An intelligent decision-making method and system for coal-fired units based on curve shape topology

Abstract

This invention discloses an intelligent decision-making method and system for coal-fired power units based on curve morphology topology. The method first utilizes a large-scale artificial intelligence model to perform time-series transformation on DCS data and then performs spatiotemporal alignment and fusion with the DCS data to construct a panoramic data view. Second, a high-precision dynamic operation model covering all operating conditions is established. Then, curve morphology topology is innovatively introduced; by extracting curve morphology primitives, constructing topological relationships, and analyzing cross-parameter domain responses, a graph neural network is used to decouple and enhance dynamic features. Next, a physically interpretable feature distillation module and a decision knowledge graph are constructed to achieve transparent intelligent decision-making. Finally, a high-fidelity digital twin parallel system driven by AI is used to perform closed-loop verification and optimization of the decisions. This invention overcomes the representational limitations of traditional mechanistic models through feature decoupling using curve morphology topology and improves the interpretability of decisions through an interpretable knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation and industrial intelligent control technology, and in particular to an intelligent decision-making method and system for coal-fired power units based on curve morphology topology. Background Technology

[0002] Against the backdrop of the accelerated global energy structure transition towards a low-carbon model, traditional coal-fired power plants, as the mainstay of energy supply, face multiple challenges in efficiency improvement, pollution control, and intelligent upgrading. Currently, my country's coal-fired power industry urgently needs to reconstruct its core competitiveness through digital and intelligent transformation to adapt to the development requirements of dual-carbon goals and Industry 4.0. Traditional coal-fired power plants rely on manual operation and monitoring, resulting in low energy efficiency, high carbon emissions, high operation and maintenance costs, and significant safety risks. With the rapid development of cutting-edge technologies such as artificial intelligence, digital twins, and the Internet of Things, building a new generation of autonomous decision-making intelligent control systems (ACS) centered on intelligent control, patrol, and monitoring, achieving a revolutionary shift from minimal human intervention to single-person operation and near-zero intervention, has become an industry consensus.

[0003] In existing technologies, the operation control of coal-fired power units mainly relies on distributed control systems (DCS) and mechanistic models. However, traditional mechanistic models have limited ability to represent data under complex operating conditions, making it difficult to reliably extract truly useful features for operational decisions from complex, variable, and cross-parameter domain operating data. This is mainly reflected in the following aspects: Severe temporal coupling: Coal-fired power units are highly nonlinear, multivariate coupled systems, with complex correlations between time-series data of various operating parameters (such as pressure, temperature, flow rate, concentration, etc.). Traditional feature extraction methods, such as principal component analysis (PCA) or simple statistical features, are difficult to effectively separate and express these deeply coupled features.

[0004] Cross-parameter domain relationships are difficult to express: Unit operation data includes not only time-series data from DCS, but also multi-source heterogeneous data from non-DCS systems such as video surveillance, equipment vibration, sound, and inspection records. Existing technologies lack an effective method to convert these data of different modalities and semantics into time-series data and to model them in a unified manner with DCS data, resulting in a large amount of data containing rich operating condition information not being fully utilized.

[0005] Lack of interpretability in decision-making results: Decisions made based on artificial intelligence models with significant black-box characteristics (such as deep neural networks) may improve accuracy, but their decision-making logic is not transparent, making it difficult for operators to understand and trust them. This is a fatal flaw in the field of power production, where safety requirements are extremely high.

[0006] Therefore, there is an urgent need for a unit intelligent decision-making method that can effectively integrate multi-source heterogeneous data, decouple complex time-series characteristics, provide interpretable decision-making basis, and have closed-loop verification capabilities, in order to overcome the shortcomings of existing technologies and promote the leap from automation to autonomy in coal-fired power plants. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent decision-making method for coal-fired power units based on curvilinear topology, addressing the challenges of integrating multi-source heterogeneous data, decoupling temporal features, and uninterpretable decision-making. Through temporal fusion of DCS data, decoupling of curvilinear topological features, and closed-loop verification using interpretable knowledge graphs and digital twins, intelligent control across all operating conditions and near-zero intervention are achieved.

[0008] To achieve the above objectives, the present invention proposes the following technical solution: an intelligent decision-making method for coal-fired power units based on curve morphology topology, comprising the following steps: Step S1: Perform time-series transformation and feature extraction on the DCS data generated during the operation of the coal-fired power unit, and align and seamlessly integrate the extracted time-series features with the DCS data to construct a panoramic data view covering the entire operating conditions of the unit; wherein, the DCS data includes at least one or more of video data, audio data, vibration data and text data; the panoramic data view contains time-series data of each operating parameter; Step S2: Based on the panoramic data view, establish a high-precision dynamic operation model covering all operating conditions, including unit start-up and shutdown, normal operation, load regulation, and abnormal operating conditions, to form a dynamic feature expression oriented towards operation cognition; Step S3: Introduce curve morphology topology, decompose the time series curve of each operating parameter in the panoramic data view into a finite number of morphological primitives, calculate the temporal topological relationship of the morphological primitives corresponding to the time series curves of different operating parameters to construct a curve morphology topology graph, and use causal analysis to construct a cross-parameter domain dynamic response correlation matrix; based on the curve morphology topology graph and the dynamic response correlation matrix, construct a graph neural network to perform feature decoupling and enhancement on the dynamic feature expression, separate and strengthen key dynamic features, and form a decoupled feature representation; Step S4: Based on the decoupled feature representation, construct a feature distillation module with physical interpretability, extract and distill operational cognitive features with clear physical meaning, and construct a decision knowledge graph based on the operational cognitive features to support intelligent decision-making for integrated control strategies of intelligent control, cruise and monitoring. Step S5: Input the control strategy generated by the intelligent decision into the AI-driven high-fidelity digital twin parallel system for closed-loop simulation verification, optimize the control strategy based on the verification results, and send the optimized control strategy to the actual unit for execution.

[0009] Furthermore, step S1 specifically includes: For video data, a video temporal behavior detection model is used to extract keyframe sequences and transform them into time series describing physical events; for audio data, Mel frequency cepstral coefficients combined with a temporal classification network are used to transform it into acoustic event sequences; for vibration data, short-time Fourier transform is used to convert it into a spectrogram, and then a temporal convolutional network is used to extract the frequency band energy change trend; for text data, a named entity recognition model is used to transform unstructured inspection records into structured event triples with timestamps. Based on a unified high-precision clock source, the time-seriesd DCS data is interpolated and aligned with the DCS data in the time dimension, and the spatial dimension association mapping is realized through a pre-established device spatial location encoding table. By utilizing a cross-modal attention mechanism, the aligned multi-source time-series data are fused at the feature level to generate the panoramic data view.

[0010] Furthermore, step S3 specifically includes: Using the continuous cohomology method in topological data analysis, key points on each time series curve are identified, and the time series curve is decomposed into a finite number of morphological primitives consisting of rising, falling, stationary, inflection points and extreme points. Calculate the temporal order relationship, inclusion relationship, intersection relationship and separation relationship of the morphological primitives corresponding to different time series curves, and construct the curve morphological topology graph with the morphological primitives of each time series curve as nodes and the temporal order relationship, inclusion relationship, intersection relationship and separation relationship as edges; Using Granger causality test or transitive entropy method, the response intensity and response delay between different operating parameters are calculated, and the cross-parameter domain dynamic response correlation matrix is ​​established; The dynamic feature representation is used as the initial feature of the nodes in the graph neural network. The curve shape topology graph and the dynamic response correlation matrix are used as the graph structure input. The message passing and aggregation of features are realized through multi-layer graph convolution operation, and the decoupled feature representation is output.

[0011] Furthermore, step S4 specifically includes: The feature distillation module employs an attention-based interpretable network to calculate a physical interpretability score for each feature dimension in the decoupled feature representation. The physical interpretability score is associated with a preset set of physical labels, and features with scores higher than a preset threshold are selected as the operational cognitive features. Using the aforementioned operational cognitive features as entity nodes, the association rule mining algorithm is used to extract causal relationships, temporal relationships, and collaborative relationships between entities from historical operational data as edges, and combined with expert experience rules, to construct the decision knowledge graph; When a decision event is triggered, path reasoning based on graph search is performed on the decision knowledge graph to generate an interpretable decision path containing decision steps, intermediate states, and expected consequences, as well as its corresponding control instruction sequence.

[0012] Furthermore, step S5 specifically includes: The high-fidelity digital twin parallel system synchronizes the operating status parameters and boundary conditions of the actual unit in real time through the OPCUA protocol. The control strategy is simulated in advance on the twin to demonstrate its control effect in a preset future time period. Compare the pre-simulation results with the preset safety, economic, and environmental constraints. If all constraints are met, the system is assessed as reliable and execution is initiated. Otherwise, a strategy optimization module based on counterfactual reasoning or reinforcement learning is triggered to adjust the control strategy parameters until all constraints are met or the preset maximum number of iterations is reached.

[0013] Furthermore, the method also includes: deeply integrating the decision knowledge graph with the intelligent control, cruise, and monitoring integrated control strategy, wherein: Intelligent control is used to directly generate setpoints or control parameters for the underlying control loop, enabling autonomous adjustment at the loop level. Intelligent cruise is used to automatically call a preset operation sequence state machine from the decision knowledge graph when the unit load changes or the operating condition switches, and to plan and execute a smooth sequence of operation steps. The intelligent monitoring panel is used to monitor the unit status in real time. When abnormal parameter trends are detected, it performs reverse reasoning from the decision knowledge graph to locate the root cause, predict the abnormal development trend, and realize automatic focusing of alarm screens and push processing suggestions.

[0014] This invention also proposes an intelligent decision-making system for coal-fired power units based on curve morphology topology, used to execute the method described in any of the above-mentioned embodiments, the system comprising: The multi-source heterogeneous data fusion module is used to realize the temporal transformation and feature extraction of DCS data, as well as the spatiotemporal alignment and seamless fusion with DCS data to generate a panoramic data view. The dynamic operation modeling module is used to establish a dynamic operation model covering all working conditions based on the panoramic data view and output dynamic feature expressions. The curve morphology topology feature decoupling module is used to introduce the correlation between curve morphology topology and cross-parameter domain response, perform feature decoupling and enhancement on the dynamic feature expression, and output the decoupled feature representation; An interpretable decision knowledge graph module is used to construct a feature distillation module and a decision knowledge graph based on the decoupled feature representation, thereby generating intelligent decisions with physical interpretability. The digital twin closed-loop verification module is used to build an AI-driven high-fidelity digital twin parallel system to perform closed-loop verification and optimization of the intelligent decision.

[0015] Furthermore, the curve morphology topology feature decoupling module specifically includes: The curve morphology primitive extraction unit is used to decompose a time series curve into a finite number of morphology primitives using the continuous homology method. The topology calculation unit is used to calculate the topological relationships of morphological primitives between different time series curves and construct a curve morphology topology graph. The cross-parameter causal analysis unit is used to analyze the response relationship between parameters using Granger causality tests and to construct a dynamic response correlation matrix. The graph neural network decoupling unit is used to perform graph convolution operations on dynamic feature representations using the curve shape topology graph and dynamic response correlation matrix as graph structures, thereby achieving feature decoupling and enhancement.

[0016] Furthermore, the multi-source heterogeneous data fusion module specifically includes: The video temporalization unit is used to extract keyframe sequences from the video using a video temporal behavior detection model and convert them into time series describing physical events. The audio temporalization unit is used to convert audio data into a sequence of acoustic events by combining Mel frequency cepstral coefficients with a temporal classification network. The vibration temporalization unit is used to convert vibration data into a spectrum using short-time Fourier transform, and then use a temporal convolutional network to extract the frequency band energy change trend. The text temporalization unit is used to convert unstructured inspection records into structured event triples with timestamps using a named entity recognition model. The spatiotemporal alignment unit is used to achieve time dimension interpolation alignment based on a unified clock source and to achieve spatial dimension association mapping through a device spatial location encoding table; The cross-modal fusion unit is used to perform feature-level fusion on aligned multi-source time-series data using a cross-modal attention mechanism to generate the panoramic data view.

[0017] Furthermore, the digital twin closed-loop verification module specifically includes: The twin synchronization unit is used to synchronize the operating status parameters and boundary conditions of the actual unit in real time via the OPCUA protocol; The strategy simulation unit is used to perform advance simulation of the control strategy on the twin, simulating its control effect in a preset future time period; The constraint assessment unit is used to compare the pre-simulation results with the preset safety constraint indicators, economic constraint indicators and environmental constraint indicators to determine whether the strategy is reliable. The strategy optimization unit is used to adjust the control strategy parameters based on counterfactual reasoning or reinforcement learning methods when the pre-exercise results do not meet the constraints, and to trigger the strategy pre-exercise unit to perform iterative optimization until all constraints are met or the preset maximum number of iterations is reached.

[0018] Beneficial effects: 1. This invention solves the problems of scattered data sources, semantic inconsistencies, and difficulty in unified modeling in traditional operation analysis by realizing the temporal transformation, spatiotemporal alignment, and seamless integration of DCS data. This facilitates the construction of a complete and accurate panoramic data view. By converting DCS data such as video, audio, vibration, and text into temporal data through a large model and deeply integrating it with DCS data, multi-dimensional and cross-modal accurate perception of the unit's operating status is achieved.

[0019] 2. This invention establishes a high-precision dynamic operation model covering all operating conditions and introduces curve morphology topology and cross-parameter domain response correlation for feature decoupling and enhancement. This enables the extraction of more effective dynamic features from complex, multivariate coupled time-series data, overcoming the limitations of traditional mechanistic models in representing complex operating conditions. Compared to existing technologies that rely on numerical sequence fitting methods such as LSTM and GRU, this invention innovatively elevates time-series analysis from the numerical domain to the geometric morphology and topological relation domains. By constructing curve morphology topology graphs and causal correlation matrices and utilizing graph neural networks for information aggregation, it fundamentally decouples complex coupled features that are difficult to handle by traditional methods, significantly improving the model's representation capability under complex operating conditions.

[0020] 3. This invention, by constructing a physically interpretable feature distillation module and a decision knowledge graph, can explicitly reveal the relationship between raw data, dynamic features, and operational decisions, thereby improving the interpretability, verifiability, and reusability of intelligent decision-making results. Compared to the black-box decision-making model of comparative documents, the decision results of this invention are no longer incomprehensible numerical values, but rather traceable and verifiable reasoning paths with clear causal chains. This greatly enhances the trust of operators in the intelligent system and provides core technical support for human-machine collaboration.

[0021] 4. This invention, by combining intelligent control, integrated control strategy of cruise and monitoring, and a high-fidelity digital twin parallel system, can improve the accuracy and reliability of unit decision-making under all operating conditions, providing interpretable decision support for single-person operation and near-zero intervention. This invention introduces a high-fidelity digital twin parallel system for closed-loop verification. By rehearsing and optimizing strategies in a secure virtual environment, it effectively avoids unreliable commands directly affecting real equipment, providing a solid safety guarantee for intelligent cruise and intelligent control. This is a key link in achieving the goals of near-zero intervention and single-person operation. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps of the temporal feature decoupling and enhancement based on curve morphology topology in this invention.

[0024] Figure 3 This is a system composition block diagram of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] This embodiment provides an intelligent decision-making method and system for coal-fired power units based on curve morphology topology, used to achieve integrated decision-making of autonomous intelligent control, cruise control, and monitoring of coal-fired power units under all operating conditions. Figure 1 As shown, the method includes the following steps S101 to S105.

[0027] Step S101: Multi-source heterogeneous data fusion and DCS data time-series processing This step aims to convert various types of DCS data (video, audio, vibration, text, etc.) generated during the operation of coal-fired power units into time-series data, and perform spatiotemporal alignment and feature-level fusion with the DCS data to construct a unified panoramic data view. Specifically, it includes the following sub-steps S1011 to S1013.

[0028] Sub-step S1011: DCS data time-series processing. This sub-step processes four different types of DCS data separately: S1011a: For video data, a video temporal behavior detection model based on the Transformer architecture (such as VideoMAE or TimeSformer) is used to extract keyframes at a sampling rate of 1-5 frames per second. The model identifies physical events such as flame flicker frequency, flame center shift, coal flow morphology changes, and belt misalignment, and outputs a sequence of feature vectors with a unified timestamp, with each feature vector having a dimension of 128-256.

[0029] S1011b: For audio data, the acoustic features (13-dimensional basic coefficients plus first-order and second-order differences, totaling 39 dimensions) are first extracted using Mel frequency cepstral coefficients (MFCC). Then, combined with a bidirectional gated cyclic unit (BiGRU) time-series classification network, the audio streams of areas such as the furnace, blower, and coal mill are transformed into time-series features such as combustion uniformity index (0-1 continuous value), bearing abnormal noise event label (0 / 1), and probability of blower surge symptoms.

[0030] S1011c: For vibration data, short-time Fourier transform is used to convert the time-domain vibration signal into a time-spectrum graph. Then, a ResNet-18 network pre-trained on ImageNet is used to extract the depth features (512 dimensions) of the spectrum graph. Finally, the dimensions are reduced to 32 dimensions through a temporal convolutional network (TCN) to form temporal features that characterize the health status of the equipment (such as 1st harmonic amplitude, 2nd harmonic / 1st harmonic ratio, high-frequency noise energy, etc.).

[0031] S1011d: For text data, a BERT-based Chinese pre-trained model combined with a BiLSTM-CRF structure is used to achieve Named Entity Recognition (NER). Unstructured text such as inspection records and operation logs are parsed into structured event triples in the form of (time, equipment, parameter, status, operation), such as (15:30, A coal mill, current, fluctuation, inspection), which are then inserted into a unified time grid as discrete event-type time series data.

[0032] Sub-step S1012: Spatiotemporal Alignment. Based on a unified high-precision clock source (NTP server, 1ms accuracy), all DCS data and the time-seriesd DCS data are subjected to segmented cubic Hermite interpolation (PCHIP) in the time dimension to ensure that all data have the same sampling period. Simultaneously, through a pre-established device spatial location encoding table, each data source is associated with its physical device ID and three-dimensional spatial coordinates (e.g., the coordinates corresponding to "#2 burner flame video" (x=12.5, y=8.3, z=5.0)), achieving spatial dimension mapping. The output is a multi-source aligned data stream with spatiotemporal labels.

[0033] Sub-step S1013: Seamless Fusion. A fusion network is constructed using a cross-modal attention mechanism. DCS data is used as the query matrix Q, and the temporal features of the DCS data are used as the key matrix K and value matrix V. The importance of each non-DCS information item under the current operating condition is dynamically calculated through attention weights, and feature-level weighted fusion is performed. For example, when judging combustion stability, if the furnace pressure signal is disturbed by soot blowing, the attention weights automatically bias towards the stability features of the flame video, generating a more robust combustion stability index after fusion. All fused features together constitute a panoramic data view, which is a fixed-length feature vector sequence with approximately 6000 dimensions and a sampling rate of 1Hz.

[0034] Step S102: Establish a high-precision dynamic operation model under all working conditions Based on the panoramic data view, a high-precision dynamic operation model covering all operating conditions, including unit start-up and shutdown, normal operation, load regulation, and abnormal operating conditions, is established to form a dynamic feature expression oriented towards operation cognition.

[0035] Specifically, a 4-layer Dilated Temporal Convolutional Network (DilatedTCN) is constructed, with each layer having a kernel size of 3 and dilation rates of 1, 2, 4, and 8, respectively. The receptive field covers the past 30 minutes (1800 time steps). The network input shape is (1800, 6000), and the output is the change trajectory of 20 core state parameters (main steam pressure, main steam temperature, reheat steam temperature, furnace outlet NOx concentration, coal consumption for power generation, etc.) within the next 5-15 minutes (300-900 time steps). The loss function is the weighted mean squared error (MSE), where physical constraint penalty terms (such as penalizing the prediction of a decrease in main steam pressure when the load increases) are added in the form of a regularization term. The training data covers one year of historical operating data of the unit, including steady-state, dynamic, and disturbance conditions (such as coal mill tripping, soot blowing, load shedding, etc.) within the 50%~100% load range. After training, the output of the penultimate layer of the TCN network (before the global average pooling layer) is extracted as a high-dimensional (2048-dimensional) dynamic feature representation vector. This vector is updated every second, summarizing the historical dynamics of the unit over the past 30 minutes and the predicted trend for the next 5-15 minutes.

[0036] Step S103: Temporal Feature Decoupling and Enhancement Based on Curve Morphology Topology like Figure 2 As shown, this step introduces curve morphology topology to decouple and enhance the dynamic feature representation, specifically including the following sub-steps S1031 to S1034.

[0037] Sub-step S1031: Curve morphology primitive extraction. For key parameters in the panoramic data view (pre-selected 50, such as coal feed rate, main steam pressure, turbine valve opening, etc.), within a 30-minute sliding window, the continuous cohomology method in topological data analysis is used to identify key points (local maxima, minima, inflection points, start and end points of flat segments) on each time series curve. Then, a top-down piecewise linear approximation is used to approximate each curve. It is decomposed into a sequence of finite morphological elements consisting of rising segments (U), falling segments (D), stable segments (S), inflection points (P), and extreme points (E). For example, the main steam pressure curve can be represented as [U,P,D,S,U,E].

[0038] Sub-step S1032: Topology construction. For any two curves... and Calculate the four basic temporal topological relationships between their corresponding morphological primitives: (i) sequence relationship (whether primitive A occurs before primitive B); (ii) inclusion relationship (whether wave A is completely contained within wave B); (iii) intersection relationship (whether the two curves intersect); and (iv) separation relationship (whether the two waves are independent). Using the morphological primitives of all curves as nodes and the above topological relationships as directed edges, construct a heterogeneous curve morphological topology graph. This figure reveals the deep dependency structure between different parameters in terms of curve shape.

[0039] Sub-step S1033: Cross-parameter domain response analysis. Using the Granger causality test, calculate the causal response strength for each pair of parameters (x, y) within a sliding time window. Set the lag order p = 10 (i.e., 10 seconds), calculate the F-statistic and p-value. If the p-value < 0.01, accept "x is a Granger cause of y". Simultaneously, calculate the causal strength using transitive entropy. (Range 0~1) and average response delay (Unit: seconds). All causal relationships form a weighted directed graph—a cross-parameter domain dynamic response correlation matrix. , of which elements .

[0040] Sub-step S1034: Decoupling and Enhancement of Graph Neural Networks. Construct a 2-layer graph isomorphic network (GIN). The initial features of the nodes in this GNN are the dynamic feature representations output in step S102. Each dimension (each node corresponds to one feature dimension). The graph structure of GNN consists of... and Combined: If the curve primitives corresponding to the two parameters are in There is a topological relationship between them, or If a causal relationship exists (strength > 0.3), an edge is added between nodes i and j, with the edge weight encoded by either the causal strength or the topological relationship type. Through multi-layer graph convolution operations, each node aggregates information from its topologically and causally related neighbors. This structured message-passing mechanism effectively separates (decouples) features that interfere with each other due to physical coupling, and enhances key features that occupy information hub positions in the network. For example, the node representing "coal feed rate" passes information to the node representing "main steam pressure," decoupling the "influence component from coal feed rate" and the "influence component from valve opening" in the latter's features to different output dimensions. The output of the GNN is the decoupled feature representation. The dimensions are the same as the input (2048 dimensions), but the mutual information between the dimensions is significantly reduced.

[0041] Step S104: Construct an interpretable decision knowledge graph Based on the decoupled feature representation, a feature distillation module and a decision knowledge graph with physical interpretability are constructed, specifically including the following sub-steps S1041 to S1043.

[0042] Sub-step S1041: Feature distillation. Construct an interpretable autoencoder based on an attention mechanism: the encoder will distill the 2048-dimensional features... The bottleneck layer is compressed to 100 dimensions, and the decoder attempts to reconstruct 20 key physical parameters (such as main steam pressure, NOx concentration, and turbine thermal stress). An attention mechanism is introduced between the bottleneck layer and the decoder, calculating a physical interpretability score for each bottleneck neuron. The score is based on the mutual information between the neuron's output and preset physical labels (such as "boiler combustion stability," "heated surface ash fouling degree," "turbine thermal stress index," and "coal feeder blockage probability"). Neurons with scores higher than a preset threshold (e.g., 0.6) are selected, and their outputs are defined as operational cognitive features. (Typical dimension is 50). Each dimension is given a specific physical meaning, such as f_t[0] = "combustion stability (high / medium / low)", f_t[1] = "heated surface cleanliness (0~1)", f_t[2] = "turbine thermal stress index", etc.

[0043] Sub-step S1042: Knowledge graph construction. To run cognitive features. Each dimension is represented as an entity node. The FP-Growth algorithm is used to mine frequent itemsets from historical operational data (over one year), extracting strong association rules between entities as directed edges. For example: (combustion stability, low) & (oxygen content, low) → (CO concentration, high), with support >10% and confidence >95%. Simultaneously, the experience of senior operation experts (such as Standard Operating Procedures (SOPs) and contingency plans) is transformed into rules (such as "load increase → increase coal feed → increase air volume → pressure increase → open the damper wider"), which are injected as prior knowledge. Finally, the Neo4j graph database is used for storage, forming a directed and labeled decision knowledge graph that can be queried online. It clearly expresses the logical relationship between "state - cause - consequence - operation".

[0044] Sub-step S1043: Decision reasoning. When a decision event is triggered (such as an AGC command increase or a main steam temperature alarm), the decision engine... Path reasoning based on graph search is performed. The A* algorithm or bidirectional breadth-first search (BFS) is used to find the optimal path from the event node to the operation node or root cause node. Forward reasoning generates interpretable, step-by-step control command sequences (e.g., "Due to an increase in load command, cruise operation is required: Step 1: A mill coal feed rate +2t / h; Step 2: After 30 seconds, the air damper +3%; Step 3: After 1 minute, gradually open the regulating valve based on the pressure deviation"). Backward reasoning locates the root cause of the anomaly and provides handling suggestions (e.g., "Abnormal bearing vibration, possible causes: high lubricating oil temperature (probability 70%), it is recommended to check the oil cooler; or bearing wear (probability 25%), it is recommended to apply for load reduction"). All reasoning results are accompanied by confidence levels and evidence paths.

[0045] Step S105: Decision-making closed-loop verification based on digital twin The control strategy generated by intelligent decision-making is input into an AI-driven high-fidelity digital twin parallel system for closed-loop verification and optimization, specifically including the following sub-steps S1051 to S1053.

[0046] Sub-step S1051: Twin Synchronization. The high-fidelity digital twin parallel system is based on a hybrid of a mechanistic model built using the Modelica language and a neural network data-driven model, capable of running at 5-10 times the real-time speed. This system uses the OPCUA protocol to synchronize the physical unit's operating status parameters (all DCS measurement points) and boundary conditions (ambient temperature, coal quality analysis data, etc.) in real-time at 100ms intervals, ensuring consistency between the twin and the physical unit's state.

[0047] Sub-step S1052: Strategy Pre-simulation. Before sending the decision strategy generated in step S104 to the physical units, input it into the digital twin, set the simulation start time to the current time, the simulation duration to 15-30 minutes, and perform a pre-simulation at 5x speed to simulate the dynamic response of the units after the strategy is executed.

[0048] Sub-step S1053: Reliability Assessment and Optimization. The system monitors key safety, economic, and environmental indicators in real time during the simulation process, including but not limited to: main steam temperature change rate ≤ ±5℃ / min, furnace pressure fluctuation ≤ ±300Pa, main steam pressure deviation ≤ ±0.5MPa, NOx emission ≤ 50mg / Nm³, and power supply coal consumption increase rate ≤ 1%. If all indicators are within the preset constraints, the strategy is assessed as reliable and executed. If any indicator exceeds the limit, the strategy optimization module is triggered. This optimization module first uses counterfactual reasoning: it performs small perturbations (e.g., -0.5t / h, +5 seconds) on key parameters in the original strategy (such as coal feed increment and action delay time) to generate multiple candidate strategies for parallel simulation. If all perturbations fail to meet the constraints, it switches to the Deep Reinforcement Learning (DRL) module and uses the Proximal Policy Optimization (PPO) algorithm to explore the optimal strategy parameters online in a twin environment. After multiple iterations (generally ≤ 10 rounds), a strategy that meets all constraints is found and executed. If a feasible strategy cannot be found after exceeding the maximum number of iterations (e.g., 20 rounds), the system will issue an alarm saying "Unable to make an automatic decision, requesting manual intervention".

[0049] like Figure 3 As shown, this embodiment also provides an intelligent decision-making system for coal-fired power units based on curve morphology topology, used to execute the above method. The system includes a multi-source heterogeneous data fusion module 101, a dynamic operation modeling module 102, a curve morphology topology feature decoupling module 103, an interpretable decision knowledge graph module 104, and a digital twin closed-loop verification module 105. Detailed descriptions are as follows: Multi-source heterogeneous data fusion module 101: Used to realize the temporal transformation and feature extraction of DCS data, as well as spatiotemporal alignment and seamless fusion with DCS data to generate a panoramic data view. This module contains six sub-units: Video temporalization unit 1011: Uses a video temporal behavior detection model to extract video keyframe sequences and convert them into time series describing physical events.

[0050] Audio temporalization unit 1012: It uses Mel frequency cepstral coefficients combined with a temporal classification network to convert audio data into a sequence of acoustic events.

[0051] Vibration temporalization unit 1013: After converting vibration data into a spectrum using short-time Fourier transform, a temporal convolutional network is used to extract the frequency band energy change trend.

[0052] Text temporalization unit 1014: Uses a named entity recognition model to convert unstructured inspection records into structured event triples with timestamps.

[0053] Spatiotemporal alignment unit 1015: It realizes time dimension interpolation alignment based on a unified high-precision clock source, and realizes spatial dimension association mapping through the device spatial location encoding table.

[0054] Cross-modal fusion unit 1016: Uses a cross-modal attention mechanism to perform feature-level fusion on aligned multi-source time-series data to generate the panoramic data view.

[0055] Dynamic Operation Modeling Module 102: Based on the panoramic data view, this module establishes a high-precision dynamic operation model covering all operating conditions and outputs dynamic feature representations. Internally, this module contains a 4-layer Dilated Temporal Convolutional Network (DilatedTCN), which, after training with historical data, extracts the output of the penultimate layer as the dynamic feature representation vector.

[0056] Curve Morphology Topology Feature Decoupling Module 103: This module introduces curve morphology topology and cross-parameter domain response correlation, performs feature decoupling and enhancement on the dynamic feature representation, and outputs a decoupled feature representation. This module contains four sub-units: Curve morphology primitive extraction unit 1031: used to decompose time series curves into a finite number of morphology primitives (rising, falling, stationary, inflection point, extreme point) using the continuous coherence method.

[0057] Topology calculation unit 1032: used to calculate the topological relationships (order, inclusion, intersection, separation) of morphological primitives between different curves and construct the curve morphology topology graph.

[0058] Cross-parameter causal analysis unit 1033: used to analyze the response relationship between parameters using Granger causality test or transfer entropy analysis, and to construct a dynamic response correlation matrix.

[0059] Graph Neural Network Decoupling Unit 1034: Used to perform graph convolution operation on dynamic feature representation by utilizing the curve shape topology graph and dynamic response correlation matrix as graph structure, thereby achieving feature decoupling and enhancement.

[0060] Interpretable Decision Knowledge Graph Module 104: This module is used to construct a feature distillation module and a decision knowledge graph based on the decoupled feature representation, generating intelligent decisions with physical interpretability. Internally, this module implements feature distillation (based on an attention bottleneck layer autoencoder), knowledge graph construction (Neo4j graph database, integrating data mining rules and expert experience), and decision reasoning (graph search path reasoning).

[0061] Digital Twin Closed-Loop Verification Module 105: Used to construct an AI-driven high-fidelity digital twin parallel system to perform closed-loop verification and optimization of the intelligent decisions. This module contains four sub-units: Twin synchronization unit 1051: used to synchronize the operating status parameters and boundary conditions of the actual unit in real time via the OPCUA protocol.

[0062] Strategy Pre-playing Unit 1052: Used to perform advance pre-playing of control strategies on twins, simulating their control effects within a preset future time period.

[0063] Constraint Evaluation Unit 1053: Used to compare the pre-simulation results with the preset safety constraint indicators, economic constraint indicators and environmental constraint indicators to determine whether the strategy is reliable.

[0064] Strategy optimization unit 1054: When the pre-exercise result does not meet the constraints, it adjusts the control strategy parameters based on counterfactual reasoning or reinforcement learning methods, and triggers the strategy pre-exercise unit to perform iterative optimization until all constraints are met or the preset maximum number of iterations is reached.

[0065] The aforementioned modules and units work together to fully realize an intelligent decision-making closed loop, from multi-source perception and deep cognition to explainable decision-making and security verification.

[0066] Example 2 This embodiment provides another intelligent decision-making method for coal-fired power units based on curve morphology topology. The steps are basically the same as in Embodiment 1, but the difference lies in the deep integration of the decision knowledge graph with the integrated control strategy of "intelligent control, cruise, and monitoring." Specifically, this includes: Intelligent control: Rules from the knowledge graph are directly applied to the underlying loop control. For example, when the graph contains the rule "(Main steam pressure deviation > 0.3 MPa for 10 seconds) → rapidly increase coal feedforward", the intelligent control module instantiates this rule as a feedforward controller. When the condition is met, incremental commands are directly superimposed on the coal feeder control loop, significantly improving the response speed and anti-interference capability of pressure control.

[0067] Intelligent cruise control: Used to handle complex operating conditions such as dry-wet transitions. Standard operating procedures (SOPs) are encoded as state machines in a knowledge graph. When "load < 30% and separator inlet superheat < 5℃" is detected, the state machine automatically switches from "dry state" to "wet state transition" state, sequentially executing operations such as starting the boiler circulating pump, shutting off the superheater desuperheating water, and adjusting the feedwater bypass valve, all without manual intervention.

[0068] Intelligent monitoring module: When abnormal parameter trends are detected (such as a slow increase in the vibration of the blower bearing), the intelligent monitoring module performs reverse reasoning in the knowledge graph to locate possible causes (excessive lubricating oil temperature, bearing wear, blade imbalance), calculates the probability of each cause, automatically retrieves the trends of relevant measuring points and associates them with operation records, and pushes the most likely cause and handling suggestions (such as "check the cooling water flow of the oil cooler") to the monitoring interface. At the same time, it automatically focuses the relevant system screen to assist operators in making quick decisions.

[0069] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart decision-making method for coal-fired power units based on curve morphology topology, characterized in that, Includes the following steps: Step S1: Perform time-series transformation and feature extraction on the DCS data generated during the operation of the coal-fired power unit, and align and seamlessly integrate the extracted time-series features with the DCS data to construct a panoramic data view covering the entire operating conditions of the unit; wherein, the DCS data includes at least one or more of video data, audio data, vibration data and text data; the panoramic data view contains time-series data of each operating parameter; Step S2: Based on the panoramic data view, establish a high-precision dynamic operation model covering all operating conditions, including unit start-up and shutdown, normal operation, load regulation, and abnormal operating conditions, to form a dynamic feature expression oriented towards operation cognition; Step S3: Introduce curve morphology topology, decompose the time series curve of each operating parameter in the panoramic data view into a finite number of morphological primitives, calculate the temporal topological relationship of the morphological primitives corresponding to the time series curves of different operating parameters to construct a curve morphology topology graph, and use causal analysis to construct a cross-parameter domain dynamic response correlation matrix; based on the curve morphology topology graph and the dynamic response correlation matrix, construct a graph neural network to perform feature decoupling and enhancement on the dynamic feature expression, separate and strengthen key dynamic features, and form a decoupled feature representation; Step S4: Based on the decoupled feature representation, construct a feature distillation module with physical interpretability, extract and distill operational cognitive features with clear physical meaning, and construct a decision knowledge graph based on the operational cognitive features to support intelligent decision-making for integrated control strategies of intelligent control, cruise and monitoring. Step S5: Input the control strategy generated by the intelligent decision into the AI-driven high-fidelity digital twin parallel system for closed-loop simulation verification, optimize the control strategy based on the verification results, and send the optimized control strategy to the actual unit for execution.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: For video data, a video temporal behavior detection model is used to extract keyframe sequences and transform them into time series describing physical events; for audio data, Mel frequency cepstral coefficients combined with a temporal classification network are used to transform it into acoustic event sequences; for vibration data, short-time Fourier transform is used to convert it into a spectrogram, and then a temporal convolutional network is used to extract the frequency band energy change trend; for text data, a named entity recognition model is used to transform unstructured inspection records into structured event triples with timestamps. Based on a unified high-precision clock source, the time-seriesd DCS data is interpolated and aligned with the DCS data in the time dimension, and the spatial dimension association mapping is realized through a pre-established device spatial location encoding table. By utilizing a cross-modal attention mechanism, the aligned multi-source time-series data are fused at the feature level to generate the panoramic data view.

3. The method according to claim 1, characterized in that, Step S3 specifically includes: Using the continuous cohomology method in topological data analysis, key points on each time series curve are identified, and the time series curve is decomposed into a finite number of morphological primitives consisting of rising, falling, stationary, inflection points and extreme points. Calculate the temporal order relationship, inclusion relationship, intersection relationship and separation relationship of the morphological primitives corresponding to different time series curves, and construct the curve morphological topology graph with the morphological primitives of each time series curve as nodes and the temporal order relationship, inclusion relationship, intersection relationship and separation relationship as edges; Using Granger causality test or transitive entropy method, the response intensity and response delay between different operating parameters are calculated, and the cross-parameter domain dynamic response correlation matrix is ​​established; The dynamic feature representation is used as the initial feature of the nodes in the graph neural network. The curve shape topology graph and the dynamic response correlation matrix are used as the graph structure input. The message passing and aggregation of features are realized through multi-layer graph convolution operation, and the decoupled feature representation is output.

4. The method according to claim 1, characterized in that, Step S4 specifically includes: The feature distillation module employs an attention-based interpretable network to calculate a physical interpretability score for each feature dimension in the decoupled feature representation. The physical interpretability score is associated with a preset set of physical labels, and features with scores higher than a preset threshold are selected as the operational cognitive features. Using the aforementioned operational cognitive features as entity nodes, the association rule mining algorithm is used to extract causal relationships, temporal relationships, and collaborative relationships between entities from historical operational data as edges, and combined with expert experience rules, to construct the decision knowledge graph; When a decision event is triggered, path reasoning based on graph search is performed on the decision knowledge graph to generate an interpretable decision path containing decision steps, intermediate states, and expected consequences, as well as its corresponding control instruction sequence.

5. The method according to claim 1, characterized in that, Step S5 specifically includes: The high-fidelity digital twin parallel system synchronizes the operating status parameters and boundary conditions of the actual unit in real time through the OPCUA protocol. The control strategy is simulated in advance on the twin to demonstrate its control effect in a preset future time period. Compare the pre-simulation results with the preset safety, economic, and environmental constraints. If all constraints are met, the system is assessed as reliable and execution is initiated. Otherwise, a strategy optimization module based on counterfactual reasoning or reinforcement learning is triggered to adjust the control strategy parameters until all constraints are met or the preset maximum number of iterations is reached.

6. The method according to claim 1, characterized in that, The method further includes: deeply integrating the decision knowledge graph with the intelligent control, cruise, and monitoring integrated control strategy, wherein: Intelligent control is used to directly generate setpoints or control parameters for the underlying control loop, enabling autonomous adjustment at the loop level. Intelligent cruise is used to automatically call a preset operation sequence state machine from the decision knowledge graph when the unit load changes or the operating condition switches, and to plan and execute a smooth sequence of operation steps. The intelligent monitoring panel is used to monitor the unit status in real time. When abnormal parameter trends are detected, it performs reverse reasoning from the decision knowledge graph to locate the root cause, predict the abnormal development trend, and realize automatic focusing of alarm screens and push processing suggestions.

7. An intelligent decision-making system for coal-fired power units based on curve morphology topology, characterized in that, The system for performing the method according to any one of claims 1 to 6 comprises: The multi-source heterogeneous data fusion module is used to realize the temporal transformation and feature extraction of DCS data, as well as the spatiotemporal alignment and seamless fusion with DCS data to generate a panoramic data view. The dynamic operation modeling module is used to establish a dynamic operation model covering all working conditions based on the panoramic data view and output dynamic feature expressions. The curve morphology topology feature decoupling module is used to introduce the correlation between curve morphology topology and cross-parameter domain response, perform feature decoupling and enhancement on the dynamic feature expression, and output the decoupled feature representation; An interpretable decision knowledge graph module is used to construct a feature distillation module and a decision knowledge graph based on the decoupled feature representation, thereby generating intelligent decisions with physical interpretability. The digital twin closed-loop verification module is used to build an AI-driven high-fidelity digital twin parallel system to perform closed-loop verification and optimization of the intelligent decision.

8. The system according to claim 7, characterized in that, The curve morphology topology feature decoupling module specifically includes: The curve morphology primitive extraction unit is used to decompose a time series curve into a finite number of morphology primitives using the continuous homology method. The topology calculation unit is used to calculate the topological relationships of morphological primitives between different time series curves and construct a curve morphology topology graph. The cross-parameter causal analysis unit is used to analyze the response relationship between parameters using Granger causality tests and to construct a dynamic response correlation matrix. The graph neural network decoupling unit is used to perform graph convolution operations on dynamic feature representations using the curve shape topology graph and dynamic response correlation matrix as graph structures, thereby achieving feature decoupling and enhancement.

9. The system according to claim 7, characterized in that, The multi-source heterogeneous data fusion module specifically includes: The video temporalization unit is used to extract keyframe sequences from the video using a video temporal behavior detection model and convert them into time series describing physical events. The audio temporalization unit is used to convert audio data into a sequence of acoustic events by combining Mel frequency cepstral coefficients with a temporal classification network. The vibration temporalization unit is used to convert vibration data into a spectrum using short-time Fourier transform, and then use a temporal convolutional network to extract the frequency band energy change trend. The text temporalization unit is used to convert unstructured inspection records into structured event triples with timestamps using a named entity recognition model. The spatiotemporal alignment unit is used to achieve time dimension interpolation alignment based on a unified clock source and to achieve spatial dimension association mapping through a device spatial location encoding table; The cross-modal fusion unit is used to perform feature-level fusion on aligned multi-source time-series data using a cross-modal attention mechanism to generate the panoramic data view.

10. The system according to claim 7, characterized in that, The digital twin closed-loop verification module specifically includes: The twin synchronization unit is used to synchronize the operating status parameters and boundary conditions of the actual unit in real time via the OPCUA protocol; The strategy simulation unit is used to perform advance simulation of the control strategy on the twin, simulating its control effect in a preset future time period; The constraint assessment unit is used to compare the pre-simulation results with the preset safety constraint indicators, economic constraint indicators and environmental constraint indicators to determine whether the strategy is reliable. The strategy optimization unit is used to adjust the control strategy parameters based on counterfactual reasoning or reinforcement learning methods when the pre-exercise results do not meet the constraints, and to trigger the strategy pre-exercise unit to perform iterative optimization until all constraints are met or the preset maximum number of iterations is reached.

Images