FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing

Abstract

The invention provides an FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing, which comprises a data preprocessing module, a model training module and a recommendation output module, real-time data transmission among the modules is realized through data interfaces, the data preprocessing module is used for preprocessing obtained historical charging order data of a user to obtain preprocessed training data; the model training module is deployed on the GPU server and is used for training an FSE-OPL-DenseNet model according to the preprocessed training data so as to obtain a trained FSE-OPL-DenseNet model; and the recommendation output module is integrated on an API interface of a mobile terminal or a cloud terminal, and is used for receiving a charging request of a user in real time, generating an access probability distribution result of the user to each candidate charging interest site according to a trained FSE-OPL-DenseNet model, and generating a charging interest site recommendation list based on the access probability distribution result. According to the invention, the accuracy and reliability of charging station recommendation are improved.

Description

Technical Field

[0001] This invention relates to the field of vehicle charging technology, specifically to an FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing. Background Technology

[0002] With the rapid development of electric vehicles and the increasing demand for charging, traditional deep learning models (such as DenseNet and ResNet) have performed well in image classification and feature extraction, but they suffer from insufficient ability to recognize subtle features when dealing with charging behavior analysis in complex scenarios. Furthermore, while algorithms based on collaborative filtering and matrix factorization are widely used in industries such as e-commerce in the field of recommender systems, their application in recommending electric vehicle charging facilities is still in the exploratory stage.

[0003] Current related technologies mainly include traditional deep learning models, data-augmented deep learning models, collaborative filtering-based recommendation systems, and heterogeneous information networks (HIN)-based recommendation systems. First, traditional deep learning models (such as DenseNet and ResNet) use convolutional neural networks (CNNs) for efficient feature extraction and classification, but they lack sufficient sensitivity to handle subtle differences in charging behavior. For example, DenseNet uses densely connected layers to improve information flow, but its ability to dynamically adjust key channel features is limited. Second, data augmentation techniques in data-augmented deep learning models (such as DLACNN) are used to generate more diverse training data to alleviate overfitting. However, this method is still insufficient when dealing with the complexity and heterogeneity of charging interest sites. Third, collaborative filtering algorithms in collaborative filtering-based recommendation systems perform similarity calculations based on user rating data to recommend charging stations to users. However, this technique performs poorly on data sparsity issues, especially when user rating behavior is limited. Fourth, recommendation systems based on heterogeneous information networks (HINs) can integrate multiple types of nodes and connections (such as geographical location and social relationships), providing richer semantic information for the recommendation system. However, existing methods have limited ability to capture nonlinear relationships in charging prediction.

[0004] In summary, the existing technology has the following drawbacks: (1) Insufficient ability to identify subtle features: Traditional deep learning models have difficulty capturing subtle morphological differences between charging interest sites, resulting in low classification accuracy.

[0005] (2) Data sparsity problem: When the user rating behavior is limited, the collaborative filtering-based recommendation system cannot effectively calculate the similarity between users.

[0006] (3) Limited ability to handle uncertainty and ambiguity: In the recommendation of charging interest sites, the uncertainty of image quality and user behavior requires more powerful fuzzy logic processing capabilities.

[0007] (4) Lack of modeling of disturbance values: The charging forecast does not consider the impact of event disturbances (such as seasonal changes or sudden events) on the number of visits, resulting in inaccurate forecast results. Summary of the Invention

[0008] In view of this, the present invention provides an FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing, which aims to accurately identify and classify user charging behavior, alleviate the data sparsity problem, improve the ability to identify subtle features of charging behavior, overcome the problem of insufficient sensitivity of traditional deep learning models to key features in complex charging scenarios, thereby optimizing the accurate classification and recommendation of the next charging interest site (the charging pile that the next user may be interested in), and improving the performance and reliability of the recommendation system.

[0009] According to a first aspect of the present invention, an FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing is provided, comprising a data preprocessing module, a model training module, and a recommendation output module. The modules transmit data in real time through a data interface. Specifically: the data preprocessing module, deployed on a cloud server, preprocesses acquired historical user charging order data to obtain preprocessed training data; the model training module, deployed on a GPU server, trains an FSE-OPL-DenseNet model based on the preprocessed training data to obtain a trained FSE-OPL-DenseNet model; and the recommendation output module, integrated into a mobile terminal or cloud API interface, receives user charging requests in real time, generates a probability distribution of user visits to candidate charging interest sites based on the trained FSE-OPL-DenseNet model, and generates a recommended list of charging interest sites based on the probability distribution.

[0010] According to a second aspect of the present invention, an electric vehicle charging recommendation method based on sparse network sharing using FSE-OPL-DenseNet is provided, comprising: preprocessing acquired historical user charging order data through a data preprocessing module deployed on a cloud server to obtain preprocessed training data; training an FSE-OPL-DenseNet model based on the preprocessed training data through a model training module deployed on a GPU server to obtain a trained FSE-OPL-DenseNet model; receiving user charging requests in real time through a recommendation output module integrated into a mobile terminal or cloud API interface, generating a probability distribution result of user access to each candidate charging interest site based on the trained FSE-OPL-DenseNet model, and generating a charging interest site recommendation list based on the access probability distribution result.

[0011] Compared with the prior art, the present invention has the following beneficial effects: (1) Introduce the FSE module The FSE (Feature Selection Enhancement) module of this invention combines fuzzy logic and attention mechanisms to dynamically adjust the weights of feature channels, enhancing the model's ability to focus on key features. Compared to the traditional DenseNet, the FSE module improves the model's ability to identify subtle features of charging interest sites.

[0012] (2) Application of Orthogonal Projection Loss (OPL) The OPL of this invention reduces feature overlap between categories by enforcing orthogonality constraints on category feature vectors, thereby improving the model's ability to distinguish categories. This method is significantly superior to the traditional cross-entropy loss function.

[0013] (3) Integrating fuzzy logic and deep learning The model of this invention enhances the ability to represent complex charging interest site data by introducing a fuzzy layer to handle uncertainty. Compared with existing technologies, FSE-OPL-DenseNet performs better in handling inconsistencies in image quality and user behavior.

[0014] (4) Three-stage prediction framework The three-stage prediction method proposed in this invention, which separates perturbation values ​​from historical data, effectively captures the complex dependency between charging POI patterns and perturbation values. Compared to traditional methods that rely solely on historical data, this framework has a significant advantage in prediction accuracy. Attached Figure Description

[0015] Figure 1 This is a block diagram of the FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing, according to the present invention.

[0016] Figure 2 Example image recommended for the next charging interest site.

[0017] Figure 3 A schematic diagram illustrating the interaction process between the agent and the environment in reinforcement learning.

[0018] Figure 4 This is a schematic diagram of the overall framework of FSE-OPL-DenseNet.

[0019] Figure 5 This is a schematic diagram of the FSE module.

[0020] Figure 6 This is a schematic diagram of the Dense module and the FSE-Dense module.

[0021] Figure 7 A diagram illustrating the charging queue situation.

[0022] Figure 8 This is a schematic diagram of the ROC curve for dataset 1.

[0023] Figure 9 This is a schematic diagram of the ROC curve for dataset 2.

[0024] Figure 10 This diagram illustrates the loss values ​​of different models during the iteration process.

[0025] Figure 11 Feature maps for charging images in different scenarios.

[0026] Figure 12 Feature maps of Pro images in different scenarios.

[0027] Figure 13 This is a schematic diagram illustrating the application of the recommended solution of this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See Figure 1 This invention provides an FSE-OPL-DenseNet electric vehicle charging recommendation system based on sparse network sharing, including a data preprocessing module 110, a model training module 120, and a recommendation output module 130. The modules communicate with each other in real time via a data interface. Data preprocessing module 110, deployed on a cloud server, is used to preprocess the acquired historical charging order data of users to obtain preprocessed training data; The model training module 120 is deployed on a GPU server and is used to train the FSE-OPL-DenseNet model based on the preprocessed training data to obtain the trained FSE-OPL-DenseNet model. The recommended output module 130 is integrated into a mobile terminal or cloud API interface to receive user charging requests in real time, generate the access probability distribution results of users for each candidate charging interest site based on the trained FSE-OPL-DenseNet model, and generate a recommended list of charging interest sites based on the access probability distribution results.

[0030] Optionally, the step of preprocessing the acquired user historical charging order data to obtain preprocessed training data includes: cleaning the user historical charging order data using an ETL tool, removing invalid order data with charging duration less than a preset first threshold or greater than a preset second threshold, and obtaining valid user historical charging order data as the preprocessed training data.

[0031] Optionally, the FSE-OPL-DenseNet model training includes the following stages: Input image segmentation stage: Obtain the original charging image from the valid user historical charging order data, perform semantic segmentation on the original charging image using the U-Net network, dividing it into a charging area image and a background area image, and using the charging area image as the segmented image, wherein the original charging image includes a charging station location map, a surrounding POI distribution map, and a user's historical check-in trajectory map; Feature extraction stage based on fuzzy attention: Input the original charging image and the segmented image into two parallel FSE-Dense blocks for feature extraction. The process involves several steps: First, feature extraction is performed to obtain target features from the original charging image and the segmented image. Second, feature fusion and optimization are performed by concatenating the target features from the original charging image and the segmented image to obtain fused features. These fused features are then subjected to dimensionality reduction using a global pooling layer and enhanced using a sigmoid function to obtain enhanced fused features. Finally, the enhanced fused features are normalized using a softmax activation function to output the initial access probability distribution of users for each candidate charging interest site. The initial access probability distribution is then optimized using an orthogonal projection loss function to obtain the final access probability distribution.

[0032] Optionally, generating a charging interest site recommendation list based on the access probability distribution results includes: sorting the candidate charging interest sites from high to low according to the access probability distribution results, and selecting the top K candidate charging interest sites to form a charging interest site recommendation list, where K is a preset positive integer; Optionally, the charging interest site recommendation list includes the name of the candidate charging interest site, the distance from the user's current location to the candidate charging interest site, the estimated charging time, surrounding POI information, and the recommendation priority. The recommendation priority is determined by the distance from the user's current location to the candidate charging interest site, and the user's current location is obtained from the user's charging request.

[0033] Optionally, in the two parallel FSE-Dense blocks, each FSE-Dense block includes a Dense module and an FSE module; the Dense module includes four Dense sub-layers for extracting initial features of the original charging image and initial features of the segmented image; the FSE module is embedded after each Dense sub-layer of the Dense module for performing sparse feature filtering and enhancement on the initial features of the original charging image and the initial features of the segmented image to obtain the target features of the original charging image and the target features of the segmented image.

[0034] Optionally, the FSE module performs sparse feature filtering and enhancement in the paths of the initial features of the original charging image and the initial features of the segmented image through the following steps: taking the initial features of the original charging image and the initial features of the segmented image as the original feature maps respectively. Global average pooling: Compresses the spatial features of the original feature map of the path into channel descriptors to quantify the global contribution of each channel. The formula is:

[0035] in, For the first c Each channel is located in eigenvalues, M , H The height and width of the original feature map. For channel global features; Feature activation: Channel weights are dynamically adjusted through a gating mechanism to highlight the sparse features of the original feature map. The formula is as follows:

[0036] in, W 1. W 2 represents the weights of the fully connected layer. b 1. b 2 is the bias. δ It is the ReLU activation function. σ For the sigmoid function, s This is the channel activation value; Fuzzy logic layer processing: Introducing fuzzy member functions and rule layers to handle feature uncertainty; Member function layer: Maps channel features to fuzzy membership degrees in the range [0,1], and fits the fuzzy distribution of the user's historical charging order data, using the following formula:

[0037] Fuzzy rule layer: Integrates multi-channel fuzzy information through AND / OR logic, with the following formula:

[0038] in, For AND operation, For OR operation, R For fuzzy rule sets; Using the channel activation value s With fuzzy processing results FR ( MF ( z Together, we adjust the original feature map to enhance key channels, suppress redundant channels, and complete sparse feature selection. The formula is as follows:

[0039] in, Indicates the first c Recalibrated output feature map of each channel It is the first c The channel activation value of each channel; The recalibrated output feature map is used as the target feature of the path to obtain the target feature of the original charging image and the target feature of the segmented image.

[0040] Optionally, the orthogonal projection loss function is expressed as:

[0041] in, It is a cosine similarity generator. This represents an absolute value generator. It is a mini batch size. , This is the feature vector of the sample.

[0042] Optionally, when training the FSE-OPL-DenseNet model, the model training module uses a total loss function that is a weighted sum of the cross-entropy loss function and the orthogonal projection loss function, and the total loss function is expressed as:

[0043] in, λ This represents the hyperparameter that controls the weights of the orthogonal image loss. Represents the cross-entropy loss function. This represents the orthogonal projection loss function.

[0044] This invention also provides an FSE-OPL-DenseNet electric vehicle charging recommendation method based on sparse network sharing, comprising: The acquired historical charging order data of users is preprocessed by a data preprocessing module deployed on a cloud server to obtain preprocessed training data. The FSE-OPL-DenseNet model is trained using the model training module deployed on the GPU server based on the preprocessed training data, resulting in a trained FSE-OPL-DenseNet model. The system receives user charging requests in real time through a recommendation output module integrated into the mobile or cloud API interface, generates the user's access probability distribution for each candidate charging interest site based on the trained FSE-OPL-DenseNet model, and generates a recommended list of charging interest sites based on the access probability distribution.

[0045] It should be understood that the charging interest sites mentioned in this invention refer to charging piles that users may be interested in.

[0046] It should also be understood that POI refers to points of interest (i.e., charging points of interest) around charging stations that users may be interested in.

[0047] Specifically, POIs include the following 17 categories: food and beverage, hotels and accommodations, shopping, lifestyle services, beauty and personal care, leisure and entertainment, sports and fitness, education and training, culture and media, medical care, automotive services, transportation facilities, finance, real estate, business companies, government agencies, and tourist attractions.

[0048] Compared with the prior art, the present invention has the following beneficial effects: (1) Introduce the FSE module The FSE (Feature Selection Enhancement) module of this invention combines fuzzy logic and attention mechanisms to dynamically adjust the weights of feature channels, enhancing the model's ability to focus on key features. Compared to the traditional DenseNet, the FSE module improves the model's ability to identify subtle features of charging interest sites.

[0049] (2) Application of Orthogonal Projection Loss (OPL) The OPL of this invention reduces feature overlap between categories by enforcing orthogonality constraints on category feature vectors, thereby improving the model's ability to distinguish categories. This method is significantly superior to the traditional cross-entropy loss function.

[0050] (3) Integrating fuzzy logic and deep learning The model of this invention enhances the ability to represent complex charging interest site data by introducing a fuzzy layer to handle uncertainty. Compared with existing technologies, FSE-OPL-DenseNet performs better in handling inconsistencies in image quality and user behavior.

[0051] (4) Three-stage prediction framework The three-stage prediction method proposed in this invention, which separates perturbation values ​​from historical data, effectively captures the complex dependency between charging POI patterns and perturbation values. Compared to traditional methods that rely solely on historical data, this framework has a significant advantage in prediction accuracy.

[0052] Specifically, the solution of the present invention is further described with reference to the following examples: See Figure 2 Example image of recommended charging interest sites. Based on historical charging check-in data, the system predicts the next charging interest site a user will visit within a specific time period (the next charging station the user might be interested in), thereby benefiting the owners / operators of interest points around the charging interest site, such as hotels, shopping malls, and food streets near the user's charging interest site. Taking the weekly activities of electric vehicle users as the research object, charging events are considered as an activity constraint, not limited to distance and price, and corresponding activity routes are recommended to maximize the overall utility of electric vehicle users over a week. Users can decide the time for each charging session, that is, users can charge within the activity time without spending extra time waiting for charging to complete. The charging time will correspondingly affect the number of charging sessions within a week, which is an endogenous variable affecting overall utility.

[0053] See Figure 3 This invention describes the interaction between an agent and its environment in reinforcement learning. To maximize the utility of an electric vehicle user's activities over a week, it models the agent-environment interaction using a Markov Decision Process (MDP) to analyze changes in the user's charging state. Compared to a simple Markov process, this invention adds an action variable A and a reward variable R, meaning that state transitions are not intrinsic but are determined by the state. S t With input actions A t The state at time t+1 is jointly determined by: Decide.

[0054] See Figure 4 The computational framework, designed in three steps, includes: Step 1, Segmentation Stage: Process the input image to classify charging types and separate relevant features from unnecessary background elements, thereby improving the input for subsequent feature extraction.

[0055] Step 2: Feature extraction stage based on fuzzy attention. In this stage, the original image and the segmented image are processed through parallel paths, each path applying a series of FSE-Dense blocks. The FSE-Dense module consists of a Dense module and an FSE module. The former is a series of densely connected convolutional networks used for initial feature extraction, while the latter uses an attention mechanism to selectively enhance features.

[0056] Step 3, the feature fusion and optimization stage, is used to fuse the features extracted from the two paths. After fusion, a global pooling layer and a sigmoid function are used for dimensionality reduction and feature enhancement, respectively. Then, the merged features are processed by a softmax activation function to generate classification results. Furthermore, an orthogonal projection loss function optimizes classification by minimizing intra-class variance and maximizing inter-class distance. The output of this framework is the classification result, i.e., the preliminary access probability distribution, indicating whether the input image contains features of the next charging interest site, while OPL provides the confidence or accuracy of the classification. The overall structure aims to capture and emphasize the most discriminative features in the input image, thereby improving the classification performance of the next charging interest site recommendation.

[0057] See Figure 5 The FSE module further refines channel feature adjustment by applying attention or gating processing along the channel dimension. Building upon the SE module, it recalibrates channel features, assigning higher importance to channels exhibiting characteristics of the next charging interest site, thus focusing the model's attention on relevant morphological patterns and enhancing the discriminative power of the feature maps. The FSE module not only processes the feature maps extracted by convolution operations through compression and activation operations to obtain global channel-level features, but also extracts the relationships between different channels and adjusts their weights accordingly. Thus, multiplying the original feature map by the channel weights yields the final features. Furthermore, the FSE module can further filter and enhance the features output by the fuzzy neural network, providing targeted guidance for deep feature extraction in the neural network. Its main functions include: Channel feature recalibration: compressing spatial information through global average pooling to obtain global feature descriptors, and dynamically adjusting the importance of each channel through gating mechanisms. Fuzzy logic support: introducing fuzzy membership functions and fuzzy rule layers to enhance the model's ability to handle uncertainty and fuzziness. Feature recalibration: adjusting the weights of the original feature map using the output of activation operations to optimize the final feature representation.

[0058] See Figure 6 This paper demonstrates the difference between the traditional Dense module and the FSE-dense module of this invention. The connection method between each module in the FSE-dense module of this invention is the same as that in the traditional dense module, the difference being the addition of an FSE after the traditional convolutional block for adaptive feature reweighting. Therefore, this design allows for immediate recalibration of features after each feature transformation step, optimizing the quality and efficiency of the information flow. The introduction of the FSE module provides dynamic weight adjustment for features in each channel, which can be enhanced or suppressed based on the importance of the feature signal in distinguishing the features of the next charging interest site. This not only improves the model's adaptability and attention to complex data features but also minimizes potential information loss in the deep network structure by optimizing the information flow, ensuring that the model accurately captures subtle charging changes related to the next charging interest site. Therefore, this method significantly improves the recommendation accuracy and reliability of the model in high-precision next charging interest site analysis tasks. The structure of the DenseNet model is particularly suitable for next charging interest site analysis due to its efficient feature propagation and reuse. In charging recommendation, the DenseNet model ensures the extraction of comprehensive features from charging images, capturing both macroscopic location structure and microscopic charging details, which is crucial for accurate classification. The DenseNet model is responsible for extracting fine-grained features and passing these features to the classification layer.

[0059] This invention's strategy of deeply integrating the FSE module with the dense layers of DenseNet aims to effectively combine the inherent dense connectivity of DenseNet with the channel feature recalibration function of the FSE module. This deep integration aims to leverage the rich inter-layer information flow provided by dense connections, combined with the fine-tuning of key features by the FSE module, to improve the model's ability to identify key charging morphology features when recommending the next charging interest site. The FSE module is integrated into the DenseNet module, whose core is its densely connected convolutional layer structure. The l-th layer accepts feature maps from previous layers as input, ensuring efficient feature propagation and reuse. In FSE-OPL-DenseNet, DenseNet is used as the basic feature extractor, responsible for capturing fine-grained features from the input image. The core layer sequence of the DenseNet module consists of batch normalization (BN), ReLU activation function, and convolutional layers (Conv), with each convolutional layer immediately followed by an FSE module.

[0060] In processing images of next charging stations of interest, the Blur Squeeze & Excitement (FSE) module is designed to enhance the neural network's attention to and representation of charging features by weighting feature channels. This approach amplifies important features while suppressing unimportant ones, thereby enhancing the network's ability to capture key information in the image. However, given the morphological similarity of charging stations in images of next charging stations of interest, relying solely on FSE may lead to the model inaccurately classifying features from different categories as similar. To address this issue, this invention introduces Orthogonal Projection Loss (OPL). OPL enforces orthogonality constraints to ensure that feature vectors from different categories are orthogonal in the feature space, thereby strengthening the boundaries between categories and improving the model's ability to distinguish between them. This approach ensures that even if the FSE module amplifies certain features, OPL prevents these features from being confused with features from other categories, thus reducing the problem of feature overlap between categories. By simultaneously optimizing feature recalibration and category orthogonality, this approach enhances the model's ability to interpret complex biomarkers and improves recognition accuracy in cases of feature diversity and subtle differences. It is particularly suitable for capturing subtle but crucial feature differences in images of next charging stations of interest, thereby significantly improving classification accuracy and recommendation reliability.

[0061] The specific technical implementation is as follows: (1) Overall technical architecture: FSE-OPL-DenseNet model and recommendation system This invention achieves recommendation functionality by constructing a technical architecture of "FSE-OPL-DenseNet model + recommendation system". The core is the integration of the "Foggy Squeeze Excitation (FSE) module", "Orthogonal Projection Loss (OPL)", and "Sparse Feature Sharing Framework" to form a closed loop of "data processing - feature extraction - model optimization - recommendation output", as detailed below: The recommendation system architecture consists of three core modules, physically connected as "data preprocessing module → model training module → recommendation output module," and logically achieving real-time data transmission through a data interface. 1) Data preprocessing module: Deployed on a cloud server, it cleans user charging order data using ETL tools, removes invalid order data with charging time less than a preset first threshold or greater than a preset second threshold, for example: removes invalid orders with charging time <10 minutes or >12 hours, and obtains valid user historical charging order data as the preprocessed training data, which is then output to the model training module.

[0062] 2) Model training module: The FSE-OPL-DenseNet model is deployed on a GPU server. After loading the training data, the model training is completed through backpropagation, and the trained model is output to the recommendation output module. 3) Recommendation output module: Integrated into the mobile charging APP or cloud API interface, it receives user charging requests in real time. The user charging request records the user's current location and historical charging records. It calls the trained model to generate a Top-K recommendation list and feeds it back to the user terminal.

[0063] (2) Core Model: The three-step recommendation process of the FSE-OPL-DenseNet model The model recommendation process is divided into three stages: "input image segmentation → feature extraction based on fuzzy attention → feature fusion and optimization". Each step is logically connected to ensure the efficient transformation of features from "raw data → screening and enhancement → accurate recommendation". Step 1: Input Image Segmentation The original charging images are obtained from the valid user historical charging order data. Semantic segmentation of these images is performed using a U-Net network, dividing them into charging area images and background area images. The charging area images are used as the segmented images. The original charging images include a charging station location map, a surrounding POI distribution map, and a user's historical check-in trajectory map. For example, charging station signs and POIs within a 1km radius constitute the charging area image, while irrelevant buildings and distant roads constitute the background area image. Background interference is removed, and core features are retained. The segmented images are then transmitted via a data interface to the "feature extraction based on fuzzy attention" step, providing clean input for subsequent feature extraction.

[0064] Step 2: Feature extraction based on fuzzy attention A dual-path feature extraction structure is constructed, in which the target features of the original charging image and the segmented image are extracted through parallel "FSE-Dense blocks" to achieve the sharing and filtering of sparse features: 1) Basic Structure: Dense Blocks: The DenseNet model is a CNN-based model consisting of 5 dense blocks. Each dense block contains 4 sub-layers: "Batch Normalization (BN) → ReLU activation → 3×3 convolution". Feature sharing is achieved through "dense connections"—the input of the l-th layer is a concatenation of feature maps from the previous l-1 layers, as shown in the formula: (1) in, H l The sublayer synthesis function is (BN+ReLU+3×3 convolution). Using the feature map of the preceding layer can reduce parameter redundancy and improve the efficiency of sparse feature propagation (corresponding to...) Figure 6 (Dense module structure). Among them, Indicates the 0th layer, ..., l The invention involves connection operations on the feature maps generated at layer 1. H l Multiple inputs are concatenated into a single tensor. A synthesis function for three consecutive generators is introduced: Batch Normalization (BN) → Rectified Linear Unit (ReLU) → 3×3 Convolutional Layer (Conv). The dense block consists of four BN→ReLU→3×3 Conv layers and an input layer, each connected via a feedforward fully connected method. The input layer is connected to all other BN→ReLU→3×3 Conv layers, with the first BN→ReLU→3×3 Conv layer connected to the next three, and so on. The transition layer in the neural network is a dimensionality reduction mechanism. This layer sequentially uses 1×1 convolutional layers to reduce the depth of the feature maps, followed by spatial downsampling of the resulting feature maps using pooling layers. The term M represents a set of feature maps. c This represents the compression factor, which is usually preset to 0.5. This setting compresses the number of channels to half of the original number, thereby reducing parameters and improving the computational efficiency of the network.

[0065] 2) Core Innovation: FSE (Fuzzy Extrusion) Module: An FSE module is embedded after each Dense sub-layer, combining fuzzy logic and channel attention to achieve sparse feature selection and enhancement. The steps are as follows (corresponding to...). Figure 5 ): ① Global average pooling: Compresses spatial features into channel descriptors and quantifies the global contribution of each channel. The formula is: (2) in, For the first c Each channel is located in eigenvalues, M , H The height and width of the original feature map. For channel global features; ② Feature Incentive: Channel weights are dynamically adjusted through a gating mechanism to highlight sparse features useful for recommendations (such as channels related to "fast charging stations" and "low queuing rates"). The formula is: (3) in, W 1. W 2 represents the weights of the fully connected layer. b 1. b 2 is the bias. δ It is a ReLU activation function (suppressing invalid features). σ This is the sigmoid function (outputs weights in the range [0,1]).s This is the channel activation value; ③ Fuzzy logic layer processing: Introducing fuzzy member functions and rule layers to handle feature uncertainties (such as fluctuations in user charging time): Member function layer: Maps channel features to fuzzy membership degrees in the range [0,1], fitting the fuzzy distribution of the charging data, with the following formula: (4) in, , These are the empirically optimal parameters; The cross-entropy loss is defined as follows: (5)

[0066]

[0067] in, y This indicates the actual charging status. This indicates the charging prediction result. , This indicates that a feature extraction module is used. x It is input. , It's a classification module. , This represents the projected vector for classifiable learning, used for cross-entropy loss, and weights the ground-value features to achieve orthogonality between these categories without being obvious.

[0068] Fuzzy rule layer: Integrates multi-channel fuzzy information through "AND / OR" logic, with the following formula: (6) in, For AND operation, For OR operation, R For fuzzy rule sets; ④ Feature recalibration: using the aforementioned channel activation values s With fuzzy processing results FR ( MF ( z Together, we adjust the original feature map to enhance key channels, suppress redundant channels, and complete sparse feature selection. The formula is as follows: (7) in, Indicates the first c Recalibrated output feature map of each channel It is the first c The channel activation value of each channel.

[0069] Step 3: Feature Fusion and Optimization ① Feature fusion: The feature maps extracted from the two paths (original image path and segmented image path) are concatenated, and the dimensionality is reduced to 1×1×C feature vector through global average pooling. Then, the key features (such as the "business district + fast charging" combined feature) are enhanced by the sigmoid function. ② OPL (Orthogonal Projection Loss) Optimization: This invention improves cross-entropy loss by enforcing orthogonality in the intermediate vector space. Given a dataset A pair of inputs and outputs , This is the feature output of intermediate network operations. The goal of this invention is to analyze these features. Clustering constraints are implemented to ensure that features from different categories are orthogonal and features from the same category are highly similar. To this end, this invention defines a unified loss function that simultaneously guarantees intra-class clustering and inter-class orthogonality in mini-batches. The OPL loss function is introduced to constrain the feature space to "intra-class clustering and inter-class orthogonality," avoiding category confusion. The formula is as follows: (8) in, It is a cosine similarity generator. This represents an absolute value generator. It is a mini batch size. , The sample feature vector is shown below. The cosine similarity generator is as follows: (9) in, express Normalization generator. Normalization is crucial for ensuring that the output of orthogonal projection loss remains consistent with the intrinsic angular properties of cross-entropy loss.

[0070] ③ Total Loss and Optimization: The total loss is the weighted sum of the cross-entropy loss (classification loss) and the OPL loss, as shown in the formula: (10) in, λ This represents the hyperparameter that controls the weights of the loss in orthogonal images. λ =0.05 is the empirically optimal weight. The Adam optimizer (initial learning rate = 0.001, decaying to half its original value in rounds 10 and 20) is used to update the model parameters through backpropagation, minimizing the total loss to improve classification accuracy. The entire framework can be viewed as an optimization problem, as shown below: (11) in, This represents the model parameters. The iterative process of the SE-OPL-DenseNet framework uses backpropagation to transmit error information back to the hidden layers, enabling continuous updating and iteration of the weight matrix between the hidden layers until the network converges.

[0071] (3) Implementation of sparse sharing mechanism Feature sharing: Through dense connections of Dense blocks, features from different layers can be reused. For example, the "charging point location feature" can be shared across multiple layers, reducing parameter redundancy and improving the propagation efficiency of sparse features. Sparse filtering: The FSE module dynamically adjusts channel weights and processes fuzzy logic to retain only key sparse features that contribute to the recommendation, such as "users prefer fast charging" and "avoid outdoor charging points on rainy days," while suppressing redundant features such as background textures irrelevant to charging points, thus achieving efficient utilization of sparse features.

[0072] (4) Practical application process Taking electric vehicle charging apps as an example, the recommended workflow is as follows: When a user opens the app and authorizes their current location information, the app's background calls the data preprocessing module to obtain the user's historical charging records (such as charging preferences and commonly used charging point types in the past 3 months), current location, and charging station data and POI information within 5km of the current location in real time. The preprocessed data is input into the recommendation output module, which loads the trained FSE-OPL-DenseNet model to predict the probability distribution of the user's next charging interest site. The model outputs a Top-3 recommendation list, which includes the name of the charging point, distance (e.g., "Lianghang Creative Park Charging Station, 2.3km"), estimated charging time (e.g., "45 minutes"), nearby POIs (e.g., "including food court"), and recommendation priority. After the user selects a recommended location, the app calls a navigation interface (such as the Gaode Map API) to provide route planning and supports a charging reservation function, synchronizing the queuing status of charging points in real time.

[0073] (5) Case Design 1) Experimental data and evaluation indicators The primary task in analyzing user charging needs under multidimensional characteristics is data collection. Complete user demand analysis data types can be broadly categorized into three types: basic user information, charging element data (charging facility location, charging costs, etc.), and user preferences (user charging history, rating records, etc.). Considering whether the data changes over time, these three types of data can be further divided into static data and dynamic data. Typically, basic user information is static data, a statistical summary of a user's basic personal information, while charging information and user preferences are behavioral information that changes continuously over time, thus belonging to dynamic information. Because basic user information involves user privacy and is very difficult to obtain, this invention only uses charging element data and user preference data.

[0074] Data transformation is primarily aimed at ensuring data consistency across different scales. One-hot encoding can handle discontinuous numerical features and expand features to some extent. Sparse encoding can also achieve better spatial performance in practice, and this method is simple and easy to design and implement. Therefore, this invention employs one-hot encoding for data transformation. In classification problems, imbalanced datasets often occur; therefore, this invention uses a method of synthesizing minority class data to balance the sample size of different classes. Since this algorithm requires a large amount of text and image data related to user location selection, this invention uses keyword extraction to process text data and information comparison to process image data. Data validation and quality checks were further screened and supplemented using the pandas package in Python. Furthermore, since many features in the data are discrete variables that cannot be directly input into the model and need to be converted into vectors, this invention introduces embedding to complete the feature conversion process.

[0075] 2) Dataset This invention uses two real-world datasets for experiments. The Jiading Darongcheng charging station dataset contains 37,550 charging data entries from electric vehicle charging stations in Shanghai from 2020 to 2021. The Baoshan Baoan Road No. 188 charging station dataset contains 33,860 charging data entries from electric vehicle charging stations in Shanghai from 2020 to 2021. In the Baoshan Baoan Road No. 188 charging station dataset, short text descriptions of charging interest sites are used. Furthermore, the Baoshan Baoan Road No. 188 charging station dataset also records the social relationships between each user and other users. In the Baoshan Baoan Road No. 188 charging station dataset, each charging interest site has an average of 2.67 descriptive words, and users have an average of 4.36 friends.

[0076] Table 1. Introduction to the Experimental Dataset

[0077] The charging check-in records in the dataset all include timestamps and latitude / longitude coordinates. Table 1 presents the statistical characteristics of the Jiading Darongcheng charging station and Baoshan Baoan Highway No. 188 charging station datasets. This invention also calculates the distance between two random charging interest sites for each dataset based on the Haversine formula. The average distances between charging interest sites in the Jiading Darongcheng charging station and Baoshan Baoan Highway No. 188 charging station datasets are 14.5 km and 15.5 km, respectively. This invention removes inactive users and charging interest sites from each dataset, ensuring that each user has at least 10 charging check-in records and that each charging interest site has been visited by at least 10 users. For performance evaluation, this invention sorts each user's charging check-in records according to timestamp order, using the most recent 20% of each user's charging check-in records as the test set, the earliest 70% as the training set, and the remaining 10% as the validation set to adjust parameters.

[0078] 3) Data Fields Based on the collected and preprocessed user charging element data and user preference data, this invention establishes a multi-dimensional feature system for mining user charging demand, as shown in Table 2.

[0079] Table 2 Multidimensional Feature System for User Needs Mining

[0080] 4) Model Evaluation and Setup Accuracy: Accuracy is the most basic evaluation metric in problem classification. Accuracy is defined as the percentage of correctly predicted results out of all samples. (12) Precision: Precision refers to the prediction result, that is, the probability that a positive sample actually occurs among all predicted positive samples. (13) Recall: Recall refers to the original sample, that is, the probability that a sample is predicted as a positive sample among actual positive samples. (14) F1 score: Precision and recall metrics can fluctuate. In some cases, both precision and recall should be considered. The most commonly used method is the F1 score, which is a weighted harmonic mean of precision and recall.

[0081] (15) TP and FP represent true positive and false positive, respectively; TP and FN represent true positive and false negative, respectively.

[0082] Due to memory limitations of the graphics processing unit, a batch size of 5 was chosen. The initial learning rate was set to 0.001, and halved at the 10th and 20th epochs. The entire training process took 30 epochs. To reduce the risk of overfitting in the neural network model, this invention employs several strategies, each playing a crucial role in ensuring the model's robustness and generalization ability: Dropout: To further refine and address potential overfitting issues in the experiments, this invention employs dropout as a regularization technique during model training. After a series of preliminary tests, the invention carefully selected a dropout rate to find an optimal balance that prevents overfitting while maintaining the model's learning ability. Specifically, certain layers of the neural network use a dropout rate of 0.3. In each training iteration, this rate randomly deactivates half of the neurons in these layers, effectively reducing the risk of overfitting. This method encourages the network to learn more robust features independent of specific neurons, thereby improving the model's generalization ability to unseen data.

[0083] Fuzzy Layers: Adding fuzzy layers to this model architecture increases the ability to handle uncertainty and variability, which is particularly beneficial for complex data such as the next charging interest site. Fuzzy layers can introduce a form of regularization by managing imprecise information, enabling the model to learn more general patterns rather than simply memorizing training data.

[0084] Cross-validation: To further prevent overfitting, this invention employs a 10x cross-validation technique. This involves dividing the dataset into ten parts, nine for training and one for testing, then alternating between the two parts so that each part is used for testing. This approach helps evaluate the model's performance across different subsets of data, ensuring that the model generalizes well to new, unseen data. By combining these techniques, the model achieves a balance between learning from training data and maintaining its ability to generalize to new data, effectively reducing the risk of overfitting.

[0085] The experiments were conducted on a GPU (NVIDIA® GeForce RTX 2080) with 8GB of memory, using PyTorch. All experiments were performed according to the above settings; 60% of the data was used for training and 40% for testing.

[0086] (6) Results and Analysis: Comparison with the baseline model Baseline Models: In this invention, several baseline models were studied and compared to evaluate their performance in application environments. These models include: 1) DenseNet201 Efficient CNN (DECNN): This model is a fast and efficient convolutional neural network utilizing the DenseNet201 architecture. It is known for its fast processing speed and high efficiency. 2) DeepLeukNet Augmented CNN (DLACNN): This model uses data augmentation techniques to generate images, solving the inherent overfitting problem of the model. This approach is particularly valuable in the field of deep learning. 3) Particle Swarm Optimization (PSO) Improved ResNet50 (IRP-CNN): This model enhances the ResNet50 convolutional neural network model by integrating PSO, aiming to improve the network's performance and efficiency. 4) Latent Space Representation Learning Model (LSRLM): This model focuses on latent space learning, which is crucial for effective representation learning in complex data scenarios. 5) Ensemble Pre-trained CNN (EPTCNN): An ensemble approach that combines a classifier with a pre-trained convolutional neural network. This approach aims to leverage the strengths of multiple models to improve overall performance.

[0087] Based on the results of a series of preliminary experiments, this invention selected 0.3 as the dropout rate. In these experiments, the invention tested the impact of different dropout rates (ranging from 0.1 to 0.5, with a step size of 0.05) on model performance. The study found that setting the dropout rate to 0.3 achieves the optimal balance in terms of validation set performance, effectively mitigating overfitting while maintaining the model's predictive power. While determining the dropout rate directly from mathematical or linear algebraic theory can be challenging, this paper chooses an empirical approach combined with the model's actual performance. This method, although reliant on trials and errors, ensures that the optimal dropout rate for the model structure is determined through extensive experimental validation.

[0088] Table 3 lists the charging confusion matrix for four types of multi-level classification problems: no queuing required (benign), moderate queuing (early stage), long queue (early stage), and crowded queue (late stage), as shown in the example. Figure 7 .

[0089] Each cell in the matrix displays a percentage of the total dataset. The diagonal entries represent correct classifications for each category: 15.01% for benign, 30.10% for early, 29.41% for early, and 24.48% for late, indicating high accuracy in classifying different stages. Entries outside the diagonal show minimal misclassifications, each less than 0.16% of the total dataset, demonstrating very high precision and recall across all categories. Table 4 simplifies the problem to a binary classification, distinguishing between benign and malignant charging predictions. Values ​​are given as percentages of total observations. The model correctly predicted benign recommendations with a TN value of 15.08% and correctly identified malignant recommendations with a TP value of 84.38%. FP and FN are very low, at 0.31% and 0.23% respectively, further demonstrating the model's effectiveness. Figure 8 The receiver operating characteristic (ROC) curve is shown, with an area under the curve (AUC) of 0.907. The ROC curve for true positive rate versus false positive rate illustrates the trade-off between sensitivity and specificity. An AUC value close to 1 indicates high separation, meaning the model can accurately distinguish between different categories. The curve is significantly higher than the indiscriminate line (diagonal dashed line), indicating a good predictive model.

[0090] Table 3. Confusion matrix of prediction results for dataset 1

[0091] Table 4. Confusion matrix of prediction results for dataset 1 (binary classification)

[0092] Table 5 provides a comprehensive comparison of the proposed framework's performance metrics using Dataset 1 and the established baseline model. The proposed framework demonstrates a significant advantage across all evaluation parameters, achieving an accuracy of 99.00%, precision of 98.88%, recall of 98.94%, and an F1 score of 98.91%. These metrics not only surpass the best-performing baseline model, DLACNN (96.09% accuracy and 94.27% F1 score), but also achieve statistically higher levels. A paired-tailed t-test with a significance threshold of 0.01 shows that the proposed framework achieves a double-aster score on all metrics. The significance level indicates a significant improvement in predictive effectiveness. This level of significance strongly suggests that the improvements brought about by the proposed framework are not due to random variations in the data, but rather reflect genuine advancements in model architecture and data processing techniques. The robustness of the proposed model is particularly evident in its precision and recall scores, which are important indicators of the model's reliability in effectively classifying true positive results without increasing false positive results.

[0093] Table 5 Comparison with baseline models (Dataset 1)

[0094] Note:" "This indicates that at a significance level of 0.05, the model is statistically significant compared to the optimal baseline model via a two-tailed paired t-test." "Indicates that it has higher statistical significance at the significance level of 0.01.

[0095] Table 6. Confusion matrix of prediction results for Dataset 2 (binary classification)

[0096] Figure 9 The ROC curve for Dataset 2 is shown, with an AUC of 0.940, indicating excellent discriminative ability of the model. The curve deviates significantly from the no-discrimination line, highlighting the strong predictive power of the proposed framework. Table 7 provides a detailed comparison of the performance on Dataset 2, demonstrating the comparison of the proposed framework with other baseline models. In this comparison, the proposed framework significantly outperforms other models in accuracy, precision, recall, and F1 score, achieving 97.07%, 98.17%, 96.04%, and 97.09% respectively, far exceeding other models. For example, compared to the closest performing model, DLACNN, the proposed framework improves accuracy by 3.5 percentage points, precision by 3.2 percentage points, recall by 4.24 percentage points, and F1 score by 3.73 percentage points. These metrics are indicated by double asterisks (...). The results indicate that the significant performance improvement compared to the best baseline model is statistically significant and was verified by a two-tailed paired t-test at the 0.01 significance level.

[0097] Table 7 Comparison with baseline models (Dataset 2)

[0098] Note:" "This indicates that at a significance level of 0.05, the model is statistically significant compared to the optimal baseline model via a two-tailed paired t-test." "Indicates that it has higher statistical significance at the significance level of 0.01.

[0099] The results of the baseline model and the model of this invention are compared, covering performance metrics and structural feature analysis. The DECNN model is fast, but its ability to capture subtle differences in charging morphology is limited. DLACNN integrates data augmentation, improving generality and mitigating overfitting. The model of this invention combines fuzzy logic and the SE module, significantly improving recommendation accuracy. IRP-CNN optimizes efficiency through PSO, but its inter-class feature discrimination is not as good as the model of this invention. LSRLM focuses on latent space learning, but its ability to identify subtle differences is weak. EPTCNN uses an ensemble learning method, but the orthogonal projection loss of the model of this invention provides deeper support. In summary, the FSE-OPL-DenseNet framework significantly improves performance, and innovative designs such as fuzzy logic, orthogonal feature space, and the SE mechanism are key to addressing complex challenges.

[0100] (7) Discussion of Results To examine whether all the new settings in the SE-OPL-DenseNet model of this invention have a precise impact on prediction performance, this invention compares the complete model and four incomplete models. Table 8 shows the definition of incomplete models; 1 indicates that the model includes the corresponding operation, and 0 indicates that the model does not include these operations. Table 8 shows a comparison between SE-DenseNet-169 and the models defined in Table 9 that do not include one of the operations. Although frames 1, 2, ... and 5 perform well, it can be inferred that each setting and operation improves the effectiveness of the model. These tables highlight the progressive improvements brought by each component in the SE-OPL-DenseNet model. The comparison results show that in FSE-DenseNet-169, the highest accuracy is achieved when transfer learning, orthogonal projection loss, the FSE module, and the segmentation operation are fully integrated. The performance of the incomplete models is slightly lower, which highlights the importance of each feature in enhancing the model's predictive power, especially in the challenging task of charging recommendation.

[0101] Table 8 Definition of Incomplete Model

[0102] Table 9 Comparison between SE-DenseNet-169 and models that do not perform any operations

[0103] In addition to a comprehensive examination of each component in the SE-OPL-DenseNet model, this invention explores the optimization of hyperparameters, particularly the λ value in the loss function. This hyperparameter plays a crucial role in balancing the contribution of the orthogonal projection loss to the overall loss function.

[0104] Table 10 lists the results of testing different values ​​of λ, aiming to determine the optimal setting for this parameter in classification tasks. The model's accuracy varies with different λ values, highlighting the sensitivity of the model's performance to this particular hyperparameter. The results show that a λ value of 0.05 produces the highest accuracy and is the optimal choice for classification tasks. As the λ value increases, the accuracy gradually decreases, indicating that in this specific case, lower weights in the orthogonal projection loss are more effective. This finding is particularly insightful because it underscores the importance of fine-tuning hyperparameters in complex models like these.

[0105] Careful calibration of λ not only improves model performance but also provides insights into the behavior of the loss function and its impact on the model. Clearly, the choice of hyperparameters significantly influences model effectiveness, especially in nuanced tasks where accuracy is critical, such as charging recommendations. Research in this area highlights the importance of hyperparameter optimization in developing advanced machine learning models for the next generation of charging interest sites.

[0106] Table 10 Comparison between different hyperparameters

[0107] Figure 10 This demonstrates intuitively how the loss values ​​of different models change over training time. Initially, the loss values ​​of all models decrease sharply, a typical observation in the early stages of training. However, as training progresses, especially in the latter half, the loss trajectory of the model proposed in this invention differs significantly from other models. This differentiation is characterized by a more pronounced and stable reduction in loss values, highlighting the efficiency and effectiveness of the proposed model. An interesting observation is the impact of applying the OPL and SE modules individually versus their combined application on performance. While each component can individually enhance the base model, their synergistic integration significantly improves model performance. This joint application leverages the OPL's ability to refine the feature space and the SE module's ability to recalibrate channel features, thereby significantly improving the model's accuracy and efficiency. Visual analysis, in particular, provides an insightful window into the computer vision recognition process of the next charging interest site, laying the foundation for developing more advanced models in the future.

[0108] The algorithm model of this invention performs excellently in detecting the next charging interest site, making reliable recommendations, which is crucial for users' charging needs. With the help of this model, accurate recommendations can be made earlier, greatly increasing the likelihood of successful charging and further improving the success rate of next interest site recommendations. Admittedly, compared to some existing computational models, this framework, which includes fuzzy inference, may increase computational resource consumption. However, the evaluation of this invention using k times the computation time shows that although fuzzy inference introduces an additional computational burden, the convergence speed is only slightly improved. Specifically, for small-batch online detection, the time difference may only be a few milliseconds. The computational time requirement for next charging interest site detection is not particularly high. Therefore, although the model proposed in this invention is more computationally complex than some existing models, this is not the model's main limitation.

[0109] Next, this invention first utilizes tensor computation visualization tools to interpret the "black box" nature of CNNs. This method involves explaining the feature maps generated by the attention mechanism. Secondly, it evaluates the performance of various models by observing changes in loss values ​​over different time periods, thus providing an intuitive metric for model efficiency.

[0110] like Figure 11 As shown, feature maps of benign charging images were examined in different scenarios, including: (a) low-level feature maps; (b) high-level feature maps; (c) low-level feature maps without attention mechanisms; (d) high-level feature maps without attention mechanisms; (e) low-level feature maps for segmentation; and (f) high-level feature maps for segmentation. Similarly, Figure 12 The feature maps of the Pro image are shown in these cases. Notably, high-level feature maps are more abstract than low-level feature maps and are easier for computers to interpret. There is a clear difference between feature maps at the same level in models with and without attention mechanisms. The feature maps generated by the model with attention mechanisms show more pronounced contrast. This indicates that the attention mechanism effectively amplifies the importance of key information. Furthermore, although the feature maps of the segmented image contain relatively less information, they complement the feature maps of the original image, thereby improving the model's prediction accuracy.

[0111] Most current charging recommendation map apps are based on distance and price, neglecting users' travel needs. Furthermore, due to a lack of maintenance of charging facilities, some charging stations are unavailable, and current recommendation systems do not consider charging station availability. The main function of this invention is to plan travel routes for users based on their activity needs, under both conditions of charging constraints and those without. Figure 13 ).

[0112] In summary, the proposed FSE-OPL-DenseNet model is a pioneering method for charging recommendation. This framework cleverly combines fuzzy logic with advanced deep learning techniques to handle the complexity and variability of next charging interest sites. Empirical evaluations on a comprehensive next charging interest site dataset demonstrate the framework's superior performance. FSE-OPL-DenseNet achieves an impressive accuracy of 99.26%, significantly outperforming traditional models and improving upon the best-performing baseline by 3.01%. Precision, recall, and F1 score are also excellent, reaching 98.88%, 98.94%, and 98.91%, respectively, far exceeding comparable models such as DLACNN, which achieve 96.09% accuracy and 94.27% F1 score. These results are statistically validated using a two-tailed paired t-test, confirming the high robustness and reliability of the proposed model. Integrating fuzzy attention mechanisms and orthogonal projection loss into a deep learning framework not only improves recommendation accuracy but also ensures adaptability to complex patterns in next charging interest sites.

[0113] Specific embodiments of the invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing can be advantageous.

[0114] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims.

Claims

1. A sparse network-based FSE-OPL-DenseNet electric vehicle charging recommendation system, characterized in that, It includes a data preprocessing module, a model training module, and a recommendation output module. These modules communicate with each other in real-time via a data interface. The data preprocessing module, deployed on a cloud server, is used to preprocess the acquired historical charging order data of users to obtain preprocessed training data. The model training module, deployed on a GPU server, is used to train the FSE-OPL-DenseNet model based on the preprocessed training data, and obtain the trained FSE-OPL-DenseNet model. The recommendation output module, integrated into the mobile or cloud API interface, is used to receive user charging requests in real time, generate the access probability distribution of users for each candidate charging interest site based on the trained FSE-OPL-DenseNet model, and generate a recommended list of charging interest sites based on the access probability distribution.

2. The system according to claim 1, characterized in that, The process of preprocessing the acquired historical user charging order data to obtain preprocessed training data includes: The user's historical charging order data is cleaned using an ETL tool, and invalid order data with charging duration less than a preset first threshold or greater than a preset second threshold is removed, so as to obtain valid user historical charging order data as the preprocessed training data.

3. The system according to claim 2, characterized in that, The training of the FSE-OPL-DenseNet model includes the following stages: Input image segmentation stage: Obtain the original charging image from the valid user historical charging order data, perform semantic segmentation on the original charging image through the U-Net network, divide it into charging area image and background area image, and use the charging area image as the segmented image. The original charging image includes a charging station location map, a surrounding POI distribution map, and a user historical check-in trajectory map. Feature extraction stage based on fuzzy attention: The original charging image and the segmented image are respectively input into two parallel FSE-Dense blocks for feature extraction to obtain the target features of the original charging image and the target features of the segmented image; Feature fusion and optimization stage: The target features of the original charging image and the target features of the segmented image are concatenated to obtain fused features; the fused features are first reduced in dimensionality by a global pooling layer, and then enhanced by a Sigmoid function to obtain enhanced fused features; The enhanced fusion features are normalized using the softmax activation function to output the preliminary access probability distribution of users for each candidate charging interest site; the preliminary access probability distribution is then optimized using the orthogonal projection loss function to obtain the final access probability distribution.

4. The system according to claim 1, characterized in that, The process of generating a list of recommended charging interest sites based on the access probability distribution results includes: Based on the access probability distribution, the candidate charging interest sites are sorted from high to low access probability, and the top K candidate charging interest sites are selected to form a charging interest site recommendation list, where K is a preset positive integer.

5. The system according to claim 4, characterized in that, The charging interest site recommendation list includes the name of the candidate charging interest site, the distance from the user's current location to the candidate charging interest site, the estimated charging time, the surrounding POI information, and the recommendation priority. The recommendation priority is determined by the distance from the user's current location to the candidate charging interest site, and the user's current location is obtained from the user's charging request.

6. The system according to claim 3, characterized in that, In the two parallel FSE-Dense blocks, each FSE-Dense block includes a Dense module and an FSE module; The Dense module includes four Dense sub-layers, which are used to extract the initial features of the original charging image and the initial features of the segmented image; The FSE module is embedded after each Dense sub-layer of the Dense module and is used to perform sparse feature filtering and enhancement on the initial features of the original charging image and the initial features of the segmented image to obtain the target features of the original charging image and the target features of the segmented image.

7. The system according to claim 6, characterized in that, The FSE module performs sparse feature filtering and enhancement in the paths of initial features in the original charging image and the initial features in the segmented image through the following steps: The initial features of the original charging image and the initial features of the segmented image are respectively used as the original feature maps; Global average pooling: Compresses the spatial features of the original feature map of the path into channel descriptors to quantify the global contribution of each channel. The formula is: in, For the first c Each channel is located in eigenvalues, M , H The height and width of the original feature map. For channel global features; Feature activation: Channel weights are dynamically adjusted through a gating mechanism to highlight the sparse features of the original feature map. The formula is as follows: in, W 1. W 2 represents the weights of the fully connected layer. b 1. b 2 is the bias. δ It is the ReLU activation function. σ For the sigmoid function, s This is the channel activation value; Fuzzy logic layer processing: Introducing fuzzy member functions and rule layers to handle feature uncertainty; Member function layer: Maps channel features to fuzzy membership degrees in the range [0,1], and fits the fuzzy distribution of the user's historical charging order data, using the following formula: Fuzzy rule layer: Integrates multi-channel fuzzy information through AND / OR logic, with the following formula: in, For AND operation, For OR operation, R For fuzzy rule sets; Using the channel activation value s With fuzzy processing results FR ( MF ( z Together, we adjust the original feature map to enhance key channels, suppress redundant channels, and complete sparse feature selection. The formula is as follows: in, Indicates the first c Recalibrated output feature map of each channel It is the first c The channel activation value of each channel; The recalibrated output feature map is used as the target feature of the path to obtain the target feature of the original charging image and the target feature of the segmented image.

8. The system according to claim 1, characterized in that, The orthogonal projection loss function is expressed as: in, It is a cosine similarity generator. This represents an absolute value generator. It is a mini batch size. , This is the feature vector of the sample.

9. The system according to claim 8, characterized in that, When training the FSE-OPL-DenseNet model, the model training module uses a total loss function that is a weighted sum of the cross-entropy loss function and the orthogonal projection loss function. The total loss function is expressed as follows: in, λ This represents the hyperparameter that controls the weights of the orthogonal image loss. Represents the cross-entropy loss function. This represents the orthogonal projection loss function.

10. A method for recommending electric vehicle charging based on sparse network sharing using FSE-OPL-DenseNet, characterized in that, include: The acquired historical charging order data of users is preprocessed by a data preprocessing module deployed on a cloud server to obtain preprocessed training data. The FSE-OPL-DenseNet model is trained using the model training module deployed on the GPU server based on the preprocessed training data, resulting in a trained FSE-OPL-DenseNet model. The system receives user charging requests in real time through a recommendation output module integrated into the mobile or cloud API interface, generates the user's access probability distribution for each candidate charging interest site based on the trained FSE-OPL-DenseNet model, and generates a recommended list of charging interest sites based on the access probability distribution.

Images