Electric vehicle cluster parameter prediction method and device and computer equipment

Abstract

The invention provides an electric vehicle cluster parameter prediction method and device and computer equipment, and the method comprises the steps: carrying out the pre-training of a bidirectional long-short-term memory network according to commercial charging pile data, and enabling the bidirectional long-short-term memory network to be used for predicting the charging behavior parameters of an electric vehicle in combination with a historical charging track and a future charging trend; based on authorized household charging pile data, a transfer learning algorithm is adopted to carry out fine tuning on the pre-trained bidirectional long-short-term memory network, and a charging behavior prediction model is obtained; inputting the historical charging data of each electric vehicle into a charging behavior prediction model to obtain predicted charging data of each electric vehicle; and according to each piece of predicted charging data, constructing a monomer feasible region of each electric vehicle, and according to each monomer feasible region, obtaining a cluster feasible region of the electric vehicle cluster through sino polyhedron modeling and Minkowski summation. Therefore, the problems of limited individual data and long time span are relieved, and the accuracy and practicability of electric vehicle cluster parameter prediction are further improved.

Description

Technical Field

[0001] The present application relates to the technical field of electric vehicle scheduling, and in particular to a method, device and computer equipment for predicting electric vehicle cluster parameters. Background Art

[0002] With the increasing popularity of electric vehicles (EVs), an increasing number of charging stations are being connected to the power grid, posing a significant challenge to transformer capacity in distribution areas. To ensure stable grid operation and improve energy efficiency, accurately predicting EV charging demand and rationally scheduling charging activity at each charging station are crucial. Currently, parameter prediction for EV clusters relies primarily on in-depth analysis of their spatiotemporal distribution characteristics and the dynamics of charging demand, providing a scientific basis for grid scheduling and demand response.

[0003] Among existing technical solutions, recurrent neural networks (RNNs) and their improved long short-term memory (LSTM) networks are the mainstream prediction models. RNNs, through their recurrent structure, capture the temporal dependencies of sequences and possess a certain degree of memory capacity. LSTMs, building on this, introduce forget gates, input gates, and output gates, effectively alleviating the vanishing gradient problem and enhancing their ability to model long sequences of data. Furthermore, in recent years, prediction accuracy has been improved through the integration of robust optimization and generative adversarial networks.

[0004] It's worth noting that the overall prediction effectiveness of electric vehicle clusters relies on accurate modeling of the charging behavior of individual EV users. Due to the limited amount of individual user data and the long time span, traditional RNNs are prone to gradient explosion during training, making it difficult for the model to converge effectively. Furthermore, traditional LSTMs, due to their unidirectional temporal modeling, lack the ability to incorporate future trends. This limits the ability to accurately predict and respond to individual charging needs in real time, thus affecting the accuracy and practicality of overall cluster parameters. Summary of the Invention

[0005] The purpose of this application is to solve at least one of the above technical deficiencies, especially the technical deficiencies in the prior art of low accuracy and practicality of cluster parameters.

[0006] In a first aspect, the present application provides a method for predicting electric vehicle cluster parameters, the method comprising:

[0007] Based on the acquired commercial charging station data, a pre-built bidirectional long short-term memory network is pre-trained. The bidirectional long short-term memory network is used to combine historical charging trajectories and future charging trends to predict the charging behavior parameters of electric vehicles.

[0008] Based on the obtained authorized home charging pile data, a transfer learning algorithm is used to fine-tune the pre-trained bidirectional long short-term memory network to obtain a charging behavior prediction model;

[0009] Input the acquired historical charging data of each electric vehicle into the charging behavior prediction model to obtain the predicted charging data of each electric vehicle;

[0010] According to the predicted charging data, the single-unit feasible domain of each electric vehicle is constructed, and based on each single-unit feasible domain, the cluster feasible domain of the electric vehicle cluster is obtained through Chino polyhedron modeling and Minkowski summation.

[0011] In one embodiment, based on the obtained authorized home charging pile data, a transfer learning algorithm is used to fine-tune the pre-trained bidirectional long short-term memory network to obtain a charging behavior prediction model, including the following steps:

[0012] Based on the preset objective function, the pre-trained bidirectional long short-term memory network is iteratively updated using authorized home charging pile data until the preset iteration end conditions are met to obtain a charging behavior prediction model. The objective function is used to minimize the maximum mean difference between commercial charging pile data and home charging pile data in the feature mapping space.

[0013] In one embodiment, the objective function is expressed as:

[0014]

[0015] in, represents the model parameters of the pre-trained bidirectional long short-term memory network, Represents the empirical distribution of commercial charging pile data On the data After mapping function After the expectations, Represents the empirical distribution of home charging pile data On the data After mapping function After the expectations, Used to map input data into a high-dimensional feature space.

[0016] In one embodiment, the expression of the single-unit feasible region of each electric vehicle is:

[0017]

[0018] in, represents the maximum discharge power of the nth electric vehicle, represents the power of the nth electric car at time t, represents the maximum charging power of the nth electric vehicle, Indicates the charging start time. Indicates the charging end time. Indicates the initial charge. represents the power of the nth electric car at time t, Indicates the estimated charging capacity. represents the maximum dischargeable capacity of the nth electric vehicle, Indicates the charge and discharge efficiency, Indicates the time interval, Indicates the maximum chargeable capacity of the nth electric vehicle.

[0019] In one embodiment, the step of obtaining the cluster feasible region of the electric vehicle cluster by using Chino polyhedron modeling and Minkowski summation based on the feasible regions of each individual unit includes:

[0020] Construct a monomer Chino polyhedron model corresponding to each monomer feasible region;

[0021] The monomer symmetry center and monomer expansion range of each monomer Chino polyhedron model are superimposed to obtain the aggregate symmetry center and aggregate expansion range, and the cluster feasible domain is obtained based on the aggregate symmetry center and aggregate expansion range.

[0022] In one embodiment, the step of constructing a monomer Chino polyhedron model corresponding to each monomer feasible region includes:

[0023] Model each monomer feasible region as a convex polyhedron and construct the corresponding generating vector;

[0024] Generate the corresponding normal vector matrix according to each plane of each convex polyhedron, and calculate the extension distance of each convex polyhedron on each normal vector;

[0025] According to the generating vector, normal vector matrix and extension distance of each convex polyhedron, the largest inscribed Chino polyhedron is found in each convex polyhedron to obtain each monomer Chino polyhedron model.

[0026] In one embodiment, the bidirectional long short-term memory network includes a forward long short-term memory network and a reverse long short-term memory network, the forward long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the past time point, and the reverse long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the future time point.

[0027] In a second aspect, the present application provides an electric vehicle cluster parameter prediction device, the device comprising:

[0028] A network pre-training module is used to pre-train a pre-built bidirectional long-short-term memory network based on acquired commercial charging station data. The bidirectional long-short-term memory network is used to combine historical charging trajectories and future charging trends to predict charging behavior parameters of electric vehicles.

[0029] The charging behavior prediction model determination module is used to fine-tune the pre-trained bidirectional long short-term memory network based on the obtained authorized home charging pile data using a transfer learning algorithm to obtain a charging behavior prediction model;

[0030] A predicted charging data determination module is used to input the acquired historical charging data of each electric vehicle into the charging behavior prediction model to obtain the predicted charging data of each electric vehicle;

[0031] The cluster feasible domain determination module is used to construct the individual feasible domain of each electric vehicle according to the various predicted charging data, and obtain the cluster feasible domain of the electric vehicle cluster through Chino polyhedron modeling and Minkowski summation based on each individual feasible domain.

[0032] In a third aspect, the present application provides a storage medium: the storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the electric vehicle cluster parameter prediction method as described in any one of the above embodiments.

[0033] In a fourth aspect, the present application provides a computer device, comprising: one or more processors, and a memory;

[0034] The memory stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the steps of the electric vehicle cluster parameter prediction method in any one of the above embodiments are performed.

[0035] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0036] This application proposes a method for predicting electric vehicle cluster parameters, addressing the existing issues of sparse individual user data and one-way time series modeling. It proposes a systematic solution with enhanced predictive and generalization capabilities. By introducing a bidirectional long short-term memory (BiLSTM) network, the model can simultaneously capture historical charging trajectories and future charging trends, enabling bidirectional modeling of charging behavior over time, significantly improving the comprehensiveness and accuracy of predictions. Furthermore, a transfer learning algorithm is introduced to pre-train the model using high-quality data from commercial charging stations and fine-tune it with data from authorized home charging stations. This effectively improves the model's adaptability and generalization capabilities under conditions of limited individual data and long time spans, alleviating the data sparsity issue. After completing the prediction of individual vehicle charging behavior, the feasible region for each electric vehicle is constructed based on the prediction results. Using the Chino polyhedron as the modeling foundation and combining it with the Minkowski summation, a cluster feasible region model is formed that reflects the overall power and capacity boundaries of the cluster. This accurately characterizes the operating characteristics of the electric vehicle group, providing a scientific, controllable, and responsive decision-making basis for power grid scheduling and load management, and comprehensively improving the accuracy and practicality of electric vehicle cluster parameter prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0038] Figure 1 A flow chart of a method for predicting electric vehicle cluster parameters provided in an embodiment of the present application;

[0039] Figure 2 A schematic diagram of the structure of an electric vehicle cluster parameter prediction device provided in an embodiment of the present application;

[0040] Figure 3 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0042] The present application provides a method for predicting electric vehicle cluster parameters. The following embodiments are explained using the method applied to computer equipment as an example. It can be understood that the computer equipment can be various devices with data processing capabilities, including but not limited to single servers, server clusters, personal laptops, desktop computers, etc. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards, and provide corresponding operation portals for users to choose to authorize or refuse. Figure 1 As shown, the method may include the following steps:

[0043] S101: Based on the acquired commercial charging pile data, a pre-built bidirectional long short-term memory network is pre-trained. The bidirectional long short-term memory network is used to combine historical charging trajectories and future charging trends to predict the charging behavior parameters of electric vehicles.

[0044] Commercial charging station data refers to a structured data set related to user charging behavior, collected by electric vehicle charging stations deployed in public areas. This data includes, but is not limited to, charging start and end times, charging power, charging station location, and user identification information. This data has wide coverage and is updated frequently, making it a crucial data source for training charging behavior prediction models. The Bidirectional Long Short-Term Memory (Bi-LSTM) network is an improved recurrent neural network architecture, consisting of a forward LSTM and a backward LSTM. It can simultaneously capture both forward and backward temporal dependencies in sequential data, thereby enhancing the model's ability to model complex time series. Pretraining is the process of initially training a neural network model on large-scale historical data to obtain model parameters with general predictive capabilities, providing a foundation for subsequent transfer learning or fine-tuning. Historical charging trajectories represent the charging behavior sequences of electric vehicle users at different time points, reflecting their long-term electricity usage patterns. Future charging trends are the inference of possible future charging behavior patterns, which helps improve the model's ability to predict future demand. Charging behavior parameters are key features that characterize user charging demand, including predicted charging time, charging power, and duration.

[0045] The computer first receives real-time raw charging data uploaded from multiple commercial charging pile systems via a communication interface. This data is processed through data cleaning, format standardization, and feature extraction modules to transform it into a unified structured time series dataset. After completing data preprocessing, the computer automatically divides the data into training, validation, and test sets for subsequent model training tasks.

[0046] The computer device loads a pre-built bidirectional long short-term memory network model structure. This model consists of an input layer, a bidirectional LSTM encoding layer, a fully connected layer, and an output layer. Model structure parameters such as hidden layer dimensions and time step size are automatically adjusted based on data characteristics. Subsequently, the training engine module is called, using historical charging trajectories from commercial charging station data as the input sequence and actual charging behavior within a future time period as the prediction target to perform supervised learning. During the training process, an optimization algorithm can be used to iteratively update the network parameters, continuously improving model accuracy by minimizing the loss function between the predicted and actual values.

[0047] To further incorporate future charging trends, the computer device enhances the training process. Specifically, by introducing auxiliary features outside the time window as external input nodes, such as holiday identifiers, weather information, and electricity price fluctuations, the ability to detect sudden charging demand is improved. Furthermore, the self-attention mechanism module adjusts the importance weights of each time step in the time series, allowing the model to more accurately capture key nodes. After training is complete, the model parameters are solidified and stored in the storage module of the computing device, forming a preliminary prediction model with generalization capabilities, which can be used for subsequent migration or real-time prediction.

[0048] Pre-training a bidirectional long short-term memory network using commercial charging pile data can fully utilize the advantages of large-scale sample data coverage and high data density of commercial charging piles. Through pre-training, the model's overall understanding of electric vehicle users' charging behavior is enhanced, and a stable and universal parameter initialization state is formed. The structural characteristics of the bidirectional long short-term memory network not only enhance the model's ability to retain historical information, but also significantly improve its perception of future trends, making the prediction results more closely aligned with the changing patterns of user behavior in real-world scenarios. By acquiring key behavioral features from large-scale data, the pre-training process effectively avoids underfitting or gradient instability problems that arise in subsequent individual modeling, improving the accuracy and convergence speed of the final prediction model. Therefore, it is possible to achieve detailed modeling of complex charging behavior while ensuring computational efficiency, providing reliable support for subsequent predictions of electric vehicle clusters and grid scheduling.

[0049] S102: Based on the obtained authorized home charging pile data, a transfer learning algorithm is used to fine-tune the pre-trained bidirectional long short-term memory network to obtain a charging behavior prediction model.

[0050] Acquisition of home charging pile data can be accomplished through data exchange between a home charging pile terminal equipped with communication capabilities and a computer device executing the electric vehicle cluster parameter prediction method. Users can set authorization through the client program or local management interface, allowing the system to collect relevant data from their home charging scenarios upon initial access. Users can also modify the scope of authorization or revoke authorization at any time through the configuration page. The computer device receives relevant data via a secure communication protocol only with user authorization and uses it to train the charging behavior prediction model.

[0051] Home charging station data refers to charging behavior records generated by users using private charging stations in their homes. This data includes, but is not limited to, charging time period, charging power, user ID, meter number, and residential location. Compared to commercial charging station data, home charging station data features distinct individual user characteristics, smaller data size, and greater behavioral stability. Transfer learning algorithms are machine learning techniques that leverage existing trained models for a source task to improve learning performance through parameter transfer, structure transfer, or feature transfer when training data for the target task is insufficient. A pre-trained bidirectional long short-term memory network refers to a Bi-LSTM model that has been trained using large-scale commercial charging station data and exhibits preliminary generalization capabilities. Fine-tuning refers to an optimization process that, while retaining the original model parameters, retrains the model for a limited number of rounds using target scenario data to better align the model with the target data distribution. A charging behavior prediction model is a machine learning model fine-tuned based on home charging data that can be used to predict the future charging behavior of individual electric vehicles. Its output includes key parameters such as estimated charging time, power, and duration.

[0052] The computer receives data from multiple home charging pile terminals via the data access module. Because home charging data can be inconsistent in format and quality, the data standardization module can be used to standardize the input data and remove outliers. Next, the standardized home charging behavior sequence is combined with its corresponding contextual features to construct an input dataset for the home scenario, which is then mapped and compared with the original commercial scenario data structure.

[0053] The computer device loads a previously pre-trained bidirectional LSTM model and enables the transfer learning module. This module first freezes the model's underlying general feature extraction layers, such as the LSTM encoding layer, preserving the long-term temporal dependency structure learned in commercial scenarios. It then enables gradient updates only for the model's high-level feature fusion and output layers, gradually adapting the model to the charging behavior characteristics of home scenarios. During training, an adaptive learning rate control mechanism can be used to dynamically compensate for data sparsity in home scenarios and avoid overfitting.

[0054] Based on the model's performance on the validation set, the system continuously evaluates model accuracy and fine-tunes its structure until it meets the set prediction accuracy threshold or reaches the upper limit of the number of iterations. Finally, the fine-tuned model parameters are solidified and stored, forming a new charging behavior prediction model that can learn individual household user behavior trends and can be used for subsequent real-time predictions or integrated calculation of group parameters.

[0055] Fine-tuning the pre-trained model using a transfer learning algorithm leverages common behavioral characteristics learned in commercial scenarios while quickly adapting to the individual differences in charging behavior among home users, resulting in a highly accurate prediction model with a small sample size. Fine-tuning the model using data characteristics from home charging scenarios enhances scenario adaptability and the ability to capture individual behaviors. This approach not only improves the accuracy of individual-level predictions of charging behavior parameters but also enhances the model's practical application in grid-side cluster management tasks.

[0056] S103: Inputting the acquired historical charging data of each electric vehicle into a charging behavior prediction model to obtain predicted charging data of each electric vehicle.

[0057] Historical charging data refers to the collected and stored records of each EV's charging behavior over a period of time. These records include charging start and end times, charging duration, charging power, location information, timestamp, SOC (State of Charge), voltage, current, and other parameters. This data reflects the actual charging patterns and habits of EV users and serves as the core input for modeling individual user behavior. Predictive charging data refers to the structured information output by the prediction model regarding each EV's likely future charging behavior. This information, typically presented as a time series, is used in scenarios such as scheduling, optimization, and load forecasting.

[0058] The computer device communicates with the EV charging management platform through an interface, periodically pulling or receiving historical EV charging records in real time. This data is centrally connected to the data receiving module and preprocessed by the data cleaning module, including time series alignment, missing value filling, outlier data removal, and unit normalization. The data is then grouped according to the EV's unique identifier to ensure that model predictions are personalized for each vehicle.

[0059] Each set of preprocessed historical charging data is packaged according to a predefined input format, such as constructing a fixed-length sliding window. It is then supplemented with contextual features such as date type, weather information, and location category. This data is then fed into a fine-tuned charging behavior prediction model. Based on a bidirectional long-short-term memory network architecture, this model combines current sequence information with future trend information to generate a predicted charging behavior result for the next time period. This process can be executed in parallel for all EVs to be predicted using a batch processing mechanism, improving overall processing efficiency.

[0060] The model outputs predicted charging data in a structured manner and stores it. Predictions can include fields such as hourly charging probability, maximum load power, and expected start / end times for the next day or week. The output parsing module parses the model results into standard JSON or CSV formats and stores them in a database. This module also supports data interfacing with the power grid dispatching platform for subsequent practical applications such as cluster-level dispatch optimization, regional load forecasting, and load transfer control. To ensure data traceability and predictive performance, the corresponding relationship between model inputs and outputs is also recorded for subsequent model evaluation and retraining.

[0061] By inputting each electric vehicle's historical charging data into a charging behavior prediction model, it is possible to combine the long-term behavior habits and cycle characteristics of individual users to predict their future charging behavior trajectories, giving the model greater personalized expression capabilities. This approach can effectively improve the accuracy and timeliness of prediction data, providing a more granular and realistic input data source for grid scheduling, distribution optimization, and demand response. Furthermore, this method enables rapid modeling of individual behaviors in large-scale electric vehicle clusters, enhancing the scalability and intelligent control capabilities of the overall prediction, and improving the intelligence and stability of grid operations.

[0062] S104: constructing a single-unit feasible region for each electric vehicle based on the predicted charging data, and obtaining a cluster feasible region for the electric vehicle cluster through Chino polyhedron modeling and Minkowski summation based on each single-unit feasible region.

[0063] The single-unit feasible domain refers to the multidimensional geometric representation of each electric vehicle within a specific time-power space, constrained by its predicted charging data. It reflects the set of reasonable charging strategies that the electric vehicle may adopt during different time periods. Zonotope modeling is a method for representing and operating on uncertainty sets in multidimensional space, suitable for modeling linearly additive uncertain sets. Minkowski summation, a process of performing vector sum operations on multiple sets in mathematical geometry, is used to synthesize multiple single-unit feasible domains to form an overall cluster feasible domain. Ultimately, the cluster feasible domain, which is the overall range of charging possibilities for all electric vehicles, is constructed to support capacity analysis, dispatch optimization, and safety verification of power grids.

[0064] The computing device first analyzes the acquired predicted charging data on a vehicle-by-vehicle basis, extracting the behavioral boundaries of each electric vehicle within the predicted period, including charging power range, time period, and frequency distribution. By setting adjustable rule functions, the predicted value for each vehicle is converted into a set of interval points in the two-dimensional space of charging time and power. This is then further encapsulated into a multidimensional vector set, providing the data foundation for the subsequent construction of the individual feasible domain.

[0065] The Chino polyhedron modeling algorithm module is invoked to map each electric vehicle's time-power data vector into a zonotope structure. Specifically, based on each vehicle's predicted charging power fluctuation range and available charging period, it is represented as a geometric solid based on a center point and generated vectors, ensuring closure, additiveness, and efficient computational performance. Each vehicle's zonotope data structure is then standardized and stored for subsequent composite calculations.

[0066] All individual feasible regions are integrated using a Minkowski summation operation. This operation is performed in the computer using a vector superposition mechanism, ensuring that the feasible region of the combined cluster is quickly generated while preserving the original geometric structure. This cluster feasible region is represented as a high-dimensional data structure, providing input for subsequent grid-side load scheduling, current capacity analysis, or intelligent policy control modules. Furthermore, to improve system response speed, the Minkowski summation results can be pruned, compressed, or boundary sampled for visualization and rapid retrieval.

[0067] By constructing a single-vehicle feasible region based on predicted charging data and then synthesizing the cluster feasible region using Chino polyhedron modeling and Minkowski summation, not only can the time and power uncertainties of individual behaviors be efficiently expressed, but the overall adjustability and boundary conditions of the charging load can also be accurately reflected at the cluster level. This approach improves the combinability and interpretability of multi-vehicle cluster behavior, avoiding the shortcomings of traditional statistical models that make it difficult to characterize spatial overlap and boundary fluctuations. At the same time, the precisely synthesized cluster feasible region can provide a key basis for capacity verification and peak load forecasting of the distribution network, enabling controllable scheduling and dynamic optimization of charging behavior, thereby effectively improving the safety and load adaptability of power grid operation.

[0068] In the above-mentioned embodiment, a systematic solution with enhanced predictive and generalization capabilities is proposed to address the existing issues of sparse individual user data and unidirectional time series modeling. By introducing a bidirectional long short-term memory (BiLSTM) network, the model can simultaneously capture historical charging trajectories and future charging trends, achieving bidirectional modeling of charging behavior in the temporal dimension, significantly improving the comprehensiveness and accuracy of the prediction. Furthermore, a transfer learning algorithm is introduced to pre-train the model using high-quality data from commercial charging stations and fine-tune it with data from home charging stations. This effectively improves the model's adaptability and generalization capabilities under conditions of limited individual data and long time spans, alleviating the data sparsity issue. After completing the prediction of individual vehicle charging behavior, the feasible region for each electric vehicle is constructed based on the prediction results. Using the Chino polyhedron as the modeling foundation and combining it with the Minkowski summation, a cluster feasible region model is formed that reflects the overall power and capacity boundaries of the cluster. This accurately characterizes the operating characteristics of the electric vehicle group, providing a scientific, controllable, and responsive decision-making basis for grid scheduling and load management, and comprehensively improving the accuracy and practicality of electric vehicle cluster parameter prediction.

[0069] In one embodiment, based on the acquired authorized home charging pile data, a transfer learning algorithm is used to fine-tune the pre-trained bidirectional long short-term memory network to obtain a charging behavior prediction model, including the following steps:

[0070] Based on the preset objective function, the pre-trained bidirectional long short-term memory network is iteratively updated using authorized home charging pile data until the preset iteration end conditions are met to obtain a charging behavior prediction model. The objective function is used to minimize the maximum mean difference between commercial charging pile data and home charging pile data in the feature mapping space.

[0071] The objective function is a mathematical expression defined during the optimization process, measuring the degree of discrepancy between the model's predicted output and the desired target. In this scenario, it takes the form of Maximum Mean Discrepancy (MMD), which measures the distribution distance between data from two different domains in a high-dimensional feature map space. The preset iteration termination criteria are predefined criteria for determining training termination, such as the error convergence threshold, maximum number of iterations, and gradient change rate, ensuring an effective and stable training process.

[0072] The computer first performs standardized preprocessing on the home and commercial charging station data, unifying the data format, time window, and normalization strategy. This step allows the data acquisition module to map the raw data into a unified feature space, ensuring consistent and representative input for subsequent model training. Next, a BiLSTM model pre-trained using commercial charging station data is loaded as the base model, providing prior parameters for fine-tuning the training process.

[0073] Initialize the transfer learning module and load the preset objective function, using the Maximum Mean Difference (MMD) as the loss metric. In each iteration, samples of commercial and home charging stations are fed into the BiLSTM network. The feature extraction layer maps these samples to a high-dimensional feature space. The objective function then calculates the difference in the mean distribution between the two sample sets. This result serves as the basis for gradient backpropagation, guiding model parameter updates to minimize cross-domain discrepancies. During this process, the loss trend is continuously recorded, and training termination is dynamically determined based on preset termination criteria.

[0074] When the objective function converges or reaches the maximum number of iterations, the parameter update process is terminated and the final trained charging behavior prediction model is output. This model has the ability to generalize to both commercial and home scenarios, accurately reflecting the charging behavior of electric vehicles in different scenarios in subsequent prediction tasks. To improve model deployment efficiency, the trained model can also be quantized, compressed, and adapted for multiple platforms to support deployment on edge devices or in-vehicle computing units.

[0075] By iteratively updating the pre-trained bidirectional long short-term memory network using the maximum mean difference as the objective function, the distribution deviation between commercial charging pile data and home charging pile data in the feature space can be effectively reduced, thereby achieving domain adaptation of data between different scenarios and improving the generalization ability of the prediction model. Using home charging pile data for fine-tuning training enables the model to learn charging behavior patterns that are more representative of home usage scenarios, compensating for the deviation caused by training only with commercial data. At the same time, iterative updates until the preset termination conditions are met can prevent overfitting and ensure training efficiency, ultimately resulting in a high-performance charging behavior prediction model that retains the learning ability of the original model while adapting to new scenarios. This ensures prediction accuracy while improving the model's adaptability and robustness in multi-source data environments.

[0076] In one embodiment, the objective function is expressed as:

[0077]

[0078] in, represents the model parameters of the pre-trained bidirectional long short-term memory network, Represents the empirical distribution of commercial charging pile data On the data After mapping function After the expectations, Represents the empirical distribution of home charging pile data On the data After mapping function After the expectations, Used to map input data into a high-dimensional feature space.

[0079] In this embodiment, this expression represents the goal of maintaining knowledge transferability for commercial scenarios while simultaneously aligning the model's response to home scenario data through training, reducing inter-domain drift and thereby improving the model's generalization. By using this objective function, a computer device can transfer learning a pre-trained bidirectional long short-term memory (BiLSTM) network, not only preserving the model's predictive capabilities for commercial charging pile scenarios but also gradually adapting to the data distribution of home charging piles. This objective function minimizes the maximum mean difference between commercial and home data in the feature map space, enabling the model to continuously adjust its parameters during training, explicitly aligning the feature distributions of the different data domains and narrowing the gap between them in high-dimensional space. This process enhances the model's robustness to scenario changes, effectively avoiding performance degradation due to data domain differences, and preventing catastrophic forgetting during fine-tuning, ensuring that the model retains the effective features learned in commercial scenarios. Ultimately, this feature alignment mechanism makes the optimization process more transparent, helping to improve the model's stability and practicality in cross-scenario deployments.

[0080] In one embodiment, the expression of the single-unit feasible region of each electric vehicle is:

[0081]

[0082] in, represents the maximum discharge power of the nth electric vehicle, represents the power of the nth electric car at time t, represents the maximum charging power of the nth electric vehicle, Indicates the charging start time. Indicates the charging end time. Indicates the initial charge. represents the power of the nth electric car at time t, Indicates the estimated charging capacity. represents the maximum dischargeable capacity of the nth electric vehicle, Indicates the charge and discharge efficiency, Indicates the time interval, Indicates the maximum chargeable capacity of the nth electric vehicle.

[0083] I understand. is the power constraint, which means that at any time , No. The power of electric vehicles The power output must be within the physical limits of the device, between its maximum discharge power and maximum charging power. As a time constraint, it is stipulated that charging behavior can only occur within the predetermined charging time window, preventing the scheduling strategy from exceeding the actual allowed charging and discharging time period; It is the power constraint, which means that during the charging process, the power of the electric vehicle cannot be lower than the initial power , nor exceed the target power , to prevent overcharging or over-discharging; To achieve cumulative energy constraints, the total energy charged or released during the entire charging cycle must be between the maximum dischargeable capacity and the maximum chargeable capacity, further ensuring the rationality and stability of the charging and discharging behavior.

[0084] In this embodiment, the formula defines the feasible behavior range of each electric vehicle within a given time interval by imposing fine constraints on its charging and discharging power, state of charge, and total energy. With the help of this feasible domain, computer equipment can ensure that the operation of each vehicle meets physical and operational constraints when performing group scheduling or energy management, effectively preventing the occurrence of abnormal situations such as overcharging, over-discharging, and current overload. At the same time, this structured modeling method lays the foundation for the subsequent construction of the feasible domain of the cluster, and can further analyze and optimize the group characteristics through geometric methods. This method not only improves the algorithm's constraint processing capabilities in multi-objective scheduling scenarios, but also enhances the model's controllability, computational feasibility, and adaptability to real-world deployments, achieving optimal group energy efficiency while ensuring individual rationality.

[0085] In one embodiment, the steps of obtaining the cluster feasible region of the electric vehicle cluster by using Chino polyhedron modeling and Minkowski summation based on the feasible regions of each individual unit include:

[0086] Construct a monomer Chino polyhedron model corresponding to each monomer feasible region;

[0087] The monomer symmetry center and monomer expansion range of each monomer Chino polyhedron model are superimposed to obtain the aggregate symmetry center and aggregate expansion range, and the cluster feasible domain is obtained based on the aggregate symmetry center and aggregate expansion range.

[0088] Among them, the Chino polyhedron model is a geometric structure used for convex set modeling. It can represent a convex set by linearly scaling in all directions through the center of symmetry. In this scenario, it is used to characterize the charging and discharging behavior space of electric vehicles. The single symmetry center is the geometric center of the Chino polyhedron, representing the most representative scheduling power configuration of electric vehicles in the behavior space. The single expansion range refers to the amplitude of change allowed by the Chino polyhedron in each dimensional direction, reflecting the flexibility of the charging and discharging behavior of electric vehicles. The aggregate symmetry center is the composite center after superimposing the single symmetry centers of multiple electric vehicles in a vector manner. The aggregate expansion range is the overall behavior fluctuation boundary formed by superimposing all the single expansion ranges in the geometric space.

[0089] First, the computer acquires the feasible domain of each electric vehicle, including information such as power boundaries, power limits, and time windows. For each electric vehicle, a set of behavior space vectors is constructed from the time series power limits and their corresponding energy boundaries. This is then linearly mapped to form the input vector to be modeled.

[0090] Next, the computer constructs a Chino polyhedron model for each electric vehicle based on the individual feasible region information. Specifically, the center of symmetry of the electric vehicle is determined by calculating the geometric center of the individual feasible region. Furthermore, the maximum charge and discharge amplitudes in each time dimension are calculated to determine the generating vectors in each dimension, thereby generating a Chino polyhedron model. This model accurately describes the energy handling capabilities of each vehicle under controllable scheduling.

[0091] The computer then performs a vector summation of the individual symmetry centers of all electric vehicles to obtain the aggregate symmetry center. It also linearly superimposes the expansion and contraction ranges of each individual Chino polyhedron to obtain the aggregate expansion and contraction range. Based on these two geometric elements, the computer constructs a cluster feasible domain model according to the additive definition of Chino polyhedra. This model is then used to calculate, verify, and optimize the subsequent group scheduling strategy. This cluster model is numerically represented as a high-dimensional convex set generated by the aggregate symmetry centers and the aggregate expansion and contraction vector set, reflecting the overall energy scheduling potential and behavioral boundaries of the entire electric vehicle cluster.

[0092] By constructing a single Chino polyhedron model and superimposing its symmetry center with its telescopic range to generate a cluster feasible domain, computer equipment can uniformly represent the overall charging capacity boundary of an electric vehicle cluster in a geometric and structured manner, effectively simplifying the complexity of cluster behavior modeling. This approach not only preserves the behavioral flexibility of each electric vehicle at the individual level, but also preserves the directionality and volatility of group behavior through geometric synthesis, thereby providing an efficient and analyzable feasible solution space for subsequent scheduling optimization algorithms. Furthermore, this method improves the controllability of the scheduling center solution by summing the symmetry centers, and enhances the cluster's robustness to scheduling fluctuations by superimposing telescopic ranges, ultimately achieving efficient and reliable energy scheduling management.

[0093] In one embodiment, the step of constructing a monomer Chino polyhedron model corresponding to each monomer feasible region includes:

[0094] Model each monomer feasible region as a convex polyhedron and construct the corresponding generating vector;

[0095] Generate the corresponding normal vector matrix according to each plane of each convex polyhedron, and calculate the extension distance of each convex polyhedron on each normal vector;

[0096] According to the generating vector, normal vector matrix and extension distance of each convex polyhedron, the largest inscribed Chino polyhedron is found in each convex polyhedron to obtain each monomer Chino polyhedron model.

[0097] Among them, a convex polyhedron refers to a closed convex set constructed under the constraints of charging and discharging power, power, time, etc., which can fully describe the charging and discharging behavior boundary of an electric vehicle within a given time period. The generating vector is a linear combination basis vector used to construct the polyhedron structure, which represents the ability of the monomer feasible domain to change in all directions in the multidimensional space. The normal vector matrix is ​​a matrix composed of unit normal vectors corresponding to each face of the polyhedron, which is used to describe the geometric boundary properties of the polyhedron in all directions. The extension distance is the maximum distance covered by the polyhedron from the center point to the boundary in the direction of a specific normal vector, reflecting the size of the scheduling space of the model. The maximum inscribed Chino polyhedron refers to the Chino polyhedron with the largest volume that can be completely embedded inside a given convex polyhedron. Its structure consists of a center point and a set of adjustable and retractable generating vectors, representing the maximum scheduling flexibility model of electric vehicles within the constraints.

[0098] The computer device first further formalizes and models each feasible domain as a convex polyhedron, that is, the boundary of the set is characterized by a set of inequality matrices, thereby providing a basic structure for subsequent processing. After the modeling is completed, the computer device performs a structural analysis on each convex polyhedron. Specifically, all the faces that make up the polyhedron are identified, the unit normal vector of each face is extracted, and they are combined to form the corresponding normal vector matrix. On this basis, for each normal vector direction, starting from the geometric center of the polyhedron, its extension distance in that direction is calculated to obtain the scheduling capability boundary of the electric vehicle in different directions. These geometric parameters together reflect the spatial extension characteristics of the current electric vehicle within a given scheduling cycle.

[0099] After obtaining the normal vector matrix and extension distance, the computer device further performs a maximum inscribed model extraction operation. Specifically, with the geometric center of the polyhedron as the reference origin, under the guidance of the normal vector matrix, the largest embeddable Chino polyhedron is optimized and solved so that it does not exceed the boundary of the original convex polyhedron in any direction. By linearly scaling and combining the generated vectors, a set of scaled vectors that satisfy the boundary constraints of the convex polyhedron is constructed. This set and the center point are collectively defined as the single Chino polyhedron model of the electric vehicle. Finally, this model is saved for subsequent group model aggregation and energy scheduling analysis.

[0100] By modeling each individual feasible domain as a convex polyhedron, the computer equipment can fully express the behavioral boundaries of electric vehicles within the scheduling cycle in a geometric structure, forming a unified and processable mathematical model. Further extraction of the normal vector matrix and extension distance helps to accurately characterize the boundary restrictions of the polyhedron in all directions, thus providing a mathematical basis for the construction of the maximum inscribed Chino polyhedron. Finally, by optimizing the construction of the maximum inscribed Chino polyhedron, it not only ensures that the model always operates within the original constraints, but also maximizes the freedom of electric vehicle scheduling behavior. This approach not only improves the compactness and expressiveness of the model, but also greatly improves the synthesizability and computational efficiency of subsequent group modeling, laying a solid foundation for the unified modeling and optimized control of large-scale charging clusters.

[0101] In one example, the feasible domain modeling process for a single electric vehicle (EV) can be accurately characterized by Zonotope. This method first represents the feasible domain of each EV as a Zonotope set, which is mathematically expressed as:

[0102]

[0103] in, represents a scheduling vector, such as a vector of power or energy varying over time, is the symmetry center of the Zonotope, To generate vectors, representing the extension direction of the Zonotope in all directions, is the expansion coefficient in the corresponding direction, and its expansion range is limited by the constant .

[0104] In order to accurately describe the scheduling space of a single EV, after constructing the Zonotope structure, the normal vector matrix is ​​further constructed based on the normal vectors corresponding to each plane of the polyhedron. , and based on this, calculate the extension distance of the original convex polyhedron in each normal vector direction These extension distances are used to limit the boundaries of the maximum inscribed zonotope to ensure that the constructed model does not exceed the original feasible region.

[0105] Next, in order to find the largest inscribed zonotope from the original convex polyhedron, the following optimization problem is constructed:

[0106]

[0107] By solving this optimization problem, the largest symmetric Zonotope model of a single EV in its feasible domain is obtained, that is, the symmetry center is obtained. , generate vector matrix And the maximum telescopic range in all directions .

[0108] After obtaining the Zonotope model parameters of all monomeric EVs, the closed nature of the Zonotope can be used to directly complete the EV cluster modeling through Minkowski summation. At this time, superimposing multiple Zonotopes is equivalent to the symmetry center of each monomer model. and telescopic range Sum them separately to get the overall symmetry center of the EV cluster , and the overall scaling range after aggregation Finally, the power and capacity of the EV cluster at any time t are Can be classified into the aggregated Zonotope model , thereby depicting the power-capacity boundary area of ​​the entire EV cluster within the scheduling cycle, providing mathematical support for further cluster energy management and optimized scheduling.

[0109] In one embodiment, the bidirectional long short-term memory network includes a forward long short-term memory network and a reverse long short-term memory network. The forward long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the past time point, and the reverse long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the future time point.

[0110] Understandably, in a standard LSTM network, information can only be transferred from the past to the future along the timeline, i.e., forward sequence. This structure has achieved good results in many tasks. However, some time series tasks rely not only on past information but also on future information. To this end, a bidirectional LSTM structure is introduced. It consists of two independent LSTMs: a forward LSTM, which extracts time-dependent features from the current time point forward; and a backward LSTM, which extracts time-dependent features from the current time point forward. Ultimately, the outputs of these two directions can be concatenated, summed, or averaged to predict the data required to construct the parameters of a single EV model.

[0111] A bidirectional long short-term memory network is established to predict single EV data. The forward LSTM and reverse LSTM models are as follows:

[0112]

[0113] in, represents the input vector of the current time step; represents the hidden state of the previous time step (forward); Represents the activation value of the forget gate, which is used to control how much old cell state is retained; Represents the activation value of the input gate, which is used to control the impact of the current input on the cell state; Represents the candidate cell state (new information), which is formed after processing by the tanh function; Represents the cell state at the current time step; Represents the activation value of the output gate, which is used to control the extent to which the current cell state is exposed in the hidden state; represents the hidden state at the current time step (network output); The weight matrix corresponding to each gate (f stands for forward); Represents the bias vector corresponding to each gate; Represents the Sigmoid activation function, with an output range between 0 and 1; Represents the hyperbolic tangent activation function, with an output range between -1 and 1; Represents the multiplication or Hadamard product of a matrix and a vector.

[0114] The above formula defines how the forward LSTM passes the current input at each time step and the previous hidden state , dynamically updating the gating structure and state, thereby effectively capturing the long-term and short-term dependencies of the sequence. The bidirectional LSTM further enhances the model's temporal perception, enabling it to simultaneously consider past and future contextual information, which has significant advantages for tasks such as sentiment analysis, speech recognition, and video frame prediction.

[0115] The following describes the electric vehicle cluster parameter prediction device provided by the embodiment of the present application. The electric vehicle cluster parameter prediction device described below and the electric vehicle cluster parameter prediction method described above can be referenced to each other. Figure 2 As shown, the present application provides an electric vehicle cluster parameter prediction device, the device comprising:

[0116] The network pre-training module 201 is used to pre-train a pre-built bidirectional long short-term memory network based on the acquired commercial charging pile data. The bidirectional long short-term memory network is used to combine historical charging trajectories and future charging trends to predict charging behavior parameters of electric vehicles;

[0117] A charging behavior prediction model determination module 202 is configured to fine-tune the pre-trained bidirectional long short-term memory network using a transfer learning algorithm based on the acquired authorized home charging pile data to obtain a charging behavior prediction model;

[0118] The predicted charging data determination module 203 is used to input the acquired historical charging data of each electric vehicle into the charging behavior prediction model to obtain the predicted charging data of each electric vehicle;

[0119] The cluster feasible region determination module 204 is used to construct the individual feasible region of each electric vehicle according to the predicted charging data, and obtain the cluster feasible region of the electric vehicle cluster through Chino polyhedron modeling and Minkowski summation based on each individual feasible region.

[0120] In one embodiment, the charging behavior prediction model determination module 202 includes:

[0121] The charging behavior prediction model determination unit is used to iteratively update the pre-trained bidirectional long short-term memory network based on a preset objective function using authorized home charging pile data until the preset iteration end conditions are met to obtain a charging behavior prediction model. The objective function is used to minimize the maximum mean difference between commercial charging pile data and home charging pile data in the feature mapping space.

[0122] In one embodiment, the objective function is expressed as:

[0123]

[0124] in, represents the model parameters of the pre-trained bidirectional long short-term memory network, Represents the empirical distribution of commercial charging pile data On the data After mapping function After the expectations, Represents the empirical distribution of home charging pile data On the data After mapping function After the expectations, Used to map input data into a high-dimensional feature space.

[0125] In one embodiment, the expression of the single-unit feasible region of each electric vehicle is:

[0126]

[0127] in, represents the maximum discharge power of the nth electric vehicle, represents the power of the nth electric car at time t, represents the maximum charging power of the nth electric vehicle, Indicates the charging start time. Indicates the charging end time. Indicates the initial charge. represents the power of the nth electric car at time t, Indicates the estimated charging capacity. represents the maximum dischargeable capacity of the nth electric vehicle, Indicates the charge and discharge efficiency, Indicates the time interval, Indicates the maximum chargeable capacity of the nth electric vehicle.

[0128] In one embodiment, the cluster feasible region determination module 204 includes:

[0129] A monomer Chino polyhedron model construction unit is used to construct a monomer Chino polyhedron model corresponding to each monomer feasible domain;

[0130] The cluster feasible domain determination unit is used to superimpose the monomer symmetry center and monomer expansion range of each monomer Chino polyhedron model to obtain the aggregate symmetry center and aggregate expansion range, and obtain the cluster feasible domain based on the aggregate symmetry center and aggregate expansion range.

[0131] In one embodiment, the single Chino polyhedron model building unit includes:

[0132] Convex polyhedron modeling subunit, used to model each monomer feasible region as a convex polyhedron and construct the corresponding generating vector;

[0133] An extension distance determination subunit, configured to generate a corresponding normal vector matrix according to each plane of each convex polyhedron, and calculate the extension distance of each convex polyhedron on each normal vector;

[0134] The monomer Chino polyhedron model construction subunit is used to find the largest inscribed Chino polyhedron in each convex polyhedron according to the generation vector, normal vector matrix and extension distance of each convex polyhedron to obtain each monomer Chino polyhedron model.

[0135] In one embodiment, the bidirectional long short-term memory network includes a forward long short-term memory network and a reverse long short-term memory network. The forward long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the past time point, and the reverse long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the future time point.

[0136] In one embodiment, the present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the electric vehicle cluster parameter prediction method as described in any of the above embodiments.

[0137] In one embodiment, the present application also provides a computer device having computer-readable instructions stored therein. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the electric vehicle cluster parameter prediction method as described in any one of the above embodiments.

[0138] Schematically, as Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 can be provided as a server. Figure 3 Computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by memory 301 for storing instructions executable by processing component 302, such as an application. The application stored in memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, processing component 302 is configured to execute the instructions to perform the electric vehicle cluster parameter prediction method according to any of the above-described embodiments.

[0139] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or the like.

[0140] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0141] Finally, it should be noted that, in this article, relational terms such as first and second are merely used to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. Without further restriction, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element. Herein, "one," "said," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. A plurality refers to at least two, such as 2, 3, 5, or 8. "And / or" includes any and all combinations of the relevant listed items.

[0142] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referenced to each other.

[0143] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting electric vehicle cluster parameters, characterized in that: The method comprises: Pre-training a pre-built bidirectional long-short-term memory network based on acquired commercial charging pile data. The bidirectional long-short-term memory network is used to combine historical charging trajectories and future charging trends to predict charging behavior parameters of electric vehicles. Based on the obtained authorized home charging station data, a transfer learning algorithm is used to fine-tune the pre-trained bidirectional long short-term memory network to obtain a charging behavior prediction model; Inputting the acquired historical charging data of each electric vehicle into the charging behavior prediction model to obtain predicted charging data for each electric vehicle; According to each of the predicted charging data, a single-unit feasible region of each of the electric vehicles is constructed, and according to each of the single-unit feasible regions, a cluster feasible region of the electric vehicle cluster is obtained through Chino polyhedron modeling and Minkowski summation.

2. The electric vehicle cluster parameter prediction method according to claim 1, characterized in that: The step of fine-tuning the pre-trained bidirectional long short-term memory network using a transfer learning algorithm based on the obtained authorized home charging pile data to obtain a charging behavior prediction model includes: Based on a preset objective function, the pre-trained bidirectional long short-term memory network is iteratively updated using the authorized home charging pile data until the preset iteration end condition is met to obtain the charging behavior prediction model. The objective function is used to minimize the maximum mean difference between the commercial charging pile data and the home charging pile data in the feature mapping space.

3. The electric vehicle cluster parameter prediction method according to claim 2, characterized in that: The expression of the objective function is: in, represents the model parameters of the pre-trained bidirectional long short-term memory network, Represents the empirical distribution of the commercial charging pile data On the data After mapping function After the expectations, Represents the empirical distribution of the home charging pile data On the data After mapping function After the expectations, Used to map input data into a high-dimensional feature space.

4. The electric vehicle cluster parameter prediction method according to claim 1, characterized in that: The expression of the single feasible region of each electric vehicle is: in, represents the maximum discharge power of the nth electric vehicle, represents the power of the nth electric car at time t, represents the maximum charging power of the nth electric vehicle, Indicates the charging start time. Indicates the charging end time. Indicates the initial charge. represents the power of the nth electric car at time t, Indicates the estimated charging capacity. represents the maximum dischargeable capacity of the nth electric vehicle, Indicates the charge and discharge efficiency, Indicates the time interval, Indicates the maximum chargeable capacity of the nth electric vehicle.

5. The electric vehicle cluster parameter prediction method according to claim 1, characterized in that: The step of obtaining the cluster feasible region of the electric vehicle cluster by using Chino polyhedron modeling and Minkowski summation based on each of the monomer feasible regions includes: Constructing a monomer Chino polyhedron model corresponding to each monomer feasible region; The monomer symmetry center and the monomer expansion range of each monomer Chino polyhedron model are superimposed to obtain the aggregate symmetry center and the aggregate expansion range, and the cluster feasible domain is obtained according to the aggregate symmetry center and the aggregate expansion range.

6. The electric vehicle cluster parameter prediction method according to claim 5, characterized in that: The step of constructing a monomer Chino polyhedron model corresponding to each monomer feasible region includes: Modeling each of the monomer feasible regions as a convex polyhedron and constructing a corresponding generating vector; Generate a corresponding normal vector matrix according to each plane of each convex polyhedron, and calculate the extension distance of each convex polyhedron on each normal vector; According to the generating vector, normal vector matrix and extension distance of each convex polyhedron, the largest inscribed Chino polyhedron is found in each convex polyhedron to obtain each monomer Chino polyhedron model.

7. The electric vehicle cluster parameter prediction method according to any one of claims 1 to 6, characterized in that: The bidirectional long short-term memory network includes a forward long short-term memory network and a reverse long short-term memory network. The forward long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the past time point, and the reverse long short-term memory network is used to capture the time-dependent features of the input data from the current time point to the future time point.

8. An electric vehicle cluster parameter prediction device, characterized in that: The device comprises: A network pre-training module is used to pre-train a pre-built bidirectional long short-term memory network based on acquired commercial charging pile data. The bidirectional long short-term memory network is used to combine historical charging trajectories and future charging trends to predict charging behavior parameters of electric vehicles; The charging behavior prediction model determination module is used to fine-tune the pre-trained bidirectional long short-term memory network based on the obtained authorized home charging pile data using a transfer learning algorithm to obtain a charging behavior prediction model; A predicted charging data determination module, configured to input the acquired historical charging data of each electric vehicle into the charging behavior prediction model to obtain predicted charging data for each electric vehicle; The cluster feasible region determination module is used to construct a single feasible region of each electric vehicle according to each predicted charging data, and obtain the cluster feasible region of the electric vehicle cluster through Chino polyhedron modeling and Minkowski summation based on each single feasible region.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions, which, when executed by one or more processors, enable the one or more processors to perform the steps of the electric vehicle cluster parameter prediction method according to any one of claims 1 to 7.

10. A computer device, characterized in that: include: one or more processors, and memory; The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the electric vehicle cluster parameter prediction method according to any one of claims 1 to 7 are performed.

Images