Charging pile power distribution method and device based on artificial intelligence and medium

Abstract

The embodiment of the invention discloses a charging pile power distribution method and device based on artificial intelligence and a medium, and relates to the technical field of electric vehicle charging, and the method comprises the steps: collecting multi-dimensional charging data of a target charging pile cluster; performing data fusion based on the multi-dimensional charging data to generate a fusion data vector, extracting local features in the fusion data vector by using a convolutional neural network, and extracting time sequence features by using a gating cycle unit to generate a depth feature vector; based on the depth feature vector, a multi-objective optimization function is constructed, and the multi-objective optimization function comprises a charging efficiency objective, a power grid load fluctuation objective, a user satisfaction objective and a battery life objective; and solving the multi-objective optimization function by using a particle swarm optimization algorithm to obtain a power distribution parameter of each charging pile, and sending the power distribution parameters to a corresponding charging pile controller so as to adjust the output power of the charging piles and realize power distribution of the charging piles.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of electric vehicle charging, and in particular to a charging pile power distribution method based on artificial intelligence, equipment and medium. BACKGROUND

[0002] With the development of the electric vehicle industry, the construction and operation scale of charging infrastructure, especially charging pile clusters, continue to expand. An efficient power distribution strategy is the key to improving the operation efficiency of charging pile clusters, ensuring the stable operation of the power grid, and meeting the diverse needs of users.

[0003] Existing charging pile power distribution technologies mostly use strategies based on fixed rules. For example, some systems use a fixed power output mode of "first come, first served", that is, each connected vehicle is allocated its maximum acceptable power until the total power of the cluster reaches the upper limit. Another common strategy is to allocate power according to simple priority rules, such as according to the remaining battery level (State of Charge, SOC) of the vehicle. These methods achieve basic power distribution to some extent, but have obvious limitations. First, such strategies are usually static or semi-static and cannot respond in real time to dynamic fluctuations in power grid load, which can exacerbate power grid load during peak usage periods and result in idle power resources during low usage periods. Second, existing technical solutions use data in a single and shallow manner, typically relying only on basic operating parameters of the charging pile (such as voltage and current), and fail to fully integrate multi-dimensional information such as the battery health state (State of Health, SOH) of the vehicle, the historical charging preferences and real-time choices of the user, resulting in a lack of precision and foresight in the allocation decision. Existing technologies often focus on a single goal of "charging speed" and are unable to effectively coordinate multiple mutually conflicting goals such as power grid stability, battery life protection, and user satisfaction while pursuing charging efficiency.

[0004] Therefore, in the process of charging pile power distribution in the prior art, due to the rigidity of the allocation strategy, insufficient use of data, and insufficient multi-objective coordination capability, there is a problem of being unable to dynamically adapt to power grid and user needs, resulting in the need to improve the overall operation efficiency of the charging pile cluster. SUMMARY

[0005] One or more embodiments of the present specification provide a charging pile power distribution method based on artificial intelligence, equipment and medium, to solve the technical problem that in the process of charging pile power distribution in the prior art, due to the rigidity of the allocation strategy, insufficient use of data, and insufficient multi-objective coordination capability, there is a problem of being unable to dynamically adapt to power grid and user needs, resulting in the need to improve the overall operation efficiency of the charging pile cluster.

[0006] One or more embodiments of the present specification adopt the following technical solutions:

[0007] One or more embodiments of the present specification provide an artificial intelligence-based charging pile power distribution method, the method comprising: collecting multi-dimensional charging data of a target charging pile cluster, wherein the multi-dimensional charging data comprises charging pile operation data, charging battery data, power grid load data, and user charging preference data; performing data fusion based on the multi-dimensional charging data to generate a fusion data vector, using a convolutional neural network to extract local features in the fusion data vector, and using a gated recurrent unit to extract time series features to generate a deep feature vector; based on the deep feature vector, constructing a multi-objective optimization function, wherein the multi-objective optimization function comprises a charging efficiency objective, a power grid load fluctuation objective, a user satisfaction objective, and a battery life objective; using a particle swarm optimization algorithm to solve the multi-objective optimization function to obtain power distribution parameters for each charging pile, and sending the power distribution parameters to a corresponding charging pile controller to adjust the output power of the charging pile, thereby achieving charging pile power distribution.

[0008] One or more embodiments of the present specification provide an artificial intelligence-based charging pile power distribution device, comprising:

[0009] at least one processor; and,

[0010] a memory in communication connection with the at least one processor; wherein,

[0011] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above method.

[0012] One or more embodiments of the present specification provide a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions are configured to execute the above method.

[0013] The above-mentioned at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: Through the embodiments of this specification, by collecting and fusing data from four dimensions—charging piles, batteries, power grids, and users—the limitations of information silos in traditional methods are broken. Simultaneously collecting data on battery health status and real-time state of charge, macroscopic load stability of the power grid, and personalized user needs and preferences, richer contextual information can be perceived. The combination of convolutional neural networks and gated recurrent units is used to extract deep features. Convolutional neural networks excel at capturing local spatial correlations in data, while gated recurrent units specialize in processing time-series data, enabling the extraction of deep features from users' historical charging behavior, battery status, and other data. By learning long-term patterns and short-term dynamics from the changing trends of battery power, and combining these two to generate a deep feature vector, the predictability and accuracy of decision-making are greatly improved. The power allocation problem is explicitly formulated as a multi-objective optimization problem that includes charging efficiency, grid stability, user satisfaction, and battery life. By constructing a multi-objective optimization function, it can adapt to dynamically changing needs. For example, during the night when the grid load is light, charging efficiency and user satisfaction can be appropriately emphasized; while during the evening when electricity consumption is high, grid stability must be prioritized. The particle swarm optimization algorithm is used to solve the multi-objective function, ensuring that the final output power allocation parameters are a comprehensive optimal solution obtained through a fully intelligent search. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0015] Figure 1 A flowchart illustrating an artificial intelligence-based power allocation method for charging piles, provided as an embodiment of this specification;

[0016] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based charging pile power distribution device provided in the embodiments of this specification. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0018] This specification provides an artificial intelligence-based charging pile power allocation method. It should be noted that the execution subject in this specification embodiment can be a server or any device with data processing capabilities. Figure 1 A flowchart illustrating an artificial intelligence-based power allocation method for charging piles, as provided in this specification, is shown below. Figure 1 As shown, the main steps include the following:

[0019] Step S101: Collect multi-dimensional charging data of the target charging pile cluster.

[0020] This multi-dimensional charging data includes charging pile operation data, charging battery data, grid load data, and user charging preference data.

[0021] In one embodiment of this specification, charging pile operation data is collected from multiple charging pile controllers. This data includes output power, voltage, current, and temperature, used to assess the current operating status and available power capacity of the charging pile. Specifically, output power, voltage, current, and temperature data are read in real-time from the communication interface of the charging pile controllers to monitor the real-time load and overheating risk of the charging pile. Battery data is collected from the electric vehicle battery management system. This data includes state of charge (SOC) and state of health (SOH), used to assess the battery's charging needs and safe charging thresholds. Specifically, SOC and SOH are obtained through the electric vehicle's OBD interface or charging gun communication protocol to determine the battery's maximum allowable charging power and lifespan protection strategy. Grid load data is collected from the grid monitoring system. This data includes real-time load rate and voltage stability, used to assess the current load condition and stability limitations of the grid. Specifically, real-time load rate and voltage fluctuation data are collected from smart meters or SCADA systems to identify grid peak and off-peak periods and stability boundaries. User charging preference data is collected from user terminals. This data includes historical charging time and charging mode selections, which are used to identify user charging behavior patterns and personalized needs. Specifically, historical charging records and mode selections are obtained from user mobile apps or cloud databases to predict user charging duration and power requirements.

[0022] Step S102: Data fusion is performed based on multi-dimensional charging data to generate a fused data vector. Local features in the fused data vector are extracted using a convolutional neural network, and time series features are extracted using a gated recurrent unit to generate a deep feature vector.

[0023] Based on this multi-dimensional charging data, data fusion is performed to generate a fused data vector. Specifically, this includes: preprocessing the charging pile operation data, charging battery data, grid load data, and user charging preference data from the multi-dimensional charging data to generate preprocessed data. This preprocessing includes outlier removal, data normalization, and time synchronization. Outlier removal uses the standard deviation method to identify and remove data points that deviate from the normal range. Data normalization uses a min-max scaling method to map data of different dimensions to a unified interval. Time synchronization uses a timestamp alignment method to unify multi-source data to the same time reference. Based on this preprocessed data, an attention mechanism is used to assign data weights for weighted processing, determining the weighted processing data. The grid load data is the first weight, the charging battery data is the second weight, and the user charging preference data is the third weight. These first, second, and third weights are trained based on historical data and are used to reflect the degree of data influence on power allocation. The weighted processing data is then concatenated to generate a fused data vector, which is then input into the feature extraction network.

[0024] In one embodiment of this specification, it is first necessary to collect charging pile operation data, charging battery data, grid load data, and user charging preference data in real time from the target charging pile cluster. These data come from the sensor interface of the charging pile controller, signals transmitted by the electric vehicle battery management system through OBD or charging gun communication protocols, real-time readings provided by grid monitoring systems such as smart meters or SCADA systems, and historical records and current selections stored in the user's mobile APP or cloud database. After the data collection is completed, a preprocessing operation is performed. In the outlier removal stage, the mean and standard deviation of the historical sequence of each data type (such as charging pile output power, battery state of charge, grid load rate, and user charging duration) are calculated to establish a normal fluctuation range. Any data point that deviates from this range by more than a preset standard deviation multiple will be automatically identified and removed. For example, when the charging pile output power suddenly spikes to a level far higher than the historical average, it will be regarded as sensor noise or a temporary fault and filtered out to ensure the authenticity and reliability of the input data. Next, data normalization is performed using a min-max scaling method. This involves calculating the minimum and maximum values ​​for each data feature (such as voltage, current, and load rate) within its historical acquisition period, and then linearly mapping all real-time data to a unified numerical range, such as zero to one. This eliminates scale differences caused by different units (such as volts, amperes, and percentages), ensuring that subsequent algorithms can fairly process various features without bias towards certain dimensions due to numerical magnitude discrepancies. Simultaneously, a time synchronization step adds a high-precision timestamp to each data source and uses interpolation algorithms or window aggregation methods to align asynchronously arriving data to a common time reference point. For example, it matches grid load fluctuation data with user charging request times with second-level precision, avoiding data timing errors caused by transmission delays and ensuring consistency of multi-source data on the timeline.

[0025] The preprocessed valid data then enters the attention mechanism weight allocation stage. A lightweight neural network model is trained based on historical data. The input of the lightweight neural network model is the preprocessed multi-dimensional data fragments, and the output is the dynamic weight value of each data type. Among them, the grid load data is assigned the first weight, reflecting its key impact on grid stability; the charging battery data is given the second weight to highlight the importance of battery safety and life protection; and the user charging preference data is assigned the third weight to reinforce personalized needs. The training process optimizes the weight parameters through the backpropagation algorithm, so that those data features that have made great contributions to the decision in historical successful power allocation cases automatically receive higher weights.

[0026] During weighted processing, the value of each data point is multiplied by its corresponding weight coefficient. For example, the real-time value of the power grid load is multiplied by the first weight, and the battery health status is multiplied by the second weight, thus generating weighted data. This process is essentially a data augmentation step that emphasizes key information and suppresses secondary noise. All weighted data are concatenated into a high-dimensional fusion data vector. Weighted data from different sources are connected in a predetermined order (e.g., charging pile operation data first, then battery data, power grid data, and user data) into a continuous numerical sequence. This vector serves as the direct input to the feature extraction network, providing a structured and optimized data foundation for subsequent feature extraction by convolutional neural networks and gated recurrent units. The entire implementation process ensures a seamless transformation from raw multi-source heterogeneous data to a high-quality fusion vector. Each step is interconnected: preprocessing ensures data cleanliness and consistency, the attention mechanism enables intelligent weight adjustment, and weighting and concatenation complete feature integration.

[0027] Conventional charging pile power allocation methods often use raw data directly for decision-making or only perform simple standardization, leading to significant data heterogeneity issues. For example, grid load data and user preference data have different dimensions; direct fusion can distort the model's attention due to numerical differences. The embodiments in this specification, however, fundamentally improve data quality by removing outliers, normalizing data, and synchronizing time during the preprocessing stage. This eliminates noise and timing bias, allowing subsequent algorithms to operate based on consistent and reliable input, thus avoiding decision-making errors caused by inaccurate data. Furthermore, conventional methods often use fixed weights or empirical weighting to allocate data importance, which cannot adapt to dynamic changes in scenarios. For instance, grid load data should have a greater impact during peak hours, but fixed weights cannot reflect this. The attention mechanism introduced in the embodiments of this specification generates dynamic weights based on historical data training, adaptively adjusting the contribution of different data types. In real-time operation, it prioritizes key factors (such as grid stability or battery emergency status), thereby enhancing the system's context awareness and flexibility, making power allocation more aligned with actual needs. Conventional data fusion often overlooks the deep correlations between multi-source data, performing only shallow concatenation. However, the embodiments in this specification generate a fusion vector by weighting and emphasizing important features, providing rich and structured input to the feature extraction network. This helps convolutional neural networks and gated recurrent units more effectively capture local patterns and temporal dependencies, improving the overall model's prediction accuracy. Through refined preprocessing and intelligent weight allocation, it not only solves the problems of data inconsistency, rigid weights, and insufficient feature utilization in conventional methods, but also achieves deep integration of multi-dimensional data. This provides a more scientific, adaptive, and reliable data foundation for charging pile power allocation, indirectly optimizing the balance of multiple objectives such as charging efficiency, grid stability, and user satisfaction.

[0028] The fusion data vector is extracted using a convolutional neural network, and time-series features are extracted using a gated recurrent unit to generate a deep feature vector. Specifically, the fusion data vector is input into the convolutional neural network, and convolution operations are performed on grid load fluctuations and charging pile temperature changes through convolutional layers to extract local spatial features and generate a local feature vector. The local feature vector is then input into a gated recurrent unit, and time-dependent features are extracted by performing sequence modeling on user charging behavior and battery state of charge changes through recurrent layers to generate a time-series feature vector. The local feature vector and the time-series feature vector are concatenated and dimensionality is reduced through a fully connected layer to generate a deep feature vector, which is used to construct a multi-objective optimization function.

[0029] In one embodiment of this specification, a combined use of convolutional neural networks and gated recurrent units is employed to achieve the collaborative mining of spatial and temporal features of data. Specifically, when inputting the fused data vector into the convolutional neural network, the network structure first needs to be designed according to the data characteristics. The convolutional layers use one-dimensional convolutional kernels that slide along the time dimension, with the kernel width covering data points across multiple consecutive time steps. Each convolutional kernel is specifically designed to detect a particular type of local pattern. For example, some convolutional kernel parameters are trained to identify sudden step increases or decreases in grid load data, which typically correspond to the beginning or end of peak grid electricity consumption. Other convolutional kernels focus on identifying abnormal temperature change slopes in charging pile temperature data, which may indicate changes in the efficiency of the cooling system or potential failure risks. During the convolution operation, each convolution kernel performs local perception on different segments of the fused data vector. By linearly weighted summation and applying a non-linear activation function such as ReLU, a preliminary feature map is generated. This step can effectively capture short-term local correlations in the data. For example, there may be a correlation between the rapid fluctuation of the power grid load within a few minutes and the abnormal rise in the temperature of a charging pile within the same time period. This correlation may be subtle in the original data, but it is highlighted by the local receptive field of the convolution kernel.

[0030] Subsequently, pooling operations are performed on the generated feature maps, typically using max pooling, which retains the most significant feature responses within small local regions. This not only reduces the dimensionality of the data and the subsequent computational burden but also enhances the robustness of the features to small data translations and fluctuations. After convolution and pooling operations, the originally high-dimensional fused data vector is transformed into a set of feature maps that better represent its essential local spatial features. These feature maps are flattened and connected to form a comprehensive local feature vector, which condenses the key information related to spatial locality in multi-source data.

[0031] The local feature vector is input into a gated recurrent unit (GRU) for time-series feature extraction. The GRU, an advanced recurrent neural network structure, incorporates update and reset mechanisms to effectively capture long-term dependencies in the time series. During processing, the local feature vector is divided into chronologically ordered sequence segments, with the feature sub-vector at each time step sequentially input into the GRU unit. The update gate controls how much state information from the previous time step is retained, while the reset gate determines how much past information is ignored to better capture new patterns. For example, when processing user charging behavior sequences, the update gate can learn the user's long-standing habit of charging during specific time periods (e.g., evening) and maintain this background information during sequence modeling. The reset gate, however, might selectively "forget" a user's unusual daytime fast-charging request.

[0032] The GRU learns long-term patterns to respond more sensitively to sudden changes; similarly, for sequences of changes in battery state of charge, the GRU can learn the typical shape of its charging curve (e.g., constant current followed by constant voltage) and identify abnormal charging states that deviate from the normal curve. These time-dependent features are crucial for predicting battery charging completion time and assessing battery health.

[0033] After processing all time steps using GRU, the final hidden layer state output converges the contextual information of the entire sequence, forming a time-series feature vector. This vector deeply encodes the patterns and anomalies of dynamic processes such as user behavior preferences and battery state evolution. Finally, the local feature vector output by the convolutional neural network is concatenated with the time-series feature vector output by the gated recurrent unit, forming a more comprehensive joint feature vector. This vector simultaneously contains the local spatial patterns and temporal evolution patterns of the data. This high-dimensional joint feature vector is then subjected to dimensionality reduction through one or more fully connected layers. Each neuron in the fully connected layer is connected to all neurons in the previous layer. Through a combination of linear transformations and non-linear activation functions (such as Sigmoid or Tanh), a complex mapping from the joint features to a higher-level abstract representation is learned. The purpose of dimensionality reduction is to reduce the feature dimensionality while preserving the most discriminative information, avoiding the "curse of dimensionality" problem in subsequent optimization algorithms, and enhancing the generalization ability of the features. Finally, the fully connected layer outputs a deep feature vector.

[0034] Compared to conventional feature extraction methods, the embodiments in this specification achieve collaborative mining of spatiotemporal features in multi-source charging power allocation data by deeply fusing convolutional neural networks and gated recurrent units, resulting in significant benefits. Conventional methods often process different types of data separately or use simple feature engineering methods (such as manually calculating the mean and variance), making it difficult to capture the complex interaction between local spatial patterns such as sudden changes in grid load and temperature anomalies and long-term time dependencies such as user behavior sequences and battery state changes. This leads to limited feature representation capabilities, which in turn affects the accuracy of subsequent optimization decisions. In contrast, the embodiments in this specification, through the local perception characteristics of convolutional neural networks, can automatically and effectively identify key local correlation signals from the fused data, such as the synchronicity between rapidly fluctuating grid load and temperature changes of specific charging piles. These subtle but important patterns are easily overlooked if extracted using manual rules. At the same time, the introduction of gated recurrent units can go beyond instantaneous snapshot-style analysis and instead model the dynamic process from a continuous time series perspective. Conventional static methods cannot understand the continuity of user charging habits or the gradual trend of battery charging status. However, the embodiments in this specification, through the memory and forgetting mechanism of GRU, can flexibly learn long-term patterns and sensitively respond to short-term changes, thereby generating temporal features that better reflect the dynamic evolution of the system. By concatenating and dimensionality-reducing the local spatial features extracted by CNN and the temporal dependency features extracted by GRU, the final generated deep feature vector is a highly abstract and integrated representation. It is not a simple stacking of original information, but rather reveals the deep spatiotemporal correlation structure driving power allocation decisions behind the data.

[0035] Step S103: Construct a multi-objective optimization function based on deep feature vectors.

[0036] The multi-objective optimization function includes charging efficiency objective, grid load fluctuation objective, user satisfaction objective, and battery life objective.

[0037] Based on this deep feature vector, a multi-objective optimization function is constructed, specifically including: extracting multi-dimensional objective features from the deep feature vector, where the multi-dimensional objective features include charging efficiency features, grid load fluctuation features, user satisfaction features, and battery life features; based on the multi-dimensional objective features, determining the dimensional indicators under each dimension, and generating a multi-objective optimization function based on multiple dimensional indicators, where the dimensional indicators include charging efficiency indicators, grid load fluctuation indicators, user satisfaction indicators, and battery life indicators. Specifically, based on the multi-dimensional objective features, determining the dimensional indicators under each dimension includes: calculating the cumulative charging energy per unit time based on the output power and charging duration in the charging pile operation data to determine the charging efficiency indicator; calculating the variance of the load sequence based on the real-time load rate in the grid load data to determine the grid load fluctuation indicator; calculating the matching score based on the historical charging modes and current selection in the user charging preference data to determine the user satisfaction indicator; and calculating the battery degradation coefficient based on the health status and state of charge in the charging battery data using a pre-built battery degradation model to determine the battery life indicator. Based on multiple indicators of this dimension, a multi-objective optimization function is generated, specifically including: maximizing the charging efficiency indicator to construct a single-objective function for charging efficiency; maximizing the user satisfaction indicator to construct a single-objective function for user satisfaction; minimizing the grid load fluctuation indicator to construct a single-objective function for grid stability; and minimizing the battery life indicator to construct a single-objective function for battery life. Dynamic weighting coefficients are calculated based on real-time collected grid load data and user charging preference data. The single-objective functions for charging efficiency, grid stability, user satisfaction, and battery life are then linearly weighted using these dynamic weighting coefficients to generate the multi-objective optimization function.

[0038] In one embodiment of this specification, firstly, multi-dimensional target features are extracted from the deep feature vector. Feature components with different semantic directions are separated through a specific feature decoupling layer (typically composed of a set of parallel fully connected subnetworks). Specifically, the charging efficiency feature is a numerical sequence decoded from the deep feature vector through a fully connected branch. This branch is constrained to learn patterns related to energy transfer rates during training, and its output features reflect the achievable charging power efficiency level under a given system state. The grid load fluctuation feature is extracted by another dedicated fully connected branch. The weight parameters of this branch focus on capturing unstable patterns and fluctuation cycles in the grid data, and its output features quantify the current vulnerability of the grid and its sensitivity to power changes. User satisfaction features are obtained by analyzing latent variables related to user behavior preferences in the deep feature vector. For example, an attention mechanism is used to focus on specific dimensions in the vector representing users' historical selection patterns and urgency preferences, and these are aggregated into a comprehensive score reflecting the user's expected satisfaction. Battery life characteristics are derived from deep feature vectors by resolving potential variables related to battery health and charge / discharge stress. These variables are semantically aligned using prior knowledge from the battery electrochemical model, enabling them to characterize the impact of different charging strategies on the battery.

[0039] After obtaining the target features in these four dimensions, they need to be transformed into specific, calculable dimensional indicators. The charging efficiency indicator is calculated by multiplying the real-time output power in the charging pile's operating data by the corresponding effective charging duration to obtain the total electrical energy transmitted within that time period. Then, the cumulative charging energy per unit time is calculated. This indicator reflects the actual efficiency of power conversion into battery-stored energy, rather than simply the output power. The determination of the grid load fluctuation indicator requires based on the real-time load rate sequence in the grid load data. First, continuous load rate sampling points are collected within a certain time window. Then, the statistical variance of these sampling point sequences is calculated. This variance directly reflects the stability of the grid load; the larger the variance, the more severe the fluctuations and the higher the risk of impact on the grid. The user satisfaction indicator is calculated through a matching degree scoring process. Historical charging patterns (such as frequently used charging periods and preferred power levels) in the user's charging preference data are compared with the current user's choices (such as the selected charging mode and target power level), while also considering factors such as the urgency of the current charging. A comprehensive matching degree score is output through a pre-trained scoring model. Determining battery life indicators is the most complex step. It requires inputting the health status and current state of charge from the rechargeable battery data into a pre-built battery degradation model. This model is usually a regression model trained on a large amount of battery aging experimental data. It can predict the incremental degradation of battery capacity based on parameters such as the current charging rate, temperature, and SOC range, i.e., the battery degradation coefficient.

[0040] Next, we construct single-objective functions based on these dimensional indicators. For the charging efficiency indicator and the user satisfaction indicator, since we want their values ​​to be as high as possible, we construct a single-objective function of the maximization type, that is, the function output is positively correlated with the indicator value. For the grid load fluctuation indicator and the battery life indicator (attenuation coefficient), we want their values ​​to be as low as possible, so we construct a single-objective function of the minimization type.

[0041] The dynamic weighting coefficients are then calculated. This embodiment does not use fixed weights, but rather dynamically adjusts them based on real-time collected grid load data and user charging preference data. For example, when grid monitoring data shows the system is at peak load, the weighting coefficient of the grid stability single-objective function is automatically increased, forcing the optimization algorithm to prioritize grid safety. When it is detected that a user has selected "emergency charging" mode, the weighting coefficient of the charging efficiency single-objective function is increased accordingly. The specific values ​​of the weighting coefficients are calculated in real-time using a lightweight rule engine or a small neural network, ensuring that the weighting allocation is closely adapted to the current system operating scenario.

[0042] Finally, the four single-objective functions are linearly weighted using the calculated dynamic weight coefficients to generate the final multi-objective optimization function. This function comprehensively represents the system's overall expectation of four objectives—charging efficiency, grid stability, user satisfaction, and battery life—at a specific moment. The value of this function directly represents the overall merit of a set of power allocation parameters, thus providing a clear and adaptive optimization direction for subsequent particle swarm optimization algorithms. The above construction process realizes a complete chain from deep feature semantic parsing to specific quantitative index calculation, and then to dynamic weight adaptive integration, ensuring that the optimization objective can both reflect the deep state of the system and flexibly respond to real-time operating condition changes.

[0043] Conventional methods typically use raw, unprocessed observational data (e.g., directly taking power values ​​as efficiency targets and instantaneous load values ​​as stability targets) to construct optimization objectives. This superficial approach to defining indicators fails to capture the complex nonlinear relationships and potential risk factors behind the data, resulting in optimization objectives that do not truly reflect the system's core requirements. For example, simply maximizing power may come at the cost of sacrificing battery life or exacerbating grid fluctuations, and conventional methods, due to their simple objective function definitions, struggle to inherently balance these contradictions. The target features in the embodiments of this specification originate from deep feature vectors. These features, refined through convolutional neural networks and gated recurrent units, already contain complex spatiotemporal correlations and causal information within the data. Therefore, optimization objectives constructed based on these deep features (such as charging efficiency features and grid load fluctuation features) are no longer isolated surface indicators but rather carry the deep semantics of the overall system operating state, making the optimization direction more precise and intelligent. Secondly, in determining the dimensional indicators, the embodiments in this specification employ more scientific and refined calculation methods. For example, charging efficiency is defined using the cumulative energy value per unit time rather than simple power; grid fluctuations are defined using the variance of the load sequence rather than instantaneous values; and user satisfaction is defined using a rating based on matching historical and current behavior. These indicator definitions better reflect the essence of the problem and avoid misleading the optimization process. Conventional methods mostly use fixed weights or require manual intervention to adjust weights, which cannot adapt to the dynamic and varied nature of charging scenarios. The embodiments in this specification automatically and instantly adjust the weight ratio of each single objective function in the overall objective by analyzing grid load and user preference data in real time. This not only greatly enhances the system's adaptability to different scenarios (such as grid peaks, urgent user needs, and battery protection priority), but also achieves true multi-objective dynamic trade-offs. This ensures that the final power allocation decision can guarantee the stability and efficiency of the macro system while meeting the personalized needs of individual micro-level users. Its comprehensive performance far exceeds that of traditional optimization function construction methods based on fixed weights or simple rules.

[0044] Step S104: Use the particle swarm optimization algorithm to solve the multi-objective optimization function, obtain the power allocation parameters of each charging pile, and send the power allocation parameters to the corresponding charging pile controller to adjust the output power of the charging pile and realize the power allocation of the charging pile.

[0045] The multi-objective optimization function is solved using a particle swarm optimization algorithm to obtain the power allocation parameters for each charging station. Specifically, this includes: initializing the particle swarm, where each particle represents a power allocation scheme, particle positions are randomly generated based on the available power range in the charging station's operational data, and particle velocities are set based on historical allocation schemes; calculating the fitness value of each particle based on the output of the multi-objective optimization function, where the fitness value is inversely proportional to the particle mass; updating the individual optimal position and the global optimal position of each particle by comparing the current fitness value with historical optimal values ​​to select the optimal position; adjusting the particle position and velocity according to the particle velocity update formula, performing iterative optimization until the convergence condition is met, and outputting the global optimal position as the power allocation parameter.

[0046] In one embodiment of this specification, after constructing the multi-objective optimization function, a particle swarm optimization algorithm is used to find the optimal power allocation parameters. First, the particle swarm is initialized. Each particle in the search space represents a complete power allocation scheme, i.e., a vector containing the power values ​​of all charging piles to be allocated. The initialization of particle positions is not completely random, but rather a bounded random generation based on the available power range of each charging pile recorded in the charging pile operation data. For example, the rated power upper limit and the currently available lower limit reported by each charging pile controller are read, and then an initial power value is randomly generated within a specific power range of that pile using a uniform or normal distribution. This ensures that the initial position of each particle is within a feasible physical constraint range, avoiding the search for invalid solutions.

[0047] The initialization of particle velocity is based on historical allocation schemes. Successful power allocation schemes in similar scenarios (such as similar total grid load and similar user demand distribution) are retrieved from the database recently. The magnitude and direction of power parameter changes between adjacent iterations in these historical schemes are analyzed and used as a reference to set the mean and variance of the initial particle velocity. This allows the particle swarm to start iterating from a relatively optimal initial search trend, accelerating the convergence process.

[0048] After initialization, the iterative optimization loop begins. In each iteration, the fitness value of each particle is calculated. This is done by substituting the particle's current position (i.e., a set of power allocation parameters) into the previously constructed multi-objective optimization function. This function comprehensively evaluates the achievement of multiple objectives under these parameters, such as charging efficiency, grid load fluctuation, user satisfaction, and battery life, and outputs a comprehensive evaluation value. In this design, the fitness value is inversely proportional to the particle's mass; that is, the smaller the function value, the better the overall performance of the power allocation scheme (because the function includes objectives that need to be minimized, such as grid fluctuation and battery degradation). Therefore, the goal of the optimization process is to find the particle position that minimizes the fitness value.

[0049] Next, the individual optimal position of each particle and the global optimal position of the entire population are updated. For each particle, its current fitness value is compared with its best fitness value experienced throughout the entire search history. If the current value is better (smaller), its individual historical optimal position is updated with the current position. At the same time, among all the individual optimal positions of particles, the position with the best (smallest) fitness value is selected as the current global optimal position. This global optimal position represents the best power allocation scheme found by the algorithm so far. Then, the velocity and position of each particle are adjusted according to the particle velocity update formula. The velocity update depends on three core components: the inertial component of the particle's current velocity, the cognitive component pointing to its own historical optimal position, and the social component pointing to the current global optimal position of the group. The inertial component keeps the particle's original motion trend, the cognitive component guides the particle to explore the best area it has previously discovered, and the social component prompts the particle to move closer to the best area discovered by the group. By linearly weighting and combining these three components, the new velocity vector of the particle is calculated. Subsequently, the particle moves to a new position based on the new velocity, thus generating a new set of trial solutions for power allocation parameters. During this process, it is continuously checked whether the new position remains within the available power constraints of each charging station. If it exceeds the limits, boundary handling is performed, such as clamping the power value to the maximum or minimum allowable value. The iterative process continues until a preset convergence condition is met. The convergence condition is typically set to either of two conditions: first, the fitness value of the globally optimal position improves by less than a very small threshold over several consecutive iterations, indicating that the solution quality has stabilized and the space for further optimization is limited; second, the algorithm reaches the preset maximum number of iterations to prevent infinite loops.

[0050] Once the convergence condition is met, the algorithm terminates and outputs the final determined global optimal position as the optimal power allocation parameter. Finally, these power allocation parameters are sent to the corresponding charging pile controller via a communication network (such as RS485 or Ethernet-based protocols). Each controller, upon receiving its assigned target output power value, invokes its internal power regulation module to precisely adjust the actual output power of the charging pile to the target value by controlling the duty cycle of power electronic switching devices (such as IGBTs), thus completing the closed loop of the entire intelligent power allocation process. The entire particle swarm optimization solution process simulates social collective intelligence, efficiently finding a near-globally optimal power allocation scheme in a complex multi-objective constraint space, and transforming the theoretical calculation results into actual control commands for the physical equipment.

[0051] The conventional methods described above often rely on simple rule bases (such as fixed priority rules), traditional linear programming, or greedy algorithms with a single objective. These methods cannot meet the global optimization objective when dealing with the inherent high-dimensionality, nonlinearity, and complex optimization problems involving conflicting multiple objectives in charging pile power allocation. The particle swarm optimization algorithm used in the embodiments of this specification is a global optimization algorithm based on swarm intelligence. It can explore multiple regions of the solution space in parallel. Through information sharing and cooperation among particles, it can effectively escape local optima and search towards the global optimum or near-global optimum. Secondly, this algorithm transforms the complex multi-objective problem into a single-objective problem for optimization through weighted summation. However, by introducing dynamic weight coefficients, this single-objective function can flexibly reflect the dynamic changes in the importance of various sub-objectives (such as efficiency, stability, satisfaction, and lifespan) under different scenarios, thus realizing a practical and efficient multi-objective optimization solution path. Compared with some complex multi-objective evolutionary algorithms, the particle swarm optimization algorithm usually has a faster convergence speed while ensuring solution quality. This is crucial for real-time power allocation applications that require rapid response (such as completing decisions within minutes). Furthermore, the algorithm initializes its initial speed and position using historical data, essentially injecting prior knowledge into the search process, further accelerating convergence and improving optimization efficiency. The final output power allocation parameters represent a balanced solution derived after comprehensively considering all key constraints and objectives, ensuring that the charging pile cluster achieves superior overall operational efficiency, higher grid friendliness, a better user experience, and longer equipment lifespan. This comprehensive decision-making quality surpasses that of conventional methods relying on empirical rules or simple mathematical models.

[0052] The method further includes: receiving execution status feedback from the charging pile controller, wherein the execution status feedback includes power adjustment results and communication status; if the execution status feedback indicates that the power allocation parameters have failed to execute, then retransmitting the power allocation parameters to the charging pile controller, with the maximum number of retransmissions set based on the historical success rate; if execution still fails after the maximum number of retransmissions is reached, then switching to a backup power allocation scheme, wherein the backup power allocation scheme is generated based on the historical average power allocation parameters; and updating the weight parameters of the multi-objective optimization function according to the adjusted power allocation strategy for subsequent power allocation decisions.

[0053] In one embodiment of this specification, after the power allocation parameters are sent, the system enters the execution status monitoring and dynamic adjustment phase. It continuously listens for execution status feedback signals from the charging pile controller and encapsulates them using a predefined communication protocol (such as an application layer protocol based on TCP / IP) before transmitting them back to the server. The execution status feedback mainly includes the power adjustment result and the communication status. The power adjustment result is generated by the charging pile controller after attempting to execute the received power allocation parameters, clearly indicating whether the adjustment was successfully completed. For example, the controller internally determines success by comparing the commanded power value with the actual output power value within the allowable error range. It also includes detailed information such as the specific actual output power, error code, or adjustment timestamp. The communication status reflects the health of the data transmission link, including network layer indicators such as signal strength, data transmission error rate, and connection latency. This helps distinguish whether the power adjustment failure is due to a fault in the charging pile's actuator or a network transmission problem.

[0054] Upon receiving feedback information, it is first parsed and categorized for storage, and then associated with the corresponding power allocation command, timestamp, and charging pile identifier and recorded in the operation log database. Next, it determines whether the feedback indicates a failure in the power allocation parameter execution. The failure criteria are complex, including not only explicitly indicated failure states in the power adjustment results, but also communication anomalies (such as continuous timeouts) or feedback information loss exceeding a preset time limit. Once a failure is determined, a retransmission mechanism is immediately initiated. First, the historical command execution success rate for that specific charging pile under similar operating conditions (such as similar time periods and similar target power values) is queried from the historical database. Based on this, the maximum number of retransmissions for this operation is dynamically set. For example, for charging piles with historically stable performance, the maximum number of retransmissions may be set lower, say two, prioritizing cases of transient communication interference; while for charging piles with a history of occasional failures, a higher number of retransmissions may be set, such as three or four, to provide more recovery opportunities. Before each retransmission, a brief, random backoff delay is introduced to avoid network congestion and check the communication link status. If a link interruption is detected, an attempt is made to re-establish the connection before sending the command. Feedback is continuously monitored during the retransmission process. If, after a certain period (e.g., after the maximum number of retransmissions is reached), the feedback received from the charging pile controller still indicates execution failure, the system determines that the charging pile is unavailable or in a severely faulty state, and then triggers the fault switching logic to activate the backup power allocation scheme. The backup scheme is a degradation operation strategy dynamically generated based on historical data. From historical operating data, records of power allocation parameters that have been successfully executed under similar global system states (e.g., similar total charging demand, total grid load level) are extracted. The average or median of these historical successful parameters are calculated to generate a set of conservative but reliable power allocation parameters as the backup scheme. The backup scheme bypasses the charging pile considered faulty or assigns it an extremely low safe power value (e.g., the minimum power required to maintain communication), and intelligently redistributes its original power demand to other normally operating charging piles in the cluster to ensure that the overall power allocation target is approximately achieved and to maintain the system's basic service capacity.

[0055] Finally, based on the actual situation reflected in the execution status feedback, the weight parameters in the multi-objective optimization function are updated for subsequent power allocation decisions. For example, if the feedback indicates that power command execution is difficult due to local grid fluctuations, the weight coefficient of the grid stability objective will be appropriately increased in subsequent optimizations to make the decision more robust. If the feedback shows that a certain user preference (such as "economic charging") is difficult to satisfy under the current network conditions, the weight of the user satisfaction objective may be temporarily fine-tuned. The update of the weight parameters is usually based on a predefined rule set or a lightweight reinforcement learning model, adjusting the strategy according to the negative reward signal of "execution failure". Through this complete closed loop of monitoring, judgment, retransmission, switching, and learning, it can not only effectively deal with single-point equipment failures or temporary communication interruptions, but also learn from failures and continuously optimize its decision model, thereby significantly improving the availability, reliability, and adaptability of the entire charging pile cluster.

[0056] Conventional methods typically employ an open-loop control mode of "one transmit, one receive" or only have a simple timeout retransmission mechanism. When a charging pile fails to execute optimal power allocation parameters due to its own malfunction or communication problems, conventional methods often lead to idle resources at that pile, a decline in overall system performance, and may even cause local grid problems due to power allocation imbalance. Furthermore, they cannot learn from the failure to avoid repeating the same mistakes. The embodiments in this specification construct an intelligent closed-loop system capable of sensing execution effects, responding to execution failures, and learning from them. By receiving and analyzing detailed feedback including power adjustment results and communication status, it can accurately diagnose the root cause of the problem—whether it's a faulty actuator at the charging pile or an unstable network link—and thus take more targeted measures. Setting a maximum number of retransmissions based on historical success rates avoids the resource waste caused by infinite retries and improves the probability of fault recovery. When retransmission is ineffective, it automatically switches to a backup power allocation scheme based on historical average parameters, ensuring degraded operation even when some units fail, providing basic service continuity, and greatly enhancing the system's resilience—something that open-loop control or simple retry mechanisms cannot achieve. Updating and optimizing function weight parameters based on execution feedback enables dynamic adjustment and continuous optimization based on feedback from the actual operating environment. For example, after multiple execution failures during periods of power grid fluctuation, the weight of the power grid stability objective is automatically strengthened, making subsequent decisions more conservative and robust. This adaptive characteristic allows the system to proactively adapt to environmental changes, accumulate operational experience, and ultimately achieve synergistic optimization of overall operational efficiency, reliability, and lifespan.

[0057] Through the embodiments described in this specification, by collecting and fusing data from four dimensions—charging piles, batteries, power grids, and users—the limitations of information silos in traditional methods are overcome. Simultaneously collecting data on battery health and real-time state of charge, macroscopic load stability of the power grid, and personalized user needs and preferences allows for the perception of richer contextual information. A combination of convolutional neural networks and gated recurrent units is used to extract deep features. Convolutional neural networks excel at capturing local spatial correlations in data, while gated recurrent units specialize in processing time-series data, learning long-term patterns from users' historical charging behavior and battery charge change trends. By combining the laws and short-term dynamics, the generated deep feature vector greatly improves the predictability and accuracy of decision-making. The power allocation problem is explicitly formulated as a multi-objective optimization problem that includes charging efficiency, grid stability, user satisfaction, and battery life. By constructing a multi-objective optimization function, it can adapt to dynamically changing needs. For example, during the night when the grid load is light, charging efficiency and user satisfaction can be appropriately emphasized; while during the evening when electricity consumption is high, grid stability should be prioritized. The particle swarm optimization algorithm is used to solve the multi-objective function, ensuring that the final output power allocation parameters are a comprehensive optimal solution obtained through a fully intelligent search.

[0058] This specification also provides an embodiment of a charging pile power distribution device based on artificial intelligence, such as... Figure 2 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method.

[0059] This specification also provides a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0060] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0061] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A power allocation method for charging piles based on artificial intelligence, characterized in that, The method includes: Collect multi-dimensional charging data of the target charging pile cluster, wherein the multi-dimensional charging data includes charging pile operation data, charging battery data, grid load data and user charging preference data; Data fusion is performed based on the multi-dimensional charging data to generate a fused data vector. A convolutional neural network is used to extract local features from the fused data vector, and a gated recurrent unit is used to extract time series features to generate a deep feature vector. Based on the deep feature vector, a multi-objective optimization function is constructed, wherein the multi-objective optimization function includes charging efficiency objective, grid load fluctuation objective, user satisfaction objective, and battery life objective; The multi-objective optimization function is solved using the particle swarm optimization algorithm to obtain the power allocation parameters for each charging pile. These power allocation parameters are then sent to the corresponding charging pile controller to adjust the output power of the charging pile and achieve power allocation for the charging pile.

2. The charging pile power allocation method based on artificial intelligence according to claim 1, characterized in that, Based on the multi-dimensional charging data, data fusion is performed to generate a fused data vector, specifically including: The charging pile operation data, charging battery data, grid load data, and user charging preference data in the multi-dimensional charging data are preprocessed to generate preprocessed data. The preprocessing operations include outlier removal, data normalization, and time synchronization. Outlier removal is based on the standard deviation method to identify and remove data points that deviate from the normal range. Data normalization uses the minimum-maximum scaling method to map data of different dimensions to a unified interval. Time synchronization uses the timestamp alignment method to unify multi-source data to the same time reference. Based on the preprocessed data, an attention mechanism is used to assign data weights for weighted processing to determine the weighted data. The grid load data is the first weight, the rechargeable battery data is the second weight, and the user charging preference data is the third weight. The first weight, the second weight, and the third weight are trained based on historical data and are used to reflect the degree of influence of the data on power allocation. The weighted data is concatenated to generate a fused data vector, which is then input into the feature extraction network.

3. The charging pile power allocation method based on artificial intelligence according to claim 1, characterized in that, The fused data vector is extracted using a convolutional neural network, and time-series features are extracted using a gated recurrent unit to generate a deep feature vector, specifically including: The fused data vector is input into a convolutional neural network, and the convolutional layer performs convolution operations on the power grid load fluctuation and the charging pile temperature change to extract local spatial features and generate local feature vectors. The local feature vector is input into the gated recurrent unit, and the user charging behavior and battery state of charge change are sequentially modeled through the recurrent layer to extract time-dependent features and generate time series feature vectors. The local feature vector and the time series feature vector are concatenated and then dimensionality is reduced through a fully connected layer to generate a deep feature vector, which is used to construct a multi-objective optimization function.

4. The charging pile power allocation method based on artificial intelligence according to claim 1, characterized in that, Based on the aforementioned deep feature vectors, a multi-objective optimization function is constructed, specifically including: Multidimensional target features are extracted from the deep feature vector, wherein the multidimensional target features include charging efficiency features, grid load fluctuation features, user satisfaction features, and battery life features; Based on the multidimensional target features, dimensional indicators for each dimension are determined, and a multi-objective optimization function is generated according to the multiple dimensional indicators. The dimensional indicators include charging efficiency indicators, grid load fluctuation indicators, user satisfaction indicators, and battery life indicators.

5. The charging pile power allocation method based on artificial intelligence according to claim 4, characterized in that, Based on the aforementioned multidimensional target features, dimensional indicators for each dimension are determined, specifically including: Based on the output power and charging time in the charging pile operation data, the cumulative value of charging energy per unit time is calculated to determine the charging efficiency index. Based on the real-time load rate in the power grid load data, the variance of the load sequence is calculated to determine the power grid load fluctuation index. The matching score is calculated based on the historical charging modes and current selections in the user charging preference data, and the user satisfaction index is determined. Based on the health status and state of charge in the rechargeable battery data, the battery degradation coefficient is calculated using a pre-built battery degradation model to determine the battery life index.

6. The charging pile power allocation method based on artificial intelligence according to claim 4, characterized in that, Based on the aforementioned dimensional indicators, a multi-objective optimization function is generated, specifically including: Maximize the charging efficiency index to construct a single objective function for charging efficiency, and maximize the user satisfaction index to construct a single objective function for user satisfaction. Minimize the grid load fluctuation index to construct a single objective function for grid stability, and minimize the battery life index to construct a single objective function for battery life. Based on the real-time collected power grid load data and user charging preference data, a dynamic weighting coefficient is calculated; The single-objective functions of charging efficiency, grid stability, user satisfaction, and battery life are linearly weighted and combined using the dynamic weighting coefficients to generate the multi-objective optimization function.

7. The power allocation method for charging piles based on artificial intelligence according to claim 1, characterized in that, The multi-objective optimization function is solved using the particle swarm optimization algorithm to obtain the power allocation parameters for each charging pile, specifically including: Initialize the particle swarm, where each particle represents a power allocation scheme, the particle position is randomly generated based on the available power range in the charging pile operation data, and the particle velocity is set based on the historical allocation scheme. Based on the output of the multi-objective optimization function, the fitness value of each particle is calculated, wherein the fitness value is inversely proportional to the particle mass; Update the individual particle's optimal position and the global optimal position by comparing the current fitness value with the historical optimal value, and select the optimal position. The particle position and velocity are adjusted according to the particle velocity update formula, and iterative optimization is performed until the convergence condition is met. The global optimal position is then output as the power allocation parameter.

8. The charging pile power allocation method based on artificial intelligence according to claim 1, characterized in that, The method further includes: Receive execution status feedback from the charging pile controller, wherein the execution status feedback includes power adjustment results and communication status; If the execution status feedback indicates that the power allocation parameters have failed to be executed, the power allocation parameters will be resent to the charging pile controller. The maximum number of resentments is set based on the historical success rate. If the retransmission still fails after reaching the maximum number of retransmissions, the system switches to the backup power allocation scheme, which is generated based on historical average power allocation parameters. The weight parameters of the multi-objective optimization function are updated according to the adjusted power allocation strategy for subsequent power allocation decisions.

9. A power distribution device for charging piles based on artificial intelligence, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to perform the method as described in any one of claims 1-8.

Images