Blade electric vehicle friendly charging pile management system and method

Abstract

The invention discloses a friendly charging pile management system and method for a pure electric vehicle, and relates to the technical field of charging piles. Comprising a sensing layer, a cloud processing layer and a user interaction layer which realize data interaction through an encrypted communication protocol. The vehicle type identification module receives three frames of images of entering vehicles collected by a wide-angle camera, after Gamma gray correction, 3 * 3 Gaussian filtering and zooming to 224 * 224 and normalization preprocessing, vehicle types are classified, pure electric vehicles enter a priority queue of the double-queue scheduling module, and mixed extended-range electric vehicles enter a lag queue; the double-queue module builds queues according to Redis Zset and ranks the queues every 5 seconds according to a priority rule, and when the charging piles are idle, scheduling with the highest priority is selected; an identity verification algorithm of the parking lock control module is used for unlocking through license plate editing distance comparison and 120-second time sequence verification, queue jumping is avoided, and the problems that pure electric charging resources are occupied and queuing is disordered are solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of charging piles, in particular to a pure-electric-friendly charging pile management system and method based on vehicle type identification and double-queue scheduling. BACKGROUND

[0002] With the rapid increase in the number of new energy vehicles, the demand for direct-current charging piles, as core facilities for high-power energy supplement, is increasingly urgent. The problems exposed in the operation of charging piles during holiday peak periods are becoming increasingly prominent.

[0003] 1. Lack of differentiation in vehicle type scheduling: plug-in hybrid / electric range extended vehicles choose to occupy charging piles for charging to save costs, and share charging resources with pure electric vehicles. The former long-term occupies high-power direct-current piles with low energy supplement demand, resulting in a significant extension of the waiting time of pure electric vehicle owners and an increase in range anxiety.

[0004] 2. Inaccurate waiting time estimation: the time required for charging is difficult to predict, and users cannot effectively plan their travel and charging arrangements.

[0005] 3. Disorderly queue: when there are many people waiting, there is a lack of clear queue order and control mechanism, and the order among vehicle owners is out of control, which can easily lead to disputes due to being cut in line.

[0006] 4. Low utilization of charging resources: most electric vehicles reduce their charging power when the battery SOC value reaches 80%, but users are used to charging to full capacity, and the time spent in this stage accounts for more than 40%. This not only increases the uncertainty of waiting, but also significantly extends the overall waiting time of subsequent users, resulting in low resource utilization.

[0007] Therefore, it is necessary to provide a pure-electric-friendly charging pile management system and method based on vehicle type identification and double-queue scheduling. SUMMARY

[0008] The technical problem solved by the present application is to provide a pure-electric-friendly charging pile management system and method, which has the advantages of accurate vehicle type identification, dynamic double-queue scheduling, and accurate waiting time prediction, and solves the problems of current pure electric vehicle charging resource being occupied by plug-in hybrid / electric range extended vehicles, inaccurate waiting time estimation, disorderly queue, and low charging efficiency.

[0009] To solve the above technical problems, one technical solution adopted by the present application is:

[0010] A pure-electric-friendly charging pile management system, comprising a perception layer, a cloud processing layer, and a user interaction layer, wherein the perception layer, the cloud processing layer, and the user interaction layer each comprise a data receiving module, and the cloud processing layer realizes data interaction with the perception layer and the user interaction layer through an encrypted communication protocol.

[0011] The perception layer includes a wide-angle camera, a charging pile integration unit, and a vehicle parking lock. The wide-angle camera is used to capture images of entering vehicles. The charging pile integration unit includes a pile-side camera, a charging status sensor, and a metering module. The pile-side camera is used to identify the license plate of the vehicle in the parking space. The charging status sensor is used to collect battery SOC value, charging power, charging voltage, and charging current, with a collection frequency of ≥1 time / 3 seconds. The metering module includes a power metering mode and a duration metering mode. The vehicle parking lock receives and executes unlock or lock commands sent by the cloud processing layer. At this time, the vehicle parking lock supports remote cloud control and local triggering.

[0012] The cloud processing layer includes a vehicle model recognition module, a dual-queue scheduling module, a waiting time prediction module, a parking lock control module, and a billing module. The vehicle model recognition module uses a deep learning model to classify vehicles into two categories: pure electric vehicles and plug-in hybrid / range-extended electric vehicles. The dual-queue scheduling module constructs a priority queue and a lag queue based on vehicle type and dynamically schedules the charging order according to a priority function. The waiting time prediction module outputs the estimated waiting time for a single vehicle based on a deep learning model. The parking lock control module controls the vehicle's parking lock switch in conjunction with the license plate recognition results. The billing module adopts a SOC segmented billing strategy.

[0013] The user interaction layer is used for user scanning to enter, queue query, waiting time display, charging status push and payment settlement.

[0014] Furthermore, the vehicle model recognition module adopts a CNN-ResNet18 model, including an input layer, five residual blocks, a global average pooling layer, a fully connected layer, and a Softmax output layer. The input layer receives an image input of a preset size, which is 224×224 in this scheme. Each residual block includes two convolutional layers, a batch normalization layer, and a ReLU activation function. The fully connected layer has an output dimension of 2, corresponding to two vehicle models.

[0015] The vehicle model recognition module performs fusion preprocessing on the vehicle image, including pixel mean fusion, Gamma correction, filtering, and normalization.

[0016] Furthermore, the dual-queue scheduling module uses the Redis Zset data structure to construct a priority queue and a lag queue. The queue elements include the scheduling ID, license plate number, remaining battery power, and enqueue timestamp.

[0017] The priority function of the dual-queue scheduling module is:

[0018]

[0019] in, The weights are set for queue types: priority queues have a weight of 1, and lagging queues have a weight of 0.5. The standardized value of the enqueue timestamp is calculated using the following formula:

[0020]

[0021] and These are the minimum and maximum timestamps in the current queue, respectively.

[0022] The SOC value is a standardized value, and its standardized calculation formula is as follows:

[0023]

[0024] If the vehicle is in the lagging queue and its SOC value is ≤20%, the weight of the vehicle will be temporarily adjusted to 1, and the effective duration of this adjustment is 60 minutes.

[0025] If a vehicle is in the lagging queue for 3 consecutive times and charges for ≥60 minutes, its weight will be reduced by 0.3.

[0026] Furthermore, the waiting time prediction module adopts the ConvFormer model, which includes an input layer, a feature extraction layer, and an output layer; the input layer receives a 136-dimensional feature vector; the feature extraction layer includes a parallel structure of 6 CNN-Blocks and 6 Transformer Encoders; the output layer includes a one-dimensional convolutional layer, a ReLU activation function, and a fully connected layer, and the fully connected layer outputs the predicted waiting time value.

[0027] Furthermore, the waiting time prediction module operates as follows: First, the input features are Z-score standardized; then, the CNN-Block module of the ConvFormer model extracts local time-series features, while the TransformerEncoder captures the global dependency between the number of available charging piles and the queue length; next, the concatenated features are processed by a one-dimensional convolutional layer with a kernel size of 2, a ReLU activation function, and a one-dimensional max pooling layer with a pooling size of 2; the processed features are input to a fully connected layer to output the predicted value; finally, the result is pushed to the user terminal via the WebSocket protocol, and the prediction is re-performed every 30 seconds based on the queue update result.

[0028] Furthermore, the SOC segmented billing strategy of the billing module is as follows:

[0029] When the battery SOC value is between 0% and 80%, the power consumption billing mode is used, and the calculation formula is as follows:

[0030]

[0031] in, For the first Price per unit time period For the first The calibrated charge amount within the time period;

[0032] When the battery SOC value exceeds 80%, the time-based billing mode is adopted, and the rate is set to 1.5 times that of the power consumption billing mode;

[0033] If the battery's SOC exceeds 80% and the charging power drops to below 30% of the rated power for a duration of ≥10 minutes, the billing module will trigger a rate upgrade, adjusting the rate to twice that of the current power-based billing mode, and will push a reminder message through the user interaction layer. The reminder message includes the power change, the reason for the rate adjustment, and suggestions.

[0034] Furthermore, the billing module also includes a metering error calibration unit, which adopts a dual-dimensional calibration mechanism of power calibration and duration calibration.

[0035] The formula for power calibration is:

[0036]

[0037] in, This is the measurement value from a standard electricity meter. The value of the electrical energy output by the charging pile under test. This is a systematic error;

[0038] The duration calibration-related power curve uses a correction factor when the charging power is less than 30% of the rated power. The calculation formula is as follows:

[0039]

[0040] in, The remaining battery power is obtained in real time by the charging status sensor.

[0041] Furthermore, the ground lock control module employs a dual mechanism of license plate comparison and time-series verification:

[0042] The license plate matching uses an edit distance algorithm, allowing for 0 to 1 character mismatches;

[0043] The timing verification requires that the vehicle arrives at the corresponding parking space within 120 seconds after entering the queue. This is achieved by comparing the queuing timestamp with the parking space arrival timestamp, where the parking space arrival timestamp is the time when the pile-side camera first recognizes the vehicle.

[0044] The ground lock control module blacklists vehicles that fail verification three times in a row, restricting their entry for 15 days.

[0045] A method for managing electric vehicle-friendly charging piles, applied to the aforementioned electric vehicle-friendly charging pile management system, includes the following steps:

[0046] S1. Capture panoramic images of entering vehicles using a wide-angle camera and upload them to the cloud processing layer via an encrypted communication protocol;

[0047] S2. After fusing and preprocessing the panoramic images of the entering vehicles, the vehicle type recognition module classifies the vehicles, assigns pure electric vehicles to the priority queue, assigns plug-in hybrid / range-extended electric vehicles to the lagging queue, and generates a scheduling ID associated with the license plate.

[0048] S3. The waiting time prediction module is based on a deep learning model. It takes the number of idle charging piles, queue length, initial vehicle SOC value and historical charging power curve as input, and outputs the estimated waiting time and pushes it to the user through the user interaction layer.

[0049] S4. The dual-queue scheduling module sorts the dual queues according to the priority function. When the charging pile is idle, the vehicle is scheduled according to the priority. After verifying the identity through license plate recognition, the ground lock is unlocked.

[0050] S5. The charging status sensor collects data in real time. When the battery SOC value exceeds 80%, the billing module switches to time-based billing mode and pushes a power outage reminder to the user through the user interaction layer.

[0051] S6. After charging is complete, the metering module generates a bill, the ground lock is locked, the queue is updated, and the waiting time is re-predicted.

[0052] Furthermore, step S2 includes:

[0053] S21: Fuse the vehicle panoramic image and perform preprocessing, including Gamma correction, Gaussian filtering, size scaling to preset size, and normalization.

[0054] S22: Extract vehicle features using a deep learning model. These features include vehicle outline, wheelbase, and charging interface. Output classification results.

[0055] S23: If the maximum probability is ≥95%, then directly join the queue; if the maximum probability is <95%, then trigger manual review.

[0056] The beneficial technical effects of this invention are:

[0057] 1. This invention uses a CNN-ResNet18 network in the vehicle model recognition module to receive three frames of images of vehicles entering the parking area captured by a wide-angle camera. After Gamma grayscale correction, 3×3 Gaussian filtering, scaling to 224×224, and normalization preprocessing, the vehicle models are classified. Pure electric vehicles are entered into the priority queue of the dual-queue scheduling module, while plug-in hybrid electric vehicles are entered into the lagging queue. The dual-queue scheduling module uses Redis Zset to construct queues, sorts them every 5 seconds according to priority rules, and selects the highest priority element for scheduling when the charging pile is idle. The identity verification algorithm of the ground lock control module allows 0 to 1 character mismatches through license plate editing distance comparison and 120-second time sequence verification to control the ground lock unlocking, avoiding queue jumping and effectively solving the problems of pure electric vehicle charging resources being occupied and chaotic queuing order among waiting crowds.

[0058] 2. This invention receives 136-dimensional features, including the number of idle charging piles, queue length, initial vehicle SOC value, and power parameters of the last three charging cycles, through the ConvFormer model of the waiting time prediction module. After preprocessing, local temporal features are extracted by 6 CNN-Blocks and global dependencies are captured by 6 Transformer Encoders. The input features are updated every 3 seconds and the predicted waiting time is pushed to the user interaction layer every 30 seconds, thus solving the problem of inaccurate prediction of charging time.

[0059] 3. This invention uses a segmented billing strategy based on SOC in the billing module. When the SOC value is between 0% and 80%, the billing mode is based on the amount of electricity used. When the SOC value rises to 80%, it automatically switches to a time-based billing mode with a 1.5x rate. When the charging power is 30% lower than the rated power for 10 minutes, the rate is upgraded to 2x, prompting users to end charging in time and solving the problem of low utilization efficiency of charging resources. Attached Figure Description

[0060] Fig. 1 This is a diagram showing the overall architecture of the charging pile management system of the present invention.

[0061] Fig. 2 This is a flowchart of the vehicle identification module of the charging pile management system of the present invention;

[0062] Fig. 3 This is a logic diagram of the dual-queue scheduling module of the charging pile management system of the present invention;

[0063] Fig. 4 This is a timeline diagram of the entire charging process of the charging pile management system of the present invention. Detailed Implementation

[0064] In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0065] This specific embodiment discloses in detail a pure electric vehicle-friendly charging pile management system and method, such as... Figs. 1 to 4 As shown, it includes a perception layer, a cloud processing layer, and a user interaction layer. The cloud processing layer achieves secure data transmission with the perception layer and the user interaction layer through an encrypted communication protocol based on TLS 1.3, and the AES-256-GCM algorithm is used to encrypt the data during the transmission process to ensure data integrity and confidentiality.

[0066] The hardware configuration and data acquisition principle of the perception layer are as follows: The wide-angle camera uses a CMOS industrial camera with a resolution of 1920×1080, a horizontal shooting angle of 120°, and an image acquisition frame rate of 25fps configured by the camera control chip. Its optical lens adopts a large aperture design of F1.8 to adapt to low light environment. The RAW format data output by the image sensor is processed into JPEG format by the internal ISP module and then transmitted to the local edge gateway through the RJ45 Ethernet interface. In the charging pile integrated unit, the pile-side camera adopts a 1280×720 resolution high-definition camera with an 8mm lens focal length to ensure clear imaging of license plates within a 3-5m range directly in front of the parking space. The built-in license plate recognition algorithm is implemented through FPGA hardware acceleration. The charging status sensor adopts a multi-channel analog quantity acquisition module. The SOC value is obtained by communicating with the vehicle BMS system via the CAN bus, with a sampling frequency of 1 time / 3 seconds. The charging power, voltage, and current are acquired by high-precision Hall sensors with sensor ranges of 0-500kW, 0-1000V, and 0-800A, respectively, and an acquisition accuracy of ±0.5%FS. The acquired data is converted into digital signals by a 16-bit ADC and output as digital signals. The metering module adopts a dual-core architecture. Power metering is calculated based on the sampling data of the shunt and voltage transformer through a digital integration algorithm. Duration metering is achieved through a high-precision RTC clock module with a timing accuracy of ±1s / day. After calibration, the metering error is controlled within ±1.8%. The vehicle ground lock uses an electromagnetic actuator, and the control module uses an STM32L4 series MCU with a response time of 450ms. It supports RS485 interface cloud communication and local triggering unlocking by infrared sensors.

[0067] The cloud processing layer is deployed on a cloud server with a hardware configuration of Intel Xeon Platinum 8350C 16 vCPUs, 56GB of RAM, and an RTX 3090 GPU with 24GB of video memory. The operating system is Ubuntu 20.04.4LTS. The cloud processing layer builds a message queue based on Kafka 2.8.1, creating a topic named `perception_data` with 4 partitions and 2 replicas. The edge gateway in the perception layer acts as the producer, encapsulating the collected data in JSON format and sending it to this topic. The encapsulated content includes the device ID, data type, collection timestamp, data value, and checksum. The Kafka consumer group is configured as `cloud_consumer`, receiving data through a batch pull mechanism, with a transmission latency consistently below 85ms. The vehicle model recognition, dual-queue scheduling, waiting time prediction, ground lock control, and billing modules are all deployed using a microservice architecture, managed via Docker containers. Modules communicate via the gRPC protocol, and service registration and discovery are implemented using Consul.

[0068] The user interaction layer is developed using the Vue3 framework for the APP / Mini Program frontend, while the backend interface is implemented using the ThinkPHP 6.0 framework. UI response times are controlled within 180ms through code splitting, lazy loading, and other resource loading strategies. When a user scans a QR code to enter, the frontend uses the phone's camera to scan the QR code, parses the entry identifier, sends it to the backend for verification, and then returns it to the queue for access permissions. The entire process is transmitted via HTTPS, and the data exchange format is JSON.

[0069] The following describes the architecture design, training process, deployment optimization, and operation mechanism of the CNN-ResNet18 model to achieve accurate classification of pure electric vehicles and plug-in hybrid / range-extended electric vehicles.

[0070] The CNN-ResNet18 model architecture employs an 18-layer deep neural network structure, including an input layer, five residual blocks, a global average pooling layer, a fully connected layer, and a softmax output layer. The input layer receives a pre-processed RGB image with dimensions of 224×224×3. The first residual block contains two Conv2d layers: the first layer uses a 7×7 convolutional kernel with a stride of 2, padding of 3, and 64 output channels; the second layer uses a 3×3 convolutional kernel with a stride of 1, padding of 1, and 64 output channels. The Conv2d layers in the remaining four residual blocks all use 3×3 convolutional kernels with a stride of 1, padding of 1, and output channels of 64, 128, 256, and 512 respectively. Each residual block has short-circuit connections; when the number of input and output channels is inconsistent, a 1×1 convolutional layer is used for dimension adjustment.

[0071] The global average pooling layer converts the 7×7×512 feature map output by the last residual block into a 512-dimensional feature vector. The fully connected layer maps the output to 2-dimensional corresponding to the two types of vehicle models, and the Softmax layer converts it into a probability distribution.

[0072] The CNN-ResNet18 model training dataset consists of 100,000 images of 100 different bus models from a pilot bus stop, including 60,000 images of pure electric vehicles and 40,000 images of plug-in hybrid / range-extended electric vehicles, covering various lighting conditions, weather conditions, and shooting angles. The image preprocessing workflow includes: Gamma grayscale correction and parameter... Set to 1.2; 3×3 Gaussian filter, filter kernel is 1 / 16 multiplied by matrix [[1,2,1],[2,4,2],[1,2,1]]; bilinear interpolation to scale to 224×224; normalization processing, calculation formula is... .

[0073] The model training environment is based on the PyTorch 1.11.0 framework, using the cross-entropy loss function, calculated as follows: The optimizer used was SGD, with an initial learning rate of 5e-4, momentum of 0.9, weight decay of 1e-4 (decreasing by 1 / 10 every 50 epochs), a batch size of 64, and a total of 200 iterations. The dataset was divided into training and validation sets in a 9:1 ratio. An early stopping strategy was adopted: if the accuracy on the validation set did not improve for 10 consecutive epochs, training was stopped. The final validation accuracy was 98.5%.

[0074] The CNN-ResNet18 model deployment utilizes the ONNX format optimization, converted via PyTorch's `torch.onnx.export` interface, setting the input tensor shape to (1,3,224,224) and specifying dynamic axes. The deployment environment is an Intel Xeon Platinum 8350C 16 vCPU cloud node, using TensorRT 8.2 as the inference engine. Quantization-aware training was used to convert the model's accuracy from FP32 to FP16, resulting in a more than 3x improvement in inference speed. An HTTP interface based on FastAPI is provided, with an average response time consistently below 90 milliseconds.

[0075] When the CNN-ResNet18 model is running, the system first performs pixel mean fusion on three consecutive frames of images captured by a wide-angle camera. The calculation formula is as follows: The fused image is preprocessed and input into the model. The CNN-ResNet18 model extracts key features such as vehicle outline, wheelbase, and charging interface sequentially through residual blocks. After linear transformation by a fully connected layer, the class probability is output by a Softmax layer. If the maximum probability is ≥0.95, the vehicle model is directly determined; if the maximum probability is <0.95, it is pushed to manual review, and the results of the manual review will be used for incremental learning of the model.

[0076] The following section describes the queue construction, priority calculation, training optimization, and operation mechanism of the dual-queue scheduling algorithm to achieve dynamic allocation of charging resources.

[0077] The dual-queue scheduling algorithm is constructed using a Zset data structure based on the Redis 6.2.6 database, creating two queues: a priority queue (key: priority_queue) and a delayed queue (key: delayed_queue). Each queue element is a four-tuple, including the scheduling ID, license plate number, remaining battery power, and enqueue timestamp. The scheduling ID is a globally unique identifier generated from a UUID, and the enqueue timestamp is accurate to milliseconds.

[0078] The priority function is designed as follows ;

[0079] in, The weights are set for queue types: priority queue weight is set to 1, and lagging queue weight is set to 0.5. The standardized value of the enqueue timestamp is calculated using the following formula: , and These are the minimum and maximum timestamps in the current queue, respectively. The SOC value is a standardized value, and its standardized calculation formula is as follows:

[0080] The dual-queue scheduling algorithm was trained using 12 months of 5×10... 4 Each dispatch record contains data covering vehicle model, queuing time, SOC value, queue type, waiting time, and charging time.

[0081] Training is performed using a reinforcement learning framework. The agent acts as the scheduling module, with its state space including queue length, number of available charging stations, SOC value of each vehicle, and enqueue time; the action space contains scheduling instructions; and the reward function is... ,in To improve queue turnover efficiency, This represents the average waiting time for pure electric vehicles.

[0082] The training process iterated for 1000 rounds, with each round consisting of 1000 episodes. The action value function was updated using the Q-learning algorithm, resulting in a 42% reduction in waiting time for pure electric vehicles and a 35% improvement in queue turnover efficiency.

[0083] The dual-queue scheduling algorithm is deployed on a cloud-based Java service node, developed using the Spring Boot framework, with the service port being 8081. It communicates with the ground lock control module via the gRPC protocol, and data serialization uses the Protocol Buffers format.

[0084] The frequency of scheduling instructions is set to once every 5 seconds via a timed task thread pool. Before generating instructions, the charging pile status and queue data are obtained, and the priorities are calculated and sorted.

[0085] When the dual-queue scheduling algorithm is running, vehicles are inserted into the corresponding queues based on the vehicle type identification results: pure electric vehicles enter the priority queue, and plug-in hybrid / range-extended electric vehicles enter the lagging queue, which is implemented by the ZADD command in Redis with a score of F; every 5 seconds, the queues are sorted in descending order of F value by the ZREVRANGE command.

[0086] When a charging station is idle, the system obtains the ID of the idle charging station, calls the ZPOPMAX command to extract the element with the largest F value, generates a scheduling instruction, and sends it to the ground lock control module.

[0087] If a vehicle in the lagging queue has a SOC ≤ 20%, its weight is temporarily adjusted to 1 using the ZINCRBY command, with an effective duration of 60 minutes. Vehicles in the lagging queue that have been charged for ≥ 60 minutes three times consecutively have their weight reduced by 0.3. After adjustment, the F value is recalculated and the ranking is updated.

[0088] The following section discusses the architecture design, training process, deployment optimization, and operation mechanism of the ConvFormer model to achieve accurate prediction of the expected waiting time for a single vehicle.

[0089] The ConvFormer model architecture consists of an input layer, a feature extraction layer, and an output layer.

[0090] The input layer receives a 136-dimensional feature vector, including 20 dimensions of charging pile-related features, such as the number of idle charging piles, the total number of charging piles, and the average occupancy rate in the past hour; 15 dimensions of queue-related features, such as the length of the priority queue, the length of the lagging queue, and the queuing rate in the past 5 minutes; 12 dimensions of vehicle-related features, such as the initial SOC value, the average power of the last 3 charging cycles, the standard deviation of the charging power, and the battery capacity; and 89 dimensions of historical statistical features, such as the average waiting time in the past hour, the average charging time in different SOC intervals, and the average service time of the charging pile. All features are concatenated after preprocessing.

[0091] The feature extraction layer employs a parallel structure of 6 CNN-Block units and 6 Transformer Encoder layers. Each CNN-Block unit includes a one-dimensional convolutional layer with 3 kernels, a stride of 1, padding of 1, and 64 output channels. It uses the GELU activation function, calculated as follows: The algorithm incorporates a max-pooling layer with a pooling window of 2 and a stride of 2. The local feature vector output from the concatenated 6 CNN-Block layers has a dimension of 64×8. Each layer of the Transformer Encoder includes 8 attention heads, each with a dimension of 64; the feedforward network has a dimension of 512 and uses the ReLU activation function, with residual connections and layer normalization. The global feature vector output from the 6-layer Encoder has a dimension of 136×64.

[0092] The output layer first merges local and global features into a 100×64 vector using a Concat operation. This vector then passes through a Conv1d layer with a 2-kernel, 1-stride, and ReLU activation. After passing through a MaxPool1d layer with pooling twice, the vector is input to a fully connected layer. This fully connected layer has 64 hidden units, and output layer 1 outputs the predicted value.

[0093] The ConvFormer model was trained using 1.98×10⁻⁶ data points from October 2022 to October 2023. 7 Each charging data point is collected every 3 seconds. Preprocessing uses forward padding to fill in missing state of charge (SOC) data. The criteria for removing power outliers are as follows: ,in The power characteristic mean, The standard deviation of power characteristics; Z-Score standardization, calculated using the following formula: ,in The characteristic mean, The standard deviation is the feature value. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0094] The training environment is based on the PyTorch 1.11.0 framework and uses the MSE loss function, the calculation formula of which is as follows: The optimizer used is Adam, with a learning rate of 1e-5. , The batch size is 32, and the iterations are 50 rounds. After determining the 6+6 structure, the MAE on the test set is 0.01002, and the MSE is 0.00033.

[0095] The ConvFormer model is deployed on a cloud server with an RTX 3090 GPU, 24GB of VRAM, and 56GB of RAM. It is stored in pth format with a file size of 80MB and a loading time of 1.8 seconds. GPU-accelerated inference is used, and the inference time is ≤50ms when batch processing 32 data points.

[0096] During ConvFormer model execution, 136-dimensional features are acquired every 3 seconds. After Z-score normalization, these features are input into CNN-Block and Transformer Encoder. CNN-Block extracts local features such as charging power changes and SOC fluctuations, while Transformer Encoder captures global dependencies. The concatenated features are then processed by the output layer to obtain the waiting time.

[0097] The system calculates the prediction bias every 7 days. If the MAE increases by more than 5%, the model is fine-tuned using 7 days of newly added data, with a data volume of 6.72 × 10⁻⁶. 5 The algorithm iterates for 10 rounds with a learning rate of 1e-6.

[0098] The following section describes the two-dimensional calibration mechanism, training process, deployment method, and operating principle of the metrology error calibration algorithm to achieve accurate calibration of metrology data.

[0099] The measurement error calibration algorithm is constructed using a dual-dimensional calibration machine that considers both power consumption and duration.

[0100] The power calibration is calculated based on the power difference between the standard energy meter and the charging pile under test, using the following formula: ;in, This is the electrical energy value measured by a high-precision remote total meter, with an accuracy of ±0.1%. The electrical energy value output by the metering module of the charging pile under test; The systematic error of the standard verification device is 0.02% when it is determined by previous calibration tests that no correction is required.

[0101] The duration calibration is linked to the charging power curve. When the charging power is less than 30% of the rated power, a correction factor is used, and its calculation formula is as follows: Where SOC is the remaining battery power obtained in real time by the charging status sensor.

[0102] The training of the metrological error calibration algorithm uses 1×10 provided by the metrology institution. 4 The error data covers different load conditions including 20%, 50%, and 80% of rated power, as well as various ambient temperatures such as -20℃, 0℃, 25℃, and 50℃. and Value. Trained using a linear regression model. Parameters, to minimize To find the objective, the least squares method is used.

[0103] The metering error calibration algorithm is deployed in the billing module's PHP service, developed using the ThinkPHP framework, and encapsulated as a CalibrationService class. It provides calibrateElectricity (electricity calibration) and calibrateDuration (time-based calibration) methods. It is linked to the metering module via a database trigger: calibration is triggered when electricity data is inserted into the metering module; and it communicates with the waiting time prediction module via a RESTful API to obtain 24-hour error prediction results.

[0104] When the measurement error calibration algorithm is running, the measurement module collects data every 3 seconds. Value and duration data are obtained synchronously. The value is used to calculate the power error correction value using the `calibrateElectricity` method. The corrected power calculation formula is as follows: .

[0105] The charging status sensor collects power and SOC values ​​in real time. When the power is less than 30% of the rated power, the calibrateDuration method is used to calculate k. The corrected duration calculation formula is as follows: ;

[0106] Trigger duration calibration when SOC > 80% and power < 30% of rated power;

[0107] After charging is complete, all calibration data is stored in the calibrated_meter table of the MySQL database and synchronized to the user interaction layer to generate an invoice.

[0108] The following section describes the dual mechanism, training process, deployment method, and operating principle of the identity verification algorithm to achieve accurate vehicle identity verification and ground lock control.

[0109] The identity verification algorithm adopts a dual mechanism of license plate comparison and time-series verification: the license plate comparison uses the edit distance algorithm to calculate the minimum number of conversion operations required for the pile-side camera to recognize the license plate and the dispatch license plate, allowing 0 to 1 character mismatch;

[0110] The formula for calculating edit distance is: : To adjust the length of license plates, To identify the length of the license plate; when hour ,when hour When characters match When characters do not match .

[0111] The timing verification requires that the vehicle arrive at the corresponding parking space within 120 seconds after entering the queue. This is achieved by comparing the entry timestamp with the parking space arrival timestamp, where the parking space arrival timestamp is the time when the pile-side camera first recognizes the vehicle.

[0112] The identity verification algorithm was trained using 3×10 for 6 months. 4 The system records license plate recognition and parking space arrival information, including dispatched license plates, recognized license plates, edit distance, and arrival time. Statistical analysis determined an edit distance threshold of 1, and training the system to achieve a comparison accuracy of ≥99.5%. A 120-second time series threshold was determined based on the average travel time from the queuing point to the parking space (approximately 80 seconds) and parking time (approximately 30 seconds), ensuring that over 99% of normal vehicles arrive within the threshold.

[0113] The identity verification algorithm is deployed in the C++ service of the ground lock control module. It is developed using the Qt framework and communicates with the pile-side camera with the communication parameters set to baud rate 9600, data bits 8, stop bits 1, and no parity. It connects to the ground lock through the GPIO interface. The service initializes the serial port and GPIO. The camera trigger frequency is 1 time / 2 seconds. The ground lock is controlled by high and low level signals, with high level for unlocking and low level for locking.

[0114] During operation, the identity verification algorithm receives the charging pile ID, target license plate, and priority instruction from the scheduling module, triggering continuous recognition by the pile-side camera. It acquires a result every 2 seconds and calculates the distance twice consecutively. If both distances are ≤1 and the arrival time difference is ≤120 seconds, an unlock instruction is generated. This instruction includes the lock_id, action field, and token identifier. The token is generated using MD5 encryption and unlocked by sending a high-level signal via GPIO. If the infrared sensor detects no vehicle 10 seconds after unlocking, a low-level relock signal is sent. License plates that fail verification three times consecutively are blacklisted and restricted from entering the site for 15 days.

[0115] A management method for electric vehicle-friendly charging piles is proposed and applied to an electric vehicle-friendly charging pile management system. The following details the specific implementation steps, technical principles, and module collaboration mechanism of the management method to achieve intelligent management of the entire charging pile process.

[0116] In step S1, the wide-angle camera captures 3 panoramic images of the entering vehicles at a rate of 25 frames per second. Each image is compressed using JPEG and then encapsulated into a JSON data packet by the edge gateway in the format of device ID, acquisition timestamp, image data, and CRC32 checksum. The edge gateway, acting as a Kafka producer, sends the data to the perception_data topic through the librdkafka client and enables a message acknowledgment mechanism, with the acks parameter set to 1 to ensure reliable data transmission and a transmission latency of ≤100ms.

[0117] Step S2: The consumer thread of the vehicle model recognition module pulls image data and first performs pixel mean fusion to eliminate noise; the fused image is then corrected using Gamma parameters. The power is set to 1.2, 3×3 Gaussian filter, 224×224 scaling, and normalized before being input into the CNN-ResNet18 model. When the model output probability is ≥0.95, pure electric vehicles are put into the priority queue, and plug-in hybrid / range-extended electric vehicles are put into the lagging queue. Both generate UUID format scheduling IDs to associate with license plate information. If the probability is <0.95, manual review is triggered, and the result is simultaneously added to the queue.

[0118] Step S3: The waiting time prediction module acquires feature data every 3 seconds, and the preprocessing submodule performs Z-Score standardization. The standardized features are input into the ConvFormer model, CNN-Block extracts local temporal features, and TransformerEncoder captures global dependencies. After feature concatenation, the data is processed by Conv1d, ReLU activation functions, and MaxPool1d, and finally the waiting time is output by the fully connected layer. The data is pushed to the user end via the WebSocket protocol, and a heartbeat mechanism is used to maintain connection stability. The prediction is updated and pushed again every 30 seconds based on the queue.

[0119] Step S4: The dual-queue scheduling module presses every 5 seconds. Calculate priorities and sort them in descending order using the ZREVRANGE command; when the charging pile is idle, call the ZPOPMAX command to obtain the element with the largest F value, generate a scheduling command and send it to the ground lock control module; the ground lock module triggers continuous recognition by the camera, and if the calculated edit distance is ≤1 and the vehicle arrives within 120 seconds, it unlocks; if the verification fails, the next element is taken for scheduling.

[0120] Step S5: The charging status sensor collects the SOC value and power data every 3 seconds and transmits them to the billing module; when the SOC value is in the range of 0%-80%, the power consumption billing mode is adopted, and the calculation formula is as follows: ,in For the first Price per unit time period For the first The calibrated charging amount within the time period; when the SOC value exceeds 80%, it automatically switches to the time-based billing mode, with the rate set at 1.5 times the rate of the power-based billing mode; if the SOC value is in the range of 0-80% and the charging power drops to within 30% of the rated power for a duration of ≥10 minutes, the rate upgrade will not be triggered; if the SOC value exceeds 80% and the charging power drops to within 30% of the rated power for a duration of ≥10 minutes, the billing module will trigger a rate upgrade, adjusting the rate to twice the rate of the power-based billing mode, and pushing a reminder message through the user interaction layer, including the power change, the reason for the rate adjustment, and suggestions.

[0121] Step S6: When the user initiates a charging completion command via the APP / mini-program, or when the SOC reaches 100%, the metering module stops collecting data, calls the metering error calibration algorithm to calibrate the power and duration data, generates a final bill, which includes the calibrated power, duration, and cost details, and stores the bill in the database; after receiving the charging completion signal, the parking lock control module sends a low-level signal to the vehicle parking lock, and the vehicle parking lock performs a locking action; the queue module deletes the four-tuple information of vehicles that have completed charging and updates the queue length; the waiting time prediction module re-executes the prediction based on the updated queue data and pushes the new waiting time to the user interaction layer.

[0122] In summary, the vehicle model recognition module uses a CNN-ResNet18 network to receive three frames of images of vehicles entering the parking area from a wide-angle camera. After Gamma grayscale correction, 3×3 Gaussian filtering, scaling to 224×224, and normalization preprocessing, the vehicle models are classified. Pure electric vehicles are entered into the priority queue of the dual-queue scheduling module, while plug-in hybrid electric vehicles are entered into the lagging queue. The dual-queue scheduling module uses a Redis Zset to construct the queue, sorts it every 5 seconds according to priority rules, and selects the highest priority element for scheduling when the charging pile is idle. The identity verification algorithm of the parking lock control module allows 0-1 character mismatches through license plate editing distance comparison and controls parking lock unlocking through 120-second time sequence verification to prevent queue jumping and effectively solve the problems of pure electric vehicle charging resources being occupied and chaotic queuing order among waiting crowds.

[0123] Furthermore, the ConvFormer model of the waiting time prediction module receives 136-dimensional features including the number of idle charging piles, queue length, initial vehicle SOC value, and power parameters of the last three charging cycles. After preprocessing, local temporal features are extracted by 6 CNN-Blocks and global dependencies are captured by 6 Transformer Encoders. The input features are updated every 3 seconds and the predicted waiting time is pushed to the user interaction layer every 30 seconds, solving the problem of inaccurate prediction of charging time.

[0124] Furthermore, through the SOC segmented billing strategy of the billing module, the billing mode is based on the power consumption when the SOC value is between 0% and 80%; when the SOC value exceeds 80%, it automatically switches to the duration billing mode with a 1.5 times rate; if the SOC value exceeds 80% and the charging power is 30% lower than the rated power for 10 minutes, the rate is upgraded to 2 times, and a reminder is pushed through the user interaction layer to prompt the user to end the charging in time, thus solving the problem of low charging resource utilization efficiency.

[0125] It is worth noting that: the modules in this embodiment are functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system is the interaction or connection relationship between the modules, that is, the improvement of the overall structure of the system in order to solve the corresponding technical problems to be solved by this system.

[0126] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A pure electric vehicle-friendly charging pile management system, characterized in that, It includes a perception layer, a cloud processing layer, and a user interaction layer. Each of the perception layer, the cloud processing layer, and the user interaction layer includes a data receiving module. The cloud processing layer interacts with the perception layer and the user interaction layer through an encrypted communication protocol. The perception layer includes a wide-angle camera, a charging pile integration unit, and a vehicle parking lock; the wide-angle camera is used to capture images of entering vehicles; the charging pile integration unit includes a pile-side camera, a charging status sensor, and a metering module; the pile-side camera is used to identify the license plate of the vehicle in the parking space; the charging status sensor is used to collect battery SOC value, charging power, charging voltage, and charging current; the metering module includes a power metering mode and a duration metering mode; the vehicle parking lock receives and executes unlock or lock commands sent by the cloud processing layer. The cloud processing layer includes a vehicle model recognition module, a dual-queue scheduling module, a waiting time prediction module, a parking lock control module, and a billing module. The vehicle model recognition module uses a deep learning model to classify vehicles into two categories: pure electric vehicles and plug-in hybrid / range-extended electric vehicles. The dual-queue scheduling module constructs a priority queue and a lag queue based on vehicle type and dynamically schedules the charging order according to a priority function. The waiting time prediction module outputs the estimated waiting time for a single vehicle based on a deep learning model. The parking lock control module controls the parking lock switch in conjunction with the license plate recognition results. The billing module adopts a SOC segmented billing strategy. The user interaction layer is used for user scanning to enter, queue query, waiting time display, charging status push and payment settlement.

2. The electric vehicle-friendly charging pile management system according to claim 1, characterized in that: The vehicle model recognition module uses a CNN-ResNet18 model, which includes an input layer, five residual blocks, a global average pooling layer, a fully connected layer, and a Softmax output layer. The input layer receives an image input of a preset size. Each residual block includes two convolutional layers, a batch normalization layer, and a ReLU activation function. The fully connected layer has an output dimension of 2, corresponding to two vehicle models. The vehicle model recognition module performs fusion preprocessing on the vehicle image, including pixel mean fusion, Gamma correction, filtering, and normalization.

3. The electric vehicle-friendly charging pile management system according to claim 1, characterized in that: The dual-queue scheduling module uses the Redis Zset data structure to build a priority queue and a lag queue. The queue elements include scheduling ID, license plate number, remaining battery power, and enqueue timestamp. The priority function of the dual-queue scheduling module is: ； in, The weights are set for queue types: priority queues have a weight of 1, and lagging queues have a weight of 0.

5. The standardized value of the enqueue timestamp is calculated using the following formula: ； and These are the minimum and maximum timestamps in the current queue, respectively. The SOC value is a standardized value, and its standardized calculation formula is as follows: ； If the vehicle is in the lagging queue and its SOC value is ≤20%, the weight of the vehicle will be temporarily adjusted to 1, and the effective duration of this adjustment is 60 minutes. If a vehicle is in the lagging queue for 3 consecutive times and charges for ≥60 minutes, its weight will be reduced by 0.

3.

4. The electric vehicle-friendly charging pile management system according to claim 1, characterized in that: The waiting time prediction module adopts the ConvFormer model, which includes an input layer, a feature extraction layer, and an output layer. The input layer receives a 136-dimensional feature vector. The feature extraction layer includes a parallel structure of 6 CNN-Blocks and 6 Transformer Encoders. The output layer includes a one-dimensional convolutional layer, a ReLU activation function, and a fully connected layer. The fully connected layer outputs the predicted waiting time value.

5. The electric vehicle-friendly charging pile management system according to claim 4, characterized in that, The operation flow of the waiting time prediction module is as follows: First, the input features are Z-Score standardized; then, the CNN-Block module of the ConvFormer model extracts local time series features, while the Transformer Encoder captures the global dependency between the number of available charging piles and the queue length; next, the concatenated features are processed by a one-dimensional convolutional layer with a kernel size of 2, a ReLU activation function, and a one-dimensional max pooling layer with a pooling size of 2; the processed features are input to a fully connected layer to output the predicted value; finally, the result is pushed to the user terminal via the WebSocket protocol, and the prediction is re-performed every 30 seconds based on the queue update result.

6. The electric vehicle-friendly charging pile management system according to claim 1, characterized in that: The SOC segmented billing strategy of the billing module is as follows: When the battery SOC value is between 0% and 80%, the power consumption billing mode is used, and the calculation formula is as follows: ； in, For the first Price per unit time period For the first The calibrated charge amount within the time period; When the battery SOC value exceeds 80%, the time-based billing mode is adopted, and the rate is set to 1.5 times that of the power consumption billing mode; If the battery SOC value exceeds 80% and the charging power drops to below 30% of the rated power for a duration of ≥10 minutes, the billing module will trigger a rate upgrade, adjusting the rate to twice that of the power billing mode, and will push a reminder message through the user interaction layer.

7. The electric vehicle-friendly charging pile management system according to claim 6, characterized in that: The billing module also includes a metering error calibration unit, which adopts a two-dimensional calibration mechanism of power calibration and duration calibration. The formula for power calibration is: ； in, This is the measurement value from a standard electricity meter. The value of the electrical energy output by the tested charging pile. This is a systematic error; The duration calibration-related power curve uses a correction factor when the charging power is less than 30% of the rated power. The calculation formula is as follows: ； in, The remaining battery power is obtained in real time by the charging status sensor.

8. The electric vehicle-friendly charging pile management system according to claim 1, characterized in that: The ground lock control module employs a dual mechanism of license plate comparison and time-series verification. The license plate matching uses an edit distance algorithm, allowing for 0 to 1 character mismatches; The timing verification requires that the vehicle arrives at the corresponding parking space within 120 seconds after entering the queue. This is achieved by comparing the queuing timestamp with the parking space arrival timestamp, where the parking space arrival timestamp is the time when the pile-side camera first recognizes the vehicle. The ground lock control module blacklists vehicles that fail verification three times in a row, restricting their entry for 15 days.

9. The management method for electric vehicle-friendly charging piles according to claim 1, characterized in that, The electric vehicle-friendly charging pile management system according to any one of claims 1-8 includes the following steps: S1. Capture panoramic images of entering vehicles using a wide-angle camera and upload them to the cloud processing layer via an encrypted communication protocol; S2. After fusing and preprocessing the panoramic images of the entering vehicles, the vehicle type recognition module classifies the vehicles, assigns pure electric vehicles to the priority queue, assigns plug-in hybrid / range-extended electric vehicles to the lagging queue, and generates a scheduling ID associated with the license plate. S3. The waiting time prediction module is based on a deep learning model. It takes the number of idle charging piles, queue length, initial vehicle SOC value and historical charging power curve as input, and outputs the estimated waiting time and pushes it to the user through the user interaction layer. S4. The dual-queue scheduling module sorts the dual queues according to the priority function. When the charging pile is idle, the vehicle is scheduled according to the priority. After verifying the identity through license plate recognition, the ground lock is unlocked. S5. The charging status sensor collects data in real time. When the battery SOC value exceeds 80%, the billing module switches to time-based billing mode and pushes a power outage reminder to the user through the user interaction layer. S6. After charging is complete, the metering module generates a bill, the ground lock is locked, the queue is updated, and the waiting time is re-predicted.

10. The electric vehicle-friendly charging pile management method according to claim 9, characterized in that, Step S2 includes: S21: Fuse the vehicle panoramic image and perform preprocessing, including Gamma correction, Gaussian filtering, size scaling to preset size, and normalization. S22: Extract vehicle features using a deep learning model. These features include vehicle outline, wheelbase, and charging interface. Output classification results. S23: If the maximum probability is ≥95%, then directly join the queue; if the maximum probability is <95%, then trigger manual review.

Images