Distributed charging pile group dynamic load balancing and power optimization allocation system and method

Abstract

The present application relates to electric vehicle charging scheduling technical field, especially to distributed charging pile group dynamic load balancing and power optimization allocation system and method, including: cloud coordinator obtains the space correlation data between piles containing geographical distance and traffic similarity and fuses with historical and weather data, inputs into the spatiotemporal graph neural network prediction model, outputs probability distribution prediction value and prediction confidence interval and issues, edge control node constructs multi-objective optimization model based on prediction uncertainty, generates dynamic allocation table and backup scheme under capacity constraint to minimize charging time and maximize renewable energy consumption as the target, issues allocation table to execute power allocation and obtains actual power feedback, when the actual power exceeds the confidence interval range, switch backup scheme to preferentially reduce flexibility load. The present application effectively improves the deficiency that single-point prediction lacks defense, realizes reliable global dynamic load balancing under the premise of guaranteeing the safe operation of power grid transformer.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging scheduling technology, and in particular to a distributed charging pile group dynamic load balancing and power optimization allocation system and method. Background Technology

[0002] With the widespread adoption of new energy vehicles, the large-scale integration of distributed charging pile clusters has brought highly complex spatiotemporal coupled loads to regional power distribution networks. To ensure the stable operation of transformers and improve the service efficiency of charging stations, the energy management network needs to integrate multi-dimensional data such as historical charging curves, meteorological conditions, and road network traffic to construct predictive models. By anticipating future power demand trends, the system formulates dynamic power allocation strategies to achieve load balancing and resource scheduling of the charging pile clusters within the limited power capacity of the charging stations.

[0003] However, existing allocation methods mostly rely on deterministic single-point predictions and lack quantitative assessment of prediction uncertainty. When faced with extreme load fluctuations caused by sudden traffic flow, the predictions are prone to failure and it is difficult to avoid the physical safety risks of transformer overload. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a distributed charging pile group dynamic load balancing and power optimization allocation system and method, which aims to improve the problem that the existing technology is prone to failure under extreme load fluctuations and causes safety risks such as transformer overload.

[0005] In a first aspect, the present invention provides the following technical solution: a method for dynamic load balancing and power optimization allocation of distributed charging pile groups, comprising:

[0006] S1. The cloud coordinator acquires historical charging curves, weather data, and spatial correlation data between charging piles, wherein the spatial correlation data between charging piles includes the geographical distance and traffic similarity between each charging pile unit.

[0007] S2. The cloud coordinator integrates the historical charging curve, the weather data, and the spatial correlation data between charging piles into the charging load prediction model constructed based on the spatiotemporal graph neural network, outputs the multi-period power demand probability distribution prediction value and the corresponding prediction confidence interval of each charging pile unit, and sends it to the edge control node.

[0008] S3. The edge control node constructs a multi-objective optimization model based on the prediction uncertainty represented by the multi-time period power demand probability distribution prediction value. Under the transformer capacity constraint, the model aims to minimize the charging completion time and maximize the renewable energy consumption rate, and generates a dynamic power allocation table and backup scheme.

[0009] S4. The edge control node sends the dynamic power allocation table to each charging pile unit to perform charging power allocation.

[0010] S5. The edge control node obtains the actual power of the charging pile unit during the execution process as feedback. When the actual power exceeds the predicted confidence interval, the edge control node switches to the backup scheme and prioritizes reducing the charging power belonging to the flexibility load.

[0011] Preferably, in step S1, the step of obtaining historical charging curves, weather data, and spatial correlation data between charging piles includes:

[0012] Time step alignment and missing value imputation are performed on the collected raw historical charging data and raw weather data to generate a time series feature matrix;

[0013] The shortest physical travel distance between each charging pile unit is calculated based on the road network topology, and the overlap characteristics of interest points around each charging pile unit and the historical traffic flow tidal characteristics are extracted.

[0014] The traffic mapping relationship is calculated using the cosine similarity algorithm, and the shortest physical travel distance is weighted and fused with the traffic mapping relationship to generate the spatial correlation data between the piles.

[0015] Preferably, in step S2, the step of inputting the data into the charging load prediction model constructed based on a spatiotemporal graph neural network includes:

[0016] Causal convolution operations are performed on the historical charging curves and the weather data using a temporal convolutional network layer to extract the temporal evolution features of the load.

[0017] The spatial correlation data between the piles is transformed into an adjacency matrix, and the spatial node features of adjacent charging pile units are aggregated using graph convolutional network layers to extract load spatial coupling features.

[0018] An adaptive weight allocation is performed on the load temporal evolution features and the load spatial coupling features using a multi-head attention mechanism, and a latent variable feature vector is output.

[0019] Preferably, in step S2, the step of outputting the predicted power demand probability distribution of each charging pile unit over multiple time periods and the corresponding prediction confidence interval includes:

[0020] A Gaussian mixture distribution mapping layer is introduced into the charging load prediction model to calculate and generate the expected value, variance, and mixture weights that characterize the future probability density.

[0021] Based on the expected value, variance, and mixed weights, a continuous probability density function is reconstructed, and samples are taken in the discrete time dimension to generate the predicted value of the multi-time period power demand probability distribution.

[0022] The probability density function is integrally truncated at preset upper and lower quantiles using a quantile regression algorithm to extract the upper and lower bounds of the prediction at a given confidence level, thus forming the prediction confidence interval.

[0023] Preferably, in step S3, the step of constructing the multi-objective optimization model includes:

[0024] Calculate the statistical moment information of the predicted power demand probability distribution for the multi-period time period, and define a continuous set of uncertainties containing the worst load fluctuations within the polyhedral boundary;

[0025] A queuing theory model is introduced to construct a first objective function that minimizes the charging completion time, and a renewable energy generation curve is extracted to construct a second objective function that maximizes the renewable energy absorption rate.

[0026] The power flow equations of the distribution network and the transformer capacity constraints are established, and then coupled with the first objective function, the second objective function, and the set of uncertainties to complete the construction of the multi-objective optimization model.

[0027] Preferably, in step S3, the step of generating the dynamic power allocation table and the backup scheme includes:

[0028] The multi-objective optimization model is decomposed into a main problem and sub-problems using a column and constraint generation algorithm, and the optimal power allocation sequence under the nominal prediction state is obtained through alternating iterations.

[0029] The optimal power allocation sequence is discretized and sliced ​​according to the control period to generate the dynamic power allocation table for daily charging and discharging scheduling.

[0030] During the sub-problem solving stage, extreme load deviation scenarios are extracted based on the prediction uncertainty, and a derating power instruction set for the extreme scenario is pre-calculated to form the backup scheme.

[0031] Preferably, in step S4, the step of performing charging power allocation includes:

[0032] The local controller of each charging pile unit performs protocol parsing on the dynamic power allocation table and extracts the target output voltage and current setpoints for the current time slot.

[0033] The duty cycle signal of the power switch in the bidirectional converter circuit is dynamically adjusted by the pulse width modulation module according to the target output voltage and current set values.

[0034] The underlying voltage and current dual closed-loop control logic is activated to smooth transient power jitter during the tracking of the dynamic power allocation table and complete the physical power injection.

[0035] Preferably, in step S5, the step of preferentially reducing the charging power belonging to the flexible load includes:

[0036] Real-time acquisition of battery state of charge, battery capacity, and user's expected departure time for connected electric vehicles; and construction of a multi-dimensional flexible evaluation function.

[0037] The real-time flexibility index of each charging task is calculated based on the multidimensional flexibility evaluation function. Electric vehicles are sorted from high to low according to the real-time flexibility index to generate a load reduction priority queue.

[0038] Power derating instructions are issued sequentially according to the order of the load reduction priority queue until the total actual power falls back to the range of the predicted confidence interval.

[0039] Secondly, this invention provides the following technical solution: a distributed charging pile group dynamic load balancing and power optimization allocation system, comprising:

[0040] The data acquisition module is used to acquire historical charging curves, weather data, and spatial correlation data between charging piles. The spatial correlation data between charging piles includes the geographical distance and traffic similarity between each charging pile unit.

[0041] The load forecasting module is used to fuse the historical charging curve, the weather data and the spatial correlation data between the charging piles into the charging load forecasting model constructed based on the spatiotemporal graph neural network, and output the multi-period power demand probability distribution forecast value and the corresponding prediction confidence interval of each charging pile unit.

[0042] The optimization modeling module is used to construct a multi-objective optimization model based on the prediction uncertainty represented by the predicted value of the multi-time period power demand probability distribution. Under the transformer capacity constraint, the model aims to minimize the charging completion time and maximize the renewable energy consumption rate, and generates a dynamic power allocation table and backup scheme.

[0043] A power allocation module is used to send the dynamic power allocation table to each charging pile unit to perform charging power allocation.

[0044] The feedback control module is used to obtain the actual power of the charging pile unit during the execution process as feedback. When the actual power exceeds the predicted confidence interval, it switches to the backup scheme and prioritizes reducing the charging power of the flexible load.

[0045] The present invention has the following beneficial effects:

[0046] 1. In this invention, a cloud-edge collaborative architecture combining spatiotemporal graph neural network and multi-objective optimization is adopted. By fusing multi-dimensional data, a probability distribution prediction value with confidence interval is output, and an optimization model including severe fluctuations is constructed accordingly. This improves the shortcomings of traditional single-point prediction in dealing with sudden loads, and realizes global dynamic load balancing of distributed charging pile groups under the premise of ensuring the safety of power grid equipment.

[0047] 2. In this invention, a queuing theory model and renewable energy generation curves are introduced in the power allocation stage. The optimal allocation sequence is obtained by alternating and iteratively using column and constraint generation algorithms. On the basis of strictly adhering to the power flow and capacity constraints of the distribution network, the synergistic optimization of minimizing the charging completion time and maximizing the green electricity consumption rate is achieved, thereby improving the overall economic benefits of the power station.

[0048] 3. In this invention, a safety fallback mechanism with the predicted confidence interval as the threshold is designed. When the actual power exceeds the limit, it switches to the pre-calculated backup scheme. By extracting the battery status and off-site time to construct a flexible evaluation function, the flexible load power is reduced step by step in descending order of flexibility index, avoiding global power outage and ensuring the user's core charging needs to the greatest extent under extreme conditions. Attached Figure Description

[0049] Figure 1 The flowchart shows the dynamic load balancing and power optimization allocation method for distributed charging pile groups proposed in this invention.

[0050] Figure 2 This is an architecture diagram of the distributed charging pile group dynamic load balancing and power optimization allocation system proposed in this invention. Detailed Implementation

[0051] The technical solutions in 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.

[0052] Example 1

[0053] In the first embodiment of the present invention, the present invention provides a method for dynamic load balancing and power optimization allocation of distributed charging pile groups, such as... Figure 1 As shown, it includes:

[0054] S1. The cloud coordinator obtains historical charging curves, weather data, and spatial correlation data between charging piles. Among them, the spatial correlation data between charging piles includes the geographical distance and traffic similarity between each charging pile unit.

[0055] Further, in step S1, the steps of acquiring historical charging curves, weather data, and spatial correlation data between charging piles include: performing time step alignment and missing value imputation on the collected original historical charging data and original weather data to generate a time-series feature matrix; calculating the shortest physical travel distance between each charging pile unit based on the road network topology, and extracting the overlap characteristics of interest points around each charging pile unit and historical traffic flow tidal characteristics; calculating the traffic mapping relationship using the cosine similarity algorithm, and weightedly fusing the shortest physical travel distance and the traffic mapping relationship to generate spatial correlation data between charging piles.

[0056] Specifically, the cloud coordinator first continuously collects basic operational data and environmental meteorological information of the distributed charging pile cluster through the communication network. It acquires raw historical charging data containing timestamps and instantaneous output power, and simultaneously obtains raw weather data including temperature and humidity from the meteorological service interface. To eliminate differences in sampling frequency between heterogeneous data sources, the system sets a uniform time resolution to align the raw historical charging data and raw weather data across time steps. For missing values ​​that appear during the alignment process, a linear interpolation algorithm is used to repair them, thereby constructing a complete and continuous time-series feature matrix. Each row of this matrix represents a specific time step, and each column corresponds to the power status or various meteorological indicators of a charging pile unit.

[0057] In the spatial dimension, the cloud coordinator calculates the shortest physical travel distance between any two charging pile units based on a pre-loaded urban road network topology map and the GPS coordinates of each charging pile unit. This distance is the shortest path length for the vehicle to travel along the actual road, denoted as . ,in and These represent different charging pile unit numbers.

[0058] Further, the overlap characteristics of points of interest (POIs) and historical traffic flow tidal features around each charging station unit are extracted. The system delineates a geographical area with a fixed radius centered on each charging station unit, and counts the distribution of different types of POIs within this area, such as commercial areas and residential areas, forming POI feature vectors. Simultaneously, historical monitoring data from the city's traffic management platform is accessed, and the inbound and outbound traffic flow values ​​for each time period within the aforementioned coverage area are extracted to construct historical traffic flow tidal feature vectors. The POI feature vectors and historical traffic flow tidal feature vectors are concatenated to obtain a comprehensive feature vector representing the traffic attributes of each charging station unit.

[0059] Then, the cosine similarity algorithm is used to calculate the traffic mapping relationship between each charging pile unit. Let the charging pile unit be... The comprehensive feature vector is Charging station unit The comprehensive feature vector is The traffic mapping relationship between the two is called cosine similarity. The calculation formula is as follows:

[0060] ;

[0061] In the formula, This represents the inner product of two composite eigenvectors. and Representing vectors respectively sum vector The L2 norm. Calculated. The closer the values ​​are to one, the more similar the two charging station units are in terms of surrounding businesses and traffic flow patterns.

[0062] Finally, the shortest physical travel distance and traffic mapping relationship are weighted and fused to generate spatial correlation data between piles. Since distance and similarity have different dimensions, the shortest physical travel distance is first normalized. The generated spatial correlation matrix elements are... The calculation formula is:

[0063] ;

[0064] In the formula, This represents the maximum and shortest physical travel distance between all charging pile unit pairs in the global road network. and These are the system-preset physical distance weight coefficient and traffic similarity weight coefficient, respectively, and satisfy the following conditions: The conditions. Ultimately determined by all This forms a spatial correlation data matrix between piles.

[0065] This step, through alignment interpolation and multi-dimensional spatial feature extraction and fusion, provides standardized, high-quality feature input for subsequent load forecasting models, ensuring the system's accurate extraction of complex spatiotemporal correlation features.

[0066] S2, the cloud coordinator integrates historical charging curves, weather data and spatial correlation data between charging piles into the charging load prediction model built on spatiotemporal graph neural network, outputs the multi-period power demand probability distribution prediction value and the corresponding prediction confidence interval of each charging pile unit, and sends it to the edge control node.

[0067] Further, in step S2, the steps input into the charging load prediction model constructed based on the spatiotemporal graph neural network include: performing causal convolution operations on historical charging curves and weather data using a temporal convolutional network layer to extract load temporal evolution features; converting the spatial correlation data between charging piles into an adjacency matrix, aggregating the spatial node features of adjacent charging pile units using a graph convolutional network layer to extract load spatial coupling features; and adaptively assigning weights to the load temporal evolution features and load spatial coupling features through a multi-head attention mechanism to output a latent variable feature vector.

[0068] Further, in step S2, the step of outputting the predicted power demand probability distribution values ​​for each charging pile unit across multiple time periods and the corresponding prediction confidence intervals includes: introducing a Gaussian mixture distribution mapping layer into the charging load prediction model to calculate and generate the expected value, variance, and mixture weights that characterize the future probability density; reconstructing the continuous probability density function based on the expected value, variance, and mixture weights, and sampling it in the discrete time dimension to generate the predicted power demand probability distribution values ​​across multiple time periods; and using a quantile regression algorithm to integrally truncate the probability density function at preset upper and lower quantile points, extracting the upper and lower bounds of the prediction at a given confidence level to form the prediction confidence interval.

[0069] Specifically, the cloud coordinator takes the preprocessed temporal feature matrix as input and performs causal convolution operations using a temporal convolutional network layer. Let time... Charging station unit The input vector containing historical charging power and meteorological data is The weights of the causal convolution kernel are The kernel size is Extracted load time-series evolution characteristics The calculation formula is:

[0070] ;

[0071] In the formula, Indicates charging pile unit in the past Historical charging curves and weather data characteristics at any given time This is the bias term for the temporal convolutional layer. This operation ensures that temporal causality is strictly followed when extracting features, avoiding the leakage of information from future data.

[0072] Simultaneously, the previously generated spatial correlation data between piles is transformed into an adjacency matrix of graph structure data. The spatial node features of adjacent charging pile units are aggregated using graph convolutional network layers. Let the initial node feature matrix of all input charging pile units be... Extracted load space coupling feature matrix The calculation formula is:

[0073] ;

[0074] In the formula, This represents the adjacency matrix with self-loops. It is the identity matrix. For matrix The degree matrix, The trainable weight matrix of the graph convolutional layer. This is a non-linear activation function. This step enables the interaction of characteristic information between each charging pile node and its surrounding nodes regarding the spatial distribution of charging behavior.

[0075] Subsequently, an adaptive weight allocation method is used to apply weights to the temporal evolution features and spatial coupling features of the load. The temporal evolution features and spatial coupling features are then concatenated to obtain a fused feature matrix. And generate a query matrix through linear mapping. Key matrix Sum matrix The output latent variable eigenvectors The calculation formula is:

[0076] ;

[0077] In the formula, , , All are composed of fused feature matrices Multiply by the corresponding training weight matrix to obtain, is the scaling factor for the dimension of the key vector. The weight distribution calculated by the attention mechanism reflects the dynamic importance of temporal patterns and spatial structure in influencing future loads.

[0078] After obtaining the latent variable feature vectors, the charging load prediction model introduces a Gaussian mixture distribution mapping layer. The system maps the latent variable feature vectors to Gaussian mixture distribution through a fully connected network. The parameters of the nth Gaussian component. Let the nth Gaussian component be... The expected value of each component is variance is Mixed weights are The future charging power demand is reconstructed based on expected value, variance, and mixed weights. continuous probability density function Its formula is:

[0079] ;

[0080] The system modulates the probability density function at a set future discrete time step. Sampling is performed to generate multi-time period power demand probability distribution predictions for each charging pile unit.

[0081] Finally, the quantile regression algorithm is used to integrate and truncate the reconstructed probability density function at the preset upper and lower quantiles. Let the required prediction confidence level of the system be... The predicted lower bound is The upper bound of the prediction is Its truncation integral formula is:

[0082] ;

[0083] ;

[0084] By solving the above integral equation, the upper and lower bounds of the prediction at a given confidence level are extracted, forming a prediction that includes the lower bound. and upper limit The cloud coordinator then sends the predicted confidence interval, along with the probability distribution prediction value, to the edge control node.

[0085] This step deeply integrates the nonlinear coupling characteristics of the charging network in both spatiotemporal dimensions, and accurately quantifies the uncertainty of load forecasting results in the form of probability distribution, thereby improving the reliability of forecast data in actual scheduling.

[0086] S3. The edge control node constructs a multi-objective optimization model based on the prediction uncertainty characterized by the probability distribution of power demand in multiple time periods. Under the constraint of transformer capacity, the model aims to minimize the charging completion time and maximize the renewable energy consumption rate, thereby generating a dynamic power allocation table and backup scheme.

[0087] Furthermore, in step S3, the step of constructing a multi-objective optimization model includes: calculating the statistical moment information of the predicted power demand probability distribution values ​​for multiple time periods; defining a continuous uncertainty set containing the worst load fluctuations within the polyhedral boundary; introducing a queuing theory model to construct a first objective function that minimizes the charging completion time; and extracting renewable energy generation curves to construct a second objective function that maximizes the renewable energy absorption rate; establishing the power flow equations of the distribution network and transformer capacity constraints, and coupling them with the first objective function, the second objective function, and the uncertainty set to complete the construction of the multi-objective optimization model.

[0088] Further, in step S3, the steps of generating the dynamic power allocation table and backup scheme include: using a column and constraint generation algorithm to decompose the multi-objective optimization model into a main problem and sub-problems, and obtaining the optimal power allocation sequence under the nominal prediction state through alternating iterations; discretizing and slicing the optimal power allocation sequence according to the control cycle into time windows to generate a dynamic power allocation table for daily charging and discharging scheduling; in the sub-problem solving stage, extracting extreme load deviation scenarios based on prediction uncertainty, pre-calculating the derating power instruction set for the extreme scenario, and forming a backup scheme.

[0089] Specifically, after receiving the predicted power demand probability distribution values ​​for multiple time periods, the edge control node extracts statistical moment information such as the expected value and variance. Based on this statistical moment information, a continuous uncertainty set containing the worst-case load fluctuations is constructed within the polyhedral boundary. Let time... The expected load forecast value is The maximum fluctuation deviation is The actual load is represented as the superposition of the expected value and the uncertainty fluctuation, which is the set of uncertainties. The calculation formula is:

[0090] ;

[0091] In the formula For the actual load, As an uncertainty scaling factor, The conservative parameters preset by the system to constrain the total fluctuation space. This represents the total number of control periods.

[0092] We introduce a queuing theory model to construct a first objective function that minimizes the charging completion time. Assume the vehicle arrival rate within the charging station follows a Poisson distribution with parameters . Charging service rate and the charging power allocated to each charging station unit They are positively correlated. First objective function The calculation formula is:

[0093] ;

[0094] In the formula This represents the total number of charging station units. For the first The target electricity demand of a vehicle connected to a charging pile unit is represented by the latter part of the formula, which characterizes the average waiting time of the system under the queuing theory model.

[0095] Extract the renewable energy generation curve to construct a second objective function that maximizes the renewable energy absorption rate. Let time... The predicted total renewable energy power generation capacity of the system is Second objective function The calculation formula is:

[0096] ;

[0097] The smaller function inside the formula is used to map the boundary of renewable energy power that is actually fully absorbed by each charging pile unit.

[0098] Establish the power flow equations for the distribution network and the transformer capacity constraints. Assume the node voltage amplitude is... The branch current is The line resistance and reactance are respectively and The calculation formulas for the power flow equality constraints and transformer capacity inequality constraints of the distribution network branches are as follows:

[0099] ;

[0100] ;

[0101] In the formula This is the voltage at the first node. and These represent the active and reactive power transmitted through the line, respectively. The rated apparent capacity of the regional transformer. The system power factor is denoted as . The edge control node couples the aforementioned power flow equality constraints and capacity inequality constraints with the first objective function, the second objective function, and the set of uncertainties to complete the construction of the multi-objective optimization model.

[0102] A column and constraint generation algorithm is used to decompose the constructed multi-objective optimization model into a main problem and sub-problems, which are then solved iteratively and alternately. The main problem minimizes the running objective under nominal prediction conditions to find the optimal basic allocation variables, while the sub-problems, given the basic allocation variables, are solved under a set of uncertainties. The system identifies extreme load fluctuation scenarios that polarize the system's operating costs and adds the corresponding new constraint column back to the main problem. After a finite number of iterations, until the convergence difference between the upper and lower bound objective function values ​​is less than the set minimum value, the system outputs the optimal power allocation sequence under the nominal predicted state.

[0103] The converged optimal power allocation sequence is discretized and sliced ​​according to the system's control cycle, and converted into the execution power setpoint of each charging pile unit in each time slot, generating a dynamic power allocation table for daily charging and discharging scheduling. In the last iteration of the sub-problem solving stage, the edge control node directly extracts the extreme load deviation scenario at the boundary of the uncertainty set based on the prediction uncertainty, pre-solves the derating power instruction set that strictly meets the transformer capacity safety constraint under this adverse scenario, and packages and solidifies this instruction set to form a backup plan.

[0104] This step enables the setting of a safe defense boundary for the charging network under the worst load fluctuation prediction deviation, and obtains the optimal allocation and scheduling strategy between improving the green energy consumption rate and reducing user charging waiting time.

[0105] S4. The edge control node sends the dynamic power allocation table to each charging pile unit to execute the charging power allocation.

[0106] Further, in step S4, the step of performing charging power allocation includes: the local controller of each charging pile unit performs protocol parsing of the dynamic power allocation table and extracts the target output voltage and current setting values ​​of the current time slot; using the pulse width modulation module to dynamically adjust the duty cycle signal of the power switch in the bidirectional conversion circuit according to the target output voltage and current setting values; activating the underlying voltage and current dual closed-loop control logic to smooth transient power jitter during the tracking of the dynamic power allocation table and complete the physical power injection.

[0107] Specifically, the edge control node distributes the dynamic power allocation table to each charging pile unit via the industrial communication bus. Each charging pile unit's local controller receives the data, parses the communication protocol, and extracts the target power allocation value for the current time slot. Combining this with the real-time battery terminal voltage fed back by the battery management system, the local controller divides it by the current battery terminal voltage to convert it into the target output current setting value for the current time slot, and extracts the corresponding target output voltage setting value according to the safe charging boundary protocol.

[0108] To accurately execute the above setpoints, the local controller activates the underlying voltage and current dual-loop control logic. The outer loop voltage control of this logic calculates the voltage deviation by subtracting the target output voltage setpoint from the actual output voltage acquired by the sensor. This deviation is then used in the outer loop discrete proportional-integral algorithm to generate the reference current signal required by the inner loop. Let the current discrete sampling time be... The target output voltage setting value is The actual collected output voltage is The outer ring proportional coefficient and the outer ring integral coefficient are respectively and Reference current signal The calculation formula is:

[0109] ;

[0110] Then, the inner loop current control is entered. The current deviation is calculated by subtracting the reference current signal from the actual output current acquired by the sensor. This deviation is then used to generate a modulation control voltage signal for driving the system via an inner loop discrete proportional-integral algorithm. Let the actual acquired output current be... The inner ring proportional coefficient and the inner ring integral coefficient are respectively and Modulation control voltage signal The calculation formula is:

[0111] ;

[0112] The pulse width modulation module dynamically adjusts the duty cycle signal of the power switching transistors in the bidirectional converter circuit based on the modulation control voltage signal. Assume the peak amplitude of the internally provided high-frequency triangular carrier wave is... The generated duty cycle signal The calculation formula is:

[0113] ;

[0114] The local controller uses this duty cycle signal A gate drive pulse sequence of corresponding width is generated to directly control the turn-on and turn-off times of the insulated-gate bipolar transistor (IGBT) devices inside the bidirectional converter circuit. During the closed-loop tracking of the dynamic power allocation table, this mechanism can quickly offset transient power fluctuations caused by grid voltage fluctuations or changes in battery internal resistance, thus completing the injection of physical electrical energy into the electric vehicle battery.

[0115] This step transforms the upper-level abstract optimization allocation strategy into specific control actions of the lower-level power electronic converter, ensuring high-precision tracking and stable execution of power allocation commands.

[0116] S5. The edge control node obtains the actual power of the charging pile unit during the execution process as feedback. When the actual power exceeds the prediction confidence interval, the edge control node switches to the backup plan and prioritizes reducing the charging power of the flexible load.

[0117] Furthermore, in step S5, the step of prioritizing the reduction of charging power belonging to flexible loads includes: acquiring the battery state of charge, battery capacity, and expected user departure time of the connected electric vehicles in real time, and constructing a multi-dimensional flexibility evaluation function; calculating the real-time flexibility index of each charging task based on the multi-dimensional flexibility evaluation function, sorting the electric vehicles according to the real-time flexibility index from high to low, and generating a load reduction priority queue; issuing power derating instructions in the order of the load reduction priority queue until the total actual power falls back to the range of the predicted confidence interval.

[0118] Specifically, the edge control node collects the actual output power of each charging pile unit in real time during execution through underlying sensors such as smart meters, and calculates the total actual power by summing the actual output power of all charging pile units. Let the current sampling time be... The total actual power is The upper bound of the prediction confidence interval extracted from the predicted values ​​of the multi-period power demand probability distribution is: When discovered If the system determines that an extreme load change or predictive failure has led to an overload risk for the transformer, the edge control node immediately interrupts the normal dynamic power allocation table distribution process and switches to the pre-calculated backup scheme operation mode, preparing to prioritize reducing the charging power of flexible loads.

[0119] In the priority reduction step, the edge control node obtains the underlying status data of electric vehicles currently connected to each charging pile unit in real time through the vehicle-to-charging-pile communication protocol. Specifically, the obtained data includes the... Current battery state of charge of an electric vehicle Target battery state of charge Rated battery capacity And the user's expected departure time set via the charging terminal Combined with the current system time and the maximum output power limit of the corresponding charging pile unit A multidimensional flexibility evaluation function is constructed. This function quantifies the adjustability margin of each charging task in the time dimension, and its output is a real-time flexibility index. The calculation formula is:

[0120] ;

[0121] In the formula, the numerator represents the total amount of electricity required to meet the vehicle's remaining charging needs, and the denominator represents the theoretical maximum amount of electricity that can be provided by continuous charging at maximum power during the remaining parking time. Real-time flexibility index The closer the value is to one, the greater the charging margin in terms of time, that is, the higher the flexibility.

[0122] The edge control node uses the aforementioned multidimensional flexibility evaluation function to calculate the real-time flexibility index of all tasks currently in the charging state, and then sorts all charging tasks in descending order according to the calculated real-time flexibility index values ​​to generate a load reduction priority queue. The target total power to be reduced is calculated, denoted as [value missing]. Its value is the total actual power. Subtract the upper boundary of the prediction confidence interval The difference. Edge control nodes, following the load reduction priority queue, sequentially issue power derating or charging pause commands to the charging pile units with the highest flexibility at the front of the queue. Each derating command reduces the target total power. The corresponding power reduction value is deducted until the total target power is less than or equal to zero, so that the total actual power strictly falls back to the safe range of the prediction confidence interval.

[0123] This step provides a precise safety fallback mechanism when the system encounters a sudden power limit exceedance. By quantifying the charging task margin, it avoids a crude global power outage, protecting the transformer from overload damage while maximizing the charging needs of users with urgent time constraints.

[0124] Example 2

[0125] Traditional load balancing methods rely heavily on deterministic single-point predictions, lacking quantitative assessment of prediction uncertainties. These methods are prone to failure when faced with extreme load fluctuations caused by sudden traffic surges, and struggle to mitigate the physical safety risks of transformer overload. To address these issues, this invention provides a distributed charging pile group dynamic load balancing and power optimization allocation system, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this system is as follows:

[0126] The data acquisition module is used to acquire historical charging curves, weather data, and spatial correlation data between charging piles. The spatial correlation data between charging piles includes the geographical distance and traffic similarity between each charging pile unit.

[0127] The load forecasting module is used to fuse the historical charging curve, the weather data and the spatial correlation data between the charging piles into the charging load forecasting model constructed based on the spatiotemporal graph neural network, and output the multi-period power demand probability distribution forecast value and the corresponding prediction confidence interval of each charging pile unit.

[0128] The optimization modeling module is used to construct a multi-objective optimization model based on the prediction uncertainty represented by the predicted value of the multi-time period power demand probability distribution. Under the transformer capacity constraint, the model aims to minimize the charging completion time and maximize the renewable energy consumption rate, and generates a dynamic power allocation table and backup scheme.

[0129] A power allocation module is used to send the dynamic power allocation table to each charging pile unit to perform charging power allocation.

[0130] The feedback control module is used to obtain the actual power of the charging pile unit during the execution process as feedback. When the actual power exceeds the predicted confidence interval, it switches to the backup scheme and prioritizes reducing the charging power of the flexible load.

[0131] Specifically, the data acquisition module first performs time step alignment and missing value imputation on the collected original historical charging data and original weather data to generate a time-series feature matrix. Based on the road network topology, it calculates the shortest physical travel distance between each charging pile unit, extracts the overlap of surrounding points of interest and traffic flow tidal features, and then uses a cosine similarity algorithm to calculate traffic mapping relationships. Finally, it weights and fuses the physical distance and traffic mapping relationships to generate spatial correlation data between charging piles. The load prediction module uses a temporal convolutional network layer to perform causal convolution operations to extract load temporal evolution features. It then converts the spatial correlation data into an adjacency matrix and extracts load spatial coupling features through a graph convolutional network layer. After outputting latent variable feature vectors through multi-head attention mechanism adaptive weight allocation, it introduces a Gaussian mixture distribution mapping layer to reconstruct the continuous probability density function and samples it in the discrete time dimension. Simultaneously, it uses a quantile regression algorithm to perform integral truncation at preset quantile points, extracting the upper and lower bounds of the prediction to form the prediction confidence interval. The optimization modeling module calculates the statistical moment information of the predicted values ​​and defines polyhedral edges. Based on the continuous uncertainty set within the boundary, a multi-objective optimization model including dual objectives and distribution network power flow and capacity constraints is constructed by combining queuing theory model and renewable energy generation curves. The model is decomposed into principal and sub-problems through alternating iterations using column and constraint generation algorithms. This not only obtains a dynamic power allocation table after discretization of time windows, but also extracts a pre-calculated derating instruction set for extreme load deviation scenarios based on prediction uncertainty to form a backup plan. The power allocation module uses the local controller to perform protocol parsing on the allocation table to extract voltage and current setpoints. It uses a pulse width modulation module to dynamically adjust the duty cycle signal of the conversion circuit and activates the underlying dual closed-loop control logic to smooth transient power jitter and complete power injection. When the total actual power exceeds the limit, the feedback control module obtains the battery state of charge, capacity, and expected departure time of the connected vehicles in real time to construct a multi-dimensional flexibility evaluation function. It calculates the real-time flexibility index of each charging task and arranges them in descending order to generate a load reduction priority queue. Power derating instructions are issued sequentially until the total actual power smoothly falls back to the predicted confidence interval.

[0132] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic load balancing and power optimization allocation of distributed charging pile groups, characterized in that, include: S1. The cloud coordinator acquires historical charging curves, weather data, and spatial correlation data between charging piles, wherein the spatial correlation data between charging piles includes the geographical distance and traffic similarity between each charging pile unit. S2. The cloud coordinator integrates the historical charging curve, the weather data, and the spatial correlation data between charging piles into the charging load prediction model constructed based on the spatiotemporal graph neural network, outputs the multi-period power demand probability distribution prediction value and the corresponding prediction confidence interval of each charging pile unit, and sends it to the edge control node. S3. The edge control node constructs a multi-objective optimization model based on the prediction uncertainty represented by the multi-time period power demand probability distribution prediction value. Under the transformer capacity constraint, the model aims to minimize the charging completion time and maximize the renewable energy consumption rate, and generates a dynamic power allocation table and backup scheme. S4. The edge control node sends the dynamic power allocation table to each charging pile unit to perform charging power allocation. S5. The edge control node obtains the actual power of the charging pile unit during the execution process as feedback. When the actual power exceeds the predicted confidence interval, the edge control node switches to the backup scheme and prioritizes reducing the charging power belonging to the flexibility load.

2. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, Step S1, the step of acquiring historical charging curves, weather data, and spatial correlation data between charging piles, includes: Time step alignment and missing value imputation are performed on the collected raw historical charging data and raw weather data to generate a time series feature matrix; The shortest physical travel distance between each charging pile unit is calculated based on the road network topology, and the overlap characteristics of interest points around each charging pile unit and the historical traffic flow tidal characteristics are extracted. The traffic mapping relationship is calculated using the cosine similarity algorithm, and the shortest physical travel distance is weighted and fused with the traffic mapping relationship to generate the spatial correlation data between the piles.

3. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, Step S2, the step of inputting the data into the charging load prediction model constructed based on a spatiotemporal graph neural network, includes: Causal convolution operations are performed on the historical charging curves and the weather data using a temporal convolutional network layer to extract the temporal evolution features of the load. The spatial correlation data between the piles is transformed into an adjacency matrix, and the spatial node features of adjacent charging pile units are aggregated using graph convolutional network layers to extract load spatial coupling features. An adaptive weight allocation is performed on the load temporal evolution features and the load spatial coupling features using a multi-head attention mechanism, and a latent variable feature vector is output.

4. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, Step S2, the step of outputting the predicted power demand probability distribution of each charging pile unit over multiple time periods and the corresponding prediction confidence interval, includes: A Gaussian mixture distribution mapping layer is introduced into the charging load prediction model to calculate and generate the expected value, variance, and mixture weights that characterize the future probability density. Based on the expected value, variance, and mixed weights, a continuous probability density function is reconstructed, and samples are taken in the discrete time dimension to generate the predicted value of the multi-time period power demand probability distribution. The probability density function is integrally truncated at preset upper and lower quantiles using a quantile regression algorithm to extract the upper and lower bounds of the prediction at a given confidence level, thus forming the prediction confidence interval.

5. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, Step S3, the step of constructing the multi-objective optimization model, includes: Calculate the statistical moment information of the predicted power demand probability distribution for the multi-period time period, and define a continuous set of uncertainties containing the worst load fluctuations within the polyhedral boundary; A queuing theory model is introduced to construct a first objective function that minimizes the charging completion time, and a renewable energy generation curve is extracted to construct a second objective function that maximizes the renewable energy absorption rate. The power flow equations of the distribution network and the transformer capacity constraints are established, and then coupled with the first objective function, the second objective function, and the set of uncertainties to complete the construction of the multi-objective optimization model.

6. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, Step S3, the step of generating the dynamic power allocation table and backup scheme, includes: The multi-objective optimization model is decomposed into a main problem and sub-problems using a column and constraint generation algorithm, and the optimal power allocation sequence under the nominal prediction state is obtained through alternating iterations. The optimal power allocation sequence is discretized and sliced ​​according to the control period to generate the dynamic power allocation table for daily charging and discharging scheduling. During the sub-problem solving stage, extreme load deviation scenarios are extracted based on the prediction uncertainty, and a derating power instruction set for the extreme scenario is pre-calculated to form the backup scheme.

7. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, In step S4, the step of performing charging power allocation includes: The local controller of each charging pile unit performs protocol parsing on the dynamic power allocation table and extracts the target output voltage and current setpoints for the current time slot. The duty cycle signal of the power switch in the bidirectional converter circuit is dynamically adjusted by the pulse width modulation module according to the target output voltage and current set values. The underlying voltage and current dual closed-loop control logic is activated to smooth transient power jitter during the tracking of the dynamic power allocation table and complete the physical power injection.

8. The method for dynamic load balancing and power optimization allocation of distributed charging pile groups according to claim 1, characterized in that, In step S5, the step of prioritizing the reduction of charging power belonging to flexible loads includes: Real-time acquisition of battery state of charge, battery capacity, and user's expected departure time for connected electric vehicles; and construction of a multi-dimensional flexible evaluation function. The real-time flexibility index of each charging task is calculated based on the multidimensional flexibility evaluation function. Electric vehicles are sorted from high to low according to the real-time flexibility index to generate a load reduction priority queue. Power derating instructions are issued sequentially according to the order of the load reduction priority queue until the total actual power falls back to the range of the predicted confidence interval.

9. A distributed charging pile group dynamic load balancing and power optimization allocation system, characterized in that, The system used for the distributed charging pile group dynamic load balancing and power optimization allocation method according to any one of claims 1-8, the system comprising: The data acquisition module is used to acquire historical charging curves, weather data, and spatial correlation data between charging piles. The spatial correlation data between charging piles includes the geographical distance and traffic similarity between each charging pile unit. The load forecasting module is used to fuse the historical charging curve, the weather data and the spatial correlation data between the charging piles into the charging load forecasting model constructed based on the spatiotemporal graph neural network, and output the multi-period power demand probability distribution forecast value and the corresponding prediction confidence interval of each charging pile unit. The optimization modeling module is used to construct a multi-objective optimization model based on the prediction uncertainty represented by the predicted value of the multi-time period power demand probability distribution. Under the transformer capacity constraint, the model aims to minimize the charging completion time and maximize the renewable energy consumption rate, and generates a dynamic power allocation table and backup scheme. A power allocation module is used to send the dynamic power allocation table to each charging pile unit to perform charging power allocation. The feedback control module is used to obtain the actual power of the charging pile unit during the execution process as feedback. When the actual power exceeds the predicted confidence interval, it switches to the backup scheme and prioritizes reducing the charging power of the flexible load.

Images