New energy charging SaaS operation management method based on artificial intelligence

Abstract

The invention discloses a new energy charging SaaS operation management method based on artificial intelligence. The method comprises the steps of collecting new energy charging facility data and generating a structured data set; constructing a charging state map to form a map structure sequence; inputting an improved space-time residual image convolutional network, and outputting an embedded vector; a price strategy is generated by combining a multi-source information input condition variational auto-encoder; inputting the equipment state embedded vector and the operation sequence into an improved Transform model, identifying the health state of the equipment and generating a maintenance suggestion; and packaging a final result into a service module, deploying the service module to an SaaS platform, and providing multi-end calling. The new energy charging SaaS operation management method is realized by fusing an improved space-time residual image convolutional network and introducing a Transform structure of a gating fusion mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy charging operation management automation, and particularly relates to a new energy charging SaaS operation management method based on artificial intelligence. BACKGROUND

[0002] Under the background of the rapid development of the current electric vehicle industry, as a key link to support the operation of electric transportation systems, the intelligent operation level of new energy charging infrastructure has become a core factor restricting industry efficiency and user experience. In order to improve charging pile utilization, reduce operation and maintenance costs, and optimize user charging behavior, many enterprises and research institutions have carried out research on charging network scheduling, user behavior modeling, equipment fault detection, and price strategy optimization. At present, the mainstream technical solutions generally rely on rule-based engines or traditional machine learning models, combined with fixed time granularity data collection and centralized strategy publishing mechanisms, to preliminarily intelligently manage new energy charging systems.

[0003] In terms of data modeling, some methods have attempted to introduce graph neural networks to model the topological distribution of charging piles, device connectivity, and user paths, or use time series convolution networks, long short-term memory networks, etc. to predict load changes and user behavior. However, these methods mostly separate spatial and temporal modeling processes, and cannot model dynamic relationships in charging behavior. At the same time, they lack flexible response mechanisms for sudden device abnormalities or dramatic fluctuations in charging status over time, resulting in insufficient state identification accuracy. In addition, in terms of device state identification and maintenance recommendations, existing methods often rely on threshold-based alarm systems, which cannot capture device sub-health states or trend faults in advance, making it difficult to support predictive maintenance needs in large-scale operations.

[0004] In terms of personalized pricing, current commercial platforms generally use discount methods based on static segments or user levels, lacking the ability to model deep user individual behavior characteristics, making it difficult to balance revenue optimization and user satisfaction. Some research attempts to introduce reinforcement learning models for pricing strategy generation, but due to poor strategy generalization ability and uncontrollable user response characteristics, it is difficult to promote in practical applications. In terms of operation scheduling, traditional platforms mostly use static scheduling tables or manual adjustment based on rules, which cannot jointly model and optimize multi-role operation goals, lacking flexibility and adaptability.

[0005] In addition, the existing SaaS type charging platforms mostly focus on data display and basic control functions, and have limited support for intelligent functions such as multi-source data fusion, real-time modeling and strategy feedback, and there is a split between various business modules, making it difficult to realize closed-loop operation management driven by strategies. In terms of system structure, there is a lack of unified modeling entry and standardized strategy output mechanism, which is not conducive to supporting efficient collaboration between multiple roles such as operators, users and device manufacturers.

[0006] In summary, the existing new energy charging operation management method has significant deficiencies in data space-time modeling, personalized strategy generation, device health management and multi-role collaboration, and is difficult to meet the actual needs of current charging site large-scale deployment and operation efficiency improvement. In particular, there is a lack of unified space-time perception capability at the modeling level, resulting in scattered strategy generation and slow response, low precision of device anomaly detection, lagging maintenance response, and lack of intelligent linkage mechanism in platform management, which has become a core bottleneck restricting the performance improvement of existing systems.

[0007] Therefore, how to provide a new energy charging SaaS operation management method based on artificial intelligence is a problem that those skilled in the art need to solve. SUMMARY

[0008] One object of the present application is to provide a new energy charging SaaS operation management method based on artificial intelligence. The present application fully integrates structured data modeling, space-time graph neural network, personalized generation model and device state recognition technology, and describes in detail the joint modeling of device state and user behavior through a space-time residual graph convolution network, the generation of personalized charging price strategies using a conditional variational autoencoder, and the introduction of an improved Transformer structure to realize the whole process of device health state evaluation and predictive maintenance, with the advantages of accurate strategy response, intelligent device management, user experience optimization and platform scheduling collaboration.

[0009] According to the new energy charging SaaS operation management method based on artificial intelligence, the method comprises the following steps:

[0010] Collecting data of new energy charging facilities and performing data preprocessing to generate a structured charging operation dataset;

[0011] Based on the structured charging operation dataset, a charging state atlas is constructed, edge connection relationships are established according to geographical adjacency relationships, user use association relationships and load conduction relationships, and a graph structure sequence is formed according to time sequence;

[0012] The graph structure sequence is input into an improved space-time residual graph convolution network model for modeling, and device state embedding vectors and user behavior embedding vectors are output;

[0013] According to the user behavior embedding vector and the user historical charging record, a user portrait data is constructed, current period electricity price, power grid load state and user portrait information are combined, and the combined features are input into a conditional variational autoencoder model to output an individualized charging price strategy;

[0014] The device state embedding vector and the device running time sequence are input into an improved time sequence modeling network based on the Transformer structure, short-term fluctuations and long-term trend features of the device running state are extracted through a multi-head attention mechanism, device health state recognition and risk level evaluation are performed, and a predictive maintenance suggestion is generated;

[0015] The individualized charging price strategy, the device health state evaluation result and the user portrait data are packaged into a standardized service module and deployed to a new energy charging SaaS platform, which is provided to the operation management end, the user end and the device supplier end through the platform interface.

[0016] Optionally, the data of the new energy charging facility is collected, and data preprocessing is performed to generate a structured charging operation dataset, which specifically includes:

[0017] An edge data collection device is deployed on site at the new energy charging facility to collect device operation data of the charging pile in real time, and the device operation data includes voltage, current, power, working state code, running temperature, running time, interface state and device exception identifier;

[0018] User charging behavior data is collected synchronously, including user identification information, charging start and end time, selected charging mode, historical charging behavior data and transaction record data, and the data is time-labeled and bound to the corresponding device operation data;

[0019] Environment data is collected from the power grid interface and external data sources, including current period power grid load information, electricity price floating curve, meteorological data and regional traffic congestion indicators, and is uniformly associated to the corresponding site and time window;

[0020] Field standardization and data type conversion operations are performed on the collected device operation data, user charging behavior data and environment data, a uniform field structure table is established, and resampling and time sequence alignment are performed according to a uniform time stamp granularity;

[0021] The time series aligned data is subjected to missing value filling, repeated value elimination and abnormal value identification processing, and the continuous data segments are subjected to slicing processing in a sliding window segmentation reconstruction manner to form a structured charging operation dataset with time sequence attributes.

[0022] Optionally, the structured charging operation dataset is used to construct a charging state atlas, edge connection relationships are established according to geographical adjacency relationships, user use correlation relationships and load conduction relationships, and a graph structure sequence is formed according to a time sequence, and the method specifically comprises the following steps:

[0023] Based on the equipment operation data and the user charging behavior data in the structured charging operation dataset, an initial node set of the charging state atlas is constructed, the initial node set comprises charging pile equipment nodes and user behavior nodes, and each node is assigned a unique identifier and an attribute vector;

[0024] According to the geographical adjacency relationship between the equipment, the power grid access topological structure and the load conduction path, the spatial correlation edge connection relationship between the equipment nodes is constructed, and according to the continuous charging behavior path of the user between the equipment, the behavior interaction edge connection relationship between the user behavior nodes and the charging pile equipment nodes is constructed;

[0025] According to a time sequence index, the structured charging operation dataset is divided into continuous time segments, and a graph structure with a time stamp is constructed for each time segment to form a graph structure sequence of the charging state atlas;

[0026] The node set and the edge set in the graph structure sequence are normalized, and graph structure sparsification, edge weight normalization and node attribute vector encoding operations are performed to obtain a graph structure sequence with time sequence characteristics and structure attributes.

[0027] Optionally, the graph structure sequence is input into an improved spatio-temporal residual graph convolutional network model for modeling, and device state embedding vectors and user behavior embedding vectors are output, and the method specifically comprises the following steps:

[0028] The graph structure sequence dataset is input into the improved spatio-temporal residual graph convolutional network model, the graph structure sequence dataset contains a set of graph snapshots arranged in time sequence, each graph snapshot is composed of a node set, an edge set, a node attribute vector and an edge connection weight;

[0029] For each graph snapshot at each time, a spatial graph convolution operation is performed, a normalized adjacency matrix is constructed according to the edge connection weight between the nodes, and the attribute vector of the target node and the attribute vector of the neighbor node are linearly combined and nonlinearly activated through a graph convolution operator to extract the spatial correlation characteristics of the node in the topological structure;

[0030] A cross-layer residual connection structure is introduced after each spatial graph convolution layer, the node representation input by the previous layer and the graph convolution output of the current layer are element-level residual fused, the gradient disappearance and overfitting problems in the deep network are alleviated, and the stability of the node representation in the deep modeling process is enhanced;

[0031] The residual fused node representation is arranged in time index order and input into a time dimension modeling module, in which one-dimensional time convolution operation is performed on the graph snapshot sequence to extract the dynamic trend of node state evolution over time, and a node state maintaining module is embedded in the time convolution layer to maintain the short-term change response ability and long-term state stability of key nodes;

[0032] A graph time alternation processing unit is arranged between the spatial graph convolution operation and the time dimension modeling module, which alternately performs spatial feature extraction of each graph snapshot and time dimension modeling of node state in the sequence, so that the graph structure evolution features and state evolution trend can be jointly captured;

[0033] Through the stacking combination of multiple spatial convolution layers, time convolution layers, cross-layer residual connection structures and graph time alternation processing units, the device state embedding vector of each device node in the current time window and the behavior embedding vector of each user node are output.

[0034] Optionally, the user portrait data is constructed according to the user behavior embedding vector and the user historical charging record, the combined features are input into the conditional variational autoencoder model combined with the current period electricity price, the grid load state and the user portrait information, and the personalized charging price strategy is output.

[0035] Based on the user behavior embedding vector and the historical charging behavior data in the structured charging operation data set, the user charging frequency, charging period distribution, average charging capacity and site preference information are extracted to construct a user portrait representation vector.

[0036] The user portrait representation vector is spliced with the grid load information, the electricity price floating curve and the meteorological data of the current period to construct a joint input feature vector for the pricing task, which is input into the conditional variational autoencoder model.

[0037] In the conditional variational autoencoder model, an encoder module receives the joint input feature vector and extracts a latent pricing preference distribution, outputs a mean vector and a variance vector as hidden space parameters, samples a hidden variable vector and passes it to a decoder module.

[0038] The hidden variable vector is mapped to a personalized charging price strategy output vector by the decoder module, which includes a floating single value, a recommended price adjustment coefficient and a peak-valley response intensity factor for a specified user in the current period.

[0039] The personalized charging price strategy output vector is numerically modified and boundary cropped to generate an executable format of charging pricing strategy scheme, and a pricing mapping pair is formed for each user and charging site, which is provided to the SaaS platform for pricing execution and user display.

[0040] Optionally, the device state embedding vector and the device runtime sequence data are input into the improved Transformer structure-based time series modeling network, short-term fluctuations and long-term trend features of the device running state are extracted through the multi-head attention mechanism, device health state recognition and risk level assessment are performed, and predictive maintenance suggestions are generated, specifically including:

[0041] The device state embedding vector and the runtime sequence of the same device in the structured charging operation data set are time-indexed and aligned, and a joint time series input sequence containing device running state information and historical running indicators is constructed;

[0042] The joint time series input sequence is input into the Transformer structure-based time series modeling network with a gating factor fusion mechanism, time position information is added to the input data of each time step through the position encoding module, and a state input sequence with time dependence is constructed;

[0043] In each Transformer encoding unit, the state dependence relationship between the current time step and other time steps is extracted based on the multi-head attention mechanism, and a gating factor structure is introduced after the output of the feedforward subnetwork to perform weighted fusion operation on the current layer output and the previous layer input;

[0044] During attention calculation, a state mutation perception module is introduced to automatically adjust the attention allocation weight according to the change amplitude between the current state and the previous state;

[0045] The time series features encoded by the Transformer structure are input into the parallel health state recognition branch and the risk trend scoring branch, the first branch outputs the device health category label of the current period, the second branch outputs the future risk score level and trend change indicators, and the complete device health state assessment result is formed;

[0046] Based on the device health state assessment result, the predictive maintenance suggestions are automatically generated, including recommended maintenance time interval, potential fault level and involved maintenance module type;

[0047] The device health state assessment result and the predictive maintenance suggestions are packaged in a structured format and stored synchronously in the device state database.

[0048] Optionally, the personalized charging price strategy, device health state assessment result and user portrait data are packaged as a standardized service module and deployed to a new energy charging SaaS platform, and provided to the operation management end, user end and device supplier end through the platform interface, specifically including:

[0049] Index matching and feature alignment are performed on the personalized charging price strategy, the equipment health state evaluation result, the predictive maintenance suggestion and the user portrait data to construct a joint input data set integrating user, equipment and operation state;

[0050] According to the operation management target, a composite operation strategy target set including a benefit efficiency indicator, a user satisfaction indicator and a system stability indicator is set, and a strategy mapping space is established;

[0051] Based on the joint input data set and the composite operation strategy target set, a strategy generation rule engine is used to schedule and plan the operation behavior of each charging station in the current period, and the scheduling plan content includes dynamic pricing strategy execution instructions, charging pile start-stop state adjustment instructions and maintenance time window configuration instructions;

[0052] For charging pile equipment with a risk of failure, a scheduling suspension instruction and a failure avoidance path are generated according to the risk level and the remaining available time interval marked in the equipment health state evaluation result, and the maintenance suggestion is synchronized and bound to the scheduling strategy;

[0053] The strategy generation result is subjected to effectiveness verification and data consistency check, and the strategy priority is sorted and adjusted according to the user historical feedback data and the current load prediction information, and a final executable dynamic operation strategy instruction set is constructed;

[0054] The dynamic operation strategy instruction set is encapsulated as a service interface that can be called by the SaaS platform, and is pushed to the operation management terminal, the user terminal prompt system and the equipment supplier interface according to the role permission, so that each role can respond and execute the strategy in real time.

[0055] The beneficial effects of the present application are:

[0056] The new energy charging SaaS operation management method based on artificial intelligence proposed by the present application overcomes the problems of state modeling fragmentation, strategy response lag, maintenance relying on artificial experience and insufficient platform collaboration ability in existing charging systems, and has significant technical advantages and practical application value. First, by constructing a structured charging operation data set and introducing a unified graph structure modeling method, the spatial topological relationship and time evolution law between charging pile equipment and user behavior are uniformly expressed, the problem of separate modeling of spatial information and time information in existing methods is solved, and the overall coherence and topological perception ability of charging behavior modeling are enhanced.

[0057] Secondly, the application introduces an innovative space-time residual graph convolution network, which introduces a cross-layer residual structure and a state retention mechanism on the basis of traditional space-time graph modeling, significantly improves the modeling robustness of the model to node state mutation, frequent fluctuation and structural disturbance, and effectively enhances the stability and prediction accuracy in the device state modeling process. This structure enables the system to identify abnormal device operation trends in real-time during actual operation, providing a reliable basis for subsequent predictive maintenance.

[0058] In addition, the application introduces a conditional variational autoencoder model in the pricing strategy generation aspect, combines user portraits with current electricity prices and load states to output strategies, realizes a dynamic pricing mechanism for individual users, makes up for the static and lack of personalization of existing platform pricing strategies, and takes into account the dual goals of platform revenue regulation and user behavior guidance. Further, the application also introduces a residual Transformer structure with a gating factor fusion mechanism in the device state modeling, combines attention mechanisms and state mutation perception capabilities, realizes deep modeling of device time series data and fault trend identification, and improves the accuracy and timeliness of predictive maintenance recommendations.

[0059] Finally, the application integrates the above key modules into a new energy charging SaaS platform, builds a multi-role collaborative management mechanism for operators, users and device manufacturers, supports structured packaging and real-time response of strategy output, and improves the intelligent operation capability and service delivery efficiency of the platform. Overall, the application realizes intelligent optimization of the whole process from state modeling, strategy generation to closed-loop scheduling in the new energy charging operation scenario, significantly improves the reliability, scalability and operation intelligence level of the system. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application are used to explain the application, and do not constitute a limitation of the application. In the drawings:

[0061] Fig. 1 A flowchart of a new energy charging SaaS operation management method based on artificial intelligence proposed by the application;

[0062] Fig. 2 A structure diagram of an improved time series modeling network based on a Transformer structure proposed by the application. DETAILED DESCRIPTION

[0063] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0064] ReferenceFigs. 1-2 A new energy charging SaaS operation management method based on artificial intelligence, comprising:

[0065] Step one: collect data of new energy charging facilities, including charging pile equipment operation data, user charging behavior data, power grid load data, electricity price change information and environmental parameter data, and perform data cleaning, missing completion, format unification and time sequence alignment on the data of new energy charging facilities to generate structured charging operation data set;

[0066] Step two: based on the structured charging operation data set, a charging state atlas containing charging pile equipment nodes and user behavior nodes is constructed, edge connection relationships are established according to geographical adjacency relationship, user use association relationship and load conduction relationship, and a graph structure sequence is formed according to time sequence;

[0067] Step three: input the graph structure sequence into the improved spatio-temporal residual graph convolution network model for modeling, and output the equipment state embedding vector and the user behavior embedding vector;

[0068] Step four: according to the user behavior embedding vector and the user historical charging record, the user portrait data is constructed, the combined features are input into the conditional variational autoencoder model combined with the current period electricity price, the power grid load state and the user portrait information, and the personalized charging price strategy is output;

[0069] Step five: input the equipment state embedding vector and the equipment running time sequence data into the improved time sequence modeling network based on the Transformer structure, extract the short-term fluctuation and long-term trend features of the equipment running state through the multi-head attention mechanism, and perform equipment health state recognition and risk level evaluation to generate predictive maintenance suggestions;

[0070] Step six: encapsulate the personalized charging price strategy, the equipment health state evaluation result and the user portrait data into a standardized service module, and deploy it to the new energy charging SaaS platform, and provide it to the operation management end, the user end and the equipment supplier end through the platform interface.

[0071] The application realizes joint modeling and deep analysis of charging pile equipment state and user behavior information by constructing structured charging operation dataset, fusing improved spatiotemporal residual graph convolution network, conditional variational autoencoder and time series modeling network based on Transformer structure, on the basis of ensuring data time sequence consistency and structure stability, improving the integrity of equipment state embedding expression and user behavior feature extraction. Combined with individual user portrait and dynamic price information, personalized charging price strategy is generated, taking into account platform operation efficiency and user experience. At the same time, the Transformer structure with state mutation perception and gate fusion mechanism can model the fine trend of equipment operation state, realize high-precision health state recognition and predictive maintenance suggestion generation. Finally, the multi-class output result is standardized and packaged and deployed to the new energy charging SaaS platform, realizing intelligent response, closed-loop control and multi-role collaboration of the whole charging service process, with the advantages of accurate modeling, timely response, flexible deployment and strong platform universality.

[0072] In the embodiment, the data of the new energy charging facility includes charging pile equipment operation data, user charging behavior data, power grid load data, price change information and environmental parameter data, and the data of the new energy charging facility is subjected to data cleaning, missing value completion, format unification and time sequence alignment to generate a structured charging operation dataset, which specifically includes:

[0073] An edge data acquisition device is deployed on site of the new energy charging facility to acquire the equipment operation data of the charging pile in real time, and the equipment operation data includes voltage, current, power, working state code, operating temperature, interface state and equipment abnormality identifier.

[0074] User charging behavior data is synchronously acquired, including user identifier information, charging start and end time, selected charging mode, historical charging behavior data and transaction record data, and the user charging behavior data is time label bound with the corresponding equipment operation data.

[0075] Environmental data is acquired from a power grid interface and an external data source, including power grid load information, price fluctuation curve, meteorological data and regional traffic congestion index of a current time period, and is uniformly associated to corresponding sites and time windows.

[0076] Field standardization and data type conversion operations are performed on the acquired equipment operation data, user charging behavior data and environmental data, a uniform field structure table is established, and resampling and time sequence alignment are performed according to a uniform time stamp granularity.

[0077] The time sequence aligned data is subjected to missing value filling, repeated value elimination and abnormal value identification processing, and a sliding window segmentation reconstruction method is used to slice continuous data segments to form a structured charging operation dataset with time sequence attributes.

[0078] The embodiment realizes comprehensive collection and unified management of multi-source heterogeneous data by deploying edge data collection devices on the site of new energy charging facilities, fusing user behavior data, power grid load, electricity price information and environmental parameters, and combining field standardization, time series alignment and sliding window slicing processing mechanisms to generate structured charging operation data sets with high time sequence consistency and clear structure, providing a high-quality, low-latency data foundation for subsequent graph modeling, strategy generation and health assessment, and significantly improving the intelligent level and real-time response capability of charging operation driven by data.

[0079] In the embodiment, based on the structured charging operation data set, a charging state graph containing charging pile device nodes and user behavior nodes is constructed, edge connection relationships are established according to geographical adjacency relationships, user use correlation relationships and load conduction relationships, and a graph structure sequence is formed according to time sequence, which specifically includes:

[0080] Based on the device operation data and user charging behavior data in the structured charging operation data set, an initial node set of the charging state graph is constructed, the node set includes charging pile device nodes and user behavior nodes, and each node is assigned a unique identifier and an attribute vector;

[0081] According to the geographical adjacency relationship between the devices, the power grid access topology structure and the load conduction path, the spatial correlation edge connection relationship between the device nodes is constructed, according to the continuous charging behavior path of the user among multiple devices, the behavior interaction edge connection relationship between the user nodes and the device nodes is constructed, and the edge set of the initial graph structure is formed;

[0082] According to the time sequence index, the structured charging operation data set is divided into continuous time segments, and a graph structure with a timestamp is constructed for each time segment to form a graph structure sequence of the charging state graph;

[0083] The node set and the edge set in the graph structure sequence are normalized, and graph structure sparsification, edge weight normalization and node attribute vector encoding operations are performed to obtain a graph structure sequence with time sequence characteristics and structure attributes.

[0084] The embodiment constructs a charging state graph containing charging pile device nodes and user behavior nodes, establishes multi-type edge connection relationships according to geographical adjacency relationships, user behavior paths and load conduction logic, and generates a graph structure sequence according to time index, which not only realizes comprehensive modeling of spatial correlation and behavior dependence of the charging operation process, but also improves the expression ability and calculation efficiency of the graph structure data through normalization processing such as node attribute encoding and edge weight normalization, providing a high-quality graph structure input basis for subsequent device state evolution modeling and user behavior feature extraction based on graph neural networks.

[0085] In the embodiment, the input of the graph structure sequence into the improved spatio-temporal residual graph convolutional network model is modeled, and the output device state embedding vector and the user behavior embedding vector specifically include:

[0086] The graph structure sequence dataset is input into the improved spatio-temporal residual graph convolutional network model, the graph structure sequence includes a set of graph snapshots arranged in time sequence, each graph snapshot is composed of a set of nodes, a set of edges, a node attribute vector and an edge connection weight;

[0087] For each graph snapshot at each time, a spatial graph convolution operation is performed, a normalized adjacency matrix is constructed according to the edge connection weight between nodes, and a linear combination and a nonlinear activation of the attribute vector of the target node and the attribute vector of the neighbor node are performed by a graph convolution operator to extract the spatial correlation feature of the node under the topological structure;

[0088] A cross-layer residual connection structure is introduced after each spatial graph convolution layer, the node representation input by the previous layer and the graph convolution output of the current layer are element-level residual fused to alleviate the gradient vanishing and overfitting problems in the deep network and enhance the stability of the node representation in the deep modeling process;

[0089] The node representation after residual fusion is arranged in time index order and input into the time dimension modeling module, the time dimension modeling module includes a one-dimensional time convolution layer, a node state maintaining module and a feature normalization activation unit, the one-dimensional time convolution layer performs convolution operation on the state representation of each node in the multi-graph snapshot along the time index dimension, the node state maintaining module is used to maintain the historical state statistical value of each node in the sliding time window and splice with the current state, and the feature normalization activation unit is used to perform batch normalization and activation function processing, the graph snapshot sequence is one-dimensional time convolution operated in the time dimension modeling module, the dynamic trend of the node state evolution with time is extracted, and the node state maintaining module is embedded in the time convolution layer to maintain the short-term change response ability and the long-term state stability of the key node;

[0090] The graph time alternation processing unit is arranged between the spatial graph convolution and the time convolution module, and is composed of an alternation controller, an intermediate feature cache and a fusion update mechanism. The alternation controller is used to set the execution order of the spatial feature extraction stage and the time sequence modeling stage in each round of processing. The intermediate feature cache is used to store the spatial processing output result and provide the time modeling stage with a call. The fusion update mechanism is used to fuse the spatial and time features and update the current node state representation. The graph time alternation processing unit alternately performs spatial feature extraction and time dimension modeling of the sequence of nodes for each graph snapshot. The topological correlation features between nodes are extracted by inputting the graph structure information of each time step into the spatial graph convolution network, and the extracted result is input into the time modeling module to update the node representation in combination with the historical state information. Meanwhile, the updated node state is taken as the input of the spatial convolution of the next graph snapshot, so as to jointly capture the evolution features of the graph structure and the evolution trend of the state.

[0091] Through the stacking combination of the multi-layer spatial convolution layer, the time convolution layer, the cross-layer residual connection structure and the graph time alternation processing unit, the device state embedding vector of each device node and the behavior embedding vector of each user node in the current time window are output.

[0092] In the embodiment, the improved space-time residual graph convolution network model is introduced to perform spatial-time joint modeling operation on the graph structure sequence of the charging state atlas. First, the spatial graph convolution module is used to extract the spatial correlation features of the nodes in the topological structure, and the cross-layer residual connection structure is combined to enhance the transmission stability of the node features in the deep network. Then, the time convolution module is used to model the time evolution features of the node state, and the node state retention module is introduced to improve the modeling ability of the sensitivity and long-term stability of the state change. Meanwhile, the graph time alternation processing unit is arranged to realize ordered and alternating coupling between the spatial feature extraction and the time feature modeling, so that the device state embedding vector and the user behavior embedding vector have stronger dynamic representation ability and global structure perception ability, thereby providing sufficient expression and context consistent input basis for subsequent personalized price strategy generation and device health state evaluation, and effectively improving the prediction accuracy and strategy adaptability of the system as a whole.

[0093] In the embodiment, the user portrait data is constructed according to the user behavior embedding vector and the user historical charging record. The combined features are input into the conditional variational autoencoder model in combination with the current period electricity price, the power grid load state and the user portrait information, and the personalized charging price strategy is output. Specifically, the personalized charging price strategy includes:

[0094] Based on the user behavior embedding vector and the historical charging behavior data in the structured charging operation data set, the user charging frequency, the charging period distribution, the average charging capacity and the site preference information are extracted, and the user portrait representation vector is constructed.

[0095] The user portrait representation vector is spliced with the current period electricity price, power grid load state and weather influence factor to construct a joint input feature vector for the pricing task as the conditional information input into the conditional variational autoencoder model;

[0096] In the conditional variational autoencoder structure, the encoder module receives the joint input feature vector and extracts a latent pricing preference distribution, and outputs a mean vector and a variance vector as hidden space parameters, and a hidden variable vector is sampled and transmitted to the decoder module together with the conditional input;

[0097] The decoder module maps the hidden variable vector into a personalized charging price strategy output vector, which includes a floating single value, a recommended price adjustment coefficient and a peak-valley response intensity factor for the specified user in the current period;

[0098] The personalized charging price strategy output vector is numerically modified and boundary cropped to generate an executable format charging pricing strategy scheme, and a pricing mapping pair is formed for each user and charging site, which is provided to the SaaS platform for pricing execution and user display.

[0099] The embodiment constructs a user portrait representation vector, combines the current period electricity price, power grid load state and external environmental factors, and uses a conditional variational autoencoder model for modeling and reasoning, effectively realizing joint modeling and adaptive matching of user charging preferences and power system state, and outputting a charging price strategy with personalized and dynamic response capability, which not only improves user satisfaction and price acceptance, but also enhances the pricing flexibility and intelligent control capability of the new energy charging operation platform in a complex supply and demand environment.

[0100] In the embodiment, the device state embedding vector and the device operation time series data are input into the improved time series modeling network based on the Transformer structure, the short-term fluctuations and long-term trend features of the device operation state are extracted through the multi-head attention mechanism, and the device health state is identified and the risk level is evaluated, and the predictive maintenance suggestion is generated, which specifically includes:

[0101] The device state embedding vector and the operation time series data of the same device in the structured charging operation data set are time-indexed and aligned, and a joint time series input sequence containing device operation state information and historical operation indicators is constructed;

[0102] The joint time series input sequence is input into the time series modeling network based on the Transformer structure with the introduction of the gating factor fusion mechanism, the position encoding module adds time position information to the input data of each time step, and a state input sequence with time dependence is formed;

[0103] In each layer of the Transformer encoding unit, the state dependence between the current time step and other time steps is extracted based on the multi-head attention mechanism, and a gating factor structure is introduced after the output of the feedforward subnetwork to perform a weighted fusion operation on the current layer output and the previous layer input, thereby enhancing the network's response capability to sudden changes in operating state or non-stationary characteristics;

[0104] During the attention calculation process, a state mutation perception module is introduced to automatically adjust the attention allocation weight according to the change amplitude between the current state and the previous state, so that the model has stronger attention ability at time points with significant state changes;

[0105] The time series features encoded by the Transformer structure are input into parallel health state recognition branches and risk trend scoring branches. The first branch outputs the device health category label for the current period, and the second branch outputs the future risk score level and trend change indicators, forming a complete device health state evaluation result.

[0106] Based on the state fluctuation amplitude, risk score value and state trend, a predictive maintenance recommendation is automatically generated, which includes recommended maintenance time interval, potential failure level and involved maintenance module type.

[0107] The device health state evaluation result and the predictive maintenance recommendation are packaged in a structured format and stored synchronously in the device state database.

[0108] The embodiment aligns the device state embedding vector with the running time series data by time indexing, constructs a joint time series input sequence, and inputs it into an improved Transformer structure-based time series modeling network that introduces a gating factor fusion mechanism and a state mutation perception module. Under the joint action of the multi-head attention mechanism and the position encoding, the short-term fluctuation characteristics and long-term trend information of the device operating state are comprehensively extracted. Through the modeling output of the parallel health state recognition branch and the risk trend scoring branch, the current device health level and the future risk trend are simultaneously predicted, and a predictive maintenance recommendation containing the recommended maintenance time, failure level and maintenance module type is automatically generated. Finally, the evaluation result is packaged in a structured format and written into the device state database. This method significantly improves the fine-grained level of new energy charging facility health management, the real-time and predictive nature of risk early warning, and the execution efficiency of predictive maintenance, and has good practical value and promotion prospects.

[0109] In the embodiment, the personalized charging price strategy, device health state evaluation result and user portrait data are packaged into a standardized service module and deployed to a new energy charging SaaS platform, which is provided to the operation management end, user end and device supplier end through the platform interface, specifically including:

[0110] Index matching and feature alignment are performed on the personalized charging price strategy, the equipment health state evaluation result, the predictive maintenance suggestion, and the user portrait data to construct a joint input data set that integrates users, equipment, and operation states;

[0111] According to the operation management target, a composite operation strategy target set including a benefit efficiency index, a user satisfaction index, and a system stability index is set, and a strategy mapping space is established;

[0112] Based on the joint input data set and the composite operation strategy target set, a strategy generation rule engine is used to schedule and plan the operation behavior of each charging station in the current period, and the scheduling plan content includes a dynamic pricing strategy execution instruction, a charging pile start-stop state adjustment instruction, and a maintenance time window configuration instruction;

[0113] For charging pile equipment at risk of failure, a scheduling suspension instruction and a failure avoidance path are preferentially generated according to the risk level and the remaining available time interval marked in the equipment health state evaluation result, and a maintenance suggestion is synchronously bound to the scheduling strategy;

[0114] The strategy generation result is subjected to effectiveness verification and data consistency check, and the strategy priority is sorted and adjusted according to the user historical feedback data and the current load prediction information to construct a final executable dynamic operation strategy instruction set;

[0115] The dynamic operation strategy instruction set is encapsulated as a service interface that can be called by the SaaS platform, and is pushed to the operation management terminal, the user terminal prompt system, and the equipment supplier interface according to the role permission, for real-time scheduling response and strategy execution by each role.

[0116] The embodiment unifies the personalized charging price strategy, the equipment health state evaluation result, and the user portrait data, forms a standardized service module, and deploys it to the new energy charging SaaS platform, effectively realizing the multi-role cooperation of the user side, the equipment side, and the operation side. By setting a composite operation target and combining a strategy generation rule engine, the system can dynamically generate multi-dimensional scheduling instructions such as pricing, start-stop, and maintenance according to the real-time operation state, improving the flexibility and execution efficiency of the operation strategy. At the same time, different roles can obtain the corresponding strategy through the platform interface, ensuring the accuracy and timeliness of service distribution, and further enhancing the adaptability and intelligent management level of the platform to complex operation scenarios.

[0117] Embodiment 1:

[0118] In the operation and management practice of new energy public charging stations in some places, the charging station has more than 50 new energy charging piles, and the daily service vehicles are more than 500. There are a series of problems such as difficulty in real-time monitoring of equipment running state, lack of personalized electricity price strategy, large fluctuation of user waiting time and delay in response to equipment sudden failure. The traditional operation and maintenance mode mainly relies on manual inspection and fixed period equipment maintenance plan, which leads to poor equipment failure prediction ability, unstable service user satisfaction, and also causes unreasonable allocation of operation and maintenance resources and cost redundancy.

[0119] In order to verify the application effect of the application in actual operation, the new energy charging SaaS operation and management method based on artificial intelligence is deployed in the SaaS platform of the site. The system first collects the running parameters (such as voltage, current, temperature, etc.) of the charging pile equipment, user charging behavior records, power grid load data, electricity price change trend and environmental data, and generates a unified structured charging operation data set through data cleaning, time alignment and format unification. Subsequently, based on the data set, a charging state atlas containing device nodes and user nodes is constructed, and an improved spatio-temporal residual graph convolution network is used to model the graph structure sequence, and the device state embedding vector and user behavior embedding vector are extracted.

[0120] On the basis of generating user portraits, combined with electricity price, load and other information input conditions, the system dynamically generates personalized charging price strategies perceived by users. At the same time, the device state embedding vector and device historical running time sequence are jointly input into an improved time series modeling network based on the Transformer structure, and the device health state recognition and risk level evaluation are automatically completed, and then the predictive maintenance suggestions are output. All results are uniformly packaged and deployed in the platform standard interface, providing health scores and maintenance priorities for operation personnel, providing real-time personalized prices for user end, and synchronously pushing to the equipment supplier end for spare parts allocation.

[0121] After 180 days of continuous deployment and operation, compared with the traditional manual management mode, the average response time is shortened by 38%, the maintenance frequency is reduced by more than 40%, the user waiting time is reduced by nearly 50%, the user satisfaction is improved to more than 8.2 points on average, which is significantly better than the traditional mode of 5.9 points. At the same time, in terms of pricing flexibility, the personalized strategy improves the income elasticity space, and the charging price difference rate is controlled between 10%-13%, which is recognized by users.

[0122] The specific effect data is shown in the following table:

[0123] Table 1: Operation data based on the method of the application

[0124]

[0125] Table 2: Comparison data of traditional method

[0126]

[0127]

[0128] It can be seen by comparison that the method significantly reduces the maintenance frequency and response time, improves the user service efficiency and satisfaction, and proves the feasibility and practical value of the application in new energy charging operation management.

[0129] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered in the protection scope of the application.

Claims

1. An artificial intelligence-based new energy charging SaaS operation management method, characterized in that, The method comprises the following steps: Collecting data of new energy charging facilities and performing data preprocessing to generate a structured charging operation dataset; Based on the structured charging operation dataset, a charging state atlas is constructed, edge connection relationships are established according to geographical adjacency relationships, user use correlation relationships and load conduction relationships, and a graph structure sequence is formed according to time sequence; The graph structure sequence is input into an improved spatiotemporal residual graph convolution network model to output device state embedding vectors and user behavior embedding vectors; According to the user behavior embedding vectors and the user historical charging records, a user portrait data is constructed, combined features are input into a conditional variational autoencoder model in combination with current period electricity price, power grid load state and user portrait information, and an individualized charging price strategy is output; The device state embedding vectors and the device operation time sequence are input into an improved time series modeling network based on a Transformer structure, short-term fluctuations and long-term trend features of the device operation state are extracted through a multi-head attention mechanism, device health state recognition and risk level evaluation are performed, and a device health state evaluation result is generated; The individualized charging price strategy, the device health state evaluation result and the user portrait data are packaged into a standardized service module and deployed to a new energy charging SaaS platform, which is provided to an operation management end, a user end and a device supplier end through a platform interface. 2.The new energy charging SaaS operation management method based on artificial intelligence according to claim 1, characterized in that, The collecting data of new energy charging facilities and performing data preprocessing to generate a structured charging operation dataset specifically comprises: An edge data collection device is deployed on site of the new energy charging facilities to collect device operation data of the charging piles in real time, the device operation data including voltage, current, power, working state code, operation temperature, operation time, interface state and device exception identifier; User charging behavior data including user identifier information, charging start and end time, selected charging mode, historical charging records and transaction record data are collected synchronously and time-labeled bound to corresponding device operation data; Environment data including current period electricity price, power grid load state, meteorological data and regional traffic congestion indicators are collected from a power grid interface and external data sources and uniformly associated to corresponding sites and time windows; Field standardization and data type conversion operations are performed on the collected device operation data, user charging behavior data and environment data, a uniform field structure table is established, and resampling and time series alignment are performed according to a uniform time stamp granularity; Missing value filling, repeated value elimination and abnormal value identification processing are performed on the time series aligned data, sliding window segmentation reconstruction is adopted to slice continuous data segments, and a structured charging operation dataset with time sequence attributes is formed. 3.The new energy charging SaaS operation management method based on artificial intelligence according to claim 1, characterized in that, The based on the structured charging operation dataset, the charging state atlas is constructed, edge connection relationships are established according to geographical adjacency relationships, user use correlation relationships and load conduction relationships, and a graph structure sequence is formed according to time sequence specifically comprises: Based on the device operation data and user charging behavior data in the structured charging operation dataset, an initial node set of a charging state atlas is constructed, the initial node set includes charging pile device nodes and user behavior nodes, and each node is assigned a unique identifier and an attribute vector; According to the geographical adjacency relationship between devices, the power grid access topology structure and the load conduction path, the spatial correlation edge connection relationship between device nodes is constructed, and according to the continuous charging behavior path of users between devices, the behavior interaction edge connection relationship between user behavior nodes and charging pile device nodes is constructed; According to the time sequence index, the structured charging operation dataset is divided into continuous time segments, and a time-stamped graph structure is constructed for each time segment to form a graph structure sequence of the charging state atlas; The node set and edge set in the graph structure sequence are normalized, and graph structure sparsification, edge weight normalization and node attribute vector encoding operations are performed to obtain a graph structure sequence with time sequence characteristics and structure attributes. 4.The new energy charging SaaS operation management method based on artificial intelligence according to claim 1, characterized in that, The improved spatio-temporal residual graph convolutional network model is inputted with the graph structure sequence, and outputs device state embedding vectors and user behavior embedding vectors, which specifically includes: The graph structure sequence dataset is inputted into the improved spatio-temporal residual graph convolutional network model, and the graph structure sequence dataset contains a set of graph snapshots arranged in chronological order, each graph snapshot is composed of a node set, an edge set, a node attribute vector and an edge connection weight; For each graph snapshot at a time, a spatial graph convolution operation is performed, a normalized adjacency matrix is constructed according to the edge connection weight between nodes, and the attribute vector of the target node and the attribute vector of the neighbor node are linearly combined and nonlinearly activated through a graph convolution operator to extract the node state representation of the node in the topology structure; A cross-layer residual connection structure is introduced after each spatial graph convolution layer to perform element-level residual fusion on the node state representation input by the previous layer and the graph convolution output of the current layer; The node state representation after element-level residual fusion is arranged in chronological order and input into the time dimension modeling module, the time dimension modeling module includes a one-dimensional time convolution layer, a node state maintaining module and a feature normalization activation unit, the one-dimensional time convolution layer performs convolution operation on the state representation of each node in multiple graph snapshots along the time index dimension, the node state maintaining module is used to maintain the historical state statistical value of each node in a sliding time window and splice with the current state, and the feature normalization activation unit is used to perform batch normalization and activation function processing, and the node state maintaining module is embedded in the time convolution layer, the node state maintaining module records and updates the key state indicators of the node through a sliding window mechanism. A graph time alternation processing unit is arranged between the spatial graph convolution operation and the time dimension modeling module, and is composed of an alternation controller, an intermediate feature buffer and a fusion update mechanism. The alternation controller is used to set the execution order of the spatial feature extraction stage and the time sequence modeling stage in each processing round. The intermediate feature buffer is used to store the spatial processing output result and provide the time modeling stage with the spatial processing output result. The fusion update mechanism is used to fuse the spatial and time features and update the current node state representation. The graph time alternation processing unit alternately performs spatial feature extraction and time dimension modeling of the sequence node state of each graph snapshot. The graph structure information of each time step is input into the spatial graph convolution network to extract the topological correlation features between nodes, and the extraction result is input into the time modeling module to update the node state representation in combination with the historical state information. Meanwhile, the updated node state representation is taken as the spatial convolution input of the next graph snapshot. Through the stacking combination of the spatial convolution layer, the time convolution layer, the cross-layer residual connection structure and the graph time alternation processing unit, the device state embedding vector of each device node under the current time window and the user behavior embedding vector of each user node are output. 5.The new energy charging SaaS operation management method based on artificial intelligence according to claim 1, characterized in that, The user portrait data is constructed according to the user behavior embedding vector and the user historical charging record. The combined features are input into the conditional variational autoencoder model in combination with the current period electricity price, the power grid load state and the user portrait information, and the personalized charging price strategy is output. Specifically, the user portrait representation vector is constructed based on the user behavior embedding vector and the historical charging record in the structured charging operation data set, the user charging frequency, the charging period distribution, the average charging capacity and the site preference information are extracted, and the user portrait representation vector is constructed. The user portrait representation vector is spliced with the current period electricity price, the power grid load state and the meteorological data to construct a joint input feature vector for the pricing task, and the joint input feature vector is input into the conditional variational autoencoder model. In the conditional variational autoencoder model, the encoder module receives the joint input feature vector and extracts the latent pricing preference distribution, and outputs the mean vector and the variance vector as the hidden space parameters. The hidden variable vector is sampled and transmitted to the decoder module. The hidden variable vector is mapped to the personalized charging price strategy output vector by the decoder module. The personalized charging price strategy output vector includes the floating single value, the recommended price adjustment coefficient and the peak-valley response intensity factor for the specified user in the current period. The personalized charging price strategy output vector is numerically modified and boundary cropped to generate an executable format charging pricing strategy scheme, and a pricing mapping pair is formed for each user and charging site, which is provided to the SaaS platform for pricing execution and user display. The device state embedding vector and the device running time series data are input into the improved time series modeling network based on the Transformer structure, the short-term fluctuation and long-term trend features of the device running state are extracted through the multi-head attention mechanism, the device health state is identified and the risk level is evaluated, and the predictive maintenance suggestion is generated. Specifically, 6.The new energy charging SaaS operation management method based on artificial intelligence according to claim 1, wherein, ​ The device state embedding vector is time-indexed and aligned with the runtime sequence of the same device in the structured charging operation data set, and a joint time sequence input sequence containing device operation state information and historical operation indicators is constructed; The joint time sequence input sequence is input into a time sequence modeling network based on the Transformer structure with a gating factor fusion mechanism, a position encoding module is used to add time position information to the input data of each time step, and a state input sequence with time dependence is formed; In each Transformer encoding unit, the state dependence relationship between the current time step and other time steps is extracted based on the multi-head attention mechanism, and a gating factor structure is introduced after the output of the feedforward subnetwork to perform weighted fusion operation on the current layer output and the previous layer input; During attention calculation, a state mutation perception module is introduced to automatically adjust the attention allocation weight according to the change amplitude between the current state and the previous state; The time sequence features encoded by the Transformer structure are input into parallel health state recognition branches and risk trend scoring branches, the first branch outputs the device health category label of the current period, the second branch outputs the future risk score level and trend change indicators, and the complete device health state evaluation result is formed; Based on the device health state evaluation result, a predictive maintenance suggestion is automatically generated, which includes recommended maintenance time interval, potential failure level and involved maintenance module type; The device health state evaluation result and the predictive maintenance suggestion are packaged in a structured format and stored synchronously in the device state database. 7.The new energy charging SaaS operation management method based on artificial intelligence according to claim 1, characterized in that, The personalized charging price strategy, device health state evaluation result and user portrait data are packaged as a standardized service module and deployed to a new energy charging SaaS platform, which is provided to the operation management end, user end and device supplier end through the platform interface, specifically including: The personalized charging price strategy, device health state evaluation result and predictive maintenance suggestion, and user portrait data are indexed, matched and aligned, and a joint input data set integrating users, devices and operation states is constructed; According to the management target of the operator, a set of complex operation strategy targets including revenue efficiency indicators, user satisfaction indicators and system stability indicators is set, and a strategy mapping space is established; Based on the joint input data set and the set of complex operation strategy targets, a strategy generation rule engine is used to schedule and plan the operation behavior of each charging station in the current period, and the scheduling plan content includes dynamic pricing strategy execution instructions, charging pile start-stop state adjustment instructions and maintenance time window configuration instructions; For charging pile devices with fault risk, scheduling suspension instructions and fault avoidance paths are generated according to the risk level and remaining available time interval marked in the device health state evaluation result, and maintenance suggestions are synchronized and bound to the scheduling strategy; The strategy generation result is subjected to effectiveness verification and data consistency check, and the strategy priority is sorted and adjusted according to the user historical feedback data and current load prediction information, and a final executable dynamic operation strategy instruction set is constructed; The dynamic operation strategy instruction set is packaged as a service interface that can be called by the SaaS platform, and is pushed to an operation party management terminal, a user terminal prompt system and a device supplier interface according to role permissions, so that each role can respond and execute the strategy in real time.

Images