Distributed intelligent sensing based power plant equipment collaborative control system and method

Abstract

The application discloses a power plant equipment collaborative control system and method based on distributed intelligent sensing, relates to the field of equipment collaborative control, and specifically comprises the following steps: firstly, obtaining original distributed data streams such as furnace acoustic wave temperature measurement, pipeline wireless pressure and rotating equipment vibration and power grid instructions; performing asynchronous data alignment through vectorization processing and space-time atlas construction; reconstructing discrete and non-equivalent time sensor data into a space-time alignment state matrix containing physical correlation; further extracting a system potential coupling feature tensor; predicting fast and slow dynamic trajectories of the system respectively by adopting a double-time scale mechanism; accurately mapping dynamic differences between boiler energy accumulation and steam turbine rapid response; finally, performing multi-objective rolling optimization based on the predicted trajectories, calculating optimal control increments under the premise of ensuring system stability, and realizing rapid response and accurate collaboration to the power grid dispatching instructions.

Description

Technical Field

[0001] This application relates to the field of equipment collaborative control technology, specifically to a collaborative control system and method for thermal power plant equipment based on distributed intelligent sensing. Background Technology

[0002] With the large-scale grid connection of clean energy, the role of thermal power plants in the power system is gradually shifting from baseload power to regulating power. The power grid is placing extremely high demands on the rapid load change capability (AGC) and deep peak-shaving performance of these units. To improve the accuracy of sensing complex thermal systems, modern thermal power plants have introduced monitoring systems based on distributed intelligent sensing, such as spatial acoustic temperature measurement arrays in the furnace combustion zone, wireless pressure transmitters for key steam and water pipelines, and wireless vibration monitoring nodes for rotating equipment. These distributed sensing technologies can provide more refined spatial distribution information and equipment status data than traditional DCS systems, aiming to achieve refined coordinated control of the boiler and turbine systems through comprehensive status perception, thereby unlocking the rapid regulation potential of the units while ensuring safety.

[0003] However, in existing collaborative control schemes for thermal power plant equipment, directly applying distributed intelligent sensor data to real-time control still faces significant challenges, primarily due to the spatiotemporal asynchronicity of the data and the strong coupling characteristics between the boiler and turbine. On one hand, unlike the synchronous periodic sampling of wired sensors in traditional DCS systems, distributed intelligent sensors often employ low-power transmission protocols or event-triggered mechanisms, resulting in inconsistent timestamps of data arrival at the controller. Furthermore, since thermal power plants are typically large-scale distributed parameter systems, sensor data from different spatial locations (such as the pulverizer outlet and economizer inlet) inherently experience physical transmission delays. Most existing control systems employ simple zero-order hold (ZOH) or linear interpolation to handle this asynchronous data, ignoring the spatiotemporal correlation of data during physical transmission. This leads to the constructed "current state" often being a patchwork of past states, failing to accurately reflect the system's real-time dynamics. On the other hand, the coordinated control system of a thermal power plant's boiler and turbine exhibits significant multi-capacity inertia and multi-timescale characteristics. The boiler side experiences slow energy accumulation and high inertia, and under large time delays and asynchronous data interference, traditional PID or model predictive control (MPC) algorithms struggle to accurately predict the dynamic coupling effects between the boiler and turbine. In order to avoid system overshoot or even oscillation caused by data lag, engineering practice often forces the reduction of control gain. While this maintains stability, it leads to a slow adjustment rate, which cannot meet the grid dispatching requirements for rapid response. Summary of the Invention

[0004] To address the problems in the prior art, according to one aspect of this application, a method for coordinated control of thermal power plant equipment based on distributed intelligent sensing is provided. The method includes: acquiring raw distributed sensor data streams and grid dispatch instructions; the raw distributed sensor data streams include spatial acoustic temperature measurement data from the furnace combustion zone, wireless pressure transmitter data from steam and water pipelines, vibration spectrum data from rotating equipment, and actuator feedback from the DCS system; vectorizing the raw distributed sensor data streams to obtain a standardized sensor vector set; constructing a spatiotemporal map and aligning the standardized sensor vector set with asynchronous data to obtain a spatiotemporally aligned state matrix; extracting coupling features and reconstructing the state from the spatiotemporally aligned state matrix to obtain a system latent coupling feature tensor; performing dual-timescale state sequence prediction on the system latent coupling feature tensor to obtain a predicted state trajectory sequence; based on the grid dispatch instructions, performing multi-objective rolling optimization on the predicted state trajectory sequence to obtain an optimal control increment sequence; and extracting the first element from the optimal control increment sequence to obtain the final actuator instruction.

[0005] According to another aspect of this application, a collaborative control system for thermal power plant equipment based on distributed intelligent sensing is provided, comprising: a sensor scheduling data acquisition module for acquiring raw distributed sensor data streams and power grid dispatching instructions, wherein the raw distributed sensor data streams include spatial acoustic temperature measurement data of the furnace combustion zone, wireless pressure transmitter data of the steam and water pipelines, vibration spectrum data of rotating equipment, and actuator feedback from the DCS system; a data stream vectorization module for vectorizing the raw distributed sensor data streams to obtain a standardized sensor vector set; and a spatiotemporal aligned state matrix generation module for generating a standardized sensor vector set. The system employs a spatiotemporal graph construction and asynchronous data alignment module to obtain a spatiotemporally aligned state matrix. A state reconstruction module extracts coupled features from the spatiotemporally aligned state matrix and reconstructs the state to obtain the system's potential coupled feature tensor. A predicted state trajectory generation module performs dual-timescale state sequence prediction on the system's potential coupled feature tensor to obtain a predicted state trajectory sequence. An optimal control increment analysis module performs multi-objective rolling optimization on the predicted state trajectory sequence based on power grid dispatch instructions to obtain the optimal control increment sequence. Finally, a final actuator instruction generation module extracts the first element from the optimal control increment sequence to obtain the final actuator instruction.

[0006] Compared with existing technologies, this application provides a collaborative control system and method for thermal power plant equipment based on distributed intelligent sensing, aiming to solve the problems of control lag and overshoot caused by the spatiotemporal misalignment of distributed heterogeneous data and multi-scale coupling of the boiler and turbine systems. Specifically, it first acquires raw distributed data streams such as furnace acoustic temperature measurement, pipeline wireless pressure, and vibration of rotating equipment, along with grid commands. Asynchronous data alignment is performed through vectorization processing and spatiotemporal graph construction, reconstructing discrete and non-uniformly timed sensor data into a spatiotemporally aligned state matrix containing physical correlations, thereby eliminating state perception bias caused by data transmission delays and asynchronous sampling. Considering the large inertia and strong coupling characteristics of the boiler and turbine, the potential coupling feature tensor of the system is further extracted, and a dual-time-scale mechanism is used to predict the fast and slow dynamic trajectories of the system, accurately mapping the dynamic differences between boiler energy accumulation and turbine rapid response. Finally, multi-objective rolling optimization is performed based on the predicted trajectory, calculating the optimal control increment while ensuring system stability, achieving rapid response and precise coordination to grid dispatch commands. Attached Figure Description

[0007] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0008] Figure 1 This is a flowchart of a collaborative control method for thermal power plant equipment based on distributed intelligent sensing, according to an embodiment of this application.

[0009] Figure 2 This is a flowchart of step S2 in the collaborative control method for thermal power plant equipment based on distributed intelligent sensing according to an embodiment of this application.

[0010] Figure 3 This is a schematic diagram of the data flow in step S3 of the collaborative control method for thermal power plant equipment based on distributed intelligent sensing according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of a distributed intelligent sensing-based collaborative control system for thermal power plant equipment according to an embodiment of this application. Detailed Implementation

[0012] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0013] To address the technical challenges mentioned above, this application proposes a collaborative control method for thermal power plant equipment based on distributed intelligent sensing. Figure 1 This is a flowchart of a collaborative control method for thermal power plant equipment based on distributed intelligent sensing, according to an embodiment of this application. Figure 1As shown, the collaborative control method for thermal power plant equipment based on distributed intelligent sensing according to an embodiment of this application includes: S1, acquiring raw distributed sensor data streams and grid dispatch instructions, wherein the raw distributed sensor data streams include spatial acoustic temperature measurement data of the furnace combustion zone, wireless pressure transmitter data of the steam and water pipelines, vibration spectrum data of rotating equipment, and actuator feedback from the DCS system; S2, vectorizing the raw distributed sensor data streams to obtain a standardized sensor vector set; S3, constructing a spatiotemporal map and aligning asynchronous data on the standardized sensor vector set to obtain a spatiotemporally aligned state matrix; S4, extracting coupling features and reconstructing the state of the spatiotemporally aligned state matrix to obtain a system potential coupling feature tensor; S5, performing dual-time-scale state sequence prediction on the system potential coupling feature tensor to obtain a predicted state trajectory sequence; S6, based on the grid dispatch instructions, performing multi-objective rolling optimization on the predicted state trajectory sequence to obtain an optimal control increment sequence; S7, extracting the first element from the optimal control increment sequence to obtain the final actuator instruction.

[0014] In step S1, the raw distributed sensor data stream and grid dispatch instructions are acquired. The raw distributed sensor data stream includes spatial acoustic temperature measurement data of the furnace combustion zone, wireless pressure transmitter data of the steam-water pipeline, vibration spectrum data of rotating equipment, and actuator feedback from the DCS system. It should be understood that, to meet the urgent needs of modern power systems for the flexible adjustment capabilities of thermal power plants, the unit operation mode has shifted from single baseload generation to deep peak shaving and rapid load adjustment. In this context, relying solely on traditional single-point thermal instruments is insufficient to comprehensively capture the fine-grained state of the complex and large-scale boiler-turbine coupled system in dynamic processes, especially the real-time perception of the spatial distribution characteristics of the furnace combustion field and the health status of key auxiliary equipment under varying operating conditions. The energy release on the combustion side exhibits a significant lag, while the load response on the turbine side is relatively rapid. This difference in physical characteristics means that when facing automatic generation control (AGC) commands from the grid, a lack of real-time monitoring of the full-dimensional status of the equipment can easily lead to control overshoot or insufficient response. Therefore, introducing distributed sensing methods, including acoustic temperature measurement and wireless sensing, and effectively acquiring these raw data streams containing rich spatiotemporal information, becomes the primary prerequisite for constructing a high-precision collaborative control model, eliminating the impact of spatiotemporal misalignment, and ultimately achieving precise optimization of the furnace system across multiple time scales.

[0015] In an exemplary operation, step S1 follows this process: This step serves as the input interface for the entire collaborative control method, responsible for establishing data connections with the physical world's perception layer and power grid dispatching layer. During implementation, a multi-protocol parallel data acquisition channel is first constructed to receive heterogeneous data synchronously or asynchronously from different sources. Regarding acquiring power grid dispatching instructions, a secure connection is established with the power grid dispatching center via a dedicated remote communication device, using IEC 60870-5-104 or other power dispatching standard protocols. The received power grid dispatching instructions are primarily in the form of Automatic Generation Control (AGC) instructions. These instructions include not only the target load value (e.g., set to 600MW), but also parameters such as the load regulation rate (e.g., set to 12MW / min) and the dead zone range. These instruction data are parsed in real time and stored as a control target sequence with absolute timestamps, serving as a benchmark reference trajectory for subsequent rolling optimization solutions.

[0016] In acquiring raw distributed sensor data streams, the process encompasses four types of information streams with significantly different physical characteristics and data structures. The first is spatial acoustic temperature measurement data from the furnace combustion zone. Using several acoustic transceivers deployed around the boiler furnace walls—for example, 8 pairs or more transceiver probes in the core combustion area at an elevation of 20 to 40 meters—the difference in sound wave propagation speed in different temperature media is utilized, combined with acoustic tomography algorithms, to inversely calculate the two-dimensional temperature field distribution of the furnace cross-section. This data is acquired as a real-time updated two-dimensional matrix. For example, if the furnace cross-section is gridded into an m×n region, the collected data is an m×n temperature value matrix, where each element represents the real-time temperature of the corresponding grid point, such as 1350℃. This two-dimensional matrix can clearly characterize the deflection of the combustion flame center and the distribution of high-temperature areas. The data update frequency is usually limited by the sound wave travel time and computation time, possibly remaining on the order of seconds. The second is wireless pressure transmitter data from the steam and water pipelines. These sensors are mainly installed at critical locations such as the coal mill outlet, economizer inlet, and superheater and reheater pipes. They communicate with the aggregation gateway using low-power wireless transmission protocols such as WirelessHART or ISA100.11a. Because the wireless sensors operate in event-triggered or periodic sleep modes to extend battery life, their data stream is a series of non-uniformly sampled time-series data. Each acquired data packet contains a unique sensor ID, a physical pressure value (in MPa), and a corresponding data generation timestamp. The sampling times of sensors at different locations in the data stream are not aligned, and there is random delay caused by data packet transmission, which constitutes the source of the asynchronous characteristics that need to be processed later. The third type is the vibration spectrum data of rotating equipment. For critical auxiliary equipment such as feedwater pumps, induced draft fans, and forced draft fans, data is collected through wireless vibration monitoring nodes installed on their bearing housings. Unlike conventional vibration passband values, the vibration spectrum data obtained here is preprocessed by edge computing nodes. This data is represented as a vector in the frequency domain, with each sampling time containing a set of frequency-amplitude pairs, reflecting the vibration energy distribution of the equipment at different frequencies. For example, the data stream contains vibration velocity or acceleration amplitudes at the power frequency of 50Hz, the second harmonic of 100Hz, and their higher harmonics. The purpose of acquiring vibration spectrum data is to assess the mechanical health constraints of equipment during rapid load changes in real time, preventing safe shutdowns caused by forced vibration. Finally, there is feedback from the actuators of the DCS system. Communication with the power plant's existing distributed control system (DCS) is established using OPCUA or Modbus TCP protocols. This data stream is periodically and synchronously sampled, mainly including the real-time status of actuators such as turbine valve opening (0%-100%), boiler main control commands, coal feeder speed feedback, and damper positions.Although DCS data has high temporal determinism, it only represents the end-stage execution state of the control loop and needs to be combined with the aforementioned distributed sensing data to form a complete system state description.

[0017] In the specific implementation architecture, the above four types of data streams are aggregated into a time-series database or in-memory data buffer capable of processing multimodal data. At this point, the data is not aligned and retains its original sampling timestamps and data format. For example, at a certain moment, the buffer might simultaneously contain a furnace temperature matrix that has just arrived (timestamp t1), coal mill outlet pressure data that arrived three milliseconds ago (timestamp t2), and DCS valve feedback from one second ago (timestamp t3). This mixed data set, containing spatial two-dimensional information of the furnace combustion field, discrete-time series of pipeline pressure, frequency domain characteristics of rotating equipment, and actuator feedback, together constitutes the original distributed sensor data stream. This data stream, along with the parsed power grid dispatch instructions, is directly transmitted to provide the most basic and unmodified physical observation input for subsequently constructing a standardized sensor vector set.

[0018] In step S2, the original distributed sensor data stream is vectorized to obtain a standardized sensor vector set. Correspondingly, since the original distributed sensor data stream originates from sensing nodes with dispersed physical locations, heterogeneous communication protocols, and vastly different sampling mechanisms, the unstructured nature of this data and the inherent random noise severely hinder subsequent high-order feature extraction. Especially in high-frequency electromagnetic environments, wireless transmission is highly susceptible to interference, resulting in dead pixels or packet loss, and acoustic temperature measurement data may also experience instantaneous distortion due to combustion noise. Directly using this raw data, mixed with glitches and errors, to construct a spatiotemporal map will inevitably introduce computational bias, leading to misjudgments and oscillations in control commands. Therefore, before performing complex spatiotemporal alignment and feature reconstruction, it is necessary to standardize and vectorize this heterogeneous data to filter out noise interference unrelated to the actual system state, while simultaneously mapping information from different physical dimensions and data structures to a homogeneous mathematical space.

[0019] Figure 2 This is a flowchart of step S2 in the collaborative control method for thermal power plant equipment based on distributed intelligent sensing according to an embodiment of this application. Figure 2 As shown, in an exemplary operation, step S2, which vectorizes the original distributed sensor data stream to obtain a standardized sensor vector set, includes: S21, performing multidimensional data cleaning and outlier removal on the original distributed sensor data stream to obtain a valid measurement dataset; S22, performing heterogeneous data standardization and encoding on the valid measurement dataset to obtain a normalized feature set; and S23, performing spatial topological embedding and vectorization on the normalized feature set to obtain a standardized sensor vector set.

[0020] In the above exemplary operation, the specific process of step S2 is as follows: For step S21: First, an adaptive filtering strategy is applied to each type of data in the data stream. For data presented as a continuous time series, specifically including wireless pressure transmitter data of steam and water pipelines (e.g., coal mill outlet pressure) and single-point temperature data of each grid point parsed from the acoustic temperature measurement matrix of the furnace space, outliers are identified and processed using a statistical criterion based on a sliding window. Specifically, a time sliding window of length L is set, which is determined comprehensively based on the sampling frequency of the sensor and the dynamic response characteristics of the measured physical quantity, such as L=10, and the local mean of the data within the window is calculated in real time. and standard deviation For the current sampling point value... The outlier identification and correction method is applied. If... If the absolute value of the deviation from the mean is within three standard deviations, it is considered a valid measurement and retained. Otherwise, if it exceeds this range, the data point is considered to be transient noise caused by electromagnetic interference or sensor malfunction, and the arithmetic mean of the valid data at adjacent time points is immediately used. Perform interpolation replacement. and The data values ​​are taken from the previous valid time point and the next predicted time point (or the most recent valid time point), respectively, to smoothly remove hardware glitches. For the vibration spectrum data of rotating equipment included in the original data stream, since it is inherently a discrete distribution in the frequency domain, the cleaning strategy focuses on removing background noise to highlight the effective vibration characteristics. In this process, a lower limit for the energy threshold needs to be set. This threshold is determined by analyzing the base noise level of the equipment in a stopped or unloaded state, for example, by taking 1.2 times the average energy of the base noise (e.g., set to 0.05 mm / s). During processing, the amplitude of each frequency component in the vibration spectrum data is iterated. ,like If the component is set to 0 or a minimum value, then invalid background noise caused by environmental electromagnetic radiation or weak mechanical disturbances is filtered out. After the above cleaning processes targeting different data characteristics, the original distributed sensor data stream, which was originally mixed with noise and outliers, is transformed into a clean, reliable, and effective measurement dataset. This dataset retains the true physical fluctuation characteristics (such as pressure rise caused by load changes) while eliminating transmission errors and outlier interference.

[0021] For step S22: First, the valid measurement dataset is categorized according to its physical properties and then sent to the numerical normalization module and the non-numerical encoding module respectively. For continuous variables with clear physical dimensions in the dataset, such as spatial acoustic temperature measurement data (temperature, unit °C) in the furnace combustion zone, wireless pressure transmitter data (pressure, unit MPa) in the steam-water pipeline, and vibration spectrum data (vibration velocity, unit mm / s) of rotating equipment, the Min-Max normalization method is used to eliminate dimensional differences. During this process, the historical extreme value range for each type of physical quantity needs to be pre-defined. This range is obtained based on extreme operating condition records from the unit's historical operating database over the past year. For example, for furnace temperature, the lowest historical temperature is used. =800℃ and the highest historical temperature =1500℃. For a specific effective temperature measurement at that moment... For example, for 1200℃, its normalized value can be calculated using the following formula. : In the formula These are actual physical measurement values. and These represent the corresponding historical minimum and maximum values. Through this calculation, the temperature value of 1200℃ is transformed into a dimensionless 0.5714. All physical quantities are linearly mapped to the closed interval [0,1], ensuring that data of different magnitudes have equal weight in subsequent calculations. For discontinuous numerical information contained in the effective measurement data set from the actuator feedback of the DCS system, especially the logical states of dampers such as local / remote control, fault / normal, etc., discrete modes cannot be directly algebraically calculated. Therefore, One-Hot encoding is required to convert them into numerical vectors. For example, for the operating status of the coal feeder, if there are three discrete states: running, stopped, and fault, they are encoded as three-dimensional vectors of [1,0,0], [0,1,0], and [0,0,1] respectively. For valve opening data that is originally in percentage form, such as 85%, it is directly divided by 100 to convert it into the value 0.85 in the interval [0,1], without the need for One-Hot encoding. After processing, all physical quantity data that has undergone numerical normalization and the state vector that has undergone encoding are reintegrated into a unified data structure. At this point, the original data set containing heterogeneous descriptions such as 1200℃, 15MPa, and operating conditions has been completely transformed into a normalized feature set composed of dimensionless floating-point numbers and binary vectors. This feature set eliminates the barriers of physical units and accurately expresses the current state of the equipment in pure numerical form.

[0022] Regarding step S23: First, a pre-built, plant-wide digital 3D mapping library needs to be invoked. This mapping library is a static database constructed during the control system initialization phase based on the power plant's digital delivery data, BIM model (Building Information Modeling), and KKS coding (Power Plant Identification System). It establishes a unified Cartesian coordinate system with a fixed point at the zero-meter level of the unit (such as the geometric center projection point of the boiler room) as the origin. The specific architecture of the library is a hash table structure, where the key is the hardware unique identifier (Sensor ID) of all sensors in the plant, and the value is a tuple of the precise 3D coordinates of that sensor in physical space. The unit is uniformly set to meters. In the specific processing flow, the processor traverses each data element in the normalized feature set, first parsing out the hardware identifier corresponding to the data. For example, for the normalized temperature value of 0.5714 at a certain measuring point in the furnace calculated in the previous step, its Sensor ID is identified as S_Furnace_L2_04. Then, using this ID as an index, a real-time query is performed in the plant-wide digital 3D mapping library to retrieve the physical spatial coordinates of the acoustic temperature measuring probe's installation location. If the query result is (12.5, -4.0, 32.0), it indicates that the measuring point is located at 12.5 meters on the X-axis, -4.0 meters on the Y-axis, and 32.0 meters at an elevation. At the same time, the absolute timestamp t when the measurement value was generated is extracted from the original data stream, such as t=1699850025.5s. Subsequently, a vector concatenation operation is performed to normalize the value. timestamp t and spatial coordinates By combining them in a predetermined order, a five-dimensional feature vector containing spatiotemporal information is constructed. . This represents the standardized feature vector generated by the i-th sensor at time t. The physical meaning and numerical range of each component in the vector are clearly defined, forming a unified mathematical description. The above extraction, mapping, and splicing process is repeated for all key measuring points throughout the plant (covering furnaces, pipelines, rotating equipment, etc.). Finally, all generated five-dimensional feature vectors are aggregated to output a standardized sensor vector set, which is a multi-dimensional information matrix integrating state strength, temporal progression, and spatial topology.

[0023] In step S3, a spatiotemporal graph is constructed from the standardized sensor vector set and aligned with asynchronous data to obtain a spatiotemporally aligned state matrix. It is understandable that the standardized sensor vector set, mathematically, still represents an isolated set of discrete points, failing to directly reflect the intricate physical connections and thermodynamic coupling relationships within a thermal power plant. As a strongly coupled fluid network, the interaction between devices in a thermal power plant follows a strict working fluid flow direction (e.g., pulverized coal combustion generates hot flue gas, which drives steam flow after heat exchange). This causal relationship is accompanied by specific physical delays and is not entirely dependent on the Euclidean distance in geometric space. If these data are merely considered as independent feature vectors input into the model, the control algorithm will struggle to understand the inherent topological correlation between the increase in furnace temperature and the delayed increase in main steam pressure, leading to deviations in the dynamic prediction of the system. Therefore, this application constructs a graph structure that can explicitly describe the physical connections and data correlations between nodes. Through the graph, discrete sensors are mapped into a topological network with interactive relationships, thus providing structured constraints consistent with physical mechanisms for subsequent alignment and state deduction of asynchronous data streams.

[0024] Figure 3 This is a schematic diagram of the data flow in step S3 of the collaborative control method for thermal power plant equipment based on distributed intelligent sensing according to an embodiment of this application. Figure 3 As shown, in an exemplary operation, step S3, which involves constructing a spatiotemporal graph and aligning asynchronous data to the standardized sensor vector set to obtain a spatiotemporally aligned state matrix, includes: S31, constructing a weighted sensor topology graph based on the standardized sensor vector set; S32, performing time-axis gridding and synchronous projection on the standardized sensor vector set to obtain a sparse temporal network matrix; and S33, performing graph-based hybrid imputation and state alignment on the weighted sensor topology graph and the sparse temporal network matrix to obtain a spatiotemporally aligned state matrix.

[0025] In the above exemplary operation, the specific process of step S3 is as follows: For step S31: First, the nodes are defined and initialized. The processing logic traverses each five-dimensional feature vector in the standardized sensor vector set and extracts the hardware identifier (Sensor ID) contained therein. For each independent element in the vector set, it is instantiated as a graph model. one of the vertices For example, when a vector with the ID Temp_Reheater_Inlet_02 is identified, a corresponding node is generated in the graph. This node not only indexes the sensor's identity but also binds its corresponding spatial coordinates. Attributes. At this point, if there are N key sensor nodes in the entire plant (e.g., N=500), a node set V containing N vertices is constructed. Then, the edge connections between nodes are established, employing a dual-determination strategy combining physical mechanisms and spatial geometry. On one hand, pre-digitized process flow diagram data is introduced as a logical constraint. Based on the PID (Pipeline and Instrumentation Diagram) drawings and KKS coding rules provided by the power plant design institute, it is transformed into a computer-recognizable directed graph network model using graph theory algorithms, establishing the flow direction and upstream / downstream connections of the physical media between all equipment in the plant. This data details the pipeline connections of the power plant's steam-water system, flue gas system, and pulverizing system. For any two sensor nodes i and j, their upstream / downstream relationships in the process flow are queried. If a direct media flow path exists, such as the coal mill outlet air pressure sensor pointing to the burner inlet air pressure sensor, a directed edge is established between the two nodes. This characterizes the direct causal coupling caused by matter transport. On the other hand, it calculates the Euclidean distance between any two nodes in physical space. Using the 3D coordinate data carried by the nodes, the calculation formula is as follows: Set a spatial proximity threshold. Based on the physical density of the equipment and empirical values ​​for the effective attenuation distance of heat radiation or vibration transmission, such as setting it to 5.0 meters, it is used to capture local coupling effects of non-pipe connections, such as radiative heat transfer or equipment vibration transmission. If the calculated... Even if two devices are geographically distant in their piping (e.g., adjacent superheater tube banks), a spatial connection is established, thus creating an edge link. After determining the topology, the weight of each edge is calculated to quantify the connection strength. For each pair of connected nodes i and j, the weight calculation integrates information from both spatial proximity and statistical correlation. First, based on physical distance... Calculate spatial proximity weights The Gaussian kernel function is used, with greater weight given to closer data. Secondly, to capture the consistency of dynamic data changes, sliding time window data stored in the historical database, such as sampling sequences from the past 300 seconds, is used to calculate the Pearson correlation coefficient between the two sensor data sequences. As a statistical relevance weight This coefficient reflects the degree of synchronization between two physical quantities during historical fluctuations; its absolute value is taken. To measure the strength of the correlation. The final edge weights. Calculated using a weighted fusion method. Let represent the combined connection weight between node i and node j, with values ​​ranging from [0,1]. By performing the above calculation, a sparse adjacency weight matrix A is constructed, where the non-zero elements are _i_ and _j_. This matrix, together with the node set V and the edge set E, constitutes the weighted sensor topology graph.

[0026] For step S32: First, a globally unified time grid needs to be set. The grid is divided according to the real-time requirements of the control system, and discrete sampling time sequences are set. ,in Defined as a standard control cycle. Considering the execution frequency of the Coordinated Control System (CCS) in thermal power plants, The default setting is 1 second. This setting provides a common time scale for all asynchronous data, discretizing the continuous time axis into a series of equally spaced time nodes. For any specified time step k, i.e., the target alignment time... The processor performs a truncation operation on a standardized set of sensor vectors. To cover valid data from around that time, a small time tolerance window is set. For example, 0.5 seconds. Retrieve and extract all timestamps falling within the interval. Data records within the system. For each specific sensor node. For example, for the main steam pressure transmitter with the ID Press_MainSteam_01, check if there is sampled data within that time window. If one or more sampling points are found within the window, such as in Seconds and Each second contains one data point, indicating that the sensor reported data within that control cycle. At this point, in order to obtain the closest possible target time... The true state estimation does not employ simple averaging or nearest neighbor selection, but instead uses Gaussian kernel weighted interpolation to project these neighboring data onto the accurate time. Above. The specific projection calculation formula is as follows: In this formula, That is, the calculated synchronous projection value of the i-th sensor in the k-th control cycle; This represents the normalized value of the m-th original sampling point that falls within the window; This is the actual timestamp corresponding to the sampling point; It is the target moment for alignment. Let be the Gaussian kernel function, defined as ,in The bandwidth parameter, such as 0.2 seconds, determines how quickly the weights decay over time. This means that the distance from the target time... The closer the sampling point, the greater its weight contribution to the final projected value, thus accurately reconstructing the state at the target time from local data. Conversely, for a given sensor node... For example, at a low-frequency acoustic temperature measurement point in the furnace, within a set time window... No data records were found, which could be due to the sensor's sampling period being longer than the control period, or due to communication packet loss. In this case, the system does not perform forced interpolation or retain the value from the previous moment, but instead explicitly updates the value for that moment. The corresponding location is marked as a missing value. This preserves the fact that the data is missing and avoids misleading information caused by using outdated data. After traversing and processing all N sensor nodes in the entire plant, for each control moment... A column vector of dimension N×1 is generated, containing some calculated projection values ​​and some placeholders for missing values. Arranging the column vectors from multiple consecutive control moments in chronological order forms a sparse temporal network matrix. This matrix intuitively reflects the observed system state under a unified time grid; each row represents the time series of a single sensor, and each column represents a plant-wide snapshot at a synchronization moment. Although the matrix contains a large number of blanks (sparseness), its structure is completely normalized.

[0027] For step S33: First, the processor receives and loads the two key inputs mentioned above: a weighted sensor topology graph defining node connections and strengths (including an adjacency weight matrix A), and the sparse temporal network matrix to be filled. The processor uses a point-by-point scanning mechanism to traverse each element in the sparse matrix, that is, to check the numerical state of each sensor node i at each discrete time step k. For any selected target data point First, determine whether it is a valid observation. If If the value is non-missing, it means that the actual sampled data at that moment was successfully acquired in the synchronous projection in step S32. In this case, the original projected value is directly retained. This ensures the highest level of confidence in the actual observation data and prevents any human intervention. However, if any abnormalities are detected... Marked as a missing value, indicating that the node is missing. Data gaps can occur due to low-frequency sampling or packet loss during transmission. In such cases, a graph-based inference mechanism is immediately initiated. First, in the weighted sensor topology graph, taking the node i with the currently missing data as the center, all its first-order neighbor nodes are retrieved. These neighboring nodes are sensors identified as having a direct and strong correlation in step S31 based on the upstream / downstream relationship of the medium flow or the Euclidean distance in physical space. For example, if a node with missing data is generated by the coal mill #B outlet air pressure sensor, its neighbor set may include nodes such as the primary air fan #B outlet pressure (upstream), the #2 corner burner inlet pressure (downstream), and the coal mill #A outlet air pressure at the same elevation (spatial adjacency). Subsequently, the valid observations of these neighboring nodes at the same time k are extracted. And combined with the pre-calculated edge weights in the topology graph The missing values ​​are calculated using a graph-space weighted smoothing algorithm. This process leverages prior knowledge of the similarity of physical nearest neighbors, estimating the state of the unknown central node based on the states of surrounding known nodes, effectively compensating for insufficient temporal sampling by utilizing spatial correlation. After traversing and filling all missing points in the sparse temporal network matrix, all NaN vacancies are filled with values ​​inferred from the physical mechanism, thus transforming it into a completely dense network. Numerical matrix, This is the length of the time window. Finally, to maintain the integrity of the information, the dense numerical sequence calculated for each node is paired with its corresponding normalized spatial coordinate features, i.e., those from step S23. The dimensions are concatenated using timestamp features. The final output is a spatiotemporally aligned state matrix. This matrix is ​​a dimensional... The three-dimensional tensor (or considered as a constant at every moment) A two-dimensional matrix sequence, where N is the total number of sensors and F is the feature dimension containing normalized state values ​​and spatial coordinates.

[0028] In step S4, coupled feature extraction and state reconstruction are performed on the spatiotemporally aligned state matrix to obtain the system's potential coupled feature tensor. It should be understood that although the sensor data has achieved synchronization and continuity in time and value after spatiotemporal alignment, the current feature representation remains at a shallow physical observation level. In the boiler-turbine coupled system of a thermal power plant, the mutual influence between equipment is often nonlinear and dynamically changing. For example, the impact of changes in feedwater flow rate on the main steam temperature depends not only on the flow rate but also on various potential factors such as combustion intensity and the degree of ash accumulation on the heated surface. This complex nonlinear coupling relationship is difficult to describe directly by linearly superimposing the original temperature or pressure values. To enable the control system to truly understand and quantify these deep-seated dynamic interaction mechanisms, it is necessary to project the data from the intuitive physical space to a higher-dimensional hidden feature space. This process aims to endow each node with richer expressive capabilities through dimensionality enhancement, enabling it to not only carry its own measurement information but also potentially encode its associated environmental context information.

[0029] In an exemplary operation, step S4, which involves extracting coupled features and reconstructing the state of the spatiotemporally aligned state matrix to obtain the system's latent coupled feature tensor, includes: S41, performing high-dimensional feature linear mapping on the spatiotemporally aligned state matrix to obtain a high-dimensional projected feature set; S42, calculating attention coefficients on the high-dimensional projected feature set to obtain an attention coefficient matrix; and S43, based on the attention coefficient matrix, performing neighborhood feature aggregation and state reconstruction on the high-dimensional projected feature set to obtain the system's latent coupled feature tensor.

[0030] In the above exemplary operation, the specific process of step S4 is as follows: For step S41: First, extract the data in the matrix row by row according to the time. For the current time k, obtain the original feature vector corresponding to each sensor node i. This vector It is a column vector of dimension F, i.e., F=5, containing normalized values. timestamp and three-dimensional spatial coordinates This directly carries the physical observation state and spatiotemporal attributes of the node at that moment. To uncover the complex nonlinear patterns hidden behind this low-dimensional physical data, a learnable shared linear transformation matrix is ​​initialized. The matrix These are core weight parameters in neural network models (such as Graph Attention Networks, GAT layers). Their dimensions are set to F'×F, where F' is the mapped high-dimensional hidden layer feature dimension, much larger than the input dimension F. For example, setting F'=64 allows the data to unfold into a manifold structure in a higher-dimensional space. This matrix... This matrix is ​​obtained through iterative optimization using a backpropagation algorithm based on a large amount of historical running sample data during the offline training phase of the entire neural network model. During the online inference phase, this matrix contains fixed, pre-trained weight parameters used to capture the linear combination relationships between input features. The specific mapping operation is as follows: For each node i among the N sensor nodes in the entire plant, its original feature vector... With shared weight matrix Performing matrix multiplication, i.e., executing a linear transformation Here, This operation applies to all nodes, embodying the idea of ​​parameter sharing. This means that regardless of a node's location, the basic rules for feature extraction (such as a combination of temperature and pressure) are universal. Through this operation, the original 5-dimensional physical feature vector (F) is mapped to a 64-dimensional feature vector (F'). This high-dimensional vector no longer has intuitive physical units (such as °C or MPa), but instead becomes an abstract semantic code that the neural network can understand. Repeating the above operation for all N nodes generates a set containing N high-dimensional vectors. This set is the output high-dimensional projected feature set.

[0031] Regarding step S42: This process utilizes a graph attention mechanism to quantify the dynamic relationships between nodes. The input is the high-dimensional projected feature set generated in step S41, which contains high-dimensional representations of all nodes in the plant. For each central node i in the graph, such as the node representing the main steam pressure, the set of all its first-order neighbor nodes is first located based on the weighted sensor topology graph. Iterate through each neighbor node j in the set, such as the node representing fuel quantity, to prepare for calculating the attention weight of node j to node i. The calculation process first performs a feature concatenation operation. The high-dimensional feature vector of the center node i is then... The high-dimensional feature vector of neighbor node j By concatenating the first and last elements in sequence, a concatenated vector of dimension 2F' is constructed. This concatenated vector contains the current state information of both nodes, forming a complete context for determining their relationship. Subsequently, a pre-trained, learnable weight vector is introduced. This vector It is a column vector of dimension 2F', a core parameter of the graph attention network model layer, and is also optimized by minimizing the prediction error during offline training. It acts as an evaluator, measuring the importance of the concatenated features to the system's dynamics. The concatenated vector is then compared with the weight vector. Perform the dot product operation, that is, calculate... This maps complex high-dimensional feature relationships to a scalar value, which initially represents the correlation energy between two nodes. To introduce nonlinearity and enhance the model's ability to fit complex conditions, the LeakyReLU activation function is applied to this scalar value. LeakyReLU retains a small negative slope (e.g., 0.2) when the input is less than 0, avoiding the neuron death problem. The formula is expressed as follows: Finally, to ensure the calculated weights are probabilistically comparable, the Softmax function is used to normalize the nonlinear scores of all neighboring nodes within the neighborhood of node i, thus obtaining the final attention coefficients. . Strictly restricted to the interval (0,1), and satisfying In a physical sense, This intuitively represents the proportion of the contribution of the state change of neighbor node j to the state evolution of central node i under a specific working condition at the current time k. For example, if the calculation result is... =0.7, indicating that the influence of fuel quantity changes on main steam pressure has a weight as high as 70%. Performing the above calculations on all nodes throughout the plant, the final generated... The set constitutes the attention coefficient matrix, which dynamically depicts the intensity distribution of the mutual influence between equipment throughout the plant.

[0032] For step S43: The processing is performed with each node i in the graph as the center node. For a specific node i, all its related neighbor nodes are found based on the attention coefficient matrix. and their corresponding weighting coefficients Next, a weighted summation operation is performed to project the high-dimensional features of each neighboring node. Multiply by its corresponding attention coefficient The physical significance of this step lies in selectively absorbing surrounding information based on the strength of the actual influence of each neighboring node on the central node i at the current moment. For example, if the turbine-side parameters, such as the valve opening, exhibit strong coupling with the boiler-side parameters, such as the drum pressure, then... If the pressure is relatively large, the turbine-side state information will dominate the aggregation process, thereby correcting or enhancing the state representation of the steam drum pressure, making it not only reflect the current pressure reading but also implicitly include the pull effect from the turbine. After completing the linear weighted aggregation, in order to capture the prevalent nonlinear dynamic characteristics in the system, the aggregation result is input into a nonlinear activation function. In this embodiment, ELU or ReLU is selected as the activation function. The ELU function can handle negative inputs and has better robustness, helping to prevent gradient vanishing. After this step, the feature vector of node i is no longer just an encoding of its own sensor readings, but has evolved into a comprehensive descriptor that integrates its own and all associated device states. To further enhance the robustness of the model, a multi-head attention mechanism is adopted, that is, the above attention calculation and aggregation process is executed in parallel K times, such as K=8, with each head focusing on capturing different types of coupling relationships (such as one head focusing on thermal parameter coupling and the other focusing on vibration mechanical coupling), and finally the K output vectors are concatenated or averaged. In this embodiment, a single head is used as an example. Finally, the above aggregation and update operation is performed in parallel on all N nodes in the entire plant, and the updated feature vectors of all nodes are then used. Reorganize according to node numbering order. For each time step k, form a state matrix of dimension N×F' (the dimension may increase if multiple heads are considered). Stack the state matrices within the continuous time window along the time dimension to output the system's latent coupling feature tensor. Mathematically, this tensor is a high-dimensional array; physically, it is a holographic image of the current operating state of the thermal power plant's boiler and turbine system. It transcends simple physical connections and profoundly reveals the deep coupling mechanism containing dynamic weights.

[0033] In step S5, a dual-timescale state sequence prediction is performed on the system's potential coupling characteristic tensor to obtain the predicted state trajectory sequence. Correspondingly, the coordinated control system of a thermal power plant's boiler and turbine is a typical multivariable, strongly coupled dynamic object with significant multi-timescale characteristics. The boiler side involves a lengthy process of pulverized coal preparation, transportation, combustion, and heat transfer to the working fluid, exhibiting slow dynamic characteristics with large inertia and large delays, where energy accumulation often occurs in the order of seconds. Conversely, the turbine side controls steam flow to drive rotor rotation, with rapid valve action, exhibiting fast response and low inertia, with energy release occurring only in the order of seconds. If a single timescale model is used to predict the overall system, it often leads to a situation where one aspect is overlooked: setting a large prediction step size or smoothing parameter to adapt to the slow dynamics of the boiler will miss the transient response details of the turbine; conversely, focusing on the fast dynamics of the turbine makes it difficult to accurately capture the long-term energy accumulation trend of the boiler. Furthermore, the pure time delay in pulverized coal transportation and combustion processes means that the current control action will not immediately cause a state change. Without targeted time-series matching, the model will learn incorrect input-output causal relationships. Therefore, this application decouples the system state on the time scale and constructs prediction inputs for different dynamic characteristics.

[0034] In an exemplary operation, step S5, performing dual-timescale state sequence prediction on the system's latent coupling feature tensor to obtain a predicted state trajectory sequence, includes: S51, performing timescale decomposition and lag processing on the system's latent coupling feature tensor to obtain a fast-channel input set and a slow-channel input set; S52, performing dual-channel RNN parallel inference on the fast-channel input set and the slow-channel input set to obtain a fast-timescale predicted trajectory and a slow-timescale predicted trajectory; S53, performing multi-resolution trajectory fusion on the fast-timescale predicted trajectory and the slow-timescale predicted trajectory to obtain a predicted state trajectory sequence.

[0035] In the above operations, the specific process of step S5 is as follows: For step S51: First, perform feature decoupling. The features in the tensor are filtered and distributed according to a pre-built equipment attribute mapping table. This mapping table is a static configuration file, stored in key-value pairs, where the key is the sensor ID or feature index, and the value is the dynamic category label to which the feature belongs, such as Fast or Slow. The construction of this table is based on the prior knowledge and mechanism analysis of power plant experts: for example, turbine valve opening feedback, generator active power, turbine speed, etc., are marked as fast variable features; coal mill feed rate, furnace outlet flue gas temperature, steam drum water level, main steam pressure (although affected by the turbine quickly, energy accumulation is slow, classified as slow process dominant or mixed), etc., are marked as slow variable features. The processor traverses the input system potential coupling feature tensor, extracting rows or slices corresponding to fast variable labels to form a fast variable feature set. Similarly, extract the slices corresponding to the slow variable labels to form the slow variable feature set. For intermediate variables with mixed characteristics, they can be placed into both sets simultaneously. Next, control input is injected. Control command sequences are extracted from the power plant's historical database, such as the PI database or the eDNA real-time database. This database stores all historical operation records issued by the DCS system. For the fast channel, the real-time turbine control command at the current time k is extracted. The most typical example is the high-pressure control valve (GV) opening command. Because the turbine response is extremely fast, it can be assumed that the current valve action will immediately affect the system state at the next moment; therefore, the current moment's k is directly used. Features of fast variables The process involves splicing. For slow-moving processes, the core issue is pure time delay compensation. Boiler-side control actions, such as adjusting the coal feeder speed, require a series of processes including coal grinding, primary air conveying, and heat release from combustion before being reflected in the steam parameters. This process involves a fixed pure time delay. For example, 60 seconds. It's unreasonable to directly use the coal feeding command at time k to predict the state at the next time k+1. Therefore, this lag parameter needs to be identified or preset using an algorithm. And retrieve back from the historical database Historical boiler control commands at any time For example, a coal feed instruction from 60 seconds ago. This time-shifted old instruction is used as a valid input for the current state evolution. Finally, two independent input sets are constructed. The fast-channel input set consists of the current fast variable features. With current turbine control commands Combining, in the form of The goal is to predict fast responses on a timescale of seconds. The slow channel input set consists of current slow variable features (or full features). Historical boiler instructions aligned with causality Combining, in the form of The aim is to predict energy accumulation trends on a minute-scale timescale.

[0036] Regarding step S52: First, the internal dual-core parallel computing engine is activated, allocating independent computing resources (such as GPU stream processors) to the fast and slow channels to ensure that the two inference paths do not block each other and proceed synchronously. For the fast channel inference, the fast channel input set prepared in step S51 is received. The dataset is fed into a specially designed and pre-trained LSTM-Fast network model. LSTM-Fast is a variant of a Long Short-Term Memory (LSTM) recurrent neural network, optimized for capturing high-frequency dynamics. Specifically, the network consists of an input layer, two to three hidden layers, and an output layer. The input layer dimension is adapted to the fast-channel feature dimension, such as 32-dimensionality. The hidden layers employ forget gates, input gates, and output gates, with a relatively small number of neurons, such as 64, to ensure extremely fast inference speed. Its core lies in the time step. Set to a very small value, for example =0.1s, meaning that each step forward in the model spans only 0.1 seconds of physical time. During offline training, the network uses historical high-frequency data with sampling frequencies exceeding 10Hz, such as rotational speed fluctuations and power oscillations recorded by the DEH system, for supervised learning. The backpropagation (BPTT) algorithm minimizes short-term prediction errors (such as the mean squared error within the next 5 seconds), thus determining the network's internal weight matrix and bias vector. During online inference, the LSTM-Fast network uses the input at the current time k to recursively predict the future. The state of the step. For example, if we need to predict the turbine dynamics in the next 5 seconds, and =0.1s, then =50 steps. In each recursion, the network returns the hidden state from the previous time step. and current input Combined, calculate the current hidden state. and output and will This serves as one of the inputs for the next time step. Ultimately, the network outputs a high-temporal-resolution sequence, namely, a fast-timescale predicted trajectory. This trajectory contains finely detailed variation curves of key turbine-side state parameters (such as generator active power, turbine speed, and high-pressure cylinder exhaust pressure) at 0.1-second intervals over a future period, clearly reflecting transient processes such as rapid power reversal caused by a primary frequency regulation action. Meanwhile, in terms of slow-channel simulation, the receiving slow-channel input set... This set is then fed into a separately trained LSTM-Slow network. While the LSTM-Slow network architecture is also based on LSTM units, its hyperparameter settings are drastically different. To accommodate the large inertia of the boiler, it has a larger number of hidden layer neurons, such as 128 or 256, to provide greater memory capacity for storing energy accumulation information over long periods. More importantly, its time step... Set it to a large value, for example =1.0s or longer. During the training phase, the network is trained using historical operating data with a low sampling frequency (e.g., 1Hz) but a long time span (e.g., several hours), focusing on learning the slow response curves of parameters such as main steam pressure and drum water level after changes in fuel quantity. During inference, the LSTM-Slow network recursively predicts the future with a step size of 1 second. The state of the step. If you need to predict the boiler trend for the next 5 minutes (300 seconds), then =300 steps. Although the number of steps is relatively high, it can cover a much more distant future due to the large physical time span. The sequence output by this network is the predicted trajectory for the slow time scale. This trajectory depicts the smooth evolution trend of key boiler-side state parameters (such as main steam pressure, furnace outlet flue gas temperature, and intermediate point temperature) at 1-second intervals over a relatively long period, ignoring high-frequency noise and focusing on the macroscopic trend of energy balance. Through a dual-channel parallel extrapolation mechanism, both the transient details of the steam turbine and the macroscopic trend of the boiler are obtained simultaneously within one calculation cycle.

[0037] For step S53: First, trajectory alignment based on the fast time axis is performed. Since subsequent control optimization is based on finer time granularity, the fast time axis is selected as the unified benchmark after fusion. For the slow time scale predicted trajectory... The data points it contains are located in Equal to integer seconds. To fill in its... For values ​​at intermediate time points, a linear interpolation algorithm is used. Compared to zero-order hold, linear interpolation better reflects the continuous changing trend of physical parameters (such as main steam pressure) between two sampling points. Specifically, for any two adjacent slow trajectory data points... and For a fast time point t' in between, calculate the interpolation. After this processing, the originally sparse slow trajectories are upsampled into dense sequences with the same time resolution as the fast trajectories, i.e., 0.1s intervals. However, the numerical changes still retain the original low-frequency smooth characteristics without introducing high-frequency noise. After unifying the time dimension, the data concatenation along the channel dimension is then performed. A new prediction matrix container is created, with the number of rows equal to the total number of time steps corresponding to the prediction time domain (e.g., ...). The number of columns represents the total number of state variables in the entire system (i.e., the number of fast variables + the number of slow variables). The original fast-timescale predicted trajectory... The data is filled into the first few columns of the matrix (corresponding to turbine-side variables), and the slow-time-scale predicted trajectory is then interpolated and aligned. The data is filled into the last few columns of the matrix (corresponding to boiler-side variables). For example, the first row of the concatenated matrix contains the predicted generator power (fast variable) and the predicted main steam pressure (interpolated slow variable) at time t=0.1s. Finally, this unified matrix is ​​encapsulated as a predicted state trajectory sequence. Mathematically, this sequence is represented as a... The two-dimensional matrix physically describes the expected evolution of the state of all critical equipment in the system over a future period (from time k+1 to k+P) at a given two-scale time resolution.

[0038] In particular, when processing multi-timescale trajectory fusion, simply using traditional zero-order hold (ZOH) or linear interpolation methods to align slow-timescale trajectories to fast-timescale nodes presents a serious mechanistic consistency defect in the physical scenario of turbine-boiler coordinated control. The energy system of a thermal power plant exhibits a significant energy potential-work causal chain characteristic. Slow variables such as main steam pressure serve as energy potential sources, and their rate of change (derivative) directly determines the response potential of fast variables such as turbine power. Linear interpolation artificially creates abrupt inflection points in the first derivative at time nodes, while zero-order hold results in the derivative being zero most of the time and infinite at the transition points. This data form, which disrupts differential continuity, contradicts the smooth energy flow characteristics of the physical system. When input to a model predictive controller highly sensitive to the rate of change, the optimizer misinterprets it as an energy abrupt change in the physical system, inducing high-frequency jitter in the actuators that violates actual physical requirements, severely accelerating equipment wear. Furthermore, the original mechanism neglects the dynamic damping effect across scales. That is, under unsteady conditions with drastic perturbations on fast timescales, the enormous inertia of the physical system should suppress the propagation of slow variables. Therefore, to address these shortcomings, this scheme employs a Hermite cooperative projection method based on dynamic damping constraints, aiming to generate state inputs that combine numerical accuracy with physical interpretability.

[0039] Based on this, in an exemplary preferred operation, step S53, which involves multi-resolution trajectory fusion of the fast-timescale predicted trajectory and the slow-timescale predicted trajectory to obtain a predicted state trajectory sequence, includes: performing cross-scale dynamic damping and gradient correction on the fast-timescale predicted trajectory and the slow-timescale predicted trajectory to obtain a corrected physical gradient sequence. This step aims to eliminate the deviation between the geometric derivative and the physical derivative. The implementation process no longer blindly trusts the geometric trend of the slow trajectory, but instead uses the volatility of the fast trajectory to impose physical constraints. Within the corresponding interval of each slow time step, the system first calculates the first-order difference variance of the fast-timescale predicted trajectory, using it as a dynamic perturbation operator characterizing the current system instability. Subsequently, the initial geometric gradient of the slow-timescale predicted trajectory is obtained using the central difference method, and the geometric gradient is nonlinearly scaled and corrected in conjunction with a preset physical damping coefficient.

[0040] For example, at a certain time k, the predicted main steam pressure rises from 15.0 MPa to 15.2 MPa, with a geometric slope of 0.1 MPa / s. However, at this time, the turbine-side power experiences severe oscillations due to grid interference (large variance of fast variables), causing the exponential term to drop to 0.2. The physical gradient calculated at this point... It will be corrected to 0.1 × 0.2 = 0.02 MPa / s. This treatment introduces a soft threshold gating mechanism through an exponential decay term. When the fast variable on the turbine side oscillates violently, it forcibly flattens the derivative trend of the slow variable, preventing the MPC controller from making aggressive adjustments using the spurious rate of change of the slow variable under unstable operating conditions.

[0041] Based on the corrected physical gradient sequence, the predicted trajectory on a slow timescale is interpolated to obtain a physically smooth slow trajectory. This step generates a dense trajectory that meets the energy continuity requirement at the microscopic timescale. Energy-consistent cubic Hermitian interpolation is performed, using the physical correction slope calculated in the previous step. And the slow trajectory node values, to construct a high-resolution smooth projection operator. For any interpolation time t, between and First, a normalized time variable is introduced. Then substitute , This represents the value of the generated physically smooth slow trajectory at time t; and These are the endpoint values ​​of the original slow trajectory; and This process corrects the gradient at the endpoints using physical methods. By constructing cubic polynomial basis functions, it ensures that the interpolated curve is not only numerically continuous, but also that its first derivative (i.e., the rate of energy change) is smooth and continuous at the connection points and strictly controlled by physical damping constraints. This eliminates numerical spikes during optimizer solving and achieves alignment between the data and physical layers.

[0042] Heterogeneous state manifold concatenation is performed on physically smooth slow trajectories and fast-timescale predicted trajectories to obtain a sequence of predicted state trajectories. The ultimate goal is to fuse the heterogeneous data into a standardized object containing confidence information. The heterogeneous state manifold concatenation and encapsulation steps are performed to generate physically smooth slow trajectories. Trajectory prediction with fast time scale The sequence is concatenated along the feature dimension, and the local confidence weights of the concatenated sequence are calculated simultaneously. In an exemplary operation, heterogeneous state manifold concatenation is performed on the physically smooth slow trajectory and the fast timescale predicted trajectory to obtain the predicted state trajectory sequence, including: heterogeneous state manifold concatenation of the physically smooth slow trajectory and the fast timescale predicted trajectory using the following formula:

[0043] in, Represents the predicted state trajectory sequence. To predict trajectories on a fast timescale This represents a vector concatenation operation. For physically smooth slow trajectories, This represents the confidence weight vector composed of the diagonal elements of the constructed diagonal matrix. Let be the absolute value of the vector of gradients of all slow variables at the current time. To prevent small constants with a denominator of zero, such as Taking main steam pressure as a slow variable and turbine power as a fast variable as an example, at time t, the predicted power value with a resolution of 0.1s, the interpolated pressure value, and the weights calculated based on the pressure change rate are combined into a vector. The introduction of this weight term ensures that when the derivative of the slow variable is severely damped, i.e. As the value approaches zero, the weight increases significantly, which mathematically tells the subsequent optimizer that the slow variable here is in a passively damped state and should prioritize ensuring system stability. This allows for flexibility in adjustment when the system is stable, while automatically applying inertial damping when there is severe disturbance.

[0044] In step S6, based on the grid dispatch instructions, a multi-objective rolling optimization solution is performed on the predicted state trajectory sequence to obtain the optimal control increment sequence. It can be understood that after obtaining the accurate predicted future state trajectory of the system and the real-time dispatch instructions from the grid, the core task of the controller becomes decision-making: how to find a balance between "fast and slow"? It must respond quickly to the grid's AGC (Automatic Generation Control) power commands, ensure that the fluctuations in the boiler's main steam pressure remain within a safe range, and avoid shortening the lifespan of regulating valves due to drastic movements. Traditional PID control only provides feedback based on the current deviation, making it difficult to consider these mutually constraining multiple objectives, and it is prone to overshoot when dealing with large time delays. Simply pursuing response speed may lead to excessive thermal stress, while excessive conservatism will fail to meet grid performance requirements. Therefore, this application introduces an optimization mechanism with a forward-looking perspective, capable of comprehensively balancing tracking accuracy, system safety, and operational economy in the prediction time domain.

[0045] In an exemplary operation, the specific process of step S6 is as follows: First, construct a comprehensive objective function to quantify the control effect. This function aims to transform different control requirements into a computable algebraic form. Objective function The penalty is mainly composed of three weighted components. The first component is the load tracking deviation, which is the difference between the predicted actual generator power and the grid's AGC dispatch command. This reflects the accuracy of the power plant's response to grid demand and has a relatively large weight. The second component is the main steam pressure fluctuation penalty, which is the degree to which the predicted main steam pressure deviates from the set value (such as the rated pressure). This is related to the safe and stable operation of the unit, preventing pressure exceeding limits from causing a shutdown. The third component is the control variable change rate penalty, used to constrain the magnitude of fuel quantity increments, feedwater increments, and turbine valve opening increments, preventing excessively drastic control actions that could lead to actuator wear or system oscillations.

[0046] and These are the state error weight matrix and the control increment weight matrix, respectively, both positive definite diagonal matrices. For example, if load response is given more emphasis, The weighting element corresponding to the power deviation is set to 100; if valve protection is required, The weight element corresponding to the valve increment is set to 10. These two matrices are obtained through offline simulation or genetic algorithm tuning during the controller debugging phase.

[0047] In addition to the objective function, strict constraints must be set to ensure physical feasibility. These mainly include: amplitude constraints on control variables (e.g., the opening of the turbine control valve is strictly limited to between 0% and 100%, and the feedwater pump speed cannot exceed the rated speed), rate constraints on control increments (e.g., the coal feeder's coal feeding rate cannot exceed 10 t / h / min), and safety constraints on key state variables (e.g., the upper and lower limits of the steam drum water level, and the temperature rise rate of the screen superheater cannot exceed 1.5℃ / min to prevent thermal stress cracking). These constraints constitute the feasible region of the optimization problem.

[0048] To solve the multi-objective optimization problem with complex physical constraints established above, the built-in numerical optimization solver strategy is dynamically selected based on the current operating conditions. First, the processor performs local linearization on the dual-scale nonlinear prediction model generated in step S5 near the current operating point, calculating the sensitivity matrix (Jacobi matrix) of the predicted output relative to the control input. This transforms the originally complex nonlinear rolling optimization problem into a standard quadratic programming (QP) mathematical model, i.e., constructing a model of the form... The optimization equation, where It is a Hessian matrix. The gradient vector is used to transform the physical constraints into a system of linear inequalities. If the system operates near stable conditions and only requires fine-tuning, the transformed standard model is directly fed into the QP solver, using the effective set method or interior point method for rapid iterative solution. These deterministic algorithms leverage the convexity of the objective function and the analytical properties of the gradient, converging to an exact solution within milliseconds. This greatly satisfies the stringent requirements of real-time control in thermal power plants regarding computational latency, ensuring the timeliness of control commands. However, if the unit is in a large-scale variable operating condition phase, such as deep peak shaving, rapid start-up and shutdown, or accident handling, the nonlinear characteristics of the boiler and turbine are significantly enhanced. Simple linear approximation may introduce non-negligible errors. Therefore, the system automatically switches to a nonlinear programming solution path, selecting the Particle Swarm Optimization (PSO) or Sequential Quadratic Programming (SQP) algorithm. These algorithms directly handle non-convex objective functions and complex nonlinear constraints by performing multi-particle parallel search or multi-step quadratic approximation within the feasible region. Although the computation time increases slightly, they ensure the global optimality of the solution and the robustness of the control strategy under large disturbance conditions.

[0049] In each control cycle, such as time k, the solver, under the premise of strictly satisfying all physical constraints such as valve opening limit and temperature rise rate protection, traverses the search space and finally calculates and outputs a set of optimal control increment sequences. , so that the objective function Reaching the minimum value. This sequence contains a series of optimal operational actions for the next M time points, such as: increasing the fuel quantity by 2% at the current time, increasing it by 1.5% at the next time, and simultaneously quickly closing the regulating valve to store pressure... This sequence is the optimal control increment sequence, which not only reflects the rapid following of grid commands, but also implicitly compensates for system inertia and delays in advance, achieving a global optimal balance of multiple objectives.

[0050] In step S7, the first element is extracted from the optimal control increment sequence to obtain the final actuator command. That is, although the solver has already spent computational resources generating an optimal control increment sequence covering the entire future prediction time domain (e.g., dozens of future control cycles) in the previous multi-objective rolling optimization step, theoretically this sequence can guarantee optimal control performance over a considerable period. However, the actual operating environment of thermal power plants is full of randomness and uncertainty, such as instantaneous fluctuations in coal calorific value, random disturbances in coal mill output, and sudden changes in grid frequency. These real-time disturbances, not fully captured by the prediction model, cause the system state to gradually deviate from the predicted trajectory over time. If the entire pre-calculated long sequence is executed in an open loop, the control commands at subsequent times will become invalid due to outdated state information, resulting in accumulated errors. Therefore, it is necessary to follow the basic principle of rolling time domain in model predictive control, that is, only adopting the most direct and certain decision calculated based on the latest state, and discarding subsequent predictive control quantities. This process aims to transform long-cycle open-loop optimization into real-time closed-loop control, ensuring that every action is an optimal correction based on the system's latest perception feedback, thereby achieving adaptive correction to complex dynamic environments.

[0051] In an exemplary operation, step S7 proceeds as follows: The process directly receives the optimal control increment sequence containing action plans for the next M time points, output from step S6. The processor first performs a sequence truncation operation, extracting only the first vector element from the sequence. And immediately discard subsequent data. Extracted It is a vector containing the adjustment amplitude of each channel within the current control cycle, for example... .

[0052] In order to convert this relative increment into an absolute position command that the actuator can recognize, the controller needs to acquire the current reference command. The reference command originates from the real-time feedback channel of the DCS system, i.e., the command value issued in the previous control cycle and held by the actuator, or directly reads the current actual position feedback of the actuator, such as the valve displacement measured by the LVDT sensor. For example, by reading the DCS memory address, the actual comprehensive opening command of the turbine high-pressure regulating valve at the current moment is obtained as 85.5%, and the average speed command of the coal feeder is 600 rpm.

[0053] Subsequently, an overlay calculation is performed to extract the optimal increment. Compared with current benchmark instructions Perform algebraic summation. For example, the fuel quantity increment calculated by the algorithm. Corresponding to an increase of 5 rpm in the coal feeder speed, the steam turbine valve opening increment is... The value is -0.1%. The calculation process is as follows: new feeder speed command = 600 + 5 = 605 rpm; new valve opening command = 85.5% + (-0.1%) = 85.4%. Before output, the calculation result needs to be verified by the amplitude and rate limiting module to ensure that the command value is within the safe range allowed by the hardware (such as 0-100%).

[0054] After verification, the values ​​are sent to the signal conversion module to generate the final actuator commands. For analog control devices, the values ​​are mapped to standard 4-20mA current signals or 0-10V voltage signals via a D / A converter; for digital intelligent devices, they are encapsulated as data packets from a fieldbus (such as Profibus-DP) or industrial Ethernet (such as Profinet). Finally, this set of physical signals is sent in parallel to the underlying actuator hardware via hardwiring or a communication network. Specifically, the current signal directly drives the servo valve coil of the turbine electro-hydraulic control system to fine-tune the hydraulic actuator stroke, while the bus message instructs the coal mill frequency converter to adjust the output frequency to change the coal feed rate. The execution of this series of physical actions marks the end of the current control cycle and triggers a change in the system state, awaiting the sensing and optimization of the next cycle.

[0055] In summary, the distributed intelligent sensing-based collaborative control method for thermal power plant equipment, based on the embodiments of this application, is explained, aiming to solve the problems of control lag and overshoot caused by the spatiotemporal misalignment of distributed heterogeneous data and the multi-scale coupling of the boiler and turbine systems. Specifically, firstly, raw distributed data streams such as furnace acoustic temperature measurement, pipeline wireless pressure, and vibration of rotating equipment, along with grid commands, are acquired. Asynchronous data alignment is performed through vectorization processing and spatiotemporal graph construction, reconstructing discrete and non-uniformly timed sensor data into a spatiotemporally aligned state matrix containing physical correlations, thereby eliminating state perception deviations caused by data transmission delays and asynchronous sampling. Considering the large inertia and strong coupling characteristics of the boiler and turbine, the potential coupling feature tensor of the system is further extracted, and a dual-time-scale mechanism is used to predict the fast and slow dynamic trajectories of the system, accurately mapping the dynamic differences between boiler energy accumulation and turbine rapid response. Finally, multi-objective rolling optimization is performed based on the predicted trajectory, calculating the optimal control increment while ensuring system stability, achieving rapid response and precise coordination to grid dispatch commands.

[0056] Figure 4 This is a block diagram of a distributed intelligent sensing-based collaborative control system for thermal power plant equipment according to an embodiment of this application. Figure 4As shown, the distributed intelligent sensing-based collaborative control system 100 for thermal power plant equipment according to an embodiment of this application includes: a sensor scheduling data acquisition module 110, used to acquire raw distributed sensor data streams and power grid scheduling instructions, wherein the raw distributed sensor data streams include spatial acoustic temperature measurement data of the furnace combustion zone, wireless pressure transmitter data of the steam and water pipelines, vibration spectrum data of rotating equipment, and actuator feedback from the DCS system; a data stream vectorization module 120, used to vectorize the raw distributed sensor data streams to obtain a standardized sensor vector set; and a spatiotemporal alignment state matrix generation module 130, used to perform spatiotemporal alignment of the standardized sensor vector set. The graph construction and asynchronous data alignment are used to obtain a spatiotemporally aligned state matrix; the state reconstruction module 140 is used to extract coupled features and reconstruct the state from the spatiotemporally aligned state matrix to obtain the system's potential coupled feature tensor; the predicted state trajectory generation module 150 is used to predict the system's potential coupled feature tensor using a dual-timescale state sequence to obtain a predicted state trajectory sequence; the optimal control increment analysis module 160 is used to perform multi-objective rolling optimization on the predicted state trajectory sequence based on power grid dispatch instructions to obtain the optimal control increment sequence; and the final actuator instruction generation module 170 is used to extract the first element from the optimal control increment sequence to obtain the final actuator instruction.

[0057] Here, those skilled in the art will understand that the specific operations of each step in the above-described distributed intelligent sensing-based collaborative control system for thermal power plant equipment have been referenced above. Figures 1 to 3 The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing has been described in detail, and therefore, its repeated description will be omitted.

[0058] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of this disclosure, and this disclosure is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this disclosure, and these modifications and improvements are also considered to be within the scope of protection of this disclosure.

Claims

1. A method for collaborative control of thermal power plant equipment based on distributed intelligent sensing, characterized in that, include: The raw distributed sensor data stream and power grid dispatch instructions are acquired. The raw distributed sensor data stream includes spatial acoustic temperature measurement data of the furnace combustion zone, wireless pressure transmitter data of the steam and water pipeline, vibration spectrum data of rotating equipment, and actuator feedback from the DCS system. The raw distributed sensor data stream is vectorized to obtain a standardized sensor vector set; Spatiotemporal mapping of standardized sensor vector sets is constructed and asynchronous data is aligned to obtain a spatiotemporally aligned state matrix. Coupled feature extraction and state reconstruction are performed on the spatiotemporally aligned state matrix to obtain the system's latent coupled feature tensor; A dual-time-scale state sequence prediction is performed on the system's latent coupling feature tensor to obtain a predicted state trajectory sequence; Based on power grid dispatch instructions, multi-objective rolling optimization is performed on the predicted state trajectory sequence to obtain the optimal control increment sequence. The first element is extracted from the optimal control increment sequence to obtain the final executor instruction.

2. The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing according to claim 1, characterized in that, The raw distributed sensor data stream is vectorized to obtain a standardized sensor vector set, including: Multidimensional data cleaning and outlier removal are performed on the raw distributed sensor data stream to obtain an effective measurement dataset. Heterogeneous data standardization and encoding are performed on the effective measurement dataset to obtain a normalized feature set; Spatial topological embedding and vectorization are performed on the normalized feature set to obtain a standardized sensor vector set.

3. The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing according to claim 1, characterized in that, Spatiotemporal mapping of the standardized sensor vector set is constructed and asynchronous data is aligned to obtain a spatiotemporally aligned state matrix, including: Construct a weighted sensor topology graph based on a standardized sensor vector set; A sparse temporal network matrix is ​​obtained by time-axis gridding and synchronous projection of the standardized sensor vector set. A graph-based hybrid imputation and state alignment are performed on the weighted sensor topology graph and the sparse temporal network matrix to obtain a spatiotemporally aligned state matrix.

4. The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing according to claim 1, characterized in that, Coupled feature extraction and state reconstruction are performed on the spatiotemporally aligned state matrix to obtain the system's latent coupled feature tensor, including: A high-dimensional feature linear mapping is performed on the spatiotemporal aligned state matrix to obtain a high-dimensional projected feature set; Attention coefficients are calculated on the high-dimensional projected feature set to obtain the attention coefficient matrix; Based on the attention coefficient matrix, neighborhood feature aggregation and state reconstruction are performed on the high-dimensional projected feature set to obtain the system's latent coupled feature tensor.

5. The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing according to claim 1, characterized in that, The system's latent coupling feature tensor is subjected to dual-time-scale state sequence prediction to obtain a predicted state trajectory sequence, including: The system's potential coupling feature tensor is decomposed by time scale and lag-processed to obtain the fast channel input set and the slow channel input set; A dual-channel RNN is used to perform parallel extrapolation on the fast-channel input set and the slow-channel input set to obtain the fast-time-scale predicted trajectory and the slow-time-scale predicted trajectory. Multi-resolution trajectory fusion is performed on the fast-timescale predicted trajectory and the slow-timescale predicted trajectory to obtain the predicted state trajectory sequence.

6. The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing according to claim 5, characterized in that, Multi-resolution trajectory fusion is performed on the fast-timescale predicted trajectory and the slow-timescale predicted trajectory to obtain a predicted state trajectory sequence, including: Cross-scale dynamic damping and gradient correction are applied to the fast-timescale and slow-timescale predicted trajectories to obtain the corrected physical gradient sequence. Based on the corrected physical gradient sequence, the predicted trajectory at a slow time scale is interpolated to obtain a physically smooth slow trajectory; Heterogeneous state manifolds are spliced ​​together from physically smooth slow trajectories and fast timescale predicted trajectories to obtain a sequence of predicted state trajectories.

7. The method for coordinated control of thermal power plant equipment based on distributed intelligent sensing according to claim 6, characterized in that, Heterogeneous state manifold concatenation is performed on physically smooth slow trajectories and fast timescale predicted trajectories to obtain a sequence of predicted state trajectories. This includes concatenating the physically smooth slow trajectories and fast timescale predicted trajectories using the following formula: in, Represents the predicted state trajectory sequence. To predict trajectories on a fast timescale This represents a vector concatenation operation. For physically smooth slow trajectories, This represents the confidence weight vector composed of the diagonal elements of the constructed diagonal matrix. Let be the absolute value of the vector of gradients of all slow variables at the current time. To prevent tiny constants with a denominator of zero.

8. A collaborative control system for thermal power plant equipment based on distributed intelligent sensing, characterized in that, include: The sensor scheduling data acquisition module is used to acquire raw distributed sensor data streams and power grid scheduling instructions. The raw distributed sensor data streams include spatial acoustic temperature measurement data of the furnace combustion zone, wireless pressure transmitter data of the steam and water pipeline, vibration spectrum data of rotating equipment, and feedback from the actuators of the DCS system. The data stream vectorization module is used to vectorize the raw distributed sensor data stream to obtain a standardized sensor vector set; The spatiotemporal alignment state matrix generation module is used to construct a spatiotemporal map of a standardized sensor vector set and align asynchronous data to obtain a spatiotemporal alignment state matrix. The state reconstruction module is used to perform coupled feature extraction and state reconstruction on the spatiotemporally aligned state matrix to obtain the system's latent coupled feature tensor. The predicted state trajectory generation module is used to perform dual-time-scale state sequence prediction on the system's potential coupling feature tensor to obtain the predicted state trajectory sequence. The optimal control increment analysis module is used to perform multi-objective rolling optimization on the predicted state trajectory sequence based on power grid dispatch instructions to obtain the optimal control increment sequence; The final executor instruction generation module is used to extract the first element from the optimal control increment sequence to obtain the final executor instruction.

Images