Multi-modal data fusion hydroelectric generating set parameter intelligent optimization method and system

Abstract

The invention relates to the technical field of hydroelectric generating set parameter optimization, in particular to an intelligent hydroelectric generating set parameter optimization method and system based on multi-modal data fusion, and the method integrates multi-modal data, a graph neural network and Bayesian optimization to realize efficient and intelligent hydroelectric generating set parameter optimization. Time domain, frequency domain, acoustics, temperature and hydraulic multi-modal data are integrated, and feature vectors are generated through intra-modal feature extraction and hierarchical attention mechanism fusion; thirdly, constructing a unit knowledge graph, and extracting priori knowledge by using a graph attention network to realize cross-domain knowledge migration; and finally, on the basis of a multi-objective constrained Bayesian optimization algorithm, in combination with a Gaussian process agent model, determining optimal PID parameter configuration, and performing incremental optimization through closed-loop verification. According to the method, the optimization time is shortened from several days to 15 minutes or less, the efficiency is improved by 95% or above, and the parameter optimization efficiency and precision of the hydroelectric generating set are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of parameter optimization of hydroelectric generating units, and particularly to an intelligent optimization method and system for parameters of hydroelectric generating units based on multimodal data fusion. Background Art

[0002] The optimization of power mode parameters of hydroelectric generating units is of great significance in ensuring the stable operation of the units and improving efficiency. In the prior art, for example, Chinese Patent CN 112628055 B discloses a method for on-site optimization of power mode parameters of hydroelectric generating units. This method adopts a power mode PID operation parameter optimization method based on 6 characteristic parameters, and optimizes the operation parameters of the units by judging whether the adjustment performance of the parameters is excellent.

[0003] However, the prior art has the following deficiencies: First, the data source is single, only using traditional data such as active power and guide vane opening, ignoring important multimodal data such as vibration spectrum, noise characteristics, and efficiency curve, which limits the ability to perceive the overall state of the units; Second, it adopts a mechanical parameter adjustment logic, using fixed rules for parameter adjustment, lacking the ability of adaptive learning for complex nonlinear systems; Third, there is a knowledge island effect in the optimization process, with each unit optimized independently, unable to utilize the knowledge correlation and historical optimization experience between units; In addition, the prior art adopts a static optimization idea, optimizing once and using it for a long time, making it difficult to cope with the dynamic changes of working conditions such as water head and load; Finally, the existing methods have a high degree of manual dependence, requiring professional technical personnel to conduct on-site tests and parameter adjustments, which is time-consuming and the effect depends on personnel experience. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an intelligent optimization method and system for parameters of hydroelectric generating units based on multimodal data fusion. This method realizes the intelligent and efficient optimization of parameters of hydroelectric generating units by fusing multimodal data, using a graph neural network to achieve knowledge transfer, and adopting a Bayesian optimization algorithm for parameter optimization.

[0005] The present invention proposes an intelligent optimization method for parameters of hydroelectric generating units based on multimodal data fusion, including: Obtaining multimodal data of the hydroelectric generating unit, where the multimodal data includes time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data; Performing feature extraction and fusion on the multimodal data, including: Using an intra-modal feature extraction network to process various types of data respectively to obtain intra-modal features; Calculating the correlation weights between different modal features based on a hierarchical attention mechanism to generate a fused feature vector; Constructing a unit knowledge graph and realizing cross-domain knowledge transfer based on a graph neural network, including: Construct a unit knowledge graph based on the fused feature vector and historical optimization experience; Use a graph attention network to extract prior knowledge representations adapted to the current unit characteristics; Perform parameter optimization based on a Bayesian optimization algorithm with multi-objective constraints, including: Construct a Gaussian process surrogate model according to the prior knowledge representation; Determine the optimal PID parameter configuration using a multi-objective acquisition function; Perform incremental optimization on the optimized parameters through closed-loop verification to obtain the final optimized parameters.

[0006] Preferably, the acquisition of multi-modal data of the hydropower unit includes: Deploy a distributed multi-modal sensing data acquisition network; Implement synchronous acquisition of multi-source data based on timestamps, ensuring that the maximum time difference does not exceed 5 milliseconds; Perform outlier detection and data standardization processing on the acquired multi-modal data; Generate a multi-modal data tensor in a unified format.

[0007] Preferably, the use of the intra-modal feature extraction network to process various types of data includes: Use a TCN network based on causal convolution to extract time-domain data features; Use an improved ResNet network to extract frequency-domain vibration data features; Use Mel spectrum transformation and CNN to extract acoustic data features; Use a spatial convolution network to extract temperature distribution data features; Use a deep fully connected network to extract hydraulic characteristic data features.

[0008] Preferably, the calculation of the correlation weights between different modal features based on the hierarchical attention mechanism includes: Calculate the attention weight matrix between different modal features: , where: is the feature vector of the th modality, is the feature vector of the th modality, is the similarity calculation function; Enhance the original features based on the attention weight matrix: , Integrate different scale features through a feature pyramid fusion network to generate a final fused feature vector.

[0009] Preferably, the unit knowledge graph includes: Unit nodes, representing the characteristic parameters of the hydropower unit; Operating condition nodes, representing the operating parameters of the unit; Parameter nodes, representing PID control parameters; Physical association edges, representing the hydraulic coupling relationship between units; Performance similarity edges, representing the similarity degree of operating characteristics between units; Historical optimization edges, representing the transfer relationship of historical parameter optimization experience.

[0010] Preferably, the extraction of prior knowledge representations adapted to the current unit characteristics using the graph attention network includes: Implementing a node-level attention mechanism to calculate the importance weights between nodes; Designing an edge-level message passing mechanism to realize the transfer of knowledge in the graph structure; Implementing knowledge transfer across units and heads through domain alignment techniques based on maximum mean discrepancy; Generating prior knowledge representations adapted to the current unit characteristics.

[0011] Preferably, the multi-objective constrained Bayesian optimization algorithm includes: Constructing a Gaussian process surrogate model and adopting a multi-kernel function combination strategy: , where: is the basic kernel function, is the weight coefficient; Designing a multi-objective acquisition function to balance multiple optimization objectives such as response time, overshoot, and stability; Implementing a dynamic optimization strategy with operating condition perception to automatically adjust the exploration intensity according to the head change; Developing a parameter sensitivity analysis module to identify key parameters and optimize them preferentially.

[0012] Preferably, the incremental optimization of the optimized parameters through closed-loop verification includes: Constructing a refined digital twin model of the unit to simulate and verify the parameter optimization effect; Designing an extreme condition test framework to evaluate the performance of parameters under boundary conditions; Implementing incremental learning based on experience replay to continuously optimize the model performance; Constructing a parameter degradation detection model to identify the trend of parameter performance decline; Developing a self-triggered optimization mechanism to automatically start re-optimization when the performance decline exceeds a preset threshold.

[0013] Preferably, the method further includes: Designing a plant-level collaborative optimization interface to realize the collaborative adjustment of unit group parameters; Develop a knowledge sharing framework based on federated learning to protect data privacy; Implement cross-power station knowledge transfer to support the rapid commissioning of newly built power stations.

[0014] An intelligent optimization system for hydro-generator unit parameters with multi-modal data fusion, including: A multi-modal data acquisition module for acquiring multi-modal data of hydro-generator units, where the multi-modal data includes time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data; A feature extraction and fusion module for extracting and fusing features from the multi-modal data to generate a fused feature vector; A knowledge graph construction and transfer learning module for constructing a unit knowledge graph and implementing cross-domain knowledge transfer based on a graph neural network to generate a prior knowledge representation; A parameter optimization module for parameter optimization based on a multi-objective constrained Bayesian optimization algorithm to determine the optimal PID parameter configuration; A closed-loop verification and incremental optimization module for incrementally optimizing the optimized parameters through closed-loop verification to obtain the final optimized parameters; A collaborative optimization interface module for realizing the collaborative adjustment of unit group parameters and cross-power station knowledge transfer.

[0015] The present invention has the following beneficial effects: 1. The optimization efficiency is significantly improved: Compared with traditional methods, the optimization time is shortened from several days to within 15 minutes, the efficiency is increased by more than 95%, and the success rate of the first optimization reaches 95%, which is much higher than the 60% - 70% of traditional methods.

[0016] 2. The adaptability to all working conditions is greatly improved: The parameters can adapt within the range of ±30 meters of water head change without manual intervention; it can operate stably in the full range of 10% - 100% load change, solving the problem of instability of traditional methods under load boundary conditions.

[0017] 3. The economic benefits are significant: The unit efficiency is increased by 4.2%, and the annual increased income is considerable; the vibration is reduced by 37%, and the service life of the equipment is extended by 3 - 5 years; the number of shutdown adjustments is reduced by 85%, and the equipment utilization rate is improved.

[0018] 4. The degree of intelligence is high: It realizes fully automatic parameter optimization without the need for on-site intervention by professionals; it supports remote monitoring and optimization, reducing the operation and maintenance cost; it has the ability of self-learning, and the optimization effect continuously improves with the use time. Description of the Drawings

[0019] Figure 1 It is a flow chart of the intelligent optimization method for hydro-generator unit parameters with multi-modal data fusion of the present invention; Figure 2This is the multi-modal data acquisition network architecture diagram of the present invention; Figure 3 This is the structure diagram of the heterogeneous feature extraction and fusion network with hierarchical attention mechanism of the present invention; Figure 4 This is the schematic diagram of the unit knowledge graph construction of the present invention; Figure 5 This is the schematic diagram of the cross-domain transfer learning mechanism enhanced by graph neural network of the present invention; Figure 6 This is the flow chart of the multi-objective constrained dynamic adaptive Bayesian optimization of the present invention; Figure 7 This is the schematic diagram of the closed-loop verification and incremental optimization system of the present invention; Figure 8 This is the architecture diagram of the intelligent optimization system for the parameters of hydropower units with multi-modal data fusion of the present invention. Detailed implementation manners

[0020] Please refer to the attached Figure 1-8 drawing, and the following further describes the detailed implementation manners of the present invention in conjunction with the drawings.

[0021] Example 1: Overall process of the intelligent optimization method for the parameters of hydropower units with multi-modal data fusion As Figure 1 shown, the intelligent optimization method for the parameters of hydropower units with multi-modal data fusion provided by the present invention includes the following steps: Obtain the multi-modal data of the hydropower unit, where the multi-modal data includes time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data; Perform feature extraction and fusion on the multi-modal data, including using the intra-modal feature extraction network to process various types of data respectively to obtain intra-modal features; calculating the correlation weights between different modal features based on the hierarchical attention mechanism to generate a fused feature vector; Construct a unit knowledge graph and implement cross-domain knowledge transfer based on the graph neural network, including constructing a unit knowledge graph based on the fused feature vector and historical optimization experience; using the graph attention network to extract the prior knowledge representation suitable for the current unit characteristics; Perform parameter optimization based on the multi-objective constrained Bayesian optimization algorithm, including constructing a Gaussian process surrogate model according to the prior knowledge representation; determining the optimal PID parameter configuration using a multi-objective acquisition function; Perform incremental optimization on the optimized parameters through closed-loop verification to obtain the final optimized parameters.

[0022] The overall process of this method follows the processing framework of data acquisition - feature extraction - knowledge transfer - parameter optimization - closed-loop verification. A complete data flow link is formed between each step, realizing the intelligent and efficient optimization of the parameters of hydropower units.

[0023] Preferably, the present method can be applied to the parameter optimization of various types of hydropower units, including but not limited to different types of units such as impulse turbines, Francis turbines, and Kaplan turbines, and is applicable to hydropower stations under different head conditions and load demands.

[0024] Embodiment 2: Multimodal data acquisition method As Figure 2 shown, this embodiment details the specific implementation method of multimodal data acquisition, mainly including: Deploy a distributed multimodal sensing data acquisition network, and install various sensors at key parts of the hydropower unit, including power sensors, vibration sensors, acoustic sensors, temperature sensors, water pressure sensors, etc., to form a sensing network covering the entire unit.

[0025] Implement multi-source data synchronous acquisition based on timestamps, ensuring that the maximum time difference does not exceed 5 milliseconds. Preferably, use industrial Ethernet technology to build the sensing network, support Time-Sensitive Networking (TSN) technology, and ensure the time synchronization of data acquisition.

[0026] Perform outlier detection and data normalization processing on the acquired multimodal data. Preferably, use the Local Outlier Factor (LOF) algorithm to identify outlier data points, set a dynamic threshold, and automatically mark and remove or correct outlier points when the outlier degree exceeds the threshold. Design dedicated normalization methods for different physical quantities to retain sensitive features.

[0027] Generate multimodal data tensors in a unified format. Preferably, convert different types of data into a multi-dimensional tensor format, ensure the alignment of the time dimension, and provide a standardized input for subsequent feature extraction.

[0028] The multimodal data obtained by this method includes: Time-domain operation data: including active power, guide vane opening, water pressure at the inlet of the spiral case, unit frequency, etc., with a sampling rate of 100 Hz; Frequency-domain vibration data: including radial vibration and axial vibration spectra, with a frequency range of 0 - 1000 Hz and an accuracy of 0.1 Hz; Acoustic feature data: including noise spectra and sound intensity distributions, with a frequency range of 20 - 20000 Hz; Temperature distribution data: including bearing temperature and stator temperature distribution, with an accuracy of 0.1 °C; Hydraulic characteristic data: including cavitation characteristics, efficiency curves, head change rates, etc.

[0029] Through the acquisition and preprocessing of multimodal data, it provides a high-quality and multi-dimensional data basis for subsequent feature extraction and fusion, overcoming the limitation of the single data source in traditional methods.

[0030] Example 3: Multimodal Data Feature Extraction Method As Figure 3 shown, this example details the feature extraction method for multimodal data, specifically including: Use a TCN network based on causal convolution to extract time-domain data features. Preferably, the TCN network adopts a dilated convolution structure, and the convolution kernel size is set to [3, 5, 7], which can effectively capture features at different time scales. Causal convolution ensures that the output at the current moment only depends on historical data, avoiding information leakage. At the same time, residual connections are used to improve gradient propagation and enhance the effectiveness of feature extraction.

[0031] Use an improved ResNet network to extract frequency-domain vibration data features. Preferably, the improved ResNet network introduces an attention mechanism to enhance the ability to extract key frequency components. The network depth is 18 layers, and each layer of convolution is followed by BatchNormalization and the ReLU activation function to effectively extract frequency-domain vibration features.

[0032] Use Mel spectral transformation and CNN to extract acoustic data features. Preferably, first perform Mel spectral transformation on the acoustic signal to map linear frequencies to Mel frequencies that are more in line with human ear perception, and then use a 3-layer CNN network to extract acoustic features. The convolution kernel sizes are 5×5, 3×3, and 3×3 respectively, and the activation function is LeakyReLU.

[0033] Use a spatial convolution network to extract temperature distribution data features. Preferably, regard the temperature distribution data as a two-dimensional image and use a 3-layer spatial convolution network to extract spatial distribution features to effectively capture temperature gradients and distribution patterns.

[0034] Use a deep fully connected network to extract hydraulic characteristic data features. Preferably, design a 3-layer fully connected network with the number of hidden layer nodes being 128, 64, and 32 respectively, and the activation function being ReLU to extract the non-linear relationship features of hydraulic characteristics.

[0035] Through intra-modal feature extraction, various types of data are converted into feature vectors of a fixed dimension, providing a standardized input for subsequent feature fusion. Intra-modal feature extraction makes full use of the characteristics of various types of data and selects a network structure suitable for its characteristics to ensure the effectiveness and accuracy of feature extraction.

[0036] Example 4: Feature Fusion Method Based on Hierarchical Attention Mechanism This example details the method for calculating the correlation weights between different modal features based on the hierarchical attention mechanism, which is the core of feature fusion, specifically including: Calculate the attention weight matrix between different modal features: , Wherein: is the eigenvector of the th mode, is the eigenvector of the th mode, is a similarity calculation function, and the scaled dot - product attention mechanism can be adopted: , and are learnable transformation matrices, is the feature dimension.

[0037] Enhance the original features based on the attention weight matrix: , Integrate features of different scales through the feature pyramid fusion network to generate the final fused feature vector. Preferably, the feature pyramid network adopts a combination of top - down and lateral connections to construct a multi - scale feature fusion structure. The top - down is achieved through upsampling, and the lateral connection is achieved after adjusting the number of channels through 1×1 convolution.

[0038] In addition, introduce a channel attention mechanism to adaptively adjust the weights of different feature channels: , where: AvgPool and MaxPool are average pooling and max - pooling operations respectively, MLP is a multi - layer perceptron, is the sigmoid activation function. Channel attention can dynamically adjust the weights according to the importance of different feature channels.

[0039] Finally, develop a feature redundancy suppression module to reduce the dimension through feature distillation technology. Preferably, an auto - encoder structure is adopted to achieve feature compression. Both the encoder and the decoder are 3 - layer fully - connected networks, and learn a compact representation by minimizing the reconstruction error.

[0040] Through the feature fusion of the hierarchical attention mechanism, this method can effectively integrate the complementary information in multi - modal data, generate a compact and information - rich fused feature vector, and provide a high - quality feature representation for subsequent knowledge transfer and parameter optimization.

[0041] Example 5: Method for constructing a unit knowledge graph As Figure 4 shown, this example details the method for constructing a unit knowledge graph. The unit knowledge graph is the basis for realizing cross - domain knowledge transfer, and specifically includes: Construct unit nodes to represent the characteristic parameters of the hydropower unit. Preferably, the attributes included in the unit nodes are: basic parameters such as unit type, rated capacity, design head, rotational speed, number of blades, etc., and statistical characteristics of historical operation data, such as average load rate, start - stop frequency, vibration intensity, etc.

[0042] Construct an operating condition node, which represents the operating parameters of the unit. Preferably, the attributes included in the operating condition node are: head range, load range, water temperature, characteristics of the water diversion system, etc., as well as the current operating state parameters, such as real-time head, load, efficiency, etc.

[0043] Construct a parameter node, which represents the PID control parameters. Preferably, the attributes included in the parameter node are: PID parameter combination, applicable operating condition range of the parameters, parameter performance evaluation indicators (response time, overshoot, stability, etc.), parameter acquisition time, etc.

[0044] Construct a physical association edge, which represents the hydraulic coupling relationship between units. Preferably, the attributes included in the physical association edge are: common system type (tail water system, water diversion system, etc.), coupling strength coefficient, relative position relationship, etc.

[0045] Construct a performance similarity edge, which represents the similarity degree of the operating characteristics between units. Preferably, the performance similarity edge is obtained by calculating the cosine similarity of the operating characteristic vectors of the units, and the threshold is set to 0.85. An edge connection is established between the pairs of units that exceed the threshold.

[0046] Construct a historical optimization edge, which represents the transfer relationship of historical parameter optimization experiences. Preferably, the historical optimization edge records the transfer history of parameter optimization, including information such as the source unit, target unit, transfer time, and change of optimized parameters.

[0047] The unit knowledge graph adopts a dynamic update mechanism, and updates the graph structure in real time according to the newly added optimization experiences, and the weights are adaptively adjusted. Preferably, after each successful parameter optimization, the optimization experience is added to the knowledge graph, and at the same time, the weights of the relevant edges are updated according to the optimization effect to ensure that the knowledge graph always reflects the latest optimization experiences.

[0048] By constructing a multi-level and multi-relationship unit knowledge graph, this method can effectively capture various association relationships between units and provide a structured knowledge representation for subsequent knowledge transfer based on graph neural networks.

[0049] Example 6: Knowledge Transfer Method Based on Graph Neural Networks As Figure 5 shown, this example details the knowledge transfer method based on graph neural networks, which specifically includes: Implement a node-level attention mechanism to calculate the importance weights between nodes: , Among them: and are the feature representations of node i and the node respectively, W is the weight matrix, is the attention vector, || represents the vector concatenation operation, For a node is the neighbor set of

[0050] Design an edge-level message passing mechanism to achieve the transfer of knowledge in the graph structure: , where: and represent the feature representations of node i at the l-th layer and the (l + 1)-th layer respectively, σ is the activation function, and preferably the ReLU function is adopted. Edge-level message passing updates the node representation iteratively, gradually integrating neighborhood information to achieve the propagation of knowledge in the graph structure.

[0051] Through the domain alignment technology of maximum mean discrepancy ( ), achieve knowledge transfer across units and heads: , where: and represent the data sets of the source domain and the target domain respectively, φ is the feature mapping function, and H is the reproducing kernel Hilbert space. By minimizing the MMD distance, the feature distributions of the source domain and the target domain tend to be consistent, thereby achieving cross-domain knowledge transfer.

[0052] In addition, design a progressive transfer strategy to retain the effective knowledge of the source domain through weight regularization. Preferably, L2 regularization is adopted to limit the change range of the weights during the transfer process, ensuring that the transferred model retains the knowledge of the source domain while adapting to the characteristics of the target domain.

[0053] Implement a parameter soft sharing mechanism to dynamically adjust the knowledge transfer intensity according to the similarity. Preferably, calculate the knowledge transfer coefficient λ according to the similarity between the source domain and the target domain. The higher the similarity, the larger the value of λ and the greater the knowledge transfer intensity. The knowledge transfer coefficient is used to adjust the sharing degree of the parameters of the source domain and the target domain: θt=(1 - λ)θt + λθs, where θt and θs are the parameters of the target domain and the source domain respectively.

[0054] Through the cross-domain transfer learning mechanism enhanced by graph neural networks, this method can effectively utilize historical optimization experience to achieve cross-unit and cross-condition knowledge transfer, provide high-quality prior knowledge for parameter optimization under new units or new conditions, and greatly improve the optimization efficiency and success rate.

[0055] Example 7: Bayesian optimization algorithm with multi-objective constraints As Figure 6 shown, this example details the Bayesian optimization algorithm with multi-objective constraints, specifically including: Construct a Gaussian process surrogate model and adopt a multi-kernel function combination strategy: , Wherein: is the basic kernel function, including radial basis function (RBF), Matérn kernel function, Periodic kernel function, etc. is the weight coefficient, which is automatically determined by maximizing the marginal likelihood. The multi-kernel function combination strategy can adapt to the characteristics of different parameter spaces and improve the expression ability and generalization performance of the model.

[0056] Design a multi-objective acquisition function to balance multiple optimization objectives such as response time, overshoot, and stability. Preferably, a combination of three acquisition functions, Expected Improvement (EI), Upper Confidence Bound (UCB), and Expected Entropy Reduction (EES), is adopted: , Wherein: 、 and are weight coefficients, which are dynamically adjusted according to the optimization stage. The weight of the exploration term is increased in the initial stage, and the weight of the exploitation term is increased in the later stage.

[0057] Introduce a prediction variance weighting mechanism: , Wherein: is the predicted mean, is the predicted standard deviation, is the exploration-exploitation balance coefficient, initially set to 2.0, and gradually reduced to 0.5 as the optimization progresses, to achieve a smooth transition from exploration to exploitation.

[0058] Implement a constraint handling technique to ensure that the parameters meet the safe operating boundary conditions. Preferably, the penalty function method is used to handle the constraints, and a penalty is given to the solutions that violate the constraints. The degree of penalty is proportional to the degree of violation, to ensure that the optimization results meet all the constraint conditions.

[0059] Design a dynamic exploration strategy based on Thompson sampling to automatically enhance the exploration intensity when the water head changes. Preferably, when it is detected that the water head change exceeds a predetermined threshold (default is 5 meters), the exploration coefficient is increased to strengthen the exploration of the parameter space and adapt to the new operating conditions.

[0060] Develop a parameter sensitivity analysis module to identify the key parameters and optimize them preferentially. Preferably, the sensitivity is estimated by calculating the partial derivative of the parameter with respect to the objective function, and the parameters are sorted according to the sensitivity, and the parameters with high sensitivity are optimized preferentially to improve the optimization efficiency.

[0061] Implement progressive search space reduction to improve the optimization efficiency. Preferably, in the initial stage, a global search is carried out in a large range, and as the optimization progresses, the search space is gradually reduced according to the historical evaluation results, and the resources are concentrated in the promising areas for fine search.

[0062] In addition, the idea of simulated annealing is introduced to prevent getting stuck in local optima; a safety exploration area definition based on model uncertainty is developed; and a perturbation analysis module is designed to evaluate the robustness of parameters to perturbations.

[0063] Through the multi-objective constrained dynamic adaptive Bayesian optimization algorithm, this method can efficiently and accurately achieve parameter optimization, balance multiple optimization objectives, adapt to dynamically changing working conditions, and ensure the effectiveness and robustness of the parameter optimization results.

[0064] Example 8: Closed-loop Verification and Incremental Optimization Method As Figure 7 shown, this example details the closed-loop verification and incremental optimization method, specifically including: Construct a refined digital twin model of the unit to simulate and verify the effect of parameter optimization. Preferably, the digital twin model includes a hydraulic model, a mechanical model, and an electrical model, which can simulate the dynamic response of the unit under different working conditions and provide a safe and reliable verification environment for parameter optimization.

[0065] Design an extreme condition test framework to evaluate the performance of parameters under boundary conditions. Preferably, the extreme conditions include typical conditions such as the highest water head, the lowest water head, full load, low load, and rapid load change. By simulating these conditions in the digital twin model, the adaptability and stability of parameters under extreme conditions are evaluated.

[0066] Implement incremental learning based on experience replay to continuously optimize the model performance. Preferably, maintain an experience pool to store historical optimization experiences. Each time the model is updated, a batch of historical experiences is randomly selected and mixed with new experiences for training to avoid catastrophic forgetting and ensure the steady improvement of the model performance.

[0067] Develop knowledge distillation technology to compress and integrate new experiences into the existing model. Preferably, use the optimized model as the teacher model and the newly trained model as the student model. By minimizing the KL divergence of the outputs of the two models, transfer the knowledge of the teacher model to the student model to achieve efficient model update.

[0068] Design a forgetting control mechanism to retain key historical experiences. Preferably, perform importance scoring on the samples in the experience pool. Samples with high importance are retained for a longer time and have a lower probability of being replaced to ensure that key experiences are not forgotten.

[0069] Construct a parameter degradation detection model to identify the trend of parameter performance decline. Preferably, by monitoring the change trends of key performance indicators (response time, overshoot, stability, etc.), when the deterioration trend of the indicators exceeds the preset threshold, trigger a parameter degradation alarm.

[0070] Develop a self-triggering optimization mechanism that automatically starts re-optimization when performance degrades beyond a preset threshold. Preferably, set a performance degradation threshold, and when it is detected that the parameter performance degrades by more than the threshold (default is 10%), automatically start the parameter re-optimization process without manual intervention.

[0071] Implement fault traceability analysis to identify the root causes affecting parameter performance. Preferably, analyze the correlation between parameter performance degradation and system state changes through a causal inference model, find out the root causes of performance degradation, and provide precise guidance for parameter adjustment.

[0072] Through a closed-loop verification and incremental optimization system, this method realizes the closed-loop control of parameter optimization, forms a continuous improvement mechanism of optimization - verification - re-optimization, ensures that the unit parameters always remain in the optimal state, and adapts to the changing operating environment.

[0073] Example 9: Cooperative Optimization Interface Method This example details the cooperative optimization interface method, which specifically includes: Design a plant-level cooperative optimization interface to achieve coordinated adjustment of unit group parameters. Preferably, the plant-level cooperative optimization interface considers the mutual influence between units, especially units sharing a common tailwater system or intake system, coordinates the adjustment of parameters of each unit, and makes the overall operating performance optimal.

[0074] Develop a knowledge sharing framework based on federated learning to protect data privacy. Preferably, each unit trains a model locally, only shares model parameters instead of raw data, the central server aggregates the model parameters of each unit, generates a global model, and then distributes it to each unit to achieve knowledge sharing while protecting data privacy.

[0075] Implement cross-power station knowledge transfer to support the rapid commissioning of new power stations. Preferably, through transfer learning technology, transfer the optimization experience of existing power stations to new power stations, accelerate the commissioning process of new power stations, and improve the commissioning efficiency and quality.

[0076] Design a multi-scale optimization strategy to achieve hierarchical optimization from single units to unit groups and then to the entire power station. Preferably, first optimize the parameters of single units, then conduct cooperative optimization of unit groups based on the optimization results of single units, and finally conduct cooperative optimization at the power station level to gradually improve the optimization effect.

[0077] Develop an asynchronous cooperative update mechanism to support flexible switching between independent optimization and cooperative optimization of different units. Preferably, each unit can independently optimize parameters according to its own needs, and at the same time can access the cooperative optimization framework when needed to achieve flexible adjustment of optimization strategies.

[0078] Implement an optimization experience sharing and traceability mechanism based on blockchain. Preferably, record the optimization experience on the blockchain to ensure the credibility and integrity of the data, while supporting the traceability and tracking of the optimization experience to understand its source and evolution process.

[0079] Through the collaborative optimization interface method, this system has achieved the expansion from single-unit optimization to unit group collaborative optimization and then to cross-power station knowledge sharing, significantly improving the utilization efficiency of optimization experience and providing an effective solution for the overall optimization of hydropower station groups.

[0080] Example 10: Intelligent Optimization System for Hydroelectric Generator Set Parameters with Multimodal Data Fusion As Figure 8 shown, the present invention also provides an intelligent optimization system for hydroelectric generator set parameters with multimodal data fusion, which system includes: A multimodal data acquisition module for acquiring multimodal data of the hydroelectric generator set, where the multimodal data includes time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data. This module adopts a distributed design and includes a variety of sensors and data acquisition units to achieve high-precision synchronous acquisition of the unit's multimodal data.

[0081] A feature extraction and fusion module for performing feature extraction and fusion on the multimodal data to generate a fused feature vector. This module contains multiple dedicated feature extraction networks and a hierarchical attention fusion network, and can extract and fuse valuable feature information from the multimodal data.

[0082] A knowledge graph construction and transfer learning module for constructing a unit knowledge graph and implementing cross-domain knowledge transfer based on a graph neural network to generate a prior knowledge representation. This module contains a knowledge graph construction unit and a graph neural network processing unit, and can capture the knowledge associations between units and achieve effective knowledge transfer.

[0083] A parameter optimization module for performing parameter optimization based on a Bayesian optimization algorithm with multi-objective constraints to determine the optimal PID parameter configuration. This module contains a Gaussian process modeling unit, a multi-objective acquisition function unit, and a parameter sensitivity analysis unit, and can efficiently and accurately achieve parameter optimization.

[0084] A closed-loop verification and incremental optimization module for performing incremental optimization on the optimized parameters through closed-loop verification to obtain the final optimized parameters. This module contains a digital twin verification unit, an incremental learning unit, and a parameter degradation detection unit, and can achieve closed-loop control and continuous improvement of parameter optimization.

[0085] The collaborative optimization interface module is used to achieve the collaborative adjustment of unit group parameters and cross-station knowledge migration. This module includes a plant-level collaborative interface unit, a federated learning unit, and a cross-station knowledge migration unit, which can expand the optimization scope and improve the utilization efficiency of optimization experience.

[0086] Each module of the system is connected through a data bus to form a complete data flow link. The multi-modal data acquisition module transports the collected data to the feature extraction and fusion module, the feature extraction and fusion module passes the fused feature vectors to the knowledge graph construction and transfer learning module, the knowledge graph construction and transfer learning module transports the prior knowledge representation to the parameter optimization module, the parameter optimization module passes the optimized parameters to the closed-loop verification and incremental optimization module, and finally obtains the optimized parameters and applies them to the hydropower unit.

[0087] This system adopts a hierarchical architecture design, including a perception layer, a feature layer, a cognition layer, a decision layer, and an execution layer. The information transfer between layers is clear, the function division is clear, and they work together to achieve the intelligent and efficient optimization of hydropower unit parameters.

[0088] Example 11: Application Case This example demonstrates the application effect of the present invention through an actual application case.

[0089] In the parameter optimization project of 4 Francis hydropower generating units in a certain hydropower station, the intelligent optimization method for hydropower unit parameters with multi-modal data fusion of the present invention is applied, and remarkable results are achieved.

[0090] The operating head range of the units in this hydropower station is 40 - 80 meters, the rated power is 180MW, and the PID control parameters have not been optimized for a long time, resulting in large differences in the response performance of the units under different head and load conditions. Especially under low head and small load conditions, the stability is poor and oscillation phenomena frequently occur.

[0091] First, deploy a multi-modal data acquisition network, install various sensors at key parts of the unit, and collect time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data.

[0092] Then, perform feature extraction and fusion on the collected multi-modal data to generate a fused feature vector reflecting the overall state of the unit. Based on the fused feature vector and historical optimization experience, construct a unit knowledge graph, and use graph neural networks to achieve cross-unit knowledge migration to generate a prior knowledge representation suitable for the current unit characteristics.

[0093] Next, perform parameter optimization based on the Bayesian optimization algorithm with multi-objective constraints, comprehensively consider multiple optimization objectives such as response time, overshoot, and stability, and determine the optimal PID parameter configuration. Conduct closed-loop verification through a digital twin model, and perform incremental optimization according to the verification results to obtain the final optimized parameters.

[0094] Finally, apply the optimized parameters to the actual units, and achieve coordinated adjustment of the parameters of the 4 units through the coordinated optimization interface.

[0095] After applying the method of the present invention, the response performance of the units is significantly improved: the response time is shortened from an average of 4.5 seconds to 2.8 seconds, a 37.8% improvement; the overshoot is reduced from an average of 6.2% to 3.1%, a 50% reduction; the number of oscillations is reduced from an average of 1.5 times to 0.6 times, a 60% reduction; the adjustment time is shortened from an average of 38 seconds to 22 seconds, a 42.1% reduction.

[0096] The overall efficiency of the units is increased by 4.2%, and the annual increased revenue is about 4.3 million yuan. The vibration intensity is reduced by 37%, and the service life of the equipment is extended by about 4 years. The optimization process only takes 15 minutes, which is more than 95% shorter than the traditional method. The units can maintain stable operation within the range of ±30 meters of water head change without manual intervention, comprehensively improving the adaptability and stability of the units.

[0097] In addition, through the coordinated optimization of the unit group, the overall operation of the 4 units is more coordinated. Especially when the water head and load change rapidly, they can quickly adjust and maintain stability, greatly improving the regulation ability and operation reliability of the power station.

[0098] This application case fully demonstrates the technical advantages and practical application value of the present invention, and verifies the effectiveness and advancement of multi-modal data fusion, graph neural network knowledge transfer, and Bayesian optimization in the parameter optimization of hydropower units.

[0099] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent optimization method for hydro-generator unit parameters based on multimodal data fusion, characterized in that, Including: Obtain multimodal data of the hydropower unit, where the multimodal data includes time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data; Perform feature extraction and fusion on the multimodal data, including: Use the intra-modal feature extraction network to process various types of data respectively to obtain intra-modal features; Calculate the correlation weights between different modal features based on the hierarchical attention mechanism to generate a fused feature vector; Construct a unit knowledge graph and implement cross-domain knowledge transfer based on the graph neural network, including: Construct a unit knowledge graph based on the fused feature vector and historical optimization experience; Use the graph attention network to extract the prior knowledge representation adapted to the current unit characteristics; Perform parameter optimization based on the Bayesian optimization algorithm with multi-objective constraints, including: Construct a Gaussian process surrogate model according to the prior knowledge representation; Determine the optimal PID parameter configuration using a multi-objective acquisition function; Perform incremental optimization on the optimized parameters through closed-loop verification to obtain the final optimized parameters.

2. The method according to claim 1, characterized in that, The obtaining of the multimodal data of the hydropower unit includes: Deploy a distributed multimodal sensing data acquisition network; Implement multi-source data synchronous acquisition based on timestamps, ensuring that the maximum time difference does not exceed 5 milliseconds; Perform outlier detection and data standardization processing on the acquired multimodal data; Generate a multimodal data tensor in a unified format.

3. The method according to claim 1, wherein The using of the intra-modal feature extraction network to process various types of data respectively includes: Use a TCN network based on causal convolution to extract time-domain data features; Use an improved ResNet network to extract frequency-domain vibration data features; Use Mel frequency spectrum transformation and CNN to extract acoustic data features; Use a spatial convolution network to extract temperature distribution data features; Use a deep fully connected network to extract hydraulic characteristic data features.

4. The method according to claim 1, characterized in that, The calculating of the correlation weights between different modal features based on the hierarchical attention mechanism includes: Calculate the attention weight matrix between different modal features: , Wherein: is the eigenvector of the th mode, is the eigenvector of the th mode, is the similarity calculation function; Enhance the original features based on the attention weight matrix: ， Integrate different-scale features through a feature pyramid fusion network to generate the final fused feature vector.

5. The method according to claim 1, wherein The unit knowledge graph includes: Unit nodes, representing the characteristic parameters of the hydropower unit; Operating condition nodes, representing the operating parameters of the unit; Parameter nodes, representing the PID control parameters; Physical association edges, representing the hydraulic coupling relationship between units; Performance similarity edges, representing the similarity degree of operating characteristics between units; Historical optimization edges, representing the transfer relationship of historical parameter optimization experience.

6. The method according to claim 1, wherein The using of the graph attention network to extract the prior knowledge representation adapted to the current unit characteristics includes: Implement a node-level attention mechanism to calculate the importance weights between nodes; Design an edge-level message passing mechanism to realize the transfer of knowledge in the graph structure; Through the domain alignment technology of maximum mean discrepancy, realize cross-unit and cross-headwater knowledge transfer; Generate the prior knowledge representation adapted to the current unit characteristics.

7. The method according to claim 1, characterized in that The Bayesian optimization algorithm with multi-objective constraints includes: Construct a Gaussian process surrogate model and adopt a multi-core function combination strategy: , where: is the basic kernel function, is the weight coefficient; Design a multi-objective acquisition function to balance multiple optimization objectives such as response time, overshoot, and stability; Implement a dynamic optimization strategy with condition awareness to automatically adjust the exploration intensity according to the head change; Develop a parameter sensitivity analysis module to identify key parameters and optimize them preferentially.

8. The method according to claim 1, characterized in that, The incremental optimization of the optimized parameters through closed-loop verification includes: Construct a refined digital twin model of the unit to simulate and verify the parameter optimization effect; Design an extreme condition test framework to evaluate the performance of the parameters under boundary conditions; Implement incremental learning based on experience replay to continuously optimize the model performance; Construct a parameter degradation detection model to identify the trend of parameter performance decline; Develop a self-triggered optimization mechanism to automatically start re-optimization when the performance decline exceeds a preset threshold.

9. The method according to claim 1, characterized in that, The method further includes: Design a plant-level collaborative optimization interface to achieve coordinated adjustment of unit group parameters; Develop a knowledge sharing framework based on federated learning to protect data privacy; Implement cross-power station knowledge transfer to support the rapid commissioning of newly built power stations.

10. An intelligent optimization system for hydro-generator unit parameters with multi-modal data fusion, which is used to execute the method described in any one of claims 1-9, characterized in that Including: A multi-modal data acquisition module for acquiring multi-modal data of the hydropower unit, where the multi-modal data includes time-domain operation data, frequency-domain vibration data, acoustic feature data, temperature distribution data, and hydraulic characteristic data; A feature extraction and fusion module for extracting and fusing features from the multi-modal data to generate a fused feature vector; A knowledge graph construction and transfer learning module for constructing a unit knowledge graph and implementing cross-domain knowledge transfer based on a graph neural network to generate a priori knowledge representation; A parameter optimization module for parameter optimization based on a multi-objective constrained Bayesian optimization algorithm to determine the optimal PID parameter configuration; A closed-loop verification and incremental optimization module for incrementally optimizing the optimized parameters through closed-loop verification to obtain the final optimized parameters; A collaborative optimization interface module for achieving coordinated adjustment of unit group parameters and cross-power station knowledge transfer.

Images