Power generation control system collaborative optimization data processing method and system based on large model

Abstract

The application relates to the technical field of power generation control, and discloses a power generation control system collaborative optimization data processing method and system based on a large model, multi-source heterogeneous data in a power generation control system is collected, the multi-source heterogeneous data is subjected to abnormal value detection and standardization processing, and then a standardized data set is obtained through a Min-Max normalization algorithm; a power generation industry knowledge graph is constructed, entity recognition and linking are carried out, the standardized data set is fused and trained with the power generation industry knowledge graph, and a power generation large model is obtained; a high-dimensional feature vector is extracted by using the power generation large model, dimension reduction processing is carried out by using a UMAP algorithm, target feature vectors are screened out by combining a multi-head attention mechanism and a mutual information method; the target feature vectors are input into a CNN-LSTM fusion model to identify real-time working conditions of the power generation control system, and control parameters are dynamically optimized based on a DRL algorithm; and the application significantly improves the operation efficiency of the power generation control system.

Description

Technical Field

[0001] This invention relates to the field of power generation control technology, specifically to a collaborative optimization data processing method and system for power generation control systems based on a large model. Background Technology

[0002] As the power generation industry transforms towards digitalization and intelligence, the operation of power generation control systems relies on massive amounts of heterogeneous data from multiple sources. This includes data generated by various systems such as production control systems, enterprise resource planning, equipment management systems, and energy management systems. The data encompasses various types, including structured operating parameters, unstructured equipment logs, and semi-structured monitoring reports. The data sources are diverse, formats vary, and quality is inconsistent, with issues such as incompatible interfaces and data silos, posing significant challenges to data fusion and processing. Existing data processing methods for power generation control systems mostly employ traditional data cleaning and fusion techniques, resulting in low efficiency in multi-source data fusion and a lack of targeted solutions. Standardized processing algorithms are insufficient for quickly adapting to heterogeneous data of different formats and precisions, and key features are easily lost during data cleaning, failing to meet real-time control requirements. They rely heavily on manual experience or simple statistical methods, making it difficult to accurately extract core features related to power generation conditions and control parameters from high-dimensional data, resulting in low accuracy in subsequent condition identification. Decision optimization lacks coordination, failing to achieve deep linkage between data processing, condition identification, and parameter optimization, and often employs a single control strategy, which cannot adapt to the nonlinear, time-varying, and strongly coupled operating characteristics of power generation systems, making it difficult to simultaneously improve system operating efficiency, safety, stability, and economy. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a collaborative optimization data processing method and system for a power generation control system based on a large model.

[0004] The first aspect of this invention provides a collaborative optimization data processing method for a power generation control system based on a large model, the method comprising the following steps: Collect multi-source heterogeneous data from the power generation control system, perform outlier detection and standardization on the multi-source heterogeneous data, and then obtain a standardized data set through the Min-Max normalization algorithm; A knowledge graph of the power generation industry is constructed, and entity recognition and linking are performed. The standardized dataset is integrated with the knowledge graph of the power generation industry for training to obtain a large power generation model. High-dimensional feature vectors are extracted using a large-scale power generation model, and dimensionality reduction is performed using the UMAP algorithm. The target feature vectors are then selected by combining multi-head attention mechanism and mutual information method. The target feature vector is input into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and the control parameters are dynamically optimized based on the DRL algorithm.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the step of collecting multi-source heterogeneous data from the power generation control system, performing outlier detection and standardization on the multi-source heterogeneous data, and then obtaining a standardized data set through a Min-Max normalization algorithm includes: Collect multi-source heterogeneous data within the power generation control system, and aggregate structured operating parameters, unstructured equipment logs, and semi-structured scheduling instructions to form a raw dataset to be processed; The data space in the original dataset to be processed is randomly and recursively divided, the isolated path length of each data point is calculated, and abnormal data in the original dataset to be processed is determined and removed based on the path length. The numerical data after outlier removal is scaled; the unstructured text data after outlier removal is converted into numerical features through word frequency statistics and feature quantization; and the semi-structured data after outlier removal is extracted with effective numerical information through hierarchical parsing. The standardized data is input into the Min-Max normalization algorithm, and all data are scaled according to a unified interval to obtain a standardized dataset.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of constructing a power generation industry knowledge graph, performing entity recognition and linking, and fusing a standardized dataset with the power generation industry knowledge graph for training to obtain a large-scale power generation model includes: Acquire data on equipment, operating conditions, control parameters, and fault types in the power generation field, and build a power generation industry knowledge graph framework that includes an entity layer, a relationship layer, and an attribute layer. Entity extraction is performed on text and data in the power generation field to identify the entity information required for the power generation industry knowledge graph framework. The identified entities are vector-encoded, and entity alignment and linking are performed through similarity matching to obtain a knowledge graph of the power generation industry. Standardized datasets are used as training inputs, and entity and relationship information from the power generation industry knowledge graph is incorporated into the model training process as prompts. The LoRA low-rank adaptation algorithm is used to perform lightweight fine-tuning on the basic large model, enabling the model to learn the data patterns of the power generation industry and the logical connections with the knowledge graph. After training, a large power generation model is obtained.

[0007] Optionally, in the third implementation of the first aspect of the present invention, the standardized dataset is used as training input, and entity and relation information from the power generation industry knowledge graph is integrated into the model training process in the form of prompts. The LoRA low-rank adaptation algorithm is used to perform lightweight fine-tuning of the basic large model, enabling the model to learn the data patterns of the power generation industry and the logical association with the knowledge graph. After training, a large power generation model is obtained, including: The knowledge graph of the power generation industry is analyzed to extract all entity information and the relationships between entities. Entity names, attributes and relationship types are converted into prompt statements to form a set of prompt information. The set of prompt information is matched with the standardized data set. Each piece of standardized data corresponds to the knowledge graph prompt information and is used as the training input for the basic large model. Initialize the parameters of the basic large model, and insert a low-rank adaptation matrix into the Transformer layer of the basic large model; The correlated training input data is fed into the base model, and only the parameters of the low-rank matrix inserted by LoRA are updated during the training process. The model is continuously fine-tuned according to the preset training rounds, and the large power generation model is obtained after training is stopped.

[0008] Optionally, in the fourth implementation of the first aspect of the present invention, the step of extracting high-dimensional feature vectors using a large-scale power generation model, performing dimensionality reduction using the UMAP algorithm, and combining multi-head attention mechanism and mutual information method to select target feature vectors includes: The standardized dataset is input into the trained large-scale power generation model, and the feature encoding layer of the large-scale power generation model is used to extract global features from the data to generate high-dimensional feature vectors. The UMAP uniform manifold approximation and projection algorithm is used to perform dimensionality reduction on high-dimensional feature vectors to obtain feature components in low-dimensional space. By using a multi-head attention mechanism, attention weights are calculated for each feature component in the low-dimensional space, highlighting the feature components related to operating conditions and control. The mutual information method is used to calculate the degree of correlation between each feature and the system operating state. Redundant and irrelevant features are eliminated according to the degree of correlation. The remaining effective features are integrated and sorted to form the target feature vector.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the step of using the UMAP uniform manifold approximation and projection algorithm to perform dimensionality reduction on the high-dimensional feature vector to obtain feature components in the low-dimensional space includes: Initialize the parameters of the UMAP algorithm, and determine the number of nearest neighbors, the distance metric, and the target low-dimensional space dimension; Calculate the distance between each high-dimensional feature vector and other vectors, select the nearest neighbors of each vector, establish a nearest neighbor graph in high-dimensional space, and capture the local structural features of high-dimensional data; Based on the constructed nearest neighbor graph, the local connectivity weight of each high-dimensional feature vector is calculated, and different weight values ​​are assigned according to the distance between the vectors. Initialize the coordinates of the feature vectors in the low-dimensional space, and randomly assign an initial coordinate in the low-dimensional space to each high-dimensional feature vector. Using the connectivity weights of the high-dimensional nearest neighbor graph as constraints, the coordinates of each feature vector in the low-dimensional space are iteratively adjusted. When the adjustment range of the feature vector coordinates in the low-dimensional space is less than a preset threshold, the iterative optimization stops. The low-dimensional space feature vectors after iterative optimization are processed to remove outlier coordinate points, thus obtaining the feature components in the low-dimensional space.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the step of inputting the target feature vector into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system and dynamically optimizing the control parameters based on the DRL algorithm includes: The target feature vector is input into the CNN-LSTM fusion model, and the CNN module extracts spatial dimension features from the target feature vector to capture the static correlation information between different features. The features extracted by the CNN module are input into the LSTM module network to perform temporal dimension feature mining on the feature sequence, obtain the dynamic pattern of feature changes over time, obtain spatiotemporal features, and identify the real-time operating condition type of the power generation control system through fully connected layer mapping. A control parameter optimization environment is constructed based on the DRL algorithm. The real-time operating conditions identified are used as the environmental state input. The control parameters are used as the action space to continuously interact with the power generation control system environment and update the strategy network parameters. Based on the output of the strategy network, the control parameters of the power generation control system are dynamically adjusted.

[0011] A second aspect of the present invention provides a collaborative optimization data processing system for a power generation control system based on a large model, the system comprising: The data processing module is used to collect multi-source heterogeneous data from the power generation control system, perform outlier detection and standardization on the multi-source heterogeneous data, and then obtain a standardized data set through the Min-Max normalization algorithm. The model training module is used to construct a knowledge graph of the power generation industry and perform entity recognition and linking. It integrates the standardized dataset with the knowledge graph of the power generation industry for training to obtain a large power generation model. The feature selection module is used to extract high-dimensional feature vectors using the large-scale power generation model, perform dimensionality reduction using the UMAP algorithm, and select target feature vectors by combining multi-head attention mechanism and mutual information method. The parameter optimization module is used to input the target feature vector into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and dynamically optimize the control parameters based on the DRL algorithm.

[0012] A third aspect of the present invention provides a data processing device for collaborative optimization of a power generation control system based on a large model. The data processing device includes a memory and at least one processor. The memory stores instructions. The at least one processor invokes the instructions in the memory to cause the data processing device to perform the steps of the data processing method for collaborative optimization of a power generation control system based on a large model as described in any of the preceding claims.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the collaborative optimization data processing method for a power generation control system based on a large model as described in any of the preceding claims.

[0014] The technical solution provided by this invention involves collecting multi-source heterogeneous data from a power generation control system, performing outlier detection and standardization on the multi-source heterogeneous data, and then obtaining a standardized data set through a Min-Max normalization algorithm. A power generation industry knowledge graph is constructed, and entity recognition and linking are performed. The standardized data set and the power generation industry knowledge graph are fused and trained to obtain a large-scale power generation model. High-dimensional feature vectors are extracted using the large-scale power generation model, and dimensionality reduction is performed using the UMAP algorithm. Combined with a multi-head attention mechanism and mutual information method, target feature vectors are selected. The target feature vectors are input into a CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and control parameters are dynamically optimized based on the DRL algorithm. This invention achieves efficient standardization processing, accurate feature selection, and collaborative dynamic optimization decision-making for multi-source heterogeneous data, significantly improving the operating efficiency, safety, stability, and economy of the power generation control system. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart illustrating the collaborative optimization data processing method for a power generation control system based on a large model, provided in an embodiment of the present invention. Figure 2 A schematic diagram of the structure of a collaborative optimization data processing system for a power generation control system based on a large model, provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a collaborative optimization data processing device for a power generation control system based on a large model, provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the collaborative optimization data processing method for a power generation control system based on a large model provided in this embodiment of the invention includes the following steps: Step 101: Collect multi-source heterogeneous data from the power generation control system, perform outlier detection and standardization on the multi-source heterogeneous data, and then obtain a standardized data set through the Min-Max normalization algorithm. In this embodiment, multi-source heterogeneous data from the power generation control system are collected, and structured operating parameters, unstructured equipment logs, and semi-structured scheduling instructions are aggregated to form an original dataset to be processed. The data space in the original dataset to be processed is randomly and recursively divided, and the isolated path length of each data point is calculated. Based on the path length, abnormal data in the original dataset to be processed is identified and removed. The numerical data after removing outliers is scaled, and the unstructured text data after removing outliers is converted into numerical features through word frequency statistics and feature quantization. The semi-structured data after removing outliers is extracted with effective numerical information through hierarchical parsing. The standardized data is input into the Min-Max normalization algorithm, and all data are scaled according to a unified interval to obtain a standardized dataset.

[0019] Step 102: Construct a knowledge graph for the power generation industry, perform entity recognition and linking, and integrate the standardized dataset with the knowledge graph for power generation to obtain a large-scale power generation model. In this embodiment, data on equipment, operating conditions, control parameters, and fault types within the power generation field are acquired to construct a power generation industry knowledge graph framework comprising entity, relation, and attribute layers. Entity extraction is performed on text and data from the power generation field to identify the entity information required for the power generation industry knowledge graph framework. The identified entities are vector-encoded, and entity alignment and linking are performed through similarity matching to obtain the power generation industry knowledge graph. A standardized dataset is used as training input, and entity and relation information from the power generation industry knowledge graph are integrated into the model training process as prompts. The LoRA low-rank adaptation algorithm is used to perform lightweight fine-tuning of the basic large model, enabling the model to learn the patterns of power generation industry data and the logical connections of the knowledge graph. After training, a large power generation model is obtained.

[0020] In this embodiment, the knowledge graph of the power generation industry is parsed to extract all entity information and the relationships between entities. Entity names, attributes, and relationship types are converted into prompt statements to form a set of prompt information. The set of prompt information is then matched with a standardized data set, with each standardized data point corresponding to a knowledge graph prompt, which is used as the training input for the basic large-scale model. The parameters of the basic large-scale model are initialized, and a low-rank adaptation matrix is ​​inserted into the Transformer layer of the basic large-scale model. The associated training input data is then input into the basic large-scale model, and only the parameters of the low-rank matrix inserted by LoRA are updated during training. The model is continuously fine-tuned according to a preset training cycle, and the power generation large-scale model is obtained after training stops.

[0021] A deep analysis of the power generation industry knowledge graph is performed to comprehensively extract all entity information and relationships between entities. Entity names, attributes, and relationship types are meticulously analyzed and transformed into natural language prompts, forming a set of prompt information. For example, if the entity is "generator," the attribute is "power," and the relationship is "connected to the grid," it can be transformed into the prompt "The generator has a power attribute and is connected to the grid." The set of prompt information is then precisely matched with a standardized data set. This ensures that each standardized data point corresponds to a relevant knowledge graph prompt, and this matched data serves as the training input for the basic large-scale model. For instance, a standardized data point about generator operating parameters, linked to relevant generator knowledge graph prompts, is input into the model. A low-rank adaptation matrix is ​​inserted into the Transformer layer of the basic large-scale model. The introduction of the low-rank adaptation matrix enables lightweight fine-tuning, reduces the number of training parameters, and improves training efficiency. The matched training input data is then fed into the basic large-scale model for training.

[0022] During training, only the parameters of the low-rank matrix inserted by LoRA are updated, while other parameters of the basic large model remain unchanged. Through continuous iterative training, the model gradually learns the data patterns of the power generation industry and the logical connections of the knowledge graph. The model is continuously fine-tuned according to the preset training rounds. When the preset number of rounds is reached or the model performance meets the requirements, training stops, and the resulting model is the large-scale power generation model.

[0023] Training parameters: Learning rate: set to 0.001. This parameter controls the step size of model parameter updates. A smaller learning rate makes model training more stable and avoids excessive parameter updates that could lead to model non-convergence. Batch size: set to 32, which is the number of data samples input to the model for each training iteration. An appropriate batch size can balance training efficiency and memory usage. Training epochs: set to 100 epochs based on model convergence and performance to ensure the model fully learns data features and knowledge graph information. LoRA rank: selected as 8. The rank of a low-rank matrix affects the flexibility of model fine-tuning and the number of parameters. An appropriate rank can achieve lightweight fine-tuning while ensuring model performance.

[0024] Step 103: Extract high-dimensional feature vectors using the large-scale power generation model, perform dimensionality reduction using the UMAP algorithm, and combine multi-head attention mechanism and mutual information method to select target feature vectors. In this embodiment, a standardized dataset is input into the trained large-scale power generation model. The model's feature encoding layer performs global feature extraction to generate a high-dimensional feature vector. The UMAP uniform manifold approximation and projection algorithm is used to perform dimensionality reduction on the high-dimensional feature vector to obtain feature components in the low-dimensional space. A multi-head attention mechanism is used to calculate attention weights for each feature component in the low-dimensional space, highlighting feature components related to operating conditions and control. The mutual information method is used to calculate the correlation between each feature and the system's operating state. Redundant and irrelevant features are eliminated based on the correlation degree, and the remaining effective features are integrated and sorted to form the target feature vector.

[0025] In this embodiment, the parameters of the UMAP algorithm are initialized, determining the number of nearest neighbors, the distance metric, and the target low-dimensional space dimension. The distance between each high-dimensional feature vector and other vectors is calculated, and the nearest neighbors of each vector are selected to establish a nearest neighbor graph in the high-dimensional space, capturing the local structural features of the high-dimensional data. Based on the constructed nearest neighbor graph, the local connectivity weight of each high-dimensional feature vector is calculated, and different weight values ​​are assigned according to the distance between vectors. The coordinates of the feature vectors in the low-dimensional space are initialized, and an initial coordinate in the low-dimensional space is randomly assigned to each high-dimensional feature vector. Using the connectivity weight of the nearest neighbor graph in the high-dimensional space as a constraint, the coordinates of each feature vector in the low-dimensional space are iteratively adjusted. When the adjustment range of the feature vector coordinates in the low-dimensional space is less than a preset threshold, the iterative optimization stops. The low-dimensional space feature vectors after iterative optimization are sorted out, and abnormal coordinate points are removed to obtain the feature components in the low-dimensional space.

[0026] Step 104: Input the target feature vector into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and dynamically optimize the control parameters based on the DRL algorithm.

[0027] In this embodiment, the target feature vector is input into the CNN-LSTM fusion model. The CNN module extracts spatial dimension features from the target feature vector to capture the static correlation information between different features. The features extracted by the CNN module are then input into the LSTM module network to mine temporal dimension features of the feature sequence, obtain the dynamic pattern of feature changes over time, and obtain spatiotemporal features. The real-time operating condition type of the power generation control system is identified through mapping by a fully connected layer. A control parameter optimization environment is constructed based on the DRL algorithm. The identified real-time operating condition is used as the input environment state, and the control parameters are used as the action space to continuously interact with the power generation control system environment and update the strategy network parameters. The control parameters of the power generation control system are dynamically adjusted according to the output results of the strategy network.

[0028] Spatial dimensional features are extracted from the target feature vector using a CNN module. After the target feature vector is input into the CNN module, it first passes through a convolutional layer. The convolutional layer uses multiple convolutional kernels of different sizes to perform sliding convolution operations on the feature vector. Each convolutional kernel acts as a feature detector, capturing spatial features of different local regions within the target feature vector. For example, for target feature vectors related to equipment operating parameters in a power generation control system, the convolutional kernel can detect local correlation patterns between different parameters, such as the coordinated changes in voltage and current within a certain range. Through parallel computation of multiple convolutional kernels, rich spatial dimensional features can be extracted from different angles, generating multiple feature maps. Subsequently, the feature maps enter the pooling layer. The pooling layer uses max pooling or average pooling operations to downsample the feature maps. Max pooling selects the maximum value within each pooling window as the output, while average pooling calculates the average value within the window. Pooling operations not only reduce the amount of data and computational complexity but also enhance the model's robustness to changes in feature location, further highlighting important spatial feature information. Through multiple alternating operations of convolutional and pooling layers, the CNN module can effectively capture the static correlation information between different features in the target feature vector.

[0029] The LSTM module performs temporal feature mining on feature sequences, inputting the features extracted by the CNN module into the LSTM network. The LSTM unit has a unique gating mechanism, including an input gate, a forget gate, and an output gate. The input gate determines which new information can enter the cell state, the forget gate controls which information in the cell state needs to be forgotten, and the output gate determines which information in the cell state will be used as the output at the current time step. When processing feature sequences, the LSTM module processes the features at each time step sequentially. For each time step, the LSTM unit updates the cell state and hidden state based on the current input features, the hidden state of the previous time step, and the cell state through the gating mechanism. In this way, the LSTM module can remember important information from historical time steps and dynamically adjust according to the current input, thereby mining the dynamic patterns of feature changes over time and obtaining spatiotemporal features that integrate spatial and temporal information.

[0030] The real-time operating condition type of the power generation control system is identified by mapping using a fully connected layer. The spatiotemporal features output from the LSTM module are input into the fully connected layer. Each neuron in the fully connected layer is connected to all neurons in the previous layer. Through weighted summation and activation function operations, the high-dimensional spatiotemporal features are mapped to a low-dimensional operating condition category space. After training, the fully connected layer can accurately identify the real-time operating condition type of the power generation control system based on the input spatiotemporal features, such as normal operation condition, fault warning condition, etc.

[0031] A control parameter optimization environment is constructed by inputting the identified real-time operating conditions as the environment state into the environment built by the DRL algorithm. Simultaneously, control parameters are defined as the action space, such as generator power regulation parameters and valve opening control parameters. Based on the current state and the executed actions, the environment performs state transitions according to preset physical rules and the system model, and returns corresponding reward signals. The reward signals are designed to guide the agent to learn the optimal control strategy; for example, a positive reward is given when adjusting control parameters makes the system more stable and efficient, and a negative reward is given otherwise.

[0032] The policy network parameters are updated by selecting actions to interact with the environment based on the current state, thereby obtaining new states and rewards. This process is repeated continuously to collect a large number of state-action-reward-new-state samples. Using these samples, a deep reinforcement learning algorithm (such as DDPG) is employed to update the policy network parameters. DDPG includes a policy network and a value network. The policy network is responsible for generating actions based on the state, and the value network is used to evaluate the value of state-action pairs. The gradient ascent method is used to update the policy network parameters, enabling the agent to obtain higher cumulative rewards when selecting actions.

[0033] The control parameters of the power generation control system are dynamically adjusted based on the updated strategy network output. The actions output by the strategy network, after appropriate transformation and mapping, directly act on the actuators of the power generation control system, enabling real-time adjustment of the control parameters. This allows the system to automatically optimize its operation based on real-time conditions, improving system stability and efficiency.

[0034] Please see Figure 2 A schematic diagram of the structure of a collaborative optimization data processing system for a power generation control system based on a large model provided in this embodiment of the invention. The system includes: The data processing module is used to collect multi-source heterogeneous data from the power generation control system, perform outlier detection and standardization on the multi-source heterogeneous data, and then obtain a standardized data set through the Min-Max normalization algorithm. The model training module is used to construct a knowledge graph of the power generation industry and perform entity recognition and linking. It integrates the standardized dataset with the knowledge graph of the power generation industry for training to obtain a large power generation model. The feature selection module is used to extract high-dimensional feature vectors using the large-scale power generation model, perform dimensionality reduction using the UMAP algorithm, and select target feature vectors by combining multi-head attention mechanism and mutual information method. The parameter optimization module is used to input the target feature vector into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and dynamically optimize the control parameters based on the DRL algorithm.

[0035] Figure 3 This is a schematic diagram of the structure of a large-model-based power generation control system collaborative optimization data processing device 300 provided in an embodiment of the present invention. This large-model-based power generation control system collaborative optimization data processing device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the large-model-based power generation control system collaborative optimization data processing device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the large-model-based power generation control system collaborative optimization data processing device 300 to implement the method provided in the above embodiment.

[0036] The collaborative optimization data processing device 300 for a large-scale power generation control system may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The structure of the collaborative optimization data processing device for the power generation control system based on the large model shown does not constitute a limitation on the computer device provided by the present invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0037] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the various steps of the collaborative optimization data processing method for a power generation control system based on a large model provided in the above embodiments.

[0038] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0039] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A collaborative optimization data processing method for a power generation control system based on a large model, characterized in that, The method includes the following steps: Collect multi-source heterogeneous data from the power generation control system, perform outlier detection and standardization on the multi-source heterogeneous data, and then obtain a standardized data set through the Min-Max normalization algorithm; A knowledge graph of the power generation industry is constructed, and entity recognition and linking are performed. The standardized dataset is integrated with the knowledge graph of the power generation industry for training to obtain a large power generation model. High-dimensional feature vectors are extracted using a large-scale power generation model, and dimensionality reduction is performed using the UMAP algorithm. The target feature vectors are then selected by combining multi-head attention mechanism and mutual information method. The target feature vector is input into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and the control parameters are dynamically optimized based on the DRL algorithm. The process involves extracting high-dimensional feature vectors using a large-scale power generation model, performing dimensionality reduction using the UMAP algorithm, and combining multi-head attention mechanism and mutual information method to select target feature vectors, including: The standardized dataset is input into the trained large-scale power generation model, and the feature encoding layer of the large-scale power generation model is used to extract global features from the data to generate high-dimensional feature vectors. The UMAP uniform manifold approximation and projection algorithm is used to perform dimensionality reduction on high-dimensional feature vectors to obtain feature components in low-dimensional space. By using a multi-head attention mechanism, attention weights are calculated for each feature component in the low-dimensional space, highlighting the feature components related to operating conditions and control. The mutual information method is used to calculate the degree of correlation between each feature and the system operating state. Redundant and irrelevant features are eliminated according to the degree of correlation. The remaining effective features are integrated and sorted to form the target feature vector. The method employs the UMAP uniform manifold approximation and projection algorithm to perform dimensionality reduction on high-dimensional feature vectors, obtaining feature components in the low-dimensional space, including: Initialize the parameters of the UMAP algorithm, and determine the number of nearest neighbors, the distance metric, and the target low-dimensional space dimension; Calculate the distance between each high-dimensional feature vector and other vectors, select the nearest neighbors of each vector, establish a nearest neighbor graph in high-dimensional space, and capture the local structural features of high-dimensional data; Based on the constructed nearest neighbor graph, the local connectivity weight of each high-dimensional feature vector is calculated, and different weight values ​​are assigned according to the distance between the vectors. Initialize the coordinates of the feature vectors in the low-dimensional space, and randomly assign an initial coordinate in the low-dimensional space to each high-dimensional feature vector. Using the connectivity weights of the high-dimensional nearest neighbor graph as constraints, the coordinates of each feature vector in the low-dimensional space are iteratively adjusted. When the adjustment range of the feature vector coordinates in the low-dimensional space is less than a preset threshold, the iterative optimization stops. The low-dimensional space feature vectors after iterative optimization are sorted out, and outlier coordinate points are removed to obtain the feature components in the low-dimensional space. The step of inputting the target feature vector into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and dynamically optimizing the control parameters based on the DRL algorithm, includes: The target feature vector is input into the CNN-LSTM fusion model, and the CNN module extracts spatial dimension features from the target feature vector to capture the static correlation information between different features. The features extracted by the CNN module are input into the LSTM module network to perform temporal dimension feature mining on the feature sequence, obtain the dynamic pattern of feature changes over time, obtain spatiotemporal features, and identify the real-time operating condition type of the power generation control system through fully connected layer mapping. A control parameter optimization environment is constructed based on the DRL algorithm. The real-time operating conditions identified are used as the environmental state input. The control parameters are used as the action space to continuously interact with the power generation control system environment and update the strategy network parameters. Based on the output of the strategy network, the control parameters of the power generation control system are dynamically adjusted.

2. The collaborative optimization data processing method for a power generation control system based on a large model as described in claim 1, characterized in that, The system collects multi-source heterogeneous data from the power generation and control system, performs outlier detection and standardization on the multi-source heterogeneous data, and then obtains a standardized data set through the Min-Max normalization algorithm, including: Collect multi-source heterogeneous data within the power generation control system, and aggregate structured operating parameters, unstructured equipment logs, and semi-structured scheduling instructions to form a raw dataset to be processed; The data space in the original dataset to be processed is randomly and recursively divided, the isolated path length of each data point is calculated, and abnormal data in the original dataset to be processed is determined and removed based on the path length. The numerical data after outlier removal is scaled; the unstructured text data after outlier removal is converted into numerical features through word frequency statistics and feature quantization; and the semi-structured data after outlier removal is extracted with effective numerical information through hierarchical parsing. The standardized data is input into the Min-Max normalization algorithm, and all data are scaled according to a unified interval to obtain a standardized dataset.

3. The collaborative optimization data processing method for a power generation control system based on a large model as described in claim 1, characterized in that, The process involves constructing a knowledge graph for the power generation industry, performing entity recognition and linking, and then fusing a standardized dataset with the power generation industry knowledge graph for training to obtain a large-scale power generation model, including: Acquire data on equipment, operating conditions, control parameters, and fault types in the power generation field, and build a power generation industry knowledge graph framework that includes an entity layer, a relationship layer, and an attribute layer. Entity extraction is performed on text and data in the power generation field to identify the entity information required for the power generation industry knowledge graph framework. The identified entities are vector-encoded, and entity alignment and linking are performed through similarity matching to obtain a knowledge graph of the power generation industry. Standardized datasets are used as training inputs, and entity and relationship information from the power generation industry knowledge graph is incorporated into the model training process as prompts. The LoRA low-rank adaptation algorithm is used to perform lightweight fine-tuning on the basic large model, enabling the model to learn the data patterns of the power generation industry and the logical connections with the knowledge graph. After training, a large power generation model is obtained.

4. The collaborative optimization data processing method for a power generation control system based on a large model as described in claim 3, characterized in that, The process involves using standardized datasets as training input, incorporating entity and relationship information from the power generation industry knowledge graph as prompts into the model training process, and employing the LoRA low-rank adaptation algorithm for lightweight fine-tuning of the basic large model. This allows the model to learn the patterns in power generation industry data and the logical connections within the knowledge graph. Upon completion of training, a large power generation model is obtained, including: The knowledge graph of the power generation industry is analyzed to extract all entity information and the relationships between entities. Entity names, attributes and relationship types are converted into prompt statements to form a set of prompt information. The set of prompt information is matched with the standardized data set. Each piece of standardized data corresponds to the knowledge graph prompt information and is used as the training input for the basic large model. Initialize the parameters of the basic large model, and insert a low-rank adaptation matrix into the Transformer layer of the basic large model; The correlated training input data is fed into the base model, and only the parameters of the low-rank matrix inserted by LoRA are updated during the training process. The model is continuously fine-tuned according to the preset training rounds, and the large power generation model is obtained after training is stopped.

5. A collaborative optimization data processing system for a power generation control system based on a large model, characterized in that, The system includes: The data processing module is used to collect multi-source heterogeneous data from the power generation control system, perform outlier detection and standardization on the multi-source heterogeneous data, and then obtain a standardized data set through the Min-Max normalization algorithm. The model training module is used to construct a knowledge graph of the power generation industry and perform entity recognition and linking. It integrates the standardized dataset with the knowledge graph of the power generation industry for training to obtain a large power generation model. The feature selection module is used to extract high-dimensional feature vectors from a large-scale power generation model. It employs the UMAP algorithm for dimensionality reduction, combined with a multi-head attention mechanism and mutual information method, to select target feature vectors. The process involves: inputting a standardized dataset into the trained large-scale power generation model; performing global feature extraction through the model's feature encoding layer to generate high-dimensional feature vectors; using the UMAP uniform manifold approximation and projection algorithm to perform dimensionality reduction on the high-dimensional feature vectors, obtaining feature components in the low-dimensional space; calculating attention weights for each feature component in the low-dimensional space using the multi-head attention mechanism, highlighting features related to operating conditions and control; calculating the correlation between each feature and the system's operating state using the mutual information method; eliminating redundant and irrelevant features based on the correlation strength; integrating and sorting the remaining effective features to form the target feature vector; and initializing the parameters of the UMAP algorithm. The process involves several steps: First, determining the number of nearest neighbors, the distance metric, and the target low-dimensional space dimension. Then, calculating the distance between each high-dimensional feature vector and other vectors, selecting the nearest neighbors for each vector, and establishing a nearest neighbor graph in the high-dimensional space to capture the local structural features of the high-dimensional data. Based on the constructed nearest neighbor graph, calculating the local connectivity weight of each high-dimensional feature vector and assigning different weight values ​​according to the distance between vectors. Next, initializing the feature vector coordinates in the low-dimensional space by randomly assigning an initial coordinate in the low-dimensional space to each high-dimensional feature vector. Using the connectivity weight of the nearest neighbor graph in the high-dimensional space as a constraint, iteratively adjusting the coordinates of each feature vector in the low-dimensional space until the adjustment range is less than a preset threshold, at which point the iterative optimization stops. Finally, the iteratively optimized low-dimensional space feature vectors are processed to remove abnormal coordinate points, yielding the feature components in the low-dimensional space. The parameter optimization module is used to input the target feature vector into the CNN-LSTM fusion model to identify the real-time operating conditions of the power generation control system, and dynamically optimize the control parameters based on the DRL algorithm: The target feature vector is input into the CNN-LSTM fusion model, and the CNN module extracts spatial dimension features from the target feature vector to capture the static correlation information between different features; the features extracted by the CNN module are input into the LSTM module network, and the feature sequence is mined for temporal dimension features to obtain the dynamic pattern of feature changes over time, resulting in spatiotemporal features. The real-time operating condition type of the power generation control system is identified through mapping using a fully connected layer; a control parameter optimization environment is constructed based on the DRL algorithm, using the identified real-time operating conditions as the environmental state input, and continuously interacting with the power generation control system environment with the control parameters as the action space to update the strategy network parameters; the control parameters of the power generation control system are dynamically adjusted according to the output results of the strategy network.

6. A collaborative optimization data processing device for a power generation control system based on a large model, characterized in that, The large-model-based power generation control system collaborative optimization data processing device includes a memory and at least one processor. The memory stores instructions. The at least one processor calls the instructions in the memory to cause the large-model-based power generation control system collaborative optimization data processing device to execute each step of the large-model-based power generation control system collaborative optimization data processing method as described in any one of claims 1-4.

7. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement each step of the collaborative optimization data processing method for a power generation control system based on a large model as described in any one of claims 1-4.

Images