Thermal power plant scheduling optimization method based on multi-source data fusion

Abstract

The embodiment of the invention provides a thermal power plant scheduling optimization method based on multi-source data fusion, which belongs to the technical field of power system operation and control, and comprises the following steps: accessing multi-source data through a containerized micro-service cluster and a hybrid communication protocol, generating a high-precision spatio-temporal characteristic matrix in combination with cubic spline interpolation optimized by sub-millisecond clock synchronization and a genetic algorithm; constructing a hybrid model fusing data driving and a thermodynamic mechanism, extracting deep features by using wavelet and improved empirical mode decomposition, and realizing collaborative optimization of economy and safety by using a multi-objective optimization algorithm and an intelligent evolutionary strategy; deploying a hierarchical fault-tolerant architecture and a simulation environment based on a high-speed distributed computing platform and a digital twinning technology; kalman filtering closed-loop feedback and an online self-learning mechanism are introduced, so that the model has a self-evolution capability. Prediction precision, adjusting performance and operation economy of thermal power generating unit scheduling can be remarkably improved, and meanwhile safety and stability of the system are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, specifically to a method for optimizing the dispatch of thermal power plants based on multi-source data fusion. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, the penetration rate of renewable energy sources such as wind and solar power in the power grid is rapidly increasing. Their inherent volatility and intermittency pose unprecedented challenges to the safe and stable operation of the power system. Against this backdrop, the role of thermal power units, as an important regulating resource in the traditional power system, is undergoing a profound transformation from "baseload power" providing baseload power to "regulating power" providing flexible regulation capabilities, and is being given more stringent requirements for deep peak shaving, rapid response, and start-up and shutdown capabilities. Therefore, the optimization of thermal power plant dispatch is no longer simply about pursuing the operational economy of the units themselves, but is a key link directly related to grid frequency stability, voltage support, and reserve capacity configuration, becoming one of the core technologies supporting the high proportion of renewable energy consumption.

[0003] Traditional thermal power plant dispatch optimization mainly relies on mechanistic models built upon first principles of thermodynamics and fluid mechanics. While these models have clear physical meanings, they face significant bottlenecks in practical applications: Firstly, mechanistic models typically contain a large number of complex partial differential equations and nonlinear algebraic equations, requiring massive computation and time to solve, making it difficult to meet the real-time decision-making needs of modern power markets at the minute or even second level. Secondly, model parameters are highly dependent on equipment design values ​​and offline calibration, making it difficult to accurately characterize the complex dynamic characteristics of units under unsteady conditions such as variable load and low load, leading to deviations between optimization results and actual operation, and potentially causing safety issues such as equipment exceeding limits and unstable combustion.

[0004] In recent years, with the development of the Industrial Internet of Things (IIoT) and big data technologies, data-driven scheduling optimization methods have become a research hotspot. Modern thermal power plants have accumulated massive amounts of operational data through distributed control systems (DCS) and plant-level monitoring information systems (SIS), providing a foundation for building high-precision prediction and optimization models. However, effectively utilizing this data still faces significant challenges: The "multi-source heterogeneity" problem at the data level: Data originates from systems of different manufacturers, from different eras, and following different communication protocols (such as Modbus, OPC, IEC 61850, etc.), leading to complex data access and parsing. Simultaneously, the acquisition frequencies at different measurement points vary significantly (e.g., millisecond-level pressure versus second-level temperature), and the inconsistencies in timestamp benchmarks across systems cause severe spatiotemporal mismatch. Existing preprocessing methods such as linear interpolation and simple alignment often ignore the dynamic characteristics of the system, easily introducing noise or losing crucial information, making it difficult to generate high-quality modeling input.

[0005] The "black box" and "generalization" problems at the model level: While purely data-driven models (such as deep neural networks) possess powerful nonlinear fitting capabilities, their "black box" nature means they lack explicit constraints on fundamental physical laws (such as energy conservation and mass conservation). Under boundary conditions not covered by training data (such as emergency hill climbing and deep peak shaving), the model may produce predictions that violate physical laws, generating unsafe or even dangerous scheduling instructions, resulting in insufficient generalization ability and reliability.

[0006] The "Static" vs. "Single-Objective" Problem at the Decision-Making Level: Power plant dispatching is a multi-objective optimization problem that requires balancing conflicting goals such as fuel costs, pollutant emissions, equipment wear and tear, and grid command response accuracy. Traditional optimization algorithms often use manual weighting to transform multiple objectives into a single objective, making it difficult to dynamically adjust optimization preferences based on real-time grid demand and unit status. Furthermore, most models employ offline training, lacking online feedback and self-correction mechanisms, making them unable to adapt to characteristic drift caused by equipment aging and changes in operating conditions, resulting in a continuous decline in model performance over time.

[0007] In summary, there is an urgent need to develop a new intelligent dispatching and optimization method for thermal power plants that can efficiently integrate multi-source heterogeneous data, deeply combine the advantages of mechanism constraints and data-driven approaches, and possess dynamic multi-objective optimization and online self-evolution capabilities. This method aims to maximize the regulation performance and economic benefits of the units while ensuring their safe and stable operation, and to provide key technical support for building a new power system. Summary of the Invention

[0008] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a method for optimizing the scheduling of thermal power plants based on multi-source data fusion.

[0009] This invention provides a method for optimizing the scheduling of thermal power plants based on multi-source data fusion, comprising the following steps: Step 1: Data fusion and processing. Access operational data from multiple heterogeneous data sources within the thermal power plant, align the operational data to time based on a unified clock synchronization mechanism, and generate a spatiotemporal feature matrix containing multiple feature variables. Step 2: Model and algorithm optimization. Construct a hybrid model that integrates data-driven model and thermodynamic mechanism constraints. Use the spatiotemporal feature matrix as input and employ a multi-objective optimization algorithm to solve the hybrid model, so as to achieve synergistic optimization of the economic efficiency and safety of thermal power unit operation and output dispatch instructions. Step 3: System architecture and deployment. Deploy and optimize the scheduling system based on a distributed computing platform and digital twin technology. The digital twin technology is used to build a simulation environment consistent with the physical units and to verify the scheduling instructions generated in Step 2 before going online. Step 4: Apply closed-loop and adaptive mechanisms to establish a closed-loop feedback and online self-learning mechanism. Feed back the actual operating data of the thermal power plant to Step 1 to update the spatiotemporal feature matrix and correct the model parameters of the hybrid model in Step 2 online, so that the hybrid model has self-evolution capability.

[0010] Furthermore, the data fusion and processing in step 1 further includes: Design a containerized microservice cluster architecture, where each microservice independently encapsulates the access and parsing functions of a specific data source; The heterogeneous data source is accessed through a hybrid communication protocol architecture and a protocol conversion gateway; A sub-millisecond clock synchronization mechanism based on a precision clock synchronization protocol is introduced to provide a unified high-precision timestamp for all access data; A cubic spline interpolation algorithm combining a genetic algorithm to dynamically adjust the interpolation window is used to resample and precisely align multi-frequency data, ultimately generating the spatiotemporal feature matrix.

[0011] Furthermore, the specific implementation method of the sub-millisecond clock synchronization mechanism in step 1 is as follows: The clock source uses dual-mode timing from the Global Positioning System and the BeiDou satellite, and is equipped with a highly stable rubidium atomic clock as a local time reference to ensure that the clock drift is less than 1 millisecond within 24 hours when the satellite signal fails. A hierarchical synchronization network is constructed between the master clock and the slave clocks of each data acquisition node to achieve a synchronization accuracy better than ±0.5 milliseconds; The unified clock synchronization mechanism ensures that data from different systems, such as distributed control systems, plant-level monitoring information systems, and edge sensors, have a comparable time reference.

[0012] Furthermore, the process of generating the spatiotemporal feature matrix in step 1 further includes: The characteristic variables include boiler main steam pressure, turbine speed, flue gas oxygen content, coal feed rate, and auxiliary machine current; During the matrix generation process, the time series of each feature variable is standardized to make its mean 0 and variance 1, so as to eliminate the influence of dimensions. For missing values ​​in the data, a k-nearest neighbor algorithm-based imputation method is used, where the value of k is adaptively selected in the range of 5 to 20 according to the local data density to ensure the accuracy and robustness of the imputation.

[0013] Furthermore, the hybrid model in step 2 is specifically constructed as follows: The data-driven part employs a deep neural network structure with at least five hidden layers to learn the non-linear mapping relationships between running data. The thermodynamic mechanism constraint part will transform the mass conservation equation, energy conservation equation and entropy increase inequality based on the first and second laws of thermodynamics into soft constraint terms and embed them into the loss function of the deep neural network through the Lagrange multiplier method to ensure that the model output conforms to physical laws. Before inputting the spatiotemporal feature matrix into the hybrid model, wavelet multi-scale decomposition and improved adaptive noise complete set empirical mode decomposition are used to process the input data in order to extract deep time-frequency features.

[0014] Furthermore, the multi-objective optimization algorithm used in step 2 includes at least one of the following: The non-dominated sorting genetic algorithm with an elitist strategy performs global optimization by simulating binary crossover and polynomial mutation operators, and uses an elitist retention strategy to ensure that the Pareto optimal solution set is not lost. The Crowned Porcupine optimization algorithm explores the solution space by simulating the foraging and defense behaviors of a Crowned Porcupine population; and... A hybrid algorithm combining encoded genetics and deep reinforcement learning is proposed, wherein the encoded genetic algorithm is used to perform a coarse global search to generate a diverse initial solution set, and the deep reinforcement learning agent uses the initial solution set as an experience pool to finely optimize the policy through interaction with the environment.

[0015] Furthermore, the distributed computing platform in step 3 consists of at least 10 computing nodes, each configured with no less than 16-core central processing unit, 64 gigabytes of memory and 1 terabyte of solid-state drive, supporting a parallel processing framework based on memory computing. The simulation environment constructed by the digital twin technology uses computational fluid dynamics and finite element analysis methods to perform a 1:1 digital mapping of key physical processes such as boiler combustion, steam-water circulation, and turbine work, achieving a simulation step size of 0.1 seconds, and ensuring that the model accuracy verification error of the hybrid model in the simulation environment is less than 5%.

[0016] Furthermore, the specific process for verifying the scheduling instructions before they go online in step 3 is as follows: The set of candidate scheduling instructions generated in step 2 is input into the digital twin simulation environment; The simulation environment is driven to simulate the dynamic response process of the physical unit under the candidate scheduling command, and to monitor whether the core safety indicators, including the temperature and pressure of key equipment, exceed the preset safety threshold. Only when the simulation results show that the execution process of a certain command among the candidate scheduling commands meets all safety constraints is the command judged as verified and allowed to be sent to the actual thermal power unit control system.

[0017] Furthermore, the closed-loop feedback mechanism in step 4 is implemented based on the extended Kalman filter algorithm. This mechanism uses the weight parameters and bias parameters of the hybrid model as state variables, sets the state transition matrix as an identity matrix, and adaptively adjusts the process noise covariance matrix according to the rate of change of the model parameters. The observation matrix is ​​obtained by linearizing the input-output relationship of the hybrid model. Each time a new batch of actual operating data is received, a prediction and update step is performed, thereby realizing the recursive correction of the model parameters.

[0018] Furthermore, the online self-learning mechanism in step 4 is implemented based on an online sequence extreme learning machine, which is configured with 1000 hidden layer nodes and a regularization coefficient of 0.001. When new operational data is received, the mechanism only updates the output weights analytically based on the new data and some historical data, without iteratively retraining the entire model. It also supports the dynamic addition and deletion of the number of hidden layer nodes in the learning machine, thereby enabling the learning machine to be lightweight and rapidly evolve online to adapt to the slow changes in the unit's status.

[0019] Compared with existing technologies, the thermal power plant scheduling optimization method based on multi-source data fusion in this invention not only solves the core pain points such as multi-source data fusion, model inaccuracy, and system fragility through the organic coordination of each step, but also realizes the paradigm shift of thermal power plant scheduling from "static rule-driven" to "dynamic intelligent optimization". Under the premise of ensuring the safe and stable operation of the units, it significantly improves their regulation performance, economic benefits and environmental protection level. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the thermal power plant scheduling optimization method based on multi-source data fusion proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework for multi-source heterogeneous data fusion and spatiotemporal feature matrix construction in this invention; Figure 3 This is a logical flowchart of the hybrid modeling and multi-objective optimization algorithm that integrates data-driven and physical mechanisms in this invention. Figure 4This is a schematic diagram of the multi-level interaction relationship and data flow of the closed-loop adaptive evolution and digital twin verification of the scheduling optimization system in this invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0024] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0025] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0026] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0027] refer to Figure 1 The schematic diagram of the overall technical solution architecture shown illustrates that the execution flow of the method described in this invention begins with the access and fusion of multi-source heterogeneous data, proceeds through hybrid modeling and multi-objective optimization, and finally achieves the generation and execution of scheduling instructions through high-availability system deployment and a closed-loop adaptive mechanism. The entire architecture is clearly divided into four logical layers: a data fusion layer, a model optimization layer, a system deployment layer, and an application closed-loop layer. Each layer interacts with data and instructions through standardized interfaces, ensuring the modularity, scalability, and engineering feasibility of the system.

[0028] In the aforementioned method for optimizing the scheduling of thermal power plants based on multi-source data fusion, step 1, data fusion and processing, aims to address the challenges of accessing, aligning, and processing heterogeneous multi-source data from thermal power plants, as well as the resulting poor data quality. This provides a highly consistent and complete spatiotemporal feature matrix for subsequent modeling. Specifically, step 1 first designs a containerized microservice cluster architecture based on container orchestration technology. (See [link to relevant documentation]). Figure 2 The diagram illustrates the core principle framework. In this architecture, each microservice independently encapsulates the access, parsing, and preprocessing functions of a specific type of data source, supporting hot-swapping and elastic scaling. The number of cluster nodes is dynamically adjusted according to the real-time data source scale, with a minimum configuration of no less than 3 nodes to ensure high availability and fault tolerance. Microservices communicate asynchronously through a lightweight message bus, with end-to-end data transmission latency controlled within 10 milliseconds, effectively supporting the needs of high concurrency and low latency data flow. This architecture significantly improves the system's ability to quickly integrate new data sources, avoiding the high coupling and poor scalability defects of traditional monolithic architectures. At the engineering implementation level, each microservice instance runs in an independent Linux container, encapsulated using a Docker image. Its resource configuration (number of CPU cores, memory size) can be dynamically allocated according to the throughput of the data source being processed, ensuring maximum resource utilization efficiency. The message bus adopts a publish-subscribe pattern. Data producer microservices publish parsed structured data to specific topics, while data consumer microservices subscribe to the required topics, achieving decoupling between producers and consumers and enhancing the system's flexibility and robustness.

[0029] Secondly, addressing the fragmentation of industrial field communication protocols, this invention constructs a hybrid communication protocol system, integrating the OPC UA protocol for process control and the MQTT protocol for the Internet of Things (IoT). The OPC UA protocol interfaces with industrial automation systems such as DCS and SIS, supporting complex data types and semantic information models to ensure the complete semantic transmission of key control parameters. The MQTT protocol connects edge sensors and smart meters deployed near critical equipment such as boilers and turbines; its lightweight nature enables efficient data upload in low-bandwidth, high-packet-loss industrial wireless environments. To achieve seamless protocol interoperability, the system incorporates a protocol conversion gateway, integrating a mapping rule base containing at least 15 mainstream industrial protocols, supporting automatic identification, parsing, and conversion, significantly reducing system integration complexity. The protocol conversion gateway maintains a dynamic protocol registry. When a new data source is connected, the gateway first attempts to identify its communication protocol type through feature byte matching or port scanning, and then loads the corresponding parser and converter from the mapping rule base. For OPC UA clients, the gateway simulates a standard OPC UA server, providing information model nodes compliant with the IEC 62541 standard. For MQTT clients, the gateway acts as a message broker, receiving JSON-formatted telemetry data from edge devices and converting it into a unified internal data structure. This hybrid protocol architecture design enables the system to seamlessly integrate a full spectrum of data sources, from traditional industrial control systems to modern IoT devices, laying the foundation for building a comprehensive profile of unit operating status.

[0030] To address the issue of inconsistent timestamps across multiple data sources, this invention introduces a sub-millisecond-level clock synchronization mechanism based on the IEEE 1588 precision clock synchronization protocol. A hierarchical synchronization network is constructed between the master and slave clocks. The master clock utilizes GPS and BeiDou dual-mode satellite timing, supplemented by a highly stable rubidium atomic clock as the local time reference. Even in the event of complete satellite signal failure, the system can still control clock drift within 1 millisecond for 24 hours, achieving a synchronization accuracy better than ±0.5 milliseconds. This high-precision clock system provides a reliable time reference for subsequent data alignment. In practical deployment, the master clock server is deployed in the power plant's information center, distributing precise time signals to downstream switches and key data acquisition nodes (such as the DCS master control station, SIS server, and edge computing gateway) via a fiber optic network. Each slave clock node has a built-in PTP protocol stack, periodically exchanging timestamps with the upstream clock, calculating network latency and local clock offset, and adjusting the local system clock accordingly. This synchronization mechanism ensures that all data points from all data sources have unified and accurate timestamps, providing a prerequisite for subsequent spatiotemporal alignment operations.

[0031] Based on this, this invention proposes a dynamic interpolation window optimization method combining cubic spline interpolation and a genetic algorithm. First, outlier detection is performed on the original time series, using the Laida criterion to identify and remove outliers exceeding three times the standard deviation, ensuring data quality. Then, with minimizing interpolation error as the objective function, a genetic algorithm dynamically searches for the optimal interpolation window size within a range of 0.5 to 5 times the original sampling interval. The algorithm sets the population size to 100, the crossover probability to 0.8, and the mutation probability to 0.01, iterating for 50 generations before outputting the optimal window parameters. Finally, cubic spline interpolation is applied within this window for resampling, generating a second-order continuous and smooth interpolation curve, effectively preserving the dynamic characteristics of the original signal. This method overcomes the insufficient adaptability of fixed-window interpolation in high-frequency abrupt changes or low-frequency stationary scenarios. In terms of algorithm execution details, for each variable time series to be aligned, the system first calculates the mean and standard deviation within its sliding window, identifying and marking outliers. Subsequently, the individual encoding of the genetic algorithm is a real number representing the interpolation window size (in units of the original sampling interval). The fitness function is defined as the root mean square error (RMSE) between the interpolated points and the true observation points after cubic spline interpolation using this window on the validation set. After several generations of evolution, the algorithm converges to the optimal window parameter, which is used to resample the entire time series to generate an aligned data stream with a uniform time granularity (e.g., 1 second).

[0032] Finally, the system organizes the aligned multi-source data into a spatiotemporal feature matrix, with dimensions equal to the time series length multiplied by the number of feature variables. These feature variables cover key operating parameters such as boiler main steam pressure, turbine speed, flue gas oxygen content, coal feed rate, and auxiliary machine current. During matrix generation, each variable is standardized (mean 0, variance 1), and missing values ​​are imputed using a k-nearest neighbor (kNN) strategy. The k value is adaptively selected based on the local data distribution density (range 5–20), ensuring imputation accuracy while avoiding human bias. This spatiotemporal feature matrix provides a structured, high-quality input foundation for subsequent modeling. The spatiotemporal feature matrix is ​​stored in memory or cache as a two-dimensional array, with each row corresponding to a timestamp and each column corresponding to a feature variable. Standardization eliminates scale differences between variables of different dimensions, making the model training process more stable. For kNN imputation, the system first calculates the local density of the points to be imputed in the feature space. Higher density results in a smaller k value (e.g., 5) to preserve local details; lower density results in a larger k value (e.g., 20) to enhance the robustness of the imputation. The filling process is completed synchronously during the matrix construction stage, ensuring that the output spatiotemporal feature matrix has no missing values ​​and can be directly used for model training and inference.

[0033] In the aforementioned method for optimizing the scheduling of thermal power plants based on multi-source data fusion, step 2, model and algorithm optimization, aims to construct a hybrid optimization model that combines high prediction accuracy with physical interpretability, and to achieve coordinated scheduling that balances economy and safety through advanced multi-objective optimization algorithms. Its logical flow framework can be found in [reference needed]. Figure 3 Specifically, step 2 proposes a hybrid modeling method that deeply integrates data-driven and mechanism analysis: a deep neural network is responsible for learning complex nonlinear mapping relationships from massive operational data, while a physical constraint module based on the first and second laws of thermodynamics ensures that the model output conforms to fundamental physical laws such as energy conservation, mass conservation, and entropy increase. Specifically, the DNN adopts a 5-layer multilayer perceptron structure (with 128, 256, 256, 128, and 64 neurons respectively), with ReLU as the activation function; physical constraints are embedded in the loss function using the Lagrange multiplier method as soft constraints, with a constraint violation penalty coefficient set to 0.1. This design retains the strong fitting ability of the data-driven model while avoiding the physical inconsistencies that may occur in a pure black-box model, significantly improving the reliability of the model under extreme operating conditions. The input of the hybrid model is the spatiotemporal feature matrix generated in step 1, and the output is the expected operating state of the unit under a given scheduling instruction (such as coal consumption, emissions, and temperature / pressure of key equipment). After each forward propagation, the physics constraint module calculates the physical plausibility score of the model output based on the laws of thermodynamics and adds this score as a penalty term to the total loss function. During backpropagation, this penalty term guides the network weights to be updated in a direction that conforms to physical laws, thus embedding a safety guarantee during the training phase.

[0034] To uncover deep time-frequency features in data, this invention combines wavelet multi-scale decomposition with an improved adaptive noise-complete ensemble empirical mode decomposition (EMD). Wavelet decomposition uses the db4 wavelet basis to perform a 5-level decomposition of the input signal, separating approximation and detail components in different frequency bands. ICEEMDAN effectively suppresses mode aliasing in traditional EMD through adaptive noise injection and ensemble averaging strategies. The number of intrinsic mode functions (IMFs) obtained from the decomposition is automatically determined based on the signal complexity (typically 8–12). These two types of features together constitute the enhanced input of the model, improving its ability to characterize the dynamic behavior of the generator unit. In the feature engineering stage, the original spatiotemporal feature matrix is ​​first fed into the wavelet decomposition module to obtain 5 levels of approximation coefficients and 5 levels of detail coefficients, which reflect the energy distribution of the signal at different time scales. Simultaneously, the original signal is also fed into the ICEEMDAN module, decomposed into several IMF components and a residual trend term. All these decomposition products (wavelet coefficients, IMF components) are reorganized and concatenated into a new, higher-dimensional feature vector, which serves as the final input to the DNN model. This multi-scale feature fusion strategy enables the model to simultaneously capture the long-term trend, medium-term fluctuations, and short-term transients of the unit's operating status, greatly enriching the connotation of the input information.

[0035] At the optimization level, this invention employs a non-dominated sorting genetic algorithm with an elitist strategy to handle multi-objective optimization problems (such as minimizing coal consumption, minimizing emissions, and maximizing load response). The population size is set to 200, the crossover operator uses simulated binary crossover (distribution exponent 20), the mutation operator uses multinomial mutation (distribution exponent 20), and the evolution lasts for 100 generations. An elitist retention strategy ensures that the Pareto front solution set is not lost. Furthermore, for high-dimensional nonlinear problems, a porcupine optimization algorithm is introduced to simulate its swarm intelligence behavior during foraging (population size 50, perception probability 0.7, escape probability 0.3, 100 iterations), demonstrating excellent performance in global search capability and convergence speed. Further, this invention designs a hybrid algorithm combining encoded genetics and deep reinforcement learning: first, a coarse search is performed using an encoded genetic algorithm to generate a diverse initial solution set; then, this initial set is used as an experience pool for a DRL agent (using the Deep Deterministic Policy Gradient (DDPG) algorithm). Both the actor and critic networks are 3-layer fully connected layers (256 neurons per layer, learning rate 0.001), achieving fine optimization through interaction with the environment. This hybrid strategy combines global exploration with local exploitation, significantly improving the quality and stability of the solution. The multi-objective optimizer takes the current grid dispatch instructions (such as target load) and the current state of the generating units as inputs, and outputs a set of Pareto-optimal dispatch instructions (such as coal feed, air volume, valve opening, etc.). The NSGA-II algorithm first generates a random initial population, and then performs selection, crossover, and mutation through non-dominated sorting and congestion calculation to gradually approach the Pareto front. The CPO algorithm serves as an alternative or supplementary scheme, and its unique search mechanism may be more advantageous in handling certain types of non-convex and discontinuous optimization problems. The hybrid algorithm of encoded genetics and DRL represents a higher-order optimization strategy: the encoded genetic algorithm quickly explores the solution space and finds multiple potential regions; the DRL agent then performs refined policy gradient search within these regions, using its powerful function approximation ability to find local optima. This hierarchical optimization architecture ensures that the system can always find high-quality feasible solutions in complex and variable dispatch scenarios.

[0036] In the aforementioned method for optimizing the scheduling of thermal power plants based on multi-source data fusion, step 3, system architecture and deployment, aims to support the efficient operation of the algorithm by constructing a highly available and secure system deployment architecture. Options include building a new high-speed distributed computing platform or repurposing existing private cloud resources. The new platform, based on open-source frameworks such as Apache Spark, consists of at least 10 computing nodes (each node ≥16 CPU cores, 64GB memory, 1TB SSD), supporting in-memory computing and pipelined parallelism. Data storage utilizes a combination of columnar databases and a distributed file system, along with data compression and partitioned indexing, ensuring a retrieval response time of less than 2 seconds in 95% of query scenarios. The repurposing solution integrates existing servers using KVM virtualization technology, creating resource-isolated virtual machine instances. The cloud management platform provides automatic scaling and fault migration capabilities within 30 seconds, maximizing the utilization of existing assets. (Reference) Figure 1 The entire system is deployed in the power plant's information management area, interacting with the DCS / SIS system in the production control area via a one-way isolation gateway to ensure production safety. The computing platform adopts a master-slave architecture, with the master node responsible for task scheduling and resource management, and the slave nodes responsible for specific computing tasks. Spark's in-memory computing capabilities allow large-scale data processing and model training to be completed in memory, avoiding frequent disk I / O and significantly improving processing speed. Columnar databases (such as Apache Parquet) are particularly suitable for storing wide-table data such as spatiotemporal feature matrices; their column-based storage characteristic means that when querying specific feature variables, only the relevant columns need to be read, greatly reducing I / O overhead.

[0037] Simultaneously, a high-fidelity simulation testing environment was constructed using digital twin technology. This environment includes a physics engine based on computational fluid dynamics and finite element analysis, enabling a 1:1 digital mapping of processes such as boiler combustion, steam-water circulation, and turbine operation. Real-time data-driven models synchronously reproduce the unit's dynamic behavior, with a simulation step size of 0.1 seconds and a model verification error of <5%. This environment is used for safety verification before new strategies are implemented, avoiding the risks associated with direct intervention in the production system. See [link to digital twin environment documentation] for details. Figure 4The diagram illustrates the multi-level interaction relationships and data flow. This environment consists of a physical model layer, a data-driven model layer, and a visualization layer. The physical model layer, built on CFD / FEA software, accurately simulates the combustion process, flue gas flow, heat and mass transfer, and other physical phenomena within the furnace. The data-driven model layer deploys the hybrid model trained in step 2, enabling rapid prediction of unit response based on input scheduling commands. Before deploying a new scheduling strategy, the system first conducts closed-loop simulation testing in the digital twin environment: the commands generated by the new strategy are input into the digital twin, and its dynamic response is observed to ensure it meets safety boundaries (e.g., temperature and pressure do not exceed limits). Only strategies that pass verification are allowed to be deployed to the real unit. This "simulate first, then execute" mechanism provides dual protection for the safe operation of the system.

[0038] In terms of security architecture, the system deploys a hierarchical fault tolerance and closed-loop feedback mechanism in the management and production control areas: dual-machine hot standby for critical servers, redundant fiber optic network links, 1-second heartbeat detection at the software layer (with a 3-second timeout threshold), and a "daily full backup + hourly incremental backup" strategy for data (retaining data for ≥30 days). This design significantly improves the system's continuous service capability under abnormal conditions such as hardware failures and network interruptions. The dual-machine hot standby adopts a primary-backup mode; if the primary server fails, the backup server can take over all services within 30 seconds. The network layer uses dual uplink fiber optic links, achieving bandwidth aggregation and fault switching through link aggregation technology. The software heartbeat detection mechanism is implemented by an independent monitoring process; once a core service process is detected as unresponsive, it immediately triggers a service restart or fault migration process. The data backup strategy ensures that the system can recover data at any point within 30 days under any circumstances, maximizing the protection of historical operational data and model assets.

[0039] In the aforementioned method for optimizing the dispatch of thermal power plants based on multi-source data fusion, step 4 applies closed-loop and adaptive mechanisms. To achieve continuous model evolution, this invention establishes a closed-loop feedback mechanism based on Kalman filtering. The weights and bias parameters of the hybrid model are used as state variables to construct a state-space model. The state transition matrix is ​​an identity matrix, and the process noise covariance is adaptively adjusted according to the parameter change rate. The observation matrix is ​​obtained by linearizing the model. After receiving a batch of new operating data every minute, an extended Kalman filter prediction-update step is executed once, recursively correcting the model parameters to ensure that they always closely reflect the actual unit characteristics. (Reference) Figure 4The actual operational data is written back to the feature matrix after integrity verification, millisecond-level timestamp alignment, and version management, forming a data closed loop. Version management records the operation time, content, and responsible person, supporting data backtracking and difference analysis. Based on this, the model achieves incremental learning through an online sequence extreme learning machine: 1000 hidden layer nodes, a regularization coefficient of 0.001, updating parameters only based on new data and a portion of old data, eliminating the need for full retraining, and supporting dynamic structural adjustments (such as adding or deleting nodes), achieving lightweight online evolution. Furthermore, the system automatically evaluates model performance (indicators including RMSE, MAPE, and R²) every 24 hours. If any indicator deteriorates by more than 10%, a retraining process is triggered, retraining using data from the most recent 30 days, and only going live after passing security verification in a digital twin environment. This mechanism ensures the model maintains high accuracy and high reliability over the long term. The closed-loop feedback mechanism is the core of this invention's long-term adaptive capability. The EKF mechanism enables the model to compensate in real time for changes such as slow equipment aging and sensor drift. OS-ELM offers a more efficient incremental learning approach. Its single-hidden-layer feedforward network structure and analytical solution method allow parameter updates to be completed within milliseconds, making it ideal for online applications. The performance evaluation module acts as the system's "health monitor," periodically performing "check-ups" on the model. If a performance decline is detected, a retraining process is immediately initiated to ensure the system is always in optimal working condition.

[0040] To illustrate the application effects of this invention, a typical application example is constructed: A 600 MW supercritical coal-fired unit needs to reduce its load from 600 MW to 300 MW within 30 minutes when participating in a deep peak-shaving task for the power grid. Traditional dispatching methods generate instructions based on static coal consumption characteristic curves, leading to unstable boiler combustion and severe fluctuations in main steam pressure during load reduction, triggering multiple safety protection actions. After applying the method of this invention, the system first constructs a high-precision spatiotemporal feature matrix by fusing DCS, SIS, and edge sensor data in step 1; the hybrid model in step 2 accurately predicts the combustion stability and pressure response under different load reduction rates; the multi-objective optimizer comprehensively considers coal consumption, emissions, and combustion stability to generate a smooth and safe load reduction curve; this curve is then validated in a digital twin environment before being deployed for execution. Actual operation results show that the main steam pressure fluctuation amplitude is reduced by 60%, no further protection actions occur, and coal consumption increases by only 2%, significantly better than the traditional method.

[0041] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for optimizing the scheduling of thermal power plants based on multi-source data fusion, characterized in that, Includes the following steps: Step 1: Data fusion and processing. Access operational data from multiple heterogeneous data sources within the thermal power plant, align the operational data to time based on a unified clock synchronization mechanism, and generate a spatiotemporal feature matrix containing multiple feature variables. Step 2: Model and algorithm optimization. Construct a hybrid model that integrates data-driven model and thermodynamic mechanism constraints. Use the spatiotemporal feature matrix as input and employ a multi-objective optimization algorithm to solve the hybrid model, so as to achieve synergistic optimization of the economic efficiency and safety of thermal power unit operation and output dispatch instructions. Step 3: System architecture and deployment. Deploy and optimize the scheduling system based on a distributed computing platform and digital twin technology. The digital twin technology is used to build a simulation environment consistent with the physical units and to verify the scheduling instructions generated in Step 2 before going online. Step 4: Apply closed-loop and adaptive mechanisms to establish a closed-loop feedback and online self-learning mechanism. Feed back the actual operating data of the thermal power plant to Step 1 to update the spatiotemporal feature matrix and correct the model parameters of the hybrid model in Step 2 online, so that the hybrid model has self-evolution capability.

2. The method according to claim 1, characterized in that, The data fusion and processing in step 1 further includes: Design a containerized microservice cluster architecture, where each microservice independently encapsulates the access and parsing functions of a specific data source; The heterogeneous data source is accessed through a hybrid communication protocol architecture and a protocol conversion gateway; A sub-millisecond clock synchronization mechanism based on a precision clock synchronization protocol is introduced to provide a unified high-precision timestamp for all access data; A cubic spline interpolation algorithm combining a genetic algorithm to dynamically adjust the interpolation window is used to resample and precisely align multi-frequency data, ultimately generating the spatiotemporal feature matrix.

3. The method according to claim 2, characterized in that, The specific implementation method of the sub-millisecond-level clock synchronization mechanism in step 1 is as follows: The clock source uses dual-mode timing from the Global Positioning System and the BeiDou satellite, and is equipped with a highly stable rubidium atomic clock as a local time reference to ensure that the clock drift is less than 1 millisecond within 24 hours when the satellite signal fails. A hierarchical synchronization network is constructed between the master clock and the slave clocks of each data acquisition node to achieve a synchronization accuracy better than ±0.5 milliseconds; The unified clock synchronization mechanism ensures that data from different systems, such as distributed control systems, plant-level monitoring information systems, and edge sensors, have a comparable time reference.

4. The method according to claim 2, characterized in that, The process of generating the spatiotemporal feature matrix in step 1 further includes: The characteristic variables include boiler main steam pressure, turbine speed, flue gas oxygen content, coal feed rate, and auxiliary machine current; During the matrix generation process, the time series of each feature variable is standardized to make its mean 0 and variance 1, so as to eliminate the influence of dimensions. For missing values ​​in the data, a k-nearest neighbor algorithm-based imputation method is used, where the value of k is adaptively selected in the range of 5 to 20 according to the local data density to ensure the accuracy and robustness of the imputation.

5. The method according to any one of claims 1 to 4, characterized in that, The hybrid model in step 2 is specifically constructed as follows: The data-driven part employs a deep neural network structure with at least five hidden layers to learn the non-linear mapping relationships between running data. The thermodynamic mechanism constraint part will transform the mass conservation equation, energy conservation equation and entropy increase inequality based on the first and second laws of thermodynamics into soft constraint terms and embed them into the loss function of the deep neural network through the Lagrange multiplier method to ensure that the model output conforms to physical laws. Before inputting the spatiotemporal feature matrix into the hybrid model, wavelet multi-scale decomposition and improved adaptive noise complete set empirical mode decomposition are used to process the input data in order to extract deep time-frequency features.

6. The method according to claim 5, characterized in that, The multi-objective optimization algorithm used in step 2 includes at least one of the following: The non-dominated sorting genetic algorithm with an elitist strategy performs global optimization by simulating binary crossover and polynomial mutation operators, and uses an elitist retention strategy to ensure that the Pareto optimal solution set is not lost. The Crowned Porcupine optimization algorithm explores the solution space by simulating the foraging and defense behaviors of a Crowned Porcupine population; and... A hybrid algorithm combining encoded genetics and deep reinforcement learning is proposed, wherein the encoded genetic algorithm is used to perform a coarse global search to generate a diverse initial solution set, and the deep reinforcement learning agent uses the initial solution set as an experience pool to finely optimize the policy through interaction with the environment.

7. The method according to any one of claims 1 to 4, characterized in that, The distributed computing platform in step 3 consists of at least 10 computing nodes, each configured with no less than 16-core central processing unit, 64 gigabytes of memory and 1 terabyte of solid-state drive, and supports a parallel processing framework based on memory computing. The simulation environment constructed by the digital twin technology uses computational fluid dynamics and finite element analysis methods to perform a 1:1 digital mapping of key physical processes such as boiler combustion, steam-water circulation, and turbine work, achieving a simulation step size of 0.1 seconds, and ensuring that the model accuracy verification error of the hybrid model in the simulation environment is less than 5%.

8. The method according to claim 7, characterized in that, The specific process for verifying the scheduling instructions before they go online in step 3 is as follows: The set of candidate scheduling instructions generated in step 2 is input into the digital twin simulation environment; The simulation environment is driven to simulate the dynamic response process of the physical unit under the candidate scheduling command, and to monitor whether the core safety indicators, including the temperature and pressure of key equipment, exceed the preset safety threshold. Only when the simulation results show that the execution process of a certain instruction among the candidate scheduling instructions meets all safety constraints will the instruction be deemed to have passed verification and be allowed to be sent to the actual thermal power unit control system.

9. The method according to any one of claims 1 to 4, characterized in that, The closed-loop feedback mechanism in step 4 is based on the extended Kalman filter algorithm. This mechanism uses the weight parameters and bias parameters of the hybrid model as state variables, sets the state transition matrix as the identity matrix, and adaptively adjusts the process noise covariance matrix according to the rate of change of the model parameters. The observation matrix is ​​obtained by linearizing the input-output relationship of the hybrid model. Each time a new batch of actual operating data is received, a prediction and update step is performed, thereby realizing the recursive correction of the model parameters.

10. The method according to claim 9, characterized in that, The online self-learning mechanism in step 4 is implemented based on an online sequence extreme learning machine, which is configured with 1000 hidden layer nodes and a regularization coefficient of 0.

001. When new operational data is received, the mechanism only updates the output weights analytically based on the new data and some historical data, without iteratively retraining the entire model. It also supports the dynamic addition and deletion of the number of hidden layer nodes in the learning machine, thereby enabling the learning machine to be lightweight and rapidly evolve online to adapt to the slow changes in the unit's status.

Images