A neural network-based charging pile intelligent guiding system and method

Abstract

The application relates to the technical field of charging pile guiding, in particular to a charging pile intelligent guiding system and method based on a neural network, which standardizes and generates structured input by fusing charging pile states, user positions, weather, regional events and time period multi-source information through an Internet of Things data acquisition module, and improves system environment perception capability; a high-dimensional space-time feature vector is constructed by using a feature processing module, and feature expression effect is enhanced; a time sequence prediction model is established based on a long short-term memory network, future available charging pile numbers are predicted, and available charging pile number prediction is enhanced. A navigation guiding strategy model is constructed by introducing a reinforcement learning algorithm, user successful charging and charging pile utilization rate are taken as reward bases, guiding strategies are dynamically optimized, navigation time and predicted resource states are comprehensively considered, intelligent guiding and resource reasonable allocation are realized, and user experience and overall scheduling efficiency are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of charging pile guiding, in particular to a charging pile intelligent guiding system and method based on a neural network. BACKGROUND

[0002] With the rapid popularization of new energy vehicles, as the core supporting infrastructure, the importance of the layout optimization and guiding service of charging piles is increasingly prominent. At present, a large number of public charging pile sites are widely deployed in cities, but due to uneven regional distribution, concentrated use peak, insufficient real-time scheduling capability and other problems, users often face the dilemma of charging pile resource shortage, long queuing time, no available pile after arrival and other problems in actual use. This not only affects the user experience, but also leads to the problem of idle or overloaded resources of some charging piles due to the difference in use frequency, resulting in the decline of energy utilization efficiency.

[0003] The traditional charging guiding system mostly adopts a static matching strategy based on the current state, which only recommends the target charging station according to the user's current location, straight-line distance or historical use data, and cannot adapt to dynamic factors such as complex and changeable urban traffic, weather, regional events and user concentrated travel behavior in real time. In addition, the existing system lacks effective prediction ability of future charging pile use trend, and it is difficult to make accurate judgment on the availability when the user arrives, thereby leading to the invalidation of the guiding strategy, and leading to the use of some charging stations crowded and the utilization rate of some charging stations extremely low. SUMMARY

[0004] To solve the above problems, the present application provides a charging pile intelligent guiding system and method based on a neural network, which predicts the available number of charging piles by constructing a time series prediction model with multi-dimensional data, and realizes effective guiding of charging piles and improves the utilization rate of charging piles by combining reinforcement learning.

[0005] To achieve the above purpose, the technical solution adopted by the present application is:

[0006] A charging pile intelligent guiding system based on a neural network, comprising: an Internet of Things data acquisition module, a feature processing module, a charging pile state prediction module and a navigation guiding decision module;

[0007] The Internet of Things data acquisition module is used to acquire charging pile use state data, user location information, weather information, regional event information and time period information, and to perform standardized processing to generate data input of a unified structure;

[0008] The feature processing module is used to perform feature extraction and fusion processing on the data input of a unified structure to generate an output multi-dimensional spatio-temporal feature vector;

[0009] The charging pile state prediction module is configured to model the space-time feature vector output by the feature processing module based on a time sequence neural network, and construct a prediction model of the number of available charging piles.

[0010] The navigation guidance decision module is configured to calculate the estimated arrival time of each charging station from the current location of the user, calculate the number of available charging piles of each charging station within the preset range of the user by the prediction model of the number of available charging piles, and determine the guidance charging station based on the number of available charging piles of each charging station by a pre-trained reinforcement learning model, wherein the reward function of the reinforcement learning model is constructed by the success of charging of the user and the utilization rate of the charging pile.

[0011] Further, the collection of charging pile usage state data, user location information, weather information, regional event information and time period information includes the following steps:

[0012] The usage state data of each charging pile is collected in real time from the charging station terminal device, including the current idle state, occupancy time, number of queued vehicles and device running state;

[0013] The current location coordinate information of the user is obtained through the mobile terminal, and the region where the user is located is coded in combination with the geographic information system;

[0014] The real-time weather information of the region where the user is located is obtained by calling the weather service interface, including temperature, precipitation, wind force level and weather condition type;

[0015] The regional event information of traffic control, performance, assembly and vehicle behavior in a specific region within the prediction period is obtained from the city event information platform;

[0016] The time period information of the hour, whether it is a working day or not, and whether it is a holiday corresponding to the time of the short message request is extracted according to the server system time.

[0017] Further, the standardization processing to generate data input of uniform structure includes the following steps:

[0018] The missing items of the collected charging pile usage state data are filled and the abnormal values are removed, and the normalized processing is performed to generate the charging pile state feature vector;

[0019] The location feature vector is generated according to the latitude and longitude of the user location information and according to the regional numbering of the geographic information system;

[0020] The temperature, precipitation and wind force level in the weather information are numerically coded, and the weather condition type is one-hot coded to generate the weather feature vector;

[0021] The event type recognition and regional association mapping are performed on the regional event information, binary identification is performed on the traffic control, performance and assembly related to the region where the user is located within the prediction period, and an event feature vector is generated;

[0022] Periodic encoding is performed on the time period information, including hour number sine and cosine mapping, workday and holiday classification identification, and a time feature vector is generated;

[0023] The charging pile state feature vector, the location feature vector, the weather feature vector, the event feature vector and the time feature vector are spliced in a preset order to construct a unified structure data input.

[0024] Further, the feature extraction and fusion processing of the unified structure data input includes the following steps:

[0025] The unified structure data input is input as a high-dimensional vector as a whole to an embedding layer, and is mapped to a continuous feature space through embedding transformation to obtain an initial embedding feature representation;

[0026] The initial embedding feature representation is constructed into a time series data segment using a sliding time window mechanism, and a time series relationship of data evolution over time is established;

[0027] The time series data segment is input to a multi-layer perception network, and a deep feature vector is extracted using a nonlinear activation function, and a residual connection mechanism is used to maintain the stability between the original features and the high-order features;

[0028] The deep feature vector is normalized and compressed to generate a unified dimension fusion representation vector;

[0029] The fusion representation vector is output as a multi-dimensional spatio-temporal feature vector.

[0030] Further, the modeling of the spatio-temporal feature vector output by the feature processing module based on the time series neural network includes the following steps:

[0031] The multi-dimensional spatio-temporal feature vector is arranged in time sequence to construct a sliding time window sequence for representing the state change characteristics in continuous time periods;

[0032] The time window sequence is input to a long short-term memory network, and a time-dependent feature is extracted using its time gate mechanism to obtain a hidden state vector corresponding to each time step;

[0033] The hidden state vector is subjected to attention weighting processing to generate a context-related global semantic representation;

[0034] input the global semantic representation into a fully connected neural network to output a predicted value of the number of available charging piles of each candidate charging station at the target time point.

[0035] Further, the calculation of the estimated arrival time of the user to each charging station according to the current location of the user includes:

[0036] obtaining the latitude and longitude coordinate information of the current location of the user and the departure time of the user, and obtaining the geographic location information of each charging station within the target range;

[0037] calling a navigation service API to obtain the required driving time corresponding to the optimal navigation path from the current location of the user to each candidate charging station;

[0038] adding the required driving time and the departure time of the user to obtain the estimated arrival time of the user to each charging station.

[0039] Further, the construction of the reinforcement learning model includes the following steps:

[0040] establishing a state space composed of the current location of the user, the estimated number of available charging piles of each charging station, the navigation time, the queuing state and the historical utilization rate to represent the current environment state;

[0041] defining the action space as a plurality of candidate charging stations that can be recommended by the system to the user, and each action corresponds to a guidance strategy for guiding the user to one of the charging stations;

[0042] constructing a reward function, wherein if the user successfully starts charging after arriving at the target charging station, a positive reward value is given, and if the user fails to complete charging due to the charging pile being full, the queuing timeout or the navigation failure, a negative reward value is given, and the reward value is adjusted in combination with the actual charging pile utilization rate change of the target charging station after being recommended;

[0043] using a reinforcement learning algorithm based on a policy optimization objective function to train the policy network, selecting the optimal action according to the current state, and comparing with the actual feedback reward value, updating the model parameters using the back propagation algorithm until converging to a stable state, obtaining the trained reinforcement learning model.

[0044] Further, the policy optimization objective function of the reinforcement learning model is as follows:

[0045] ;

[0046] wherein, is the gradient of the objective function of the policy parameter θ; is the policy probability distribution of taking action under state , and the parameter is θ; is the advantage function, used to define the state and action advantage value relative to the average behavior under the current policy; is the gradient of the log probability of the policy network with respect to the parameters, used for backpropagation to update the policy network; is the expectation of the joint distribution of the state and action.

[0047] Further, the reward function of the reinforcement learning model is as follows:

[0048] ;

[0049] wherein, is the immediate reward value at time step t; is an indicator function of whether the user successfully completes charging at time step t, taking the value 1 when successful and 0 when failed; is the actual charging pile utilization rate of the target charging station to which the user is guided at time step t; is the average charging pile utilization rate of all candidate charging stations of the system at time step t; is a stability constant to prevent the denominator from being zero; is the time or distance cost of the user's navigation path; is the value of the improvement of the regional charging resource scheduling efficiency due to the current guidance policy; 、 、 and are the charging success incentive coefficient, the utilization rate optimization coefficient, the navigation cost penalty coefficient, and the resource scheduling efficiency reward coefficient, respectively.

[0050] A neural network-based intelligent charging pile guidance method, applied to the neural network-based intelligent charging pile guidance system of any one of the preceding items, comprising the following steps:

[0051] The acquisition module is used to collect charging pile usage state data, user location information, weather information, regional event information, and time period information, and to perform standardization processing to generate data input of a unified structure;

[0052] The unified structure data input is subjected to feature extraction and fusion processing to generate an output multi-dimensional spatio-temporal feature vector;

[0053] The spatio-temporal feature vector output by the feature processing module is modeled based on a time series neural network to construct a usable charging pile number prediction model;

[0054] The expected arrival time of the user to each charging station is calculated according to the current position of the user, the future available charging pile number of each charging station within the preset range of the user is calculated through a prediction model of the available charging pile number, and a guided charging station is determined according to the future available charging pile number of each charging station based on a pre-trained reinforcement learning model, wherein a reward function of the reinforcement learning model is constructed by the successful charging condition of the user and the charging pile utilization rate.

[0055] The application has the advantages that: the application fuses real-time charging pile state, user position, weather change, regional event and time period and the like multi-source dynamic information through the Internet of Things data acquisition module, and generates standardized structured input data, thereby ensuring that the system comprehensively perceives the current environment state; the feature processing module is used to construct a high-dimensional space-time feature vector fusing position, time and environment, thereby improving the accuracy and robustness of feature expression; further, a long short-term memory network (LSTM) based time sequence modeling mechanism is introduced to predict the available condition of the charging pile in the future period, thereby making up for the lack of foresight of the traditional system. Meanwhile, a reinforcement learning algorithm is introduced to construct a navigation guide strategy model, the successful charging condition of the user and the charging pile utilization rate are used as the design basis of the reward function, and the dynamic optimization of the guide strategy is realized through continuous interactive learning. The reinforcement learning model not only considers the time consumption of the user arrival path, but also takes the predicted resource state at the arrival time into account, thereby ensuring that the system reasonably allocates charging demand under the condition of limited resources and avoiding user congestion and resource waste. Through the synergistic effect of the above multi-layer neural network modeling and strategy optimization mechanism, the charging pile guide system realizes the intelligent transformation from static recommendation to prediction guidance to adaptive scheduling, thereby effectively improving the user charging experience and overall resource allocation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0056] Fig. 1 Fig. 1 is a structural schematic diagram of a neural network-based charging pile intelligent guide system in the application.

[0057] Fig. 2 Fig. 4 is a step flowchart of the construction of a reinforcement learning model in the application. DETAILED DESCRIPTION

[0058] Please refer to Figs. 1-2 Fig. 1 is a structural schematic diagram of a neural network-based charging pile intelligent guide system in the application, which comprises an Internet of Things data acquisition module, a feature processing module, a charging pile state prediction module and a navigation guide decision module.

[0059] The Internet of Things data acquisition module is used to acquire charging pile use state data, user position information, weather information, regional event information and time period information, and perform standardized processing to generate data input of a unified structure.

[0060] The feature processing module is configured to perform feature extraction and fusion processing on the data input of the unified structure to generate an output multidimensional spatiotemporal feature vector.

[0061] The charging pile state prediction module is configured to model the spatiotemporal feature vector output by the feature processing module based on a time series neural network to construct an available charging pile number prediction model.

[0062] The navigation guidance decision module is configured to calculate the estimated arrival time of each charging station from the current location of the user, calculate the future available charging pile number of each charging station within the preset range of the user through the available charging pile number prediction model, and determine the guidance charging station based on the future available charging pile number of each charging station according to a pre-trained reinforcement learning model, wherein the reward function of the reinforcement learning model is constructed by the successful charging condition of the user and the charging pile utilization rate.

[0063] In some embodiments, first, the Internet of Things data acquisition module collects charging pile usage state data, user location information, weather information, regional event information and time period information in real time by accessing charging station device terminals, user mobile positioning systems, third-party weather interfaces and city event platforms. The collected data is standardized, including normalization and missing data filling for numerical data, one-hot encoding for discrete data, and periodic mapping for time data (including using sine and cosine functions to encode hour features) to generate structured data tensors in a unified format that can be used as input for neural networks. The feature processing module receives the above data input and jointly models the multi-source heterogeneous features through a multi-channel embedding mechanism and a deep neural network. Specifically, discrete variables are first mapped to a low-dimensional continuous vector space through an embedding layer, and then concatenated with continuous variables to form an initial feature representation; then a sliding window strategy is used to generate fixed-length time series samples, forming an input sequence in the time dimension. This sequence is input into a multi-layer perceptron (MLP), which extracts and fuses deep representations through multiple layers of nonlinear transformations and residual connection structures, outputs high-order spatio-temporal feature vectors with consistent dimensions, and serves as input for the subsequent prediction module. The charging pile state prediction module uses a long short-term memory network (LSTM) to model the time series of spatio-temporal feature vectors, capturing the long-term dependencies of charging pile usage states over time. In the LSTM calculation, the hidden state update includes the joint control of the input gate, the forget gate and the output gate, which can effectively alleviate the gradient vanishing problem and improve the modeling ability of long-time dependent information. To enhance the model's ability to perceive key time segments, an attention mechanism is introduced to assign attention weights to the hidden states of each time step, extracting a global context-related feature vector. This vector is regressed through a fully connected neural network to obtain the predicted number of available charging piles at each charging station time point, achieving parallel modeling and output of multiple target point future states. The navigation guidance decision module is based on the available charging pile prediction results and user navigation information to build a reinforcement learning-driven guidance strategy optimization model. First, the navigation path calculation module obtains the estimated arrival time of the user from the current location to each target charging station, which is used as a query index to obtain the predicted number of piles at the corresponding time point. The state space of the guidance strategy includes user current location encoding, navigation time consumption, future available state of target pile, historical site congestion level, etc.; the action space is the set of candidate charging stations that the system can recommend at the current time step.

[0064] Further, the collection of charging pile usage state data, user location information, weather information, regional event information and time period information includes the following steps:

[0065] Collecting the use state data of each charging pile from the charging station terminal equipment in real time, including current idle state, occupancy duration, number of queued vehicles and equipment running state;

[0066] Obtaining the current position coordinate information of the user through the mobile terminal, and encoding the region where the user is located in combination with the geographic information system;

[0067] Calling the weather service interface to obtain the real-time weather information of the region where the user is located, including temperature, precipitation, wind force level and weather condition type;

[0068] Obtaining the regional event information of traffic control, performance, assembly and vehicle behavior in the specific region within the predicted period from the city event information platform;

[0069] According to the server system time, extracting the time period information of the hour period, whether it is a workday or not and whether it is a holiday corresponding to the short message request time.

[0070] It should be noted that the collection of charging pile usage state data relies on the edge gateway devices deployed at each charging site. The devices periodically (e.g., every 30 seconds) synchronize data with the server and upload indicators including the following: current status flag (idle, occupied, fault), occupied time for this charging session (in seconds), cumulative number of queued vehicles, device operating status (e.g., online, offline, fault code), etc. For abnormal states, such as occupied but no current change, long time without update, etc., the system triggers threshold judgment strategy to identify invalid states through data continuity rules within the set time window and performs data rejection or interpolation completion to improve training sample quality and prediction stability. User location information is obtained by the mobile terminal of the access client through the Global Navigation Satellite System (GNSS) module, collecting latitude, longitude and elevation data, and combining with network base station assisted positioning (Assisted GPS, A-GPS) to improve the location accuracy in complex urban environments. After obtaining the coordinate information, the system accesses the Geographic Information System (GIS) to complete the spatial coding of the location. To enhance the expression ability of spatial information in the neural network, the spatial coordinates are mapped to unique regional identifiers based on the GeoHash or S2 hierarchical grid partitioning algorithm, and the appropriate precision (such as 6-level or 7-level coding) is set according to the regional level, so as to perform regional feature aggregation and adjacency relationship modeling subsequently. Weather information is obtained by calling third-party meteorological data interfaces (such as National Meteorological Bureau API or commercial platforms such as and Wind Weather, OpenWeather, etc.), covering the current user's area instant weather state, including temperature (unit: °C), precipitation (unit: mm / h), wind speed (unit: m / s), wind force level (according to meteorological standard 0 to 12 levels) and weather phenomenon (such as sunny, cloudy, light rain, snowstorm, etc.). Among them, the numerical variables are directly used as model inputs after standardization; the category variables are converted into sparse vectors using One-Hot Encoding. To ensure the timeliness of the information, the system uses a cache mechanism for weather data and sets the minimum polling interval, such as synchronizing every 5 minutes, to improve the prediction accuracy. The collection of regional event information is based on the interface connection with the government data platform or the city traffic event center to obtain structured event information stream. Event data includes event type (such as traffic accident, large-scale performance, road regulation), event occurrence time, expected end time, event center point coordinates and event influence radius, etc. The system determines whether the user's current location falls within the event influence area based on the spatial calculation module (such as using the spherical distance algorithm or R-tree index), and if the spatial constraint and time overlap conditions are met, the corresponding event label is added for the user. In feature construction, the event type is enumerated and represented, and the spatial relationship between the event occurrence state and the user's location is encoded into a Boolean regional event feature input.The time period information is extracted based on the server timestamp, and the extracted items include the number of hours (0-23), the day of the week (0-6), whether it is a weekday (Boolean value), whether it is a holiday (Boolean value), etc. Finally, the five types of information are converted into vector form after the above processing, and are combined into a unified structure input tensor in a set order during the feature splicing stage, serving as the input data of the feature processing module. This structured input fully retains the spatial-temporal context information of user behavior, external disturbance factors, and resource state evolution trends, providing a foundation guarantee for subsequent neural networks in terms of prediction accuracy and policy robustness.

[0071] Further, the standardized processing to generate the data input of the unified structure includes the following steps:

[0072] The collected charging pile usage state data is filled in missing items and outliers are removed, and normalized to generate a charging pile state feature vector;

[0073] According to the latitude and longitude of the user location information, and according to the geographic information system, a location feature vector is generated;

[0074] The temperature, precipitation, and wind force level in the weather information are numerically encoded, and the weather condition type is one-hot encoded to generate a weather feature vector;

[0075] The regional event information is identified by event type and region association mapping, and the traffic control, performance, and assembly related to the user's region within the prediction period are identified by binary, to generate an event feature vector;

[0076] The time period information is periodically encoded, including hour sine and cosine mapping, weekday and holiday classification identification, to generate a time feature vector;

[0077] The charging pile state feature vector, location feature vector, weather feature vector, event feature vector, and time feature vector are spliced in a predetermined order to construct a data input of a unified structure.

[0078] In some embodiments, first, for charging pile usage state data, the system performs integrity verification and anomaly detection on its collected values. For missing fields, the mean or median of historical data at the same site within a sliding time window is used for interpolation to improve feature continuity; for abnormal values (such as occupancy time exceeding the maximum allowable upper limit, abnormal sudden increase in the number of queued vehicles, etc.), the system automatically removes outliers based on the box plot method (IQR) or 3σ criterion to ensure the stability of the feature distribution. After cleaning, the numerical fields (such as occupancy time, number of queues, etc.) are mapped to the [0, 1] interval using the Min-Max Scaling method, and the standardized results are used as the charging pile state feature vector input. For user location data, the latitude and longitude coordinates obtained by the system are mapped through the GIS spatial mapping service, and based on the GeoHash or S2 cell encoding algorithm, the continuous geographic location is quantized into discrete region numbers, and further encoded into numerical features. In order to enhance the spatial representation capability, region embedding vectors or region adjacency graphs can be introduced based on the encoding as input structure reference for subsequent graph neural networks or spatio-temporal networks. In this embodiment, the sparse one-hot vector of region number is used as the location feature representation to ensure the uniformity of the input structure. For weather information processing, the system distinguishes between continuous and discrete variables for modeling. Temperature, precipitation, and wind grade are continuous variables, which are standardized. Wind grade can be considered as an ordered category for numerical mapping because it has a fixed interval. Weather condition types (such as "sunny", "moderate rain", "heavy snow", etc.) are treated as unordered categories and are converted into sparse vectors using one-hot encoding, and are then concatenated into the weather feature vector. The standardization process of regional event information depends on the event recognition model and spatial correlation rules. The system first classifies events based on keyword recognition, event labels, and location coordinates (such as performances, traffic controls, exhibitions, and gatherings, etc.), and uses the spatial overlap relationship between the radius buffer and the user's location to determine whether the event is affected. Each type of event is defined as a binary label, forming a sparse event vector, representing the possibility of the user's region being affected by the event in the current prediction period, enhancing the model's ability to perceive sudden factors. Time period information processing is based on periodic mapping and semantic hierarchical rules. The system extracts the current hour, day of the week, and holiday flag from the timestamp. The hour information is periodically encoded using the sine and cosine functions. This encoding can effectively enhance the model's ability to model the fluctuations in charging demand within a 24-hour period. Holiday and weekday information is determined using a pre-defined calendar API, and a Boolean value is used to represent the holiday status, forming a time feature vector. After the standardization and encoding of the above-mentioned features, the system combines the five types of feature vectors (charging pile state, location, weather, event, and time) in the predetermined concatenation order to construct a unified input feature tensor.

[0079] Further, the feature extraction and fusion processing of the unified structure data input comprises the following steps:

[0080] The unified structure data input is input as a high-dimensional vector as a whole to an embedding layer, mapped to a continuous feature space through embedding transformation, and an initial embedding feature representation is obtained;

[0081] The initial embedding feature representation is adopted to construct a time sequence data segment using a sliding time window mechanism, and a time sequence relationship of data evolution over time is established;

[0082] The time sequence data segment is input to a multi-layer perception network, a deep feature vector is extracted using a nonlinear activation function, and a stability between original features and high-order features is maintained through a residual connection mechanism;

[0083] The deep feature vector is normalized and compressed to generate a dimension-unified fusion representation vector;

[0084] The fusion representation vector is output as a multi-dimensional spatiotemporal feature vector.

[0085] In some embodiments, first, the standardized unified structure input data is input as a whole high-dimensional vector to the embedding layer for encoding transformation. Since the input vector contains multiple discrete features such as position encoding, weather category, event type, time mark, etc., to avoid the dimension expansion and sparsity problem caused by one-hot encoding, the system introduces a trainable embedding matrix to map each discrete feature type in this stage. For example, for area number encoding, the embedding layer can map it to a dense embedding vector through a lookup table. After concatenating multiple embedding vectors with continuous features, an initial feature representation that is semantically unified and numerically continuous is formed. After obtaining the initial embedding representation, in order to introduce time series dependency and state evolution relationship, the system constructs a sliding time window mechanism to input the features of continuous time steps in the historical time period into a time series data segment. This construction method can explicitly model the dynamic trend of input variables in multiple time steps, which is beneficial to the subsequent network to identify periodic behavior or mutation trend. Then, the time series segment is input to a multi-layer perceptron (MLP) network for feature extraction. The MLP network contains an alternating stack structure of multiple linear layers and nonlinear activation functions (such as ReLU, GELU or Swish), which can capture high-order nonlinear combination relationships between original feature vectors. In the deep structure, a residual connection mechanism (Residual Connection) is introduced, that is, the low-layer output is directly added to the high-layer input. In order to ensure the consistency of feature dimensions between different input samples and improve the numerical stability during neural network training, the system introduces a normalization operation such as batch normalization (Batch Normalization) or layer normalization (Layer Normalization) after feature extraction to uniformly adjust the distribution of different batch samples. After normalization, the high-dimensional feature vector is compressed to a specified dimension through a linear transformation layer to generate a unified length fusion representation vector. Finally, the fusion representation vector, as one of the core modeling results of the system, is defined as a multi-dimensional spatio-temporal feature vector, which is used as input for the subsequent charging pile state prediction model. This vector not only contains data embedding from space, time, environment and user behavior, but also integrates high-order feature transformation information across time series, has strong expression ability, and can effectively drive high-precision modeling of charging pile dynamic resource state in the subsequent model.

[0086] Further, the time series neural network based on the spatio-temporal feature vector output by the feature processing module includes the following steps:

[0087] The multi-dimensional spatio-temporal feature vector is arranged in time sequence to construct a sliding time window sequence for representing state change characteristics in a continuous time period;

[0088] input the time window sequence to a long short-term memory network, extract time-dependent features by using a time gating mechanism thereof, and obtain a hidden state vector corresponding to each time step;

[0089] perform attention weighting processing on the hidden state vector to generate a context-related global semantic representation;

[0090] input the global semantic representation to a fully connected neural network, and output a predicted value of the number of available charging piles of each candidate charging station at a target time point.

[0091] Specifically, a time series neural network is introduced to model the multi-dimensional spatio-temporal feature vector output by the feature processing module, and a prediction model for the number of available charging piles in a future period is constructed. Specifically, the model first organizes the input spatio-temporal feature vector in chronological order, constructs continuous time segments through a sliding time window mechanism, and thus forms an input sequence with a time series structure to express the dynamic changes of charging pile-related states, environmental factors, and user behaviors, etc. in the past period. The sliding window mechanism is used to extract continuous feature vectors within a fixed length of time span, and to generate a context information segment for each time point on which the current prediction depends. Each time segment contains feature representations corresponding to multiple historical time points, which constitute the input basis required for prediction modeling. On the basis of the constructed time series structure, a long short-term memory network (LSTM) is introduced as the core modeling structure, which uses its time gating mechanism to handle long-term dependencies and short-term mutations in the input sequence. This neural network structure dynamically selects and updates the input features of each time step in the sequence through memory cells and multiple gating mechanisms, which can effectively preserve the time series of charging pile state change trends and external influencing factors (such as weather or events). While processing each time segment, the system further introduces an attention mechanism to weight and sum the hidden states of different time steps to obtain a global context feature representation for the current prediction task. This mechanism automatically adjusts the weight of each time point in the final semantic representation according to its influence degree in the sequence, thereby improving the model's ability to perceive key time nodes, especially when facing input sequences with non-stationary or sudden disturbances, the model has stronger adaptability.

[0092] To address the problem of decreased prediction accuracy as the number of users increases, the system can introduce a user-guided intervention feedback mechanism and a dynamic distributed state correction strategy in the time series modeling process, and embed a system-level influence modeling factor in the model architecture to enhance the model's sensitivity and adaptability to changes in group behavior, thereby effectively alleviating the prediction bias caused by the guidance strategy itself.

[0093] Specifically, in constructing the time window sliding input data, the system not only contains the state change of the charging pile at the historical moment, the external environmental factors and the individual user behavior characteristics, but also synchronously introduces the "number of guided users" which reflects the global characteristics of the system intervention intensity. This feature is dynamically generated by the navigation guidance decision module, records the number of users who accept system guidance in each time period, as a time series variable and other features are input into the LSTM network, helping the model to capture the resource distribution trend caused by the guidance behavior when modeling. For example, when a certain charging station is continuously guided by the system to a large number of users, its short-term availability may quickly decrease, and if the prediction model cannot timely perceive the guidance intensity, it will cause the problem of systematic over-recommendation. In order to further improve the convergence stability of the prediction model, the system can also introduce a periodic retraining and fine-tuning mechanism, that is, to feedback sample the actual guidance results at a fixed period (such as every 2 hours), and calculate the error between the predicted value and the actual available pile number at the arrival time. When the error exceeds the threshold, trigger the local model update process, and preferentially update the parameter subset representing the influence of user-intensive guidance behavior, to improve the modeling response speed to crowded dynamics. At the same time, a penalty term is introduced during the model training stage, which models the coupling degree of prediction and actual error with user guidance density, so that the model has a certain "anti-regulation ability", that is, actively avoids the risk of hot resource aggregation caused by system guidance in prediction. In addition, the system adds an "estimation factor of guidance influence" in the semantic weighting process of the attention mechanism, which adjusts the attention weight of the time step by coupling with the guidance user density in the historical time period, so that the model tends to learn the resource change rule from the historical behavior under the low guidance interference state, avoiding the erosion of guidance behavior interference on the long-term semantic expression of the model.

[0094] Through the above structural and algorithmic improvements, the embodiment not only constructs a prediction framework with time-dependent modeling capability, but also solves the prediction deviation problem caused by large-scale system guidance behavior, ensuring that the prediction of available charging piles can still maintain high accuracy in the scenario of growing user numbers, significantly improving the stability, reliability and robustness of the model facing group interference.

[0095] Further, the calculating the estimated arrival time of the user to each charging station according to the current location of the user comprises:

[0096] obtaining the latitude and longitude coordinate information of the current location of the user and the departure time of the user, and obtaining the geographic position information of each charging station in the target range;

[0097] calling a navigation service API to obtain the driving time required for the optimal navigation path from the current location of the user to each candidate charging station;

[0098] The required driving time is added to the user departure time to obtain the user's estimated arrival time at each charging station.

[0099] Specifically, first, the system obtains the user's current accurate geographic location through the mobile terminal, extracts the longitude and latitude coordinates thereof, and records the system timestamp when the user triggers the guidance request as the actual departure time. In order to improve the spatial consistency of the location information, the system converts and regionally encodes the static geographic coordinates of all charging stations into a unified coordinate system, constructs a spatial index structure of the target station, so as to quickly filter out a candidate charging station set within the user-specified navigation radius. Subsequently, the system calls a high-precision navigation service API (such as an online map service path calculation interface optimized based on A* or Dijkstra algorithm), takes the user's current location as the starting point, and sequentially constructs a shortest path or optimal path query request for each candidate charging station. This path planning process is not only based on static map topological data, but also integrates dynamic information such as road traffic state, real-time traffic congestion condition, and traffic restriction rules, to obtain the optimal driving path and the corresponding required time under the current period. In order to improve the concurrent performance of path planning, the system uses an asynchronous batch request mechanism to submit all candidate paths to the navigation server at one time, and receives the optimal time results of each path through callback, effectively reducing the API response delay.

[0100] Further, the construction of the reinforcement learning model includes the following steps:

[0101] A state space composed of the user's current location, the number of predicted available charging piles of each charging station, the navigation time, the queuing state, and the historical utilization rate is established to represent the current environment state.

[0102] The action space is defined as a plurality of candidate charging stations that the system can recommend to the user, and each action corresponds to a guidance strategy for guiding the user to one of the charging stations.

[0103] A reward function is constructed, wherein if the user successfully starts charging after arriving at the target charging station, a positive reward value is given; if the user fails to complete charging due to the charging pile being full, the queuing timeout, or the navigation failure, a negative reward value is given, and the reward value is adjusted in combination with the actual charging pile utilization rate change of the target charging station after being recommended;

[0104] A reinforcement learning algorithm based on a policy optimization objective function is used to train the policy network, the optimal action is selected according to the current state, and the model parameters are updated using the backpropagation algorithm by comparing with the actual feedback reward value until convergence to a stable state, thereby obtaining the trained reinforcement learning model.

[0105] Specifically, first, the system establishes a state space to represent the current environment of the user in multi-dimensional features. The state not only includes the real-time geographic location of the user, but also combines the predicted available charging pile number of each candidate charging station at the time when the user is expected to arrive, the predicted driving time from the current location of the user to the charging station, the current queuing condition of the charging station, the historical utilization rate and other dynamic and static information. These features are uniformly encoded to form a state vector, which is input into the policy model to represent the current decision-making background. Correspondingly, the action space is defined as a number of candidate charging stations that can be recommended by the system at the current time, and each action is a specific strategy to guide the user to go to a certain charging station. The system uses a neural network to model the strategy function, outputs the probability distribution of each action according to the current state, and selects the target charging station with the highest probability or performs the strategy by sampling. In reinforcement learning, the design of the reward function is particularly important, and it is used as the core evaluation index for guidance optimization in this embodiment. Specifically, when the user goes to a certain charging station according to the system recommendation and successfully completes the charging operation in a short time, it is considered as an effective guidance, and the system gives a positive reward; if the target charging pile resource is insufficient, the queuing time is too long or the navigation path is interrupted, resulting in guidance failure, a negative reward is applied to punish the bad strategy. At the same time, after the guidance is successful, the system will also evaluate the resource utilization rate trend of the site, and if the reasonable transfer of resource load is realized due to the guidance, an additional reward is given to encourage the strategy model to learn the behavior of promoting the overall resource balance of the system. In order to realize the training of the strategy model, the system uses a policy gradient-based optimization algorithm, combines the feedback reward generated by the actual guidance behavior, and continuously adjusts the parameters of the strategy network, so that in a similar state, the charging station guidance strategy can be selected more optimally in the future. The training process is carried out in multiple rounds of simulated interaction, and each round of guidance behavior forms a state-action-reward sequence. The system evaluates the pros and cons of the current strategy according to the cumulative return, and optimizes the neural network parameters through the back propagation mechanism, so that the strategy gradually converges to the optimal. In addition, in order to solve the guidance interference problem caused by the increase in the number of users, the system introduces a system behavior influence factor, which is incorporated into the recent number of users who accept the system guidance as a global feature in the state modeling stage, so that the model can perceive the disturbance of the guidance density to the resource distribution, thereby avoiding excessive concentration of recommending a popular site. In addition, the system also combines the real-time feedback mechanism and the regular fine-tuning mechanism to incrementally update the strategy model, so that it can adapt to the user behavior patterns and environmental changes in the long term.

[0106] Further, the policy optimization objective function of the reinforcement learning model is as follows:

[0107] ;

[0108] wherein, is the gradient of the objective function of the policy parameter θ; is the state Take action the policy probability distribution, parameterized by θ; is the advantage function, used to define the advantage value of an action under the current policy relative to the average behavior; is the advantage function, used to define the advantage value of an action under the current policy relative to the average behavior; is the gradient of the log probability of the policy network with respect to the parameters, used for backpropagation to update the policy network;

[0109] Specifically, the core of the construction of the objective function lies in the quantification of the pros and cons of policy behavior through the introduction of the advantage function, and the weighting of the log probability gradient of the policy network, so as to guide the iterative update of the policy network parameters in the direction most conducive to improving the long-term return, and realize the effective learning and optimization of the guiding policy. Specifically, the policy network outputs a probability distribution under any given state, which represents the system's selection preference for different guiding actions. For example, when the user is in a certain geographical location, the system needs to select the optimal guiding target among multiple charging stations, and at this time the policy network outputs the selection probability of each target station according to the state input. The optimization objective function measures the degree of response of the current policy to a specific action in that state by taking the derivative of the log probability function of the policy. In order to determine whether the response is worth reinforcing, the system further introduces the advantage function to evaluate the relative advantage of the action in the current state compared to the average policy behavior. The calculation of the advantage function is based on the return evaluation mechanism, that is, the actual or estimated long-term revenue brought by a certain action is observed after its execution, and this revenue is compared with the expected return of all possible behaviors in that state. If the effect brought by the action is better than the average, the advantage function is positive; if it is lower than the average, it is negative. The optimization objective function uses the advantage function as a weight item to act on the log probability gradient to form the parameter update direction. In this way, the policy network will continuously increase the selection tendency of high-advantage actions and suppress low-advantage actions during the training process, thereby achieving the improvement of the overall performance of the policy. In addition, the system uses sample sampling and expectation approximation methods in the process of implementing the objective function, and uses the state-action pairs obtained by batch sampling to construct the empirical estimate of the objective function, and updates the parameters of the policy network through the backpropagation mechanism. This optimization process has the characteristics of fast convergence speed and strong stability of policy convergence, and is particularly suitable for complex guiding decision-making tasks in high-dimensional and variable environments. Through the design and implementation of the above policy optimization objective function, the system not only can realize reasonable guidance to users in static state, but more importantly, has the ability to continuously learn and adapt in dynamic environment. With the changes in the spatio-temporal distribution of charging demand, the evolution of user behavior patterns, and the dynamic fluctuations of resource state, the system can adjust the guiding policy in real time to ensure the improvement of user success charging rate and the efficient use of overall charging pile resources, thereby realizing a guiding optimization mechanism with high intelligence and robustness.

[0110] Further, the reward function of the reinforcement learning model is as follows:

[0111] ;

[0112] wherein, is the immediate reward value at time step t; is an indicator function of whether the user successfully completes charging at time step t, taking value 1 for success and 0 for failure; is the actual charging pile utilization rate of the target charging station to which the user is guided at time step t; is the average charging pile utilization rate of all candidate charging stations of the system at time step t; is a stability constant to prevent the denominator from being zero; is the time or distance cost of the user navigation path; is the value of the improvement of the regional charging resource scheduling efficiency due to the current guidance strategy; , , and are respectively a charging success incentive coefficient, a utilization rate optimization coefficient, a navigation cost penalty coefficient, and a resource scheduling efficiency reward coefficient.

[0113] Specifically, first, the reward function takes the user's charging success as the core evaluation basis. When the user reaches the target charging station according to the system guidance and successfully completes the charging operation, the system gives a positive reward value to this behavior; if the user fails to charge due to the target station pile being full, queuing timeout or other factors, no reward is given. This part is realized by a Boolean indicator function, and its result directly reflects the basic service effect of the strategy execution, representing the most direct feedback of user satisfaction. Second, in order to guide the system to achieve more reasonable resource allocation, the reward function further introduces the ratio between the utilization rate of the target charging station and the average utilization rate of all candidate stations as an adjustment factor. This ratio is used to measure whether the current guidance behavior helps to balance the resource load distribution in the region. If the target station utilization rate is moderate or relatively low, this guidance behavior helps to alleviate the congestion in hot spot areas or improve the resource utilization efficiency in edge areas; on the contrary, if it is guided to a high load area, the ratio will inhibit the growth of the reward value, so as to encourage the strategy to avoid centralized recommendation. In addition, in order to ensure that the user experience is not damaged due to resource scheduling optimization, the reward function also introduces the navigation path cost as a penalty term. The system will quantify the time or distance of the user's navigation from the current location to the target charging station, and if the path is too long or the expected time is too high, a penalty coefficient will be weighted and deducted in the reward function. This part ensures that the strategy can also consider the user's travel cost and convenience while optimizing the resources. Finally, the system also includes the resource scheduling efficiency improvement brought by the strategy execution at the regional level into the reward function calculation, which is used to quantify the positive contribution of a guidance behavior to the overall regional resource balance. This part is realized by comparing the charging pile load distribution or standby time before and after guidance, and if the strategy effectively reduces the congestion level or improves the resource turnover rate, the system gives additional positive incentives to guide the strategy to optimize the system's overall performance. The above indicators are all set with adjustable weight coefficients in the reward function to adapt to the strategy preferences and operation goals in different cities and different scenarios. Overall, this reward function takes into account the success rate of individual behavior and the efficiency of system resource allocation, has good feedback orientation and optimization driving ability, and provides a stable and business value learning signal source for the effective training of the reinforcement learning strategy model. Through this reward mechanism, the system can ultimately achieve high-robust intelligent guidance capability for complex and variable urban environments, improve user experience and ensure efficient use of public resources.

[0114] The application also includes a neural network-based intelligent charging pile guidance method applied to any of the neural network-based intelligent charging pile guidance systems described above, comprising the following steps:

[0115] The acquisition module is used to collect charging pile usage state data, user location information, weather information, regional event information and time period information, and perform standardization processing to generate data input with a unified structure;

[0116] Feature extraction and fusion processing are performed on the data input of the unified structure to generate an output multidimensional spatiotemporal feature vector;

[0117] A spatiotemporal feature vector output by the feature processing module is modeled based on a time series neural network to construct a prediction model for the number of available charging piles;

[0118] The estimated arrival time of a user to each charging station is calculated according to the current location of the user, the number of available charging piles of each charging station within a preset range of the user is calculated through the prediction model for the number of available charging piles, and a guide charging station is determined according to the number of available charging piles of each charging station in the future based on a pre-trained reinforcement learning model, wherein a reward function of the reinforcement learning model is constructed by the successful charging condition of the user and the charging pile utilization rate.

[0119] The above embodiments merely describe the preferred embodiments of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by ordinary engineering technicians in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A charging pile intelligent guidance system based on neural networks, characterized in that, include: The system includes an IoT data acquisition module, a feature processing module, a charging pile status prediction module, and a navigation guidance decision-making module. The IoT data acquisition module is used to collect charging pile usage status data, user location information, weather information, regional event information and time period information, and perform standardized processing to generate data input with a unified structure. The feature processing module is used to extract and fuse features from the unified structure of the data input to generate a multi-dimensional spatiotemporal feature vector. The charging pile status prediction module is used to model the spatiotemporal feature vector output by the feature processing module based on a time-series neural network, combined with the number of users receiving guidance in each time period, to construct a prediction model for the number of available charging piles. The navigation guidance decision module is used to calculate the estimated arrival time of the user to each charging station based on the user's current location, calculate the number of future available charging stations for each charging station within the user's preset range through the available charging station number prediction model, and determine the guiding charging station based on the number of future available charging stations for each charging station based on the pre-trained reinforcement learning model. The reward function of the reinforcement learning model is constructed by the user's successful charging status and the charging station utilization rate. The available charging pile number prediction model updates locally based on the error between the calculated prediction value and the actual number of available charging piles at the arrival time; The construction of the reinforcement learning model includes the following steps: Establish a state space consisting of the user's current location, the estimated number of available charging piles at each charging station, navigation time, queuing status, and historical utilization rate to represent the current environmental state; The action space is defined as multiple candidate charging stations that the system can recommend to the user, and each action corresponds to a guidance strategy that directs the user to one of the charging stations. Construct a reward function where a positive reward is given if a user successfully starts charging after arriving at the target charging station; a negative reward is given if the user is unable to complete charging due to a full charging station, queuing timeout, or navigation failure. The reward value is adjusted based on the actual change in the utilization rate of the charging stations after the target charging station is recommended. The policy network is trained using a reinforcement learning algorithm based on the policy optimization objective function. The optimal action is selected according to the current state, and the model parameters are updated by comparing the action with the actual reward value. The process continues until the model converges to a stable state, resulting in a trained reinforcement learning model. The collection of charging pile usage status data, user location information, weather information, regional event information, and time period information includes the following steps: Real-time data on the usage status of each charging pile is collected from the terminal equipment of the charging station, including the current idle status, occupancy time, number of vehicles in queue, and equipment operation status. The system obtains the user's current location coordinates through the mobile terminal and encodes the area where the user is located using a geographic information system. Call the weather service interface to obtain real-time weather information for the user's location, including temperature, precipitation, wind speed, and weather condition type; Obtain information from the city event information platform regarding whether there are traffic controls, performances, gatherings, and vehicle activities within a specific area during the predicted time period; Extract the hour segment, weekday / holiday / holiday time period, and other relevant information corresponding to the SMS request time based on the server system time.

2. The intelligent guidance system for charging piles based on neural networks according to claim 1, characterized in that, The standardization process to generate data input with a unified structure includes the following steps: The collected charging pile usage status data is filled with missing items and outliers are removed, and normalization is performed to generate charging pile status feature vectors. Based on the latitude and longitude of the user's location information, and by assigning a region number according to the geographic information system, a location feature vector is generated; Numerical encoding is performed on temperature, precipitation, and wind speed in weather information, and unique thermal encoding is performed on weather condition types to generate weather feature vectors; The system performs event type identification and regional association mapping on regional event information, and assigns binary labels to traffic control, performances and gatherings related to the user's region within the predicted time period to generate event feature vectors. Periodic encoding of time period information, including sine and cosine mapping of hours, and classification of weekdays and holidays, generates time feature vectors; The charging pile status feature vector, location feature vector, weather feature vector, event feature vector, and time feature vector are concatenated in a preset order to construct a unified data input structure.

3. The intelligent guidance system for charging piles based on neural networks according to claim 1, characterized in that, The feature extraction and fusion processing of the data input with the unified structure includes the following steps: The data input of the unified structure is input as a high-dimensional vector into the embedding layer, and mapped to a continuous feature space through embedding transformation to obtain the initial embedded feature representation; The initial embedded feature representation is constructed using a sliding time window mechanism to create time-series data segments and establish the time-series relationship of data evolution over time. The time-series data segment is input into a multilayer perceptron network, and deep feature vectors are extracted using a nonlinear activation function. The stability between the original features and higher-order features is maintained through a residual connection mechanism. The deep feature vectors are normalized and compressed to generate a fusion representation vector with uniform dimensions. The fused representation vector is output as a multi-dimensional spatiotemporal feature vector.

4. The intelligent guidance system for charging piles based on neural networks according to claim 1, characterized in that, The process of modeling the spatiotemporal feature vector output by the feature processing module based on a temporal neural network to construct a prediction model for the number of available charging piles includes the following steps: The multidimensional spatiotemporal feature vectors are arranged in chronological order to construct a sliding time window sequence, which is used to represent the state change features within a continuous time period. The time window sequence is input into a long short-term memory network, and its time gating mechanism is used to extract time-dependent features to obtain the hidden state vector corresponding to each time step. The hidden state vector is subjected to attention weighting to generate a context-dependent global semantic representation; The global semantic representation is input into a fully connected neural network, which outputs a predicted number of available charging piles for each candidate charging station at the target time point.

5. The intelligent guidance system for charging piles based on neural networks according to claim 4, characterized in that, The calculation of the estimated arrival time to each charging station based on the user's current location includes: Obtain the latitude and longitude coordinates of the user's current location and the user's departure time, and obtain the geographical location information of each charging station within the target area; Call the navigation service API to obtain the required travel time for the optimal navigation route from the user's current location to each candidate charging station; The required travel time is added to the user's departure time to obtain the user's estimated arrival time at each charging station.

6. The intelligent guidance system for charging piles based on neural networks according to claim 1, characterized in that, The policy optimization objective function of the reinforcement learning model is as follows: ； in, Let θ be the gradient of the objective function. In the state Take action below The policy probability distribution, with parameter θ; The advantage function is used to define the state. and actions The advantage value relative to the average behavior under the current strategy; This is the gradient of the log probability of the policy network with respect to the parameters, used for backpropagation to update the policy network; To calculate the expectation of the joint distribution of states and actions.

7. The intelligent guidance system for charging piles based on neural networks according to claim 6, characterized in that, The reward function of the reinforcement learning model is as follows: ； in, The instantaneous reward value at time step t; This is an indicator function for the user to successfully complete charging at time step t. The value is 1 for success and 0 for failure. The actual charging station utilization rate at the target charging station guided to the user at time step t; The average charging pile utilization rate of all candidate charging stations in the system at time step t; To prevent the stability constant from having a denominator of zero; The time or distance cost of navigating a user's path; This represents the improvement in regional charging resource scheduling efficiency caused by the current guidance strategy. , , and These are the charging success incentive coefficient, utilization optimization coefficient, navigation cost penalty coefficient, and resource scheduling efficiency reward coefficient, respectively.

8. A neural network-based intelligent guidance method for charging piles, applied to the neural network-based intelligent guidance system for charging piles as described in any one of claims 1-7, characterized in that, Includes the following steps: The acquisition module is used to collect charging pile usage status data, user location information, weather information, regional event information and time period information, and perform standardized processing to generate data input with a unified structure; The data input of the unified structure is subjected to feature extraction and fusion processing to generate a multi-dimensional spatiotemporal feature vector output. A model for predicting the number of available charging piles is constructed by modeling the spatiotemporal feature vectors output by the feature processing module based on a temporal neural network. The estimated arrival time to each charging station is calculated based on the user's current location. The number of available charging piles in each charging station within the user's preset range is calculated using an available charging pile prediction model. Based on a pre-trained reinforcement learning model, the guiding charging station is determined according to the number of available charging piles in each charging station in the future. The reward function of the reinforcement learning model is constructed by combining the user's successful charging status with the utilization rate of the charging piles.

Images