelligence includes the following steps: S1 collects the effective value of AC input voltage, current, charging power, station load, operating status of other charging piles, historical charging records, and environmental and time information of the charging piles. The collected data is preprocessed to form a standardized multi-source feature dataset, edge real-time electrical data, and the current operating status of the charging station. In step S1, multi-source data related to the voltage changes on the AC input side of the charging piles are collected during the operation of the charging station. The collected data is preprocessed to form a standardized multi-source feature dataset, edge real-time electrical data, and the current operating status of the charging station. The multi-source data includes charging pile operation data, charging station operation data, historical charging data, environmental data, and time information. Among them, the charging pile operation data includes at least the effective value of the AC input voltage, input current, and current charging power of each charging pile; the charging station operation data includes at least the station-level load and the operating status of other charging piles in the station; the historical charging data includes the start and end times of historical charging processes, charging power, voltage drop, or charging interruption records; the environmental data includes ambient temperature and available humidity, wind speed, and weather conditions; and the time information includes the current time period, date type, and whether it is a weekday or holiday.

[0022] Specifically, the effective value of the AC input voltage, input current, and charging power of each charging pile are obtained in real time through the charging pile's own electrical measurement unit or the energy metering device connected to the charging pile; the station-level load is obtained through the charging station's low-voltage side metering device, station-level energy management device, or station-level controller; the operating status of other charging piles within the station is obtained through the charging station's internal communication network. For charging piles currently performing charging tasks, the state of charge of the currently served vehicles can also be obtained through the charging communication interface and included as part of the current charging status. For charging stations equipped with controllable energy storage systems, the state of charge and current available charging and discharging capacity of the energy storage system are also collected; when no controllable energy storage system is configured, corresponding energy storage operation data is not generated.

[0023] Historical charging data is obtained from the charging station's local database or data storage nodes connected to the charging station, and is associated with charging piles and charging tasks. Environmental data can be obtained from the charging station's environmental sensors or through an external meteorological data interface; date type, weekday or holiday information, and time period information are determined based on system time and preset calendar information, establishing a time correspondence between the electrical status of charging piles, station load, operating status of other charging piles, environmental status, and time status within the same time interval.

[0024] Furthermore, data with different sampling periods and update times are time-aligned according to a unified time reference. For continuously collected AC input voltage RMS values, input current, and charging power data, raw time-series data are obtained according to a preset sampling frequency, and feature aggregation is performed according to a preset time window. In some implementations, a raw sampling frequency of 100Hz can be used, with a feature aggregation window of 10 seconds; the sampling frequency and aggregation window can be set according to the charging pile's measurement capabilities and edge computing resources. For data such as weather, date type, and holiday information, which have a lower update frequency than electrical data, they are matched with the corresponding electrical time-series data according to the corresponding valid time interval.

[0025] Data cleaning, missing data processing, outlier data processing, and normalization are performed on multi-source data that has been aligned to the specified time. For isolated missing values ​​in continuous electrical data, if the duration of the missing value is within a preset compensable range, interpolation can be performed based on adjacent valid sample values ​​or the most recent valid value can be used to fill the gap. For temporary missing auxiliary data such as weather and date types, the most recent valid data or preset missing value processing rules can be used to fill the gap. For data that exceeds the range of the corresponding measuring equipment, is significantly beyond the physical operating range of the equipment, or is distorted due to communication anomalies, the corresponding data is marked as outlier data and is not used directly as valid feature data.

[0026] When key real-time electrical data such as the effective value of the AC input voltage and the input current are continuously missing, and valid edge real-time electrical data or the current operating status of the charging station cannot be formed, subsequent voltage drop prediction and adaptive control processing based on this invalid data will not be initiated; subsequent processing will only begin after the key real-time electrical data becomes valid. During this period, the charging pile maintains its current valid operating status and operates according to its existing local control and protection logic.

[0027] After data cleaning, feature processing is performed on various types of valid data. For continuous time-series data such as AC input voltage RMS, input current, charging power, and station load, time-series features reflecting their changes over time are retained, and corresponding statistical features are extracted. When it is necessary to characterize periodic or frequency-varying features, frequency domain features can also be extracted. Data categorized by current time period, date type, holiday status, and weather status are numerically encoded, and the encoded results are correlated with electrical features within the same time window. Continuous numerical features are normalized according to the preprocessing rules consistent with the prediction model training phase, so that features with different dimensions and value ranges can be used as model input. Data cleaning, missing value handling, outlier detection, normalization, and extraction of time-series, frequency domain, and statistical features are all completed in this step.

[0028] After the above processing, a standardized multi-source feature dataset is formed. The standardized multi-source feature dataset includes electrical operation features, station-level load features, other charging pile operation features, historical charging features, environmental features, and time features organized in chronological order, while maintaining the temporal correspondence between various features. The standardized multi-source feature dataset is used in step S2.

[0029] Simultaneously, edge real-time electrical data is generated from the real-time sampled data after time synchronization and anomaly verification. This edge real-time electrical data includes at least the currently valid effective value of the AC input voltage of the charging pile, the input current, and the charging power. Furthermore, the current charging station operating status is generated based on the operating status of each charging pile, the station-level load, and the current charging status obtained at the same time. When a controllable energy storage system is configured, the obtained energy storage state of charge and available charge / discharge capacity are saved as optional energy storage operating states corresponding to the current charging station operating status. The edge real-time electrical data and the current charging station operating status are used in step S3, and the current charging station operating status is also used in step S4.

[0030] S2, input the standardized multi-source feature dataset into the cloud Transformer-LSTM hybrid prediction model, establish global correlation between multi-source features based on Transformer, and perform temporal evolution modeling on the correlated features based on LSTM to form a priori cloud pressure drop trend within a future preset time window. In step S2, the standardized multi-source feature dataset formed in step S1 is input into the Transformer-LSTM hybrid prediction model deployed in the cloud to predict the evolution trend of AC input voltage drop risk of the charging station within a future preset time window, forming a cloud-based voltage drop trend prior. The Transformer-LSTM hybrid prediction model is a pre-trained prediction model that has been validated for effectiveness. The model input adopts the same feature organization order and data preprocessing rules as the training phase.

[0031] Specifically, the standardized multi-source feature dataset is organized into a continuous multi-source feature sequence according to chronological order, ensuring that the electrical operation features of charging piles, station-level load features, other charging pile operation features, historical charging features, environmental features, and time features corresponding to the same time location remain consistent. The multi-source feature sequence is first input into a Transformer encoding layer, which uses a multi-head self-attention mechanism to establish global correlations between features from different sources and between features at different time locations. This yields a feature sequence that integrates multi-source correlation information, providing a unified representation of the correlation between station-level load changes, the operating status of other charging piles, environmental conditions, time conditions, and the effective value changes of the AC input voltage of the charging piles.

[0032] Subsequently, the feature sequences obtained after processing by the Transformer encoding layer are input into the LSTM layer in chronological order. The LSTM layer performs temporal evolution modeling on the correlated and enhanced feature sequences, extracting the continuity of the charging station's operating status over time and the time dependency between the effective value of the AC input voltage of the charging piles and the station-level load evolution, and based on this, forms a voltage drop trend prediction result for a preset time window in the future. Thus, the Transformer is used to establish global correlations between multi-source heterogeneous features, and the LSTM is used to extract the dynamic laws of feature evolution over time based on the global correlations. The two are processed in series in the order of Transformer first, followed by LSTM.

[0033] In some implementations, the future preset time window is set according to the charging station load change rate and control requirements, for example, set to 5–30 minutes. The prediction time window is used to define the future time range corresponding to the risk of AC input side voltage drop of charging piles caused by station-level load evolution predicted in this step.

[0034] The Transformer-LSTM hybrid prediction model generates a cloud-based voltage drop trend prior based on the temporal characteristics of the LSTM layer output. This cloud-based voltage drop trend prior characterizes the overall evolution trend of the effective value of the AC input voltage of the charging pile within a future preset time window under the current multi-source operating conditions, and includes initial cloud-based prediction information corresponding to the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of the voltage drop. As a prediction result on a longer time scale, the cloud-based voltage drop trend prior does not form a charging power adjustment command in this step.

[0035] Furthermore, verified and usable versions of the Transformer-LSTM hybrid prediction model are stored in the cloud. If the parameter file for the currently intended model is missing, the model fails to load, or the model's running state is abnormal and cannot complete effective inference, this abnormal model will not be used to form a new cloud-based pressure drop trend prior. If a previously verified and usable model version exists, it will be used to complete the prediction processing in this step. If no cloud-based model version capable of completing effective inference exists, a new cloud-based pressure drop trend prior will not be formed in this control cycle, and subsequent processing based on this prediction link will be stopped until a valid model is obtained, at which point prediction processing will resume.

[0036] The cloud pressure drop trend prior formed in this step serves as the input for step S3.

[0037] S3, input the prior information of the cloud voltage drop trend, the real-time electrical data of the edge, and the current operating status of the charging station into the edge prediction model for short-time scale correction, and obtain the corrected voltage drop prediction result including the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop; In step S3, the cloud voltage drop trend prior formed in step S2, along with the edge real-time electrical data and the current charging station operating status formed in step S1, are input into the edge prediction model deployed on the edge side. The cloud voltage drop trend prior is then corrected on a short time scale to generate a corrected voltage drop prediction result corresponding to the current real-time operating status of the charging station. The edge prediction model is a pre-trained prediction model deployed on the edge controller, and its input interface is matched with the cloud voltage drop trend prior, the edge real-time electrical data, and the current charging station operating status.

[0038] Specifically, the cloud-based voltage drop trend prior is used as prior information reflecting the evolution of station-level load and the change trend of the effective value of AC input voltage of charging piles within a preset time window. Simultaneously, the current effective value of AC input voltage, input current, charging power, and the current operating status of the charging station are input into the edge prediction model. Based on the cloud-based voltage drop trend prior, the edge prediction model performs short-term time-scale correction by combining the current voltage, current, power, and station-level operating status reflected by local real-time electrical data. This ensures that the corrected prediction result simultaneously reflects the station-level load evolution trend over a longer time scale and the local electrical state changes near the current moment.

[0039] The prior voltage drop trend data from the cloud and the real-time data from the edge do not form independent final risk assessment results. Instead, the edge prediction model uses the prior cloud voltage drop trend as the basis for prediction, processes it together with the current real-time electrical status, and outputs a unified corrected voltage drop prediction result. Thus, the cloud prediction is used to characterize the overall trend of change within a preset time window, while the edge correction is used to correct the overall trend of change based on the current real-time status.

[0040] In some implementations, the edge prediction model employs a lightweight model suitable for edge controller operation, which can be obtained by pruning and quantizing the trained prediction model to reduce model storage and real-time inference computation. The edge prediction model uses the same input feature organization and data preprocessing rules as its training phase, and updates the current correction input based on continuously acquired local real-time electrical data to adapt to changes in the effective value of the AC input voltage, current, and charging power of the charging pile over shorter time scales.

[0041] After processing by the edge prediction model, a corrected voltage drop prediction result is generated, including the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative voltage drop duration. The probability of voltage drop occurrence represents the likelihood of a voltage drop event occurring at the AC input terminal of the charging pile within a future preset time window; the predicted voltage drop amplitude represents the predicted decrease in the effective voltage value of the AC input terminal of the charging pile relative to the corresponding reference voltage within the future preset time window, in units of V; and the predicted cumulative voltage drop duration represents the cumulative time the effective voltage value of the AC input terminal of the charging pile is in a voltage drop state within the future preset time window, in units of s. When multiple separate voltage drop time segments exist within the same prediction time window, the duration of each voltage drop time segment is cumulatively included in the predicted cumulative voltage drop duration. These three prediction results correspond to the same prediction time window and the same AC input terminal monitoring object of the charging pile.

[0042] Furthermore, before the edge prediction model runs, the validity of the currently available cloud pressure drop trend priors and the running status of the edge prediction model are checked. If a new cloud pressure drop trend prior corresponding to the current control cycle is not received due to communication interruption or other reasons, but the edge side retains the most recent valid cloud pressure drop trend prior that is still within the preset validity period, the most recent valid cloud pressure drop trend prior can be used to complete the short-term correction of this control cycle; if there is no cloud pressure drop trend prior that is within the validity period, this control cycle will not enter step S4 based on the prediction link.

[0043] If the parameters of the currently loaded edge prediction model are abnormal, the model loading fails, or the inference process cannot form a complete probability of voltage drop occurrence, predicted voltage drop amplitude, and predicted cumulative voltage drop duration, the abnormal result will not be used as the corrected voltage drop prediction result. If a previously verified and usable edge prediction model version exists, the correction process in this step will be re-executed using the previously verified and usable edge prediction model version. If no edge prediction model capable of completing effective inference exists, or if step S1 does not generate valid real-time edge electrical data and the current charging station operating status, no new corrected voltage drop prediction result will be generated in this control cycle, and the process will not proceed to step S4. Normal processing will resume only after the corresponding data or model becomes valid again.

[0044] The probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop formed in this step together constitute the corrected voltage drop prediction result, which serves as the input for step S4.

[0045] S4. Based on the current operating status of the charging station and the current allowable power adjustment range of the charging pile, multiple candidate control actions with consecutive values ​​are formed. The candidate control actions include at least a charging power adjustment action. The correction voltage drop prediction result, the current operating status of the charging station, and each candidate control action are input into the electrical digital twin model of the charging station. The effective value of the AC input voltage of the charging pile after the execution of each candidate control action is predicted. A set of safe and executable candidate actions is formed by screening based on voltage safety conditions and charging pile operating constraints. When the charging station is equipped with controllable energy storage, the candidate control actions also include energy storage compensation actions. The screening also considers energy storage operating constraints. In step S4, based on the current charging station operating status formed in step S1 and the corresponding charging pile's current allowable power adjustment range, multiple continuous value candidate control actions are formed. The corrected voltage drop prediction result formed in step S3, the current charging station operating status, and each candidate control action are input into the charging station electrical digital twin model synchronized with the current charging station operating status. The electrical operating status after each candidate control action is executed is pre-simulated. The candidate control actions are screened based on the pre-simulated AC input voltage effective value status of the charging pile and the equipment operating constraints to form a set of safe and executable candidate actions.

[0046] Specifically, for a charging pile currently performing a charging task and with a current charging power greater than 0, the actual available power adjustment ratio range is determined based on the power adjustment range configured in the system and the power range allowed for the charging pile in its current operating state. The lower and upper limits are as follows: ;

[0047] in, The current charging power is expressed in kW and is obtained from the current operating status of the charging station. and These are the minimum and maximum charging power allowed for the charging pile under the current operating conditions, respectively, both in kW, and are determined by the current equipment operating parameters and control boundaries of the charging pile. and These are the lower and upper limits of the power adjustment ratio pre-configured by the system, and are dimensionless quantities. and These are the lower and upper limits of the actual power adjustment ratio obtained after considering the overall system configuration range and the allowable power range of the charging pile equipment, respectively, both of which are dimensionless quantities.

[0048] when At that time, multiple consecutive values ​​are selected as candidate power adjustment ratios within the actual power adjustment ratio range, and the first value is determined according to the following formula. The target charging power corresponding to each candidate control action:

[0049] in, For the first The candidate power adjustment ratio is a dimensionless quantity. For the first The target charging power corresponding to each candidate control action is expressed in kW. Indexing candidate control actions. In some implementations, the index is set according to a preset number of candidate actions in [...]. , Multiple candidate power adjustment ratios are selected at equal intervals within the range, so that each candidate action corresponds to a specific power value within the current continuous power adjustment range, rather than a power level that is fixed in advance by the charging pile.

[0050] when When the actual power adjustment ratio range degenerates into a single usable value, the power adjustment ratio is used as the candidate power adjustment ratio of the current charging pile, and a candidate charging power control action is formed according to the above target charging power calculation relationship.

[0051] when When this occurs, it indicates that an effective power adjustment range cannot be formed between the current system configuration range and the allowable power range of the charging pile equipment, and no effective charging power candidate action will be generated in this control cycle. When there is a valid charging request and the charging pile is in a state that allows charging to start, a candidate control action is formed based on the current allowed target charging power range; when there is no valid charging request, no new charging power candidate action is formed.

[0052] When the charging station is equipped with a controllable energy storage system and step S1 can obtain an effective energy storage operating state, the candidate control action may also include an energy storage compensation action. In this case, a candidate control action may simultaneously include the target charging power, the corresponding energy storage compensation state, and the energy storage compensation power. The energy storage compensation power is within the currently allowed charging and discharging power range of the energy storage system and is limited by the energy storage state of charge, remaining available energy, and equipment operating boundaries. If no controllable energy storage system is configured or the energy storage system is currently unavailable, no energy storage compensation candidate control action will be formed.

[0053] This results in a set of continuous value candidate control actions:

[0054] in, For a set of candidate control actions with continuous values; For the first One candidate control action; The total number of candidate control actions generated in this control cycle is a non-negative integer; when hour, It is an empty set.

[0055] Furthermore, the current operating status of the charging station is synchronized to the electrical digital twin model of the charging station. This electrical digital twin model characterizes the electrical operating relationships of the AC power supply network and energy-consuming equipment at the charging station. It includes at least the electrical characteristics of the power supply lines from the low-voltage side of the station-level distribution transformer to the AC input terminals of each charging pile, the response relationship between the power of each charging pile and the effective value of the AC input voltage, and the current operating status of each charging pile. When a controllable energy storage system is configured, it also includes the charging and discharging operating status of the energy storage system.

[0056] The electrical parameters of the power supply line are obtained based on the power distribution design parameters of the charging station or the actual line parameters. The equipment parameters of the distribution transformer, charging piles, and controllable energy storage system are obtained based on the corresponding equipment technical parameters. Real-time voltage, current, power, and equipment operating status are synchronized according to the current operating status of the charging station. In some embodiments, the electrical digital twin model of the charging station uses power flow calculation or equivalent circuit calculation to obtain the voltage changes of each AC input monitoring node under different power conditions.

[0057] The voltage drop occurrence probability, predicted voltage drop amplitude, and predicted cumulative voltage drop duration formed in step S3 are used together as the benchmark voltage drop risk state within the future preset time window and input into the electrical digital twin model of the charging station. Under the combined effect of the current actual operating state and the future voltage drop risk state, the digital twin model applies each candidate control action and pre-simulates the change in the effective value of the AC input voltage of the charging pile after the execution of each candidate control action.

[0058] For each candidate control action in the candidate control action set, the electrical digital twin model of the charging station simulates the electrical operation process after the candidate control action is applied, and obtains the trajectory of the predicted effective value of the AC input voltage of the corresponding charging pile. For candidate control actions that include energy storage compensation, the impact of controllable energy storage on the AC side operation state of the charging station is also simulated according to the energy storage compensation state and energy storage compensation power corresponding to the candidate action.

[0059] Let the future preset time window mentioned in step S2 be denoted as The unit is seconds (s); the evaluation window for the control effect of the candidate control action is denoted as... The unit is seconds (s). In this embodiment, Take and The same applies, so that the voltage safety evaluation of the candidate control action covers the future preset time window corresponding to the current voltage drop prediction.

[0060] Regarding the first For each candidate control action, the voltage safety margin is calculated based on the predicted RMS voltage trajectory of the AC input terminal obtained from its digital twin simulation:

[0061] in For the first The effective voltage value of the AC input terminal of the target charging pile at time t, predicted by the electrical digital twin model of the charging station after the execution of each candidate control action, in V; This marks the start time of the evaluation for this control cycle; The voltage safety threshold for the monitoring node at the AC input terminal of the corresponding charging pile, in V; For the first The voltage safety margin corresponding to each candidate control action, in volts (V). and They use the same AC input terminal voltage RMS value measurement caliber and correspond to the same monitoring node.

[0062] when When, it indicates that within the control effect evaluation window, the lowest predicted effective value of the AC input voltage corresponding to the candidate control action is higher than the voltage safety threshold, thus satisfying the voltage safety condition; when If the voltage threshold is not met, it indicates that there is still a risk of the voltage being touched or falling below the voltage safety threshold after the candidate control action is executed.

[0063] After obtaining the voltage safety margin corresponding to each candidate control action, the equipment operation constraints of each candidate control action are simultaneously verified. The charging pile operation constraints include at least the minimum charging power, the maximum charging power, and the power change conditions allowed under the current operating conditions; when the candidate control action includes an energy storage compensation action, the energy storage operation constraints include at least the energy storage state of charge, available charging and discharging power, and remaining available energy.

[0064] Based on voltage safety conditions and equipment operating constraints, a set of safe and executable candidate actions is obtained from the continuous value candidate control action set:

[0065] in, A set of safe and executable candidate actions; This is the set of constraints for the operation of charging piles; This is the set of constraints for energy storage operation. When no controllable energy storage system is configured or the energy storage system is currently unavailable, it will not be included. As a screening criterion for candidate control actions.

[0066] For the candidate control actions that pass the screening, retain their corresponding digital twin pre-simulation results, including the predicted voltage RMS value trajectory, voltage safety margin, and equipment operation constraint verification results, and establish a correspondence with the corresponding candidate control actions.

[0067] When key line parameters, equipment parameters, or current operating status required for the electrical digital twin model of a charging station are missing, or when the digital twin calculation of a candidate control action fails and a valid predicted voltage effective value trajectory cannot be obtained, the candidate control action will not be judged as a safe and executable action. When all candidate control actions cannot obtain valid digital twin pre-simulation results, or when all candidate control actions do not meet voltage safety conditions or equipment operating constraints, the set of safe and executable candidate actions will be empty, forming a state with no safe and executable candidate actions, and the candidate control actions that have been eliminated will not be added back to the set of safe and executable candidate actions.

[0068] The set of safe executable candidate actions formed in this step, as well as the digital twin pre-simulation results corresponding to each safe executable candidate action, serve as the input for step S5.

[0069] S5, input the set of safe executable candidate actions and the digital twin pre-simulation results corresponding to each safe executable candidate action into the PPO decision model, determine the final control action from the set of safe executable candidate actions, adjust the charging power of the charging pile according to the final control action, and control the controllable energy storage to perform compensation when the final control action includes an energy storage compensation action; In step S5, the set of safe executable candidate actions formed in step S4 and the corresponding digital twin simulation results are input into the pre-trained PPO decision model. Combined with the current operating status of the charging station, the final control action is determined from the set of safe executable candidate actions. The set of safe executable candidate actions limits the range of actions that the PPO decision model can select within this control cycle. Candidate control actions eliminated in step S4 due to not meeting voltage safety conditions or equipment operating constraints do not participate in the strategy selection in this step.

[0070] Specifically, the current charging station operating status and the digital twin pre-simulation results corresponding to each safe and executable candidate action are used as the state information of the PPO decision model. The digital twin pre-simulation results include the predicted effective value of the AC input voltage of the charging pile after the execution of the corresponding candidate control action, the voltage safety margin, and the equipment operation constraint verification results. The vehicle's state of charge and the user's set charging preferences can be used as auxiliary state information to participate in strategy decision-making, but do not change the range of safe and executable candidate actions already determined in step S4. When a controllable energy storage system is configured, the current effective energy storage operating status is used as the state information corresponding to the corresponding energy storage compensation candidate action.

[0071] The safe and executable candidate actions correspond to specific action values ​​in the continuous action space. The PPO decision model evaluates each safe and executable candidate action based on the current state and determines the final control action according to the following formula:

[0072] Where n is the control cycle index, which is a positive integer; This represents the state information input to the PPO decision model during the nth control cycle; The parameter is The PPO policy network in state Below are candidate actions that can be safely executed. Evaluation of the given strategy; This indicates the final control action. The PPO decision model only applies when... The strategy selection is performed internally without changing the voltage safety screening and equipment operation constraint screening results formed in step S4.

[0073] When the set of safe and executable candidate actions contains only one candidate control action, that candidate control action is determined as the only allowed action in the current control cycle and is therefore the final control action, without the need for comparison between candidate actions. When multiple safe and executable candidate actions have the same highest policy evaluation, the first-ranked candidate action with the same highest policy evaluation is selected as the final control action, following the fixed candidate order used in step S4 when forming candidate control actions, to ensure the determinism of the decision result under the same input conditions.

[0074] The PPO decision model acquires the ability to evaluate safe and executable candidate actions under different operating conditions through pre-training. In some implementations, its training objective comprehensively considers charging continuity, charging efficiency, and AC input voltage stability. Therefore, for multiple candidate control actions that all meet the voltage safety conditions and equipment operating constraints of step S4, this step determines the final control action based on the PPO strategy evaluation, rather than solely using the maximum voltage safety margin as the final action selection condition.

[0075] When the final control action includes a charging power adjustment action, a charging power control command is generated based on the target charging power corresponding to the final control action, and this command is sent to the charging pile through the control interface between the edge controller and the corresponding charging pile. For continuous target charging power, the edge controller converts the target charging power into a power adjustment command that the charging pile can execute according to the control interface supported by the charging pile, and the charging pile adjusts the charging power according to its own allowed power change conditions. The actual execution process is still limited by the charging pile's own minimum charging power, maximum charging power, power change capability, and local protection conditions.

[0076] When a charging station is equipped with a controllable energy storage system and the final control action includes an energy storage compensation action, an energy storage control command is generated based on the energy storage compensation status and energy storage compensation power in the final control action and sent to the energy storage control device, so that the controllable energy storage system performs power compensation according to the energy storage control command; when the final control action does not include an energy storage compensation action or the charging station is not equipped with a controllable energy storage system, only the corresponding charging pile power control is executed.

[0077] In some implementations, when the final control action causes the charging power adjustment, charging power adjustment status information can also be sent to the user terminal. The user notification does not participate in the screening of safe executable candidate actions and the determination of the final control action.

[0078] When the set of safe executable candidate actions formed in step S4 is empty, this step does not re-determine the control action from the set of continuous value candidate control actions or the candidate control actions that have been eliminated, nor does it force the PPO decision model to generate actions that exceed the set of safe executable candidate actions. Instead, the charging pile will perform conservative processing according to the existing local power reduction, charging pause or undervoltage protection logic.

[0079] When the set of safe executable candidate actions is not empty, but the PPO decision model fails to load, the reasoning is abnormal, or the output result cannot correspond to the valid candidate control action in the set of safe executable candidate actions, the abnormal output is not executed, the abnormal output is not replaced by an action outside the set of safe executable candidate actions, and the current valid operating state is maintained or taken over by the existing local control and protection logic of the charging pile.

[0080] After issuing the control command, record the final control action, the actual control command issued, and the actual execution status returned by the control interface. If there is a difference between the target control action and the actual execution result fed back by the device, record the correspondence between the two, but do not use the difference for model updating in this step. The final control action and the actual execution status serve as input for step S6.

[0081] S6, collect actual operation feedback data after the final control action is executed; for the operating range where active control to change the target charging power is not executed and energy storage compensation is not executed, form voltage drop prediction model update data based on the corresponding correction voltage drop prediction result and actual operating state; and form PPO decision model interactive update data based on the operating state before control, the actual executed action and the operating state after execution, so as to update the corresponding models respectively.

[0082] In step S6, operational feedback data after the actual execution of the final control action in step S5 is collected. A time correspondence is established between this operational feedback data and the corrected voltage drop prediction result for the corresponding control cycle, the pre-control operating state, and the actual executed action, forming voltage drop prediction model update data and PPO decision model interactive update data, respectively. The actual operational feedback data includes at least the effective value of the AC input voltage of the charging pile after the final control action is executed, the input current, the actual charging power, and the actual voltage drop status; when an energy storage compensation action is actually executed, it also includes the actual compensation status and actual compensation power of the energy storage system.

[0083] Specifically, the actual action is determined based on the control interface return result recorded in step S5, and actual operation feedback data is collected according to the corresponding control cycle. When there is a deviation between the target control command corresponding to the final control action and the actual execution result of the equipment, the actual execution action of the equipment is used as the action object for the feedback data generated in this step and the interactive update data of the PPO decision model, so that the pre-control running state, the actual execution action and the post-execution running state remain consistent.

[0084] For operating intervals where no active control to change the target charging power is implemented and no energy storage compensation is performed, the actual voltage drop event label is determined based on the actual effective value of the AC input voltage within the future preset time window corresponding to the correction voltage drop prediction result:

[0085] in, This represents the actual effective voltage value at the AC input terminal of the corresponding charging pile at physical time t, in V. This is the actual voltage drop event label, a binary variable. It is used when the lowest effective value of the AC input voltage within the corresponding prediction time window reaches or falls below... hour, ;otherwise, The actual effective value of the voltage and They use the same AC voltage measurement caliber and correspond to the same AC input terminal monitoring node of the charging pile.

[0086] Furthermore, the actual voltage drop amplitude is determined based on the actual effective value of the AC input voltage within the same prediction time window:

[0087] in, This is a reference voltage used to characterize the voltage drop at the AC input terminal, measured in volts (V). It is set according to the rated AC input voltage of the corresponding charging station or a predetermined normal operating reference voltage. This represents the actual pressure drop, expressed in V. , as well as The same voltage RMS measurement caliber and the same monitoring object are used.

[0088] The actual cumulative duration of pressure drop is determined by the following formula:

[0089] in, The actual cumulative pressure drop duration, in seconds; For the first time within the prediction time window One valid sampling time, For sampling point index; The number of valid sampling points within the prediction time window is a positive integer. For the first The sampling time interval corresponding to each valid sampling point is expressed in seconds. This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. When multiple separate voltage drop time segments exist within the same prediction time window, the duration of each voltage drop time segment is cumulatively included. This makes it have the same time meaning as the predicted cumulative duration of pressure drop in step S3.

[0090] The actual pressure drop event label, actual pressure drop amplitude, and actual pressure drop cumulative duration are respectively associated with the pressure drop occurrence probability, predicted pressure drop amplitude, and predicted pressure drop cumulative duration of the corresponding control cycle to form pressure drop prediction model update data, which is used for incremental training or periodic updates of the pressure drop prediction model.

[0091] For operating intervals where active control or energy storage compensation to change the target charging power was actually implemented within the corresponding prediction time window, a control intervention marker is added to the corresponding data. Since control actions alter the actual voltage evolution state after control execution, the absence of a voltage drop after control execution is not directly used as the basis for determining a voltage drop prediction error before control execution, nor is this result directly used as a monitoring label to update the voltage drop prediction model under uncontrolled conditions. The data subject to control intervention retains its corrected voltage drop prediction result, actual executed action, and actual operational feedback data after execution, and is stored separately from the data used for direct monitoring and updating of the voltage drop prediction model.

[0092] At the same time, according to the first The pre-control operating state, actual executed actions, and post-execution operating state of each control cycle form interactively updated data for the PPO decision model. This is the control cycle index, a positive integer. The index is determined based on the actual control effect. Reward value corresponding to each control cycle:

[0093] in, For the first The reward value for each control cycle is a dimensionless quantity. Score for charging continuity. Score the charging efficiency. The three values ​​are all normalized dimensionless quantities and are used to score the AC input voltage stability. , and The weights for charging continuity, charging efficiency, and AC input voltage stability are respectively, all of which are non-negative dimensionless quantities and satisfy the following:

[0094] When the If charging is not interrupted due to low voltage on the AC input side of the charging pile within a control cycle. Set to 1; when charging is interrupted due to the low voltage, Take 0.

[0095] The charging efficiency score is determined by the following formula:

[0096] in, For the first The actual electrical energy supplied to the vehicle within each control cycle is measured in kWh and is obtained based on the actual electrical energy metering data of the corresponding charging pile. For the first The electrical energy requested in each control cycle is in kWh and is determined based on the valid vehicle charging request at the start of the control cycle, or based on the corresponding requested charging power and the duration of the control cycle. This is a dimensionless quantity between 0 and 1. The above charging efficiency score calculation applies to... The effective charging cycle; when At that time, the control cycle is not used to form interactive update samples for the PPO decision model that includes charging efficiency evaluation.

[0097] The AC input voltage stability score is determined according to the following formula:

[0098] in, For the first The lowest effective value of the AC input voltage of the charging pile obtained by actual measurement within each control cycle, in V; Indicates when Output 0 when Time output ,when Output 1 when the time is right. The calculation is applicable to... The parameter configuration, and , and They use the same AC input terminal voltage RMS value measurement caliber and correspond to the same monitoring object.

[0099] when ,or , and When the same AC input voltage RMS value measurement caliber is not used or the same monitoring object is not used, the data of this control cycle is not used to calculate the AC input voltage stability score, nor is the PPO decision model interactive update sample containing the reward value formed using this control cycle.

[0100] In order to obtain effective , and And generate reward value At that time, based on the first The pre-control running state, actual executed actions, reward values, and post-execution running state of each control cycle form corresponding PPO decision model interactive update samples, which are used for incremental learning or periodic policy updates of the PPO decision model, enabling the PPO decision model to adjust the policy evaluation in subsequent control cycles based on the operating effects produced by the actual control actions.

[0101] Furthermore, the resulting voltage drop prediction model update data is used to update the Transformer-LSTM hybrid prediction model and the edge prediction model according to a preset update cycle. Specifically, the cloud side uses the accumulated effective voltage drop prediction model update data for iterative model training, while the edge side can use effective data corresponding to the current charging station's operating characteristics for fine-tuning model parameters. The PPO decision model updates its strategy based on the effective PPO decision model interaction update data. The updated model is used in subsequent control cycles after passing a preset validity verification. If the model training fails to converge, the verification results do not meet preset requirements, or the updated model file is abnormal, the currently verified and usable model version is maintained, and the abnormal model is not used to replace the currently running model.

[0102] When key AC input voltage RMS values, input currents, or actual execution states are continuously missing from the actual operational feedback data, making it impossible to establish an effective correspondence between the corrected voltage drop prediction results, actual execution actions, and actual operational results, the incomplete data is not used to form new model update samples. Model update data formation resumes only after complete and time-corresponding actual operational feedback data is obtained. This process generates voltage drop prediction model update data and PPO decision model interactive update data, which are then used in subsequent control cycles, thus completing the closed loop between voltage drop prediction, control execution, and actual operational feedback.

[0103] Example 2 An artificial intelligence-based charging pile voltage drop prediction and adaptive control system includes: The data acquisition and preprocessing module is used to collect the effective value of the AC input voltage of the charging pile, current, charging power, station load, operating status of other charging piles, historical charging records, and environmental and time information, and to form a standardized multi-source feature dataset, edge real-time electrical data, and current charging station operating status. The cloud-based pressure drop trend prediction module is used to perform global correlation and temporal evolution modeling of the standardized multi-source feature dataset to form a priori cloud-based pressure drop trend. The edge real-time correction module is used to generate a corrected voltage drop prediction result based on the cloud voltage drop trend prior, edge real-time electrical data and the current charging station operating status, including the probability of voltage drop occurrence, the predicted voltage drop amplitude and the predicted cumulative duration of voltage drop. The candidate action formation and digital twin pre-simulation screening module is used to generate multiple consecutive candidate control actions based on the current operating status of the charging station and the current allowable power adjustment range of the charging pile. The module inputs the corrected voltage drop prediction result, the current operating status of the charging station, and each candidate control action into the electrical digital twin model of the charging station. It then predicts the effective value of the AC input voltage of the charging pile after each candidate control action is executed. Based on voltage safety conditions and charging pile operating constraints, it screens and forms a set of safe and executable candidate actions. When the charging station is equipped with controllable energy storage, the candidate control actions also include energy storage compensation actions, and the screening also considers energy storage operating constraints. The PPO decision and control execution module is used to input the set of safe executable candidate actions and the digital twin pre-simulation results corresponding to each safe executable candidate action into the PPO decision model, determine the final control action from the set of safe executable candidate actions, adjust the charging power of the charging pile according to the final control action, and control the controllable energy storage to perform compensation when the final control action includes an energy storage compensation action. The effect feedback and model update module is used to generate pressure drop prediction model update data and PPO decision model interactive update data based on the actual operation feedback data after the final control action is executed.

[0104] The cloud-based voltage drop trend prediction module is deployed in the cloud, and the edge real-time correction module is deployed on the charging station edge controller. The cloud transmits the cloud-based voltage drop trend prior to the edge controller. The edge controller combines the edge real-time electrical data and the current operating status of the charging station to form the corrected voltage drop prediction result, and sends control commands to the charging pile. When a controllable energy storage system is configured and the final control action includes an energy storage compensation action, the controller sends an energy storage control command to the energy storage control device.

[0105] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for predicting and adaptively controlling voltage drop in charging piles based on artificial intelligence, characterized in that, Includes the following steps: S1 collects the effective value of AC input voltage, current, charging power, station load, operating status of other charging piles, historical charging records, and environmental and time information of the charging piles. The collected data is preprocessed to form a standardized multi-source feature dataset, edge real-time electrical data, and the current operating status of the charging station. S2, input the standardized multi-source feature dataset into the cloud Transformer-LSTM hybrid prediction model, establish global correlation between multi-source features based on Transformer, and perform temporal evolution modeling on the correlated features based on LSTM to form a priori cloud pressure drop trend within a future preset time window. S3, input the prior information of the cloud voltage drop trend, the real-time electrical data of the edge, and the current operating status of the charging station into the edge prediction model for short-time scale correction, and obtain the corrected voltage drop prediction result including the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop; S4. Based on the current operating status of the charging station and the current allowable power adjustment range of the charging pile, multiple candidate control actions with consecutive values ​​are formed. The candidate control actions include at least a charging power adjustment action. The correction voltage drop prediction result, the current operating status of the charging station, and each candidate control action are input into the electrical digital twin model of the charging station. The effective value of the AC input voltage of the charging pile after the execution of each candidate control action is predicted. A set of safe and executable candidate actions is formed by screening based on voltage safety conditions and charging pile operating constraints. When the charging station is equipped with controllable energy storage, the candidate control actions also include energy storage compensation actions. The screening also considers energy storage operating constraints. S5, input the set of safe executable candidate actions and the digital twin pre-simulation results corresponding to each safe executable candidate action into the PPO decision model, determine the final control action from the set of safe executable candidate actions, adjust the charging power of the charging pile according to the final control action, and control the controllable energy storage to perform compensation when the final control action includes an energy storage compensation action; S6, collect actual operation feedback data after the final control action is executed; for the operating range where active control to change the target charging power is not executed and energy storage compensation is not executed, form voltage drop prediction model update data based on the corresponding correction voltage drop prediction result and actual operating state; and form PPO decision model interactive update data based on the operating state before control, the actual executed action and the operating state after execution, so as to update the corresponding models respectively.

2. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 1, characterized in that, The cloud-based Transformer-LSTM hybrid prediction model first establishes a global correlation between multi-source features through Transformer, and then performs temporal evolution modeling on the correlated features through LSTM; the edge prediction model uses the cloud-based voltage drop trend as prior information, and combines real-time electrical data at the edge with the current operating status of the charging station for short-time scale correction, outputting the probability of voltage drop occurrence, the predicted voltage drop amplitude, and the predicted cumulative duration of voltage drop.

3. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 1, characterized in that, The electrical digital twin model of the charging station is used to characterize the electrical characteristics of the power supply line from the low-voltage side of the station-level distribution transformer to the AC input terminal of each charging pile, the response relationship between the power of each charging pile and the effective value of the AC input terminal voltage, and the current operating status of each charging pile. The corrected voltage drop prediction result is used as the benchmark voltage drop risk state within a future preset time window, and each candidate control action is applied to obtain the trajectory of the predicted effective value of the AC input terminal voltage of the charging pile after the execution of each candidate control action.

4. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 1, characterized in that, The formation of the continuous value candidate control action includes: determining the actual power adjustment ratio range based on the power adjustment range configured by the system and the current allowed charging power range of the charging pile; selecting multiple candidate power adjustment ratios within the actual power adjustment ratio range; and determining the corresponding candidate target charging power based on the current charging power.

5. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 3, characterized in that, The voltage safety margin is determined based on the lowest predicted voltage effective value in the AC input terminal predicted voltage effective value trajectory corresponding to each candidate control action and the voltage safety threshold. Candidate control actions with a voltage safety margin greater than zero and that meet the charging pile operation constraints are retained in the set of safe executable candidate actions.

6. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 3, characterized in that, When a controllable energy storage system is configured in a charging station, the candidate control actions also include energy storage compensation state and energy storage compensation power. The energy storage compensation power is limited by the energy storage state of charge, available charging and discharging power, and remaining available energy. The electrical digital twin model of the charging station also simulates the impact of energy storage compensation on the effective value of the AC input voltage of the charging pile and filters it according to the energy storage operation constraints.

7. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 5, characterized in that, The PPO decision model only evaluates the policy of candidate control actions in the set of safe and executable candidate actions; when the set of safe and executable candidate actions contains only one candidate control action, it is directly determined as the final control action; when multiple candidate control actions have the same highest policy evaluation, the final control action is determined according to a fixed candidate order.

8. The artificial intelligence-based charging pile voltage drop prediction and adaptive control method according to claim 1, characterized in that, For operating ranges where no active control to change the target charging power is executed and no energy storage compensation is executed, the voltage drop prediction model updates data based on the actual voltage drop events, actual voltage drop amplitude, and actual voltage drop cumulative duration; for operating ranges where the active control or energy storage compensation is executed, a control intervention flag is added; and the PPO decision model interactively updates data based on the operating state before control, the actual executed action, the reward value, and the operating state after execution.

9. A charging pile voltage drop prediction and adaptive control system based on artificial intelligence, characterized in that, include: The data acquisition and preprocessing module is used to collect the effective value of the AC input voltage of the charging pile, current, charging power, station load, operating status of other charging piles, historical charging records, and environmental and time information, and to form a standardized multi-source feature dataset, edge real-time electrical data, and current charging station operating status. The cloud-based pressure drop trend prediction module is used to perform global correlation and temporal evolution modeling of the standardized multi-source feature dataset to form a priori cloud-based pressure drop trend. The edge real-time correction module is used to generate a corrected voltage drop prediction result based on the cloud voltage drop trend prior, edge real-time electrical data and the current charging station operating status, including the probability of voltage drop occurrence, the predicted voltage drop amplitude and the predicted cumulative duration of voltage drop. The candidate action formation and digital twin pre-simulation screening module is used to generate multiple consecutive candidate control actions based on the current operating status of the charging station and the current allowable power adjustment range of the charging pile. The module inputs the corrected voltage drop prediction result, the current operating status of the charging station, and each candidate control action into the electrical digital twin model of the charging station. It then predicts the effective value of the AC input voltage of the charging pile after each candidate control action is executed. Based on voltage safety conditions and charging pile operating constraints, it screens and forms a set of safe and executable candidate actions. When the charging station is equipped with controllable energy storage, the candidate control actions also include energy storage compensation actions, and the screening also considers energy storage operating constraints. The PPO decision and control execution module is used to input the set of safe executable candidate actions and the digital twin pre-simulation results corresponding to each safe executable candidate action into the PPO decision model, determine the final control action from the set of safe executable candidate actions, adjust the charging power of the charging pile according to the final control action, and control the controllable energy storage to perform compensation when the final control action includes an energy storage compensation action. The effect feedback and model update module is used to generate pressure drop prediction model update data and PPO decision model interactive update data based on the actual operation feedback data after the final control action is executed.

10. The artificial intelligence-based charging pile voltage drop prediction and adaptive control system according to claim 9, characterized in that, The cloud-based voltage drop trend prediction module is deployed in the cloud, and the edge real-time correction module is deployed on the charging station edge controller. The cloud transmits the cloud-based voltage drop trend prior to the edge controller. The edge controller combines the edge real-time electrical data and the current operating status of the charging station to form the corrected voltage drop prediction result, and sends control commands to the charging pile. When a controllable energy storage system is configured and the final control action includes an energy storage compensation action, the controller sends an energy storage control command to the energy storage control device.