Electric energy grid-connected management system based on direct-current stacked storage

Abstract

The invention discloses an electric energy grid-connected management system based on direct current stack storage, and the system comprises a user aggregation information construction module which is used for obtaining the user information of each charging station in a target region, and constructing a user number knowledge graph; the user habit construction module is used for acquiring the charging habit of the user and constructing a charging habit knowledge graph; the fusion module is used for generating a fusion knowledge graph based on the user quantity knowledge graph and the charging habit knowledge graph; the information output module is used for generating an energy charging and discharging plan of all charging stations in the target area according to the fused knowledge graph according to a preset energy charging habit of the charging stations; the charging and discharging plan comprises the charging amount of each time period and the discharging amount of each time period. In the technical scheme provided by the invention, the use frequency of the battery in the charging station can be greatly increased, the window period of the battery is reduced, and the benefit of the charging station is increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of charging station grid connection technology, in particular to an electric energy grid connection management system based on direct current stacking storage. BACKGROUND

[0002] The city charging station generally includes a power acquisition module, a power distribution module and a direct current stacking storage unit. The power acquisition module is used to acquire electric energy from the power grid. The direct current stacking storage unit is used to store electric energy, and the power distribution module is used to distribute electric power to each power consumption end (charging gun). The total power that the power acquisition module can acquire is less than the total power of all power consumption ends working simultaneously due to the construction cost limitation in the construction process of the city charging station. Therefore, in practice, the direct current stacking storage unit needs to supplement the power acquisition module in the power transmission power. The direct current stacking storage can receive or release high current in the use process. Therefore, it can quickly respond to the scheduling demand and timely deliver electric power to the power distribution module.

[0003] The direct current stacking storage charging station in the city has obvious periodic characteristics in the actual use process. During the daytime power peak period, most vehicles are used for production and life, and the number of charging vehicles is small. However, a large number of vehicles will come to charge at night. Therefore, the energy storage system of the city charging station has a long idle period of equipment, which reduces the utilization rate of the equipment. SUMMARY

[0004] The summary part of the present application is used to introduce the concept in a simple form, which will be described in detail in the specific embodiment part. The summary part of the present application is not intended to indicate the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0005] In order to solve the technical problems mentioned in the background part, some embodiments of the present application provide an electric energy grid connection management system based on direct current stacking storage, comprising:

[0006] A user aggregation information construction module is configured to acquire user information of each charging station in a target area and construct a user quantity knowledge graph.

[0007] A user habit construction module is configured to acquire charging habits of users and construct a charging habit knowledge graph.

[0008] A fusion module is configured to generate a fusion knowledge graph based on the user quantity knowledge graph and the charging habit knowledge graph.

[0009] An information output module is configured to generate a charging and discharging plan of all charging stations in the target area according to the fusion knowledge graph based on a pre-set charging station charging habit.

[0010] The charge-discharge plan includes a charging amount of each time period and a discharging amount of each time period.

[0011] The present application can accurately predict the future total charge-discharge demand of all charging stations in the target area by constructing a knowledge graph that fuses regional vehicle aggregation and charging habits; based on this prediction, the system can provide a clear overall charge-discharge plan to the power grid in advance, significantly reducing the pressure on power grid dispatching and reducing the resistance of large-scale energy storage power grid connection; at the same time, this scheme effectively improves the utilization frequency of energy storage batteries in regional charging stations, significantly reduces the idle period of equipment, and significantly enhances the overall operation efficiency of the charging station.

[0012] When constructing the user quantity knowledge graph, user aggregation habits are highly cross-influenced by multiple factors such as time, weather, holidays, and temperature. A static graph constructed only by relying on historical aggregation samples (such as 24-hour station number changes + environmental factors) is difficult to effectively capture and express the deep relationships between these complex factors, resulting in a lack of correlation between graph elements and inaccurate description of user aggregation relationships, ultimately limiting the predictive ability and practicality of the graph.

[0013] Further, the user aggregation information construction module comprises:

[0014] An initial graph construction unit configured to generate a preliminary distribution graph according to input historical user aggregation information;

[0015] A graph information updating unit configured to dynamically update the preliminary distribution graph based on the correlation of adjacent elements to form a user quantity knowledge graph;

[0016] The historical user aggregation information comprises a plurality of aggregation samples, and each aggregation sample comprises user distribution information of each charging station within 24 hours, weather, holidays, and temperature.

[0017] The present application generates a preliminary distribution graph based on historical user aggregation information (containing multiple-dimensional samples such as number of people, weather, holidays, and temperature), and dynamically updates it by introducing the correlation analysis of adjacent elements, effectively solving the problem of missing correlations caused by multiple-factor cross-influence. This dynamic updating mechanism can mine and supplement the cross-factor correlation relationships that do not exist explicitly in the original data, break through the boundary limitations of the initial samples, and ultimately generate a user quantity knowledge graph that is more complete in system, has stronger correlation, and can more accurately reflect the user aggregation rules under the coupling of complex factors, significantly improving the accuracy and practicality of the graph, and laying a solid foundation for subsequent accurate prediction of regional charge-discharge demand.

[0018] The existing knowledge graph construction scheme faces significant challenges in converting multi-source heterogeneous data (such as user changes, weather, holidays, and temperature) into graph elements and relationships. On the one hand, the correspondence between elements and potential relationships in massive data is ambiguous and unclear, making it difficult to automatically and accurately identify and associate them. On the other hand, the large number of label categories leads to a decrease in the accuracy of the element clustering process, making it difficult to effectively merge elements with similar semantics or structures, which severely hinders the ability of the preliminary graph to accurately represent the complex relationship between user distribution information and multi-dimensional environmental factors.

[0019] Further, the preliminary distribution graph construction process includes:

[0020] S1: Clean all historical user aggregation information to remove abnormal data;

[0021] S2: Convert user change information, weather, holidays, and temperature at each charging station within 24 hours into elements;

[0022] S3: Use a model combining LSTM encoding and CRF decoding to identify elements to establish a candidate relationship group set containing element association relationships;

[0023] S4: Merge similar elements to generate a preliminary distribution graph.

[0024] The present application uses a LSTM encoding and CRF decoding joint model to construct a preliminary distribution graph, effectively overcoming the element recognition and association problem: 1) The LSTM-CRF model can deeply capture the potential association patterns of structurally similar elements in sequence data (such as 24-hour user changes), accurately identify the relationships between elements, and generate a high-quality candidate relationship group set; 2) The model's powerful sequence labeling ability significantly improves the accuracy of similar element recognition in a large label space, ensuring the accuracy of the merging process. The preliminary distribution graph generated thereby not only more accurately represents user number distribution information, but also clearly distinguishes and depicts the essential relationships and differences between user change information and weather, holidays, temperature, and other different dimensional factors, laying a high-precision data foundation for subsequent graph dynamic updates.

[0025] In the key step of constructing a user number knowledge graph, establishing the corresponding relationship between user change information and environmental factors such as weather, holidays, and temperature, the core problem is the existence of a significant "many-to-many" mapping relationship in historical user aggregation information (i.e., the same user change pattern can correspond to different environmental combinations, and vice versa). This inherent non-deterministic mapping relationship makes the association between elements (user changes vs. environmental factors) extremely ambiguous and difficult to directly and accurately identify, severely hindering the generation of high-quality candidate relationship groups and becoming a major bottleneck in constructing a knowledge graph that accurately reflects user aggregation patterns.

[0026] The step S3 comprises:

[0027] S31: extracting the corresponding relationship from the historical user aggregation information by a joint extraction model to generate a candidate relationship group set;

[0028] S32: extracting the association probability of the elements in the candidate relationship group set, and re-predicting the association probability of the element pairs by using a BERT relationship classification model, wherein the element is user change information, weather, holiday and temperature, and the association probability is the probability of mutual association between the elements;

[0029] S33: if the newly predicted association probability is consistent with the original association probability, the candidate relationship group set is retained; if not, the association probability is updated according to the classification result.

[0030] The application innovatively adopts a dual verification and correction mechanism of joint extraction model and BERT relationship classification model: 1) the joint extraction model first extracts a preliminary candidate relationship group set from complex historical data; 2) the BERT relationship classification model re-predicts the association probability of the element pairs (user change-environment factor) based on the deep analysis of the association of adjacent aggregation samples and verifies it. The core advantage of this mechanism is that it can effectively identify and dynamically correct the association deviation caused by the small range randomness of historical data (update when the BERT prediction probability is inconsistent with the original probability). Finally, this scheme provides a high-quality relationship data basis for knowledge graph construction, which has been verified twice and has more accurate and reliable association probability, significantly improving the accuracy and reliability of subsequent graph construction.

[0031] In constructing the corresponding relationship between user change information and environmental factors such as weather, holiday and temperature, the core challenge lies in the non-deterministic "many-to-many" mapping commonly existing in historical user aggregation information (i.e. a single user change mode can correspond to multiple environmental combinations, and vice versa). This inherent complexity leads to: (1) the potential association between elements (such as user change vs. weather) is extremely vague and difficult to directly and accurately identify and quantify. (2) Traditional methods are difficult to effectively capture the dynamic up-down dependency characteristics of the relationship between elements (i.e. the association probability needs to be judged in combination with the meaning information of adjacent aggregation samples). This directly restricts the accuracy and reliability of the generated candidate relationship group set, and becomes a key obstacle to constructing high-quality knowledge graphs.

[0032] Further, S31 comprises the following steps:

[0033] S311: inputting historical user aggregation information D, D={C1, C2, C i …C N}, C i represents the i-th aggregation sample in the historical user aggregation information, N represents the total number of aggregation samples, and C i ={e1i e 2i e 3i e 4i}, e 1i e 2i e 3i e 4i Representing the clustered samples C respectively i The first, second, third, and fourth elements are: the first element is user change information, the second element is weather, the third element is holidays, and the fourth element is temperature. The user cluster information D is transformed into a vector using the BERT relationship classification model.

[0034] S312: Aggregate sample C i The input is fed into the TPLinker network to generate a candidate relation set T, where T = {e h ,r,e t )}, where e h and e t These are the forward and backward associated elements, respectively, and r is the association probability.

[0035] Furthermore, step 32 includes the following steps:

[0036] Step S321: Input historical user aggregation information D and candidate relation group set T;

[0037] S322: Positioning e h and e t Extract the clustered sample set E containing forward-related elements and backward-related elements from the historical user clustering information D.

[0038] E = {x s-k ...x s …x e …x e+k}, where s and e represent the starting and ending numbers of the extracted clustered samples, respectively, and k represents the extraction step size;

[0039] S324: Calculate the association probability of corresponding elements;

[0040] P(r|e h e t E) = Softmax(W*h [CLS] +b);

[0041] Where W represents the classification weight matrix, b represents the bias vector, and P(r|e h e t E) represents the condition given e under the constraints of the selected relation set and the clustered sample set. h and e tWhen aggregating the sample set E, r is used as e h and e t The probability of association of effective links between them, where Softmax represents the probability transformation function;

[0042] S325: Retain the set T of candidate relation groups whose association probability is higher than the preset value.

[0043] This application effectively solves the aforementioned challenges through a refined relation construction process that combines the TPLinker network and the BERT relation classification model: the TPLinker network efficiently extracts preliminary candidate relation groups (eh,r,et) from the original aggregated samples, laying the foundation for relation identification. The BERT relation classification model, on the other hand, is based on pairs containing target elements (e...). h e t We perform in-depth analysis on the aggregated sample set E, and recalculate and verify the association probability P(r|eh,et,E).

[0044] The core advantage of this scheme lies in its introduction of a neighbor-sample awareness mechanism (via E) into relation construction. This allows the calculation of association probabilities to dynamically adapt to the inherent randomness and dependency characteristics of neighboring samples in historical data. Thus, by selecting high-probability relation groups, this scheme can not only flexibly construct complex associations with single or multiple elements (precisely controlling relation complexity), but also generate highly accurate, context-aware correspondence graphs, providing a robust and high-precision relational data foundation for subsequent knowledge graph construction.

[0045] When constructing a preliminary distribution map based on multi-source heterogeneous historical user aggregation information (user changes, weather, holidays, temperature), the core challenges faced are map redundancy and high-dimensional complexity: a large number of semantically similar or repetitive elements in the original data (such as similar weather or similar holiday types expressed in different ways) are not effectively identified and merged, resulting in a large set of elements and information redundancy. These redundant elements further generate a large number of invalid or repetitive correspondences. These problems collectively lead to an excessively complex and highly dimensional set of candidate relation groups, which not only significantly increases the difficulty and computational cost of constructing the knowledge graph, but also reduces the clarity, interpretability, and efficiency of subsequent processing of the graph.

[0046] Furthermore, S4 includes the following steps:

[0047] S41: Obtain all elements to get the set A = {e1, e2, ..., e}. n}, where n represents the total number of elements;

[0048] S42: Generate an element database based on historical user aggregation information, and map the elements to the element database;

[0049] S43: Calculate the similarity of each element, and merge the elements with similarity within a preset threshold into one element;

[0050] S44: Update the candidate relationship group set T={(e h , r, e t )} according to the merged elements, and map the updated candidate relationship group set T={(e h , r, e t )} to a low-dimensional space to generate vector representations of each element and relationship;

[0051] S45: Construct a preliminary knowledge graph according to the merged corresponding relationship and the candidate relationship group set.

[0052] The application effectively solves the graph redundancy problem by implementing the structured element merging and relationship cleaning process (S41-S45). The element merging based on semantic similarity (S43) significantly simplifies the element set and eliminates information redundancy. Simultaneous cleaning and updating of the candidate relationship group set (S44) removes redundant or invalid relationships generated by element merging. The low-dimensional vector space mapping (S44) is introduced as a key, which converts the merged elements and relationships into compact vector representations. The core advantage of this scheme is that the dimensionality is reduced from both the element and relationship levels, which systematically reduces the structural complexity and information dimensionality of the knowledge graph. The finally generated preliminary distribution graph not only significantly reduces the construction difficulty, but also has higher information density, stronger interpretability and better computational efficiency.

[0053] The user quantity knowledge graph needs to be generated using the preliminary knowledge graph. In practice, the association relationship in the preliminary knowledge graph is too complex, that is, the constructed user quantity knowledge graph is an initial balanced knowledge graph, which cannot pay too much attention to the relationship between two elements according to the demand. Different extraction pays different attention to the association between different elements, so the preliminary knowledge graph needs to be modified.

[0054] Further, the method for generating the user quantity knowledge graph comprises the following steps:

[0055] Z1: Obtain the preliminary knowledge graph and new historical user aggregation information G0;

[0056] Wherein, G0=(e0, r0, T0), e0, r0, T0 represent elements, relationships, and candidate relationship group sets respectively.

[0057] Z2: Use the TransE model to train the preliminary knowledge graph embedding and calculate the implicit relationship score of each element;

[0058] Score(h, r, t) = γ - ||h + r - t||2; wherein, h represents a vector of forward elements, r represents a vector of correlation probabilities, t represents a vector of backward elements, γ represents a margin parameter, ||h + r - t||2 represents a norm range;

[0059] Z3: supplementing a relationship with an implied relationship higher than a preset value to the preliminary knowledge graph;

[0060] Z4: generating a candidate relationship group set from new historical user aggregation information based on the manners of S31 and S32;

[0061] Z5: comparing the cosine similarity of the candidate relationship group set of the preliminary knowledge graph and the newly extracted candidate relationship group set, merging the candidate relationship group set and the newly extracted candidate relationship group set with a similarity less than a preset value into one candidate relationship group set, and generating a user quantity knowledge graph based on all candidate relationship group sets.

[0062] In the technical scheme provided in the present application, the initial knowledge graph is updated by using special historical user aggregation information, so that the candidate relationship group set with directionality can be added on the basis of the balanced preliminary knowledge graph.

[0063] When reflecting the user aggregation rule (user quantity knowledge graph) and the user charging behavior (charging habit knowledge graph) are fused, the existing method faces significant challenges: 1) there are essential differences in the knowledge structure and dynamic characteristics implied by the two graphs (user aggregation is strongly driven by external environment and changes quickly, and charging habits are relatively stable), making it difficult to effectively align and deeply integrate; 2) traditional fusion strategies (such as simple splicing or static weighting) cannot adaptively capture and weigh the contribution of relationship information in different graph levels (propagation layers), resulting in a rough fusion process, severe information loss, and the final generated fusion graph being difficult to accurately and cooperatively reflect the complex correlation between user distribution information and charging quantity.

[0064] Further, the user habit construction module generates the charging habit knowledge graph by using the preliminary knowledge graph scheme;

[0065] The fusion module fuses the charging habit knowledge graph and the user quantity knowledge graph according to the following steps:

[0066] Step 1: obtaining the propagation layer number and element expansion set of the charging habit knowledge graph;

[0067] Step 2: using neural network to learn the relationship information of the candidate relationship group set in each layer to dynamically allocate weights;

[0068] Step 3: gradually weighting and aggregating the element vectors of each layer to generate hierarchical features;

[0069] Step 4: fuse the features of each layer, construct the element link relationship, and generate a fusion graph according to the element link relationship.

[0070] Further, step 4 specifically includes:

[0071] Step 4 specifically includes:

[0072] Step 41: according to the user quantity knowledge graph and the charging habit knowledge graph, calculate the initial features for describing the user distribution information and the initial features for describing the charging quantity

[0073]

[0074] representing a set of original features related to user distribution information, representing a set of original features related to charging quantity, e represents an element in the set of original features;

[0075] Step 42: generate the initial feature set φ u and the initial feature set φ v for describing the user distribution information and the charging quantity, respectively.

[0076]

[0077] Step 43: establish a link relationship between the initial feature set for describing the user distribution information and the initial feature set for describing the charging quantity.

[0078] Step 44: fuse the charging habit knowledge graph and the user quantity knowledge graph based on the link relationship between the initial feature set for describing the user distribution information and the initial feature set for describing the charging quantity to generate a fusion graph.

[0079] The application effectively solves the fusion problem of heterogeneous knowledge graphs through an innovative hierarchical dynamic weighted fusion and double feature set linking mechanism: 1) hierarchical learning and dynamic weight distribution (steps 2-3): the information of the candidate relationship group in each propagation layer is learned by using a neural network, and the fusion weight is dynamically distributed to ensure the differentiated and adaptive integration of knowledge with different levels and different dynamic characteristics; 2) double feature set linking (steps 41-43): the link relationship between the initial feature set describing the user distribution information and the initial feature set describing the charging amount is extracted and explicitly established, providing a structured expression basis for core association. This scheme realizes the fine and hierarchical fusion of user aggregation dynamic information and stable charging habit information. The finally generated fusion knowledge graph not only significantly improves the accuracy of the fusion of the two kinds of data, but also accurately and cooperatively depicts the deep correlation between the user distribution information and the charging amount in the target area, providing solid and reliable data support for subsequent generation of high-precision regional charging and energy plans. BRIEF DESCRIPTION OF DRAWINGS

[0080] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The drawings illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application.

[0081] In addition, throughout the drawings, same or similar reference numerals are used to denote same or similar elements. It should be understood that the drawings are schematic, and the elements and elements are not necessarily drawn to scale.

[0082] In the drawings:

[0083] Figure 1 is a schematic view of the structure of a charging station.

[0084] Figure 2 is a schematic view of the structure of a direct current stacking energy grid-connected management system. DETAILED DESCRIPTION

[0085] Embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of protection of the present application.

[0086] In addition, it should be further noted that only the parts related to the invention are shown in the drawings for ease of description. The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0087] The application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0088] With reference to Figure 1 The direct-current charging station of the application generally comprises a direct-current storage unit, a grid-connected unit, a management unit and a charging unit. The grid-connected unit and the direct-current storage unit are connected with the charging unit. The charging unit is used to supply power to vehicles entering the charging station; the grid-connected unit is connected with the power grid and is used to obtain power from the power grid and deliver it to the charging unit; the direct-current storage unit is used to store power and bear part of the load when the load of the charging unit is high. The power of the direct-current storage unit is provided by the power grid.

[0089] The specific operation mode is as follows: when the electricity price is low at night, the direct-current storage unit obtains power from the power grid and stores the power; when the electricity price is high during the day, the direct-current storage unit can feed power back to the power grid, thereby obtaining the price difference income. In addition, when there are many vehicles connected to the charging unit and the grid-connected unit cannot obtain enough high charging power from the power grid at one time, the direct-current storage unit can also deliver power to the charging unit to supplement the remaining load gap of the charging unit.

[0090] Therefore, the energy storage unit (direct-current storage unit) of the conventional charging station has the functions of improving the load capacity and storing power. However, due to the large fluctuation in the number of charging vehicles and the serious disordered charging behavior, the energy storage unit only serves as a supplementary source of power for the charging unit most of the time, and it is difficult to realize large-scale and stable reverse power transmission to the power grid. The power grid itself usually also limits the access of such unstable power transmission units.

[0091] Based on this, with reference to Figure 2 A direct-current storage-based power grid management system comprises a user aggregation information construction module, a user habit construction module, a fusion module and an information output module. The user aggregation information construction module is used to obtain user information of each charging station in a target area and construct a user quantity knowledge graph; the user habit construction module is used to obtain charging habits of users and construct a charging habit knowledge graph; the fusion module generates a fusion knowledge graph based on the user quantity knowledge graph and the charging habit knowledge graph; the information output module generates a charging and discharging plan of all charging stations in the target area according to the fusion knowledge graph based on a pre-set charging habit of the charging station; the charging and discharging plan comprises a charging amount of each time period and a discharging amount of each time period.

[0092] In grid management, if the stored power is to be input into the grid, a request needs to be sent to the grid in advance to reduce the scheduling pressure of the grid. Otherwise, small-scale grid connection requirements may not be able to obtain a response from the grid. However, the power storage capacity of a single charging station is limited, and the power transmission is unstable, so it is difficult to independently initiate a grid connection application to the grid. Based on this, the application combines multiple charging stations in the target area into a whole to initiate a grid connection application. In this way, the combined charging station cluster has stronger energy storage capacity, and the time point of sending power supply demand to the grid and the power supply quantity prediction are more accurate, thereby ensuring that the charging station energy storage battery obtains more charge and discharge cycle times and reduces its idle time.

[0093] In actual grid management, whether a charging station supplies power to the grid and the size of the power supply are influenced by various factors such as electricity price, battery life, and charging station enterprise management strategy. A too rigid grid connection plan is not conducive to the flexible adjustment of the enterprise's own strategy. Therefore, instead of directly calculating the operation benefit of the charging station, the application fuses user aggregation and user charging conditions to construct a knowledge graph. After obtaining the knowledge graph that completely depicts external factors, the enterprise can adjust the charge and discharge plan according to the charging habit set by itself.

[0094] That is, the enterprise's own internal factors (management strategy) are used to set the charging habit, and the external factors (user behavior and environment) are used to generate the fused knowledge graph; after combining the fused knowledge graph and the charging habit, the charge and discharge plan can be generated. The charge and discharge plan can accurately guide the charging station when to be used for vehicle charging and when to be used for supplying power to the grid. In the application, the key lies in generating the fused knowledge graph. After generating the fused knowledge graph, the information input module can quickly locate the corresponding elements and corresponding relationships related to the charging habit input by the user from the fused knowledge graph, thereby generating the required preselected information, and generating the charge and discharge plan based on the preselected information.

[0095] As can be seen, the key of the application lies in how to generate the fused knowledge graph, and the fused knowledge graph needs to generate a user quantity knowledge graph and a charging habit knowledge graph first. The following is the generation process of the user quantity knowledge graph and the charging habit knowledge graph.

[0096] The user aggregation information construction module is used to generate the user quantity knowledge graph. The user aggregation information construction module includes an initial graph construction unit and a graph information updating unit.

[0097] The initial graph construction unit is used to generate a preliminary distribution graph according to the input historical user aggregation information.

[0098] The graph information updating unit dynamically updates the preliminary distribution graph based on the correlation of adjacent elements to form the user quantity knowledge graph.

[0099] The historical user aggregation information includes a plurality of aggregation samples, and each aggregation sample includes user distribution information of each charging station in 24 hours, weather, holidays, and air temperature. The user distribution information, the weather, the holidays, and the air temperature are elements of the aggregation sample.

[0100] The user distribution information, the weather, the holidays, and the air temperature are in a coded form. The user distribution information is the number of charging users of each charging station in the target area in each hour in 24 hours. The user distribution information can reflect the aggregation information of users in the target area at different time periods. The weather is the weather with the longest proportion in the daytime of the day. The weather includes weather labels such as rain, overcast, sun, and cloudy. The air temperature mainly refers to the average temperature of the day. The holidays mainly refer to whether it is a weekday, for example, weekdays, weekends, winter vacation, summer vacation, National Day, and the like. The holidays and the weather are discrete label coding, and the air temperature is continuous coding. The three elements of the weather, the holidays, and the air temperature will affect the user distribution information. In addition, the number of shopping malls and the number of companies in the target area can also be introduced. The present application considers the weather, the holidays, and the air temperature. The three factors are most relevant to the user's power consumption habits, so they need to be considered first. Based on this, a preliminary distribution atlas is constructed, which can directly show the relationship between the user distribution information, the weather, the holidays, and the air temperature.

[0101] The conversion of the user distribution information, the weather, the holidays, and the air temperature into codes is prior art. The weather, the holidays, and the air temperature are one-dimensional data, and the specific conversion method is not described here.

[0102] The user distribution information is three-dimensional data, including time, quantity, and charging station coordinates. Therefore, it is converted into a three-dimensional data matrix, and then the three-dimensional data matrix is coded. For example, the three-dimensional user distribution information (time, quantity, and charging station coordinates) is structured and represented as a three-dimensional data matrix (dimension: time point x charging station position x user quantity), and a coding technology (such as a neural network, tensor decomposition, etc.) is used to automatically learn and extract the spatiotemporal distribution features contained in the matrix, providing a low-dimensional and dense vector representation for subsequent knowledge graph construction or fusion.

[0103] Further, the preliminary distribution atlas construction process includes:

[0104] S1: cleaning all historical user aggregation information to remove abnormal data.

[0105] The abnormal data mainly refers to data lacking key elements, such as lacking weather or lacking air temperature. These incomplete data need to be removed to avoid affecting the subsequent weight.

[0106] S2: converting the user change information of each charging station in 24 hours, the weather, the holidays, and the air temperature into elements.

[0107] The transformation here is actually to encode the user change information, weather, holidays and temperature, and the specific encoding method has been described before.

[0108] S3: An element is identified by a model combining LSTM encoding and CRF decoding to establish a candidate relationship group set containing element association relationships.

[0109] The step S3 includes the following steps:

[0110] S31: Extract the corresponding relationship from the historical user aggregation information by a joint extraction model to generate a candidate relationship group set.

[0111] S31 mainly extracts the corresponding relationship from each element to obtain an element candidate relationship group set. The element candidate relationship group set includes (forward element, relationship, backward element).

[0112] In this application, the forward element is user change information, the backward element is one of "weather, holiday and temperature", and the relationship is the probability of the correspondence between the two.

[0113] S31 includes the following steps:

[0114] S311: Input the historical user aggregation information D, D={C1, C2, C i …C N}, C i represents the i-th aggregation sample in the historical user aggregation information, N represents the total number of aggregation samples, C i ={e 1i , e 2i , e 3i , e 4i}, e 1i , e 2i , e 3i , e 4i respectively represent the 1st element, the 2nd element, the 3rd element, the 4th element in the aggregation sample C i , the 1st element is user change information, the 2nd element is weather, the 3rd element is holiday, and the 4th element is temperature, and the user aggregation information D is converted into a vector using a BERT relationship classification model.

[0115] S312: Input the aggregation sample C i to the TPLinker network to generate a candidate relationship group set T, T={(e h , r, e t )}, wherein e h and e t are forward and backward associated elements, respectively, and r is the association probability.

[0116] In this scheme, the BERT relation classification model is to convert the encoding as a word into a vector. In this scheme, the C i is converted into a vector by the BERT relation classification model, and then the TPLinker network is used to generate the EH-to-ET matrix, the SH-to-OH matrix, and the ST-to-OT matrix, and then the intersection of the above three matrices is analyzed to generate a candidate relationship group set.

[0117] The essential function of the TPLinker network is to identify the corresponding relationship through the TPLinker network. The principle is: from the historical user aggregation information D, the probability that e 1i , e 2i , e 3i , e 4i are located in the same aggregation sample is extracted, and then the correlation probability can be generated, so that the multi-element aggregation sample can be converted into a correlation sample with only three elements, and the data dimension is reduced.

[0118] S32: Extract the correlation probability of the elements in the candidate relationship group set, and use the BERT relation classification model to re-predict the correlation probability of the element pairs, wherein the elements are user change information, weather, holidays, and temperature, and the correlation probability is the probability of the correlation between the elements.

[0119] Step 32 includes the following steps:

[0120] S321: Input the historical user aggregation information D and the candidate relationship group set T.

[0121] S322: Locate the positions of e h and e t in the historical user aggregation information D, and extract the aggregation sample set E containing the forward association elements and the backward association elements;

[0122] E = {x s-k , … x s … x e … x e+k}, wherein s, e respectively represent the starting number and the ending number of the extracted aggregation sample, and k represents the extraction step.

[0123] The ordering in the historical user aggregation information D is not disordered, but is arranged according to the coding order, for example, according to the coding order of the temperature, so there is a great similarity between adjacent aggregation samples in the historical user aggregation information D. Then, the aggregation sample set E containing the forward association elements and the backward association elements is extracted, and in the aggregation sample set E, they are all relatively similar aggregation samples.

[0124] S323: Input the clustered sample set E into the BERT relation classification model to obtain the hidden state h of CLS. [CLS] h [CLS] ∈R;

[0125] Input=[[CLS]]⊕Context⊕[[SEP]];

[0126] Where ⊕ represents the concatenation operation, Context represents the aggregated sample set E, [CLS] represents the classification label, located at the beginning number of the input sequence, and [SEP] represents the separator, located at the end number;

[0127] S324: Calculate the association probability of corresponding elements;

[0128] P(r|e h e t E) = Softmax(W*h [CLS] +b);

[0129] Where W represents the classification weight matrix, b represents the bias vector, and P(r|e h e t E) represents the condition given e under the constraints of the selected relation set and the clustered sample set. h and e t When aggregating the sample set E, r is used as e h and e t The probability of association of effective links between them, where Softmax represents the probability transformation function;

[0130] S325: Retain relationships with an association probability higher than a preset value, including the association probability of element pairs.

[0131] S33: If the newly predicted association probability is consistent with the original association probability, then retain the candidate relationship set; if they are inconsistent, then update the association probability according to the classification result.

[0132] In S31 and S32, S31 first determines which information in the database belongs to the elements, reducing the information dimensionality when calculating the correspondence in the subsequent process. Then, in S2, the Softmax function is used to further calculate the relationships between the elements, thereby obtaining accurate correspondences and updating the candidate relationship set.

[0133] S4: Merge similar elements to generate a preliminary distribution map.

[0134] S4 includes the following steps:

[0135] S41: Obtain all elements to get the set A = {e1, e2, ..., e}. n}, where n represents the total number of elements.

[0136] In S3, all candidate relation group sets are obtained, and each element in the candidate relation group sets is extracted to obtain an element set. The words in the element set are words that need to be merged based on similarity.

[0137] S42: An element database is generated based on historical user aggregation information, and elements are mapped to the element database.

[0138] The element database here is actually a new vector representation obtained by mapping each element to the element database according to the repetition rate of each element in the historical user aggregation information.

[0139] S43: Similarity of each element is calculated, and elements with similarity within a preset threshold are merged into one element.

[0140] S44: The candidate relation group set T={(e h , r, e t )} is updated according to the merged elements, and the updated candidate relation group set T={(e h , r, e t )} is mapped to a low-dimensional space to generate a vector representation of each element and relation. In this way, by calculating the cosine similarity of the elements and then comparing the similarity threshold between the elements, it can be determined whether to keep the elements. For example, the cosine similarity of element A and element B is greater than the preset threshold, so element A and element B represent the same meaning, and element A and element B are merged into one element.

[0141] S45: A preliminary knowledge graph is constructed based on the merged corresponding relation and the candidate relation group set.

[0142] After the elements are merged in S44, the number of candidate relation group sets can be further reduced. Then all candidate relation group sets are used to construct a preliminary knowledge graph. The construction of the knowledge graph using the candidate relation group set is a prior art, and therefore the specific construction scheme will not be further described here.

[0143] The method for generating the user number knowledge graph includes the following steps:

[0144] Z1: Obtain new corpus information G0 in the preliminary knowledge graph;

[0145] Wherein, G0=(e0, r0, T0), e0, r0, T0 represent elements, relations, and candidate relation group sets, respectively.

[0146] The new historical user aggregation information is actually the extracted historical user aggregation information containing T0, and essentially it is necessary to further analyze the relationship between e0 and T0.

[0147] For example: T0 represents the association degree of user change information a and weather b is 20%, and G0 is to take T0 as a whole to analyze the corresponding relationship between temperature and T0 again.

[0148] Z2: Use the TransE model to train the preliminary knowledge graph embedding and calculate the implicit relationship score of each element;

[0149] Score(h, r, t) = γ-||h+r-t||2; wherein h represents the vector of the forward element, r represents the vector of the association probability, t represents the vector of the backward element, γ represents the edge margin parameter, and ||h+r-t||2 represents the norm range;

[0150] Z3: The relationship with an implicit relationship higher than the preset value is supplemented to the preliminary knowledge graph;

[0151] Z4: Based on the S31 and S32 methods, a candidate relationship group set is generated from the new corpus information;

[0152] Z5: The cosine similarity of the candidate relationship group set of the preliminary knowledge graph and the newly extracted candidate relationship group set is compared, the candidate relationship group set with a similarity less than the preset value and the newly extracted candidate relationship group set are merged into one candidate relationship group set, and a user quantity knowledge graph is generated based on all candidate relationship group sets.

[0153] The user habit construction module generates a charging habit knowledge graph using the preliminary knowledge graph scheme.

[0154] The user habit construction module, the key lies in the need to extract the average power consumption of the user from the charging data of the user. For example, when the temperature is low, the vehicle power consumption is large, and when the vehicle reaches the charging station, the battery power is less, and the charging capacity of each vehicle is more. When constructing the charging habit knowledge graph, the data obtained is the information of the charging capacity dimension based on the historical charging data. Therefore, the structure of the charging habit knowledge graph and the user quantity knowledge graph is the same.

[0155] Therefore, the charging habit knowledge graph and the user quantity knowledge graph are essentially highly related, but these relationships are difficult to integrate into the user quantity knowledge graph. Therefore, it is necessary to generate the link relationship of the elements. Specifically, the following steps are included:

[0156] The fusion module fuses the charging habit knowledge graph and the user quantity knowledge graph according to the following steps:

[0157] Step 1: Obtain the propagation layer number and element expansion set of the charging habit knowledge graph.

[0158]

[0159] wherein, denotes a candidate relation group set of the charging habit knowledge graph, denotes an element set of the l-1th layer, L denotes the number of layers of the charging habit knowledge graph, and l denotes an index of the number of layers in the charging habit knowledge graph;

[0160] Step 2: learning the relation information of the candidate relation group set in each layer by using a neural network to dynamically assign weights;

[0161] Specifically, the attention weight of each element is calculated as follows:

[0162] z0=ReLU(W1(h i ||r i ));

[0163] π(h i ,r i )=Sigmoid(W3·ReLU(W2z0));

[0164] wherein, h i denotes a forward element, r i denotes a relation vector, t i denotes a backward element, W1, W2, and W3 respectively denote three different specifications of learnable matrices, || denotes vector concatenation, ReLu denotes a linear rectification activation function, and Sigmoid denotes an S activation function;

[0165] Step 3: gradually weighting and aggregating the element vectors of each layer to generate hierarchical features;

[0166]

[0167] wherein, α i =π(h i ,r i ), denotes the size of the candidate relation group set of the lth layer, and α i denotes the normalized attention weight.

[0168] Step 4: fusing the features of each layer, constructing an element link relation, and generating a fused graph according to the element link relation.

[0169] Step 4 specifically includes:

[0170] Step 41: calculating initial featuresused to describe user distribution information and initial features used to describe charging volume according to the user quantity knowledge graph and the charging habit knowledge graph.

[0171]

[0172] a set of original features related to user distribution information, a set of original features related to charging amount, e represents an element in the set of original features.

[0173] The set of original features related to user distribution information refers to the underlying features related to user distribution information, such as weather, temperature.

[0174] Step 42: generating an initial feature set φ describing user distribution information u and an initial feature set φ describing charging amount v ;

[0175]

[0176] Step 43: establishing a link between the initial feature set describing user distribution information and the initial feature set describing charging amount.

[0177] Step 44: based on the link between the initial feature set describing user distribution information and the initial feature set describing charging amount, fusing the charging habit knowledge graph and the user number knowledge graph to generate a fused graph.

[0178] The information output module generates a charging and discharging plan based on the fused graph. For example: in order to obtain greater economic benefits and reduce the power load service to users, the charging station will reduce the maximum output power during the peak period of electricity price, and let the direct current stacking unit supply power to the power grid to increase economic benefits.

[0179] In this way, after setting the charging habits of all charging stations in the target area, the system can accurately predict the most likely user number change trend and charging demand in the corresponding period based on the fused knowledge graph by inputting real-time temperature, weather and holiday information. Then, combined with the time-of-use electricity price strategy of the power grid (especially the valley period price), an optimized charging and discharging plan can be generated.

[0180] The above description is only some of the preferred embodiments of the present application and the explanation of the technical principles applied. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with the technical features disclosed in the embodiments of the present application (but not limited to) having similar functions to form technical solutions.

Claims

1. A power grid-connected management system based on DC superimposed energy storage, characterized in that, include: The user aggregation information construction module is used to obtain user information of each charging station in the target area and construct a user quantity knowledge graph. The user habit building module is used to obtain users' charging habits and build a charging habit knowledge graph. The fusion module generates a fused knowledge graph based on the user quantity knowledge graph and the charging habit knowledge graph. The information output module generates charging and discharging plans for all charging stations within the target area based on pre-set charging station charging habits and a fused knowledge graph. The charge / discharge plan includes the amount of charge and discharge for each time period.

2. The power grid connection management system based on DC superimposed energy storage according to claim 1, characterized in that: The user aggregation information construction module includes: The initial map construction unit is used to generate a preliminary distribution map based on the input historical user aggregation information. The graph information update unit dynamically updates the preliminary distribution graph based on the correlation between adjacent elements, forming a user quantity knowledge graph. The historical user aggregation information includes several aggregation samples. Each aggregation sample includes user distribution information, weather, holidays, and temperature for each charging station within 24 hours.

3. The power grid connection management system based on DC superimposed energy storage according to claim 2, characterized in that: The preliminary distribution map construction process includes: S1: Clean all historical user aggregation information and remove abnormal data; S2: Convert user change information, weather, holidays, and temperature at each charging station within 24 hours into elements; S3: A model combining LSTM encoding and CRF decoding is used to identify elements in order to establish a set of candidate relation groups containing element associations; S4: Merge similar elements to generate a preliminary distribution map.

4. The power grid connection management system based on DC superimposed energy storage according to claim 3, characterized in that: Step S3 includes: S31: Extract corresponding relationships from historical user aggregation information using a joint extraction model to generate a candidate relationship set; S32: Extract the association probability of elements in the candidate relationship group set, and use the BERT relationship classification model to re-predict the association probability of element pairs. Here, the elements are user change information, weather, holidays and temperature, and the association probability is the probability that the elements are related to each other. S33: If the newly predicted association probability is consistent with the original association probability, then retain the candidate relationship set; if they are inconsistent, then update the association probability according to the classification result.

5. The power grid connection management system based on DC superimposed energy storage according to claim 4, characterized in that: S31 includes the following steps: S311: Input historical user aggregation information D, D = {C1, C2, C...} i …C N }, C i Let N represent the i-th clustered sample in the historical user clustering information, and let C represent the total number of clustered samples. i ={e 1i e 2i e 3i e 4i }, e 1i e 2i e 3i e 4i Representing the clustered samples C respectively i The first, second, third, and fourth elements are: the first element is user change information, the second element is weather, the third element is holidays, and the fourth element is temperature. The user cluster information D is transformed into a vector using the BERT relationship classification model. S312: Aggregate sample C i The input is fed into the TPLinker network to generate a candidate relation group set T, T = {e h ,r,e t )}, where e h and e t These are the forward and backward associated elements, respectively, and r is the association probability.

6. The power grid connection management system based on DC superimposed energy storage according to claim 5, characterized in that: Step 32 includes the following steps: Step S321: Input historical user aggregation information D and candidate relation group set T; S322: Positioning e h and e t Extract the clustered sample set E containing forward-related elements and backward-related elements from the historical user clustering information D. E = {x s-k ...x s …x e …x e+k }, where s and e represent the starting and ending numbers of the extracted clustered samples, respectively, and k represents the extraction step size; S324: Calculate the association probability of corresponding elements; P(r|e h ,e t ,E)=Softmax(W*h [CLS] +b); Where W represents the classification weight matrix, b represents the bias vector, and P(r|e h e t E) represents the condition given e under the constraints of the candidate relation set and the clustered sample set. h and e t When aggregating the sample set E, r is used as e h and e t The probability of association of effective links between them, where Softmax represents the probability transformation function; S325: Retain the set T of candidate relation groups whose association probability is higher than the preset value.

7. The power grid connection management system based on DC superimposed energy storage according to claim 6, characterized in that: S4 includes the following steps: S41: Obtain all elements to get the set A = {e1, e2, ..., e}. n }, where n represents the total number of elements; S42: Generate an element database based on historical user aggregation information, and map the elements to the element database; S43: Calculate the similarity of each element, and merge elements with similarity within a preset threshold into 1 element; S44: Update the candidate relation group set T = {(e...} based on the merged elements. h ,r,e t )}, and update the candidate relation set T = {(e h ,r,e t Mapping to a low-dimensional space generates a vector representation of each element and relation; S45: Construct a preliminary knowledge graph based on the merged correspondences and candidate relationship sets.

8. The power grid connection management system based on DC superimposed energy storage according to claim 7, characterized in that: The method for generating the user count knowledge graph includes the following steps: Z1: Obtain a preliminary knowledge graph and new historical user aggregation information G0; Where G0 = (e0, r0, T0), e0, r0, and T0 represent elements, relations, and candidate relation sets, respectively. Z2: Use the TransE model to train the initial knowledge graph embedding and calculate the implicit relationship score of each element; Score(h,r,t)=γ-||h+rt||2; where h represents the vector of forward elements, r represents the vector of association probabilities, t represents the vector of backward elements, γ represents the margin parameter, and ||h+rt||2 represents the norm range; Z3: Add relationships with implicit relationships higher than the preset value to the preliminary knowledge graph; Z4: Based on the methods of S31 and S32, a set of candidate relationship groups is generated from new historical user aggregation information; Z5: Compare the cosine similarity between the candidate relation group set of the preliminary knowledge graph and the newly extracted candidate relation group set. Merge the candidate relation group set with a similarity less than the preset value and the newly extracted candidate relation group set into one candidate relation group set. Generate a user number knowledge graph based on all candidate relation group sets.

9. The power grid connection management system based on DC superimposed energy storage according to claim 8, characterized in that: The user habit building module uses a preliminary knowledge graph approach to generate a charging habit knowledge graph; The fusion module integrates the charging habit knowledge graph and the user number knowledge graph according to the following steps: Step 1: Obtain the propagation layers and element expansion set of the charging habit knowledge graph; Step 2: Use a neural network to learn the relational information of candidate relation groups in each layer, and dynamically allocate weights; Step 3: Gradually weight and aggregate the element vectors of each layer to generate hierarchical features; Step 4: Integrate the features of each layer, construct the element link relationships, and generate a fusion map based on the element link relationships.

10. The power grid connection management system based on DC superimposed energy storage according to claim 9, characterized in that: Step 4 specifically includes: Step 41: Calculate the initial features used to describe user distribution information based on the user quantity knowledge graph and the charging habit knowledge graph. and initial features used to describe the amount of charge This represents the original feature set related to user distribution information. Let e ​​represent the original feature set related to the charging amount, and let 'e' represent an element in the original feature set. Step 42: Generate an initial feature set φ describing user distribution information. u and the initial feature set φ describing the amount of charge v ; Step 43: Establish the link between the initial feature set describing user distribution information and the initial feature set describing charging amount; Step 44: Based on the link relationships between the initial feature set describing user distribution information and the initial feature set describing charging amount, fuse the charging habit knowledge graph and the user quantity knowledge graph to generate a fused graph.

Images