Intelligent electric vehicle charging station recommendation system and method

Abstract

The invention relates to the technical field of charging station management systems, in particular to an intelligent electric vehicle charging station recommendation system and method. Comprising a big data processing layer, a multi-path recall layer, a recall duplicate removal layer and a sorting layer, the big data processing layer collects basic information and then outputs user feature information, user context feature information, charging station feature information and charging station context feature information; the user feature information mainly comprises user basic information, preference information, statistical information and user behavior information; the charging station feature information mainly comprises charging station basic information and charging station statistical information; the multi-path recall layer comprises recall based on user context, recall based on user interestingness and recall based on charging station similarity; recall of user interestingness uses a collaborative filtering algorithm, recall based on charging station similarity is obtained through a deep learning recall model, results of the multi-path recall layer are sent to the recall duplicate removal layer, output of the recall duplicate removal layer is sent to the sorting layer, and the sorting layer outputs recommendation results.

Description

Technical Field

[0001] The present invention relates to the technical field of charging station management systems, and in particular to a smart electric vehicle charging station recommendation system and method. Background Art

[0002] In recent years, with the increasing number of electric vehicles, the number and scale of electric vehicle charging stations have also been growing. At present, due to the backward construction of charging station operation and management systems, some of the limited electric vehicle charging stations / piles in China are in an overloaded working state, while others are idle. In addition, there are different charging needs for different types of electric vehicle users. For example, some users like to go to charging stations with business districts nearby, some users are more concerned about charging prices and like to go to nearby charging stations with activities or the lowest prices, and some users always like to go to the charging station closest to them, etc.

[0003] Personalized recommendations for users have always been an important technical guarantee for improving product brand value and winning customer markets. The so-called personalized recommendation means that the system can capture the user's personal characteristics such as interests and hobbies in a timely and accurate manner, and recommend products that meet their characteristics, in order to improve user experience and increase revenue benefits. Personalized recommendations have been widely used in e-commerce and search engine fields, but are still rare in the field of electric vehicle charging station recommendations.

[0004] At present, the use of user portraits and item portraits can basically reflect the characteristics of users or commodities, but each portrait may have hundreds of features. The existing system lacks an effective model to capture users' complex preferences and real-time needs, resulting in the recommended charging stations not meeting user expectations. Faced with massive user behavior data and environmental data, existing systems are often difficult to process and analyze in real time. Secondly, the user's charging behavior is not only related to the user's current time and location, but also closely related to the current status of the charging station. The recommendation results should have a certain degree of effectiveness. Therefore, it is necessary to provide a charging pile recommendation system that can combine massive user behavior data and environmental data to recommend personalized charging piles to users in real time. Summary of the invention

[0005] The problem to be solved by the present invention is: how to combine massive user behavior data and environmental data to recommend personalized charging piles to users in real time.

[0006] In view of the deficiencies in the prior art, the present invention adopts a technical solution to solve the technical problems thereof: a smart electric vehicle charging station recommendation system, comprising a big data processing layer, a multi-channel recall layer, a recall de-recall layer and a sorting layer; the big data processing layer collects basic information and outputs user feature information, user context feature information, charging station feature information and charging station context feature information; the user feature information mainly includes user basic information, preference information, statistical information and user behavior information; the charging station feature information mainly includes charging station basic information and charging station statistical information; the multi-channel recall layer includes user context-based recall, user interest-based recall and charging station similarity-based recall; the user interest-based recall uses a collaborative filtering algorithm, the charging station similarity-based recall is obtained through a deep learning recall model, the result of the multi-channel recall layer is given to the recall de-recall layer, the output of the recall de-recall layer is given to the sorting layer, and the sorting layer outputs the recommendation result.

[0007] Preferably, the user context-based recall includes five-way recall, namely: recall the k stations closest to the charging station based on the distance between the user and the charging station; recall the k stations with the highest availability near the current moment based on the user's location; recall the k stations with the largest number of charging people in the city where the current user is located; recall the k stations with the largest number of charging people in the administrative district where the current user is located; and recall the k stations with the lowest charging prices near the current moment based on the user's location.

[0008] Preferably, the recall based on user interest is performed using a collaborative filtering algorithm, and the collaborative filtering algorithm includes user-based collaborative filtering and charging station-based collaborative filtering. The specific algorithm of collaborative filtering is as follows:

[0009]

[0010] Where v1 and v2 are two different user real-time feature vectors or two different charging station feature vectors;

[0011] Thus, in user-based collaborative filtering, user u's interest in site s can be expressed as:

[0012]

[0013] Where U is the user set, score(u i ,s) is user u i Number of charges at charging station s;

[0014] In collaborative filtering based on charging stations, user u’s interest in station s can be expressed as:

[0015]

[0016] Where S is the set of charging stations, score(u,si ) is user u at charging station s i The number of charging times.

[0017] Preferably, the deep learning recall model includes a user vector tower, a charging station vector tower and a user context vector tower. The user vector tower is obtained by passing the user feature information through its own MLP and attention mechanism, the user context feature information is passed through its own MLP to obtain the user context vector tower, and the charging station feature information is passed through its own MLP to obtain the charging station vector tower; the user vector tower and the user context vector tower are both input into the MLP to obtain the user real-time feature vector, the charging station vector tower can obtain the charging station feature vector, and the user real-time feature vector and the charging station feature vector are calculated through the inner product to obtain the charging station similarity.

[0018] Preferably, the sorting layer uses a deep learning sorting model, the input parameters of the model include user feature information, user context feature information, charging station feature information and charging station context feature information, and each parameter is respectively passed through its own FM to generate a corresponding vector tower; the user vector tower is spliced ​​with the user context vector tower to obtain the user real-time feature vector, the charging station vector tower is spliced ​​with the charging station context vector tower to obtain the charging station real-time feature vector, and the user real-time feature vector and the charging station real-time feature vector are input into the MLP to calculate the score.

[0019] Preferably, MLP uses a ReLu activation function, and the specific ReLu activation function is:

[0020]

[0021] The calculation process of FM is as follows:

[0022]

[0023] Among them, n is the number of sample data, x i is the i-th feature, <v i ,v j > is the inner product form of the second-order feature combination parameters, v i and v j is x i and x j The corresponding hidden vectors, w0 and w i First-order weight parameters for deep learning ranking models.

[0024] A recommendation method of an intelligent electric vehicle charging station recommendation system comprises the following steps:

[0025] S101: Obtaining initial data: querying user feature information based on user ID; obtaining user context feature information based on current time and location information;

[0026] S102: Multi-channel recall: According to the location in the user context feature information, query the top k nearest stations as the first recall result, query the top stations with the largest number of charging users in the current city as the second recall result, and query the top k stations with the largest number of charging users in the current administrative area as the third recall result; according to the location and time in the user context feature information, query the top k nearby charging stations with the highest availability at the current moment as the fourth recall result, and query the top k nearby charging stations with the lowest charging price at the current moment as the fifth recall result; query the top k most interesting charging stations returned by the collaborative filtering algorithm by the user ID as the sixth recall result, input the user feature information and the user context feature information into the deep learning recall model, and obtain the top k stations with the highest station similarity as the seventh recall;

[0027] S103: De-duplicating the charging stations recalled by the multi-channel recall method to obtain a set of candidate charging stations;

[0028] S104: Querying the station portrait information of the candidate object and the context feature information of the candidate charging station at the current moment;

[0029] S105: Inputting user feature information, user context feature information, charging station feature information, and charging station context feature information into a deep learning ranking model;

[0030] S106: The deep learning ranking model scores the candidate charging stations;

[0031] S107: Select k charging stations with the highest scores as recommendation results.

[0032] The beneficial effects of the present invention are as follows: the present invention obtains the data features of users and charging stations in the platform based on big data processing technology, selects an appropriate number of charging stations for each charging user through a multi-channel recall method including a deep learning recall model, and then uses a deep learning model based on an attention mechanism to extract effective features of users and recommended stations, and then returns the charging station set that the user is most interested in through sorting.

[0033] (1) The present invention comprehensively lists the relevant features of users and charging stations in the field of charging station recommendation, as well as the features of real-time context information of users and stations;

[0034] (2) The existing recall mechanism is single and mostly based on distance, and fails to fully consider multi-dimensional information such as user historical behavior and the real-time status of charging stations. The present invention uses a multi-channel recall mechanism including a deep recall model to enrich the recall strategy;

[0035] (3) Faced with massive amounts of user behavior data and environmental data, existing systems often find it difficult to process and analyze in real time. The present invention uses big data processing technology to obtain accurate and comprehensive user / charging station feature data by analyzing massive amounts of user and charging station data; it uses big data streaming processing technology to analyze the real-time status of users and charging stations, and accurately obtains the current context information of users and charging stations, which can make the recommendation results of this recommendation method more timely;

[0036] (4) Traditional deep learning recall / ranking models have problems such as large model weights and slow reasoning speed. This paper divides the weights of the deep recall model into user vector towers, context vector towers, and charging station vector towers, where the data of user vector towers and charging station vector towers can be processed offline and stored in a vector database. In this way, the model only needs to calculate a small number of dimensions of the context vector tower each time it recalls to obtain the real-time vector features of the user / charging station, and then obtain the results based on the similarity, which greatly speeds up the reasoning speed of the model; the ranking model also uses this idea;

[0037] (5) The present invention adds an attention mechanism to the sorting / recall deep learning model to analyze the user's historical behavior data, rather than a simple pooling operation. This enables the model to more accurately understand the user's behavior habits and charging intentions, thereby improving the accuracy of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the structure of the big data processing layer of the present invention;

[0039] Figure 2 It is a deep learning recall model involved in the present invention;

[0040] Figure 3 It is a deep learning sorting model involved in the present invention;

[0041] Figure 4 It is a flowchart of the system work flow of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are provided for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

[0043] As shown in the figure, the present invention relates to an intelligent electric vehicle charging station recommendation system, which is based on big data processing, deep learning, and attention mechanism recommendation technology, and is suitable for improving the charging experience of electric vehicle users. The system of the present invention includes a big data processing layer, a multi-channel recall layer, a recall re-recall layer, a splicing recall charging station context feature information layer, and a sorting layer, and the sorting layer is based on the deep learning technology of the attention mechanism.

[0044] Specifically, the big data processing layer is as follows: Figure 1 As shown in the figure, the following data features can be obtained by processing basic information using big data processing technology: user feature information, user context feature information, charging station feature information, and charging station context feature information. Basic information includes user location, user orders, sensor data, user behavior, charging station status reporting, and charging station surrounding data.

[0045] The user characteristic information mainly includes user basic information, preference information, statistical information, and user behavior information, as follows:

[0046] a) Basic user information: user ID, gender, age, registration city, registration channel, registration days, number of all electric vehicles, and all electric vehicle brands;

[0047] b) Preference information: frequently visited provinces, frequently visited cities, frequently visited administrative regions, frequently visited power stations, frequently visited days of the week for charging, frequently visited charging station types, commonly visited payment channels, frequently visited time periods of the day for charging, commonly visited car brands, frequently visited parking lot charging types, and commonly used charging pile power;

[0048] c) Statistical information: number of provinces where electricity has been charged, number of cities where electricity has been charged, number of administrative regions where electricity has been charged, number of time periods where electricity has been charged, number of charging pile power types, number of charging stations where electricity has been charged, charging frequency, activity participation rate, card and voucher usage rate, cumulative charging costs, cumulative number of charging times, historical lowest unit price for charging orders, historical highest unit price for charging orders, historical median unit price for charging orders, historical highest power level for charging orders, historical lowest power level for charging orders, historical median power level for charging orders, historical longest duration for charging orders, historical shortest duration for charging orders, historical median market for charging orders, historical highest cost for charging orders, historical lowest cost for charging orders, historical median cost for charging orders.

[0049] d) User behavior information: a list of charging stations that the user has recently browsed, collected, or used, sorted in ascending order by time.

[0050] User context feature information mainly includes the user's Amap location coordinates; current time information: week, day of the week, holiday type, time period, month, and hour.

[0051] Charging station characteristic information mainly includes basic information of charging station and statistical information of charging station, as follows:

[0052] a) Basic information of charging stations: charging station ID, operation type, power station attributes, power station type, charging mode, whether there is a ground lock, owner, operation, cooperation type, province, city, administrative district, floor, parking fee type, number of charging piles, number of AC piles, number of DC piles, total power of charging piles, whether reservation is supported, whether anti-occupancy is supported, and whether interconnection is supported;

[0053] b) Charging station statistics: cumulative number of charging orders, frequency of use of charging stations, frequency of use of single charging piles, maximum number of charging orders per day, minimum number of charging orders per day, average number of charging orders per day, median number of charging orders per day, maximum daily charging volume, minimum daily charging volume, average daily charging volume, median daily charging volume, maximum daily charging time, minimum daily charging time, average daily charging time, median daily charging time, maximum daily charging income, minimum daily charging income, average daily charging income, median daily charging income, order comment rate, charging rate using own channels, card and voucher usage rate, and event participation rate;

[0054] The context feature information of the charging station includes the occupancy status, fault status, current charging price of the charging station, number of currently available charging piles, and activities of the charging station.

[0055] The multi-channel recall layer includes recall based on user context and recall based on user interest and charging station similarity. The recall based on user context includes five channels:

[0056] The first is to recall the k nearest charging stations based on the distance between the user and the charging station;

[0057] The second is to recall the k stations with the highest availability near the current moment according to the user's location;

[0058] The third is to recall the k stations with the largest number of charging users in the current user's city;

[0059] The fourth is to recall the k stations with the largest number of charging users in the administrative area where the current user is located;

[0060] Fifth, based on the user's location, recall the k charging stations with the lowest prices near the current moment.

[0061] The recall based on user interest uses collaborative filtering algorithm, and the recall based on charging station similarity uses deep learning recall model. The collaborative filtering algorithm is a very simple "birds of a feather flock together" idea. Birds of a feather flock together means calculating the most similar list of charging stations, and then recommending similar charging stations to the user's favorite charging station. This is collaborative filtering based on charging stations. People are divided into groups, which means recommending charging stations where users similar to the user like to charge to the user. This is collaborative filtering based on users.

[0062] Among them, the specific algorithm of collaborative filtering is as follows:

[0063]

[0064] Where v1 and v2 are two different user real-time feature vectors or two different charging station feature vectors. In this way, in user-based collaborative filtering, the interest of user u in station s can be expressed as follows:

[0065]

[0066] Where U is the user set, score(u i ,s) is user u i Number of charges at charging station s.

[0067] In collaborative filtering based on charging stations, user u’s interest in station s can be expressed as follows:

[0068]

[0069] Where S is the set of charging stations, score(u,s i ) is user u at charging station s i The number of charging times.

[0070] With the user's interest in each charging station, the top k charging stations can be returned in descending order as the recall result of the route.

[0071] The calculation of charging station similarity is obtained through the deep learning recall model, such as Figure 2 As shown in Figure 2, the deep learning recall model mainly consists of three parts: user vector tower, charging station vector tower and user context vector tower. Figure 2As shown. The user feature information is passed through the perceptual neural machine MLP (multi-layer deep neural network DNN) and the attention mechanism to obtain the user vector tower. The user context feature information is passed through its own MLP to obtain the user context vector tower, and the charging station feature information is passed through its own MLP to obtain the charging station vector tower. Then, the user vector tower and the user context vector tower are input into the MLP to obtain the user real-time feature vector. The charging station vector tower can obtain the charging station feature vector. The user real-time feature vector and the charging station feature vector are used to calculate the similarity through the inner product. The higher the similarity, the more interested the current user is in this charging station.

[0072] The user vector tower and charging station feature vector are not updated in real time but on a daily basis, so they can be calculated offline and stored in the vector database in advance. In this way, when the model is recalled online for reasoning, only the user context vector tower needs to be calculated, which can significantly reduce the amount of calculation and speed up the reasoning of the model.

[0073] After the multi-channel recall is completed, it is necessary to deduplicate these recalled charging stations, query the context information of these charging stations respectively after deduplication, and then sort the deduplicated charging stations. The specific structure of the deep learning sorting model is as follows: Figure 3 As shown. Compared with the recall model, the number of charging stations that need to be sorted by the sorting model has been greatly reduced, and a complex network structure can be used. The model input parameters include user feature information, user context feature information, charging station feature information, and charging station context feature information. Similarly, the user feature information is given to FM and the attention mechanism respectively, and the attention mechanism is output to FM. After FM outputs the user vector tower, the charging station feature information outputs the charging station vector tower through FM. The user vector tower and the charging station vector tower can be calculated in advance and stored in the vector database due to their non-real-time nature. In this way, when the model is inferred online, it is only necessary to calculate the user context vector tower and the charging station context vector tower respectively. The user context vector tower is the output of FM receiving the user context feature information, and the charging station context feature information outputs the charging station context vector tower through FM. The user context vector tower and the charging station context vector tower are spliced ​​with the corresponding stored vectors to obtain the user real-time feature vector and the charging station real-time feature vector. The user real-time feature vector and the charging station real-time feature vector are calculated by MLP to obtain the scores of the candidate recommended stations and users.

[0074] The deep learning ranking model integrates the neural network framework of the FM model, MLP, and attention mechanism, and its idea is similar to Google's Wide & Deep model. The Wide part of the deep learning ranking model is the FM model, and the Deep part is composed of MLP and attention mechanism. The MLP layer uses the ReLu activation function, and the deep learning ranking model uses the Sigmoid activation function as the output function for CTR estimation.

[0075] Specifically, the ReLu activation function:

[0076]

[0077] Specifically, in the deep learning sorting model, the calculation process of FM is as follows:

[0078]

[0079] Among them, n is the number of sample data, x i is the i-th feature, <v i ,v j > is the inner product form of the second-order feature combination parameters, v i and v j is x i and x j The corresponding hidden vectors, w0 and w i First-order weight parameters for deep learning ranking models.

[0080] like Figure 4 As shown, the working method of a smart electric vehicle charging station recommendation system provided by an embodiment of the present invention specifically includes the following steps:

[0081] S101: Obtaining initial data: querying user feature information based on user ID; obtaining user context feature information based on current time and location information;

[0082] S102: Multi-channel recall: According to the location in the user context feature information, query the top k nearest stations as the first recall result, query the top stations with the largest number of charging users in the current city as the second recall result, and query the top k stations with the largest number of charging users in the current administrative area as the third recall result; according to the location and time in the user context feature information, query the top k nearby charging stations with the highest availability at the current moment as the fourth recall result, and query the top k nearby charging stations with the lowest charging price at the current moment as the fifth recall result; query the top k most interesting charging stations returned by the collaborative filtering algorithm by the user ID as the sixth recall result, input the user feature information and user context feature information into the deep learning recall model, and obtain the top k stations with the highest user similarity as the seventh recall;

[0083] S103: De-duplicating the charging stations recalled by the multi-channel recall method to obtain a set of candidate charging stations;

[0084] S104: Querying the station portrait information of the candidate object and the context feature information of the candidate charging station at the current moment;

[0085] S105: Inputting user feature information, user context feature information, charging station feature information, and charging station context feature information into a deep learning ranking model;

[0086] S106: The deep learning ranking model scores the candidate charging stations;

[0087] S107: Select k charging stations with the highest scores as recommendation results.

Claims

1. A smart electric vehicle charging station recommendation system, characterized in that: It includes a big data processing layer, a multi-channel recall layer, a recall de-duplication layer and a sorting layer; the big data processing layer collects basic information and outputs user feature information, user context feature information, charging station feature information and charging station context feature information; User characteristic information mainly includes user basic information, preference information, statistical information, and user behavior information; Charging station feature information mainly includes basic information of charging stations and statistical information of charging stations; the multi-channel recall layer includes recall based on user context, recall based on user interest and recall based on charging station similarity; The recall based on user interest uses a collaborative filtering algorithm, and the recall based on charging station similarity is obtained through a deep learning recall model. The results of the multi-way recall layer are given to the recall re-weighting layer, and the output of the recall re-weighting layer is given to the sorting layer, which outputs the recommendation results.

2. The intelligent electric vehicle charging station recommendation system as claimed in claim 1, characterized in that: The user context-based recall includes five recalls, namely: recalling the k stations closest to the user based on the distance between the user and the charging station; recalling the k stations with the highest availability near the current moment based on the user's location; recalling the k stations with the largest number of charging users in the city where the current user is located; recalling the k stations with the largest number of charging users in the administrative district where the current user is located; and recalling the k stations with the lowest charging prices near the current moment based on the user's location.

3. The intelligent electric vehicle charging station recommendation system as claimed in claim 2, characterized in that: The recall based on user interest uses collaborative filtering algorithm. The collaborative filtering algorithm includes user-based collaborative filtering and charging station-based collaborative filtering. The specific algorithm of collaborative filtering is as follows: Where v1 and v2 are two different user real-time feature vectors or two different charging station feature vectors; Thus, in user-based collaborative filtering, user u's interest in site s can be expressed as: Where U is the user set, score(u i ,s) is user u i Number of charges at charging station s; In collaborative filtering based on charging stations, user u’s interest in station s can be expressed as: Where S is the set of charging stations, score(u,s i ) is user u at charging station s i The number of charging times.

4. The intelligent electric vehicle charging station recommendation system as claimed in claim 3, characterized in that: The deep learning recall model includes a user vector tower, a charging station vector tower and a user context vector tower. The user vector tower is obtained by passing the user feature information through its own MLP and attention mechanism, the user context feature information is passed through its own MLP to obtain the user context vector tower, and the charging station feature information is passed through its own MLP to obtain the charging station vector tower; the user vector tower and the user context vector tower are both input into the MLP to obtain the user real-time feature vector, the charging station vector tower can obtain the charging station feature vector, and the user real-time feature vector and the charging station feature vector are calculated through the inner product to obtain the charging station similarity.

5. The intelligent electric vehicle charging station recommendation system as claimed in claim 4, characterized in that: The sorting layer uses a deep learning sorting model. The input parameters of the model include user feature information, user context feature information, charging station feature information, and charging station context feature information. Each parameter is passed through its own FM to generate a corresponding vector tower; the user vector tower is spliced ​​with the user context vector tower to obtain the user real-time feature vector, and the charging station vector tower is spliced ​​with the charging station context vector tower to obtain the charging station real-time feature vector. The user real-time feature vector and the charging station real-time feature vector are input into the MLP to calculate the score.

6. The intelligent electric vehicle charging station recommendation system as claimed in claim 5, characterized in that: MLP uses the ReLu activation function. The specific ReLu activation function is: The calculation process of FM is as follows: Among them, n is the number of sample data, x i is the i-th feature, <v i ,v j > is the inner product form of the second-order feature combination parameters, v i and v j is x i and x j The corresponding hidden vectors, w0 and w i First-order weight parameters for deep learning ranking models.

7. A recommendation method for a smart electric vehicle charging station recommendation system, characterized in that: The following steps are involved: S101: Obtaining initial data: querying user feature information based on user ID; obtaining user context feature information based on current time and location information; S102: Multi-channel recall: According to the location in the user context feature information, query the top k nearest stations as the first channel recall result, query the top stations with the largest number of charging users in the current city as the second channel recall result, and query the top k stations with the largest number of charging users in the current administrative area as the third channel recall result; According to the location and time in the user context feature information, query the top k nearby charging stations with the highest availability at the current moment as the fourth recall result, and query the top k nearby charging stations with the lowest charging price at the current moment as the fifth recall result; The top k charging stations that the current user is most interested in are queried by the collaborative filtering algorithm through the user ID as the sixth recall result. The user feature information and user context feature information are input into the deep learning recall model to obtain the top k stations with the highest station similarity as the seventh recall result. S103: De-duplicating the charging stations recalled by the multi-channel recall method to obtain a set of candidate charging stations; S104: Querying the station portrait information of the candidate object and the context feature information of the candidate charging station at the current moment; S105: Inputting user feature information, user context feature information, charging station feature information, and charging station context feature information into a deep learning ranking model; S106: The deep learning ranking model scores the candidate charging stations; S107: Select k charging stations with the highest scores as recommendation results.

Images