Hydroelectric generating set vibration prediction method based on multivariable time sequence

Abstract

The invention discloses a multi-variable time sequence-based hydroelectric generating set vibration prediction method. The method comprises the steps of forming a vibration data set comprising a plurality of time sequences; forming a time variable feature extraction module; forming an inter-variable feature extraction module; hyper-parameters of the model are adjusted, model training is carried out, and a prediction result is obtained. According to the method, a data set is provided for a subsequent input time variable feature extraction module and an inter-variable feature extraction module through a preprocessing and data driving-based time data segmentation method; different-scale feature extraction and fusion of a single time variable are realized through a multi-scale gating unit, feature information of a time sequence can be effectively extracted, noise interference is effectively suppressed, and integration and interaction of feature information among different time variables are realized through point convolution; semantic information loss before model data input is reduced through deformable sampling, gradient disappearance or explosion is prevented through learnable residuals and batch normalization, and the overfitting degree is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower station operation detection, and in particular to a vibration prediction method and prediction system for hydropower unit based on multivariable time series. Background Art

[0002] A hydroelectric unit is a device that converts the kinetic energy of water flow into mechanical energy, and then converts it into electrical energy through a generator. It is mainly composed of a turbine, a generator, a speed regulator, a control system, and other parts. The turbine drives the blades to rotate through the water flow, driving the generator to rotate, thereby generating electricity. According to the drop and flow of the water flow, hydroelectric units can be divided into different types, such as high-head units and low-head units. As a clean and renewable energy equipment, it is widely used in large hydropower stations around the world, providing important support for power supply. In order to ensure the safe and stable operation of modern hydropower stations, it is necessary to predict the vibration characteristics of the equipment by analyzing the historical operation data of the hydroelectric unit, thereby avoiding potential failures and performance degradation. The vibration prediction of the hydroelectric unit can provide early fault warnings and maintenance references, effectively reducing downtime and operation and maintenance costs.

[0003] At present, the main method for time series prediction of hydropower units is to predict the state of a single variable based on the historical data of a certain sensor. Most of the methods used for state prediction of hydropower units are based on the cascade deep learning network of recursive neural network LSTM and GRU. These methods basically only consider the time scale or spatial scale information of a single variable in the vibration data. However, the hydropower unit is a complex nonlinear time-varying system, which contains a variety of transient processes. Therefore, the use of a single variable for analysis cannot fully analyze the overall state information of the unit; and the above time prediction methods rely on the noise-free monitoring data obtained after denoising and decomposition of the original monitoring data. Although decomposition or denoising post-processing is more convenient, it will cause semantic information loss, resulting in the data obtained cannot fully and accurately reflect the actual operating status of the hydropower unit, but can only show the dynamic behavior of the unit under ideal conditions, ignoring the complexity and variability in actual applications, so the accuracy is low. Summary of the invention

[0004] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a vibration prediction method and prediction system for hydropower units based on multivariable time series.

[0006] In order to achieve the above object, in a first aspect, the present invention provides a method for predicting vibration of a hydropower unit based on a multivariate time series, comprising:

[0007] S100, collecting the operation data of different monitoring points of the hydropower unit, arranging and combining them to form a unit operation monitoring data set;

[0008] S200, merging the unit operation monitoring data set, performing preprocessing and segmentation, and forming a vibration data set containing multiple time series;

[0009] S300, extracting multi-scale feature information of each time series one by one through a multi-scale gating unit, performing feature extraction and fusion of time variables, and forming a time variable feature extraction module;

[0010] S400 extracts the feature relationship between different variables through point-wise convolution, integrates and interacts information across variables, and forms a feature extraction module between variables;

[0011] S500, adjust the hyperparameters of the model, divide the data set into a training set, a validation set, and a test set, perform model training, and obtain prediction results.

[0012] In some embodiments, the S200 includes:

[0013] S210, reading the unit operation monitoring data set and merging it into a data file;

[0014] S220, using deformable sampling to segment the input data file, and changing the position and size of the segmented data set based on data driving.

[0015] In some embodiments, in S200, the unit operation monitoring data set is X∈R C×T ,pass Split the data set into N blocks;

[0016] Among them, C is the number of channels or sequences, T is the length of the time series, N is the number of splits, S is the step size, P is the size of the block, and the size of the block obtained by splitting is x∈R P×C , the split data set is X∈R C×N×T .

[0017] In some embodiments, the S300 includes:

[0018] S310, normalizing the segmented data set and inputting it into a multi-scale gating unit;

[0019] S320, performing convolution processing on the normalized data set to extract multi-scale feature information of each time series one by one;

[0020] S330, using a concatenation function to fuse multiple scale feature information of a single variable, adding residual connection and activation function, extracting and fusing the features of the time variable, and forming a time variable feature extraction module.

[0021] In some embodiments, in S300, the data set after convolution processing is:

[0022]

[0023] Among them, 1×S I ∈(i=1,2,3) is the size of the convolution kernel;

[0024] Then the same monitoring signal is divided into two parts by evenly dividing it in the channel direction:

[0025]

[0026] Among them, E i and F i They are the first and second halves relative to the channel dimension, respectively;

[0027] Then control the weight of each part and multiply them point by point to obtain the long-term dependent data set X′ in the time series data i :

[0028]

[0029] Where φ(E i ) is the tanh activation function, σ(F i ) is the sigmoid activation function.

[0030] In some embodiments, in S300, the long-term dependent data set X′ i Perform pooling operation and pass Concat(Pooling(X′ i ))Splicing functions for joining:

[0031] Y=Concat(Pooling(X′1),Pooling(X′2),Pooling(X′3)),

[0032] Where Y is a feature dataset with a dimension of 3T-(S1+S2+S3-3). After the pooling operation, the dimension becomes (3T-(S1+S2+S3-3)) / W, where W is the step size.

[0033] Then, the (3T-(S1+S2+S3-3)) / W dimension is mapped to the T dimension, and the dataset Y∈R C×N×T ;

[0034] Then, the final output dataset Z is obtained by adding residual connections and ReLU activation functions. out , build a time-dependent prediction model:

[0035] Z out =ReLU(X+Y)∈R C×N×T .

[0036] In some embodiments, the S400 includes:

[0037] S410, adding pointwise convolution to extract feature relationships between different variables, integrating and interacting information across variables, and forming an inter-variable feature extraction module;

[0038] S420, performing a denormalization process to restore the output result of the inter-variable feature extraction module to its original size, and mapping to obtain an output result suitable for the prediction window length.

[0039] In some embodiments, in S400, the point-by-point convolution is a 1×1 point-wise convolution, and a connection relationship function model Z of different variable feature values ​​is established. ′ :

[0040]

[0041] Among them, K d is the weight value of the one-dimensional convolution kernel at the dth input channel, Z out,d is input Z out At the value of the dth channel, b is the bias term;

[0042] In some embodiments, in S400, the connection relationship function model Z ′ Through the GeLU (Gaussian Error Linear Units) activation function (with smooth characteristics, properties of approximate normal distribution and nonlinear characteristics, the gradient can be propagated more stably during the training of the neural network, increasing the nonlinear expression ability of the model), add the Dropout layer (in each training iteration, neurons in the neural network are randomly discarded with a certain probability, reducing the complex collaborative adaptability between neurons, avoiding overfitting, and improving generalization ability), perform a point-wise convolution, denormalize the results, and restore the model output to the original size:

[0043]

[0044] Finally, the fully connected layer is used for mapping to obtain the output result suitable for the prediction window length.

[0045] In some of the embodiments, in S500: the monitoring data set of the operating data of different monitoring points of the hydropower unit is divided into a training set, a validation set and a test set, with the training set divided into 70%, the validation set divided into 20%, and the test set divided into 10%, and then the hyperparameters are adjusted to perform model training and vibration prediction of the corresponding monitoring points.

[0046] The present invention has the following beneficial effects:

[0047] 1. The present invention provides a data set for the subsequent input time variable feature extraction module and the inter-variable feature extraction module through preprocessing and data-driven time data segmentation method, so that the model can learn more monitoring point operation data, grasp the dynamic operation characteristics of each part of the unit, grasp the dynamic globality of the unit, and at the same time, take into account the prediction of the monitoring point data of incomplete monitoring information or important components;

[0048] 2. The present invention realizes the extraction and fusion of different scale features of a single time variable through a multi-scale gating unit, which can effectively extract the feature information of the time series. For convolution kernels with different receptive fields, sufficient feature mining can be performed to ensure that more weights are assigned to important features, while effectively suppressing noise interference. The integration and interaction of feature information between different time variables is realized through point-wise convolution, and the operating characteristics between different detection signals are integrated;

[0049] 3. The present invention uses deformable sampling to reduce the loss of semantic information before the model inputs data, and prevents gradient disappearance or explosion through learnable residuals and batch normalization, thereby reducing the degree of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The process of the hydropower unit vibration prediction method proposed by the present invention is as follows: Figure 1 ;

[0051] Figure 2 The process of the hydropower unit vibration prediction method proposed by the present invention is as follows: Figure 2 ;

[0052] Figure 3 The process of the hydropower unit vibration prediction method proposed by the present invention is as follows: Figure 3 ;

[0053] Figure 4 The process of the hydropower unit vibration prediction method proposed by the present invention is as follows: Figure 4 ;

[0054] Figure 5 It is an operating principle diagram of the prediction model in the hydropower unit vibration prediction method proposed by the present invention;

[0055] Figure 6 This is a principle block diagram of the hydropower unit vibration prediction system proposed by the present invention;

[0056] Figure 7 This is the overall module architecture diagram of the hydropower unit vibration prediction method proposed by the present invention;

[0057] Figure 8 are the performance parameters of different models in predicting water conductance swing and corresponding data sets;

[0058] Fig. 9 It is a loss value curve of the prediction model and other models in the hydropower unit vibration prediction method proposed by the present invention during training;

[0059] Fig.10 The prediction results of the prediction model in the 720 length in the hydropower unit vibration prediction method proposed in the present invention are compared with the prediction results of other models. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The embodiment of the present application provides a vibration prediction method for a hydropower unit based on a multivariate time series, which solves the problems in the prior art: the main method based on the time series prediction of a hydropower unit is to predict the state of a single variable based on the historical data of a certain sensor, and the methods used for the state prediction of the hydropower unit are mostly based on the cascade deep learning network of the recursive neural network LSTM and GRU. These methods almost only consider the time scale or spatial scale information of a single variable in the vibration data. However, the hydropower unit is a complex nonlinear time-varying system, which contains a variety of transient processes. It is difficult to consider the overall state information of the unit by analyzing a single variable; at the same time, most time prediction methods rely on the noise-free monitoring data obtained after denoising and decomposition of the original monitoring data. Although decomposition or denoising post-processing is more convenient, it often causes semantic information loss, resulting in the inability of the obtained data to fully and accurately reflect the actual operating state of the hydropower unit, and can only show the dynamic behavior of the unit under ideal conditions, ignoring the complexity and variability in practical applications. The present application predicts the data of a certain monitoring point by inputting data from multiple monitoring points of the unit, so that the model can learn more operating data of monitoring points, grasp the dynamic operating characteristics of each part of the unit, grasp the dynamic globality of the unit, and at the same time take into account the prediction of monitoring point data of incomplete monitoring information or important components; it can effectively extract the characteristic information of time series, and for convolution kernels with different receptive fields, it can perform sufficient feature mining to ensure that more weights are assigned to important characteristics, and effectively suppress noise interference, use point-wise convolution to obtain feature information across variables, and integrate the operating characteristics between different detection signals; use deformable sampling to reduce the loss of semantic information before the model inputs data, accelerate convergence through learnable residuals, normalization and denormalization, facilitate comparison of the importance of different features, and prevent gradient disappearance or explosion, ensure the generalization ability and robustness of the model, and effectively reduce the degree of overfitting.

[0062] Please refer to the following examples for details:

[0063] Reference Figure 1-Figure 5 The present invention provides an embodiment of a method for predicting vibration of a hydropower unit based on a multivariate time series, comprising:

[0064] S100, collecting the operation data of different monitoring points of the hydropower unit, arranging and combining them to form a unit operation monitoring data set;

[0065] In step S100, the vibration data / pressure pulsation data / swing data and other detection data of the hydropower unit at different monitoring points are collected, and the original data obtained at each monitoring point are arranged and combined into a unit operation monitoring data set. The data set here can be 2 columns or more, and it is ensured that the data lengths obtained in all columns are equal;

[0066] Through preprocessing and data-driven time data segmentation method (deformable sampling), a data set can be provided for the subsequent input time variable feature extraction module and inter-variable feature extraction module, which enables the model to learn more monitoring point operation data, grasp the dynamic operation characteristics of each part of the unit, grasp the dynamic globality of the unit, and take into account the prediction of monitoring point data with incomplete monitoring information or important components;

[0067] S200, merging the unit operation monitoring data set, performing preprocessing and segmentation, and forming a vibration data set containing multiple time series;

[0068] Please continue reading Figure 2 In this embodiment, step S200 includes:

[0069] S210, reading the unit operation monitoring data set and merging it into a data file;

[0070] S220, using deformable sampling to segment the input data file, and changing the position and size of the segmented data set based on data-driven, can effectively reduce the semantic information loss of hard segmented data.

[0071] In step S200, the unit operation monitoring data set is X∈R C×T ,pass Divide the data set into N blocks; where C is the number of channels or sequences, T is the length of the time series, N is the number of splits, S is the step size, P is the size of the block, and the size of the block obtained by splitting is x∈R P×C , the split data set is X∈R C×N×T ;

[0072] S300, extracting multi-scale feature information of each time series one by one through a multi-scale gating unit, performing feature extraction and fusion of time variables, and forming a time variable feature extraction module;

[0073] The multi-scale gating unit is used to realize the extraction and fusion of features of different scales of a single time variable, which can effectively extract the feature information of the time series. For convolution kernels with different receptive fields, sufficient feature mining can be performed to ensure that more weights are assigned to important features, while effectively suppressing noise interference.

[0074] Please continue reading Figure 3 In this embodiment, step S300 includes:

[0075] S310, normalizing the segmented data set and inputting it into a multi-scale gating unit;

[0076] S320, performing convolution processing on the normalized data set to extract multi-scale feature information of each time series one by one;

[0077] S330, using a concatenation function to fuse multiple scale feature information of a single variable, adding residual connection and activation function, extracting and fusing the features of the time variable, and forming a time variable feature extraction module.

[0078] In step S300, a unit composed of three gated activation units (GTU) with different receptive fields (RF) is used to model the time dependency. The unit is mainly composed of three GTUs with different receptive fields. The data set is normalized before inputting the model to facilitate the model to extract features: first, the channel is mapped to 2C (C is the number of channels or sequences) through the convolution kernel, and then three 1D causal convolution layers (1D CausalConvolution), each layer has a convolution kernel of a different size, and then the time dependency on different time scales is fully mined from the input data, providing time feature representation for subsequent data analysis, model prediction and other tasks. The data set after convolution processing is:

[0079]

[0080] Among them, 1×S I ∈(i=1,2,3) is the size of the convolution kernel;

[0081] Then the convolution processed dataset X is transformed into i Average division, dividing the same monitoring signal into two parts:

[0082]

[0083] Among them, E i and F i They are the first and second halves relative to the channel dimension, respectively;

[0084] Then, the weight of each part is controlled by GTU, and point-by-point multiplication is performed to obtain the long-term dependent data set X′ in the time series data. i :

[0085]

[0086] Where φ(E i ) is the tanh activation function (a nonlinear function, symmetric about the origin, with a relatively simple derivative form, and different derivatives with different values; when x is close to 0, the derivative is close to the maximum value, which is conducive to the rapid propagation of the gradient and can converge faster), σ(F i) is the sigmoid activation function (a nonlinear function that can convert the input real number into a probability value; when x is equal to 0, the derivative is the maximum value, which is conducive to the rapid propagation of the gradient and can converge faster); by superimposing GTUs with different receptive fields, the ability to obtain long-term dependencies in time series data can be obtained;

[0087] Furthermore, the long-term dependence on the dataset X′ i Perform pooling operation (which can reduce the data dimension while retaining the key feature information in the data as much as possible, thereby improving computational efficiency and reducing the risk of overfitting), and pass Concat(Pooling(X′ i )) The concatenation function is used to join (multiple data structures can be concatenated along the specified dimension, combining different feature representations or data sets to form a new data structure containing more information):

[0088] Y=Concat(Pooling(X′1),Pooling(X′2),Pooling(X′3)),

[0089] Among them, Y is a dataset with features of dimension 3T-(S1+S2+S3-3). After the pooling operation, the dimension becomes (3T-(S1+S2+S3-3)) / W, where W is the step size. Then, a linear layer is used to map the (3T-(S1+S2+S3-3)) / W dimension to T dimension. The dataset Y∈R C×N×T ;

[0090] Then, by adding residual connections (residual connections are also called residual blocks. As the number of network layers increases, the gradient may become smaller and smaller during the back propagation process, making it difficult for training to converge. This is the gradient vanishing problem. Residual connections provide an additional, more direct path for gradient backpropagation, which helps to maintain effective gradient updates in deep networks, so that the network can be trained smoothly. Learnable residual connections can accelerate the convergence process and improve training stability) and ReLU activation function (when the input is greater than 0, the output is itself; when the input is less than or equal to 0, the output is 0, which enhances the nonlinear representation of the model, achieves better performance faster during training, and reduces the time and computing resource consumption required for training) to obtain the final output data set Z out , a time-dependent prediction model is established so that the model can learn the different scale characteristics of vibration information:

[0091] Z out =ReLU(X+Y)∈R C×N×T ;

[0092] S400 extracts the feature relationship between different variables through point-wise convolution, integrates and interacts information across variables, and forms a feature extraction module between variables;

[0093] Through point-wise convolution, the integration and interaction of feature information between different time variables is realized, and the operation characteristics between different detection signals are integrated. Through learnable residuals and batch normalization, the gradient disappearance or explosion is prevented, and the degree of overfitting is reduced;

[0094] Please continue reading Figure 4 In this embodiment, step S400 includes:

[0095] S410, adding pointwise convolution to extract feature relationships between different variables, integrating and interacting information across variables, and forming an inter-variable feature extraction module;

[0096] S420, performing a denormalization process to restore the output result of the inter-variable feature extraction module to its original size, and mapping to obtain an output result suitable for the prediction window length.

[0097] In step S400, since the sensors installed on the hydropower unit are not independent of each other, there is an inherent causal relationship in space between the state parameters. Therefore, after extracting the features of a single time series, it is necessary to model the coupling relationship between the variables. In order to capture the coupling relationship between the state parameters and improve the performance and efficiency of the model, a point-by-point convolution (1×1 point-wise convolution, which only considers a single position but extends the convolution operation of all channels) is added on the basis of the time-dependent prediction model to establish a connection relationship function model Z of different variable feature values. ′ , to achieve cross-variable information integration and interaction, thereby learning the complex relationship between variables:

[0098]

[0099] Among them, K d is the weight value of the one-dimensional convolution kernel at the dth input channel, Z out,d is input Z out At the value of the dth channel, b is the bias term;

[0100] Connection relationship function model Z ′ The GeLU activation function is used to increase the nonlinear expression ability of the model, a random inactivation layer (Dropout layer) is added to avoid overfitting, a normalization layer (Batch Normalization layer) is added to alleviate the gradient disappearance, and then a point-wise convolution is performed. Then, a denormalization layer is added to denormalize the result and restore the model output to its original size:

[0101]

[0102] Finally, the fully connected layer is used to map the output results suitable for the prediction window length.

[0103] S500, adjusting the hyperparameters of the model, dividing the data set into a training set, a validation set, and a test set, performing model training, and obtaining prediction results;

[0104] In step S500: the monitoring data set of the operating data of different monitoring points of the hydropower unit is collected and divided into a training set, a validation set and a test set; wherein, 70% of the training set is used for the model to learn the characteristics of the existing data set, 20% of the validation set is used to help verify the performance of the model on data outside the training set, and 10% of the test set is used to provide a fair data environment to measure the performance of the model on unknown data; then, the learning rate, the number of training rounds and other related hyperparameters are adjusted repeatedly, the model is trained and the prediction results are obtained, and the vibration prediction of the corresponding monitoring points is performed.

[0105] Reference Figure 6 The present invention also provides an embodiment of a vibration prediction system for hydropower generator units based on a multivariable time series, which is used to execute the vibration prediction method for hydropower generator units based on a multivariable time series in the above embodiment, and the prediction system includes:

[0106] (1) Data monitoring module, including multiple groups of detection sensors (vibration sensors, swing sensors and pressure pulsation sensors), used to monitor the bearing swing, frame vibration, hydraulic component pressure pulsation and other data of the hydropower unit, so as to facilitate the subsequent data collection to form a data set;

[0107] (2) Data operation module, including a multi-scale gated Tanh unit, which is mainly composed of three gated activation units (GTU) based on three different receptive fields (RF) to better process input data with different scale features;

[0108] It is understandable that the multi-scale gated Tanh unit is a structure that may be used in fields such as neural networks or signal processing, which can better process input data with different scale features; and the Tanh function (hyperbolic tangent function) is a common activation function with an output range between -1 and 1. It has nonlinear characteristics and can bring nonlinear mapping capabilities to neural networks; in neural networks, especially convolutional neural networks (CNN), the receptive field refers to the area range of input data that neurons can perceive: for example, for a neuron in a convolutional layer, its corresponding receptive field is the area covered by the convolution kernel sliding on the input image (or other data), and different receptive field sizes can capture features of different scales.

[0109] (3) A data storage module, comprising multiple memory groups for storing training sets, validation sets, and test sets.

[0110] Exemplarily, the memory can be independent or integrated with the processor. The above-mentioned processor may include one or more processing units, for example: the processor may include an application processor (application processor, AP), a modem processor, a graphics processor (graphics processing unit, GPU), an image signal processor (image signal processor, ISP), a controller, a video codec, a digital signal processor (digital signal processor, DSP), a baseband processor, and / or a neural network processor (neural-network processing unit, NPU), etc. Among them, different processing units can be independent devices or integrated in one or more processors. The controller can generate an operation control signal according to the instruction opcode and the timing signal to complete the control of fetching and executing instructions.

[0111] When the memory is a device independent of the processor, the electronic device may further include a bus. The bus is used to connect the memory and the processor. The bus includes hardware, software or both, and the components of the online data traffic billing device are coupled to each other. The bus may include an accelerated graphics port (Accelerated Graphics Port, AGP) or other graphics bus, an enhanced industry standard architecture (Extended Industry Standard Architecture, EISA) bus, a front side bus (Front Side Bus, FSB), a hypertransport (Hyper Transport, HT) interconnection, an industry standard architecture (Industry Standard Architecture, ISA) bus, an infinite bandwidth interconnection, a low pin count (LPC) bus, a memory bus, a micro channel architecture (MCA) bus, a peripheral component interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a serial advanced technology attachment (SATA) bus, a video electronics standard association local (VLB) bus or other suitable buses or a combination of two or more of these. In appropriate cases, the bus may include one or more buses. Although the present application embodiment describes and illustrates a specific bus, the present application considers any suitable bus or interconnection.

[0112] Prediction principle:

[0113] The operating data such as bearing swing, frame vibration, hydraulic component pressure pulsation, etc. transmitted by sensors at different monitoring points of the hydropower unit are collected, and the unit operation monitoring data set is divided and normalized using the processor to prepare the training data set for the subsequent model.

[0114] For the vibration / swing / pressure pulsation signals obtained from multiple monitoring points of different hydropower units, multi-scale gating units are used to extract the multi-scale variable eigenvalues ​​of the vibration signals. Three gating units with different receptive fields are adopted, and a 2D convolutional neural network is used to extract the variable eigenvalues. The Tanh and sigmoid activation functions are used to control the outflow ratios of different gating units. Subsequently, the maximum pooling operation is used to obtain the maximum eigenvalue extracted by the current convolutional network, and the concat splicing function operation is used to realize the connection between the eigenvalues. Finally, the original data is spliced ​​with the existing data to realize the extraction and fusion of multi-scale variable eigenvalues. During this period, the learnable residual connection is used to accelerate the convergence process and improve the training stability.

[0115] On the basis of time-dependent modeling, point-by-point convolution is added to capture the coupling relationship between the state parameters of monitoring points at different locations of the hydropower unit. A connection relationship is established between the characteristic values ​​of different variables through a 2D convolutional neural network to achieve cross-variable information integration and interaction, thereby learning the complex relationship between variables. Subsequently, the vector representation containing multi-scale and multi-variable features is restored to the original scale through denormalization processing, and then mapped to obtain the output suitable for the prediction window.

[0116] According to the model structure, input the dataset path, name and predicted variable name, adjust the length of the prediction sequence, the size of the block and the step size, and gradually set the number of training rounds and learning rate, and perform multiple rounds of learning iterations through forward propagation and reverse gradient propagation. Using this neural network structure and module to predict multivariable vibration sequences can improve the vibration prediction accuracy of hydropower units.

[0117] The results of the model training are compared with the existing models as shown in Figure 6, where: UNetTSF is a time series prediction model; MSFSP is a vibration prediction model for hydropower units based on multivariate time series; ConvTimesNet is a deep learning model for time series analysis, combining convolutional neural network (CNN) and Transformer model; DLinear is a one-layer linear time series prediction model; Linear is a linear regression model, NLinear is a linear time series prediction model; MAE (Mean Absolute Error); MSE (Mean Squared Error).

[0118] Under four prediction sequences of different lengths, in three experiments with longer prediction lengths, the proposed model obtained the lowest MSE and MAE values, indicating that this model has better prediction accuracy than other prediction models: in the results of prediction length 192, the mean square error of MSFSP is reduced by 1.3% and the mean absolute error is reduced by 3.6% compared with ConvTimeNet; in the results of prediction length 336, the mean square error of MSFSP is reduced by 9.6% and the mean absolute error is reduced by 5.5% compared with ConvTimeNet; in the results of prediction length 720, the mean square error of MSFSP is reduced by 3.5% and the mean absolute error is reduced by 1.4% compared with UnetTSF.

[0119] Figure 7 This is the loss value curve of the multivariate time series hydropower unit vibration prediction model MSFSP and other models during training. All models met the early stopping strategy within 80 training epochs, among which ConvTimeNet had the fastest convergence speed, stopping at the 22nd training epoch, and the minimum model loss was 0.065; followed by MSFSP, which stopped at the 36th training epoch, but the final loss value of MSFSP was 0.049, the lowest among all models, proving that it has a more obvious advantage in prediction accuracy compared with other models; UnetTSF is similar to DLinear, NLinear, and Linear, with a very stable descent rate, but a higher loss of 0.147, 0.149, 0.155, and 0.150 respectively.

[0120] Figure 8 This is a comparison chart of the prediction results of the multivariate time series hydropower unit vibration prediction model MSFSP based on this application and the final output of other models. It can be seen that the prediction trend and error of the MSFSP model are superior to those of other models.

[0121] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting vibration of hydropower units based on multivariate time series, characterized in that: include: S100, collecting the operation data of different monitoring points of the hydropower unit, arranging and combining them to form a unit operation monitoring data set; S200, merging the unit operation monitoring data set, performing preprocessing and segmentation, and forming a vibration data set containing multiple time series; S300, extracting multi-scale feature information of each time series one by one through a multi-scale gating unit, performing feature extraction and fusion of time variables, and forming a time variable feature extraction module; S400 extracts the feature relationship between different variables through point-wise convolution, integrates and interacts information across variables, and forms a feature extraction module between variables; S500, adjust the hyperparameters of the model, divide the data set into a training set, a validation set, and a test set, perform model training, and obtain prediction results.

2. The method for predicting vibration of a hydropower unit based on a multivariate time series according to claim 1 is characterized in that: The S200 includes: S210, reading the unit operation monitoring data set and merging it into a data file; S220, using deformable sampling to segment the input data file, and changing the position and size of the segmented data set based on data driving.

3. The method for predicting vibration of a hydropower unit based on a multivariate time series according to claim 2 is characterized in that: In S200, the unit operation monitoring data set is X∈R C×T ,pass Split the data set into N blocks; Among them, C is the number of channels or sequences, T is the length of the time series, N is the number of splits, S is the step size, P is the size of the block, and the size of the block obtained by splitting is x∈R P×C , the split data set is X∈R C×N×T .

4. The method for predicting vibration of a hydropower unit based on a multivariate time series according to claim 1, characterized in that: The S300 includes: S310, normalizing the segmented data set and inputting it into a multi-scale gating unit; S320, performing convolution processing on the normalized data set to extract multi-scale feature information of each time series one by one; S330, using a concatenation function to fuse multiple scale feature information of a single variable, adding residual connection and activation function, extracting and fusing the features of the time variable, and forming a time variable feature extraction module.

5. The method for predicting vibration of a hydropower unit based on multivariate time series according to claim 4 is characterized in that: In S300, the data set after convolution processing is: Among them, 1×S I ∈(i=1,2,3) is the size of the convolution kernel; Then the same monitoring signal is divided into two parts by evenly dividing it in the channel direction: Among them, E i and F i They are the first and second halves relative to the channel dimension, respectively; Then control the weight of each part and multiply them point by point to obtain the long-term dependent data set X′ in the time series data i : Where φ(E i ) is the tanh activation function, σ(F i ) is the sigmoid activation function.

6. The method for predicting vibration of a hydropower unit based on multivariate time series according to claim 4 is characterized in that: In S300, the long-term dependent data set X′ i Perform pooling operation and pass Concat(Pooling(X′ i ))Splicing functions for joining: Y=Concat(Pooling(X′1),Pooling(X′2),Pooling(X′3)), Where Y is a feature dataset with a dimension of 3T-(S1+S2+S3-3). After the pooling operation, the dimension becomes (3T-(S1+S2+S3-3)) / W, where W is the step size. Then, the (3T-(S1+S2+S3-3)) / W dimension is mapped to the T dimension, and the dataset Y∈R C×N×T ; Then, the final output dataset Z is obtained by adding residual connections and ReLU activation functions. out , build a time-dependent prediction model: Z out =ReLU(X+Y)∈R C×N×T 。 7. The method for predicting vibration of a hydroelectric unit based on a multivariate time series according to claim 1, characterized in that: The S400 includes: S410, adding pointwise convolution to extract feature relationships between different variables, integrating and interacting information across variables, and forming an inter-variable feature extraction module; S420, performing a denormalization process to restore the output result of the inter-variable feature extraction module to its original size, and mapping to obtain an output result suitable for the prediction window length.

8. The method for predicting vibration of a hydroelectric unit based on multivariate time series according to claim 7 is characterized in that: In S400, the point-by-point convolution is 1×1 point-wise convolution, and a connection relationship function model Z of different variable eigenvalues ​​is established. ′ : Among them, K d is the weight value of the one-dimensional convolution kernel at the dth input channel, Z out,d is input Z out The value of the dth channel, b is the bias term.

9. The method for predicting vibration of a hydroelectric generator set based on multivariate time series according to claim 7, characterized in that: In S400, the connection relationship function model Z ′ The GeLU activation function is used to increase the nonlinear expression ability of the model, and the Dropout layer is added to avoid overfitting. Then, a point-wise convolution is performed, and the obtained result is denormalized to restore the output of the model to its original size: Finally, the fully connected layer is used for mapping to obtain the output result suitable for the prediction window length.

10. The method for predicting vibration of a hydropower unit based on multivariate time series according to claim 1, characterized in that: In the S500, the monitoring data set of the operation data of different monitoring points of the hydropower unit is divided into a training set, a validation set and a test set, with the training set being divided into 70%, the validation set being divided into 20%, and the test set being divided into 10%, and then the hyperparameters are adjusted to perform model training and vibration prediction of the corresponding monitoring points.

Images