Intelligent charging pile guiding system and method based on neural network

Abstract

The invention relates to the technical field of charging pile guiding, in particular to an intelligent charging pile guiding system and method based on a neural network, and the system integrates the charging pile state, user position, weather, regional event and time period multi-source information through an Internet of Things data collection module, generates structured input in a standardized manner, and improves the environment perception capability of the system. A feature processing module is utilized to construct a high-dimensional spatial-temporal feature vector, and a feature expression effect is enhanced; a time sequence prediction model is established based on a long-short-term memory network, the number of future available charging piles is predicted, and prediction of the available number of the charging piles is enhanced. A reinforcement learning algorithm is introduced to construct a navigation guidance strategy model, successful charging of a user and the utilization rate of a charging pile are taken as reward basis, a guidance strategy is dynamically optimized, navigation time and a prediction resource state are comprehensively considered, intelligent guidance and reasonable resource allocation are realized, and user experience and overall scheduling efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of charging pile guidance, and particularly to an intelligent charging pile guidance system and method based on a neural network. Background Art

[0002] With the rapid popularization of new energy vehicles, charging piles, as their core supporting infrastructure, the importance of their layout optimization and guidance services has become increasingly prominent. At present, a large number of public charging pile stations have been widely deployed in cities. However, due to problems such as uneven regional distribution, concentrated usage peaks, and insufficient real-time scheduling capabilities, users often face difficulties such as tight charging pile resources, long queuing times, and no available piles after arrival in actual use. This not only affects the user experience but also leads to problems such as resource idleness or overload of some charging piles due to differences in usage frequencies, resulting in a decline in energy utilization efficiency.

[0003] Most traditional charging guidance systems mostly adopt static matching strategies based on the current state, and only recommend target charging stations according to the user's current location, straight-line distance, or historical usage data, and cannot adapt to dynamic factors such as complex and changeable urban traffic, weather, regional events, and concentrated user travel behaviors in real time. In addition, the existing systems lack the effective prediction ability for future charging pile usage trends, and it is difficult to make accurate judgments on the availability when users arrive, resulting in the failure of guidance strategies and causing some charging stations to be crowded and some charging stations to have extremely low utilization rates with few people using them. Summary of the Invention

[0004] To solve the above problems, the present invention provides an intelligent charging pile guidance system and method based on a neural network, which constructs a time series prediction model through multi-dimensional data for predicting the available number of charging piles, and combines reinforcement learning to achieve effective guidance of charging piles and improve the utilization rate of charging piles.

[0005] To achieve the above object, the technical solution adopted by the present invention is: An intelligent charging pile guidance system based on a neural network, comprising: an Internet of Things data acquisition module, a feature processing module, a charging pile status prediction module, and a navigation guidance decision module; The Internet of Things 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 standardization processing to generate data input with a unified structure; The feature processing module is used to perform feature extraction and fusion processing on the data input with the unified structure to generate an output multi-dimensional spatio-temporal feature vector; The charging pile status prediction module is used to model the spatio-temporal feature vector output by the feature processing module based on a time series neural network to construct a prediction model for the available number of charging piles; The navigation guidance decision module is used to calculate the estimated arrival time of the user at each charging station according to the user's current location, calculate the future number of available charging piles at each charging station within the user's preset range through the available charging pile number prediction model, and determine the guidance charging station based on the future number of available charging piles at each charging station based on the pre-trained reinforcement learning model. The reward function of the reinforcement learning model is constructed through the user's successful charging situation and the charging pile utilization rate.

[0006] Furthermore, the collection of charging pile usage status data, user location information, weather information, regional event information and time period information includes the following steps: Collect the usage status data of each charging pile in real time from the charging station terminal equipment, including the current idle status, occupancy time, number of queued vehicles and equipment operation status; Obtain the user's current location coordinate information through the mobile terminal, and encode the user's location area in combination with the geographic information system; Call the weather service interface to obtain real-time weather information in the user's area, including temperature, precipitation, wind level and weather condition type; Obtain regional event information from the city event information platform regarding whether there are traffic controls, performances, gatherings, and vehicle behaviors in a specific area during the forecast period; The hour period corresponding to the SMS request time, whether it is a working day or a holiday, is extracted according to the server system time.

[0007] Furthermore, the standardization process to generate data input with a unified structure includes the following steps: Fill in missing items and remove outliers from the collected charging pile usage status data, and perform normalization processing to generate a charging pile status feature vector; According to the longitude and latitude of the user's location information, and according to the area numbering of the geographic information system, a location feature vector is generated; Numerical encoding of temperature, precipitation and wind level in weather information, unique encoding of weather condition type, and generation of weather feature vector; Perform event type identification and regional association mapping on regional event information, identify traffic control, performances and gatherings related to the user's area during the forecast period, and generate event feature vectors; Periodically encode the time period information, including the sine and cosine mapping of hours, the classification of working days and holidays, and generate a time feature vector; The charging pile state feature vector, location feature vector, weather feature vector, event feature vector and time feature vector are spliced ​​in a preset order to construct a data input with a unified structure.

[0008] Further, the feature extraction and fusion processing of the data input to the unified structure includes the following steps: Taking the data input to the unified structure as a whole high-dimensional vector and inputting it into the embedding layer, mapping it to a continuous feature space through embedding transformation to obtain an initial embedded feature representation; Using a sliding time window mechanism to construct time-series data segments for the initial embedded feature representation, and establishing the time-series relationship of data evolution over time; Inputting the time-series data segments into a multi-layer perceptron network, using a non-linear activation function to extract deep feature vectors, and maintaining the stability between the original features and high-order features through a residual connection mechanism; Performing normalization processing and feature compression on the deep feature vectors to generate a fusion representation vector with a unified dimension; Outputting the fusion representation vector as a multi-dimensional spatio-temporal feature vector.

[0009] Further, the modeling of the spatio-temporal feature vector output by the feature processing module based on a time-series neural network to construct a predicted available charging pile number model includes the following steps: Arranging the multi-dimensional spatio-temporal feature vectors in chronological order to construct a sliding time window sequence for representing the state change features within a continuous time period; Inputting the time window sequence into a long short-term memory network, using its time gating mechanism to extract time-dependent features, and obtaining the hidden state vectors corresponding to each time step; Performing attention weighting processing on the hidden state vectors to generate a context-related global semantic representation; Inputting the global semantic representation into a fully connected neural network and outputting the predicted values of the available charging pile numbers of each candidate charging station at the target time point.

[0010] Further, the calculation of the estimated arrival time of the user from the current position to each charging station includes: Obtaining the longitude and latitude coordinate information of the user's current position and the user's departure time, and obtaining the geographical location information of each charging station within the target range; Invoking the navigation service API to obtain the required driving time corresponding to the optimal navigation path from the user's current position to each candidate charging station; Adding the required driving time to the user's departure time to obtain the estimated arrival time of the user at each charging station.

[0011] Further, the construction of the reinforcement learning model includes the following steps: Establishing a state space composed of the user's current position, the estimated available charging pile numbers of each charging station, the navigation time, the queuing status, and the historical utilization rate to represent the current environmental state; Define the action space as multiple candidate charging stations that the system can recommend to the user, and each action corresponds to a guiding strategy for guiding the user to one of the charging stations; Construct a reward function. Specifically, 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 a full charging pile, queue timeout, or navigation failure, a negative reward value is given. The reward value is adjusted in combination with the actual change in the utilization rate of the charging piles at the target charging station after recommendation; Use a reinforcement learning algorithm based on the policy optimization objective function to train the policy network, select the optimal action according to the current state, and update the model parameters using the backpropagation algorithm by comparing with the actually feedback reward value until it converges to stability, obtaining a trained reinforcement learning model.

[0012] Furthermore, the policy optimization objective function of the reinforcement learning model is as follows: ; Where, is the gradient of the objective function of the policy parameter θ; is the policy probability distribution of taking action in state , with the parameter θ; is the advantage function, which is used to define the advantage value of state and action relative to the average behavior under the current policy; is the gradient of the logarithmic probability of the policy network with respect to the parameter, which is used for backpropagation to update the policy network; is to take the expectation of the joint distribution of the state and the action.

[0013] Furthermore, the reward function of the reinforcement learning model is as follows: ; Where, is the immediate reward value at time step t; is an indicator function for the user to successfully complete charging at time step t, taking the value of 1 when successful and 0 when failed; is the actual utilization rate of the charging piles at the target charging station guided by the user at time step t; is the average utilization rate of the charging piles of all candidate charging stations in 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 improvement value of the regional charging resource scheduling efficiency caused by the current guiding policy; , , and They are the charging success incentive coefficient, the utilization rate optimization coefficient, the navigation cost penalty coefficient, and the resource scheduling efficiency reward coefficient respectively.

[0014] An intelligent charging pile guiding method based on a neural network, which is applied to the intelligent charging pile guiding system based on a neural network described in any one of the foregoing, includes the following steps: The acquisition module is used to collect the charging pile usage status data, user location information, weather information, regional event information, and time period information, and perform standardization processing to generate a data input with a unified structure. Perform feature extraction and fusion processing on the data input with the unified structure to generate an output multi-dimensional spatio-temporal feature vector. Based on a time series neural network, model the spatio-temporal feature vector output by the feature processing module to construct a prediction model for the number of available charging piles. Calculate the estimated arrival time of the user at each charging station according to the user's current location, calculate the number of future available charging piles at each charging station within the user's preset range through the prediction model of the number of available charging piles, and determine the guiding charging station based on the pre-trained reinforcement learning model according to the number of future available charging piles at each charging station. The reward function of the reinforcement learning model is constructed based on the user's successful charging situation and the utilization rate of the charging pile.

[0015] The beneficial effects of the present invention are as follows: The present invention fuses multi-source dynamic information such as the real-time status of the charging pile, user location, weather changes, regional events, and time period through the Internet of Things data acquisition module, and uniformly standardizes and generates structured input data to ensure that the system comprehensively perceives the current environmental status; uses the feature processing module to construct a high-dimensional spatio-temporal feature vector that fuses location, time, and environment, improving the accuracy and robustness of feature expression; further, by introducing a time series modeling mechanism based on a long short-term memory network (LSTM), predicts the availability of charging piles in the future period, making up for the lack of foresight ability of traditional systems. At the same time, a reinforcement learning algorithm is introduced to construct a navigation guidance strategy model, using the user's successful charging situation and the utilization rate of the charging pile as the design basis for the reward function, and realizing the dynamic optimization of the guidance strategy through continuous interactive learning. The reinforcement learning model not only considers the time consumption of the user's arrival path, but also incorporates the resource status at the predicted arrival time into the decision-making basis, ensuring that the system reasonably allocates charging demands under limited resources, avoiding user concentration congestion and resource waste. Through the synergistic effect of the above multi-layer neural network modeling and strategy optimization mechanism, this solution realizes the intelligent transformation of the charging pile guidance system from static recommendation to predictive guidance to adaptive scheduling, effectively improving the user charging experience and the overall resource allocation efficiency. Description of the Drawings

[0016] Figure 1 It is a schematic structural diagram of an intelligent charging pile guiding system based on a neural network in the present invention.

[0017] Figure 2 It is a flowchart of the steps for constructing the reinforcement learning model in the present invention. Detailed implementation manners

[0018] Please refer to Figure 1-2 As shown, the present invention relates to an intelligent charging pile guiding system based on a neural network, including: an Internet of Things data acquisition module, a feature processing module, a charging pile state prediction module, and a navigation and guidance decision-making module; The Internet of Things data acquisition module is used to collect charging pile usage status data, user location information, weather information, area event information, and time period information, and perform standardization processing to generate data input with a unified structure; The feature processing module is used to perform feature extraction and fusion processing on the data input with the unified structure to generate an output multi-dimensional spatio-temporal feature vector; The charging pile state prediction module is used to model the spatio-temporal feature vector output by the feature processing module based on a time series neural network to construct a prediction model for the number of available charging piles; The navigation and guidance decision-making module is used to calculate the estimated arrival time of the user at each charging station according to the user's current location, calculate the number of future available charging piles at each charging station within the user's preset range through the prediction model for the number of available charging piles, and determine the guiding charging station based on the pre-trained reinforcement learning model according to the number of future available charging piles at each charging station. The reward function of the reinforcement learning model is constructed through the user's successful charging situation and the charging pile utilization rate.

[0019] In some embodiments, first, the Internet of Things data acquisition module accesses the charging station device terminal, the user mobile positioning system, the third-party meteorological interface, and the urban event platform to collect real-time data on the usage status of charging piles, user location information, weather information, regional event information, and time period information. The collected data is uniformly standardized, including the normalization and missing value filling of numerical data, the one-hot encoding of discrete data, and the periodic mapping of time data (including encoding the hour feature using sine and cosine functions), to generate a unified format data tensor that is structured and can be used as the input for the neural network. 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. Subsequently, a sliding window strategy is used to generate fixed-length time series sample segments, forming an input sequence in the time dimension. This sequence is input into a Multi-Layer Perceptron (MLP), and through multi-layer non-linear transformation and residual connection structure, the fused deep representation is extracted, and a high-order spatio-temporal feature vector with a consistent output dimension is output as the modeling input for the subsequent prediction module. The charging pile status prediction module uses a Long Short-Term Memory (LSTM) to perform time series modeling on the spatio-temporal feature vector, capturing the long-term dependence of the charging pile usage status over time. In the LSTM calculation, the update of the hidden state includes the joint control of the input gate, the forget gate, and the output gate, which can effectively alleviate the problem of gradient disappearance and improve the modeling ability for long-term dependent information. To enhance the model's perception ability for key time segments, an Attention Mechanism is introduced, and by assigning attention weights to the hidden states of each time step, a global context-related feature vector is extracted. This vector is regressed through a fully connected neural network to obtain the predicted value of the number of available charging piles at the time points when users are expected to arrive at each charging station, realizing the parallel modeling and output of the future states of multiple target points. The navigation guidance decision module constructs a guidance strategy optimization model driven by reinforcement learning based on the predicted results of available charging piles and the user's navigation information. First, based on the navigation path calculation module, the estimated arrival time from the user's current location to each target charging station is obtained, and the predicted number of piles at the corresponding time point is obtained as a query index. The state space of the guidance strategy includes the encoding of the user's current location, the navigation time consumption, the future available state of the target pile position, the congestion degree of historical stations, etc.; the action space is the set of candidate charging stations that the system can recommend at the current time step.

[0020] Further, the steps of collecting the usage status data of the charging piles, user location information, weather information, regional event information, and time period information include the following steps: Real-time collect the usage status data of each charging pile from the charging station terminal device, including the current idle status, occupancy duration, number of queuing vehicles, and device operation status; Obtain the current location coordinate information of the user through the mobile terminal, and encode the area where the user is located in combination with the geographic information system; Call the weather service interface to obtain the real-time weather information of the area where the user is located, including temperature, precipitation, wind force level, and weather condition type; Obtain the area event information of whether there is traffic control, performances, rallies, and vehicle behaviors in a specific area during the prediction period from the urban event information platform; Extract the time period information corresponding to the SMS request time, including the hour segment, whether it is a working day, and whether it is a holiday, according to the server system time.

[0021] It should be noted that the collection of the charging pile usage status data relies on the edge gateway devices deployed at each charging site. This device synchronizes data with the server periodically (for example, every 30 seconds) and uploads the following metrics: current status flag (idle, occupied, faulty), occupied time of the current charging session (in seconds), cumulative number of queuing vehicles, device operation status (such as online, offline, fault code), etc. For abnormal states, such as occupied but no current change, no update for a long time, etc., the system will trigger a threshold judgment strategy, identify invalid states through the data continuity rule within the set time series window, and perform data elimination or interpolation to improve the quality of training samples and prediction stability. The user location information is obtained by the mobile terminal accessing the client through the Global Navigation Satellite System (GNSS) module, collecting latitude, longitude and elevation data, and combining 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 encoding of the location. To enhance the expression ability of spatial information in the neural network, a unique area identifier is mapped from the spatial coordinates using algorithms such as GeoHash or S2 hierarchical grid division, and an appropriate accuracy (such as 6-level or 7-level encoding) is set according to the area level for subsequent regional feature aggregation and adjacency relationship modeling. The weather information is obtained by calling third-party meteorological data interfaces (such as the API of the National Meteorological Administration or commercial platforms such as WeatherChina, OpenWeather, etc.), covering the current weather status of the area where the user is located, including temperature (in °C), precipitation (in mm / h), wind speed (in m / s), wind force level (from 0 to 12 levels according to meteorological standards), and weather phenomena (such as sunny, cloudy, light rain, heavy snow, etc.). Among them, numerical variables are directly used as model inputs after being standardized; categorical variables are converted into sparse vectors using the one-hot encoding method. To ensure the timeliness of information, the system adopts a caching mechanism for weather data and sets the shortest polling interval, for example, synchronizing once every 5 minutes, to improve the prediction accuracy. The collection of regional event information is based on the interface docking with the government data platform or the urban traffic event center to obtain structured event information flows. The event data includes key fields such as event type (such as traffic accident, large-scale performance, road control), event occurrence time, estimated end time, event center point coordinates, and event influence radius. The system judges whether the user's current location falls into the event influence area based on the spatial calculation module (such as using the spherical distance algorithm or R-tree index). If the spatial constraint and time overlap conditions are met, the corresponding event label is added to the user. In feature construction, the event type is represented in an enumerated form, and then encoded as a boolean regional event feature input through the spatial relationship between the event occurrence status and the user location.The time period information is extracted and processed based on the server timestamp, and the extracted items include hours (0-23), day of the week (0-6), whether it is a working day (Boolean value), whether it is a holiday (Boolean value), etc. Finally, after the above processing, the five types of information are converted into vector form respectively, and combined into a unified structure input tensor in the set order in the feature splicing stage as the input data of the feature processing module. This structured input fully retains the spatial-temporal context information of user behavior, the influence of external disturbance factors, and the evolution trend of resource status, providing a basic guarantee for the subsequent neural network in terms of prediction accuracy and strategy robustness.

[0022] Furthermore, the standardization process to generate data input with a unified structure includes the following steps: Fill in missing items and remove outliers from the collected charging pile usage status data, and perform normalization processing to generate a charging pile status feature vector; According to the longitude and latitude of the user's location information, and according to the area numbering of the geographic information system, a location feature vector is generated; Numerical encoding of temperature, precipitation and wind level in weather information, unique encoding of weather condition type, and generation of weather feature vector; Perform event type identification and regional association mapping on regional event information, identify traffic control, performances and gatherings related to the user's area during the forecast period, and generate event feature vectors; Periodically encode the time period information, including the sine and cosine mapping of hours, the classification of working days and holidays, and generate a time feature vector; The charging pile state feature vector, location feature vector, weather feature vector, event feature vector and time feature vector are spliced ​​in a preset order to construct a data input with a unified structure.

[0023] In some embodiments, first, for the charging pile usage status 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 enhance feature continuity. For outliers (such as occupancy time exceeding the maximum allowable limit, abnormal sudden increase in the number of queuing vehicles, etc.), the system automatically removes the outliers based on the interquartile range (IQR) method or the 3σ criterion to ensure the stability of the feature distribution. After cleaning, the numerical fields (such as occupancy duration, queuing quantity, etc.) are mapped to the [0,1] interval using the min-max normalization method (Min-Max Scaling), and the standardized results are used as the input of the charging pile status feature vector. For the user location data, the longitude and latitude coordinates obtained by the system are quantized into discrete area numbers through a GIS spatial mapping service based on the GeoHash or S2 cell encoding algorithm, and further encoded into numerical features. To enhance the spatial representation ability, a regional embedding vector or a regional adjacency graph can be introduced on the basis of the encoding as a reference for the input structure of subsequent graph neural networks or spatio-temporal networks. In this embodiment, the sparse one-hot vector of the area number is used as the location feature representation to ensure the unity of the input structure. For the weather information processing, the system models continuous variables and discrete variables separately. Temperature, precipitation, and wind force level are treated as continuous variables and are all standardized. Among them, the wind force level can be regarded as an ordered category for numerical mapping because it has a fixed level interval. The weather condition types (such as "sunny", "moderate rain", "heavy snow", etc.) are treated as unordered categories and are converted into sparse vector representations using the one-hot encoding method and are uniformly concatenated into the weather feature vector. The standardization process of the regional event information depends on the event recognition model and the spatial association rules. The system first classifies the events by type (such as performances, traffic control, exhibition gatherings, etc.) based on information such as keyword recognition, event tags, and location coordinates, and uses the spatial overlap relationship between the radius buffer and the user location to determine whether the event has an impact. Each type of event defines a binary label, forming a sparse event vector, indicating the possibility that the area where the user is located is affected by the event during the current prediction period, and enhancing the model's perception ability of sudden factors. The processing of the time period information is based on the periodic mapping and semantic hierarchical rules. The system extracts the current hour, day of the week, and holiday flag from the timestamp. Among them, the hour information is periodically encoded using the sine and cosine functions. This encoding can effectively enhance the model's ability to model the charging demand fluctuations within a 24-hour cycle. The holiday and weekday information is determined through a predefined calendar API and is represented by a boolean value indicating the holiday status, forming the time feature vector. After the standardization encoding of the above various types of features is completed, the system combines the five types of feature vectors (charging pile status, location, weather, event, time) in the preset concatenation order to construct an input feature tensor with a unified structure.

[0024] Further, the feature extraction and fusion processing of the data input to the unified structure includes the following steps: Taking the data input of the unified structure as a whole high-dimensional vector and inputting it into the embedding layer, mapping it to a continuous feature space through an embedding transformation, and obtaining an initial embedded feature representation; Constructing time-series data segments for the initial embedded feature representation by adopting a sliding time window mechanism, and establishing a time-series relationship of data evolution over time; Inputting the time-series data segments into a multi-layer perceptron network, using a non-linear activation function to extract deep feature vectors, and maintaining the stability between the original features and high-order features through a residual connection mechanism; Performing normalization processing and feature compression on the deep feature vectors to generate a unified-dimensional fusion representation vector; Outputting the fusion representation vector as a multi-dimensional spatio-temporal feature vector.

[0025] In some embodiments, first, the standardized unified structured input data is taken as a whole high-dimensional vector and input into the embedding layer for encoding transformation. Since this input vector contains multiple discrete features such as position encoding, weather category, event type, time flag, etc., to avoid the dimensionality explosion and sparsity problems brought by one-hot encoding, the system introduces a trainable embedding matrix at this stage to perform embedding mapping on each discrete feature type. For example, for the area number encoding, the embedding layer can map it to a dense embedding vector through a lookup table. After multiple embedding vectors are concatenated with continuous features, an initial feature representation that is semantically unified and numerically continuous is formed. After obtaining the initial embedding representation, to introduce time series dependence and state evolution relationship, the system constructs a sliding time window mechanism to synthesize the feature inputs at multiple consecutive moments within a historical time period into a time series data segment. This construction method can explicitly model the dynamic change trend of input variables within multiple time steps, which is beneficial for subsequent networks to identify periodic behaviors or mutation trends. Subsequently, this time series segment is input into a multi-layer perceptron network (MLP) for feature extraction. This MLP network contains an alternating stacking structure of multiple linear layers and non-linear activation functions (such as ReLU, GELU or Swish), and can capture the high-order non-linear combination relationship between the original feature vectors. In the deep structure, a residual connection mechanism (Residual Connection) is introduced, that is, the low-level output is directly added to the high-level input. To ensure the consistency of feature dimensions between different input samples and improve the numerical stability during the training of the neural network, the system introduces a normalization operation after feature extraction, such as batch normalization (Batch Normalization) or layer normalization (Layer Normalization), to uniformly adjust the distributions of different batches of samples. The normalized high-dimensional feature vector then undergoes a dimensionality reduction mapping through a linear transformation layer to compress it into a specified dimension, generating a fusion representation vector of a unified length. Finally, this fusion representation vector, as one of the core modeling results of the system, is defined as a multi-dimensional spatio-temporal feature vector and is used as the input for the subsequent charging pile state prediction model. This vector not only contains data embeddings from space, time, environment and user behavior, but also integrates high-order feature transformation information across time series, has strong expressive ability, and can effectively drive the high-precision modeling of the dynamic resource state of the charging pile in the subsequent model.

[0026] Furthermore, the step of modeling the spatio-temporal feature vector output by the feature processing module based on the temporal neural network to construct a predicted available charging pile number model includes the following steps: Arrange the multi-dimensional spatio-temporal feature vectors in chronological order to construct a sliding time window sequence for representing the state change features within a continuous time period; Input the time window sequence into a long short-term memory network, and use its time gating mechanism to extract time-dependent features to obtain the hidden state vectors corresponding to each time step; Perform attention weighting on the hidden state vectors to generate a context-related global semantic representation; Input the global semantic representation into a fully connected neural network to output the predicted values of the available charging piles at each candidate charging station at the target time point.

[0027] Specifically, by introducing a temporal neural network to model the multi-dimensional spatio-temporal feature vectors output by the feature processing module, a prediction model for the available number of charging piles in the future period is constructed. Specifically, the model first organizes the input spatio-temporal feature vectors in chronological order, and constructs continuous time segments through a sliding time window mechanism, thereby forming an input sequence with a temporal structure to express the dynamic changes of features such as the charging pile-related status, environmental factors, and user behavior in the past period of time. The sliding window mechanism is used to extract continuous feature vectors within a fixed-length time span and generate context information segments on which the current prediction depends for each moment. Each time segment contains feature representations corresponding to multiple historical time points, constituting 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, and the time gating mechanism it possesses is used to handle the long-term dependencies and short-term mutations in the input sequence. This neural network structure dynamically selects and updates the input features at each time step in the sequence through memory units and multiple gating mechanisms, and can effectively retain the temporal characteristics of the charging pile state change trend and external influencing factors (such as weather or events). While processing each time segment, the system further introduces an attention mechanism to perform weighted summation on the hidden states at 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 perception ability of key time nodes, especially when facing input sequences with non-stationarity or sudden interference, it has stronger adaptability.

[0028] To address the problem of the decline in prediction accuracy caused by the increase in the number of users, the system can introduce a user-guided intervention feedback mechanism and a dynamic distributed state correction strategy during the temporal modeling process, and embed a system-level impact modeling factor in the model architecture to enhance the model's sensitivity and adaptability to changes in group behavior, thereby effectively alleviating the prediction deviation caused by the guidance strategy itself.

[0029] Specifically, when constructing the time-series input data of the sliding time window, the system not only includes the state changes of charging piles at historical moments, external environmental factors, and individual user behavior characteristics, but also synchronously introduces the global feature of "the number of guided users", which reflects the intensity of system intervention. This feature is dynamically generated by the navigation guidance decision module, recording the number of users receiving system guidance within each time period, and serving as a time-series variable to be input into the LSTM network together with other features, helping the model capture the changing trend of resource distribution caused by the guidance behavior during modeling. For example, when a large number of users are continuously guided to a certain charging station by the system, its short-term availability may rapidly decline. If the prediction model fails to perceive the guidance intensity in a timely manner, there will be a problem of systematic over-recommendation. To further improve the convergence stability of the prediction model, the system can also introduce a periodic re-training and fine-tuning mechanism, that is, feedback sampling of the actual guidance results is performed at a fixed period (such as every 2 hours), and the error between the predicted value and the available number of charging piles at the actual arrival time is calculated. When the error exceeds the threshold, the local model update process is triggered, and the subset of parameters representing the impact of user-intensive guidance behavior is preferentially updated to improve the modeling response speed to congestion dynamics. At the same time, a penalty term is introduced during the model training stage to model the coupling degree of the prediction and actual error with the user guidance density, so that the model has a certain "anti-regulation ability", that is, actively avoiding the risk of hot resource aggregation caused by system guidance in the prediction. In addition, during the semantic weighting process of the attention mechanism, the system adds a "guidance impact estimation factor", which adjusts the attention weight of the time step through the coupling degree with the guided user density in the historical time period, making the model more inclined to learn the resource change law from the historical behavior in the low-guidance interference state, and avoiding the erosion of the long-term semantic expression of the model by the guidance behavior interference.

[0030] Through the above structural and algorithmic improvements, this embodiment not only constructs a prediction framework with time-dependent modeling capabilities, but also solves the prediction deviation problem caused by the large-scale guidance behavior of the system, ensuring high-precision prediction of the available number of charging piles in the scenario of continuous growth of the number of users, and significantly improving the stability, reliability, and robustness of the model against group interference.

[0031] Further, the calculation of the estimated arrival time of the user at each charging station according to the user's current location includes: Obtain the longitude and latitude coordinate information 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 range; Call the navigation service API to obtain the required driving time corresponding to the optimal navigation path from the user's current location to each candidate charging station; Add the required driving time to the user's departure time to obtain the estimated arrival time of the user at each charging station.

[0032] Specifically, first, the system obtains the user's current precise geographical location through the mobile terminal, extracts its longitude and latitude coordinates, and simultaneously records the system timestamp when the user triggers the guidance request as the actual departure time. To improve the spatial consistency of location information, the system performs unified coordinate system conversion and regional coding on the static geographical coordinates of all charging stations, constructs a target site spatial index structure, so as to quickly screen out the set of candidate charging stations within the user-specified navigation radius. Subsequently, the system calls the high-precision navigation service API (such as the online map service path calculation interface optimized based on the A* or Dijkstra algorithm), takes the user's current location as the starting point, and constructs a shortest path or optimal path query request for each candidate charging station in turn. This path planning process is not only based on static map topology data, but also integrates multi-factor dynamic information such as road traffic status, real-time traffic congestion conditions, and traffic restriction rules, to obtain the optimal driving path and the corresponding required time at the current time period. To improve the concurrent performance of path planning, the system adopts an asynchronous batch request mechanism, submits all candidate paths to the navigation server at one time, and receives the optimal time results of each path through the callback method, effectively reducing the API response delay.

[0033] Furthermore, the construction of the reinforcement learning model includes the following steps: Establish a state space composed of the user's current location, the expected number of available charging piles at each charging station, the navigation time, the queuing status, and the historical utilization rate, to represent the current environmental state; Define the action space as multiple 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; Construct a reward function. Among them, 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, queuing timeout, or navigation failure, a negative reward value is given, and the reward value is adjusted in combination with the actual change in the utilization rate of the charging pile after the target charging station is recommended; Use a reinforcement learning algorithm based on the policy optimization objective function to train the policy network, select the optimal action according to the current state, and update the model parameters using the backpropagation algorithm by comparing with the actually feedback reward value until it converges to stability, and obtain the trained reinforcement learning model.

[0034] Specifically, first, the system constructs a state space to represent the environment where the user is currently located in multi-dimensional features. This state not only includes the user's real-time geographical location but also combines dynamic and static information such as the predicted available charging piles at each candidate charging station at the predicted arrival time of the user, the predicted driving time from the user's current location to the charging station, the current queuing status of the charging station, and the historical utilization rate. 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 several candidate charging stations that the system can recommend at the current moment, and each action is a specific strategy to guide the user to a certain charging station. The system uses a neural network to model the policy function, outputs the probability distribution of each action according to the current state, and selects the target charging station with the highest probability or executes the policy through sampling. In reinforcement learning, the design of the reward function is particularly crucial. In this embodiment, it is used as the core evaluation index for guiding optimization. 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 regarded as an effective guidance, and the system gives a positive reward; if the guidance fails due to insufficient resources of the target charging pile, too long queuing time, or navigation path interruption, a negative reward is imposed to punish the bad policy. At the same time, after the guidance is successful, the system will also evaluate the change trend of the resource utilization rate of this station. If a reasonable transfer of the resource load is achieved due to the guidance, an additional reward is given to encourage the policy model to learn behaviors that promote the overall resource balance of the system. To train the policy model, the system adopts an optimization algorithm based on policy gradients, combines the feedback rewards generated by the actual guidance behavior, and continuously adjusts the parameters of the policy network, so that in a similar environmental state, a better charging station guidance strategy can be selected in the future. The training process is carried out in the form of multiple rounds of simulated interactions. Each round of guidance behavior forms a state-action-reward sequence. The system evaluates the quality of the current policy according to the cumulative return and optimizes the neural network parameters through the backpropagation mechanism, so that the policy gradually converges to the optimal. In addition, to solve the problem of guidance interference caused by the growth of the user population, the system introduces a system behavior influence factor, incorporates the number of users who have recently received system guidance as a global feature in the state modeling stage, so that the model can perceive the perturbation of the guidance density on the resource distribution, and thus avoid over-concentrating on recommending a certain popular station. In addition, the system also combines a real-time feedback mechanism and a regular fine-tuning mechanism to perform incremental updates on the policy model, so that it can long-term maintain adaptability to the user behavior pattern and environmental changes.

[0035] Furthermore, the policy optimization objective function of the reinforcement learning model is as follows: ; Wherein, is the gradient of the objective function of the policy parameter θ; is the action taken in the state and The policy probability distribution with parameter θ; is the advantage function, which is used to define the state and the action as the advantage value relative to the average behavior under the current policy; is the gradient of the policy network's log probability with respect to the parameter, which is used for backpropagation to update the policy network; is to take the expectation of the joint distribution of the state and the action.

[0036] Specifically, the core of constructing this objective function lies in quantifying the quality of policy behavior by introducing the advantage function, and weighting it in combination with the log probability gradient of the policy network, so as to guide the iterative update of the policy network parameters in the direction that is most conducive to improving the long-term return, and achieve effective learning and optimization of the guiding policy. Specifically, the policy network outputs a probability distribution at any given state, which is used to represent the system's preference for different guiding actions. For example, when the user is at a certain geographical location and the system needs to select the optimal guiding target among multiple charging stations, the policy network outputs the selection probabilities of each target site according to the state input. The optimization objective function measures the response degree of the current policy to a specific action at this state by taking the derivative of the log probability function of the policy. To determine whether this response is worthy of reinforcement, the system further introduces the advantage function to evaluate the relative advantage of this action compared to the average policy behavior in the current state. The calculation of the advantage function is based on the reward evaluation mechanism, that is, observing the actual or estimated long-term benefits brought after performing an action, and comparing this benefit with the expected rewards of all possible behaviors in this state. If the effect brought by this action is better than the average level, the advantage function is positive; if it is lower than the average level, it is negative. The optimization objective function uses the advantage function as a weight term and acts on the log probability gradient to form the parameter update direction. In this way, the policy network will continuously increase the selection tendency for high-advantage actions and suppress low-advantage actions during the training process, so as to achieve the improvement of the overall performance of the policy. In addition, during the implementation of this objective function, the system uses the sample sampling and expectation approximation method to construct the empirical estimate of the objective function through the state-action pairs obtained by batch sampling, and updates the parameters of the policy network through the backpropagation mechanism. This optimization process has the characteristics of fast convergence speed and strong policy convergence stability, and is especially 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, this system can not only achieve reasonable guidance for users in a static state, but more importantly, it has the ability to continuously learn and adapt in a dynamic environment. With the changes in the spatio-temporal distribution of charging demands, the evolution of user behavior patterns, and the dynamic fluctuations of resource states, the system can adjust the guiding policy in real time to ensure the improvement of the user's successful charging rate and the efficient utilization of the overall charging pile resources, thus realizing a highly intelligent and robust guiding optimization mechanism.

[0037] Furthermore, the reward function of the reinforcement learning model is as follows: ; where is the immediate reward value at time step t; is an indicator function for whether the user successfully completes charging at time step t, taking the value of 1 when successful and 0 when failed; is the actual charging pile utilization rate of the target charging station where the user is guided at time step t; is the average charging pile utilization rate of all candidate charging stations in 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 improvement value of the regional charging resource scheduling efficiency caused by the current guiding policy; , , and are the charging success incentive coefficient, utilization rate optimization coefficient, navigation cost penalty coefficient, and resource scheduling efficiency reward coefficient, respectively.

[0038] Specifically, first, the reward function takes the user's successful charging situation as the core evaluation basis. When the user arrives at the target charging station according to the system guidance and successfully completes the charging operation, the system assigns a positive reward value to this behavior; if the user fails to charge due to full chargers at the target station, queue timeout, or other factors, no reward is given. This part is implemented through a boolean indicator function, and its result directly reflects the basic service effect of the policy execution and represents the most direct feedback of user satisfaction. Secondly, 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 a regulatory factor. This ratio is used to measure whether the current guiding behavior helps to balance the resource load distribution in the region. If the utilization rate of the target station is moderate or relatively low, this guiding behavior helps to relieve congestion in hot spots or improve the resource utilization efficiency in marginal areas; conversely, if it guides to a high-load area, this ratio will inhibit the growth of the reward value, thus prompting the policy to avoid concentrated recommendations. In addition, 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 quantifies the time or distance for the user to navigate from the current location to the target charging station. If the path is too long or the estimated time consumption is too high, it is weighted and deducted in the reward function with a penalty coefficient. This part ensures that while the policy optimizes resources, it also takes into account the user's travel cost and convenience. Finally, the system also incorporates the amount of improvement in resource scheduling efficiency brought about by the policy execution at the regional level into the reward function calculation, which is used to quantify the positive contribution of a guiding behavior to the overall regional resource balance. This part is evaluated by comparing the changes in the charger load distribution or standby time before and after the guidance. If the policy effectively reduces the regional congestion or improves the resource turnover rate, the system gives an additional positive incentive to guide the policy to optimize towards the maximum overall system efficiency. Adjustable weight coefficients are set for the above indicators in the reward function to adapt to the policy preferences and operation goals in different cities and different scenarios. Overall, this reward function takes into account both the success rate of individual behaviors and the system resource allocation efficiency, has good feedback orientation and optimization driving ability, and provides a stable and business-valued learning signal source for the effective training of the reinforcement learning policy model. Through this reward mechanism, the system can ultimately achieve high-robustness intelligent guiding ability for complex and changing urban environments, improve the user experience, and ensure the efficient use of public resources.

[0039] The present invention also includes a neural network-based intelligent charging pile guiding method, which is applied to the neural network-based intelligent charging pile guiding system described in any one of the foregoing items, and 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 a data input with a unified structure; Extract and fuse the features of the data input of the unified structure to generate an output multi-dimensional spatio-temporal feature vector; Based on a temporal neural network, model the spatio-temporal feature vector output by the feature processing module to construct a prediction model for the number of available charging piles; Calculate the estimated arrival time of the user at each charging station according to the user's current location, calculate the number of future available charging piles at each charging station within the user's preset range through the prediction model for the number of available charging piles, and determine the guiding charging station based on the number of future available charging piles at each charging station according to a pre-trained reinforcement learning model. The reward function of the reinforcement learning model is constructed based on the user's successful charging situation and the utilization rate of the charging piles.

[0040] The above embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A charging pile intelligent guidance system based on neural network, characterized in that: include: IoT data acquisition module, feature processing module, charging pile status prediction module and navigation guidance decision module; The IoT data collection 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 perform feature extraction and fusion processing on the data input of the unified structure to generate an output multi-dimensional spatiotemporal feature vector; The charging pile state prediction module is used to model the spatiotemporal feature vector output by the feature processing module based on a time series neural network 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 at each charging station according to the user's current location, calculate the future number of available charging piles at each charging station within the user's preset range through the available charging pile number prediction model, and determine the guidance charging station based on the future number of available charging piles at each charging station based on the pre-trained reinforcement learning model. The reward function of the reinforcement learning model is constructed through the user's successful charging situation and the charging pile utilization rate.

2. According to the neural network-based charging pile intelligent guidance system of claim 1, it is characterized in that: The collection of charging pile usage status data, user location information, weather information, regional event information and time period information includes the following steps: Collect the usage status data of each charging pile in real time from the charging station terminal equipment, including the current idle status, occupancy time, number of queued vehicles and equipment operation status; Obtain the user's current location coordinate information through the mobile terminal, and encode the user's location area in combination with the geographic information system; Call the weather service interface to obtain real-time weather information in the user's area, including temperature, precipitation, wind level and weather condition type; Obtain regional event information from the city event information platform regarding whether there are traffic controls, performances, gatherings, and vehicle behaviors in a specific area during the forecast period; The hour period corresponding to the SMS request time, whether it is a working day or a holiday, is extracted according to the server system time.

3. The neural network-based charging pile intelligent guidance system according to claim 2 is characterized in that: The standardization process to generate data input with a unified structure includes the following steps: Fill in missing items and remove outliers from the collected charging pile usage status data, and perform normalization processing to generate a charging pile status feature vector; According to the longitude and latitude of the user's location information, and according to the area numbering of the geographic information system, a location feature vector is generated; Numerical encoding of temperature, precipitation and wind level in weather information, unique encoding of weather condition type, and generation of weather feature vector; Perform event type identification and regional association mapping on regional event information, identify traffic control, performances and gatherings related to the user's area during the forecast period, and generate event feature vectors; Periodically encode the time period information, including the sine and cosine mapping of hours, the classification of working days and holidays, and generate a time feature vector; The charging pile state feature vector, location feature vector, weather feature vector, event feature vector and time feature vector are spliced ​​in a preset order to construct a data input with a unified structure.

4. The neural network-based charging pile intelligent guidance system according to claim 1, characterized in that: The feature extraction and fusion processing of the unified structure data input comprises the following steps: Inputting the unified structure data input as a high-dimensional vector into the embedding layer as a whole, mapping it to a continuous feature space through embedding transformation, and obtaining an initial embedded feature representation; Using a sliding time window mechanism to construct time series data segments for the initial embedded feature representation, and establishing a time series relationship of data evolution over time; Inputting the time series data fragments into a multi-layer perceptron network, extracting deep feature vectors using a nonlinear activation function, and maintaining the stability between the original features and the high-order features through a residual connection mechanism; Normalizing and compressing the deep feature vector to generate a fusion representation vector with unified dimension; The fused representation vector is output as a multi-dimensional spatiotemporal feature vector.

5. The neural network-based charging pile intelligent guidance system according to claim 1, characterized in that: The method of modeling the spatiotemporal feature vector output by the feature processing module based on a time series neural network and constructing a prediction model for the number of available charging piles includes the following steps: Arrange the multi-dimensional spatiotemporal feature vectors in chronological order to construct a sliding time window sequence for representing state change characteristics within a continuous time period; Inputting the time window sequence into a long short-term memory network, using its time gating mechanism to extract time-dependent features, and obtaining a hidden state vector corresponding to each time step; Performing attention weighted processing on the hidden state vector to generate a context-dependent global semantic representation; The global semantic representation is input into a fully connected neural network, and a predicted value of the number of available charging piles for each candidate charging station at a target time point is output.

6. The neural network-based charging pile intelligent guidance system according to claim 5, characterized in that: Calculating the estimated arrival time of the user at each charging station according to 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 geographic location information of each charging station within the target range; Call the navigation service API to obtain the optimal navigation path from the user's current location to each candidate charging station and the corresponding driving time; The required driving time is added to the user's departure time to obtain the user's estimated arrival time at each charging station.

7. The neural network-based charging pile intelligent guidance system according to claim 1, characterized in that: 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, queue status, and historical utilization rate to represent the current environmental state; Define the action space as multiple 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; Construct a reward function, where a positive reward value is given if the user successfully starts charging after arriving at the target charging station; a negative reward value is given if the user cannot complete charging due to the charging pile being full, the queue timeout, or navigation failure. The reward value is adjusted based on the actual change in the utilization rate of the charging piles at the target charging station after the recommendation; 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 using the back propagation algorithm by comparing it with the actual feedback reward value until it converges to stability, thus obtaining a trained reinforcement learning model.

8. The neural network-based charging pile intelligent guidance system according to claim 7, characterized in that: The strategy optimization objective function of the reinforcement learning model is as follows: ； in, is the objective function gradient of the policy parameter θ; For the status Take action The strategy probability distribution of , with parameter θ; is the advantage function, which is used to define the state and actions The advantage of the current strategy relative to the average behavior; The gradient of the log probability of the policy network with respect to the parameters is used for back propagation to update the policy network; Find the expectation of the joint distribution of states and actions.

9. The neural network-based charging pile intelligent guidance system according to claim 8, characterized in that: The reward function of the reinforcement learning model is as follows: ； in, is the immediate reward value at time step t; is the indicator function of whether the user successfully completes charging at time step t, with a value of 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 in the system at time step t; To prevent the denominator from being zero, a stability constant is used; The time or distance cost of navigating a path for the user; is the improvement value of regional charging resource scheduling efficiency caused by the current guidance strategy; , , and They are charging success incentive coefficient, utilization optimization coefficient, navigation cost penalty coefficient and resource scheduling efficiency reward coefficient.

10. A charging pile intelligent guidance method based on a neural network, applied to a charging pile intelligent guidance system based on a neural network according to any one of claims 1 to 9, characterized in that: The following steps are involved: 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; Performing feature extraction and fusion processing on the data input of the unified structure to generate an output multi-dimensional spatiotemporal feature vector; Modeling the spatiotemporal feature vector output by the feature processing module based on a time series neural network to construct a prediction model for the number of available charging piles; The estimated arrival time at each charging station is calculated based on the user's current location, and the future number of available charging piles at each charging station within the user's preset range is calculated through the available charging pile number prediction model. The guide charging station is determined based on the future number of available charging piles at each charging station based on the pre-trained reinforcement learning model. The reward function of the reinforcement learning model is constructed through the user's successful charging situation and the charging pile utilization rate.

Images