Energy-saving steam turbine water supply system

Abstract

The invention relates to the technical field of steam turbine water supply, and discloses an energy-saving steam turbine water supply system. A control instruction analysis module of the system receives a control instruction input by a user, a deep intention is mined through a semantic understanding engine, an energy-saving control target is recognized in combination with a real-time operation log and an equipment topology library, and a system state feature vector is generated. The dynamic graph generation module collects multi-source heterogeneous data of a water supply system assembly, constructs an equipment interaction graph with equipment entities as nodes and energy flow or instruction transfer relations as edges, and maintains the topological consistency of the equipment interaction graph in real time. The feature alignment engine module converts the system state feature vector into an embedded representation that is embedded in a shared space with the atlas node. And the neighborhood exploration module calculates the similarity between the embedded node and the map node, selects a root node and performs multi-hop traversal to form an energy-saving operation candidate area. And the strategy optimization module integrates the equipment state constraint and the energy efficiency index, generates a weighted operation sequence and outputs the weighted operation sequence.

Description

Technical Field

[0001] This invention relates to the field of steam turbine feedwater technology, specifically an energy-saving steam turbine feedwater system. Background Technology

[0002] As a crucial component of the thermal cycle, the turbine feedwater system's operating status directly impacts the overall thermal economy and safety of the unit. Traditional feedwater control systems often employ PID regulation based on classical control theory or simple logic interlocking schemes. Their design focus is typically on ensuring stable operation under rated conditions, with insufficient consideration given to exploring energy-saving potential during operation. These systems generally respond directly and superficially to control commands, lacking a deep understanding of the energy-saving intentions behind the commands issued by operators or higher-level systems. The command parsing process has a low correlation with the actual operating status of the equipment and the system topology, easily leading to deviations between control actions and overall energy efficiency targets.

[0003] In system modeling, existing technologies often treat water supply systems as simple series or parallel combinations of independent devices, analyzing them by establishing independent mathematical models for each device. This approach struggles to effectively characterize and utilize the complex energy interactions and information transfer relationships between devices. For example, there are close thermodynamic and hydraulic connections between devices such as water pumps, high-pressure heaters, and deaerators; changes in the operating parameters of one device can affect the operating status of other devices through working fluid flow and heat transfer. Traditional isolated models cannot holistically describe this dynamic relationship, leading to system optimization often being limited to local aspects and making it difficult to find optimal solutions from a global perspective.

[0004] When energy-saving adjustments are required, operators or automated systems typically rely on experience or predefined rule bases. These rule bases are often outdated, lagging behind gradual changes in equipment performance or system topology modifications, resulting in poor adaptability. Furthermore, the formulation of adjustment strategies lacks an intuitive, structured information support platform, failing to clearly demonstrate the "domain of influence" comprised of the equipment to be adjusted and its associated equipment. This leads to incomplete strategy evaluations and may produce unexpected chain reactions.

[0005] The decision-making process of existing systems has limited integration with real-time operational status. The information upon which decisions are based is not updated in a timely manner, or is incompatible with the current system topology, resulting in low feasibility of the generated strategies, and potentially even jeopardizing the safe operation of equipment. Furthermore, strategy optimization algorithms often focus on a single objective, lacking the ability to comprehensively weigh multiple conflicting or constraining criteria. Summary of the Invention

[0006] The purpose of this invention is to provide an energy-saving steam turbine feedwater system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an energy-saving steam turbine feedwater system, the system comprising: The control command parsing module is used to receive control commands for the turbine feedwater system input by the user. It uses a semantic understanding engine to perform deep intent mining on the control commands, and combines real-time operation logs and equipment topology library to identify energy-saving control targets and generate system state feature vectors. The dynamic graph generation module is used to collect multi-source heterogeneous data of water supply system components, construct equipment interaction graphs, where nodes represent water supply system equipment entities, edges represent energy flow paths or control command transmission relationships between equipment, and the topological consistency of the graph is maintained in real time based on a streaming data processing framework. The feature alignment engine module is used to convert system state feature vectors into graph-compatible embedding representations, so that they share the same vector space with the node embeddings in the device interaction graph. The neighborhood exploration module is used to calculate the cosine similarity between the embedded representation of the system state feature vector and the embedded representation of each node in the device interaction graph, select the node with the highest similarity as the root node, and perform multi-hop traversal along the edge to collect the set of associated nodes and form the energy-saving operation candidate area. The strategy optimization module is used to evaluate the operational feasibility of each node within the energy-saving operation candidate area using a multi-criteria decision algorithm, integrate equipment state constraints and energy efficiency indicators, generate a weighted operation sequence, and output it.

[0008] Preferably, the control command parsing module is specifically used for: Natural language processing technology is used to parse the text content of control commands, extract key operation parameters and action instructions, and form structured command objects. Access real-time operation logs, retrieve historical operation sequences and device response records, and build an operation context model; Query the equipment topology library to obtain the connection relationships and energy flow rules of the water supply system equipment, and generate an equipment dependency graph; The structured instruction object, operation context model, and device dependency graph are input into the intent recognition network, which uses an attention mechanism to fuse multi-source information and outputs an energy-saving control target. Based on the energy-saving control target coding system state feature vector, the vector dimension corresponds to the key operating parameters of the equipment.

[0009] Preferably, the dynamic map generation module is specifically used for: Asynchronously acquire operating parameters, equipment attributes, and relationship configuration data of water supply system components from sensor networks and data warehouses; The collected data is cleaned and normalized to remove outliers and redundant information, ensuring data quality. Define the architecture of the device interaction graph, where node attributes include device type, operating status and energy efficiency level, and edge attributes include energy flow direction, control link weight and real-time load; An interactive graph is stored using a graph database, and event listeners are deployed to monitor data changes. When device data updates are detected, the graph incremental update process is triggered to adjust node attributes and edge weights to maintain topology real-time performance.

[0010] Preferably, the feature alignment engine module is specifically used for: The device interaction graph is preprocessed using a graph representation learning algorithm to generate a low-dimensional embedding vector for each node. The embedding vector captures the structural features and attribute relationships of the node. Design a feature projection network, which includes fully connected layers and normalized layers, and takes the system state feature vector as input; The feature projection network is optimized by adversarial training, so that the projected output of the system state feature vector is aligned with the node embedding vector in terms of distribution. Finally, the projected vector is used as a graph-compatible embedding representation.

[0011] Preferably, the neighborhood exploration module is specifically used for: Calculate the cosine similarity between the embedded representation of the system state feature vector and the embedded representation of all nodes in the device interaction graph, sort them, and select the first k nodes as the initial root nodes. Starting from each root node, perform a breadth-first traversal along the edges, collecting direct and indirect neighbor nodes, until the traversal depth reaches the preset number of levels. During the traversal, the degree of similarity decay between each neighbor node and the system state feature vector is dynamically evaluated. If the decay exceeds the threshold, the branch traversal is terminated. Merge all nodes obtained from the traversal to form a candidate region for energy-saving operations, and record the path relationships between nodes.

[0012] Preferably, the strategy optimization module is specifically used for: For each node in the energy-saving operation candidate area, extract its corresponding equipment operating parameters, energy efficiency historical data, and maintenance status; Construct a multi-criteria decision-making model, with criteria including energy consumption indicators, equipment lifespan impact, and operational complexity; Each criterion is assigned a dynamic weight, which is adjusted based on real-time system load and priority strategy. A weighted summation algorithm is used to calculate the overall score of each node, and nodes with scores higher than the threshold value are selected. The selected nodes are serialized and arranged to form an operation sequence, and then output in descending order of score.

[0013] Preferably, the intent recognition network is implemented through the following steps: A multi-head self-attention layer is constructed to extract features from structured instruction objects, operation context models, and device dependency graphs; The extracted features are input into a gated fusion unit, which controls the information flow through a sigmoid function to fuse multi-source features. A classifier is applied to fuse multi-source features, and the classifier outputs the probability distribution of the energy-saving control target. The target with the highest probability is selected as the final energy-saving control target based on the probability distribution. The final energy-saving control target is encoded as a fixed-dimensional system state feature vector.

[0014] Preferably, the incremental map update process adopts a distributed transaction mechanism: When the event listener detects a data change, it generates a data change event and publishes it to the message queue. The consumer process retrieves events from the message queue, parses the changed content, and locks the affected nodes; The propagation range of the influence of a node is calculated, and the edges and neighboring nodes that need to be updated are determined by graph traversal algorithm; Batch update node attributes and edge weights within a transaction to ensure atomicity and consistency; After the update, verify the connectivity of the graph and record a version snapshot.

[0015] Preferably, the graph representation learning algorithm employs a hierarchical graph attention network: Perform community detection on the device interaction graph and divide the nodes into communities; Within each community, a graph attention layer is applied to aggregate neighbor node information and generate local community embeddings. At the global level, pooling operations are applied, and local community embeddings are integrated to generate global node embeddings. The community local embedding and the node global embedding are concatenated, and the final node embedding vector is generated through nonlinear transformation.

[0016] Preferably, the multi-criteria decision model employs fuzzy logic optimization: Define the membership function for each criterion to map the device operating parameters to a fuzzy set; Apply fuzzy inference rules and adjust the criterion weights based on the real-time system state; Use deblurring methods to convert the fuzzy output into an accurate score; Nodes are sorted based on scores, and operation sequence suggestions are generated.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention deepens the interpretation of user control commands through the semantic understanding engine of the control command parsing module, elevating simple command operations to the level of energy-saving target identification, making system behavior more aligned with users' actual expectations. The dynamic graph generation module constructs a device interaction graph, transforming previously isolated device data into a network model with rich relationships, intuitively presenting the energy and information flow structure within the system, providing underlying support for global analysis. The feature alignment engine module unifies command intent representation and physical system graph representation in vector space, enabling abstract control targets to be directly mapped to specific physical device networks for location and analysis.

[0018] The neighborhood exploration module, based on similarity calculation and multi-hop traversal, intelligently filters out the core devices and their associated device sets that are most relevant to the current control objective and most likely to produce energy-saving effects from a vast device network. This precisely defines the scope of optimization operations, avoiding the inefficiency of blindly searching the entire network. The strategy optimization module, supported by a multi-criteria decision-making algorithm, comprehensively evaluates each operation option within the candidate region. Its evaluation dimensions include not only energy efficiency indicators but also device state constraints, ensuring that the final generated operation sequence pursues energy-saving effects while also considering system safety and stability.

[0019] The entire system forms a complete closed loop from instruction understanding, system modeling, target mapping, scope delineation to strategy generation. This system organically integrates operators' experiential knowledge, the system's real-time operating status, and the inherent relationships between equipment, making the energy-saving operation decision-making process more systematic, scientific, and adaptive. This approach overcomes the limitations of relying on isolated models and static rules, helping to uncover deeper energy-saving potential while maintaining the overall coordination of the feedwater system operation, promoting the economical operation of turbine units, and greatly contributing to the development of the turbine industry. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the energy-saving steam turbine feedwater system described in this invention. Figure 2 A flowchart illustrating the workflow of the control instruction parsing module; Figure 3 A flowchart of the workflow for the dynamic map generation module; Figure 4 A comparison chart of the energy efficiency of different equipment in a steam turbine feedwater system; Figure 5 A comparison chart showing the optimization effects of the turbine feedwater system operation sequence. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides an energy-saving steam turbine feedwater system. The system includes: a control command parsing module that receives control commands input by the user, performs deep intent mining through a semantic understanding engine, identifies energy-saving control targets by combining real-time operation logs and an equipment topology library, and generates a system state feature vector. A dynamic graph generation module collects multi-source heterogeneous data of feedwater system components, constructs an equipment interaction graph, where nodes represent equipment entities, edges represent energy flow paths or control command transmission relationships, and maintains the topological consistency of the graph in real time based on a streaming data processing framework. A feature alignment engine module converts the system state feature vector into a graph-compatible embedding representation, making it share the same vector space with the node embeddings in the equipment interaction graph. A neighborhood exploration module calculates the cosine similarity between the embedding representation of the system state feature vector and the embeddings of each node in the equipment interaction graph, selects the node with the highest similarity as the root node, and performs multi-hop traversal along the edges to collect a set of associated nodes, forming an energy-saving operation candidate region. A strategy optimization module evaluates the operational feasibility of each node within the energy-saving operation candidate region using a multi-criteria decision algorithm, integrates equipment state constraints and energy efficiency indicators, generates a weighted operation sequence, and outputs it.

[0023] Example 1: See Figure 2 In its implementation, the control command parsing module uses natural language processing (NLP) technology to parse the user-input control command text. This process involves lexical and syntactic analysis of the input string. NLP employs pre-trained language models, such as those based on the Transformer architecture, to segment and tag the control commands, identifying key components such as verbs, nouns, and numerical parameters. Key operational parameters and action commands are extracted. Action commands include operation types such as start, stop, and adjust, while key operational parameters involve specific values ​​such as flow rate settings and pressure thresholds. A structured command object is then formed, encapsulated in JSON or XML format, containing fields such as operation type, target device identifier, and parameter value list. This transforms unstructured natural language commands into standardized, machine-readable, and processable data objects.

[0024] Access the real-time operation log, stored in a time-series database, which records historical operation data and event sequences of various components of the turbine feedwater system. Retrieve historical operation sequences and equipment response records. Historical operation sequences refer to the sequence of control commands executed within a past period, while equipment response records include information such as the status of equipment execution results and parameter change curves. Construct an operation context model, a dynamic data structure used to characterize the current operational stage and historical behavior patterns of the system. By analyzing the success rate of recent operation sequences, equipment response latency, and parameter fluctuations, the model infers the system's current load, stability, and potential risks, providing crucial background information for subsequent intent recognition.

[0025] The system queries the equipment topology database, a graphical database storing the connection relationships, energy flow logic, and control hierarchy of all equipment in the steam turbine feedwater system. It retrieves the connection relationships and energy flow rules of the feedwater system equipment. Connection relationships describe the physical or logical links between equipment, such as the pipe connection between pumps and valves. Energy flow rules define the paths and conditions for energy (such as thermal and mechanical energy) transfer between equipment. An equipment dependency graph is generated. This is a directed graph data structure where nodes represent equipment entities and edges represent dependencies between equipment. For example, the operating status of one device directly affects the availability or performance of another. This graph clearly depicts the mutual influence relationships among the components of the system. The structured instruction object, operation context model, and equipment dependency graph are input into the intent recognition network. This intent recognition network is a deep neural network model designed to fuse multi-source heterogeneous inputs and output control intents. The intent recognition network uses an attention mechanism to fuse multi-source information. This attention mechanism calculates the importance weights of different input features for the final decision, thus focusing on key information. Output energy-saving control targets. Energy-saving control targets are generalized outputs of the intention to identify the deeper meaning of the user's original instructions, such as "minimize pump energy consumption while ensuring output" or "prioritize the use of low-grade heat energy" and other specific optimization targets.

[0026] The energy-saving control target encoding system state feature vector is used to map the abstract energy-saving control target into a fixed-dimensional numerical vector. Each dimension corresponds to a key operating parameter of the equipment, such as the feedwater pump speed, heater outlet temperature, and condenser pressure. The vector's numerical value reflects the desired state or constraints of the energy-saving control target for these parameters. The intent recognition network extracts features from multi-source inputs by constructing a multi-head self-attention layer. This layer performs multiple self-attention calculations in parallel, each focusing on input information from a different representation subspace, thus more comprehensively capturing the complex relationships within and between structured instruction objects, operational context models, and equipment dependency graphs. The extracted features are input to a gated fusion unit, a network structure with a gating mechanism, such as a variant of the gated recurrent unit (GRU). The gated fusion unit controls the information flow through a sigmoid function, which generates a gating value between 0 and 1 to adjust the contribution of different input features in the fusion process, achieving selective information fusion, suppressing noise, and enhancing the transmission of useful signals.

[0027] A classifier is applied to fuse multi-source features. The classifier is typically a softmax function or a fully connected layer combined with a softmax function. The classifier outputs a probability distribution of the energy-saving control target. This probability distribution is a vector, where each element corresponds to a predefined energy-saving control target category, and its value represents the likelihood that the input feature belongs to that category. The target with the highest probability is selected as the final energy-saving control target based on the probability distribution; that is, the category label corresponding to the maximum value in the probability vector is selected as the network's exact output. The final energy-saving control target is encoded into a fixed-dimensional system state feature vector. This encoding operation is performed by a fully connected neural network layer. The vector dimension is predefined during the system design phase and is consistent with the embedding dimension of nodes in the device interaction graph, ensuring that subsequent modules can perform operations in the vector space. The system state feature vector, as the final output of the control command parsing module, is passed to the downstream feature alignment engine module.

[0028] In some embodiments, when natural language processing (NLP) parses the text content of control commands, a fine-tuning method based on the BERT model is employed. The BERT model is fine-tuned on turbine control corpora in a specific domain to improve the accuracy of understanding technical terms and command phrases. After extracting key operating parameters and action commands, a logical consistency check is performed to verify whether the parameter values ​​are within a reasonable range and whether the action commands match the equipment capabilities. The training process of the intent recognition network adopts a supervised learning approach. The training data consists of historical control commands, corresponding system operating status records, and real energy-saving control targets annotated by experts. During training, a cross-entropy loss function is used, and the Adam optimizer is employed to update the network parameters. Through iterative training, the network can accurately identify energy-saving control targets from the input.

[0029] Optionally, the operational context model can be constructed using time series modeling techniques, such as Long Short-Term Memory (LSTM) networks. LSTM networks can handle temporal dependencies in real-time runtime logs, capturing both long-term and short-term patterns in system operation, thus enabling the operational context model to more accurately reflect the dynamic behavior of the system.

[0030] Optionally, the generation of the device dependency graph can leverage a graph database query language, such as Cypher. By writing graph traversal query statements, the connections and energy flow paths between devices can be automatically extracted from the device topology library and instantiated into in-memory graph structure objects for use by the intent recognition network. The structured instruction objects, operational context models, and device dependency graphs, as inputs to the intent recognition network, need to be vectorized. Structured instruction objects can be converted into vectors using word embedding techniques, operational context models can be represented using the hidden states of sequence models, and device dependency graphs can have their graph-level representations extracted using graph neural networks. These vectorized representations are ultimately fed into a multi-head self-attention layer. The gating fusion unit in the intent recognition network can be replaced with other feature fusion mechanisms.

[0031] Example 2: See Figure 3In practical implementation, the dynamic map generation module asynchronously collects multi-source heterogeneous data from the water supply system components from a sensor network and a data warehouse. The sensor network consists of physical sensors deployed at the turbine water supply system site, such as temperature sensors, pressure sensors, flow sensors, and vibration sensors. The data warehouse integrates records from monitoring and data acquisition systems, historical databases, and equipment management systems. The collected operating parameters include real-time equipment readings such as temperature, pressure, flow rate, and speed; equipment attributes include static information such as model, rated power, energy efficiency rating, and installation location; and relational configuration data defines the logical connections, energy transfer paths, and control dependencies between equipment. The asynchronous acquisition mechanism allows different data sources to send data to the dynamic map generation module at their own speeds and timings. The module receives and buffers these data streams through a message middleware, thereby avoiding data congestion and improving system throughput. The collected data undergoes cleaning and normalization. The data cleaning process aims to identify and handle outliers, missing values, and noise in the raw data. For example, outlier detection algorithms based on statistical quantiles are used to identify and remove sensor readings that significantly deviate from the normal range. For missing values, interpolation or historical averages from the same device are used to fill in the gaps based on data characteristics. Normalization scales operating parameters of different dimensions and magnitudes to a uniform numerical range. For example, min-max normalization maps parameter values ​​to the range [0,1], or Z-score normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1. Removing outliers and redundant information ensures the consistency and comparability of the data upon which subsequent map construction relies, thus fundamentally guaranteeing data quality.

[0032] The architecture of the device interaction graph is defined. This graph is an attribute graph where nodes represent specific equipment entities in the water supply system, such as water pumps, high-pressure heaters, deaerators, and valves. Node attributes are designed to include multiple dimensions, such as equipment type, operating status, and energy efficiency level. Equipment type is used for classification and identification; operating status describes whether the equipment is currently running, stopped, faulty, or under maintenance; and energy efficiency level indicates the equipment's energy utilization efficiency under the current operating conditions. Edge attributes are designed to include energy flow direction, control link weight, and real-time load. Energy flow direction indicates the direction of energy transfer between devices; control link weight represents the strength or priority of control command transmission; and real-time load characterizes the magnitude of energy or material flow along the current connection path. This architecture endows the device interaction graph with rich semantic information, enabling it to accurately depict the static structure and dynamic operating status of the water supply system.

[0033] A graph database is used to store the device interaction graph. A graph database is a database management system specifically designed for storing and querying graph-structured data. Graph databases utilize nodes, edges, and attributes to efficiently represent and store complex relationships in the device interaction graph. Event listeners are deployed to monitor data changes. An event listener is a background process that continuously monitors data operation logs or changes to specific attribute fields in the graph database. When an event listener detects a device data update event from the data source, such as a change in the power reading of a water pump or an adjustment of the valve opening, it triggers an incremental graph update process.

[0034] The graph incremental update process employs a distributed transaction mechanism to maintain data consistency. When an event listener detects a data change, it generates a structured data change event object containing information such as the device identifier, change type, new value, and timestamp. This event is then published to a message queue, such as Kafka or RabbitMQ. The consumer process, acting as a subscriber to the message queue, retrieves the data change event, parses its details, and locks the affected nodes in the graph database based on the device identifier, preventing concurrent modifications by other processes during the update process. The propagation range of the node's impact is calculated using a graph traversal algorithm. For example, starting from the locked node, a breadth-first or depth-first traversal is performed to analyze its connecting edges and neighboring nodes, determining which edge attributes and neighboring node states may need to be updated. Batch updates of node attributes and edge weights within the distributed transaction ensure atomicity and consistency. Atomicity means the entire update operation either succeeds completely or is rolled back entirely, preventing intermediate partial updates. Consistency guarantees that the updated graph data meets predefined business rules and constraints. After the update operation is completed, the system will verify the connectivity of the graph, check for any isolated nodes or broken connections caused by the update, and record a version snapshot of the current graph state. The version snapshot can be used for data backtracking, version comparison, and fault recovery.

[0035] In some embodiments, the data cleaning and normalization module is implemented using a streaming processing framework, such as the streaming capabilities of Apache Flink. Streaming frameworks allow data to be cleaned and normalized immediately upon entering the dynamic graph generation module, rather than being stored in batches and then processed. This significantly reduces data processing latency, enabling the device interaction graph to reflect the latest system state more quickly. Event listeners can be configured to be based on a complex event processing engine. This engine can identify patterns in the data stream; for example, it only determines a meaningful event requiring a graph update when a set of associated device parameters changes continuously within a specific time window. This filters out isolated, insignificant data fluctuations and reduces unnecessary graph update operations.

[0036] Optionally, the graph database selection can consider databases that support native graph storage and query languages. Native graph databases offer performance advantages when handling deep join queries and graph traversal operations. For example, using query languages ​​like Cypher or Gremlin can efficiently express and execute the complex graph traversal logic required in incremental graph update processes. Distributed transaction mechanisms can be implemented using a two-phase commit protocol. In a two-phase commit protocol, a coordinator process coordinates multiple participant processes. In the first phase, it queries all participants whether the update can be committed. In the second phase, based on the feedback from all participants, it decides whether to commit globally or roll back globally, thus ensuring the atomicity of update operations across multiple nodes.

[0037] It is understandable that the architecture definition of the device interaction graph is the foundation of the entire dynamic graph generation module. The definition of node and edge attributes needs to be closely integrated with the domain knowledge of the turbine feedwater system. The completeness and accuracy of the attribute design directly determine whether the device interaction graph can effectively support subsequent advanced analysis functions such as neighborhood exploration and strategy optimization. The combination of asynchronous acquisition mechanism and streaming data processing framework enables the dynamic graph generation module to handle high-frequency, high-volume industrial real-time data streams. This design avoids the data lag problem caused by traditional batch processing methods, providing a timely and accurate data foundation for energy-saving control applications with high real-time requirements.

[0038] See Figure 4 This chart is a core visualization of the energy efficiency assessment of an energy-saving steam turbine feedwater system. It focuses on six core pieces of equipment: feedwater pumps, high-pressure heaters, low-pressure heaters, deaerators, condensers, and the steam turbine, comparing the energy efficiency percentages of traditional and energy-saving systems. Yellow bars represent traditional systems, and blue bars represent energy-saving systems, clearly showing the energy efficiency differences of each piece of equipment under the two systems. This energy efficiency improvement across the entire equipment chain verifies the technical value of the dynamic graph generation module in collecting multi-source heterogeneous data, constructing equipment interaction graphs, and integrating energy efficiency level attributes. Through the collection, cleaning, and analysis of energy efficiency data from each piece of equipment, it intuitively demonstrates the energy efficiency optimization results of the energy-saving system at the equipment level. This provides fundamental data support for subsequent domain exploration and strategy optimization, and also reflects the system's global energy efficiency optimization logic from the equipment layer to the system layer. It is a key basis for understanding how this system taps into energy efficiency potential through equipment interaction graphs.

[0039] Example 3: In specific implementation, the feature alignment engine module uses a graph representation learning algorithm to preprocess the device interaction graph. Graph representation learning is a machine learning technique that maps nodes in a graph to a low-dimensional vector space. A low-dimensional embedding vector is generated for each node. This low-dimensional embedding vector is a fixed-length array of real numbers, with a dimension much lower than the dimension of the node attributes in the original graph. The embedding vector captures the structural features and attribute relationships of the nodes. Structural features refer to the node's position, connectivity, and role in the overall graph topology, while attribute relationships reflect the intrinsic connection between the node's own attributes and its vector representation. The graph representation learning algorithm is implemented using a hierarchical graph attention network. This network learns node representations through hierarchical aggregation, first aggregating information within a local community and then integrating it at the global level.

[0040] Community detection is performed on the device interaction graph. Community detection is a graph partitioning algorithm that aims to discover subsets of nodes in a graph that are tightly connected internally and sparsely connected externally. The process involves dividing the entire device interaction graph into several overlapping or non-overlapping communities. Each community contains a group of device nodes that are functionally or connectivity-related. Community detection uses criteria such as modularity optimization to find a partitioning scheme that maximizes the edge density within each community. A graph attention layer is applied within each community. Graph attention layers are a type of layer in graph neural networks that assign different importance weights to each neighbor of a node through an attention mechanism. Neighbor node information is aggregated. This aggregation operation involves weighted summation of the feature vectors of the central node's neighbors according to the attention weights, generating a new feature vector that incorporates neighborhood information. Finally, a community local embedding is generated. The community local embedding is a vector representation of a node within its own community, primarily encoding the node's structural and attribute information within the local community.

[0041] At the global level, pooling is applied. Pooling is a downsampling technique used to extract more representative global features from the community-local embedding set. Community-local embeddings are then fused. This process integrates node-local embeddings from different communities, for example, by calculating the mean or maximum value of all community-local embeddings, generating a representation that reflects the node's features across the entire graph. A global node embedding is generated. This global node embedding is a comprehensive vector representation that captures the node's global position and importance relative to the entire device interaction graph. The community-local embeddings and global node embeddings are concatenated. This concatenation operation joins the two vectors dimensionally to form a longer new vector. Finally, a nonlinear transformation is used to generate the final node embedding vector. This nonlinear transformation is typically implemented using a fully connected layer and an activation function. The goal is to fuse the concatenated vector and introduce nonlinear expressive power. The resulting node embedding vector contains both local and global information about the node.

[0042] Design a feature projection network, a feedforward neural network, whose function is to project the system state feature vector from the control instruction parsing module into the same vector space as the node embedding vectors. The feature projection network contains fully connected layers and normalization layers. The fully connected layers perform linear transformations, mapping the dimension of the input vector to the target dimension. The normalization layers stabilize the training process, accelerate convergence, and improve generalization ability. The system state feature vector, which is the output of the control instruction parsing module, is used as the input to the feature projection network, representing the semantics of the user instructions and the current state requirements of the system. The feature projection network is optimized through adversarial training, a training strategy that introduces a discriminator network to distinguish the projected vectors generated by the feature projection network from the real node embedding vectors. The goal is to align the distribution of the projected output of the system state feature vector with that of the node embedding vector. Distribution alignment means that the statistical distribution characteristics of the system state feature vector transformed by the feature projection network in the vector space are as similar as possible to the distribution characteristics of the node embedding vectors generated by the hierarchical graph attention network.

[0043] Finally, the projected vector is used as a graph-compatible embedding representation. This graph-compatible embedding representation shares the same vector space as the node embeddings in the device interaction graph, allowing the subsequent neighborhood exploration module to directly calculate metrics such as cosine similarity between the system state feature vector embedding representation and the node embeddings. The optimization objective during adversarial training can be expressed as minimizing the projection error of the feature projection network while maximizing the discrimination error of the discriminator network. The loss function of the feature projection network... It can be designed as follows: in: This represents the input system state feature vector. The mapping function represents the feature projection network. This represents the output of the discriminator network (a scalar between 0 and 1, representing the probability that the input vector is identified as a true node embedding). This represents the mathematical expectation. This loss function encourages the feature projection network. Generate a device that can "deceive" the discriminator. The projection vector is what makes the discriminator think that the projection vector is the real node embedding.

[0044] In some embodiments, community detection in a hierarchical graph attention network can employ the Louvain algorithm. The Louvain algorithm is a heuristic algorithm based on modularity optimization, capable of efficiently discovering hierarchical community structures in large-scale graphs, suitable for industrial scenarios such as device interaction graphs that may contain hundreds or even thousands of nodes. Attention weight calculation in the graph attention layer can employ an additive attention mechanism. The additive attention mechanism calculates attention scores through a single-layer feedforward neural network. This network concatenates or adds the feature vectors of the center node and its neighbors, mapping them to a scalar score, which is then normalized to weights using a softmax function. Adversarial training can use the Wasserstein distance to measure distribution differences. Adversarial training based on Wasserstein distance typically has better training stability. It uses a discriminator network to estimate the Wasserstein distance between the projected vector distribution and the true node embedding distribution, guiding the feature projection network to minimize this distance.

[0045] Optionally, the feature projection network structure can include multiple residual connections. Residual connections can alleviate the gradient vanishing problem in deep networks, facilitating the training of deeper feature projection networks and potentially learning more complex nonlinear mappings from the system state feature space to the node embedding space. Compared to traditional graph convolutional networks, hierarchical graph attention networks have the advantage of better capturing structural information at different scales in the graph through community partitioning and hierarchical aggregation. The attention mechanism within communities allows nodes to focus on the most relevant information among their direct neighbors, while global pooling integrates broader graph context information, resulting in generated node embedding vectors with richer semantics. The core challenge of the feature alignment engine module lies in how to map two vectors from different sources and with different representational meanings to the same meaningful vector space. Adversarial training provides an effective, data-driven method to enforce this distribution alignment without explicitly defining complex mapping rules between the two spaces.

[0046] Example 4: In specific implementation, the neighborhood exploration module calculates the cosine similarity between the embedded representation of the system state feature vector and the embeddings of all nodes in the device interaction graph. Cosine similarity is a metric that measures the difference in direction between two vectors, with a value range between -1 and 1. The calculation process involves performing a dot product operation on the two vectors and then dividing by the product of their respective L2 norms. The embedded representation of the system state feature vector comes from the output of the feature alignment engine module, while the embeddings of all nodes in the device interaction graph are dynamically read from the graph database. After sorting, the first k nodes are selected as the initial root nodes. The sorting operation is performed in descending order according to the cosine similarity values, where k is a positive integer hyperparameter. Its value needs to be manually set or adaptively adjusted according to the size of the device interaction graph and the actual application's requirements for computational efficiency and coverage. Selecting the first k nodes means selecting the k device nodes that are semantically closest to the current system state feature vector as the starting point for subsequent graph traversal.

[0047] Starting from each root node, a breadth-first search (BFS) traversal is performed along the edges. BFS is a graph traversal algorithm whose core principle is to prioritize visiting all direct neighbors of the current node, then the neighbors of those neighbors, and so on. A first-in, first-out (FIFO) queue data structure is used to manage the order of nodes to be visited, with the root node initially added to the queue. Direct and indirect neighbors are collected. Direct neighbors are nodes directly connected to the current node by an edge, while indirect neighbors are nodes that require two or more edges to reach. This continues until a preset depth is reached. The preset depth defines the maximum number of hops that can be traversed from the root node. For example, setting it to 3 levels means that the root node's direct neighbors (1 hop), neighbors of those neighbors (2 hops), and neighbors of those neighbors of those neighbors (3 hops) can be explored. During the traversal, the similarity decay between each neighbor node and the system state feature vector is dynamically evaluated. This similarity decay is calculated by comparing the cosine similarity between the currently visited neighbor node and the system state feature vector with the initial cosine similarity of the root node, and calculating the percentage decrease in similarity. If the decay exceeds a threshold, the branch traversal is terminated. The threshold is a predefined percentage value, such as 50%. This means that if the similarity of a node on a branch decreases by more than 50% compared to the root node, the traversal will not continue from that node, and the exploration of that branch will end prematurely. This is a pruning strategy used to avoid unnecessary calculations in graph regions where semantic relevance has been significantly reduced, thereby improving search efficiency.

[0048] All nodes obtained from traversal are merged to form an energy-saving operation candidate region. This merging operation removes nodes that appear repeatedly in traversal paths from different root nodes, ensuring that each node in the candidate region is unique. Path relationships between nodes are recorded, detailing which root node each node was accessed through and which intermediate nodes it passed through. This information is stored as an edge sequence, aiding in understanding the association patterns between nodes. The energy-saving operation candidate region is essentially a subgraph in the device interaction graph, containing a group of topologically adjacent device nodes related to the current control command intent. The breadth-first traversal algorithm ensures that we prioritize exploring node regions closer to the root node and generally with higher relevance, then gradually expand to the outer regions. The similarity decay threshold and maximum traversal depth work together to achieve a balance between search scope and computational cost. Table 1 shows a configuration of the similarity decay threshold and maximum traversal depth for different device types; the actual configuration values ​​need to be determined based on specific system debugging.

[0049] Table 1: Neighborhood Exploration Parameter Configuration The entire exploration process is based on a graph traversal algorithm, which requires processing large-scale graph structure data stored in a graph database. The neighborhood exploration module integrates a caching mechanism. This mechanism embeds frequently accessed nodes into vectors and temporarily stores traversal results of high-frequency paths in memory. When encountering the same root node or similar system states again, partial results can be retrieved directly from the cache, thus accelerating similarity calculation and traversal. Traversal depth and threshold parameters are managed through configuration files and can be adjusted online. This allows system administrators to dynamically optimize the breadth and depth of neighborhood exploration without restarting the system, adapting to different operating scenarios and performance requirements. The module outputs an energy-saving operation candidate region, which is a set of nodes and their inter-node relationships. This candidate region serves as input to the strategy optimization module, ensuring that subsequent decision analysis is based on a set of devices closely related to the current control objective.

[0050] In some embodiments, breadth-first traversal can be implemented iteratively rather than recursively. Iterative methods use loops and queue data structures, which, compared to recursive depth-first search, make it easier to control the traversal depth and avoid the risk of stack overflow during recursion, making them particularly robust when dealing with large-scale graphs. When selecting the top k root nodes, diversity constraints can be introduced. These constraints consider not only similarity ranking but also the distribution of the k root nodes in the graph, preventing the selected root nodes from concentrating in the same region and ensuring that the explored candidate regions cover a wider variety of device types and topology paths. When dynamically evaluating similarity decay, a moving average method can be used. The moving average method calculates the average similarity of the current node and its most recently visited parent nodes, then compares it with the root node similarity. This smooths out fluctuations in single-point similarity, making branch termination decisions more stable and reducing the possibility of premature pruning due to accidental low similarity of a single node. The core function of the neighborhood exploration module is to quickly locate a connected subgraph region highly relevant to the current control command within the vast device interaction graph. This method combines semantic similarity-based node retrieval with graph topology-based neighborhood expansion, which greatly improves search efficiency compared to full graph search.

[0051] Example 5: In specific implementation, the strategy optimization module extracts the corresponding equipment operating parameters, historical energy efficiency data, and maintenance status for each node in the energy-saving operation candidate area. Equipment operating parameters include current values ​​read from the real-time database, such as the speed, outlet pressure, and motor current of the feedwater pump, and the inlet steam temperature and condensate level of the high-pressure heater. Historical energy efficiency data refers to long-term operating efficiency records and unit output energy consumption curves stored in the historical database. The maintenance status comes from the equipment management system, indicating whether the equipment is within its planned maintenance period and whether there are recent fault records or performance degradation alarms. The data extraction process is implemented by concurrently querying multiple data source interfaces to ensure the completeness and timeliness of the information obtained, providing a detailed data foundation for subsequent decision-making. A multi-criteria decision-making model is constructed. This model is an evaluation framework used to comprehensively consider influencing factors from different dimensions. The criteria include energy consumption indicators, equipment lifespan impact, and operational complexity. Energy consumption indicators quantify the expected changes in system energy consumption after an operation is performed, such as the power savings in kilowatt-hours that might occur if the speed of a pump is reduced. Equipment lifespan impact assesses the degree to which the operation accelerates or slows down aging factors such as mechanical wear and thermal fatigue. Operational complexity measures the number of manual intervention steps, technical requirements, and potential execution risks required to implement the operation. Dynamic weights are assigned to each criterion, and these weights are adjusted based on real-time system load and priority strategies. For example, the weight of energy consumption indicators is significantly increased during peak electricity consumption periods, while the weight of equipment lifespan impact increases at the end of equipment lifespan or in a warning state. Priority strategies are pre-set by system maintenance personnel, defining the optimization focus under different operating conditions.

[0052] A weighted summation algorithm is used to calculate the comprehensive score for each node. This algorithm multiplies the score of each criterion by its dynamic weight and then sums the results to obtain a comprehensive score representing the priority of the node's operation. Nodes with scores higher than a configurable threshold are filtered out. This threshold removes operation options with insignificant overall benefits or excessive negative impacts, ensuring that the final recommended operation sequence is practically feasible and has a positive expected effect. The filtered nodes are then serialized and arranged to form an operation sequence. This serialization and arrangement needs to consider the technical logical order and time dependencies between operations. For example, a standby pump must be started before the load of the original operating pump can be reduced, or adjusting the heater inlet valve opening must be done before adjusting the drain valve opening to avoid system disturbance. The arrangement process generates an ordered list of operation steps and outputs them in descending order of score. The operation suggestion corresponding to the node with the highest score is placed at the top of the sequence, indicating its priority for execution.

[0053] In practical implementation, the multi-criteria decision-making model employs fuzzy logic optimization to handle uncertainties and qualitative knowledge in the evaluation process. Fuzzy logic optimization is a method based on fuzzy set theory for processing imprecise information. A membership function is defined for each criterion. This function maps precise equipment operating parameter values ​​to membership degrees in a fuzzy set. For example, for the "operational complexity" criterion, a fuzzy set ranging from "very simple" to "extremely complex" can be defined. The membership function transforms inputs such as the number of specific operation steps and required skill levels into the degree to which the values ​​belong to each fuzzy linguistic value. Triangular or trapezoidal functions are typically used for membership functions. Fuzzy inference rules are applied. These rules are "if-then" rules based on expert experience, used to describe the relationship between the input and output fuzzy sets. For example, a rule might be: "If energy savings are high and equipment lifespan is minimally affected, then the operation priority is high." The inference process activates different rules based on the real-time system state and synthesizes the conclusions of all activated rules. Defuzzing methods are used to convert fuzzy outputs into precise scores. Defuzzing is the inverse process of fuzzy inference. It transforms the fuzzy set of output generated by fuzzy inference into a precise numerical value that can be used for sorting and calculation. Commonly used defuzzing methods include the centroid method or the maximum membership method.

[0054] Taking a specific feedwater pump node as an example, assuming the energy-saving operation candidate area includes an electric feedwater pump numbered FP-101, the strategy optimization module extracts the current operating parameters of FP-101: speed 3000rpm, power consumption 2.5MW, and outlet pressure 15MPa. Historical energy efficiency data shows that the pump's efficiency under current operating conditions is 82%, slightly lower than the design value of 85%. Maintenance status shows that the pump's operating time is nearing its overhaul cycle, but there have been no recent abnormal alarms. The multi-criteria decision model begins to evaluate the operational suggestion for FP-101: "reduce the speed to 2950rpm". For the energy consumption index criterion, based on the pump's similarity characteristic curve, it is estimated that the power consumption can be reduced to 2.38MW after the speed reduction, saving 0.12MW. The membership function maps this 0.12MW saving amount to the "significant energy-saving effect" fuzzy set with a membership degree of 0.7. For the equipment life impact criterion, reducing the speed usually helps reduce bearing wear and extend the pump's life. The membership function maps this positive impact to the "extended life" fuzzy set with a membership degree of 0.8. For the operational complexity criterion, speed reduction can be achieved by adjusting the inverter commands, which is simple to operate and has low risk. The membership function maps it to the "simple operation" fuzzy set with a membership degree of 0.9. Assuming the current system load is high, the priority strategy assigns a weight of 0.5 to energy consumption indicators, 0.3 to equipment lifespan impact, and 0.2 to operational complexity. Relevant rules in the fuzzy inference rule base are activated, such as rule 1: "If the energy-saving effect is significant, lifespan is extended, and operation is simple, then the priority is high"; rule 2: "If the energy-saving effect is moderate and lifespan is extended, then the priority is high." The inference engine aggregates the output of these rules to form a fuzzy set about "operational priority." Finally, defuzzification is performed using the centroid method to obtain a precise comprehensive score, such as 85 points (out of 100). If this score is higher than the set threshold of 70 points, the speed reduction operation of the FP-101 pump is included in the candidate sequence.

[0055] In some embodiments, the construction of the fuzzy inference rule base can be optimized using machine learning methods. By collecting historical operation records and their actual energy-saving effects, as well as equipment status change data, a model is trained to automatically adjust or generate inference rules. This allows the fuzzy logic optimization model to learn more accurate evaluation patterns from the data, reducing reliance on the completeness of initial expert knowledge. The defuzzification process can employ the area bisector method. The area bisector method divides the total area under the membership function curve of the output fuzzy set into two equal parts, and the x-coordinate value corresponding to the dividing point is used as the precise output value. This method may be simpler than the centroid method in some cases and is less sensitive to the shape of the membership function.

[0056] Optionally, the dynamic adjustment of criterion weights can be based on reinforcement learning algorithms. Through interaction with the water supply control system, reinforcement learning algorithms automatically learn the optimal weight configuration for each criterion under different system states, based on long-term accumulated energy-saving benefits and equipment reliability returns, achieving adaptive optimization. The arrangement of operation sequences can incorporate graph programming algorithms. Graph programming algorithms represent the preconditions, mutual exclusions, and temporal constraints between operation steps as a graph structure, searching the graph for the operation path that satisfies all constraints and has the minimum cost, thereby generating logically more rigorous and safer operation sequences. For the assessment of equipment lifespan impact, a physics-based degradation model can be integrated. The physics degradation model quantitatively predicts the impact of operations on the remaining service life of the equipment by simulating physical processes such as stress, strain, and wear under specific operating conditions, providing more accurate, mechanism-based input data for equipment lifespan impact criteria.

[0057] It is understandable that the strategy optimization module combines multi-criteria decision-making with fuzzy logic optimization, effectively addressing the inherent uncertainties and multi-objective trade-offs in energy-saving operation assessments. Fuzzy logic allows for the use of expert knowledge described in natural language and unifies qualitative and quantitative factors within the assessment framework. Weighted summation and defuzzification ultimately produce a quantifiable decision-making basis for ranking. This implementation method ensures that the generated weighted operation sequence not only pursues optimal energy efficiency but also considers equipment asset health and operational feasibility, embodying the concept of comprehensive optimization.

[0058] See Figure 5 This graph focuses on the optimization results of the operation sequence of the energy-saving steam turbine feedwater system, comparing the effectiveness scores of five key operations—adjusting feedwater pump speed, adjusting high-pressure heaters, optimizing deaerator pressure, adjusting condensing temperature, and adjusting valve opening—before and after optimization. Orange bars represent before optimization, and blue bars represent after optimization, clearly showing the efficiency improvement of each operation. These data intuitively verify the technical value of the strategy optimization module's multi-criteria decision algorithm, which integrates equipment state constraints and energy efficiency indicators to generate weighted operation sequences. Through comprehensive evaluation and dynamic weight allocation of multiple criteria such as energy consumption indicators, equipment lifespan impact, and operational complexity, the optimization effect of each operation is significantly improved. This graph not only quantifies the actual effectiveness of operation sequence optimization but also reflects the full-process value of the strategy optimization module from candidate area selection to sequence arrangement. It is a key basis for understanding how the system achieves global energy efficiency optimization while ensuring equipment safety, providing intuitive effect verification for the economic operation and energy-saving operation decisions of the steam turbine unit.

[0059] 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.

[0060] 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. An energy-saving steam turbine feedwater system, characterized in that, include: The control command parsing module is used to receive control commands for the turbine feedwater system input by the user. It uses a semantic understanding engine to perform deep intent mining on the control commands, and combines real-time operation logs and equipment topology library to identify energy-saving control targets and generate system state feature vectors. The dynamic graph generation module is used to collect multi-source heterogeneous data of water supply system components, construct equipment interaction graphs, where nodes represent water supply system equipment entities, edges represent energy flow paths or control command transmission relationships between equipment, and the topological consistency of the graph is maintained in real time based on a streaming data processing framework. The feature alignment engine module is used to convert system state feature vectors into graph-compatible embedding representations, so that they share the same vector space with the node embeddings in the device interaction graph. The neighborhood exploration module is used to calculate the cosine similarity between the embedded representation of the system state feature vector and the embedded representation of each node in the device interaction graph, select the node with the highest similarity as the root node, and perform multi-hop traversal along the edge to collect the set of associated nodes and form the energy-saving operation candidate area. The strategy optimization module is used to evaluate the operational feasibility of each node within the energy-saving operation candidate area using a multi-criteria decision algorithm, integrate equipment state constraints and energy efficiency indicators, generate a weighted operation sequence, and output it.

2. The energy-saving steam turbine feedwater system according to claim 1, characterized in that, The control command parsing module is specifically used for: Natural language processing technology is used to parse the text content of control commands, extract key operation parameters and action instructions, and form structured command objects. Access real-time operation logs, retrieve historical operation sequences and device response records, and build an operation context model; Query the equipment topology library to obtain the connection relationships and energy flow rules of the water supply system equipment, and generate an equipment dependency graph; The structured instruction object, operation context model, and device dependency graph are input into the intent recognition network, which uses an attention mechanism to fuse multi-source information and outputs an energy-saving control target. Based on the energy-saving control target coding system state feature vector, the vector dimension corresponds to the key operating parameters of the equipment.

3. The energy-saving steam turbine feedwater system according to claim 1, characterized in that, The dynamic graph generation module is specifically used for: Asynchronously acquire operating parameters, equipment attributes, and relationship configuration data of water supply system components from sensor networks and data warehouses; The collected data is cleaned and normalized to remove outliers and redundant information, ensuring data quality. Define the architecture of the device interaction graph, where node attributes include device type, operating status and energy efficiency level, and edge attributes include energy flow direction, control link weight and real-time load; An interactive graph is stored using a graph database, and event listeners are deployed to monitor data changes. When device data updates are detected, the graph incremental update process is triggered to adjust node attributes and edge weights to maintain topology real-time performance.

4. The energy-saving steam turbine feedwater system according to claim 1, characterized in that, The feature alignment engine module is specifically used for: The device interaction graph is preprocessed using a graph representation learning algorithm to generate a low-dimensional embedding vector for each node. The embedding vector captures the structural features and attribute relationships of the node. Design a feature projection network, which includes fully connected layers and normalized layers, and takes the system state feature vector as input; The feature projection network is optimized by adversarial training, so that the projected output of the system state feature vector is aligned with the node embedding vector in terms of distribution. Finally, the projected vector is used as a graph-compatible embedding representation.

5. The energy-saving steam turbine feedwater system according to claim 1, characterized in that, The neighborhood exploration module is specifically used for: Calculate the cosine similarity between the embedded representation of the system state feature vector and the embedded representation of all nodes in the device interaction graph, sort them, and select the first k nodes as the initial root nodes. Starting from each root node, perform a breadth-first traversal along the edges, collecting direct and indirect neighbor nodes, until the traversal depth reaches the preset number of levels. During the traversal, the degree of similarity decay between each neighbor node and the system state feature vector is dynamically evaluated. If the decay exceeds the threshold, the branch traversal is terminated. Merge all nodes obtained from the traversal to form a candidate region for energy-saving operations, and record the path relationships between nodes.

6. The energy-saving steam turbine feedwater system according to claim 1, characterized in that, The strategy optimization module is specifically used for: For each node in the energy-saving operation candidate area, extract its corresponding equipment operating parameters, energy efficiency historical data, and maintenance status; Construct a multi-criteria decision-making model, with criteria including energy consumption indicators, equipment lifespan impact, and operational complexity; Each criterion is assigned a dynamic weight, which is adjusted based on real-time system load and priority strategy. A weighted summation algorithm is used to calculate the overall score of each node, and nodes with scores higher than the threshold value are selected. The selected nodes are serialized and arranged to form an operation sequence, and then output in descending order of score.

7. The energy-saving steam turbine feedwater system according to claim 2, characterized in that, The intent recognition network is implemented through the following steps: A multi-head self-attention layer is constructed to extract features from structured instruction objects, operation context models, and device dependency graphs; The extracted features are input into a gated fusion unit, which controls the information flow through a sigmoid function to fuse multi-source features. A classifier is applied to fuse multi-source features, and the classifier outputs the probability distribution of the energy-saving control target. The target with the highest probability is selected as the final energy-saving control target based on the probability distribution. The final energy-saving control target is encoded as a fixed-dimensional system state feature vector.

8. The energy-saving steam turbine feedwater system according to claim 3, characterized in that, The incremental update process for the graph adopts a distributed transaction mechanism: When the event listener detects a data change, it generates a data change event and publishes it to the message queue. The consumer process retrieves events from the message queue, parses the changed content, and locks the affected nodes; The propagation range of the influence of a node is calculated, and the edges and neighboring nodes that need to be updated are determined by graph traversal algorithm; Batch update node attributes and edge weights within a transaction to ensure atomicity and consistency; After the update, verify the connectivity of the graph and record a version snapshot.

9. The energy-saving steam turbine feedwater system according to claim 4, characterized in that, The graph representation learning algorithm employs a hierarchical graph attention network: Perform community detection on the device interaction graph and divide the nodes into communities; Within each community, a graph attention layer is applied to aggregate neighbor node information and generate local community embeddings. At the global level, pooling operations are applied, and local community embeddings are integrated to generate global node embeddings. The community local embedding and the node global embedding are concatenated, and the final node embedding vector is generated through nonlinear transformation.

10. The energy-saving steam turbine feedwater system according to claim 6, characterized in that, The multi-criteria decision-making model employs fuzzy logic optimization. Define the membership function for each criterion to map the device operating parameters to a fuzzy set; Apply fuzzy inference rules and adjust the criterion weights based on the real-time system state; Use deblurring methods to convert the fuzzy output into an accurate score; Nodes are sorted based on scores, and operation sequence suggestions are generated.

Images