Electric vehicle charging load estimation and charging mode optimization method

Abstract

The application relates to the technical field of power system optimization, and discloses a method for estimating charging load of electric vehicles and optimizing charging modes. The method initializes a source-load-storage-mid-voltage distribution network collaborative planning framework, integrates charging station geographical distribution and user charging behavior modes; real-time charging load sequence data is collected, and a fluctuation analysis algorithm is used to judge the charging load control state; when the charging load control is in a non-stable state, voltage fluctuation characteristics and charging power change characteristics of the distribution network are synchronously analyzed, and a charging efficiency evaluation value is generated; based on the evaluation value, the charging effect is divided into three levels of high efficiency, normal efficiency and low efficiency; for the normal efficiency charging effect, a time series decomposition technology is used to extract a load trend component, and an overall load stability index is calculated; the charging efficiency evaluation value and the overall load stability index are fused, and a charging power adjustment instruction is output. The method effectively improves the operation stability and charging efficiency of the distribution network.

Description

Technical Field

[0001] This invention relates to the field of power system optimization technology, specifically to a method for estimating electric vehicle charging load and optimizing charging methods. Background Technology

[0002] Against the backdrop of a global push for sustainable development, electric vehicles (EVs) have entered a golden age of development due to their significant advantages in energy conservation and environmental protection. In recent years, the global EV fleet has experienced rapid growth. Data from the International Energy Agency shows that by 2023, the global EV fleet had reached nearly 42 million vehicles, and this number continued to climb in 2024. China, as a major player in the global EV market, has vigorously promoted the development of the EV industry, making efforts in policy support, technological innovation, and other areas, resulting in a continuous increase in the number of EVs on the road.

[0003] my country has introduced a series of unprecedentedly strong policies to safeguard the development of the new energy vehicle industry. From early purchase subsidies to the later dual-credit policy, and now to the charging infrastructure construction plan, the continuous guidance and support of these policies have created a favorable environment for the development of the new energy vehicle industry. Although purchase subsidies are gradually being phased out, support continues for battery swapping models and commercial vehicle battery recycling, promoting the comprehensive development of the new energy vehicle industry. Regarding charging infrastructure construction, China is vigorously accelerating the construction of supercharging networks and supporting the deployment of battery swapping models in smart grids, committed to solving the charging challenges of new energy vehicles. The implementation of these policies has greatly stimulated consumer enthusiasm for electric vehicles and promoted their widespread adoption.

[0004] Despite the promising future of electric vehicles, charging infrastructure remains a bottleneck hindering their further development. Currently, the shortage and uneven distribution of charging facilities are significant issues. In cities, charging facilities are relatively concentrated in bustling commercial areas and near large shopping malls, while they are severely lacking in older residential areas and peri-urban areas. For example, in a first-tier city, charging stations are densely packed in the city center, with perhaps 3-5 stations per square kilometer, while in the city's outskirts, there might only be one station every tens of square kilometers. This forces some electric vehicle users to specifically plan routes to find charging facilities, increasing their time and effort. In rural and remote areas, charging facilities are even scarcer, severely limiting the usability of electric vehicles and significantly impacting consumer willingness to purchase them.

[0005] Long charging times are a major problem for electric vehicle users. Generally, it takes 6-8 hours to fully charge an electric vehicle at a regular charging station, and even fast charging takes about half an hour to an hour, a huge difference compared to the few minutes it takes to refuel a traditional gasoline car. For users with urgent travel needs, the long charging time is undoubtedly a significant inconvenience, reducing the convenience of using electric vehicles. Moreover, high charging costs are also a factor that consumers must consider, increasing the cost of ownership and weakening the market competitiveness of electric vehicles.

[0006] Large-scale, unregulated charging of electric vehicles poses a serious threat to the stable and safe operation of the power grid. Electric vehicle charging loads are random and fluctuating; when a large number of electric vehicles charge simultaneously, the grid load increases dramatically, potentially causing voltage fluctuations, frequency instability, and other problems. This can affect the normal operation of the power grid and may even lead to grid failures, posing a significant challenge to the power system.

[0007] Given the numerous challenges associated with electric vehicle charging, accurately estimating charging load and optimizing charging methods are crucial. However, existing charging load estimation methods have significant limitations. Currently, common charging load estimation methods mainly include statistical prediction based on historical data and machine learning prediction methods.

[0008] Historical data-based statistical forecasting methods primarily rely on past electric vehicle charging data, such as charging time and charging volume, to predict future charging load using statistical analysis. This method assumes that future charging behavior patterns will be similar to the past. However, in reality, electric vehicle users' travel habits and charging needs are influenced by various factors, such as differences between weekdays and holidays, weather changes, and different travel purposes. These factors introduce significant uncertainty into charging behavior, making historical data-based statistical forecasting methods difficult to accurately predict future charging load.

[0009] While machine learning prediction methods utilize advanced algorithms and can handle massive amounts of data, they also face challenges. They require large amounts of accurate and comprehensive data for model training, but in practice, data gaps and errors often occur during data acquisition. Furthermore, the training process for machine learning models is complex, demanding significant computational resources, and the model's accuracy and generalization ability are constrained by various factors. For example, when encountering new scenarios or changes in data distribution, the model may fail to accurately predict charging load and struggle to precisely account for the dynamic changes in complex factors such as user behavior, grid conditions, and market environment.

[0010] Regarding charging methods, while current conventional charging methods cause less damage to the battery, their extremely slow charging speeds fail to meet users' demands for fast charging, significantly reducing the efficiency of electric vehicles. Fast charging, on the other hand, can replenish a large amount of energy to an electric vehicle in a short time, but the high current and power during fast charging generate high temperatures inside the battery, accelerating battery aging and shortening its lifespan. Furthermore, fast charging has a significant impact on the power grid, potentially causing voltage fluctuations and increased harmonics, affecting the normal operation of other equipment on the grid.

[0011] While battery swapping, as an emerging charging method, can alleviate the problem of long charging times to some extent, it still faces numerous challenges. One of the most prominent issues is the lack of standardized battery specifications. Batteries from different manufacturers vary in specifications, capacity, and interfaces, making it difficult to achieve standardized operation of swapping stations and increasing construction and operating costs. Furthermore, battery swapping requires a large battery storage facility and an efficient battery management system, which further increases operating costs and limits its large-scale adoption. Summary of the Invention

[0012] The purpose of this invention is to provide a method for estimating electric vehicle charging load and optimizing charging methods to solve the problems mentioned in the background art.

[0013] To achieve the above objectives, the present invention provides a method for estimating electric vehicle charging load and optimizing charging methods, the method comprising:

[0014] Initialize the source-load-storage-medium voltage distribution network collaborative planning framework, integrating the geographical distribution of charging stations and user charging behavior patterns;

[0015] Real-time acquisition of charging load sequence data; determination of charging load control status through fluctuation analysis algorithm.

[0016] When the charging load control is in an unstable state, the voltage fluctuation characteristics of the distribution network and the charging power change characteristics are analyzed simultaneously to generate a charging efficiency evaluation value.

[0017] Based on the charging efficiency evaluation value, the charging effect is divided into high efficiency level, normal level and low efficiency level.

[0018] To assess the charging performance at standard levels, time series decomposition techniques are used to extract load trend components and calculate overall load stability indicators.

[0019] It integrates charging efficiency assessment values ​​and overall load stability indicators to output charging power adjustment commands.

[0020] Preferably, the initialization source-load-storage-medium voltage distribution network collaborative planning framework includes constructing a diversified user resource classification system, which covers industrial user load curves, commercial user charging preferences, and residential user electricity consumption habits, and integrates renewable energy output data with charging pile configuration parameters to form a multi-timescale interactive dataset; the multi-timescale interactive dataset generates a basic distribution map of charging load through a spatial interpolation algorithm for subsequent real-time data collection.

[0021] Preferably, the real-time acquisition of charging load sequence data specifically includes deploying a smart meter cluster and a charging pile monitoring terminal to obtain voltage sampling values, current sampling values, and charging timestamps at a fixed sampling frequency; removing outliers and imputing missing values ​​from the original sampling data; and generating a normalized load sequence using standardization processing; the normalized load sequence is then transmitted as input to a volatility analysis algorithm.

[0022] Preferably, the step of determining the charging load control state through the volatility analysis algorithm includes using a sliding window variance calculation method to divide the normalized load sequence into continuous time windows and calculate the variance of the load value in each window; if the variance of three consecutive time windows exceeds the adaptive threshold, the charging load control is determined to be in an unstable state, and the distribution network voltage fluctuation characteristic analysis is triggered; the adaptive threshold is dynamically updated through the statistical quantiles of historical stable load sequences.

[0023] Preferably, the synchronous analysis of distribution network voltage fluctuation characteristics and charging power change features includes extracting the frequency domain energy distribution of the voltage fluctuation signal and the first-order difference sequence of charging power, merging the frequency domain energy distribution and the difference sequence into a multi-dimensional feature vector using feature fusion technology, inputting the multi-dimensional feature vector into a pre-trained classification neural network model, and outputting a charging efficiency evaluation value; the classification neural network model is trained offline using the backpropagation algorithm, and the training data includes labeled voltage fluctuation and power change samples.

[0024] Preferably, the specific method for classifying the charging effect into high-efficiency, normal, and low-efficiency levels based on the charging efficiency evaluation value is as follows: set a high-efficiency level threshold and a low-efficiency level threshold; if the charging efficiency evaluation value is greater than the high-efficiency level threshold, it is classified as high-efficiency; if the charging efficiency evaluation value is less than the low-efficiency level threshold, it is classified as low-efficiency; if the charging efficiency evaluation value is between the high-efficiency level threshold and the low-efficiency level threshold, it is classified as normal; for normal-level charging effects, time series decomposition technology is automatically activated.

[0025] Preferably, the step of extracting load trend components using time series decomposition technology includes applying an empirical mode decomposition algorithm to decompose the load sequence into intrinsic mode functions and residual terms, selecting low-frequency intrinsic mode functions as trend components; calculating the approximate entropy value of the trend components as an overall load stability index; and using the overall load stability index for subsequent intelligent optimization algorithms.

[0026] Preferably, the calculation of the overall load stability index includes comparing it with a stability reference threshold; if the overall load stability index is greater than the stability reference threshold, a load instability flag is generated; otherwise, a load stability flag is generated; the load instability flag or the load stability flag participates in the decision-making process of the intelligent optimization algorithm.

[0027] Preferably, the output charging power adjustment command includes designing a differentiated electricity price incentive model, generating an initial charging power scheme based on the user response probability curve; iteratively adjusting the charging power scheme using a particle swarm optimization algorithm, with the minimization of the distribution network voltage deviation as the objective function, and outputting an optimal charging power command sequence after convergence; the optimal charging power command sequence is directly sent to the charging pile actuator.

[0028] Preferably, the particle swarm optimization algorithm iterative process combines cooperative game theory to establish a multi-stakeholder interest allocation model and quantify the weights of user satisfaction and power grid stability; Shapley value calculation is introduced when updating particle positions in each iteration to ensure Pareto optimal solution; the finally generated charging power adjustment command is integrated into the source-load-storage-medium voltage distribution network collaborative planning framework to form a closed-loop optimization system.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This patented technology provides strong support for the stable operation of the power grid through precise control of charging load and optimization of charging methods. In traditional electric vehicle charging modes, the power grid often faces significant impacts due to the randomness and concentration of charging behavior. Especially in densely populated residential areas, the simultaneous charging of a large number of electric vehicles at night can cause a sudden and sharp increase in the load on the distribution network, potentially exceeding its capacity and leading to voltage fluctuations, frequency instability, and even grid failures in severe cases. However, this patented technology, when initializing the source-load-storage-medium-voltage distribution network collaborative planning framework, fully considers the geographical distribution of charging stations and user charging behavior patterns, enabling reasonable planning and arrangement of charging load in advance.

[0031] By collecting real-time charging load sequence data and employing volatility analysis algorithms to accurately determine the charging load control status, rapid measures are taken once an unstable charging load is detected. Simultaneously, the voltage fluctuation characteristics of the distribution network and the charging power variation features are analyzed to generate a charging efficiency assessment value, which is then used to classify the charging effect into different levels. For different levels of charging effect, especially when the charging load is unstable, precise charging power adjustment commands can be output to rationally regulate the charging power. This effectively avoids the impact on the power grid caused by a large number of electric vehicles charging simultaneously, ensuring stable operation of the power grid even during peak charging periods, guaranteeing the normal operation of other electrical equipment in the power system, and improving the reliability and stability of the entire power grid.

[0032] This patented technology offers significant advantages in energy efficiency. It cleverly and rationally schedules charging time and power based on grid load and electric vehicle charging needs, achieving peak shaving and valley filling. During periods of low grid load, it encourages electric vehicles to charge, storing excess energy; while during periods of high grid load, it appropriately reduces charging power or postpones charging time, minimizing pressure on the grid. In this way, it effectively balances the grid load, improves grid operating efficiency, and reduces energy waste caused by grid load fluctuations.

[0033] This technology also promotes the close integration of electric vehicle charging and renewable energy generation. When there is a surplus of renewable energy, such as abundant solar power during sunny days or strong winds during periods of high wind, this excess electricity can be used to charge electric vehicles, storing the energy in their batteries. Conversely, during peak electricity demand periods, electric vehicles can release the stored energy back into the grid to meet the electricity needs of other users. This two-way energy interaction achieves efficient energy utilization, increases the proportion of renewable energy consumption, reduces dependence on traditional fossil fuels, and promotes the optimization of the energy structure and sustainable development. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the working principle of the electric vehicle charging load estimation and charging mode optimization method described in this invention.

[0035] Figure 2 A flowchart for initializing the source-load-storage-medium voltage distribution network collaborative planning framework;

[0036] Figure 3 A flowchart for real-time acquisition and processing of charging load sequence data;

[0037] Figure 4 A graph showing the characteristics of voltage fluctuations and charging power;

[0038] Figure 5This is a diagram showing the empirical mode decomposition and stability analysis of the load sequence. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 This invention provides a method for estimating electric vehicle charging load and optimizing charging methods. The method includes: intelligent management of charging load by initializing a source-load-storage-medium-voltage distribution network collaborative planning framework, which integrates charging station geographical distribution information and user charging behavior pattern data. After real-time acquisition of charging load sequence data, a volatility analysis algorithm is used to determine the charging load control state. When the charging load control is in an unstable state, the voltage fluctuation characteristics of the distribution network and the charging power change characteristics are analyzed simultaneously to generate a charging efficiency evaluation value. The charging efficiency evaluation value is used to classify the charging effect into high-efficiency, normal, and low-efficiency levels. For the normal level charging effect, time series decomposition technology is used to extract the load trend component and calculate the overall load stability index. The charging efficiency evaluation value and the overall load stability index are fused and processed to finally output a charging power adjustment command to optimize the charging method and improve the stability of the distribution network.

[0041] Example 1: See Figure 2The process of initializing the source-load-storage-medium-voltage distribution network collaborative planning framework begins with constructing a comprehensive and structured diversified user resource classification system. This system is the cornerstone of all subsequent data integration and analysis. The design of this diversified user resource classification system needs to cover characteristic data of three main user types: industrial user load curve data points, commercial user charging preference parameters, and residential user electricity consumption habit profiles. Industrial user load curve data points are directly extracted from the historical database of the distribution automation system. These data points record power values ​​at one-minute intervals, forming a high-precision load change trajectory that accurately reflects the start-up and shutdown patterns of large electrical equipment and the power consumption patterns brought about by periodic production activities. The acquisition of commercial user charging preference parameters adopts a hybrid approach, combining a sampling questionnaire survey of electric vehicle owners in commercial areas with a real-time monitoring system deployed at commercial charging stations. The questionnaire survey collected information such as users' usual charging time periods, charging frequency, and sensitivity to charging prices. The real-time monitoring system continuously records the start time, duration, power consumption, and charging power level of each charging event. These data are cross-validated and integrated with the questionnaire results to generate a statistically representative set of commercial user charging behavior parameters. The construction of residential users' electricity consumption habits profiles relies on a smart meter network covering a wide area. Smart meters collect electricity consumption data at fixed intervals. By processing massive amounts of historical electricity consumption data through cluster analysis algorithms, different electricity consumption pattern categories are identified, such as morning peak type, evening peak type, balanced type, or night-dominant type. Each type corresponds to a detailed electricity consumption habit profile, which includes typical daily load curves, electricity consumption differences between weekends and weekdays, and seasonal variation patterns.

[0042] While establishing a diversified user resource classification system, the system needs to integrate output data from renewable energy power generation units and configuration parameters of charging infrastructure. Renewable energy output data is acquired in real-time from photovoltaic power plant monitoring systems, wind farm central controllers, and distributed energy storage system management platforms via data interfaces. This data includes not only real-time active and reactive power output values ​​but also ultra-short-term power forecast data, statistical analysis of the deviation between forecasts and actual output, and equipment operating status information. Charging pile configuration parameters are downloaded from the charging pile operation management platform. The parameter set covers each charging pile's unique identifier, geographical coordinates, interface type, rated output voltage and current range, communication protocol version, maximum supported charging power, and the total capacity limit of its respective site. All these heterogeneous data sources are aggregated into a unified data management module. This module organizes and indexes the data according to different time scales, forming a multi-time-scale interactive dataset. The multi-time-scale interactive dataset clearly distinguishes between hourly, daily, and monthly data granularities. Hourly data is mainly used for real-time monitoring and short-term control; daily data is used to analyze intraday variation patterns and formulate daily plans; and monthly data serves for long-term trend analysis and planning optimization. Each data entry is accompanied by a precise timestamp and spatial location tag, ensuring the traceability and relevance of the data in both time and space.

[0043] The application of spatial interpolation algorithms is a crucial step in transforming multi-timescale interactive datasets into spatial visualizations. Its purpose is to convert discrete monitoring point data distributed across the power distribution network coverage area into a continuous spatial distribution map. The Kriging interpolation method is chosen as the algorithm. Kriging interpolation is a spatial prediction technique based on statistical principles. It constructs a semi-variogram model by calculating the spatial autocorrelation between known sample points, and uses this model to provide the optimal unbiased estimate of values ​​at unknown locations. The processing first maps spatial coordinate information such as industrial user load curve data points, commercial user charging preference parameters, residential user electricity consumption profiles, and charging pile configuration parameters onto an electronic map, assigning each point a corresponding attribute value. Then, based on the attribute values ​​and spatial relationships of these discrete points, the Kriging interpolation method calculates the predicted value for each grid cell within the entire area. Simultaneously, it generates an accompanying prediction error variance plot to represent the uncertainty of the interpolation results. The final output is a basic charging load distribution map, stored in raster data format, where each grid cell value represents the estimated charging load intensity or potential demand density of the geographical area. The basic distribution map of charging load not only reflects the spatial distribution of load at the current moment, but also contains the characteristics of load distribution changes at different time scales because it is generated based on multi-time-scale interactive datasets. For example, the high load in commercial areas during the daytime on weekdays and the high load in residential areas at night.

[0044] After the basic charging load distribution map is generated, it is loaded into the system's spatial database for hosting, serving as a benchmark reference map for data matching and spatial registration in subsequent real-time data acquisition stages. When the real-time data acquisition module starts working, the real-time data streams uploaded from the smart meter cluster and charging pile monitoring terminals contain the geographical coordinate information of the devices. The system overlays and analyzes the coordinates of the real-time data points with the grid of the basic charging load distribution map to determine the grid cell in which each real-time data point falls, and compares and correlates the real-time measured values ​​with the basic predicted values ​​of the grid cells. For example, a sudden increase in the power of a charging pile occurring in a grid cell with a already high basic predicted load may be far more significant than a similar event occurring in a grid cell with a low predicted load. The diversified user resource classification system itself is not static; it is designed with a dynamic update mechanism to adapt to the slow evolution of electricity user behavior patterns and the onboarding of new users. The update mechanism employs an incremental learning algorithm. When a new industrial user connects, a commercial user's charging record is updated, or a residential user installs a new smart meter, the system incorporates the new data samples into the calculation, re-runs the clustering analysis or parameter fitting process, and adjusts the central feature vectors or parameter distributions of each category in the diversified user resource classification system. This ensures that the classification system accurately reflects the actual situation of users within the current distribution network service area. The storage architecture of the multi-timescale interactive dataset uses a specially optimized time-series database. This database is optimized for efficient writing, compression, and fast range queries of time-series data, supporting second-level retrieval and complex aggregation analysis of massive historical data. It provides stable and efficient data access services for the real-time data acquisition module, ensuring a smooth transition from the initialization framework to real-time operation. The entire initialization process, from the construction of the diversified user resource classification system to the generation of the charging load basic distribution map, constitutes a complete data preparation stage, providing an indispensable static data foundation and spatial context information for subsequent real-time monitoring, status judgment, and optimized control.

[0045] Example 2: See Figure 3Real-time acquisition of charging load sequence data relies on a layered sensor network architecture. Smart meter clusters are centrally installed at key nodes of the distribution network, such as the outlet of a 10 kV feeder and the low-voltage side of the distribution transformer. Charging pile monitoring terminals are directly integrated into the control unit of each AC or DC charging pile. The smart meter clusters use high-precision metering chips to synchronously acquire three-phase voltage and current sampling values ​​at a fixed sampling frequency of 1000 times per second. Each voltage and current sampling value is accompanied by a charging timestamp accurate to milliseconds, generated by a high-stability clock chip. The charging timestamp not only marks the time when the data was generated but also provides a benchmark for subsequent time synchronization alignment of multi-source data. The charging pile monitoring terminal has a more specialized function. It continuously monitors the output voltage, output current, current charging power, real-time state of charge of the vehicle battery, and the working status of the charging connector. This data is sampled at a slightly lower frequency and transmitted to the regional data aggregator via a built-in 4G or Ethernet communication module using an encrypted protocol. Before entering the processing flow, the raw sampled data must undergo a rigorous data cleaning phase. The data cleaning module first runs an outlier removal algorithm based on the Laida criterion, which considers data points exceeding three standard deviations to be low-probability events and should be removed. For voltage and current sampled value sequences, the algorithm dynamically calculates the mean and standard deviation within a sliding window, marking any sample point deviating from the mean by more than three standard deviations as an outlier and removing it. Next, the missing value imputation module handles data gaps caused by communication interruptions or outlier removal. Missing value imputation uses a linear interpolation method, constructing a straight line using two valid data points before and after the missing point, calculating an estimated value for the missing location, and filling in the gap to ensure the temporal continuity of the data sequence. After cleaning and interpolation, the data stream enters the standardization processing module. This module converts the original voltage, current, and power samples, each with different dimensions, into a dimensionless normalized load sequence. The normalization process involves dividing the measured load value at each moment by a preset rated capacity benchmark, typically the rated capacity of the transformer or line at the monitoring point. The normalized load sequence's value range is compressed to the [0,1] interval, facilitating subsequent joint analysis and comparison of signals with different dimensions and magnitudes. The normalized load sequence is stored in a first-in, first-out (FIFO) queue within a circular buffer of a set length. The size of the circular buffer needs to accommodate at least several minutes of sampling data, providing temporary storage space for new data to enter and old data to exit.

[0046] The core of the volatility analysis algorithm is the sliding window variance calculation method. This method divides a continuous normalized load sequence into a series of temporally continuous and potentially partially overlapping intervals, each called a time window. The length of the time window is set to include 30 consecutive normalized load sequence sampling points, and the sliding step size is set to one sampling point. This means that each time a new sampling point is obtained, the time window slides forward one position, and the data within the window is updated once. For each time window, the algorithm calculates the variance of all 30 normalized load values ​​within the window. The variance is a scalar that quantitatively describes the dispersion of the load values ​​around their average value within the time window. A larger variance indicates more drastic load fluctuations within the window, while a smaller variance indicates more stable loads. The volatility analysis algorithm compares the variance calculated for each time window with a dynamically adjusted adaptive threshold. This adaptive threshold is not fixed; its value is derived from statistical analysis of historical operating data.

[0047] Historical stable load sequences are defined as load data segments whose variance values ​​have consistently fallen below a certain initial threshold over the past 24 hours. The system extracts these historical stable load sequences from the historical database, calculates the 90th percentile of all variance values ​​in these sequences, and sets this quantile as the current adaptive threshold. This dynamic update mechanism allows the adaptive threshold to follow the inherent stability of the load. For example, during periods of generally stable load at night, the adaptive threshold is automatically lowered to improve monitoring sensitivity, while during periods of greater natural load fluctuation during the day, the adaptive threshold is raised accordingly to avoid excessive false alarms. The volatility analysis algorithm checks whether the variance values ​​of three consecutive time windows exceed the current adaptive threshold. If the variance values ​​of three consecutive time windows exceed the threshold, the algorithm determines that the charging load control is in an unstable state and immediately triggers a status flag signal transition. This unstable state flag signal acts like a switch, automatically activating the downstream distribution network voltage fluctuation characteristic analysis subroutine. The distribution network voltage fluctuation characteristic analysis subroutine and the volatility analysis algorithm execute in parallel under system scheduling, achieving synchronous correlation analysis of load fluctuations and voltage fluctuations, striving to capture the complete characteristics of the unstable state in its early stages. The sliding window variance calculation method itself also participates in the adaptive threshold update loop. The newly calculated window variance value, which is considered to be in a stable state, will be included in the sample library of historical stable load sequences for subsequent recalculation of the adaptive threshold, thus forming a closed-loop feedback system with self-learning ability, enabling the entire volatility analysis algorithm to gradually adapt to the unique load fluctuation characteristics of specific distribution network nodes.

[0048] Example 3: The process of synchronously analyzing the voltage fluctuation characteristics and charging power variation features of the distribution network begins with the acquisition and preprocessing of multi-source data signals. The distribution network voltage fluctuation signal is captured by power quality analyzers installed at substations and key nodes. The signal records minute changes in the instantaneous voltage value at a high-frequency sampling rate. The charging power variation features are directly obtained from the real-time data stream of the charging pile monitoring terminal. The voltage fluctuation signal passes through a bandpass filter to remove power frequency components and extremely high-frequency noise. The preprocessed voltage time series is then fed into a Fast Fourier Transform (FFT) processor. The FFT converts the time-domain voltage signal into a frequency-domain representation, outputting a spectrum containing each frequency component and its corresponding amplitude. The frequency-domain energy distribution is calculated from this spectrum. This distribution is obtained by integrating the square of the amplitude within each frequency interval, reflecting the magnitude distribution of voltage fluctuation energy in different frequency bands and serving as an important characteristic for characterizing voltage quality degradation modes. The processing of charging power variation characteristics is relatively straightforward. The first-order difference is calculated for the continuous charging power sampling sequence. The first-order difference sequence is obtained by subtracting the power values ​​of two adjacent sampling points. This difference sequence intuitively depicts the rate of increase or decrease of charging power in a short period of time, reflecting the drastic degree of load change.

[0049] Feature fusion technology integrates the two different types of features into a unified multidimensional feature vector. The frequency domain energy distribution is represented as a vector of length N, where each element represents the energy value within a specific frequency range. The first-order difference sequence of charging power is represented as a vector of length M, where the elements are sequential power difference values. Feature fusion employs vector concatenation, linking the frequency domain energy distribution vector and the first-order difference sequence vector of charging power end-to-end to form a multidimensional feature vector of dimension N+M. To eliminate the influence of differences in the physical dimensions and numerical ranges of the different features, the concatenated multidimensional feature vector undergoes a normalization step. Normalization linearly scales the values ​​of each dimension of the vector to the [0,1] interval. The normalized multidimensional feature vector serves as the input data for the pre-trained classification neural network model. The pre-trained classification neural network model uses a fully connected feedforward neural network structure containing one input layer, three hidden layers, and one output layer. The number of neurons in the input layer is strictly equal to the dimension N+M of the multidimensional feature vector. The number of neurons in the three hidden layers decreases sequentially, set to 128, 64, and 32 respectively. Each hidden layer is followed by a ReLU activation function layer, which introduces non-linear transformation capability into the network. The output layer contains only one neuron, which uses the Sigmoid activation function to limit the final output value to the range of 0 to 1. This output value is the charging efficiency evaluation value. The classification neural network model is trained offline using the backpropagation algorithm. The training dataset synthesizes a large number of voltage fluctuation and power change samples collected during historical operation and labeled by expert knowledge or more accurate simulation models. Each sample corresponds to a real charging efficiency label value. The training process adjusts the weights and biases in the network by minimizing the cross-entropy loss function between the model's predicted values ​​and the true labels. Specifically, in each training iteration, the model's forward propagation generates predicted charging efficiency evaluation values, which are constrained to the range of 0 to 1 by a sigmoid activation function. The cross-entropy loss is calculated based on the log-likelihood relationship between the predicted probabilities and the true labels, calculating the loss value for each sample and aggregating it into an overall loss metric. The overall loss value drives the backpropagation algorithm to perform error backpropagation, calculating the gradient components of the weights and biases layer by layer along the network topology from the output layer to the input layer. The gradient information guides the direction of parameter updates, and the values ​​of the weight matrix and bias vector in the fully connected layers are adjusted through an iterative optimization algorithm to continuously converge the cross-entropy loss value. A dropout method is used during training to suppress overfitting and improve the generalization ability of the classification neural network model. After the charging efficiency evaluation value is calculated by the classification neural network model, it is transmitted to the ranking logic module.

[0050] The grading logic module classifies charging performance into three discrete levels based on preset threshold boundaries: high-efficiency, normal, and inefficient. The module internally stores two key threshold parameters: the high-efficiency threshold and the inefficient threshold. The high-efficiency threshold is set to a relatively high value, such as 0.85, while the inefficient threshold is set to a relatively low value, such as 0.30. The grading logic module compares the received charging efficiency evaluation value with these two thresholds. The specific grading rules are: if the charging efficiency evaluation value is greater than the high-efficiency threshold, the current charging performance is determined to be at the high-efficiency level; if the charging efficiency evaluation value is less than the inefficient threshold, the current charging performance is determined to be at the inefficient level; if the charging efficiency evaluation value is between the inefficient and high-efficiency thresholds, the current charging performance is determined to be at the normal level. For charging performance classified as normal, the system automatically triggers a control signal to initiate a time-series decomposition process for more in-depth trend analysis of the current load sequence. The classification neural network model is not static. The system employs a periodic model update mechanism. Every 24 hours, the system incrementally trains the classification neural network model using new data collected the previous day, fine-tuning the network parameters to ensure that the output of the charging efficiency evaluation value can adapt to the slow drift of the system's operating state. The entire process, from signal analysis, feature fusion, neural network calculation to classification, constitutes a complete analysis chain, and its mathematical relationships can be expressed by the following feature fusion formula:

[0051] ;

[0052] in: This represents the fused, normalized multidimensional feature vector. This represents a vector concatenation operation. This represents the frequency domain energy distribution vector extracted from the voltage fluctuation signal. The first-order difference sequence vector representing the charging power. and These represent the functions that take the minimum and maximum values ​​of the vectors, respectively. This formula clarifies the process of normalizing two feature vectors from different sources and then concatenating them into a unified vector, which serves as the direct input to the classification neural network model.

[0053] See Figure 4This chart demonstrates the complete technical process of power distribution network voltage signal processing and charging power feature extraction. It integrates information from multiple dimensions, including time-domain voltage signal, filtered signal, charging power variation trend, and power change rate, into a single visualization interface. The blue curve shows the fluctuations of the original voltage signal, while the red curve displays the voltage signal after bandpass filtering, effectively removing power frequency components and high-frequency noise. The green curve presents the temporal variation of charging power, reflecting the power distribution characteristics at different charging stages. The orange curve displays the first-order difference sequence of power, visually depicting the rate of change of charging power over a short period. The heatmap in the background shows the frequency domain energy distribution characteristics of the voltage signal, with color depth reflecting the energy intensity in different frequency ranges. The chart also labels the charging efficiency evaluation values ​​calculated based on the neural network model and the corresponding efficiency level classification results, providing a quantitative basis for charging strategy optimization.

[0054] Example 4: Extracting Load Trend Components Using Time Series Decomposition Techniques. The Empirical Mode Decomposition (EMD) algorithm is applied. EMD is an adaptive signal processing method suitable for analyzing non-stationary and nonlinear load series data. The algorithm processes a pre-processed normalized load series. The core objective of EMD is to decompose the complex original load series into a finite number of intrinsic mode function components arranged from high to low frequency, and a residual term representing the long-term trend. The decomposition process is achieved through an iterative sieving process. Each sieving step aims to extract the oscillating component with the highest instantaneous frequency in the sequence. The sieving steps include identifying all local maxima and local minima in the sequence, using a cubic spline interpolation function to connect all local maxima to form an upper envelope, and connecting all local minima to form a lower envelope. The mean curves of the upper and lower envelopes are calculated, and the original sequence is subtracted from these mean curves to obtain an intermediate component. This sieving process is repeated until the intermediate component satisfies two criteria for the intrinsic mode function: the number of extreme points in the entire data range is equal to or at most differs from the number of zero-crossing points; and at any given time, the mean of both the upper envelope defined by the local maxima and the lower envelope defined by the local minima is zero. The intermediate component that meets these criteria is accepted as an intrinsic mode function component. After separating this intrinsic mode function component from the original sequence, the remaining sequence is used as a new original sequence, and the above sieving process is repeated until the remaining sequence becomes a monotonic function or only one extreme point remains. The remaining sequence at this point is the residual term, also known as the residual component.

[0055] Low-frequency intrinsic mode function (IMF) components are selected from a series of IMF components obtained from the decomposition as trend components representing the load trend. The selection criterion is based on the ratio of the average period of each IMF component to the total length of the original load sequence. The average period of each IMF component is calculated, which can be estimated by averaging the zero-crossing intervals of the component. The average period is compared with the total duration of the original sequence, and components with a high proportion of average period to the total sequence length are selected. Typically, one or two of the lowest-frequency IMF components with a proportion exceeding a certain threshold are selected and either superimposed or defined separately as trend components. The trend components reflect the slow-changing long-term patterns in the load sequence, filtering out high-frequency components such as daily periodic fluctuations and random noise. The residual term contains the remaining ultra-long-term trend or constant after decomposition. The overall load stability index is quantified by calculating the approximate entropy value of the trend components. Approximate entropy is a metric for measuring the complexity and regularity of a time series; a lower value indicates a more regular and predictable sequence, while a higher value indicates a more random and complex sequence. Calculating the approximate entropy of the trend components requires setting two key parameters: the embedding dimension *m* and the tolerance parameter *r*. The embedding dimension *m* determines the length of the vectors being compared, typically taking a value of 2. The tolerance parameter *r* defines the threshold for determining whether two vectors are "similar," usually taking a scaling factor of the time series standard deviation, such as 0.2 times the standard deviation. The calculation process is as follows: Reconstruct the trend component sequence {X(i)} into a set of m-dimensional vectors, calculate the proportion of vectors whose distance to the others is less than *r*, take the logarithm, and average to obtain Φ(m,r). Then, increment *m* by 1 and repeat the above steps to obtain Φ(m+1,r). The approximate entropy value is the difference between Φ(m,r) and Φ(m+1,r). This calculated approximate entropy value serves as the overall load stability index.

[0056] Referring to Table 1, after the overall load stability index is calculated, the system compares it with a preset stability reference threshold. The stability reference threshold is an empirical threshold value derived from historical data analysis. It is determined by extracting a large number of load intervals marked as "stable" by experts or the system from historical data from a period of stable operation, calculating the approximate entropy values ​​of the trend components corresponding to these intervals, and then statistically analyzing the distribution of these approximate entropy values, taking a certain percentile as the stability reference threshold. The comparison logic is as follows: if the calculated overall load stability index is greater than the stability reference threshold, it means that the complexity or randomness of the current load trend exceeds the typical level of historical stable periods, and the system generates a load instability flag; if the overall load stability index is less than or equal to the stability reference threshold, it means that the regularity of the current load trend is within an acceptable stability range, and the system generates a load stability flag. The load instability flag or load stability flag is sent as an important state variable to the decision logic of the subsequent intelligent optimization algorithm. The intelligent optimization algorithm adjusts the weight allocation or constraints of its optimization strategy according to the type of flag received. For example, when a load instability indicator is received, the intelligent optimization algorithm may focus more on suppressing power fluctuations and improving grid stability; when a load stability indicator is received, the algorithm may focus more on improving economic efficiency and optimizing user satisfaction. The stopping criteria for empirical mode decomposition algorithms are usually based on the properties of the residual components or the limitation of the number of sieving steps. The decomposition process terminates when the residual components become monotonic functions or the number of sieving steps reaches a preset upper limit. The update frequency of the overall load stability index is synchronized with the sampling and analysis window of the load sequence to ensure that the system can respond to changes in load trends in real time.

[0057] Table 1: Characteristics of modal components generated by empirical mode decomposition;

[0058]

[0059] The table above illustrates the characteristics of several typical modal components that may be generated by empirical mode decomposition. IMF1 and IMF2, due to their small ratio of average period to total sequence length, represent high-frequency details and are usually not selected as trend components. IMF3, with an average period reaching half the total sequence length, effectively reflects the medium- to long-term trend of load changes and is selected as the main trend component. IMF4, with an average period equal to or exceeding the sequence length, may be included as a trend component if the analysis window is long enough. Residual components directly represent monotonic trends and are part of the trend components. The determination of trend components directly affects the accuracy of subsequent approximate entropy calculations and stability assessments.

[0060] See Figure 5This chart demonstrates a method for load characteristic analysis and stability assessment based on Empirical Mode Decomposition (EMD). It integrates the original load sequence, extracted trend components, and multiple modal components obtained from EMD into a unified visual framework. The blue curve shows the complete trajectory of the original load sequence, reflecting the complex patterns of charging load fluctuations over time. The red curve highlights the trend components extracted by the EMD algorithm, which filters out high-frequency components such as daily fluctuations and random noise, clearly presenting the long-term variation patterns of the load. The chart also displays multiple intrinsic mode function components generated by EMD, arranged from high to low frequency, each representing load fluctuation characteristics at different time scales. Green, orange, and purple curves correspond to modal components in different frequency ranges, with vertical offset ensuring visual distinguishability. The chart uses prominent circular markers to indicate the current load stability status: red indicates instability, and green indicates stability. It also labels the approximate entropy value calculated based on the trend components. This indicator quantifies the complexity and regularity of the load sequence; a lower value indicates a more regular and predictable sequence.

[0061] Example 5: The operation of the output charging power adjustment command module begins with a specific power distribution network operation scenario. Assume that a charging station in a city's commercial area experiences concentrated load access during the evening peak hours, and the voltage fluctuation at a certain node in the power distribution network approaches its upper limit. The differentiated electricity price incentive model is activated first. This model incorporates response probability curves for different user types. For example, commercial area users have a 60% probability of accepting a charging power adjustment when the electricity price decreases by 0.2 yuan / kWh, while the probability increases to 85% when the price decreases by 0.5 yuan / kWh. Based on the current load level, overall load stability indicators, and historical user behavior data, the system generates an initial charging power scheme. This initial charging power scheme exists in matrix form, with rows representing different time windows and columns representing different user groups or individual charging piles. Matrix elements include the suggested charging power value for that time period and the corresponding user's electricity price discount rate. This initial scheme aims to guide user behavior through economic signals, initially alleviating grid pressure. The particle swarm optimization algorithm then initiates an iterative optimization process for the initial charging power scheme. Particle Swarm Optimization (PSO) treats each possible charging power adjustment scheme as a particle, with the entire swarm searching the solution space. Each particle's position vector encodes a sequence of power setpoints for all charging stations over a future period, while its velocity vector determines the direction and step size of its position update. The objective function is set to minimize the sum of squared voltage deviations at key nodes in the distribution network, considering both line capacity and transformer load rate constraints; schemes violating these constraints are penalized. In each iteration, PSO evaluates the fitness value of each particle's corresponding scheme. Each particle records its historical best position, and the entire swarm records its global best position. Particles update their velocity and position based on their own experience and the swarm's experience, gradually moving towards a better solution region.

[0062] Cooperative game theory is introduced to coordinate the conflicting interests of multiple stakeholders in the optimization process. Cooperative game theory defines grid operators, charging station operators, and electric vehicle users as participants in the game. Grid operators pursue grid stability and operational safety, with their interests reflected in minimizing voltage deviation; charging station operators pursue maximizing operating revenue; and electric vehicle users pursue minimizing charging costs and satisfying their charging needs. The multi-stakeholder interest allocation model assigns a utility function to each stakeholder, quantifying their satisfaction. The utility function for grid stability is negatively correlated with voltage deviation. Specifically, the voltage deviation value is calculated by the absolute difference between the real-time collected voltage sample value and the rated voltage value. The utility function is designed as a decreasing function of voltage deviation, meaning that the utility value of grid stability decreases as voltage deviation increases. This negative correlation ensures that the optimization algorithm prioritizes suppressing voltage fluctuations. The utility function for user satisfaction is positively correlated with electricity price discounts and charging completion rates. Specifically, the electricity price discount originates from an incentive scheme generated by the user response probability curve. Charging completion is quantified by the ratio of the vehicle's real-time battery state of charge to a target value. The utility function exhibits an increasing relationship with both the electricity price discount and charging completion; user satisfaction utility increases synchronously with either a higher discount or a higher completion rate. The objective function of the particle swarm optimization algorithm is thus expanded from a single objective to a weighted summation multi-objective function, with weight coefficients reflecting the relative importance of different entities under the current operating state. For example, when system stability is significantly threatened, the weight for grid stability increases.

[0063] The Shapley value plays a crucial role in each particle position update decision, fairly distributing the "total benefit" resulting from the cooperative game among each participant. For each newly generated candidate solution, the algorithm calculates how the objective function value would change if the participation of the grid operator, charging station operator, or user group were absent. The Shapley value calculates each participant's share of the credit based on these marginal contributions. When updating particle positions, the algorithm considers not only the absolute improvement of the objective function but also the fairness of the benefit distribution indicated by the Shapley value, guiding the search towards solutions that achieve Pareto improvement—that is, improving overall benefit without significantly harming the interests of any party. This ensures that the final solution is Pareto optimal, preventing any party from further improving its own benefit without harming the interests of others.

[0064] The particle swarm optimization algorithm iterates continuously until a convergence condition is met. The convergence condition can be that the fitness value improves by less than a minimum threshold across multiple iterations, or that a preset maximum number of iterations is reached. After convergence, the charging power scheme represented by the globally optimal particle is selected as the optimal charging power command sequence. This sequence includes the specific power command value for each charging pile in each subsequent time interval, as well as the electricity price information to be provided to users. These commands are sent to the charging pile actuators via standard communication protocols. The charging pile controller receives the commands and adjusts its output power. The source-load-storage-medium-voltage distribution network collaborative planning framework receives real-time execution feedback from the optimal charging power command sequence, including actual power changes, voltage recovery status, and user response results. This feedback data is used to update the user response probability curve in the differentiated electricity price incentive model and to fine-tune the weight coefficients or cooperative game model parameters in the particle swarm optimization algorithm. This process, from command generation and execution to feedback adjustment, forms a closed-loop optimization system, enabling the control strategy to dynamically adapt to changes in grid conditions and user behavior.

[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for estimating electric vehicle charging load and optimizing charging methods, characterized in that, The method includes: initializing the source-load-storage-medium voltage distribution network collaborative planning framework, and integrating the geographical distribution of charging stations and user charging behavior patterns; Real-time acquisition of charging load sequence data; determination of charging load control status through fluctuation analysis algorithm. When the charging load control is in an unstable state, the voltage fluctuation characteristics of the distribution network and the charging power change characteristics are analyzed simultaneously to generate a charging efficiency evaluation value. Based on the charging efficiency evaluation value, the charging effect is divided into high efficiency level, normal level and low efficiency level. To assess the charging performance at standard levels, time series decomposition techniques are used to extract load trend components and calculate overall load stability indicators. It integrates charging efficiency assessment values ​​and overall load stability indicators to output charging power adjustment commands; The initialization source-load-storage-medium voltage distribution network collaborative planning framework includes constructing a diversified user resource classification system, which covers industrial user load curves, commercial user charging preferences, and residential user electricity consumption habits. It integrates renewable energy output data with charging pile configuration parameters to form a multi-timescale interactive dataset. The multi-timescale interactive dataset generates a basic distribution map of charging load through a spatial interpolation algorithm for subsequent real-time data collection. The synchronous analysis of distribution network voltage fluctuation characteristics and charging power change features includes extracting the frequency domain energy distribution of the voltage fluctuation signal and the first-order difference sequence of charging power, and using feature fusion technology to merge the frequency domain energy distribution and the difference sequence into a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a pre-trained classification neural network model, and outputting a charging efficiency evaluation value; the classification neural network model is trained offline using the backpropagation algorithm, and the training data includes labeled voltage fluctuation and power change samples.

2. The method for estimating electric vehicle charging load and optimizing charging methods according to claim 1, characterized in that, The real-time acquisition of charging load sequence data specifically includes deploying a cluster of smart meters and charging pile monitoring terminals to obtain voltage sampling values, current sampling values, and charging timestamps at a fixed sampling frequency; removing outliers and imputing missing values ​​from the original sampling data; and generating a normalized load sequence through standardization processing; the normalized load sequence is then transmitted as input to a volatility analysis algorithm.

3. The method for estimating electric vehicle charging load and optimizing charging methods according to claim 2, characterized in that, The method of determining the charging load control status through volatility analysis algorithm includes using a sliding window variance calculation method to divide the normalized load sequence into continuous time windows and calculate the variance of the load value in each time window; if the variance of three consecutive time windows exceeds the adaptive threshold, the charging load control is determined to be in an unstable state, and the analysis of the distribution network voltage fluctuation characteristics is triggered; the adaptive threshold is dynamically updated through the statistical quantiles of historical stable load sequences.

4. The method for estimating electric vehicle charging load and optimizing charging methods according to claim 1, characterized in that, The specific method for classifying charging performance into high-efficiency, normal, and low-efficiency levels based on charging efficiency evaluation values ​​is as follows: set high-efficiency level thresholds and low-efficiency level thresholds; if the charging efficiency evaluation value is greater than the high-efficiency level threshold, it is classified as high-efficiency; if the charging efficiency evaluation value is less than the low-efficiency level threshold, it is classified as low-efficiency; if the charging efficiency evaluation value is between the high-efficiency level threshold and the low-efficiency level threshold, it is classified as normal; for normal-level charging performance, time series decomposition technology is automatically activated.

5. The method for estimating electric vehicle charging load and optimizing charging mode according to claim 4, characterized in that, The extraction of load trend components using time series decomposition technology includes applying empirical mode decomposition algorithm to decompose the load series into intrinsic mode functions and residual terms, and selecting low-frequency intrinsic mode functions as trend components; The approximate entropy value of the trend component is calculated as an overall load stability index; the overall load stability index is used in subsequent intelligent optimization algorithms.

6. The method for estimating electric vehicle charging load and optimizing charging methods according to claim 5, characterized in that, The calculation of the overall load stability index includes comparing it with a stability reference threshold; if the overall load stability index is greater than the stability reference threshold, a load instability flag is generated; otherwise, a load stability flag is generated; the load instability flag or the load stability flag participates in the decision-making process of the intelligent optimization algorithm.

7. The method for estimating electric vehicle charging load and optimizing charging methods according to claim 6, characterized in that, The output charging power adjustment command includes designing a differentiated electricity price incentive model, generating an initial charging power scheme based on the user response probability curve, iteratively adjusting the charging power scheme using a particle swarm optimization algorithm, taking the minimization of the distribution network voltage deviation as the objective function, and outputting the optimal charging power command sequence after convergence; the optimal charging power command sequence is directly sent to the charging pile actuator.

8. The method for estimating electric vehicle charging load and optimizing charging mode according to claim 7, characterized in that, The iterative process of the particle swarm optimization algorithm is combined with cooperative game theory to establish a multi-stakeholder interest allocation model and quantify the weights of user satisfaction and power grid stability. The Shapley value is calculated during each iteration to update the particle position, ensuring the Pareto optimal solution; the final generated charging power adjustment command is integrated into the source-load-storage-medium voltage distribution network collaborative planning framework to form a closed-loop optimization system.

Images