A new energy station electric-hydrogen coupling energy storage system and method

Abstract

The application relates to the technical field of new energy storage regulation, and discloses a new energy station electric-hydrogen coupling energy storage system and method. The system comprises a feature analysis unit, a topology construction unit, a regulation domain positioning unit and a decision optimization unit. The feature analysis unit receives new energy station operation condition information, generates a working condition feature vector according to real-time load fluctuation, historical output curve and equipment operation database by using a deep feature extraction network; the topology construction unit constructs an energy topology graph for distributed energy storage equipment and hydrogen production equipment in the electric-hydrogen coupling system, dynamically updates the graph structure and node state based on a real-time monitoring mechanism, and deploys the graph structure and node state in a heterogeneous computing framework; the regulation domain positioning unit positions a matched regulation domain in a dynamic representation learning mode according to the working condition feature vector; and the decision optimization unit filters regulation instructions and generates a scheduling sequence in the regulation domain by using a multi-source decision fusion algorithm, thereby improving the adaptability and operation efficiency of the system to the working condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy storage regulation, in particular to a new energy station electric-hydrogen coupling energy storage system and method. BACKGROUND

[0002] With the rapid development of the new energy industry, the proportion of renewable energy such as wind energy and solar energy in the power system continues to increase, but such energy is significantly affected by natural conditions, and the output has strong intermittency and volatility, which brings serious challenges to the stable operation of the power system. In order to suppress the output fluctuation of new energy and improve energy utilization efficiency, energy storage system has become an indispensable part of new energy station. Traditional energy storage methods mostly rely on single electrochemical energy storage, which has problems such as limited capacity, short service life, and high long-term energy storage cost, and is difficult to meet the large-scale and long-time energy storage needs of new energy stations.

[0003] Electric-hydrogen coupling energy storage technology has gradually attracted attention. This technology converts electrical energy into hydrogen energy for storage, combining the fast response advantage of electrochemical energy storage, and is expected to achieve multi-energy complementary energy storage effect. However, the current electric-hydrogen coupling energy storage system still faces many technical bottlenecks in practical application. On the one hand, the existing system lacks the ability to accurately perceive and extract the characteristics of the operating conditions of the new energy station, and cannot dynamically adjust the system operation strategy according to real-time load fluctuations, historical output data and equipment operating state, resulting in low system operation efficiency and difficulty in adapting to complex and variable working conditions. On the other hand, the electric-hydrogen coupling system contains distributed energy storage devices, hydrogen production devices and other heterogeneous devices, and the energy conversion and power transmission relationship between devices is complex. The existing topology structure is mostly statically designed and cannot update the device state and energy flow in real time, making it difficult to achieve collaborative optimization operation between devices.

[0004] In terms of regulation strategy formulation, traditional regulation methods mostly use single decision model, which cannot fuse multi-source information for comprehensive decision making, and is difficult to accurately locate and match the regulation range with the current working condition, resulting in lack of pertinence of regulation instruction and inability to meet the optimization operation demand of system under multiple constraint conditions. At the same time, the existing system mostly uses single computing framework, which is difficult to efficiently process massive working condition data and complex topology calculation tasks, resulting in system response delay and further affecting the overall operation performance of the energy storage system. The existence of these problems seriously restricts the large-scale promotion and application of electric-hydrogen coupling energy storage technology in new energy stations, and an electric-hydrogen coupling energy storage system capable of realizing accurate working condition perception, dynamic topology construction, accurate regulation domain positioning and multi-source decision optimization is urgently needed. SUMMARY

[0005] The present application aims to provide a new energy station electric-hydrogen coupling energy storage system and method to solve the problems raised in the background.

[0006] To achieve the above object, the application provides a new energy station electric-hydrogen coupling energy storage system, which comprises:

[0007] A feature analysis unit is configured to receive new energy station operation condition information, utilize a deep feature extraction network, and generate a condition feature vector according to real-time load fluctuation, historical output curve and equipment operation database in the operation condition information;

[0008] A topology construction unit is configured to construct an energy topology graph for distributed energy storage equipment and hydrogen production equipment in the electric-hydrogen coupling system, wherein the energy topology graph represents energy storage units or hydrogen production units through nodes and represents energy conversion or power transmission relationships through edges, dynamically updates graph structure and node state based on real-time monitoring mechanism, and deploys the energy topology graph on a heterogeneous computing framework;

[0009] A regulation domain positioning unit is configured to position a regulation domain matched with a current condition in the energy topology graph in a dynamic representation learning manner according to the condition feature vector;

[0010] A decision optimization unit is configured to utilize a multi-source decision fusion algorithm to screen a regulation instruction set meeting operation constraints and generate a scheduling sequence in the regulation domain.

[0011] Preferably, the feature analysis unit comprises:

[0012] A condition mapping subunit is configured to map the operation condition information to a feature space through a deep feature extraction network to generate a load fluctuation embedding vector;

[0013] A historical analysis subunit is configured to extract a charging and discharging mode, a hydrogen production efficiency curve and a scheduling response record in the historical output curve, construct an equipment operation portrait and convert it into an equipment state embedding vector;

[0014] A database access subunit is configured to query an equipment operation database, extract a power grid scheduling strategy and an equipment operation threshold, and generate a strategy constraint embedding vector;

[0015] An intention fusion subunit is configured to utilize a feature alignment algorithm to fuse the load fluctuation embedding vector, the equipment state embedding vector and the strategy constraint embedding vector to generate the condition feature vector, wherein the condition feature vector contains power grid peak shaving demand and equipment operation boundary.

[0016] Preferably, the topology construction unit comprises:

[0017] A node generation subunit is configured to obtain charging and discharging characteristics of energy storage units and conversion efficiency of hydrogen production units through an operation parameter extraction technology, take each equipment unit as a node of the energy topology graph, and associate corresponding dynamic operation parameters;

[0018] The relationship construction subunit is configured to create edges according to power transmission paths between device units and energy conversion efficiencies, and set initial transmission weights, wherein the weight values are initialized based on energy conversion loss rates between devices;

[0019] The dynamic updating subunit is configured to track device output data, grid-connected states and connection relationship changes based on a real-time monitoring mechanism, and adjust associated nodes and edge weights using an incremental updating strategy when new devices or parameters are updated.

[0020] The heterogeneous deployment subunit is configured to select a distributed storage engine in a heterogeneous computing framework, and deploy a real-time energy topology graph.

[0021] Preferably, the regulation domain positioning unit comprises:

[0022] The topology embedding subunit is configured to pre-train an energy topology graph using a dynamic representation learning algorithm, and generate dynamic state embeddings of each node, wherein the dynamic state embedding space synchronously reflects real-time output capabilities of devices and topology connection relationships.

[0023] The vector alignment subunit is configured to map the operating condition feature vector to the dynamic state embedding space, so that the operating condition feature vector and the node dynamic state embedding are in the same vector space.

[0024] The seed positioning subunit is configured to calculate matching degrees of the mapped operating condition feature vector and all node dynamic state embeddings, and select a preset number of nodes with the highest matching degrees as regulation seed nodes.

[0025] The neighborhood expansion subunit is configured to expand adjacent nodes layer by layer according to connection edges of the regulation seed nodes, and include the adjacent nodes in the regulation domain if state matching degrees of the adjacent nodes reach an operating threshold, until a preset topology depth or a regulation capacity threshold is met.

[0026] The domain generation subunit is configured to determine the expanded node set as the regulation domain, wherein the regulation domain comprises device dynamic state embeddings and an energy topology subgraph.

[0027] Preferably, the decision optimization unit comprises:

[0028] The feature extraction subunit is configured to extract a multi-dimensional operating feature vector for each device unit in the regulation domain.

[0029] The fusion computing subunit is configured to aggregate the multi-dimensional operating feature vectors using a multi-source decision fusion algorithm, and generate a device comprehensive state vector, wherein the fusion process uses a weighted decision tree or a deep belief network.

[0030] The instruction screening subunit is configured to calculate an operating fit degree of the device comprehensive state vector and the operating condition feature vector, and screen devices with a fit degree higher than an operating threshold to generate a regulation instruction set.

[0031] The sequence generating subunit is configured to sort the control instruction set based on the operation fitness, the power grid frequency deviation and the device response delay, and generate a scheduling sequence.

[0032] Preferably, the system further comprises:

[0033] The data alignment unit is configured to collect historical fault data of the electric-hydrogen coupling system, acquire multi-source operation parameters and fault device types in a preset time period before the fault occurs, label timestamps for the multi-source operation parameters and the fault device types, and perform spatio-temporal alignment.

[0034] Preferably, the system further comprises:

[0035] The evaluation interval division unit is configured to divide the multi-source operation parameters according to a fixed sampling interval, extract operation modal components through a signal decomposition technique, and divide evaluation intervals using an adaptive sliding window according to the spectral distribution characteristics of the modal components.

[0036] Preferably, the system further comprises:

[0037] The state evaluation unit is configured to, in each evaluation interval, calculate a device risk coefficient and a health index based on the number of grid-connected point voltage fluctuations, the number of power out-of-limit events and the protection action records, in combination with a time decay coefficient of the evaluation interval, through a dynamic weight distribution algorithm.

[0038] Preferably, the system further comprises:

[0039] The prediction optimization unit is configured to establish an operation state prediction model, input the device risk coefficient and the health index of the current evaluation interval into the trained operation state prediction model, and output a predicted fault probability distribution of the next evaluation interval, wherein the operation state prediction model adopts a coding-decoding neural network architecture.

[0040] Preferably, the application further comprises a new energy station electric-hydrogen coupling energy storage method applied to the new energy station electric-hydrogen coupling energy storage system, and the method comprises:

[0041] Step 1: receiving new energy station operation condition information, using a deep feature extraction network, and generating a condition feature vector according to real-time load fluctuations, historical output curves and device operation databases in the operation condition information;

[0042] Step 2: constructing an energy topology graph for distributed energy storage devices and hydrogen production devices in the electric-hydrogen coupling system, wherein the energy topology graph represents energy storage units or hydrogen production units through nodes, and represents energy conversion or power transmission relationships through edges, dynamically updates the graph structure and node state based on a real-time monitoring mechanism, and deploys the energy topology graph on a heterogeneous computing framework.

[0043] Step 3: According to the working condition characteristic vector, locate the regulation domain matching the current working condition in the energy topology graph in a dynamic representation learning manner;

[0044] Step 4: In the regulation domain, use a multi-source decision fusion algorithm to screen a set of regulation instructions that meet the operation constraints and generate a scheduling sequence.

[0045] Compared with the prior art, the beneficial effects of the present application are:

[0046] The feature analysis unit uses a deep feature extraction network to integrate real-time load fluctuations, historical output curves, and device operation databases to generate a working condition characteristic vector, which can fully capture the dynamic change characteristics in the operation process of the new energy station, accurately reflect the current operating state of the system, and provide accurate working condition basis for subsequent regulation strategy formulation, avoiding operation strategy deviation caused by inaccurate working condition perception, so that the system can better adapt to the intermittency and volatility of new energy output, and improve the adaptability of the system to complex working conditions.

[0047] The topology construction unit constructs an energy topology graph for distributed energy storage devices and hydrogen production devices in the electricity-hydrogen coupled system, clearly represents the energy conversion and power transmission relationship between energy storage units, hydrogen production units and devices through nodes and edges, and dynamically updates the graph structure and node state based on real-time monitoring mechanism, ensuring that the topology graph can reflect the device operating state and energy flow changes in real time, breaking the limitations of traditional static topology structure. Deploying the energy topology graph in a heterogeneous computing framework can fully utilize the efficient data processing capability of heterogeneous computing, quickly process massive device state data and complex topology calculation tasks, reduce data processing delay, improve topology structure update and management efficiency, provide flexible and efficient topology support for collaborative operation between devices, and promote energy optimization distribution between distributed devices.

[0048] The regulation domain positioning unit locates the regulation domain matching the current working condition in the energy topology graph in a dynamic representation learning manner according to the working condition characteristic vector, which can accurately lock the device range that needs to be regulated, avoid resource waste caused by excessive regulation range or incomplete regulation caused by small regulation range, and make the regulation instruction accurately act on the key device, improving the pertinence and effectiveness of regulation. Through dynamic representation learning, the regulation domain range can be adjusted in real time to ensure that the regulation domain always maintains the best matching state with the current working condition, adapts to the dynamic changes of the system operating condition, and further improves the flexibility and adaptability of system regulation.

[0049] The decision optimization unit uses a multi-source decision fusion algorithm in the regulation domain to screen a regulation instruction set that meets the operation constraints and generate a scheduling sequence, can fuse multi-source information for comprehensive decision, fully considers equipment operation constraints, energy conversion efficiency, load demand and other factors, avoids the limitations of a single decision model, and generates a more optimal regulation strategy. By screening the regulation instruction set that meets the constraint conditions, the feasibility and safety of the scheduling sequence can be ensured, equipment failure or unstable system operation caused by instructions that do not meet the constraint conditions can be avoided, and the generated scheduling sequence can realize collaborative optimal operation of equipment, improve the overall energy utilization efficiency of the system, reduce the operation cost, and promote efficient application and promotion of the electric hydrogen coupling energy storage technology in the new energy station. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A timing diagram of the new energy station electric hydrogen coupling energy storage system according to the present application;

[0051] Figure 2 An architectural diagram of the feature analysis unit workflow;

[0052] Figure 3 An architectural diagram of the topology construction unit workflow;

[0053] Figure 4 An architectural diagram of the regulation domain positioning unit workflow. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0055] Please refer to Figure 1 The present application provides a new energy station electric hydrogen coupling energy storage system and method, the system comprises:

[0056] The operating condition information of the new energy station is received by the feature analysis unit, and the information includes real-time load fluctuation data, historical output curve and equipment operation database record. The deep feature extraction network processes the above information in layers, extracts high-dimensional features and generates a working condition feature vector representing the current system operating state. The topology construction unit constructs an energy topology graph for the distributed energy storage devices and hydrogen production devices in the electricity-hydrogen coupled system, wherein the nodes represent the energy storage units or hydrogen production units, and the edges represent the energy conversion or power transmission relationship. The graph is dynamically updated by a real-time monitoring mechanism for node state and connection relationship, and is deployed in a heterogeneous computing framework for efficient parallel processing. The working condition feature vector is input into the dynamic representation learning model of the energy topology graph by the regulation domain positioning unit, and the regulation domain matched with the current working condition is located by vector space matching and neighborhood expansion algorithm. The decision optimization unit uses a multi-source decision fusion algorithm within the regulation domain to screen a set of regulation instructions that meet the operating constraint conditions, and generates an optimized scheduling sequence based on the grid frequency deviation and device response delay.

[0057] Embodiment 1: refer to Figure 2 After the feature analysis unit receives the operating condition information of the new energy station, the working condition feature vector is generated by the internal sub-unit cooperative processing. The working condition mapping sub-unit first obtains the real-time load fluctuation data, which records the instantaneous change of the grid load in the form of time series. The deep feature extraction network uses a three-layer convolution structure to process the sequence, the first layer of convolution kernel width is set to 5 minutes to capture the short-term fluctuation trend; the second layer of convolution kernel is extended to 30 minutes to extract the medium period fluctuation feature; the third layer uses a global pooling layer to compress the time dimension, and finally outputs a 128-dimensional load fluctuation embedding vector. The vector encodes the amplitude, frequency and change mode features of the load fluctuation.

[0058] The historical analysis sub-unit synchronously processes the historical output curve data, and the analysis objects include the charge and discharge records of the energy storage devices, the efficiency curve of the hydrogen production devices and the historical scheduling response log. The charge and discharge records are divided into multiple charge and discharge period segments by a sliding window, and the peak power, duration and charge and discharge efficiency slope features are extracted for each segment. The hydrogen production efficiency curve is fitted by piecewise linear fitting, and the efficiency inflection point and slope change value under different load rates are recorded. The scheduling response log statistics the average delay time and success rate of the device response instructions. The above features are fused by an attention mechanism: dynamic weight is assigned to the charge and discharge mode, and the feature weight of the high fluctuation period is improved; the hydrogen production efficiency inflection point feature is weighted by a Gaussian kernel function; the response delay data is normalized. After fusion, a 256-dimensional device state embedding vector is generated, which fully represents the device historical operation image.

[0059] The database access subunit connects the equipment operation database, and the query content includes the power grid dispatching strategy file, the equipment safe operation threshold parameter and the real-time constraint condition. The dispatching strategy file is parsed into a priority rule set, such as peak shaving period division and power adjustment upper limit; the equipment operation threshold extracts the boundary value of the voltage, current, temperature and other parameters; the constraint condition includes the grid-connected point frequency deviation tolerance and the power change rate limit. The above data is processed by using an encoder structure: the strategy file is converted into a rule vector by using a natural language processing technology; the threshold parameter is directly mapped into a numerical vector; and the constraint condition is converted into a logic expression vector. After compression by a full connection layer, the three types of vectors generate a 64-dimensional strategy constraint embedding vector, wherein the power grid peak shaving demand is embedded in the form of one-hot encoding.

[0060] The intention fusion subunit receives the load fluctuation embedding vector (128 dimensions), the equipment state embedding vector (256 dimensions) and the strategy constraint embedding vector (64 dimensions). The feature alignment algorithm first normalizes the space of the three vectors: the dimensions are unified to 256 dimensions by linear transformation, and the key features of the original vectors are retained in the transformation process. The alignment process measures the similarity between vectors by using the Euclidean distance: the distance between the load fluctuation vector and the equipment state vector reflects the matching degree of the real-time working condition and the equipment capacity; and the distance between the strategy constraint vector and the equipment state vector represents the compliance degree of the operation boundary. According to the distance calculation result, the fusion weight is dynamically allocated: when the power grid peak shaving demand is enhanced, the weight of the strategy constraint vector is increased; and when the equipment operation state fluctuates greatly, the weight of the equipment state vector is increased. The weighted fusion process adopts element-by-element addition and a nonlinear activation function, and finally generates a 256-dimensional working condition feature vector. The vector contains three core information segments: the first 85 dimensions encode the load fluctuation characteristics, the middle 128 dimensions represent the historical state of the equipment, and the last 43 dimensions carry the strategy constraint condition. After the vector is generated, the verification module verifies whether the numerical range meets the preset boundary, and the abnormal value triggers the feature recalculation process.

[0061] The entire processing flow adopts a pipeline architecture: the four subunits are executed in sequence, and the subunits exchange data through shared memory. When the real-time load fluctuation data is updated, only the working condition mapping subunit is recalculated; when the equipment database is updated, the database access subunit incrementally updates the strategy constraint embedding vector. The historical analysis subunit triggers full calculation every morning, and uses incremental update mode in the remaining period. All embedding vector generation processes retain intermediate result logs, supporting running state backtracking analysis. The finally output working condition feature vector is transmitted to the control domain positioning unit through a message queue, and the transmission protocol adopts binary encoding to reduce the time delay.

[0062] In terms of exception handling mechanism, when the deep feature extraction network input data is abnormal, it automatically switches to the backup data processing channel: using historical synchronous load fluctuation data to fill the real-time data missing section; when the device database connection is interrupted, the latest strategy constraint data in the local cache is enabled. In the feature alignment process, if conflicts between vectors are detected (such as device capacity cannot meet the peak shaving demand), a conflict flag bit is generated and written into the metadata section of the working condition feature vector for subsequent unit decision reference. The time performance of the entire feature analysis process is optimized through parallel computing: the load fluctuation processing and historical analysis subunits are executed simultaneously on different computing cores of a heterogeneous computing framework, and the final fusion stage is accelerated through GPU matrix operations.

[0063] Embodiment 2: refer to Figure 3 The topology construction unit accesses the monitoring network of the electricity-hydrogen coupled system through the node generation subunit, and real-time collects the operating parameters of the distributed energy storage devices and hydrogen production devices. The energy storage unit parameters include the charge-discharge efficiency curve of the lithium battery pack, the response time characteristic of the super capacitor, and the speed-power mapping relationship of the flywheel energy storage; the hydrogen production unit parameters cover the current-hydrogen production amount correspondence table of the alkaline electrolyzer, the start-stop loss rate of the PEM electrolyzer, and the pressure-capacity conversion coefficient of the hydrogen storage tank. Each physical device is mapped as an independent node in the energy topology graph, and the node attributes include device type identifier, real-time operating state code, and dynamic parameter set. The parameter set is stored in the form of key-value pairs, for example, the energy storage node includes "current SOC", "maximum charge-discharge rate", "temperature alarm threshold" and other fields, and the hydrogen production node includes "real-time hydrogen production rate", "electrolysis efficiency", "preheating state" and other fields. The node generation process adopts an automatic registration mechanism, when a new device accesses the system, the corresponding node is automatically created through the device identifier and the parameter template is initialized.

[0064] The relationship construction subunit analyzes the energy interaction relationship between devices based on the electrical wiring diagram and the energy flow topology. For the power supply link from the energy storage unit to the hydrogen production unit, a "electric energy transmission" type directed edge is created, with the direction pointing from the energy storage node to the hydrogen production node; for the hydrogen transmission pipeline from the hydrogen production unit to the hydrogen storage tank, a "hydrogen energy transmission" type directed edge is created. The attributes of each edge include transmission medium type, maximum transmission capacity and energy conversion efficiency. The initial transmission weight is set according to the energy conversion loss rate between devices: for example, the DC power supply link from the lithium battery pack to the PEM electrolyzer, the weight value is set to 0.97 according to the measured DC / DC conversion loss rate (typical value 2.5%-3.8%); for the hydrogen transmission pipeline from the alkaline electrolyzer to the hydrogen storage tank, the weight value is set to 0.98 according to the compression loss rate (typical value 1.2%-1.8%). The weight calibration adopts an online learning mechanism, and the nominal value provided by the device manufacturer is used at the initial stage of system operation, and the metering data is dynamically corrected after operation.

[0065] The dynamic updating subunit establishes a millisecond-level polling mechanism to monitor three types of key changes: device output data is collected in real time through the Modbus / TCP protocol, and the power value and efficiency value in the node attribute are updated every 200 milliseconds; the grid-connected state is monitored through the circuit breaker signal contact, and the state change triggers the node "grid-connected flag" to flip immediately; the connection relationship change is detected by the topology analysis engine, and when a new photovoltaic hydrogen generator is added, a composite node is automatically created and a bidirectional connection edge to the grid node is established. The incremental update strategy uses a version control mechanism: each node maintains a version number, and when the parameters are updated, an incremental data packet (containing only the changed fields) is generated and broadcast to the graph database through the message queue. The edge weight update uses a sliding window algorithm, taking the moving average of the last 10 transmission loss measurements to refresh the weight value. Historical version data is retained for 72 hours to support running state backtracking analysis.

[0066] The heterogeneous deployment subunit converts the energy topology graph into an attribute graph model, and the node and edge data are stored according to the device partitioning strategy. In the heterogeneous computing framework, a distributed graph database is selected to store the static topology structure, and a columnar storage engine is used to compress attribute data; real-time state data is stored in a time series database, and a mapping relationship is established between the device ID and the graph node. The graph computing service is deployed on a GPU-accelerated node, which includes three types of core functions: topology traversal service supports shortest path query for energy transmission path optimization; neighbor aggregation service calculates node degree centrality to identify key hub devices; subgraph matching service realizes fast extraction of control domains. The service interface uses the gRPC protocol, and the response delay is controlled within 50 milliseconds. The load balancing module allocates computing resources according to the query type: state update requests are routed to the CPU cluster, and complex graph computing tasks are scheduled to the GPU node. Data consistency is guaranteed through distributed transactions, and cross-partition updates use a two-phase commit protocol.

[0067] The abnormal handling mechanism includes multiple safeguards: when device communication is interrupted, an LSTM-based output prediction model is enabled to fill in missing data, with a prediction duration of no more than 5 minutes; when the graph database fails, it automatically switches to an in-memory computing mode using the latest cached full topology snapshot; abnormal fluctuations in edge weights trigger a verification process that compares the theoretical loss range with the actual value, and when the deviation exceeds the threshold, a device diagnosis program is started. The performance monitoring module tracks the topology update time, query delay, and storage load in real time, and automatically expands the computing nodes when the resource utilization rate exceeds 80%.

[0068] The construction process of the energy topology map is strictly mapped to the physical structure of the electricity-hydrogen coupling system. Each battery cluster in the energy storage array is mapped as an independent node, and multiple electrolytic chambers of a large electrolytic cell are regarded as a composite node. The edge relationship accurately reflects the electrical connection: the "direct current transmission" edge is established between devices connected by a direct current bus, and the "alternating current coupling" edge is established between the PCS node and the grid node of the alternating current grid-connected device. The topology depth supports custom configuration, and the minimum granularity can be up to the single power module level. The graph data model adopts a standardized Schema definition, and the node type, edge type, and attribute field are uniformly registered through the metadata management module, supporting third-party systems to subscribe to topology change events through API interfaces.

[0069] The entire topology construction process adopts a hierarchical decoupling architecture: the physical device layer collects raw data through the OPCUA protocol, the data processing layer performs parameter cleaning and unit conversion, the graph model layer maintains the topology structure and attribute mapping, and the service layer provides calculation and query interfaces. The layers communicate asynchronously through a message middleware, and the system throughput supports processing 100,000-level data point updates per second. The version management database records all topology change history, and supports querying the system snapshot at any time according to the timestamp.

[0070] Embodiment 3: refer to Figure 4 The regulation domain positioning unit performs dynamic representation learning on the energy topology map through a topology embedding subunit. The subunit adopts a time series graph neural network architecture to process node state data, and the network input layer receives two types of real-time data streams: node attribute update stream containing device output value, efficiency parameter and state flag; the edge weight update stream reflects the change of energy transmission loss rate. The graph convolution layer is designed in a three-order hybrid mode: first-order convolution aggregates the features of directly adjacent nodes, second-order convolution captures the influence of two-hop neighbors, and third-order convolution introduces the context information of the entire graph. The convolution kernel weight is dynamically adjusted according to the edge type (electric energy transmission / hydrogen energy transmission): the electric energy transmission edge is weighted by the power factor, and the hydrogen energy transmission edge is weighted by the mass flow factor. The node state embedding generation process integrates the time series features: each node maintains a time series state queue with a length of 60, extracts the time-dependent features through an LSTM unit, and concatenates the spatial features output by the graph convolution to form a dynamic state embedding. The embedding vector dimension is fixed at 256, of which the first 128 dimensions encode the real-time output capacity of the device (such as maximum adjustable power and response speed), and the last 128 dimensions represent the topology connection relationship (such as neighbor node type and transmission path redundancy). The embedding model performs full-graph retraining every 5 minutes, and uses a sliding window mechanism for local fine-tuning of the embedding vector during incremental update.

[0071] The vector alignment subunit receives the working condition feature vector from the feature analysis unit, which contains the grid peak shaving demand and device operation boundary information. The alignment process uses a learnable projection matrix to map the 256-dimensional working condition feature vector to the dynamic state embedding space. The projection matrix is updated by online learning, where the matrix element represents the strength of the association between the i-th dimension of the operating condition feature and the j-th dimension of the state embedding. The mapping computation is expressed as:

[0072]

[0073] where: is the original operating condition feature vector, is the bias vector, and σ is the Sigmoid activation function, is the aligned operating condition vector. This vector is in the same metric space as the node dynamic state embedding, with its dimension distribution adjusted as follows: the first 85 dimensions correspond to load fluctuation features, the middle 128 dimensions match the device capability description, and the last 43 dimensions align with the strategy constraints.

[0074] The seed positioning subunit calculates the matching degree of the aligned operating condition vector with all node dynamic state embeddings . The matching degree function is defined as a weighted cosine similarity:

[0075]

[0076] where: is the load feature weight (default 0.4), is the device capability weight (default 0.6), and the vector slicing operation extracts the corresponding dimension segment. After calculating the similarity scores of all nodes, the top N nodes with the highest scores are selected as the regulation seed nodes, with N values dynamically set according to the system size:

[0077] N = 5 for small-scale stations (number of devices < 50),

[0078] N = 8 for medium-scale stations (50-200 devices),

[0079] N = 12 for large-scale stations (> 200 devices).

[0080] The seed node set needs to satisfy the type distribution constraint: at least 2 energy storage nodes and 1 hydrogen production node.

[0081] The neighborhood expansion subunit performs multi-level expansion starting from the seed nodes. The initial activation queue contains all seed nodes, and the expansion level L starts from 1 and increments. When expanding each level, all neighbors of the nodes in the current queue are traversed: for each neighbor node expanded by the search, its state matching degree is calculated to determine whether it is included in the regulation domain. The calculation formula of the state matching degree is as follows:

[0082]

[0083] wherein: is the similarity of , is the node available capacity, is the system total capacity, is the node availability index (0-1), , , is the weight coefficient, (default =0.7, =0.2, =0.1). If (θ is the running threshold, default 0.65), the is included in the regulation domain and added to the next layer of the expansion queue. The expansion process is subject to double constraints: the topological depth limit L≤3 (i.e. at most expanded to the three-hop neighbors of the seed node), the regulation capacity limit the total capacity of the devices in the domain is not less than 40% of the total system capacity. When any constraint is triggered, the expansion is terminated.

[0084] The domain generation sub-unit integrates the expansion results, and outputs the regulation domain containing three components: the node set stores the unique identifiers of all selected devices; the dynamic state embedding matrix stores the 256-dimensional embedding vectors of these devices; the energy topology subgraph stores the node connection relationship and edge weight in the adjacency list structure. The subgraph data is transmitted in compressed binary format, and the header information records the topological depth, total capacity proportion and core seed node list. The exception handling mechanism includes: when the expansion process exceeds 500 milliseconds, the pre-cached subgraph is enabled; when the matching degree calculation appears negative value, the feature realignment is triggered; when the node fails, the standby node is automatically removed and supplemented. The transmission protocol uses a special lightweight graph data exchange format, supporting edge weight incremental update.

[0085] The real-time performance of the dynamic representation learning module is guaranteed by heterogeneous computing: the graph convolution operation is deployed on the GPU cluster, and each node is allocated an independent CUDA stream; the time series feature extraction is executed in parallel on the CPU side; the matching degree calculation is optimized using SIMD instructions. The memory database caches the last 10 regulation domain subgraphs, supporting fast historical comparison. The monitoring interface outputs the heat map of the expansion process in real time, visualizing the node matching degree distribution and topological connection strength.

[0086] The regulation domain positioning process is synchronized with the grid dispatching cycle: in normal mode, it is executed once every minute; when the peak demand mutates, the emergency positioning is triggered (response time <300 milliseconds). Historical positioning records are stored in the analysis database, supporting retrieval of typical regulation domain configurations according to time range, load characteristics, etc. The positioning result verification mechanism includes: capacity verification to ensure that the devices in the domain can meet the peak power demand; type verification to ensure that the proportion of energy storage and hydrogen production equipment is reasonable; topology verification to detect isolated nodes and automatically rebuild connections.

[0087] In embodiment 4, the decision optimization unit receives the energy topology subgraph and node dynamic state embedding data output by the regulation domain positioning unit. The feature extraction subunit analyzes the multi-dimensional operating characteristics of each device unit in the domain. For energy storage devices, the extracted features include the current state of charge, the maximum charge and discharge power instantaneous value, the health status indicator, the temperature change rate, the life decay coefficient converted from the historical cycle number, and the communication delay time. For hydrogen production devices, the extracted features include the real-time electrolysis efficiency, the hydrogen production purity percentage, the thermal management system status code, the electrolytic cell voltage fluctuation range, the hydrogen output pressure stability coefficient, and the cooling water flow deviation value. After normalization, all features are combined into a 128-dimensional multi-dimensional operating feature vector. The physical meaning and numerical range of each dimension in the vector are defined by a metadata description file.

[0088] The fusion computing subunit uses a multi-source decision fusion algorithm to process the feature vector. The algorithm includes two parallel processing paths: the weighted decision tree path first assigns initial weights according to device type, with a base value of 0.6 for energy storage devices and 0.4 for hydrogen production devices; then dynamically adjusts the weights according to feature importance, such as increasing the state of charge feature weight by 0.15 and decreasing the temperature anomaly feature weight by 0.2. The decision tree depth is set to 7 layers, and each layer node is split based on the threshold value of a specific dimension in the feature vector. The deep belief network path uses a three-layer restricted Boltzmann machine structure, with 128 visible nodes and 64 hidden nodes, and is pre-trained unsupervisedly using the contrastive divergence algorithm. The outputs of the two paths are fused through an attention mechanism: the decision tree output focuses on individual device status, and the deep belief network output focuses on inter-device correlation characteristics. The fusion weight is adaptively adjusted according to the current operating mode. Finally, a 32-dimensional device comprehensive state vector is generated, which encodes the comprehensive evaluation of device availability, response potential, and operating risk.

[0089] The instruction screening subunit calculates the operating fit degree of the device comprehensive state vector and the operating condition feature vector. The fit degree calculation uses a multi-dimensional matching algorithm: first, project the two vectors into the same dimensional space; then calculate the matching score in different regions: use cosine similarity for power-related dimensions, use Hamming distance for constraint-related dimensions, and use dynamic time warping distance for time-related dimensions. The overall fit degree value is obtained by weighting and summing the scores of each region. The weight distribution is dynamically adjusted according to the grid dispatching priority. The screening process sets double thresholds: the basic fit degree threshold is set to 0.75 for preliminary screening; the type balance threshold requires the ratio of energy storage devices to hydrogen production devices to be maintained between 1.2:1 and 1.5:1. Devices that meet the threshold conditions enter the regulation instruction set. The instruction content is generated differently according to the device type: the energy storage device instructions include charge and discharge power set values, duration parameters, and power change slope limits; the hydrogen production device instructions include electrolysis power percentage, hydrogen production target value, and safe operation boundary parameters.

[0090] The sequence generation subunit temporally and spatially sorts the regulation instruction set. The time sorting is based on the device response delay data to establish a response time model: the median lithium battery energy storage response delay is 80 milliseconds, the median super capacitor response delay is 20 milliseconds, and the median electrolytic cell power adjustment delay is 2 seconds. The spatial sorting considers the power grid frequency deviation distribution, and the device instruction priority in the frequency deviation overrun area is promoted. The final generated scheduling sequence adopts a hierarchical structure: the first layer is the emergency response instruction (response time <100 milliseconds), the second layer is the fast adjustment instruction (100 milliseconds-2 seconds), and the third layer is the steady-state adjustment instruction (>2 seconds). Each layer is arranged in descending order of fitness, and the same sub-instruction is arranged in ascending order of device ID.

[0091] The data alignment unit runs in parallel to collect historical fault data of the electricity-hydrogen coupled system. The data collection range covers the operation records of the past 365 days, and the fault types are classified into three categories: electrical faults (overvoltage, overcurrent, insulation failure), mechanical faults (compressor abnormalities, valve sticking), and control faults (communication interruption, logic error). Multi-source operation parameters include second-level data 5 minutes before the fault occurs: 32 parameter channels such as three-phase voltage effective value, DC bus current, electrolytic cell temperature, hydrogen pressure, cooling liquid flow, and insulation resistance value. Time stamp alignment uses a high-precision clock synchronization protocol, and the clock deviation of all data sources is controlled within 1 millisecond. Temporal and spatial alignment is achieved through a three-dimensional interpolation algorithm, which unifies data with different sampling rates to a 100-millisecond time grid, and maps spatial position information to a unified coordinate system. The aligned data is stored in a distributed columnar database using a fault type-time stamp composite index structure. Fault data alignment, refer to Table 1.

[0092] Table 1: Temporal and spatial alignment table of multi-source operation parameters.

[0093]

[0094] The entire implementation process adopts an event-driven architecture: the regulation instruction generation period is 1 second, and the fault data collection period is 100 milliseconds. In terms of computing resource allocation, feature extraction and fusion calculation are deployed on FPGA acceleration cards, instruction screening and sequence generation run on CPU clusters, and data alignment processing uses a dedicated time series database engine. The abnormal handling mechanism includes: enabling a backup feature set when feature extraction is abnormal, using a simplified algorithm when fusion calculation times out, and triggering clock resynchronization process when data alignment error occurs. All intermediate results are recorded in audit logs to support full-process tracing and reproduction.

[0095] Example 5: The interval division unit processes multi-source operating parameters from the data alignment unit. Parameter data is organized in fixed sampling intervals, with a typical sampling rate of 100 milliseconds. The signal decomposition technique employs an adaptive noise complete ensemble empirical mode decomposition method to decompose the time series data of each parameter channel into multiple intrinsic mode components. Each component calculates the instantaneous frequency through Hilbert transform to form a spectral distribution map. The spectral characteristic analysis focuses on three dimensions: the dominant frequency bandwidth reflects the parameter fluctuation range, the energy concentration degree represents the signal stability, and the zero-crossing point density indicates the change intensity. An adaptive sliding window algorithm dynamically adjusts the window length according to the spectral characteristics: shortens the window to capture rapid changes when the dominant frequency bandwidth exceeds the threshold; and lengthens the window to improve stability when the energy concentration degree is below the set value. The window boundary aligns with the device operating cycle to avoid cross-cycle data segmentation. The window length adjustment follows the minimum continuity constraint, ensuring that a complete event sequence is included in a single window. The generated evaluation interval label starts with a timestamp, duration, and key spectral feature vector, serving as the basic time unit for state evaluation.

[0096] The state evaluation unit performs multi-dimensional health analysis within each evaluation interval. Input data includes grid-connected point voltage fluctuation frequency statistics, power limit event counts, protection device action records, and device operation logs. Voltage fluctuation frequency is classified by fluctuation amplitude: ±5% or less is considered a slight fluctuation, ±5%-10% is considered a moderate fluctuation, and ±10% or more is considered a severe fluctuation. Power limit events are divided into overload and underload types, with duration and maximum deviation values recorded. Protection action records analyze trip reason codes and action time delays. The time decay coefficient uses an exponential decay model, with events closer to the current time having higher weights, and decay time constants set according to device type: 2 hours for power electronic devices and 8 hours for mechanical devices. The dynamic weight allocation algorithm first sets a base weight for each type of event: voltage fluctuation weight 0.3, power limit 0.4, and protection action 0.3; then adjusts the coefficients according to event severity, with severe fluctuation event weight increased by 1.5 times and short-term slight limit weight reduced by 0.7 times. The device risk coefficient is calculated as the product of the weighted event integral and the time decay, and the health index is mapped to the 0-100 scale through an S-shaped function. The evaluation results are attached with confidence labels, reflecting data integrity and event relevance.

[0097] The prediction optimization unit constructs a running state prediction model. The model adopts a coding and decoding neural network architecture. The encoder part includes a three-layer gated recurrent unit network, with 128 neurons in each layer, and inputs the device risk coefficient vector and health index matrix of the current evaluation interval. The encoder outputs a 256-dimensional context vector, which captures the temporal evolution pattern of the device state. The decoder part adopts an attention mechanism enhanced recurrent neural network, which predicts the future state through time step expansion. The output layer of the decoder is connected to a softmax classifier, which generates a predicted failure probability distribution for the next evaluation interval. The failure type is divided into 12 mutually exclusive categories, covering the subdivision scenarios of electrical, mechanical, and control failures. The model training uses a historical aligned data set, with a fixed input sequence length of 6 consecutive evaluation intervals and an output prediction of the next interval state. The training process adopts a time series cross-validation strategy, dividing the two-year data into 104 training blocks by week, and generating training samples through a sliding window. The hyperparameter optimization adopts a Bayesian parameter tuning method, focusing on adjusting the dropout rate and learning rate decay strategy of the recurrent neural network. The model update mechanism adopts a rolling training mode, with incremental training triggered by new data every week, and the whole model retrained once every quarter.

[0098] During the execution of the prediction process, the device state data of the current evaluation interval is preprocessed and input into the encoder. The context vector and the historical state of the decoder together generate attention weights, focusing on key time step features. The decoder outputs a failure probability vector at each time step, and the final prediction result is the weighted fusion of all time step outputs. The probability distribution output includes the occurrence probability value and confidence interval of each failure type, and the high probability type triggers an early warning signal. The model explanation module provides feature contribution analysis, showing the key operating parameters that affect the prediction result and their contribution direction.

[0099] The system implementation adopts a hierarchical processing architecture: the evaluation interval division is performed on the edge computing node to reduce data transmission; the state evaluation is deployed on the local server to ensure real-time performance; the prediction optimization runs on the cloud computing platform to utilize distributed training resources. The modules communicate through standardized interfaces, and the data format uses protocol buffer encoding. The fault tolerance mechanism includes three safeguards: switch to fixed window mode when signal decomposition is abnormal; use a simplified algorithm when state evaluation times out; and start a rule-based backup predictor when the prediction model fails. All intermediate states are persistently stored, supporting complete process backtracking for fault diagnosis. The version control system manages algorithm parameter and model weight changes, meeting the audit tracking requirements.

[0100] Health index and risk coefficient data are stored in a time-series database, supporting multi-dimensional query and analysis. Prediction results are pushed to the monitoring system and displayed differently according to warning levels: low-risk probabilities are displayed as status alerts, medium-risk triggers a warning tag, and high-risk generates alarm work orders. Historical prediction accuracy is tracked through an online verification module, comparing predicted failure probabilities with actual failure records; the verification results are used to automatically adjust the model's confidence threshold. The equipment maintenance system receives health index trend data to assist in developing preventative maintenance plans. The system's operational status panel displays a real-time heatmap of the assessment interval distribution, an equipment health status matrix, and a failure probability cloud map, providing global operational situation awareness.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A new energy power station electro-hydrogen coupled energy storage system, characterized in that, include: The feature parsing unit is configured to receive operating condition information of new energy power plants, and generate operating condition feature vectors based on real-time load fluctuations, historical output curves and equipment operation databases in the operating condition information using a deep feature extraction network. The topology building unit is configured to construct an energy topology graph for distributed energy storage devices and hydrogen production devices in an electro-hydrogen coupling system. The energy topology graph uses nodes to represent energy storage units or hydrogen production units and edges to represent energy conversion or power transmission relationships. The graph structure and node status are dynamically updated based on a real-time monitoring mechanism, and the energy topology graph is deployed on a heterogeneous computing framework. The control domain positioning unit is configured to locate the control domain that matches the current operating condition in the energy topology map using a dynamic representation learning method based on the operating condition feature vector. The decision optimization unit is configured to use a multi-source decision fusion algorithm within the control domain to select a set of control instructions that meet the operational constraints and generate a scheduling sequence. The topology building unit includes: The node generation sub-unit is configured to obtain the charging and discharging characteristics of the energy storage unit and the conversion efficiency of the hydrogen production unit through the operation parameter extraction technology. Each equipment unit is used as a node in the energy topology graph and associated with the corresponding dynamic operation parameters. The relationship construction sub-unit is configured to create edges based on the power transmission path and energy conversion efficiency between device units, and set the initial transmission weight, wherein the weight value is initialized based on the energy conversion loss rate between devices; The dynamic update sub-unit is configured to track changes in equipment output data, grid connection status, and connection relationships based on a real-time monitoring mechanism. When new equipment is added or parameters are updated, an incremental update strategy is used to adjust the weights of associated nodes and edges. The heterogeneous deployment subunit is configured to select a distributed storage engine in a heterogeneous computing framework and deploy a real-time energy topology graph. The control domain localization unit includes: The topology embedding subunit is configured to pre-train the energy topology graph using a dynamic representation learning algorithm to generate a dynamic state embedding for each node. The dynamic state embedding space synchronously reflects the real-time power output capability of the equipment and the topology connection relationship. The vector alignment subunit is configured to map the working condition feature vector to the dynamic state embedding space, so that the working condition feature vector and the node dynamic state embedding are in the same vector space. The seed positioning subunit is configured to calculate the matching degree between the mapped working condition feature vector and the dynamic state embedding of all nodes, and select the preset number of nodes with the highest matching degree as the control seed nodes. The neighborhood expansion sub-unit is configured to expand adjacent nodes layer by layer according to the connection edge of the control seed node. If the state matching degree of the adjacent node reaches the running threshold, it is included in the control domain until the preset topology depth or control capacity threshold is met. The domain generation sub-unit is configured to determine the expanded set of nodes as the control domain, wherein the control domain includes device dynamic state embedding and energy topology subgraph.

2. The new energy power station electro-hydrogen coupled energy storage system according to claim 1, characterized in that, The feature parsing unit includes: The operating condition mapping subunit is configured to map the operating condition information to the feature space through a deep feature extraction network to generate a load fluctuation embedding vector. The historical analysis subunit is configured to extract the charging and discharging modes, hydrogen production efficiency curves, and scheduling response records from the historical output curves, construct a device operation profile, and convert it into a device status embedding vector. The database access subunit is configured to query the equipment operation database, extract power grid dispatching strategies and equipment operation thresholds, and generate strategy constraint embedding vectors. The intent is to fuse the sub-unit, configured to use a feature alignment algorithm to fuse the load fluctuation embedding vector, the equipment status embedding vector, and the policy constraint embedding vector to generate the operating condition feature vector, wherein the operating condition feature vector includes the power grid peak-shaving demand and the equipment operating boundary.

3. The new energy power station electro-hydrogen coupled energy storage system according to claim 1, characterized in that, The decision optimization unit includes: The feature extraction subunit is configured to extract a multi-dimensional operational feature vector for each device unit within the control domain; The fusion computing subunit is configured to aggregate multi-dimensional operational feature vectors using a multi-source decision fusion algorithm to generate a comprehensive device status vector, wherein the fusion process uses a weighted decision tree or a deep belief network. The instruction filtering subunit is configured to calculate the operational fit between the comprehensive state vector of the equipment and the operating condition feature vector, and to filter equipment with a fit higher than the operating threshold to generate a set of control instructions. The sequence generation subunit is configured to sort the set of control instructions based on operational fit, grid frequency deviation, and equipment response delay to generate a scheduling sequence.

4. The new energy power station electro-hydrogen coupled energy storage system according to claim 1, characterized in that, Also includes: The data alignment unit is configured to collect historical fault data of the electro-hydrogen coupling system, obtain multi-source operating parameters and fault equipment types for a preset period before the fault occurs, and mark the multi-source operating parameters and fault equipment types with timestamps and perform spatiotemporal alignment.

5. The new energy power station electro-hydrogen coupling energy storage system according to claim 4, characterized in that, Also includes: The evaluation interval division unit is configured to divide the multi-source operating parameters according to a fixed sampling interval, extract the operating mode components through signal decomposition technology, and divide the evaluation interval using an adaptive sliding window based on the spectral distribution characteristics of the mode components.

6. The new energy power station electro-hydrogen coupling energy storage system according to claim 5, characterized in that, Also includes: The status assessment unit is configured to calculate the equipment risk coefficient and health index within each assessment interval based on the number of voltage fluctuations at the grid connection point, the number of power over-limit events, and protection action records, combined with the time decay coefficient of the assessment interval, using a dynamic weight allocation algorithm.

7. The new energy power station electro-hydrogen coupling energy storage system according to claim 6, characterized in that, Also includes: The prediction optimization unit is configured to establish an operating status prediction model, input the equipment risk coefficient and health index of the current evaluation interval into the trained operating status prediction model, and output the predicted failure probability distribution of the next evaluation interval, wherein the operating status prediction model adopts an encoding and decoding neural network architecture.

8. A method for electro-hydrogen coupling energy storage in a new energy power station, applied to an electro-hydrogen coupling energy storage system for a new energy power station as described in any one of claims 1 to 7, characterized in that, include: Step 1: Receive the operating condition information of the new energy power station, and use a deep feature extraction network to generate an operating condition feature vector based on the real-time load fluctuations, historical output curves and equipment operation database in the operating condition information; Step 2: For the distributed energy storage devices and hydrogen production devices in the electro-hydrogen coupling system, construct an energy topology graph. The energy topology graph uses nodes to represent energy storage units or hydrogen production units and edges to represent energy conversion or power transmission relationships. The graph structure and node status are dynamically updated based on a real-time monitoring mechanism, and the energy topology graph is deployed on a heterogeneous computing framework. Step 3: Based on the operating condition feature vector, locate the control domain that matches the current operating condition in the energy topology map using a dynamic representation learning method; Step 4: Within the control domain, a multi-source decision fusion algorithm is used to select a set of control instructions that meet the operational constraints and generate a scheduling sequence.

Images