Split type multi-split high-power direct-current charging pile system

Abstract

The invention provides a split type multi-split high-power direct-current charging pile system, which belongs to the technical field of split type direct-current charging piles, and is characterized in that a dynamic power balance optimization module is established, and a sliding time window method is adopted to carry out real-time monitoring and feature extraction on the power demand of each charging port; the method comprises the following steps of: constructing an upper-layer optimization model taking the maximization of the total charging efficiency as a target and a lower-layer optimization model taking the minimization of the charging time as a target, obtaining an optimal power distribution scheme by utilizing a game solving algorithm, and inputting the scheme into a multi-head attention network based on a Transform architecture to carry out accurate power regulation and control. Intelligent distribution and real-time optimization of the power of each charging port are realized through dynamic weight adjustment of a multi-head attention mechanism, and the technical problems that a multi-port charging pile system cannot dynamically optimize power distribution according to the actual demand of each charging port, so that the charging efficiency is low and the overall energy consumption of the system is too high are solved.

Description

Technical Field

[0001] This invention belongs to the field of split-type DC charging pile technology, specifically, it relates to a split-type high-power DC charging pile system that supports multiple charging points. Background Art

[0002] With the rapid development of the electric vehicle industry, high-power DC charging piles have become an important part of electric vehicle charging infrastructure. Traditional multi-port charging pile systems typically employ a fixed power allocation strategy, where each charging port charges according to a preset power ratio. This approach is widely used in charging stations, highway service areas, and commercial centers. However, the traditional fixed power allocation strategy cannot dynamically adjust based on the battery status, charging needs, and environmental conditions of different vehicles. This often results in some charging ports having excessive power while others have insufficient power, leading to overall system inefficiency and energy waste. In current multi-port charging pile applications, due to the lack of intelligent power management algorithms and real-time optimization mechanisms, the system cannot achieve power balance and coordinated control among the charging ports. Especially under high load operation, traditional technologies struggle to maximize charging efficiency and system stability. In other words, existing technologies suffer from the technical problem that multi-port charging pile systems cannot dynamically optimize power allocation according to the actual needs of each charging port, resulting in low charging efficiency and excessive overall system energy consumption. Summary of the Invention

[0003] In view of this, the present invention provides a split-type high-power DC charging pile system that can solve the technical problem in the prior art where multi-port charging pile systems cannot dynamically optimize power allocation according to the actual needs of each charging port, resulting in low charging efficiency and excessive overall system energy consumption.

[0004] This invention is implemented as follows: A split-type high-power DC charging pile system control chip incorporates a dynamic power balance optimization module. This module monitors the actual charging power demand and battery status information of each charging port. Based on the sliding time window method, it performs segmented analysis on the power demand data of each charging port, extracts power change characteristics, calculates the theoretical optimal power allocation ratio of each charging port according to a preset power allocation strategy, establishes a mapping relationship matrix between power demand and allocation ratio, and uses a charging power prediction model to predict the power demand of each charging port in the future time period, generating a power demand prediction curve. It constructs an upper-level optimization model aiming to maximize total charging efficiency and a lower-level optimization model aiming to minimize charging time. A game-theoretic algorithm is used to solve the two-level optimization problem to obtain the optimal power allocation scheme. This optimal power allocation scheme is input into the power balance adjustment model. By adjusting the attention weight parameters of the power balance adjustment model, a power adjustment command is generated. Based on the power adjustment command, the power conversion unit and voltage adjustment unit are controlled to achieve precise allocation and dynamic adjustment of power at each charging port.

[0005] Specifically, the sliding time window method involves a sliding time period with a base length of 5 minutes. The window length is adaptively adjusted according to the frequency of charging power changes. When the power change frequency is higher than 0.1Hz, the window length is shortened to 3 minutes. When the power change frequency is lower than 0.05Hz, the window length is extended to 8 minutes. Each time window contains continuous power sampling points with a sampling frequency of 1Hz.

[0006] Specifically, the power variation characteristics are power data feature parameters extracted through statistical analysis within each time window. The power peak value is the difference between the maximum and minimum power values ​​within the time window, the power average value is the arithmetic mean of all power sampling points within the time window, and the power fluctuation amplitude is the standard deviation of the power data.

[0007] The mapping matrix is ​​a 4×4 two-dimensional array structure. The row index represents the four charging ports, the column index represents the four power allocation levels, and the matrix element value represents the allocation weight coefficient of the corresponding port under the corresponding power allocation level. The allocation weight coefficient ranges from 0 to 1, and the sum of the allocation weight coefficients of all ports under the same power allocation level is equal to 1.

[0008] Specifically, the two-layer optimization problem is a nested optimization structure consisting of an upper-layer optimization model and a lower-layer optimization model. The upper-layer optimization model aims to maximize the total charging efficiency, while the lower-layer optimization model aims to minimize the charging time. The upper-layer and lower-layer optimization models are coupled through power allocation variables.

[0009] The total charging efficiency maximization function comprehensively considers the power conversion efficiency of each charging port, the heat dissipation efficiency of the heat dissipation unit, and the battery charging efficiency. The inputs include port weight parameters, port efficiency parameters, port power parameters, efficiency correction factors, and temperature compensation coefficients. The output is the overall charging efficiency value of the system.

[0010] The charging time minimization function takes into account the differences in charging rates at different charging ports and the impact of charging strategies. The inputs include time weight parameters, battery capacity parameters, charging power parameters, charging efficiency factors, and load balancing coefficients. The output is the shortest time to complete all charging tasks.

[0011] The power balance adjustment model is specifically based on a multi-head attention network of the Transformer architecture, including 6 encoder layers, each containing a multi-head attention mechanism and a feedforward neural network. The multi-head attention mechanism has 8 heads, each attention head has a dimension of 64, and the hidden layer dimension of the feedforward neural network is 512.

[0012] Specifically, the steps for establishing the training dataset of the power balance adjustment model include collecting historical charging data under different charging scenarios, preprocessing the raw data including data cleaning, outlier detection and data normalization, constructing input feature vectors and output label vectors, and dividing the dataset into training set, validation set and test set in a ratio of 7:2:1.

[0013] Specifically, the training steps of the power balance adjustment model involve updating parameters using the Adam optimizer, setting the initial learning rate to 0.001, the batch size to 32, the number of training rounds to 100, calculating the prediction error using the cross-entropy loss function, and employing an early stopping strategy to prevent overfitting when the validation set loss does not decrease for five consecutive rounds.

[0014] The attention weight parameters of the multi-head attention mechanism are dynamically adjusted according to the current number of charging ports, the total system power, and the charging priority. When there are 1 to 2 charging ports, the attention weight parameters are concentrated. When there are 3 to 4 charging ports, the attention weight parameters are evenly distributed. When the total system power exceeds 80% of the rated power, the attention weight parameters are tilted towards the high-priority ports.

[0015] The attention weight adjustment function calculates the weight adjustment factor based on the charging port load rate, system power utilization rate and charging urgency. When the weight adjustment factor is in the range of 0 to 0.25, a linear growth weight adjustment function is used, and when the weight adjustment factor is in the range of 0.25 to 0.5, a square root growth weight adjustment function is used.

[0016] Specifically, when the weight adjustment factor is in the range of 0.5 to 0.75, a logarithmic growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the base weight value and the adjustment factor plus one natural logarithm. When the weight adjustment factor is in the range of 0.75 to 1, an exponential decay weight adjustment function is used, and the attention weight parameter is adjusted according to the product of the maximum weight value and the adjustment factor twice the negative exponent.

[0017] The power distribution unit and the power conversion unit are connected via a high-voltage DC bus. The output end of the power conversion unit is connected to the charging interface unit via a power output cable. The charging interface unit includes four independent charging ports, each equipped with an independent power adjustment module.

[0018] The heat dissipation unit is located inside the housing of the power conversion unit and includes an air-cooled heat sink and a liquid-cooled heat sink. The air-cooled heat sink is connected to the power devices of the power conversion unit through a heat-conducting copper pipe, and the liquid-cooled heat sink is connected to the main heat-generating devices of the power conversion unit through a coolant circulation pipeline.

[0019] The current detection unit is located in the output circuit of each charging port. The communication unit includes a wireless communication module and a wired communication module. The wireless communication module adopts 4G / 5G communication technology, and the wired communication module adopts Ethernet communication protocol. The insulation monitoring unit is located in the charging circuit.

[0020] This invention effectively solves the problem of unreasonable power allocation in multi-port charging pile systems by constructing a power demand analysis mechanism based on a sliding time window and a two-layer optimization model, combined with a multi-head attention network based on the Transformer architecture to achieve dynamic power balance adjustment. The invention uses a sliding time window method to analyze and predict the power demand of each charging port in real time, establishing an upper-layer optimization model aimed at maximizing total charging efficiency and a lower-layer optimization model aimed at minimizing charging time. The optimal power allocation scheme is obtained through a game-theoretic algorithm, overcoming the shortcomings of traditional fixed power allocation strategies that cannot adapt to dynamic demand changes. This invention utilizes a power balance adjustment model based on the Transformer architecture and dynamically adjusts the weight parameters of each charging port through a multi-head attention mechanism, achieving precise control and real-time optimization of power allocation. This allows the system to adaptively adjust the power allocation strategy according to actual charging demand, significantly improving charging efficiency and reducing system energy consumption. In summary, this invention solves the technical problem mentioned in the background art: multi-port charging pile systems cannot dynamically optimize power allocation according to the actual demand of each charging port, resulting in low charging efficiency and excessive overall system energy consumption. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system provided by the present invention.

[0022] Figure 2 A flowchart illustrating the steps performed by the dynamic power balance optimization module.

[0023] Figure 3 This is a schematic diagram of the power balance regulation model involved in the present invention.

[0024] Figure 4 This is a power demand prediction curve from Example 2.

[0025] Figure 5 This is a system temperature distribution monitoring diagram from Example 2. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0027] like Figure 1The diagram shows the composition of a split-type multi-port high-power DC charging pile system provided by this invention. The system includes: a power distribution unit, a power conversion unit, a charging interface unit, a heat dissipation unit, a current detection unit, a voltage regulation unit, a communication unit, an insulation monitoring unit, and a control chip. The power distribution unit and the power conversion unit are connected via a high-voltage DC bus. The output of the power conversion unit is connected to the charging interface unit via a power output cable. The charging interface unit includes four independent charging ports, each equipped with an independent power regulation module. The power distribution unit and the power conversion unit are integrated within the main control cabinet, while the charging interface units are separately located at charging terminal positions 5 to 50 meters away from the main control cabinet. The heat dissipation unit is located inside the power conversion unit's housing and includes an air-cooled radiator and a liquid-cooled radiator. The air-cooled radiator is connected to the power devices of the power conversion unit via a heat-conducting copper pipe, and the liquid-cooled radiator is connected to the main heat-generating devices of the power conversion unit via a coolant circulation pipeline. The current detection unit is located in the output circuit of each charging port for real-time monitoring of charging. The voltage regulation unit, located inside the power conversion unit, regulates the stability and accuracy of the output voltage, controlling the magnitude and direction of the current. The communication unit includes a wireless communication module and a wired communication module; the wireless communication module uses 4G / 5G communication technology, and the wired communication module uses the Ethernet communication protocol. The insulation monitoring unit, located in the charging circuit, monitors the insulation resistance and leakage current. The control chip, located in the control cavity within the power distribution unit, is electrically connected to and interacts with the power distribution unit, power conversion unit, charging interface unit, heat dissipation unit, current detection unit, voltage regulation unit, communication unit, and insulation monitoring unit. The control chip includes a dynamic power balance optimization module, which dynamically allocates power based on the actual load requirements of each charging port and optimizes the charging strategy using predictive algorithms, ultimately achieving power balance and maximizing efficiency for multi-port charging. The control cavity also contains a data storage device, which is electrically connected to the control chip and stores historical data and optimization parameters during the charging process.

[0028] like Figure 2 As shown, the dynamic power balance optimization module is used to perform the following steps: S01. Monitor the actual charging power demand and battery status information of each charging port. The battery status information includes the remaining battery capacity, charging current, charging voltage and battery temperature. At the same time, obtain the output power and efficiency parameters of the power conversion unit. S02. Based on the sliding time window method, the power demand data of each charging port is segmented and analyzed to extract the power change characteristics within each time window. The power change characteristics include power peak, power average and power fluctuation amplitude. S03. Based on the preset power allocation strategy, calculate the theoretical optimal power allocation ratio of each charging port and establish a mapping relationship matrix between power demand and allocation ratio. S04. Use the charging power prediction model to predict the power demand of each charging port in the future time period, generate a power demand prediction curve, and identify the peak and trough times of power demand in the power demand prediction curve. S05. Construct an upper-level optimization model with the goal of maximizing total charging efficiency and a lower-level optimization model with the goal of minimizing charging time. Solve the two-level optimization problem through a game theory algorithm to obtain the optimal power allocation scheme. S06. Input the optimal power allocation scheme into the pre-trained power balance adjustment model, and generate accurate power adjustment instructions by adjusting the attention weight parameters of the power balance adjustment model. The power adjustment instructions include the target power value and adjustment rate of each port. S07. Control the power conversion unit and voltage regulation unit according to the power regulation command to realize the precise allocation and dynamic adjustment of power of each charging port, and at the same time monitor the efficiency change and stability index during the power regulation process.

[0029] Specifically, the time window refers to a sliding time period with a base length of 5 minutes. The window length is adaptively adjusted according to the frequency of charging power change. When the power change frequency is higher than 0.1Hz, the window length is shortened to 3 minutes. When the power change frequency is lower than 0.05Hz, the window length is extended to 8 minutes. Each time window contains continuous power sampling points with a sampling frequency of 1Hz.

[0030] Specifically, the power variation characteristics refer to the power data feature parameters extracted through statistical analysis methods within each time window. The power peak value is the difference between the maximum and minimum power values ​​within the time window, the power average value is the arithmetic mean of all power sampling points within the time window, and the power fluctuation amplitude is the standard deviation of the power data.

[0031] The mapping matrix is ​​a 4×4 two-dimensional array structure. The row index represents the four charging ports, the column index represents the four power allocation levels, and the matrix element value represents the allocation weight coefficient of the corresponding port under the corresponding power allocation level. The allocation weight coefficient ranges from 0 to 1, and the sum of the allocation weight coefficients of all ports under the same power allocation level is equal to 1.

[0032] Specifically, the two-layer optimization problem refers to a nested optimization structure that includes an upper-layer optimization model and a lower-layer optimization model. The upper-layer optimization model aims to maximize the total charging efficiency, while the lower-layer optimization model aims to minimize the charging time. The upper-layer optimization model and the lower-layer optimization model are coupled through power allocation variables.

[0033] The total charging efficiency maximization function is used to calculate the overall charging efficiency index of the system. It comprehensively considers the power conversion efficiency of each charging port, the heat dissipation efficiency of the heat dissipation unit, and the battery charging efficiency. The inputs include port weight parameters, port efficiency parameters, port power parameters, efficiency correction factors, and temperature compensation coefficients. The output is the overall charging efficiency value of the system. The port weight parameters are derived from the monitoring data of the current detection unit, the port efficiency parameters are derived from the efficiency monitoring of the power conversion unit, the port power parameters are derived from the power output of the charging interface unit, the efficiency correction factor is calculated through the sine function relationship between the port weight parameters and the port efficiency parameters, and the temperature compensation coefficient is calculated through the exponential function relationship between the temperature monitoring data of the heat dissipation unit and the reference temperature.

[0034] The charging time minimization function is used to calculate the shortest time required to complete all charging tasks. It considers the differences in charging rates at different charging ports and the impact of charging strategies. The inputs include time weight parameters, battery capacity parameters, charging power parameters, charging efficiency factors, and load balancing coefficients. The output is the shortest time to complete all charging tasks. The time weight parameters are derived from the charging priority information received by the communication unit. The battery capacity parameters are derived from the remaining battery capacity in the battery status information. The charging power parameters are derived from the target power value in the power adjustment command. The charging efficiency factor is calculated through the logarithmic function relationship between the time weight parameters and the battery capacity parameters. The load balancing coefficient is calculated through the cosine function relationship between the charging power parameters and the standard power.

[0035] like Figure 3As shown, the specific structure of the power balance regulation model is a multi-head attention network based on the Transformer architecture, including 6 encoder layers. Each layer contains a multi-head attention mechanism and a feedforward neural network. The multi-head attention mechanism has 8 heads, and each attention head has a dimension of 64. The hidden layer dimension of the feedforward neural network is 512. The model input is the power demand vector of each charging port and the system state vector, and the output is the power allocation weight vector and the regulation strategy vector. The steps for establishing the training dataset of the power balance regulation model specifically include collecting historical charging data under different charging scenarios. The historical charging data includes charging curves of different types of batteries, charging performance data under different ambient temperatures, and efficiency data under different charging powers. Preprocessing of the raw data includes data cleaning, heterogeneous processing, etc. Constant value detection and data normalization are performed to construct input feature vectors and output label vectors. The input feature vectors include power demand, battery state information, environmental parameters, and system state. The output label vectors include the optimal power allocation scheme and adjustment strategy. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training steps of the power balance adjustment model specifically include updating parameters using the Adam optimizer, setting the initial learning rate to 0.001, the batch size to 32, and the number of training epochs to 100. The prediction error is calculated using the cross-entropy loss function. The model performance is evaluated on the validation set. If the validation set loss does not decrease for 5 consecutive epochs, an early stopping strategy is adopted to prevent overfitting. The generalization performance of the final model is evaluated on the test set. The optimal model parameters are saved for practical applications.

[0036] The attention weight parameters of the multi-head attention mechanism in the power balance regulation model are dynamically adjusted according to the current number of charging ports, the total system power, and the charging priority. When there are 1 to 2 charging ports, the attention weight parameters are concentrated and mainly focus on the charging characteristics of a single port. When there are 3 to 4 charging ports, the attention weight parameters are evenly distributed to balance the charging needs of each port. When the total system power exceeds 80% of the rated power, the attention weight parameters are tilted towards high-priority ports to ensure the completion of important charging tasks.

[0037] The attention weight adjustment function is used to dynamically adjust the attention weight parameters of the power balance adjustment model according to the system operating status. A weight adjustment factor is calculated based on the charging port load rate, system power utilization rate, and charging urgency. When the weight adjustment factor is in the range of 0 to 0.25, a linear growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the product of the base weight value and twice the adjustment factor. When the weight adjustment factor is in the range of 0.25 to 0.5, a square root growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the base weight value and twice the square root of the adjustment factor. When the weight adjustment factor is in the range of 0.5 to 0.75, a logarithmic growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the base weight value and twice the natural logarithm of the adjustment factor. When the weight adjustment factor is in the range of 0.75 to 1, an exponential decay weight adjustment function is used, and the attention weight parameter is adjusted according to the product of the maximum weight value and twice the negative exponent of the adjustment factor.

[0038] The specific implementation methods of the above steps are described in detail below.

[0039] The specific implementation of step S01 involves real-time acquisition of operating parameters from each charging port via a sensor network of the current detection unit and voltage regulation unit. First, a data acquisition task scheduling mechanism is established to synchronously acquire charging current, charging voltage, and charging power data from all ports at a sampling frequency of 1Hz. Next, status information such as remaining battery capacity, internal resistance, and temperature is obtained from the battery management system via the CAN bus communication protocol. Simultaneously, parameters such as input power, output power, conversion efficiency, and device temperature of the power conversion unit are monitored. A Kalman filter algorithm is used to filter the acquired raw data, eliminating noise interference and abnormal fluctuations to ensure data accuracy and stability. The purpose of this step is to provide reliable basic data support for subsequent power analysis and optimization.

[0040] The specific implementation of step S02 involves using a sliding time window method to perform time-series analysis and feature extraction on power demand data. First, the time window length is adaptively adjusted based on the power change frequency. The dominant frequency component of the power signal is calculated using a Fast Fourier Transform. When the dominant frequency is higher than 0.1Hz, the window length is set to 3 minutes; when the dominant frequency is lower than 0.05Hz, the window length is set to 8 minutes; the default window length is 5 minutes. Next, statistical characteristic parameters of the power data are calculated within each time window, including maximum, minimum, average, standard deviation, skewness, and kurtosis. Then, a moving average algorithm is used to smooth the power change curve and identify the trend characteristics of power changes. Finally, principal component analysis is used to reduce the dimensionality of the multidimensional feature vectors and extract the most representative power change features. The purpose of this step is to extract valuable pattern features from a large amount of time-series data, providing a data foundation for power prediction and optimization decisions.

[0041] The specific implementation of step S03 is to establish a mathematical model for the power allocation strategy based on multi-objective optimization theory. First, four power allocation levels are defined, corresponding to 25%, 50%, 75%, and 100% of the system's rated power output, respectively. Next, weight allocation rules are established based on the battery type, charging demand, and priority of each charging port, and the analytic hierarchy process (AHP) is used to determine the allocation weight coefficients for each port at different power levels. Then, a 4×4 mapping matrix is ​​constructed. The matrix model inputs are the power demand vector and system state vector of each charging port, and the outputs are the power allocation weight vector and adjustment strategy vector. The steps for establishing the training dataset for the power balance adjustment model specifically include collecting historical charging data under different charging scenarios. This historical charging data includes charging curves for different battery types, charging performance data under different ambient temperatures, and efficiency data under different charging powers. Preprocessing of the raw data includes data cleaning, outlier detection, and data normalization. An input feature vector and an output label vector are constructed. The input feature vector includes power demand, battery state information, and environmental parameters. The system parameters and system state are considered. The output label vector contains the optimal power allocation scheme and adjustment strategy. The dataset is divided into training set, validation set and test set in a ratio of 7:2:1. The training steps of the power balance adjustment model specifically include updating parameters using the Adam optimizer, setting the initial learning rate to 0.001, the batch size to 32, the number of training rounds to 100, calculating the prediction error using the cross-entropy loss function, evaluating the model performance on the validation set, using an early stopping strategy to prevent overfitting when the validation set loss does not decrease for 5 consecutive rounds, evaluating the generalization performance of the final model on the test set, and saving the optimal model parameters for practical applications.

[0042] The attention weight parameters of the multi-head attention mechanism in the power balance regulation model are dynamically adjusted according to the current number of charging ports, the total system power, and the charging priority. When there are 1 to 2 charging ports, the attention weight parameters are concentrated and mainly focus on the charging characteristics of a single port. When there are 3 to 4 charging ports, the attention weight parameters are evenly distributed to balance the charging needs of each port. When the total system power exceeds 80% of the rated power, the attention weight parameters are tilted towards high-priority ports to ensure the completion of important charging tasks.

[0043] The attention weight adjustment function is used to dynamically adjust the attention weight parameters of the power balance adjustment model according to the system operating status. A weight adjustment factor is calculated based on the charging port load rate, system power utilization rate, and charging urgency. When the weight adjustment factor is in the range of 0 to 0.25, a linear growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the product of the base weight value and twice the adjustment factor. When the weight adjustment factor is in the range of 0.25 to 0.5, a square root growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the base weight value and twice the square root of the adjustment factor. When the weight adjustment factor is in the range of 0.5 to 0.75, a logarithmic growth weight adjustment function is used, and the attention weight parameter is adjusted according to the sum of the base weight value and twice the natural logarithm of the adjustment factor. When the weight adjustment factor is in the range of 0.75 to 1, an exponential decay weight adjustment function is used, and the attention weight parameter is adjusted according to the product of the maximum weight value and twice the negative exponent of the adjustment factor.

[0044] The specific implementation methods of the above steps are described in detail below.

[0045] The specific implementation of step S01 involves real-time acquisition of operating parameters from each charging port via a sensor network of the current detection unit and voltage regulation unit. First, a data acquisition task scheduling mechanism is established to synchronously acquire charging current, charging voltage, and charging power data from all ports at a sampling frequency of 1Hz. Next, status information such as remaining battery capacity, internal resistance, and temperature is obtained from the battery management system via the CAN bus communication protocol. Simultaneously, parameters such as input power, output power, conversion efficiency, and device temperature of the power conversion unit are monitored. A Kalman filter algorithm is used to filter the acquired raw data, eliminating noise interference and abnormal fluctuations to ensure data accuracy and stability. The purpose of this step is to provide reliable basic data support for subsequent power analysis and optimization.

[0046] The specific implementation of step S02 involves using a sliding time window method to perform time-series analysis and feature extraction on power demand data. First, the time window length is adaptively adjusted based on the power change frequency. The dominant frequency component of the power signal is calculated using a Fast Fourier Transform. When the dominant frequency is higher than 0.1Hz, the window length is set to 3 minutes; when the dominant frequency is lower than 0.05Hz, the window length is set to 8 minutes; the default window length is 5 minutes. Next, statistical characteristic parameters of the power data are calculated within each time window, including maximum, minimum, average, standard deviation, skewness, and kurtosis. Then, a moving average algorithm is used to smooth the power change curve and identify the trend characteristics of power changes. Finally, principal component analysis is used to reduce the dimensionality of the multidimensional feature vectors and extract the most representative power change features. The purpose of this step is to extract valuable pattern features from a large amount of time-series data, providing a data foundation for power prediction and optimization decisions.

[0047] The specific implementation of step S03 involves establishing a mathematical model for the power allocation strategy based on multi-objective optimization theory. First, four power allocation levels are defined, corresponding to 25%, 50%, 75%, and 100% of the system's rated power output, respectively. Next, weight allocation rules are established based on the battery type, charging demand, and priority of each charging port, and the analytic hierarchy process (AHP) is used to determine the allocation weight coefficients for each port at different power levels. Then, a 4×4 mapping matrix is ​​constructed, with matrix element values ​​calculated using fuzzy comprehensive evaluation, comprehensively considering factors such as charging urgency, battery charging characteristics, and system load balance. Finally, a genetic algorithm is used to optimize and adjust the weight coefficients, ensuring that the sum of the weight coefficients for each port is 1, while simultaneously satisfying the fairness and efficiency constraints of power allocation. The purpose of this step is to establish a scientific and reasonable power allocation framework, providing a decision-making basis for subsequent dynamic optimization.

[0048] The specific implementation of step S04 involves constructing a charging power demand prediction model using machine learning methods. First, a time-series prediction model based on a Long Short-Term Memory (LSTM) neural network is established. The network structure contains three LSTM layers, each with 128 neurons, and dropout technology is used to prevent overfitting. Next, historical power data, battery status information, ambient temperature, and user behavior patterns are used as input features, and training samples are constructed using a sliding window approach. Then, the prediction model is trained using the Adam optimization algorithm, with a learning rate of 0.001, a batch size of 64, and 200 training epochs. The prediction results are then post-processed, using cubic spline interpolation to smooth the prediction curve, and a threshold detection algorithm to identify peak and trough times of power demand. The peak threshold is set to 1.2 times the average value, and the trough threshold is set to 0.8 times the average value. The purpose of this step is to accurately predict future power demand trends, providing a scientific basis for forward-looking power scheduling.

[0049] The specific implementation of step S05 involves constructing a two-layer optimization model and using a game theory algorithm to solve for the optimal power allocation scheme. First, an upper-layer optimization model is established with the objective function of maximizing total charging efficiency, constrained by system power balance constraints, port power limitation constraints, and thermal balance constraints. Next, a lower-layer optimization model is established with the objective function of minimizing charging time, constrained by battery charging characteristics constraints, user demand constraints, and safety protection constraints. Then, Stackelberg game theory is used to describe the master-slave relationship between the upper and lower-layer optimization problems. The upper-layer optimization acts as the leader, formulating the power allocation strategy, while the lower-layer optimization acts as the follower, responding to the power allocation results. Next, a hybrid optimization algorithm combining genetic algorithm and particle swarm optimization is used to solve the two-layer optimization problem, finding a Nash equilibrium solution through multiple iterations. Finally, the solution results are verified for feasibility and stability analysis to ensure the implementability of the optimal power allocation scheme. The purpose of this step is to find the best balance point among multiple optimization objectives to achieve the optimization of the overall system performance.

[0050] The specific implementation of step S06 involves inputting the optimization solution into a pre-trained power balance regulation model to generate precise control commands. First, the optimal power allocation scheme is encoded, converting the power allocation results into an input format recognizable by the model, including parameters such as the target power value, adjustment priority, and time constraints for each port. Next, the weight parameters of the multi-head attention mechanism are dynamically adjusted according to the current system state; when the number of charging ports is small, the attention weight of a single port is increased, and when the system load is high, attention resources are tilted towards high-priority ports. Then, a nonlinear transformation is performed on the feature vector through a feedforward neural network layer, and the ReLU activation function is used to enhance the model's expressive power. Next, the output vector is normalized using the softmax function to generate power regulation commands for each port, including parameters such as the target power value, adjustment rate, and adjustment direction. Finally, the generated control commands undergo safety checks and feasibility verification to ensure that the command parameters are within a safe range and will not cause system instability. The purpose of this step is to transform the abstract optimization results into concrete, executable control commands, achieving intelligent power regulation control.

[0051] The specific implementation of step S07 involves precisely controlling each hardware unit to achieve dynamic power allocation and regulation based on power regulation commands. First, a PWM control signal is sent to the power conversion unit via a digital signal processor to adjust the IGBT switching frequency and duty cycle, precisely controlling the output power of each port. Next, the closed-loop control system of the voltage regulation unit stabilizes the output voltage, employing a PI control algorithm to eliminate voltage fluctuations and ripple, maintaining voltage regulation accuracy within ±0.5%. Then, key parameters during the power regulation process are monitored in real time, including conversion efficiency, device temperature, output voltage stability, and current harmonic distortion. Next, an adaptive control algorithm is used to optimize the control parameters online based on the actual regulation effect, automatically adjusting the control strategy when efficiency degradation or stability deterioration is detected. Finally, an abnormal protection mechanism is established; when abnormal conditions such as overcurrent, overvoltage, or overtemperature are detected, the power output of the corresponding port is immediately cut off to ensure the safe operation of the system and equipment. The purpose of this step is to translate intelligent decision-making into precise physical control actions, achieving high-precision execution of power allocation and safety protection.

[0052] Furthermore, the power balancing adjustment model employs a deep neural network based on the Transformer architecture. The overall model architecture includes an input embedding layer, a position encoding layer, a 6-layer encoder layer, and an output layer. The input embedding layer converts the power demand data and system state information of each charging port into a 512-dimensional feature vector, using linear transformation and LayerNorm normalization. The position encoding layer generates position encoding vectors using sine and cosine functions, which are added to the input feature vectors and then input to the encoder layer. Each encoder layer contains a multi-head attention sublayer and a feedforward neural network sublayer, both of which use residual connections and layer normalization. The multi-head attention mechanism uses 8 attention heads, each with a dimension of 64, to capture long-range dependencies in the input sequence by parallel computation of attention weights in different subspaces. The feedforward neural network contains two linear transformation layers with a hidden layer dimension of 512. The activation function is ReLU, and the dropout ratio is set to 0.1 to prevent overfitting. The output layer uses two independent linear classifiers to output a power allocation weight vector and an adjustment policy vector, respectively. The weight vectors are normalized using a softmax function to ensure that the sum of the weights for each port is 1.

[0053] The training dataset creation process begins with historical data collection, gathering charging data from various charging scenarios, including passenger vehicle charging data, commercial vehicle charging data, and energy storage device charging data, covering charging curves for different battery types such as lithium iron phosphate batteries, ternary lithium batteries, and lithium titanate batteries. Next, charging performance data under different environmental conditions is collected, including charging efficiency data within a temperature range of -20℃ to 50℃, insulation performance data under different humidity conditions, and heat dissipation performance data at different altitudes. Finally, system efficiency data at different power levels is collected, including conversion efficiency, heat loss, and harmonic distortion data at different power outputs such as 25kW, 50kW, 75kW, 100kW, and 125kW.

[0054] The data preprocessing stage employs a multi-step cleaning strategy. First, outliers are detected using statistical methods, with the 3σ criterion used to identify data points outside the normal range. These points are then repaired or deleted using interpolation methods. Next, data integrity is checked, identifying missing values ​​and filling them using methods such as forward imputation, backward imputation, or mean imputation. Then, data consistency verification is performed, checking the logical relationships between data from different sensors and identifying and correcting data conflicts. Finally, data normalization is conducted, using Z-score standardization to unify data from different dimensions to the same scale range.

[0055] The feature engineering phase constructs a multi-dimensional input feature vector, encompassing four dimensions: time features, state features, environmental features, and historical features. Time features include periodic encoding of time information such as moment, date, and season. State features include real-time status information such as power demand at each port, remaining battery capacity, charging current, and charging voltage. Environmental features include environmental parameters such as ambient temperature, humidity, and atmospheric pressure. Historical features extract historical information such as power change trends and charging behavior patterns over the past 24 hours using a sliding window method. The output label vector contains the optimal power allocation scheme and the corresponding adjustment strategy. The power allocation scheme is represented by the power allocation ratio at each port, and the adjustment strategy is represented by a combination of adjustment rate and direction.

[0056] The dataset was partitioned using stratified sampling to ensure a consistent proportion of samples across the training, validation, and test sets. The training set comprised 70% of the total data and was used for model parameter learning; the validation set comprised 20% and was used for hyperparameter tuning and model selection; and the test set comprised 10% and was used for final performance evaluation. To prevent data leakage, the data was partitioned chronologically, ensuring that the test set's timeframe was later than that of the training and validation sets.

[0057] It should be noted that the first key technical concept of this invention is a power demand feature extraction mechanism based on a sliding time window. Traditional charging pile systems typically employ static power allocation strategies, which cannot adapt to the dynamic changes in user charging demand. This invention, by adaptively adjusting the time window length, can accurately capture power change patterns at different time scales, improving the accuracy of power demand prediction. The sliding window mechanism combined with Fast Fourier Transform (FFT) enables real-time analysis of the frequency domain characteristics of the power signal, providing a scientific basis for dynamic power allocation. Compared to traditional fixed-time interval sampling, this method better adapts to the nonlinear changes in power demand during charging, significantly improving the system's response capability to sudden power demands.

[0058] The second key technical approach is a power allocation strategy combining a two-layer optimization model and game theory. Traditional methods often consider only a single optimization objective, making it difficult to balance the conflict between system efficiency and user experience. This invention constructs an upper-layer optimization model with the objective of maximizing total charging efficiency and a lower-layer optimization model with the objective of minimizing charging time. Stackelberg game theory is used to describe the competition and cooperation between the two optimization objectives. This two-layer optimization architecture can ensure overall system efficiency while also considering the user's charging time requirements, achieving a globally optimal solution for multi-objective optimization. Compared to traditional single-objective optimization methods, two-layer optimization can better handle objective conflicts and avoid local optima traps.

[0059] The third key technological approach is an intelligent power balance regulation model based on the Transformer architecture. Traditional control methods mainly rely on manually set control rules, which are difficult to adapt to complex and ever-changing charging scenarios. This invention employs a multi-head attention mechanism that can process information from multiple charging ports in parallel, automatically learn the mutual influence relationships between ports, and achieve intelligent coordination of power allocation between ports. The dynamic adjustment mechanism of attention weights can adaptively adjust the importance of each port according to the system's operating state, improving the flexibility and adaptability of the control strategy. Compared with traditional rule-based control, this deep learning method has stronger generalization ability and robustness.

[0060] The fourth key technological approach is a multi-dimensional integrated heat dissipation and efficiency optimization mechanism. Traditional charging piles often treat heat dissipation and power control as independent subsystems, lacking a unified coordination mechanism. This invention integrates the air-cooling and liquid-cooling heat dissipation systems with the power conversion unit, dynamically adjusting the power output through a temperature compensation coefficient to achieve synergistic optimization of heat dissipation efficiency and charging efficiency. This synergistic mechanism maximizes the system's power output capacity while ensuring safe operation of the equipment, significantly improving the power density and reliability of the charging pile.

[0061] The synergistic effect of these four key technological approaches forms a complete intelligent power management system. Sliding time window feature extraction provides an accurate data foundation for two-layer optimization, which in turn provides a scientific decision-making framework for the Transformer model. The Transformer model provides an intelligent control strategy for thermal optimization, which in turn provides stable operation assurance for the entire system. This synergistic effect of multi-technology integration enables the system to maintain efficient and stable operation in complex and ever-changing real-world application environments. Compared with traditional methods, it significantly improves power utilization, charging efficiency, and system reliability, providing crucial technical support for the development of fast-charging technology for electric vehicles.

[0062] Specifically, the principle of this invention is as follows: The key to solving the problem of unreasonable power allocation in multi-port charging pile systems lies in establishing a complete dynamic power balance optimization mechanism. This mechanism achieves real-time power adjustment and system optimization based on data-driven intelligent algorithms. First, this invention continuously monitors and extracts features of the power demand of each charging port using a sliding time window method. This accurately identifies the temporal and frequency characteristics of power changes, providing a reliable data foundation for subsequent power prediction and allocation strategy formulation. This time series analysis method has stronger adaptability and prediction accuracy compared to traditional static power allocation. Second, the dual-layer optimization model constructed in this invention avoids conflicts in multi-objective optimization by treating the two objectives of maximizing system efficiency and minimizing charging time in a hierarchical manner. The upper-layer model focuses on the overall system performance, while the lower-layer model focuses on user charging needs. A game-theoretic algorithm is used to achieve a coordinated balance between the two optimization objectives, ensuring the scientific validity and feasibility of the power allocation scheme. Finally, this invention employs a multi-head attention network based on the Transformer architecture as the power balance regulation model. This model can process multi-dimensional input information and capture the complex relationships between charging ports. Through the dynamic adjustment mechanism of attention weights, the system can adaptively adjust the power allocation strategy according to the real-time charging status and priority requirements, realizing the transformation from passive response to active prediction. This intelligent control method ensures that the system can maintain the optimal power allocation state under various operating conditions.

[0063] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0064] In this embodiment, the power distribution unit uses the Huichuan Technology MD380T550G DC inverter with a rated power of 500kW, an input voltage range of 380V to 690V, an output voltage range of 200V to 1000V, a power factor greater than 0.95, and a total harmonic distortion of less than 5%. The power conversion unit uses the BYD Semiconductor DM4.0-1200A IGBT power module with a switching frequency of 20kHz, a peak current of 1200A, and integrated temperature sensor and overcurrent protection. The charging interface unit uses the NARI CHARGEV-125K DC charging interface with a maximum single-port output power of 125kW, an output voltage range of 200V to 1000V, an output current range of 0 to 125A, and an IP54 protection rating.

[0065] The air-cooled radiator in the heat dissipation unit uses a Jiamusi Electric DCF450-4500 centrifugal fan with an air volume of 4500. The system operates at a constant speed of 1200 Pa with a static pressure of 1200 Pa and a noise level below 65 dB. The liquid-cooled radiator uses a Huadian Heavy Industry HE-BR110 plate heat exchanger with a heat exchange area of ​​3.2 square meters. The maximum flow rate is 15 The operating pressure is 1.6 MPa. The heat-conducting copper tube is made of pure copper, with a diameter of 12 mm, a wall thickness of 1.5 mm, and a thermal conductivity of 401 W / m·K.

[0066] The current detection unit uses the Acrel Hall effect current sensor AHKC-EKAA, with a measurement range of 0 to 500A, an accuracy class of 0.2, a response time of less than 1μs, and an operating temperature range of -40℃ to 85℃. The voltage regulation unit uses the Huada Semiconductor HC32F460 digital signal processor with a main frequency of 200MHz, an integrated 12-bit analog-to-digital converter, a sampling frequency of 1MHz, and a voltage regulation accuracy of ±0.5%.

[0067] The wireless communication module in the communication unit adopts the Huawei MH5000-31 5G industrial module, supporting the Sub-6GHz frequency band, with a downlink peak rate of 2.3Gbps and an uplink peak rate of 230Mbps, and an operating temperature range of -30℃ to 70℃. The wired communication module adopts the Realtek RTL8111H Ethernet controller, supporting Gigabit Ethernet, a transmission rate of 1Gbps, and TCP / IP protocol stack.

[0068] The insulation monitoring unit uses the Acrel AIM-T300 insulation monitoring device, with a measurement range of 1kΩ to 100MΩ, a response time of 2s to 30s, an operating voltage range of AC85V to 264V, and a measurement accuracy of ±5%. The control chip uses the Rockchip RK3588 application processor, based on an ARM Cortex-A76 quad-core architecture with a main frequency of 2.4GHz, and integrates a Neural Processing Unit (NPU) to support machine learning acceleration. The data storage device uses a Yangtze Memory Technologies PC005 NVMe solid-state drive with a storage capacity of 1TB, a read speed of 3500MB / s, a write speed of 2900MB / s, and supports AES 256-bit hardware encryption.

[0069] The specific implementation of the dynamic power balance optimization module in this embodiment is described in detail below: The specific implementation of step S01 is the same as described above, and will not be repeated here.

[0070] The specific implementation of step S02 involves using a sliding time window method to perform time-series analysis and feature extraction on the power demand data. The calculation process for power change feature extraction is as follows: In the formula, Peak power The sampling point number and , For the first time window One power sampling point, This represents the maximum power within the window. This represents the minimum power within the window. The average power is calculated as follows: In the formula, This is the average power value. This represents the total number of sampling points within the time window. The power fluctuation amplitude is calculated as follows: In the formula, This represents the power fluctuation amplitude, i.e., the standard deviation. The current is acquired in real time using a current detection unit, with a sampling frequency of 1Hz. The time window length is determined based on the time window length, with a default value of 300 sampling points corresponding to a 5-minute window length.

[0071] The specific implementation of step S03 is to establish a mathematical model for the power allocation strategy based on multi-objective optimization theory. The mapping relationship matrix is ​​constructed as follows: In the formula, It is a 4×4 mapping matrix. For the first The charging port is at the first Allocation weighting coefficients under each power allocation level The weighting coefficient constraints are expressed as follows: , In the formula, the constraints ensure that the sum of the weight coefficients of each port under the same power allocation level is equal to 1. The weight coefficients are obtained by calculating using the analytic hierarchy process (AHP), including step 1: constructing a judgment matrix to evaluate the relative importance of each port; step 2: calculating the maximum eigenvalue and corresponding eigenvector of the judgment matrix; and step 3: performing a consistency check to ensure the rationality of the judgment matrix.

[0072] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0073] The specific implementation of step S05 involves constructing a two-layer optimization model and using a game-theoretic algorithm to solve for the optimal power allocation scheme. The total charging efficiency maximization function is expressed as follows: In the formula, This represents the overall charging efficiency value of the system. Number the port and , For the first Port weight parameters for each port. For the first Port efficiency parameters for each port. For the first Port power parameters of each port This is an efficiency correction factor. This is the temperature compensation coefficient. For the first Temperature monitoring values ​​at each port. The efficiency correction factor is calculated as follows: In the formula, the sinusoidal function relationship reflects the nonlinear coupling effect between port weight and efficiency. The temperature compensation coefficient is calculated as follows: In the formula, The reference temperature is set to 25℃. The temperature compensation parameter is set to 10℃. The function for minimizing charging time is expressed as follows: In the formula, To complete all charging tasks in the shortest time, For the first Time weighting parameters for each port For the first Battery capacity parameters for each port, For the first The charging efficiency coefficient of each port. This is the charging efficiency factor. This represents the load balancing coefficient. The charging efficiency factor is calculated as follows: In the formula, the logarithmic function reflects the synergistic effect of time weighting and battery capacity. The load balancing coefficient is calculated as follows: In the formula, The standard power is set at 100kW, and a cosine function relationship ensures balanced power distribution. The monitoring data from the current detection unit is obtained through normalization processing. Efficiency monitoring data from the power conversion unit. The power output measurement value originates from the charging interface unit. The charging priority information is received from the communication unit. The remaining battery capacity data is derived from the battery status information. The value is obtained in real time through the battery management system and ranges from 0.8 to 0.95.

[0074] The specific implementation of step S06 involves inputting the optimization solution into a pre-trained power balance adjustment model to generate precise control commands. The attention weight adjustment function is expressed as follows: when hour, ;when hour, ;when hour, ;when hour, In the formula, The adjusted attention weight parameters, The base weight value is set to 0.25. Set the maximum weight value to 1.0. This is the weighting adjustment factor. The weighting adjustment factor is calculated as follows: In the formula, The normalization coefficients are all set to 1. For the charging port load rate, For system power utilization, The urgency of charging is indicated. The load rate is calculated as follows: In the formula, The current total output power is equal to , The total rated power of the system is given. The power utilization rate is calculated as follows: In the formula, This represents the current total output power. This is the system's rated power. Among them, It is obtained through real-time calculation based on power monitoring data from each port. It is obtained through real-time calculation based on the total power monitoring data of the system. The data is derived from the user charging urgency rating received by the communication unit, with a value range of 0 to 1.

[0075] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0076] The principles and effects of each formula are explained below. The power variation feature extraction formula is based on statistical principles. By calculating peak power, average power, and standard deviation, it comprehensively depicts the power variation pattern. Compared to traditional single-index analysis methods, this multi-dimensional feature extraction method can accurately identify the fluctuation characteristics and trends of power demand, providing a richer data foundation for subsequent power prediction and optimization, and significantly improving the accuracy and adaptability of power allocation decisions. The mapping relationship matrix construction formula uses matrix form to express the allocation weight of each port under different power levels. Constraints ensure the rationality and consistency of weight allocation. Compared to the traditional fixed-ratio allocation method, this dynamic weight matrix can flexibly adjust the power allocation ratio of each port according to the actual operating status, achieving optimized power resource allocation and load balancing. The total charging efficiency maximization function integrates port weights, efficiency parameters, power parameters, efficiency correction factors, and temperature compensation coefficients. The efficiency correction factor... A sine function is used to represent the nonlinear coupling relationship between port weights and efficiency, with a temperature compensation coefficient. The influence of temperature on charging efficiency is considered using a Gaussian function. Compared to traditional linear efficiency calculation methods, this composite function more accurately reflects the synergistic effect of multiple factors during actual charging, significantly improving the calculation accuracy and optimization effect of the overall system charging efficiency. The charging time minimization function determines the overall charging completion time by taking the maximum value of the charging time at each port, where the charging efficiency factor... A logarithmic function is used to reflect the synergistic effect of time weighting and battery capacity, and the load balancing coefficient is used. A cosine function is employed to ensure balanced power distribution. Compared to the traditional simple time accumulation method, this function comprehensively considers the differences in charging characteristics and priority requirements of each port, achieving global optimization of charging time. The attention weight adjustment function adopts a piecewise function form, selecting corresponding adjustment strategies according to different ranges of the weight adjustment factor, including four adjustment modes: linear growth, square root growth, logarithmic growth, and exponential decay. Compared to the traditional fixed weight allocation method, this adaptive adjustment mechanism can dynamically optimize the allocation of attention resources according to the system operating state, significantly improving the response speed and control accuracy of the power balance adjustment model, and realizing intelligent adaptive control of the charging system.

[0077] To better understand and implement this invention, the following is a specific application scenario of embodiment 2: A split-type multi-port high-power DC charging pile system includes a main control cabinet and four charging terminals. The main control cabinet is 25 meters away from the charging terminals, and the total rated power of the system is 500kW. The system hardware configuration uses the Huichuan Technology MD380T550G DC inverter as the power distribution unit, with a rated power of 500kW. The power conversion unit uses the BYD Semiconductor DM4.0-1200A IGBT power module, with a switching frequency set to 20kHz. All four charging interface units use the NARI CHARGEV-125K DC charging interface, with a maximum output power of 125kW per port. The cooling system is equipped with a Jiamusi Motor DCF450-4500 centrifugal fan and a Huadian Heavy Industry HE-BR110 plate heat exchanger. The control system uses a Rockchip RK3588 processor with a main frequency of 2.4GHz and an integrated neural network processing unit (NPU). Data storage uses a Yangtze Memory Technologies PC005 solid-state drive with a capacity of 1TB.

[0078] The technical team selected four different types of electric vehicles for charging tests, including the BYD Han EV, Tesla Model Y, NIO ES6, and Li Auto ONE. The test environment was set for hot summer weather, with an ambient temperature of 35℃ and relative humidity of 85%. The charging tests were conducted during peak hours from 10:00 AM to 4:00 PM, for a total of 6 hours of continuous operation. Vehicle parameters and charging demand information are shown in Table 1. Table 1. Charging Parameters of Test Vehicles

[0079] The technical team first performed system initialization configuration, setting the time window length to 5 minutes and the power sampling frequency to 1Hz. The mapping matrix was pre-configured based on the vehicle battery type and charging characteristics, with high-priority ports having a weight coefficient of 0.4, medium-priority ports 0.3, and low-priority ports 0.2. After the system started running, the dynamic power balance optimization module began monitoring the power demand of each charging port in real time.

[0080] During the charging process, the power change data collected by the system showed obvious time-varying characteristics. The BYD Han EV maintained a power demand of around 110kW for the first 30 minutes, gradually decreasing to 85kW as the battery temperature rose and the state of charge increased. The Tesla Model Y's power demand curve was relatively stable, remaining between 90kW and 105kW. The NIO ES6, due to its liquid-cooled battery pack, exhibited less fluctuation in power demand, remaining stable at around 95kW. The Li Auto ONE, as a plug-in hybrid, had a lower charging power demand due to its small-capacity battery, maintaining between 35kW and 40kW.

[0081] The sliding time window method was used to analyze and process the power demand of each port, extracting characteristic parameters such as peak power, average power, and fluctuation amplitude. The system calculated that the peak power of the BYD Han EV port was 25kW, the average power was 97.5kW, and the power fluctuation amplitude was 8.3kW. The peak power of the Tesla Model Y port was 15kW, the average power was 97.5kW, and the power fluctuation amplitude was 4.2kW. The peak power of the NIO ES6 port was 8kW, the average power was 95kW, and the power fluctuation amplitude was 2.1kW. The peak power of the Li Auto ONE port was 5kW, the average power was 37.5kW, and the power fluctuation amplitude was 1.5kW.

[0082] like Figure 4 As shown, a general power prediction model based on a long short-term memory neural network is used to predict power demand over the next 15 minutes. This model is trained based on historical charging data. The prediction results show that the power demand of the BYD Han EV will begin to decrease significantly after 45 minutes, dropping from 110kW to 70kW. The power demand of the Tesla Model Y remains relatively stable throughout the charging process. The power demand of the NIO ES6 will decrease slightly after 60 minutes. Due to its smaller battery capacity, the Li Auto ONE is expected to complete charging after 75 minutes.

[0083] In the solution process of the two-level optimization model, the upper-level optimization aims to maximize the total charging efficiency of the system, while the lower-level optimization aims to minimize the charging time. The calculated port weight parameters are 0.28, 0.26, 0.25, and 0.21, and the port efficiency parameters are 0.94, 0.92, 0.93, and 0.89. The temperature compensation coefficients, considering the actual temperature of each port, are 0.89, 0.91, 0.93, and 0.95. The charging efficiency factor, comprehensively considering time weight and battery capacity, is 0.92, 0.91, 0.93, and 0.88. The dynamic power allocation results are shown in Table 2. Table 2 Power Allocation Scheme for Each Time Period

[0084] The power balance adjustment model uses a multi-head attention network based on the Transformer architecture to generate power adjustment commands. In the early stages of charging, due to the high power demands of all four ports, the attention weight parameters are evenly distributed at 0.25, 0.25, 0.25, and 0.25 respectively. As the Li Auto ONE completes charging at 75 minutes, the system automatically adjusts the attention weight parameters, allocating more resources to the remaining three ports, with the weight parameters adjusted to 0.35, 0.33, and 0.32. When the total system power exceeds 80% of the rated power, the attention weight shifts towards higher-priority ports, with the weight parameter for the BYD Han EV increasing to 0.40.

[0085] like Figure 5 As shown, the heat dissipation system exhibits excellent temperature control throughout the charging process. The temperature of the power conversion unit is consistently kept below 65℃, the air intake temperature of the air-cooled radiator is 35℃, the exhaust temperature is 42℃, and the temperature rise is controlled within 7℃. The coolant inlet temperature of the liquid-cooled radiator is 30℃, the outlet temperature is 38℃, and the heat exchange efficiency reaches 92%. Temperature monitoring at each charging port shows that port 1 has the highest temperature at 63℃, port 2 at 61℃, port 3 at 59℃, and port 4 at 55℃, with uniform temperature distribution within a safe range.

[0086] The current detection unit monitors the charging current changes at each port in real time. The charging current at port 1 gradually decreases from an initial 125A to 65A, with a smooth change and no abnormal fluctuations. The charging current at port 2 remains between 100A and 115A, with minimal fluctuations. The charging current at port 3 remains stable between 95A and 120A. The charging current at port 4 is between 30A and 40A, consistent with its battery capacity characteristics. The voltage regulation unit precisely controls the output voltage of each port within ±0.5% of the target value, demonstrating good voltage stability.

[0087] The communication system operated stably throughout the test. The data transmission latency of the 5G wireless communication module remained below 20ms, and the packet loss rate of the Ethernet wired communication was zero. Communication between the system and the vehicle's battery management system was normal, enabling real-time acquisition of battery status information and adjustments to the charging strategy. The insulation monitoring unit continuously monitored the insulation resistance value of the charging circuit, with measurements consistently above 10MΩ and leakage current less than 1mA, ensuring charging safety.

[0088] Throughout the test, the system's data storage device recorded over 2GB of operational data, including power variation curves, temperature trends, efficiency statistics, and fault diagnosis data. Data analysis results show that the dynamic power balance optimization module achieved a prediction accuracy of over 94%, with power allocation errors controlled within 3%. The multi-head attention mechanism's weight adjustment response time is less than 100ms, enabling rapid adaptation to changes in charging demand.

[0089] Test results showed that the charging times for the four vehicles were 89 minutes, 95 minutes, 88 minutes, and 74 minutes, respectively. The overall charging efficiency of the system reached 95.6%, an improvement of 2.3 percentage points compared to the traditional fixed power distribution method. The power utilization rate remained above 92%, avoiding the problem of some ports being idle in traditional systems. No abnormalities such as overheating protection or overcurrent protection occurred during the charging process, and the system operated stably and reliably.

[0090] This invention represents a significant technological advancement over traditional charging pile systems. Traditional systems employ a fixed power allocation method, failing to dynamically adjust power output based on actual charging demand, leading to wasted power resources and low charging efficiency. This invention utilizes a sliding time window method to analyze power demand changes in real time, accurately grasping the dynamic characteristics of the charging process and providing a scientific basis for intelligent power allocation. A dual-layer optimization model organically combines efficiency maximization and time minimization, using game theory to solve for the optimal power allocation scheme, overcoming the limitations of traditional single-objective optimization. The power balance adjustment model based on the Transformer architecture employs a multi-head attention mechanism, enabling parallel processing of information from multiple charging ports and automatically learning the mutual influence relationships between ports, achieving intelligent coordination of power allocation between ports. The split design separates the power conversion unit from the charging interface unit, reducing cable loss and improving system reliability. A multi-dimensional integrated cooling system organically combines air cooling and liquid cooling, achieving synergistic optimization of heat dissipation efficiency and charging efficiency, significantly improving the system's power density and operational stability.

[0091] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0092] Table 3. Variable Explanation Table

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A split-type high-power DC charging pile system, comprising a power distribution unit, a power conversion unit, a charging interface unit, a heat dissipation unit, a current detection unit, a voltage regulation unit, a communication unit, an insulation monitoring unit, and a control chip, characterized in that, The control chip incorporates a dynamic power balance optimization module. This module monitors the actual charging power demand and battery status information of each charging port. Based on the sliding time window method, it performs segmented analysis on the power demand data of each charging port, extracts power change characteristics, calculates the theoretical optimal power allocation ratio of each charging port according to a preset power allocation strategy, establishes a mapping relationship matrix between power demand and allocation ratio, and uses a charging power prediction model to predict the power demand of each charging port in the future time period, generating a power demand prediction curve. It constructs an upper-level optimization model with the goal of maximizing total charging efficiency and a lower-level optimization model with the goal of minimizing charging time. The two-level optimization problem is solved using a game theory algorithm to obtain the optimal power allocation scheme. The optimal power allocation scheme is input into the power balance adjustment model. By adjusting the attention weight parameters of the power balance adjustment model, a power adjustment command is generated. Based on the power adjustment command, the power conversion unit and voltage regulation unit are controlled to achieve precise allocation and dynamic adjustment of power at each charging port.

2. The system according to claim 1, characterized in that, The sliding time window method involves a sliding time period with a base length of 5 minutes. The window length is adaptively adjusted according to the frequency of charging power changes. When the power change frequency is higher than 0.1Hz, the window length is shortened to 3 minutes. When the power change frequency is lower than 0.05Hz, the window length is extended to 8 minutes. Each time window contains continuous power sampling points with a sampling frequency of 1Hz.

3. The system according to claim 2, characterized in that, The power variation characteristics are specifically the power data feature parameters extracted through statistical analysis within each time window. The power peak value is the difference between the maximum and minimum power values ​​within the time window, the power average value is the arithmetic mean of all power sampling points within the time window, and the power fluctuation amplitude is the standard deviation of the power data.

4. The system according to claim 3, characterized in that, The mapping matrix is ​​specifically a 4×4 two-dimensional array structure. The row index represents the four charging ports, the column index represents the four power allocation levels, and the matrix element value represents the allocation weight coefficient of the corresponding port under the corresponding power allocation level. The allocation weight coefficient ranges from 0 to 1, and the sum of the allocation weight coefficients of all ports under the same power allocation level is equal to 1.

5. The system according to claim 4, characterized in that, The aforementioned two-layer optimization problem is specifically a nested optimization structure comprising an upper-layer optimization model and a lower-layer optimization model. The upper-layer optimization model aims to maximize the total charging efficiency, while the lower-layer optimization model aims to minimize the charging time. The upper-layer and lower-layer optimization models are coupled through power allocation variables.

6. The system according to claim 5, characterized in that, The total charging efficiency maximization function comprehensively considers the power conversion efficiency of each charging port, the heat dissipation efficiency of the heat dissipation unit, and the battery charging efficiency. The inputs include port weight parameters, port efficiency parameters, port power parameters, efficiency correction factors, and temperature compensation coefficients. The output is the overall charging efficiency value of the system.

7. The system according to claim 6, characterized in that, The charging time minimization function takes into account the differences in charging rates at different charging ports and the impact of charging strategies. The inputs include time weight parameters, battery capacity parameters, charging power parameters, charging efficiency factors, and load balancing coefficients. The output is the shortest time to complete all charging tasks.

8. The system according to claim 7, characterized in that, The power balance adjustment model is specifically based on a multi-head attention network of the Transformer architecture, including 6 encoder layers, each containing a multi-head attention mechanism and a feedforward neural network. The multi-head attention mechanism has 8 heads, each attention head has a dimension of 64, and the hidden layer dimension of the feedforward neural network is 512.

9. The system according to claim 8, characterized in that, The steps for establishing the training dataset of the power balance adjustment model are as follows: collecting historical charging data under different charging scenarios, preprocessing the raw data including data cleaning, outlier detection and data normalization, constructing input feature vectors and output label vectors, and dividing the dataset into training set, validation set and test set in a ratio of 7:2:

1.

10. The system according to claim 9, characterized in that, The training steps of the power balance adjustment model are as follows: the Adam optimizer is used to update the parameters, the initial learning rate is set to 0.001, the batch size is set to 32, the number of training rounds is set to 100, the prediction error is calculated using the cross-entropy loss function, and an early stopping strategy is adopted to prevent overfitting when the validation set loss does not decrease for 5 consecutive rounds.

Images