Interaction graph neural network prediction method and device for multi-element oscillation of hydroelectric generating set

Abstract

The invention belongs to the technical field of water conservancy projects, and particularly discloses an interaction graph neural network prediction method and device for multi-element vibration of a hydroelectric generating set, and the method comprises the steps: obtaining the steady-state data of the bearing swing during the real-time operation of the hydroelectric generating set; inputting the steady-state data into a Cross GNN model to obtain a predicted value of historical operation data of the hydroelectric generating set; and predicting the future health state of the hydroelectric generating set by combining the predicted value with an alarm threshold value of the operation of the hydroelectric generating set to obtain a health prediction result. According to the invention, the accuracy and robustness of the multi-element array pendulum data prediction of the hydroelectric generating set can be improved.

Description

Technical Field

[0001] This application belongs to the technical field of water conservancy engineering. More specifically, it relates to a prediction method and device for the interaction graph neural network of multiple vibration and swing of a hydropower unit. Background Art

[0002] Hydropower units play an important role in a high-proportion new energy power system and, as the core equipment for power conversion in power plants, undertake the important task of maintaining the stable operation of the power grid. The operating conditions of hydropower units are variable and need to be adjusted according to requirements. Therefore, the operating states need to be frequently changed, and the vibration characteristics of the units are different under various operating states. The problem of predicting the vibration signals of hydropower units has always been the focus of attention in the hydropower industry. Long-term operation leads to the deterioration of the unit state, which is reflected in the fluctuation of vibration data. This not only affects the safe and stable operation of hydropower units but may also cause serious equipment damage and even major accidents endangering personal safety.

[0003] In the prior art, data prediction of the vibration of hydropower units is mainly carried out by using big data analysis and deep learning technologies, often ignoring the noise influence of vibration signals and the interaction relationship between variables, and there are deficiencies in accuracy and robustness. Therefore, how to improve the accuracy and robustness of prediction to better ensure the safe and stable operation of hydropower units is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0004] Aiming at the defects of the prior art, the purpose of this application is to provide a prediction method and device for the interaction graph neural network of multiple vibration and swing of a hydropower unit, aiming to solve the problem of insufficient accuracy and robustness in the vibration prediction of hydropower units in the prior art.

[0005] To achieve the above purpose, in the first aspect, this application provides a prediction method for the interaction graph neural network of multiple vibration and swing of a hydropower unit, including.

[0006] Obtain the steady-state data of the bearing swing during the real-time operation of the hydropower unit; Input the steady-state data into the Cross GNN model to obtain the predicted values of the historical operation data of the hydropower unit; Combine the predicted values with the alarm threshold for the operation of the hydropower unit to predict the future health state of the hydropower unit and obtain the health prediction result; Wherein, the Cross GNN model is obtained by performing noise reduction processing, time feature extraction, variable feature extraction, and multi-layer perception prediction on the sample training set corresponding to the steady-state data samples of the real-time operation of the multiple arrays and swings of the hydropower unit to obtain sample prediction results, and based on the sample prediction results, model iteration and parameter update training are carried out.

[0007] Optionally, inputting the steady-state data into the Cross GNN model to obtain the predicted values of the historical operation data of the hydropower unit includes: Inputting the steady-state data into the Cross GNN model, and using the AMIS module of the Cross GNN model to perform noise reduction processing on the steady-state data to obtain denoised data; Analyzing the denoised data through the cross-scale graph neural network of the Cross GNN model to obtain the time-scale features of the steady-state data; Constructing a cross-scale variable graph from the time-scale features through the cross-variable interaction network of the Cross GNN model, and establishing the homogeneity and heterogeneity relationships between variables through positive and negative weight modeling to obtain cross-variable interaction information; Using a multi-layer perceptron to perform multi-step prediction on the cross-variable interaction information to determine the future operation trends of several steps of the hydropower unit operation, and obtaining the predicted values.

[0008] Optionally, using the AMIS module of the Cross GNN model to perform noise reduction processing on the steady-state data to obtain denoised data includes: Determining the swing value data in the steady-state data; Inputting the swing value data into the AMSI module to obtain a multi-scale time series; Performing noise reduction processing on the multi-scale time series using the AMSI module to generate a denoised multi-scale time series corresponding to the denoised data; Wherein, the swing value data includes the swing value data in the X direction of the upper guide bearing, the swing value data in the Y direction of the upper guide bearing, the swing value data in the X direction of the lower guide bearing, the swing value data in the Y direction of the lower guide bearing, the swing value data in the X direction of the water guide bearing, and the swing value data in the Y direction of the water guide bearing.

[0009] Optionally, the obtaining method of the multi-scale time series includes: Analyzing the time series of the hydropower unit in the frequency domain through fast Fourier transform, and calculating the amplitude of each time series at different frequencies; Selecting the target number of frequency values with the highest amplitude, calculating the corresponding period lengths of each frequency value, and performing downsampling on the original time series through average pooling operation to obtain time series of different scales; Stitching the time series of different scales to obtain a stitched multi-scale time series for constructing a cross-scale time graph.

[0010] Optionally, the construction method of the cross-scale time graph includes: Determine time points at different scales according to the nodes of the cross-scale time graph, and determine the correlation weights between the time points according to the edges of the cross-scale time graph; Generate an initial weight matrix through the product of a learnable first vector and a second vector and the ReLU activation function, and normalize the initial weight matrix using Softmax; For each time node, restrict the number of neighboring nodes associated with the fine scale to be more than that of the coarse scale, and retain the association between the time node and its previous and subsequent nodes to capture the time trend, obtaining the time node features; Obtain the aggregated neighboring time node features according to the ReLU activation function, the initial weight matrix, and the features of each time node, so as to construct the cross-scale time graph.

[0011] Specifically, in this embodiment, the denoised data is analyzed through the cross-scale graph neural network of the Cross GNN model. The construction of the cross-scale time graph is the key to this step. The specific method is as follows: Each node represents a time point at a different scale, and each edge determines the correlation weight between the time points.

[0012] Generate an initial weight matrix through the product of a learnable first vector and a second vector and the ReLU activation function, and normalize it using the Softmax function.

[0013] For each time node, restrict the number of neighboring nodes associated with the fine scale to be more than that of the coarse scale, and at the same time retain the association between the time node and its previous and subsequent nodes to capture the time trend, obtaining the time node features.

[0014] Obtain the aggregated neighboring time node features according to the ReLU activation function, the initial weight matrix, and the features of each time node, thereby constructing the cross-scale time graph, and further obtaining the time scale features of the steady-state data.

[0015] Optionally, the construction method of the cross-scale variable graph includes: Use variables as nodes and the correlation weights between variables as edges; Generate an initial correlation weight matrix through the product of a learnable third vector and a fourth vector and the ReLU activation function, and normalize the initial correlation weight matrix using Softmax; Select the top K positive neighbors and the bottom K negative neighbors with the highest correlation weights for each node to construct a set of neighboring nodes; Renormalize the correlation weight matrix based on the set of neighboring nodes to obtain a restricted set of neighboring nodes for each variable; Aggregate the neighbor variable node features based on the ReLU activation function, the initial correlation weight matrix, and the feature of each variable node to obtain the cross-scale variable graph.

[0016] Specifically, in this embodiment, the time-scale features are used to construct a cross-scale variable graph through the cross-variable interaction network of the Cross GNN model, so as to model the homogeneous and heterogeneous relationships between variables and obtain cross-variable interaction information. The construction method of the cross-scale variable graph is as follows: Use variables as nodes and the correlation weights between variables as edges.

[0017] Generate the initial correlation weight matrix through the product of the learnable third vector and the fourth vector and the ReLU activation function, and normalize it using Softmax.

[0018] Select the top K positive neighbors and the bottom K negative neighbors with the highest correlation weights for each node to construct a neighbor node set.

[0019] Re-normalize the correlation weight matrix based on the neighbor node set to obtain a restricted neighbor node set for each variable.

[0020] Aggregate the neighbor variable node features based on the ReLU activation function, the initial correlation weight matrix, and the feature of each variable node to complete the construction of the cross-scale variable graph.

[0021] Optionally, the use of a multi-layer perceptron to perform multi-step prediction on the cross-variable interaction information includes: Perform feature mapping on the cross-variable interaction information, compress the output features in the time dimension to obtain compressed features; Perform time-step mapping on the compressed features, map the compressed features to the prediction sequence length, and output the future multi-step prediction results.

[0022] Specifically, in this embodiment, a multi-layer perceptron is used to perform multi-step prediction on the cross-variable interaction information. The specific process is as follows: Perform feature mapping on the cross-variable interaction information, compress the output features in the time dimension to obtain compressed features.

[0023] Perform time-step mapping on the compressed features, map them to the prediction sequence length, and output the future multi-step prediction results, that is, the predicted values of the historical operation data of the hydropower unit.

[0024] This application also provides an interactive graph neural network prediction device for the multi-variable vibration and swing of a hydropower unit, including: An acquisition module, configured to acquire the steady-state data of the bearing swing during the real-time operation of the hydropower unit; A calculation module, configured to input the steady-state data into a Cross GNN model to obtain a predicted value of the historical operation data of the hydropower unit; A prediction module, configured to combine the predicted value with an alarm threshold for the operation of the hydropower unit to predict the future health state of the hydropower unit and obtain a health prediction result; Wherein, the Cross GNN model is obtained by performing noise reduction processing, time feature extraction, variable feature extraction, and multi-layer perception prediction on a sample training set corresponding to steady-state data samples of the real-time operation of the multivariate array pendulum of the hydropower unit to obtain a sample prediction result, and performing model iteration and parameter update training based on the sample prediction result.

[0025] In a third aspect, the present application provides an electronic device, including: at least one memory for storing a program; at least one processor for executing the program stored in the memory, and when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0026] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program, and when the computer program runs on a processor, it causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0027] In a fifth aspect, the present application provides a computer program product, and when the computer program product runs on a processor, it causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0028] It can be understood that the beneficial effects of the above second aspect to the fifth aspect can refer to the relevant descriptions in the first aspect above, and will not be repeated here.

[0029] Generally speaking, compared with the prior art by the above technical solution conceived by the present application, the following beneficial effects are achieved: (1) The present application uses the AMIS module to reduce the noise of the steady-state data, remove the noise interference brought by the complex environment, and make the data better reflect the real operation state of the unit. The fast Fourier transform and average pooling operations are used to obtain multi-scale time series, which can capture features of different time granularities and comprehensively mine data information. By exploring the long-term trend of the vibration signal of the hydropower unit and the high-dimensional coupling relationship between variables, the prediction accuracy and robustness are improved.

[0030] (2) This application analyzes time-scale features through a cross-scale graph neural network, can identify the periodic and trend changes in the unit operation, construct a cross-scale variable graph, model the homogeneous and heterogeneous relationships between variables, and comprehensively understand the unit operation status. The combination of the two avoids single-variable deviation and misjudgment of time features, accurately captures the data law, improves the prediction accuracy, and enhances the robustness of the model to data changes.

[0031] (3) This application performs multi-step prediction on the cross-variable interaction information through a multi-layer perceptron, deeply integrates different variable and time-scale information, and gradually corrects the prediction results. Combining the predicted value with the alarm threshold can timely detect anomalies and give early warnings. Providing long-term information through multi-step prediction and optimizing the results through threshold judgment enhance the prediction accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is one of the schematic flowcharts of the interactive graph neural network prediction method for the multi-variable vibration and swing of a hydropower unit provided by an embodiment of this application; Figure 2 is the schematic diagram of the model of an embodiment of this application; Figure 3 is the second schematic flowchart of the interactive graph neural network prediction method for the multi-variable vibration and swing of a hydropower unit provided by an embodiment of this application; Figure 4 is the experimental result graph of an embodiment of this application; Figure 5 is the schematic structural diagram of the interactive graph neural network prediction device for the multi-variable vibration and swing of a hydropower unit provided by an embodiment of this application; Figure 6 is the schematic structural diagram of the electronic device provided by an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] In order to make the objectives, technical solutions, and advantages of this application clearer, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0034] The term "and / or" in this article is an association relationship describing associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. The symbol " / " in this article represents an "or" relationship between associated objects. For example, A / B represents A or B.

[0035] The terms "first", "second", etc. in the description and claims of this document are used to distinguish different objects, rather than to describe a specific order of objects. For example, the first response message, the second response message, etc. are used to distinguish different response messages, rather than to describe a specific order of response messages.

[0036] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0037] In the description of the embodiments of this application, unless otherwise specified, the meaning of "a plurality" refers to two or more. For example, a plurality of processing units refers to two or more processing units, etc.; a plurality of elements refers to two or more elements, etc.

[0038] The embodiments of this application will be described below in conjunction with the accompanying drawings in the embodiments of this application.

[0039] Refer to Figure 1 , this application provides an interactive graph neural network prediction method for multiple vibration and swing of a hydropower unit, including.

[0040] S101. Obtain the steady-state data of the bearing swing during the real-time operation of the hydropower unit; S102. Input the steady-state data into the Cross GNN model to obtain the predicted values of the historical operation data of the hydropower unit; S103. Combine the predicted values with the alarm threshold for the operation of the hydropower unit to predict the future health status of the hydropower unit and obtain a health prediction result; Among them, the Cross GNN model is obtained by performing noise reduction processing, time feature extraction, variable feature extraction, and multi-layer perception prediction on the sample training set corresponding to the steady-state data samples of the real-time operation of the multiple array swings of the hydropower unit to obtain sample prediction results, and performing model iteration and parameter update training based on the sample prediction results.

[0041] Optionally, the step of inputting the steady-state data into the Cross GNN model to obtain the predicted values of the historical operation data of the hydropower unit includes: Input the steady-state data into the Cross GNN model, and use the AMIS module of the Cross GNN model to perform noise reduction processing on the steady-state data to obtain noise-reduced data; Analyze the noise-reduced data through the cross-scale graph neural network of the Cross GNN model to obtain the time-scale features of the steady-state data; Construct a cross-scale variable graph from the time-scale features through the cross-variable interaction network of the Cross GNN model, and establish the homogeneity and heterogeneity relationships between variables through positive and negative weight modeling to obtain cross-variable interaction information; Use a multi-layer perceptron to perform multi-step prediction on the cross-variable interaction information to determine the future operating trends of several steps of the hydropower unit operation, and obtain the predicted values.

[0042] Specifically, the embodiments of this application mainly include the following steps: Data acquisition: Collect the steady-state data of the bearing swing during the real-time operation of the hydropower unit. These data are the basis for subsequent analysis and prediction.

[0043] Calculation of predicted values: Input the obtained steady-state data into the trained Cross GNN model to obtain the predicted values of the historical operation data of the hydropower unit.

[0044] Health status prediction: Combine the predicted values with the alarm threshold of the hydropower unit operation to evaluate the future health status of the hydropower unit and obtain the health prediction results.

[0045] It should be noted that the training of the Cross GNN model is the key foundation of the entire solution, and its specific training process is as follows: Prepare a sample training set, and perform noise reduction processing, time feature extraction, variable feature extraction, and multi-layer perceptron prediction on the training set in sequence to obtain sample prediction results.

[0046] Based on the sample prediction results, use an appropriate loss function (such as mean square error) to calculate the difference between the predicted values and the true values, and then perform model iteration and parameter update through the backpropagation algorithm until the model converges.

[0047] Specifically, the prediction process of the embodiments of this application is as follows: First, the AMIS module of the model is used to reduce the noise of the steady-state data and handle the noise problem in the time series data through AMSI. By analyzing the time series on multiple time scales, AMSI can identify the features at different scales. The features at the coarse scale usually have less noise and more obvious trends.

[0048] The specific operation is to first determine the swing value data in the steady-state data, which cover the swing values in the X and Y directions of the upper guide bearing, the X and Y directions of the lower guide bearing, and the X and Y directions of the water guide bearing. Input these swing value data into AMSI to obtain a multi-scale time series, and then perform noise reduction processing on this multi-scale time series to generate a noise-reduced multi-scale time series.

[0049] Use the extracted multi-scale time series data to construct a cross-scale time graph using Cross-Scale GNN, where the nodes represent time points at different scales and the edges represent the correlation weights between time points. Model the dependencies between different time scales through a graph neural network (GNN) to extract scale features with clearer trends and weaker noise. This step analyzes the time series of the hydropower unit in the frequency domain through the fast Fourier transform, calculates the amplitudes of each time series at different frequencies. Select the target number of frequency values with the highest amplitudes, calculate their corresponding period lengths, then perform downsampling on the original time series using average pooling operation to obtain time series at different scales, and finally concatenate these time series at different scales to construct a cross-scale time graph.

[0050] Use Cross-Variable GNN to utilize the interaction information between variables in the multivariate time series data to extract invariant correlations including homogeneity and heterogeneity. It enhances the robustness of the model to noise and improves the accuracy of time series prediction by modeling the dynamic correlations between different variables.

[0051] After completing the above steps, the vibration characteristics of the hydropower unit under steady-state operation are obtained. Subsequently, DMS uses two multi-layer perceptrons (MLP) to map the output features of Cross-Variable GNN to the prediction target. The task of the first MLP (MLPC) is to map these high-dimensional features to a lower-dimensional space, usually one-dimensional, to capture the global trends and patterns of the time series. This step helps to simplify the feature representation, making it easier to process while retaining key information. The task of the second MLP (MLPT) is to map these dimension-reduced features to the predicted time steps, and it predicts the values of each future time step based on the features of historical data. In multivariate time series prediction, directly predicting multiple future time steps can reduce the cumulative error caused by multiple iterative predictions and improve the efficiency and accuracy of prediction.

[0052] Optionally, the step of using the AMIS module of the Cross GNN model to perform noise reduction processing on the steady-state data to obtain noise-reduced data includes: Determine the swing value data in the steady-state data; Input the swing value data into the AMSI module to obtain a multi-scale time series; Perform noise reduction processing on the multi-scale time series using the AMSI module to generate a denoised multi-scale time series corresponding to the denoised data; Among them, the swing value data includes the X-direction swing value data of the upper guide bearing, the Y-direction swing value data of the upper guide bearing, the X-direction swing value data of the lower guide bearing, the Y-direction swing value data of the lower guide bearing, the X-direction swing value data of the water guide bearing, and the Y-direction swing value data of the water guide bearing.

[0053] Optionally, the acquisition method of the multi-scale time series includes: Analyze the time series of the hydropower unit in the frequency domain through fast Fourier transform, and calculate the amplitude of each time series at different frequencies; Select the target number of frequency values with the highest amplitude, calculate the corresponding period lengths of each frequency value, and perform downsampling on the original time series through average pooling operation to obtain time series of different scales; Stitch the time series of different scales to obtain a stitched multi-scale time series to construct a cross-scale time graph.

[0054] Specifically, the acquisition process of the multi-scale time series is as follows: Analyze the time series in the frequency domain using FFT, and calculate the amplitude of each time series at different frequencies. Select frequencies: Then, select the Top-S frequency values with the highest amplitude, which correspond to the most important periodic components in the time series. Calculate the period length: For each selected frequency, calculate the corresponding period length P_s. Multi-scale time series extraction: Use average pooling (AvgPool) operation, with each period length as the Ps kernel size and stride, to perform downsampling on the original time series to obtain time series of different scales X_s. Stitch the multi-scale time series: Stitch the time series extracted at all scales in the time dimension to form a multi-scale time series , whose shape is R×L×D×C, where R is the extended dimension, L is the original input length, D is the number of variables, and C is the extended number of channels.

[0055] Optionally, the construction method of the cross-scale time graph includes: Determine the time points at different scales according to the nodes of the cross-scale time graph, and determine the correlation weights between the time points according to the edges of the cross-scale time graph; Generate an initial weight matrix through the product of the learnable first vector and second vector and the ReLU activation function, and apply Softmax to normalize the initial weight matrix; For each time node, restrict the number of neighboring nodes associated with the fine scale to be more than that of the coarse scale, and retain the associations between the time node and its previous and subsequent nodes to capture the time trend, thereby obtaining the time node features; Based on the ReLU activation function, the initial weight matrix, and the features of each time node, aggregate the neighboring time node features to construct the cross-scale time graph.

[0056] Specifically, the detailed process of constructing the cross-scale time graph in this embodiment is as follows: Cross-Scale GNN first constructs an initial cross-scale time graph, where nodes represent time points at different scales and edges represent the correlation weights between time points. Initialize the correlation weights: Use two learnable vectors vec_scale_1 and vec_scale_2 to initialize the correlation weight matrix E_scale through their product and the ReLU activation function, and then apply the Softmax function to ensure that the sum of the correlation weights of each time node is 1. Scale-sensitive restriction: For any time node, the number of its associated time nodes at the fine scale should be more than that at the coarse scale. By restricting the number of neighboring nodes at each scale, ensure that the fine-scale time series contributes more time node associations. Trend-aware selection: Retain the associations between the time node and its previous and subsequent nodes to ensure that the time trend can be captured. Re-normalize the correlation weights: Based on the selected set of neighboring nodes, re-normalize the correlation weights, filter out non-significant correlations, and retain the restricted set of neighboring nodes for each node. Cross-scale interaction: Perform cross-scale interaction based on GNN in the time dimension, and the information propagation process will stack multiple layers:

[0057] where, σ is the activation function, W is the learnable matrix, are the time node features, is the aggregation of the neighboring time node features. The output of Cross-Scale GNN is the time node features after being propagated through multiple layers of GNN. These features contain the information of cross-scale interaction and can be used for subsequent prediction tasks.

[0058] Optionally, the method for constructing the cross-scale variable graph includes: Use variables as nodes and the correlation weights between variables as edges; Generate an initial correlation weight matrix through the product of the learnable third vector and the fourth vector and the ReLU activation function, and apply Softmax to normalize the initial correlation weight matrix; Select the top K positive neighbors and the bottom K negative neighbors with the highest correlation weights for each node to construct a set of neighboring nodes; Re-normalize the correlation weight matrix based on the set of neighbor nodes to obtain a restricted neighbor node set for each variable; Based on the ReLU activation function, the initial correlation weight matrix, and the variable node features of each variable, aggregate the neighbor variable node features to obtain the cross-scale variable graph.

[0059] Specifically, in this embodiment, the time-scale features are used to construct a cross-scale variable graph through the cross-variable interaction network of the Cross GNN model, thereby modeling the homogeneity and heterogeneity relationships between variables and obtaining cross-variable interaction information. This embodiment mainly enhances the robustness of the model to noise and improves the accuracy of time series prediction by modeling the dynamic correlations between different variables.

[0060] The detailed process of constructing the cross-variable graph is as follows: Cross-Variable GNN first constructs an initial cross-variable graph, where nodes represent different variables and edges represent the correlation weights between variables. Initialize the correlation weights: Use two learnable vectors vec_var_1 and vec_var_2 to initialize the correlation weight matrix E_var through their product and the ReLU activation function, and then apply the Softmax function to ensure that the sum of the correlation weights for each variable is 1. Heterogeneity decoupling: Specifically, select the Kvar+ nodes with the highest correlation weights among the nodes as positive neighbors (homogeneous connections), and the Kvar- nodes with the lowest correlation weights as negative neighbors (heterogeneous connections). Correlation weight re-normalization: Based on the selected set of neighbor nodes, re-normalize the correlation weights, filter out non-significant correlations, and retain the restricted neighbor node set for each variable. For homogeneous edges, the weight is positively correlated with the correlation score; for heterogeneous edges, the weight is negatively correlated with the correlation score.

[0061]

[0062] Cross-variable interaction: Perform cross-variable interaction based on GNN in the variable dimension, and the information propagation process will stack multiple layers:

[0063] where σ is the activation function, W is the learnable matrix, is the variable node feature, is the aggregation of neighbor variable node features. The output of Cross-Variable GNN is the variable node features after multiple layers of GNN propagation, and these features contain the information of cross-variable interaction and can be used for subsequent prediction tasks.

[0064] Optionally, using the multi-layer perceptron to perform multi-step prediction on the cross-variable interaction information includes: Perform feature mapping on the cross-variable interaction information, compress the output features in the time dimension to obtain compressed features; Perform time step mapping on the compressed features, map the compressed features to the predicted sequence length, and output the multi-step prediction results for the future.

[0065] Specifically, in this embodiment, a multi-layer perceptron is directly used to perform multi-step prediction on six performance parameters.

[0066] The detailed description is as follows: Feature mapping: DMS uses two multi-layer perceptrons (MLPs) to map the output features of the Cross-Variable GNN to the prediction target. The first MLP (MLPC) maps the features in the time dimension from the channel dimension C to 1; the second MLP (MLPT) maps the length L' of the historical input sequence to the output sequence length.

[0067] MLPC: The first MLP compresses the output features of the Cross-Variable GNN in the time dimension to facilitate capturing the global trends and patterns of the time series.

[0068] ( ) Among them, is the output feature of the Cross-Variable GNN, and MLPC is the first MLP.

[0069] MLPT: The second MLP maps the output of the first MLP to the predicted time steps.

[0070]

[0071] Among them, is the predicted future time series, and MLPT is the second MLP.

[0072] The output of DMS is the directly predicted multi-step future time series, including the predicted values of each variable at future time steps.

[0073] The embodiments of this application have the following beneficial effects: reducing cumulative errors: directly predicting multiple future time steps can reduce the cumulative errors caused by step-by-step prediction. Improving efficiency: predicting multiple time steps at once instead of step by step improves the efficiency of the prediction process. Capturing long-term dependencies: Through the mapping of the MLP, DMS can capture the long-term dependencies in the time series.

[0074] Refer to Figure 2 , Figure 2It is a schematic diagram of the model in the embodiment of the present application. The figure includes the content of four parts, namely obtaining steady-state data, constructing a cross-scale graph and a cross-variable graph, using a multi-layer perceptron for prediction, and comparing the predicted value with the actual value.

[0075] Refer to Figure 3 , Figure 3 It is a complete flowchart of the embodiment of the present application, including the following steps: The hydro-generator unit extracts steady-state data; Use AMSI to remove noise, capture periods through FFT to divide different scales and connect in the time dimension; Initialize the correlation weight, scale-sensitive limit, and selection; re-normalize the correlation weight; Cross-Scale GNN constructs a cross-scale time graph; Initialize the decoupling of the correlation weight heterogeneity, and the weight is positively correlated with the correlation score; Cross-Variable GNN constructs a cross-variable graph; MLPC maps the features in the time dimension from the channel dimension C to a low-dimensional space; MLPT maps the historical input sequence length L′ to the output sequence length.

[0076] Refer to Figure 4 , Figure 4 It is a comparison graph of the predicted values and the actual values of the upper guide bearing X-direction swing value data, upper guide bearing Y-direction swing value data, lower guide bearing X-direction swing value data, lower guide bearing Y-direction swing value data, water guide bearing X-direction swing value data, and water guide bearing Y-direction swing value data of the ferry value data of the hydro-generator unit in the embodiment of the present application. It can be seen from the figure that the prediction result of the embodiment of the present application has a high accuracy and a small error from the actual value.

[0077] Refer to Figure 5 , the present application also provides an interactive graph neural network prediction device for the multi-variable vibration and swing of a hydro-generator unit, including: An acquisition module 510, configured to acquire the steady-state data of the bearing swing during the real-time operation of the hydro-generator unit; A calculation module 520, configured to input the steady-state data into the Cross GNN model to obtain the predicted value of the historical operation data of the hydro-generator unit; A prediction module 530, configured to combine the predicted value with the alarm threshold of the operation of the hydro-generator unit to predict the future health state of the hydro-generator unit and obtain a health prediction result.

[0078] Among them, the Cross GNN model is obtained by performing noise reduction processing, time feature extraction, variable feature extraction, and multi-layer perception prediction on the sample training set corresponding to the steady-state data samples of the real-time operation of the hydropower unit's multi-element array pendulum, and based on the sample prediction results, model iteration and parameter update training are carried out.

[0079] Optionally, inputting the steady-state data into the Cross GNN model to obtain the predicted value of the historical operation data of the hydropower unit includes: Inputting the steady-state data into the Cross GNN model, and using the AMIS module of the Cross GNN model to perform noise reduction processing on the steady-state data to obtain noise-reduced data; Analyzing the noise-reduced data through the cross-scale graph neural network of the Cross GNN model to obtain the time-scale features of the steady-state data; Constructing a cross-scale variable graph for the time-scale features through the cross-variable interaction network of the Cross GNN model, and establishing the homogeneity and heterogeneity relationships between variables through positive and negative weight modeling to obtain cross-variable interaction information; Using a multi-layer perceptron to perform multi-step prediction on the cross-variable interaction information to determine the future operation trends of several steps of the operation of the hydropower unit, and obtaining the predicted value.

[0080] Optionally, using the AMIS module of the Cross GNN model to perform noise reduction processing on the steady-state data to obtain noise-reduced data includes: Determining the swing value data in the steady-state data; Inputting the swing value data into the AMSI module to obtain a multi-scale time series; Performing noise reduction processing on the multi-scale time series using the AMSI module to generate a noise-reduced multi-scale time series corresponding to the noise-reduced data; Among them, the swing value data includes the swing value data in the X direction of the upper guide bearing, the swing value data in the Y direction of the upper guide bearing, the swing value data in the X direction of the lower guide bearing, the swing value data in the Y direction of the lower guide bearing, and the swing value data in the X direction of the water guide bearing, the swing value data in the Y direction of the water guide bearing.

[0081] Optionally, the acquisition method of the multi-scale time series includes: Analyzing the time series of the hydropower unit in the frequency domain through fast Fourier transform, and calculating the amplitude of each time series at different frequencies; Selecting the target number of frequency values with the highest amplitude, and calculating the corresponding period lengths of each frequency value, and performing downsampling on the original time series through average pooling operation to obtain time series of different scales; Stitch the time series of different scales to obtain a stitched multi-scale time series, so as to construct a cross-scale time graph.

[0082] Optionally, the method for constructing the cross-scale time graph includes: Determine the time points at different scales according to the nodes of the cross-scale time graph, and determine the correlation weights between the time points according to the edges of the cross-scale time graph; Generate an initial weight matrix through the product of a learnable first vector and a second vector and the ReLU activation function, and apply Softmax to normalize the initial weight matrix; For each time node, restrict the number of neighbor nodes associated with the fine scale to be more than that of the coarse scale, and retain the association between the time node and its front and back nodes to capture the time trend, so as to obtain the time node features; Obtain the aggregated neighbor time node features according to the ReLU activation function, the initial weight matrix, and the features of each time node, so as to construct the cross-scale time graph.

[0083] Specifically, in this embodiment, the denoised data is analyzed through the cross-scale graph neural network of the Cross GNN model. The construction of the cross-scale time graph is the key to this step. The specific method is as follows: Each node represents the time points at different scales, and each edge determines the correlation weights between the time points.

[0084] Generate an initial weight matrix through the product of a learnable first vector and a second vector and the ReLU activation function, and normalize it using the Softmax function.

[0085] For each time node, restrict the number of neighbor nodes associated with the fine scale to be more than that of the coarse scale, and at the same time retain the association between the time node and its front and back nodes to capture the time trend, so as to obtain the time node features.

[0086] Obtain the aggregated neighbor time node features according to the ReLU activation function, the initial weight matrix, and the features of each time node, thereby constructing a cross-scale time graph, and further obtaining the time scale features of the steady-state data.

[0087] Optionally, the method for constructing the cross-scale variable graph includes: Use variables as nodes and the correlation weights between variables as edges; Generate an initial correlation weight matrix through the product of a learnable third vector and a fourth vector and the ReLU activation function, and apply Softmax to normalize the initial correlation weight matrix; Select the top K positive neighbors and the bottom K negative neighbors with the highest correlation weights for each node to construct a set of neighbor nodes; Renormalize the correlation weight matrix based on the set of neighbor nodes to obtain a restricted neighbor node set for each variable; Obtain the aggregated neighbor variable node features according to the ReLU activation function, the initial correlation weight matrix, and the feature of each variable node to obtain the cross-scale variable graph.

[0088] Optionally, the method for constructing the cross-scale variable graph is as follows: Use variables as nodes and the correlation weights between variables as edges.

[0089] Generate an initial correlation weight matrix through the product of a learnable third vector and a fourth vector and the ReLU activation function, and normalize it using Softmax.

[0090] Select the top K positive neighbors and the bottom K negative neighbors with the highest correlation weights for each node to construct a set of neighbor nodes.

[0091] Renormalize the correlation weight matrix based on the set of neighbor nodes to obtain a restricted neighbor node set for each variable.

[0092] Obtain the aggregated neighbor variable node features according to the ReLU activation function, the initial correlation weight matrix, and the feature of each variable node to complete the construction of the cross-scale variable graph.

[0093] Optionally, the multi-step prediction of the cross-variable interaction information using the multi-layer perceptron includes: Perform feature mapping on the cross-variable interaction information, compress the output features in the time dimension to obtain compressed features; Perform time step mapping on the compressed features, map the compressed features to the predicted sequence length, and output the future multi-step prediction results.

[0094] It should be understood that the above device is used to execute the method in the above embodiment. For the corresponding program modules in the device, their implementation principles and technical effects are similar to those described in the above method. The working process of the device can refer to the corresponding process in the above method and will not be elaborated here.

[0095] Refer to Figure 6, based on the method in the above embodiments, an embodiment of the present application provides an electronic device, which may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communications interface 620, and the memory 630 complete communication with each other through the communication bus 640. The processor 610 can call the logical instructions in the memory 630 to execute the method in the above embodiments.

[0096] In addition, when the logical instructions in the above-mentioned memory 630 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this 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 for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application.

[0097] Based on the method in the above embodiments, an embodiment of the present application provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program runs on a processor, it causes the processor to execute the method in the above embodiments.

[0098] Based on the method in the above embodiments, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, it causes the processor to execute the method in the above embodiments.

[0099] It can be understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0100] The method steps in the embodiments of the present application can be implemented in a hardware manner or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in a random access memory (RAM), flash memory, read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), register, hard disk, removable hard disk, CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0101] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0102] It can be understood that the various numerical numbers involved in the embodiments of the present application are only for convenience of description and are not used to limit the scope of the embodiments of the present application.

[0103] Those skilled in the art can easily understand that the above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An interactive graph neural network prediction method for multivariate oscillation of a hydropower unit, characterized in that: include. Obtain the steady-state data of the bearing swing when the hydropower unit is running in real time; Inputting the steady-state data into the Cross GNN model to obtain the predicted value of the historical operation data of the hydropower unit; The predicted value is combined with the alarm threshold of the operation of the hydropower unit to predict the future health state of the hydropower unit to obtain a health prediction result; Among them, the Cross GNN model is obtained by performing noise reduction, time feature extraction, variable feature extraction and multi-layer perception prediction on the sample training set corresponding to the steady-state data samples of the real-time operation of the multi-element array of the hydropower unit to obtain the sample prediction results, and performing model iteration and parameter update training based on the sample prediction results.

2. The interactive graph neural network prediction method for multivariate oscillation of a hydropower unit according to claim 1 is characterized in that: The step of inputting the steady-state data into the Cross GNN model to obtain the predicted value of the historical operation data of the hydropower unit includes: Inputting the steady-state data into the Cross GNN model, and performing noise reduction processing on the steady-state data using the AMIS module of the Cross GNN model to obtain noise-reduced data; Analyzing the denoised data through the cross-scale graph neural network of the Cross GNN model to obtain the time scale characteristics of the steady-state data; The time scale feature is used to construct a cross-scale variable graph through the cross-variable interaction network of the Cross GNN model, and the homogeneity and heterogeneity relationship between modeling variables is modeled through positive and negative weights to obtain cross-variable interaction information; A multi-layer perceptron is used to perform multi-step prediction on the cross-variable interaction information to determine the operation trend of the hydropower unit in the next few steps and obtain the predicted value.

3. The interactive graph neural network prediction method for multivariate oscillation of a hydropower unit according to claim 2 is characterized in that: The step of performing denoising on the steady-state data using the AMIS module of the Cross GNN model to obtain denoised data includes: Determining swing value data in the steady-state data; Inputting the swing value data into the AMSI module to obtain a multi-scale time series; The multi-scale time series is subjected to denoising processing using an AMSI module to generate a denoised multi-scale time series corresponding to the denoised data; Among them, the swing value data includes the upper guide bearing X-direction swing value data, the upper guide bearing Y-direction swing value data, the lower guide bearing X-direction swing value data, the lower guide bearing Y-direction swing value data and the water guide bearing X-direction swing value data, the water guide bearing Y-direction swing value data.

4. The interactive graph neural network prediction method for multivariate oscillation of a hydropower unit according to claim 3 is characterized in that: The multi-scale time series acquisition method includes: The time series of the hydroelectric units are analyzed in the frequency domain by fast Fourier transform, and the amplitude of each time series at different frequencies is calculated; Select the target number of frequency values ​​with the highest amplitude, calculate the period length corresponding to each frequency value, and downsample the original time series through the average pooling operation to obtain time series of different scales; The time series of different scales are spliced ​​to obtain a spliced ​​multi-scale time series to construct a cross-scale time graph.

5. The interactive graph neural network prediction method for multivariate oscillation of a hydropower unit according to claim 4 is characterized in that: The method for constructing the cross-scale time graph includes: Determine time points at different scales according to each node of the cross-scale time graph, and determine correlation weights between each time point according to each edge of the cross-scale time graph; Generate an initial weight matrix by multiplying the first learnable vector and the second learnable vector and a ReLU activation function, and apply Softmax to normalize the initial weight matrix; For each time node, the number of neighbor nodes associated with fine scale is limited to be greater than that of coarse scale, and the association between the time node and its previous and next nodes is retained to capture the time trend and obtain the time node feature; Aggregated neighbor time node features are obtained according to the ReLU activation function, the initial weight matrix, and the features of each time node to construct the cross-scale time graph.

6. The interactive graph neural network prediction method for multivariate oscillation of a hydropower unit according to claim 2 is characterized in that: The method for constructing the cross-scale variable graph includes: Variables are used as nodes, and the correlation weights between variables are used as edges; Generate an initial correlation weight matrix by multiplying the third vector and the fourth vector that can be learned and using a ReLU activation function, and apply Softmax to normalize the initial correlation weight matrix; Select the first K positive neighbors and the last K negative neighbors with the highest relevance weights for each node to construct a set of neighbor nodes; Renormalize the correlation weight matrix based on the neighbor node set to obtain a restricted neighbor node set for each variable; Aggregated neighbor variable node features are obtained according to the ReLU activation function, the initial correlation weight matrix, and the features of each variable node, thereby obtaining the cross-scale variable graph.

7. The interactive graph neural network prediction method for multivariate oscillation of a hydropower unit according to claim 2 is characterized in that: The method of using a multi-layer perceptron to perform multi-step prediction on the cross-variable interaction information includes: Performing feature mapping on the cross-variable interaction information, compressing the output features from a time dimension, and obtaining compressed features; The compressed features are mapped to a time step length, the compressed features are mapped to a predicted sequence length, and future multi-step prediction results are output.

8. An interactive graph neural network prediction device for multivariate vibration of a hydropower unit, characterized in that: include. An acquisition module is used to obtain the steady-state data of the bearing swing when the hydropower unit is running in real time; A calculation module, used for inputting the steady-state data into the Cross GNN model to obtain a predicted value of the historical operation data of the hydropower unit; A prediction module, used to combine the prediction value with the alarm threshold of the operation of the hydropower unit to predict the future health state of the hydropower unit and obtain a health prediction result; Among them, the Cross GNN model is obtained by performing noise reduction, time feature extraction, variable feature extraction and multi-layer perception prediction on the sample training set corresponding to the steady-state data samples of the real-time operation of the multi-element array of the hydropower unit to obtain the sample prediction results, and performing model iteration and parameter update training based on the sample prediction results.

9. An electronic device, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program runs on a processor, the processor is caused to execute the method according to any one of claims 1 to 7.

Images