Broadband oscillation suppression method based on state space model and route optimization

Abstract

The invention relates to the technical field of power transmission, in particular to a broadband oscillation suppression method based on a state space model and route optimization. Firstly, collected multivariate time sequence metadata is preprocessed, processed data and a corresponding suppressor cooperation scheme are used as input and are sent into a deep learning model with corresponding labels, and a suppression scheme of a broadband oscillation phenomenon corresponding to the data is output. According to the method, differential modeling is carried out on oscillation data with different characteristics based on the selective state space model in combination with a dynamic routing mechanism, so that the fitting capability of the model to the data is improved, the data sequence dependence capability of the RNN-like model is reserved, the sequence modeling capability of the linear complexity of the time sequence data can be realized, and the reliability of the model is improved. Therefore, efficient and accurate modeling capability can be provided for long-time-series power data. According to the invention, the problem that various complex broadband oscillations in a power grid cannot be accurately, effectively and timely suppressed in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission, and in particular to a method for suppressing broadband oscillation based on a state space model and route optimization. Background Art

[0002] In the "dual-high" power system, interactions between power generation equipment, transmission networks, and power loads can cause destabilizing oscillations with frequencies ranging from a few Hz to several thousand Hz. This dynamic process, in which electrical quantities (such as voltage, current, and power) fluctuate periodically over time due to the interaction between power electronic equipment and between the power grid and the system itself, and the oscillation frequency varies over a wide range, is called broadband oscillation. Broadband oscillations exhibit new characteristics such as strong time-varying properties, strong nonlinearity, wide frequency range drift over time, and wide-area propagation of oscillation energy. Divergent oscillations can even lead to major safety incidents such as grid disconnection or grid collapse. In recent years, this has become a bottleneck restricting the safety of the "dual-high" power system.

[0003] To address broadband oscillations, academia and industry have achieved remarkable research results in the field of oscillation suppression. However, existing broadband oscillation suppression technologies based on control science, intelligent optimization, and intelligent control methods primarily focus on optimizing control and tuning parameters within the suppressor. However, broadband oscillation suppression requires the collaboration of multiple types of oscillation suppressors across multiple frequency bands, locations, and locations. Existing broadband oscillation suppression technologies struggle to accurately and promptly address the complex broadband oscillation issues that arise in power grids. Therefore, it is necessary to provide a method that can automatically provide suppression solutions for various broadband oscillation issues that arise in power grid lines. Summary of the Invention

[0004] To address the problem that existing technologies cannot accurately, effectively, and timely suppress various complex broadband oscillations that occur in the power grid, the present invention proposes a broadband oscillation suppression method based on a state-space model and routing optimization. The method has a reasonable design, high accuracy, and further improves computational efficiency, thereby ensuring timely processing of broadband oscillations.

[0005] The present invention proposes a broadband oscillation suppression method based on a state space model and routing optimization. First, a broadband oscillation suppression model is constructed. The suppressor parameters are generated based on the normalized results of multivariate time series data collected when the system experiences broadband oscillations. The broadband oscillation suppression model includes a sequentially connected feature extraction network, a gating network, a feature analysis network, and an output module; the feature extraction module is used to extract input data features; The feature analysis network consists of multiple layers, each with multiple Mamba blocks. From the first to the second-to-last layers, each Mamba block has a corresponding routing network behind it. The Mamba blocks process input data. The routing network calculates the probabilities of each Mamba block in the next layer, selects the transfer targets, and calculates the weights of each transfer target. The output of the Mamba blocks is distributed to each transfer target based on the weights. The gating network calculates the probability of the first-layer Mamba block in the feature analysis network based on the output of the feature extraction network, and selects the transfer objects to calculate the weights. The output of the feature extraction network is distributed to the transfer objects in the first layer of the feature analysis network according to the weights for processing. In the feature analysis network, each transfer object distributes its output to the transfer object in the next layer according to the weight calculated by the corresponding routing network. When the Mamba block is selected multiple times, it calculates independently for each input and sums all the calculated tensors as the output. The output of the last layer passes the object through weighted calculation and is converted into suppressor parameters through the output module.

[0006] Preferably, the broadband oscillation suppression model is trained on a data set {multivariate time series data samples, suppressor parameters}, the multivariate time series data samples are collected from a time period when broadband oscillation occurs, and the suppressor parameters are at a moment after the multivariate time series data and when broadband oscillation is suppressed; During the training of the broadband oscillation suppression model, the loss function L g Update the gated network in reverse order, according to the total loss function L total Reversely update the feature extraction network, feature analysis network and output module; L g The loss L contains the coupling value of the prediction error and the gating network guided load. gt , and the load balancing loss L of each Mamba block in the first layer of the balanced gating network guided feature analysis network gb ; L total The task loss L includes the evaluation prediction error t , load balancing loss L b and routing network fluctuation loss L c .

[0007] Preferred: ; Among them, y j Represents the input training sample x j The true label, E1 represents the number of Mamba blocks in the first layer of the feature analysis network, w j ,e Indicates that the gating network is a training sample x j The first layer of the allocated Mamba block M 1eThe weight of Ey e (x j ) is the model input sample x j And with Mamba block M 1e is the predicted label output by the model when the object is passed; B is the training batch size.

[0008] Preferred: ; Among them, U e Indicates the Mamba block M in the current round 1e utilization rate.

[0009] Preferred: ; Among them, K i represents the number of Mamba blocks contained in the i-th layer of the feature analysis network; M i,h represents the hth transfer object of the i-th layer in the feature analysis network, Represents processing training sample x j When passing object M i,h The utilization rate of G is 1, B is the training batch size, I is the number of layers in the feature analysis network, and G is the number of layers in the feature analysis network. i Represents the set of Mamba blocks that are passed to layer i.

[0010] Preferred: ; Among them, x r is the training sample; P jih Indicates that the input sample of the broadband oscillation suppression model is x j When the i-th layer transfer object M in the feature analysis network i,h The probability vector output by the routing network; P rih Indicates that the input sample of the broadband oscillation suppression model is x r When the i-th layer transfer object M in the feature analysis network i,h The probability vector output by the routing network; is a hyperparameter.

[0011] Preferably, the routing network includes a first linear layer, a ReLu activation function, a second linear layer, and a Softmax function connected sequentially; the output of the Mamba block first undergoes dimensionality reduction through the first linear layer, then passes through the ReLu activation function and then passes through the second linear layer to map the dimension to the number of Mamba blocks in the next layer, and then is normalized through the Softmax function to obtain the probability distribution P(x) of the Mamba blocks in the next layer; the routing network selects the transfer object based on probability, and allocates the weight of the corresponding Mamba block to the transfer object according to the probability ratio of the transfer object.

[0012] Preferably, the gating network includes a sequentially connected convolution layer, a Relu activation function, an average pooling layer, a linear layer, and a softmax layer; the output of the feature extraction network first passes through the convolution layer to reduce the dimension, then passes through the Relu activation function and global average pooling and flattening, then passes through the linear layer to obtain the routing weights of each Mamba block in the first layer of the corresponding feature analysis network, and then passes through the softmax network to normalize it to obtain the probability of each Mamba block in the first layer of the feature analysis network; the sum of the weights of each transfer object in the first layer of the feature analysis network is 1, and the weight ratio is equal to the probability ratio.

[0013] Preferably, the feature extraction network includes two Mamba blocks and a multi-layer perceptron; the two Mamba blocks are arranged in parallel between the input end of the feature extraction network and the multi-layer perceptron; the first Mamba block processes the input data, and the second Mamba block processes the input data after dimension flipping; the output of the second Mamba block is dimensionally flipped, superimposed with the output of the first Mamba block and the input data, and then input into the multi-layer perceptron to obtain the output of the feature extraction network.

[0014] The present invention proposes a broadband oscillation suppression system, comprising a memory and a processor, wherein the memory stores a computer program, the processor is connected to the memory, and the processor is configured to execute the computer program to implement the broadband oscillation suppression method based on a state space model and route optimization.

[0015] The advantages of the present invention are: (1) The present invention provides a broadband oscillation suppression method based on a state-space model and routing optimization. First, the collected multivariate time series data is preprocessed. The processed data and the corresponding suppressor collaboration scheme are fed into a deep learning model as input and corresponding labels, and the suppression scheme corresponding to the broadband oscillation phenomenon is output. In the present invention, the multivariate time series information in the transmission line is modeled based on a selective state-space model, namely the Mamba block. This method can achieve the linear complexity sequence modeling capability for time series data while retaining the data sequence dependency capability of the RNN-like model. Therefore, it can provide efficient and accurate modeling capabilities for long-time series power data.

[0016] (2) The present invention further introduces a dynamic routing mechanism similar to an expert network, which can adopt targeted modeling for various types of broadband oscillation problems that occur in transmission lines. By using the gating network and the dynamic routing mechanism to perform targeted processing on the captured timing features, different features can be modeled separately to improve accuracy, and sparse network activation can be used to reduce the required computing resources.

[0017] (3) The loss function of the present invention is composed of a combination of task loss and routing loss. In order to avoid load imbalance of each Mamba block, a load balancing loss based on the utilization rate of each Mamba block is added to the loss function and a noise mechanism is introduced to improve the exploratory nature of the network during training. At the same time, in order to avoid frequent jumps caused by unstable routing network in the early stage of training, a network fluctuation loss is added to it so that similar samples can choose the same routing route, thereby improving data utilization as much as possible and thus improving the prediction accuracy of the model.

[0018] (4) The present invention adopts a multi-channel dimension conversion feature fusion method for broadband oscillation time series data during feature extraction, and simultaneously captures the correlation between multi-channel data and the correlation between single-channel data contexts. It can not only utilize the combined features of multiple channels, but also capture the time series transformation features of single-channel data. Therefore, it has a more comprehensive and accurate feature extraction capability for multivariate time series data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a structural diagram of the broadband oscillation suppression model proposed in the present invention; Figure 2 This is a flow chart of the training method for the broadband oscillation suppression model proposed in the present invention; Figure 3 This is a flow chart of the broadband oscillation model method proposed by the present invention; Figure 4 It is the Mamba block structure in the embodiment; Figure 5 Graph 1 is a training convergence diagram of the broadband oscillation model in the embodiment. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0021] This embodiment proposes a broadband oscillation data preprocessing method for normalizing multivariate time series data.

[0022] The multivariate time series data when the system experiences broadband oscillation is denoted as X = {X1, X2, …, Xn, …, XN}; 1 ≤ n ≤ N, Xn represents the time series data of the nth data source; N is the number of data sources, including voltage data and current data. Xn={x n,1 ,x n,2 ,……, }; in, Indicates the nth data source t n The value at time t n ≥1;x n,1 Indicates the value of the nth data source at the first moment, x n,2 Indicates the value of the nth data source at the second moment; The Z-score normalization method is used to normalize the data Xn of each data source, and the normalized data of the multivariate time series data X is recorded as X'={X1',X2',…,Xn',…,XN'}; Xn'={x' n,1 ,x' n,2 ,……, }; =( -µ n ) / σ n； Where, µ n is the mean of the data source Xn, σ n is the variance of the data source data Xn.

[0023] After normalization of each data source, the data Xn' is aligned in time series and divided and spliced ​​according to the length L to form a tensor x tens ∈R N×L .

[0024] In this embodiment, a training data set {multivariate time series data, suppressor parameters} of a broadband oscillation suppression model is constructed, and the construction method is as follows: First, extract multivariate time series data containing broadband oscillation phenomena from the system's historical data, use the above-mentioned broadband oscillation data preprocessing method to obtain normalized data X' and align the time series, then find the broadband oscillation suppression moment and the corresponding suppressor parameters from the time length corresponding to X'; the suppressor parameters include the target parameters of the suppressors at different positions in the system, which is the system's suppressor collaborative control scheme. In specific implementation, extract the data of each data source at the broadband oscillation suppression moment as the multi-source data label, and extract the data of each data source with a time series length of L from this moment forward to form a multivariate time series data sample, that is, the tensor x tens ∈R N×L , thereby constructing a data set {multivariate time series data samples, suppressor parameters}; the suppressor parameters are the true labels of the multivariate time series data samples.

[0025] That is, the multivariate time series data refers to the power system data when broadband oscillation occurs, and the true label refers to the suppressor parameters taken to solve the broadband oscillation, which is after the multivariate time series data ends and the broadband oscillation is suppressed.

[0026] During specific implementation, the multivariate time series data corresponding to the broadband oscillation suppression moment can also be simulated by software to obtain the corresponding suppressor parameters, thereby enriching the number of data set samples.

[0027] Reference Figure 1 This embodiment proposes a broadband oscillation suppression model, which generates suppressor parameters based on multivariate time series data such as voltage and current when broadband oscillation occurs in the system. The suppressor suppresses broadband oscillation by executing the suppressor parameters.

[0028] The broadband oscillation suppression model includes a sequentially connected feature extraction network, a gating network, a feature analysis network and an output module.

[0029] The feature extraction network consists of two Mamba blocks and a multi-layer perceptron (MLP); the two Mamba blocks are placed in parallel between the input of the feature extraction network and the multi-layer perceptron, and the output of the multi-layer perceptron is the output of the feature extraction network. In the feature extraction network, the first Mamba block is used to extract the input sample X. in ∈R N×L For processing, the second Mamba block flips the dimension of the sample X in After processing, the output of the second Mamba block is dimensionally flipped and combined with the output of the first Mamba block and the sample X in After dimension superposition, input into the multi-layer perceptron to obtain the output of the feature extraction network ∈R N×L .

[0030] The feature analysis network consists of multiple layers, each with multiple Mamba blocks. From the first to the penultimate layer, each Mamba block has a corresponding routing network at the back end. The Mamba blocks are used to process input data, and the routing network is used to calculate the probabilities of each Mamba block in the next layer. The routing network selects transfer objects from the Mamba blocks in the next layer based on the probabilities and calculates the weight of each transfer object. The output of the Mamba block is distributed to each transfer object based on the weight.

[0031] Obviously, if there are too few Mamba blocks in the feature analysis network, fitting will be impossible; if there are too many, data utilization is low and the computational complexity is high. The size of the feature analysis network can be designed based on the data volume and requirements. In this embodiment, an optimal design method is to set the number of layers in the feature analysis network to be in the range (4, 16) and the number of Mamba blocks in each layer to be in the range (6, 16).

[0032] The routing network introduces an adjustable noise exploration mechanism to calculate the probability of each Mamba block in the next layer, adding controllable noise to the model for load balancing and preventing the dynamic routing mechanism from being overly concentrated on a few Mamba blocks.

[0033] The routing network consists of a first linear layer, a ReLu activation function, a second linear layer, and a Softmax function connected sequentially. The output of the Mamba block first undergoes dimensionality reduction through the first linear layer, then passes through the ReLu activation function and then passes through the second linear layer to map the dimension to the number of Mamba blocks in the next layer. It is then normalized through the Softmax function to obtain the probability distribution P(x) of the Mamba blocks in the next layer. The formula is expressed as: P(x)=Softmax{W p2 [ReLu(W p1 y M +ε)]} Among them, W p1 represents the first linear layer parameters, W p2 represents the second linear layer parameter, y M Represents the output of the corresponding Mamba block; ε represents the introduction of noise, which is the product of a random normalized tensor and a preset noise ratio. The noise ratio is specifically less than 1 and can be 0.1.

[0034] Specifically, the gating network is used to generate the probabilities and weights of each Mamba block in the first layer of the feature analysis network. First, the gating network generates the probabilities of each Mamba block in the first layer, then filters the transfer objects in the first layer based on the probabilities, and assigns weights based on the probability ratio of each transfer object in the first layer. The formula is as follows: Let the probability that the gated network assigns to each Mamba block in the first layer be { , g∈G1}, G1 represents the set of Mamba blocks in the first layer of the feature analysis network; the gating network selects multiple Mamba blocks from the first layer as transfer objects, and the set of transfer objects in the first layer is recorded as G1, G1∈G1; then, the weight of the hth transfer object in G1 is recorded as C(M 1,h ), M 1,h ∈G1; (1); in, Indicates that the gated network gives the h-th transfer object M in the first layer 1,h The probability of the assignment.

[0035] Transfer object M 1,h The output is denoted as y 1,h =C(M 1,h ) , is the output of the feature extraction network.

[0036] In the feature analysis network, the set of Mamba blocks in layer i is denoted as Gi; the set of Mamba blocks in layer i+1 is denoted as G(i+1); the set of Mamba blocks in layer i as the transfer object is denoted as Gi , G i ∈Gi; the set of Mamba blocks in the i+1th layer as the transfer object is denoted as G i+1 , G i+1 ∈G(i+1); 1≤ i ≤I-1, where I represents the number of layers in the feature analysis network.

[0037] The hth transfer object M in the i-th layer of the feature analysis network i,h The routing network is assigned to the g-th Mamba block M in the i+1th layer (i+1)g The probability of , M i,h ∈G i ; =1; Transfer object M i,h The weight is denoted as C(M i,h ), passing object M i,h The set of selected transfer objects is denoted as , G i+1 ={M i+1,h ;M i+1,h ∈ ,M i,h ∈G i};M i+1,h is the hth transfer object in the i+1th layer; M i+1,h The weight is: C(M i+1,h )= (2); M i,h The output is denoted as y i,h ;M i+1,h The output is y i+1,h : y i+1,h = (3); in, Indicates that M is selected i+1,h The set of transfer objects in the i-th layer as transfer objects.

[0038] The gating network consists of a sequentially connected convolutional layer, a Relu activation function, an average pooling layer, a linear layer, and a softmax layer. The output of the feature extraction network first passes through a convolutional layer to reduce its dimensionality, then passes through a Relu activation function and global average pooling and flattens it. It then passes through a linear layer to obtain the routing weights of each Mamba block in the first layer of the corresponding feature analysis network. This is then normalized through a softmax network to obtain the probabilities of each Mamba block in the first layer of the feature analysis network.

[0039] Specifically, when selecting transfer targets, the gating network and routing network can select based on probability sorting, that is, selecting a set number of transfer targets with the highest probability, namely the TOP-K mechanism; or select based on probability threshold, that is, selecting all Mamba blocks with probability greater than the set threshold as transfer targets.

[0040] The transfer object in the Mamba block of the last layer of the feature analysis network performs weighted calculation on its own output as the output of the feature analysis network and inputs it into the output module, which generates the predicted data y∈R N .

[0041] Specifically, the output module consists of two linear layers and a SoftMax activation function. The output of the feature analysis network first passes through the first linear layer to resize its dimension, then passes through the SoftMax activation function to obtain an activation value vector. The activation value vector is then converted into the suppressor parameters, i.e., the predicted data y, through the second linear layer.

[0042] Reference Figure 2 , the training method of the above broadband oscillation suppression model is: S1, initialize model parameters; S2. Extract B training samples from the training set and input them into the model to obtain the predicted labels and calculate the total loss function L. total And the loss function L of the gating network g ; (4); Among them, L t is the task loss, representing the difference between the predicted label and the true label; L b The load balancing loss is the additional loss added to avoid the routing network from being concentrated in the same Mamba block; c The loss of the routing network fluctuation is to avoid the initial instability of the routing network, which is prone to frequent jumps when selecting multiple lower-layer networks. λ1 and λ2 are both hyperparameters, and can be taken in the interval (0, 0.5]. In the subsequent embodiments, they are both set to 0.1.

[0043] (5); Among them, B is the training batch size, that is, the number of training samples in each round; y j Represents the input training sample x j The true label, Represents the training sample x j The predicted label, that is, x j The predicted value generated when the broadband oscillation suppression model is input; ||.|| 2 represents the square of the 2-norm.

[0044] (6); Among them, K i represents the number of Mamba blocks contained in the i-th layer of the feature analysis network; Represents processing training sample x j When M is used as the transfer object i,h The usage rate of Mamba blocks; that is, M i,h The total number of times selected by the transfer object of the i-1th layer in the feature analysis network; in this embodiment, only one gated network is set, when i=1, =1, because the first layer of the transfer object is selected by the gating network and can only be selected once. In specific implementation, multiple gating networks can also be set in parallel. In this way, the Mamba block of the first layer may be selected as the transfer object by multiple gating networks. Greater than 1.

[0045] (7); Among them, x r is the training sample; P jih Indicates that the input sample of the broadband oscillation suppression model is x j When the i-th layer transfer object M in the feature analysis network i,h The probability vector output by the routing network; P rih Indicates that the input sample of the broadband oscillation suppression model is x r When the i-th layer transfer object M in the feature analysis network i,h The probability vector output by the routing network; is the set hyperparameter, representing the similarity between two samples; It can take values ​​in the interval (0, 1), and in subsequent embodiments, takes a value of 0.5.

[0046] (8); Among them, L gt It represents the coupling value between the error between the predicted label (i.e., the final output of the model) and the true label and the gating network guided load, and is the main loss determined by the downstream task; L gb Represents the load balancing loss, which is used to balance the load of each Mamba block in the first layer of the gating network-guided feature analysis network; It is a set hyperparameter, which can take values ​​in the interval (0, 0.5). In the subsequent embodiments, it takes a value of 0.14.

[0047] (9); Among them, y j Represents the input training sample x j The true label, E1 represents the number of Mamba blocks in the first layer of the feature analysis network, wj ,e Indicates that the gating network is a training sample x j The first layer of the allocated Mamba block M 1e The weight of Ey e (x j ) is the model input sample x j And with Mamba block M 1e The predicted label output by the model when an object is passed.

[0048] When the training sample x j After inputting the model, M 1e If it is selected as the transfer object, the weight w is calculated according to formula (1) j ,e =C(M 1e ); if M 1e If it is not selected as the transfer object, then w j ,e =0.

[0049] (10); Among them, U e Indicates the Mamba block M in the current round 1e The utilization rate is M when processing the current batch of B training samples. 1e The number of times it was selected as a transfer object.

[0050] S3, according to L g Update the gating network in reverse order, according to L total Update the feature extraction network, feature analysis network and output module in reverse order.

[0051] S4. Repeat the above steps S2-S3 until the convergence condition is reached.

[0052] The convergence condition can be set as follows: the number of iterations of the model reaches a set value, or the total loss function L is reached for N consecutive times. total The range is less than the set difference.

[0053] Reference Figure 3 The input and output of the broadband oscillation suppression model in this embodiment are both normalized data. Therefore, in practical application, it is necessary to first normalize the multivariate time series data collected when the broadband oscillation phenomenon occurs, and then extract the input data X in ∈R N×L The broadband oscillation suppression model is input, and the output of the broadband oscillation suppression model is obtained and inverse normalized to obtain the execution parameters of each data element.

[0054] The broadband oscillation suppression model and method are verified below in conjunction with specific embodiments.

[0055] This embodiment takes the broadband oscillation data from the power grid as an example, and uses the voltage, current, active power, current converter parameters, and damping ratios of each oscillation mode collected in the line as the data source. Through time series alignment, the original multivariate time series data is constructed; then the above-mentioned broadband oscillation data preprocessing method is used to construct the data set {multivariate time series data, suppressor parameters}. The multivariate time series data in the data set is the data with the same dimension obtained after normalization of the multivariate data. The suppressor parameters are the target parameter set of multiple suppressors in the system and are the coordinated adjustment strategy of the suppressors in the system.

[0056] In this embodiment, the broadband oscillation suppression model provided by the present invention is used to extract and fuse different features between multi-channel and single-channel of multivariate time series data, and the obtained features are input into the gating network to obtain the initial routing direction, and then the obtained features are input into the deep neural network composed of the main Mamba block and the routing network to obtain the prediction results. Figure 4 In this embodiment, the feature analysis network is provided with 8 layers, and each layer is provided with 8 Mamba blocks.

[0057] In this embodiment, the data set is divided into a training set and a test set. The training set is sent to the broadband oscillation suppression model for training. After the training results converge, the accuracy of the trained model is verified using the test set. The convergence speed of the model training process of the present invention is as follows: Figure 5 shown.

[0058] In order to verify the effectiveness of the broadband oscillation suppression model proposed in the present invention, the broadband oscillation suppression model proposed in the present invention (here named DRmamba model) is simulated and compared with the long short-term memory network LSTM, the multi-head attention mechanism model Transformer, and the cross-variable information fusion linear model TSMixer.

[0059] Among them, the learning data sets of the three comparison models are {original multivariate time series data, suppressor parameters}; in order to ensure the uniformity of verification conditions, the data samples in the learning data set {original multivariate time series data, suppressor parameters} in this embodiment correspond one-to-one to the data samples in the data set {multivariate time series data, suppressor parameters}, and the acquisition time and time series length of the corresponding samples are equal; that is, the multivariate time series data of the data sample (multivariate time series data, suppressor parameters) is the normalized result of the original multivariate time series data of the corresponding data sample (original multivariate time series data, suppressor parameters).

[0060] The learning dataset {original multivariate time series data, suppressor parameters} and the dataset {multivariate time series data, suppressor parameters} are both split into training and test sets; The three comparison models were trained on the training set {original multivariate time series data, suppressor parameters}, and the loss function adopted mean square error loss; after training, they were tested on the test set {original multivariate time series data, suppressor parameters}.

[0061] The test accuracy of the model of the present invention and three comparison models is shown in Table 1 below, where the prediction accuracy is the root mean square error between the model output and the true label: Table 1 ;

[0062] Combine Figure 2 Compared with Table 1, it can be seen that although the algorithm converges slowly, its prediction accuracy after convergence is relatively large.

[0063] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0064] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0065] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A broadband oscillation suppression method based on state space model and routing optimization, characterized in that: Constructing a broadband oscillation suppression model to generate suppressor parameters based on the normalized results of multivariate time series data collected when the system has broadband oscillations; The broadband oscillation suppression model includes a sequentially connected feature extraction network, a gating network, a feature analysis network, and an output module; the feature extraction module extracts input data features; The feature analysis network has multiple layers, each with multiple Mamba blocks. From the first to the second-to-last layers, each Mamba block has a corresponding routing network behind it. The Mamba blocks process the input data. The routing network calculates the probabilities of each Mamba block in the next layer, selects the transfer targets, and calculates the weight of each transfer target. The output of the Mamba block is distributed to each transfer target based on the weight. The gating network calculates the probability of the first-layer Mamba block in the feature analysis network based on the output of the feature extraction network, and screens the transfer objects to calculate the weights; the output of the feature extraction network is distributed to the first-layer transfer objects of the feature analysis network according to the weights for processing; In the feature analysis network, each transfer object distributes its output to the next layer of transfer objects according to the weight calculated by the corresponding routing network; when the Mamba block is selected multiple times, it calculates independently for each input and sums all the calculated tensors as the output; The output of the last layer passes the object through weighted calculation and is converted into suppressor parameters through the output module.

2. The broadband oscillation suppression method based on state space model and routing optimization according to claim 1, characterized in that: The broadband oscillation suppression model is trained on the dataset {multivariate time series data samples, suppressor parameters}. The multivariate time series data samples are collected from the time period when broadband oscillation occurs, and the suppressor parameters are at the moment after the multivariate time series data and when broadband oscillation is suppressed. During the training of the broadband oscillation suppression model, the loss function L g Update the gated network in reverse order, according to the total loss function L total Reversely update the feature extraction network, feature analysis network and output module; L g Contains the loss L that represents the coupling between the prediction error and the gating network guided load gt , and the load balancing loss L of the load of each Mamba block in the first layer of the balanced gating network guided feature analysis network gb ; L total The task loss L includes the evaluation prediction error t , load balancing loss L b and routing network fluctuation loss L c .

3. The broadband oscillation suppression method based on state space model and routing optimization according to claim 2, characterized in that: ； Among them, y j Represents the input training sample x j The true label, E1 represents the number of Mamba blocks in the first layer of the feature analysis network, w j ,e Indicates that the gating network is a training sample x j The first layer of the allocated Mamba block M 1e The weight of Ey e (x j ) is the model input sample x j And with Mamba block M 1e is the predicted label output by the model when the object is passed; B is the training batch size.

4. The broadband oscillation suppression method based on state space model and routing optimization according to claim 3, wherein: ； Among them, U e Indicates the Mamba block M in the current round 1e utilization rate.

5. The broadband oscillation suppression method based on state space model and routing optimization according to claim 2, wherein: ； Among them, K i represents the number of Mamba blocks contained in the i-th layer of the feature analysis network; M i,h represents the hth transfer object of the i-th layer in the feature analysis network, Represents processing training sample x j When passing object M i,h The utilization rate of G is 1, B is the training batch size, I is the number of layers in the feature analysis network, and G is the number of layers in the feature analysis network. i Represents the set of Mamba blocks that are passed to layer i.

6. The broadband oscillation suppression method based on state space model and routing optimization according to claim 5, characterized in that: ； Among them, x r is the training sample; P jih Indicates that the input sample of the broadband oscillation suppression model is x j When the i-th layer transfer object M in the feature analysis network i,h The probability vector output by the routing network; P rih Indicates that the input sample of the broadband oscillation suppression model is x r When the i-th layer transfer object M in the feature analysis network i,h The probability vector output by the routing network; is a hyperparameter.

7. The broadband oscillation suppression method based on state space model and routing optimization according to claim 1, characterized in that: The routing network consists of a first linear layer, a ReLu activation function, a second linear layer, and a Softmax function connected sequentially. The output of the Mamba block first undergoes dimensionality reduction through the first linear layer, then passes through the ReLu activation function and then passes through the second linear layer to map the dimension to the number of Mamba blocks in the next layer. It is then normalized through the Softmax function to obtain the probability distribution P(x) of the Mamba blocks in the next layer. The routing network selects the transfer object based on probability and assigns the weight of the corresponding Mamba block to the transfer object according to the probability ratio of the transfer object.

8. The broadband oscillation suppression method based on state space model and routing optimization according to claim 7, characterized in that: The gating network consists of a sequentially connected convolutional layer, a Relu activation function, an average pooling layer, a linear layer, and a softmax layer. The output of the feature extraction network first passes through a convolutional layer to reduce its dimensionality, then passes through a Relu activation function and global average pooling and flattening. It then passes through a linear layer to obtain the routing weights of each Mamba block in the first layer of the corresponding feature analysis network. This is then normalized through a softmax network to obtain the probabilities of each Mamba block in the first layer of the feature analysis network. The sum of the weights of each transfer object in the first layer of the feature analysis network is 1, and the weight ratio is equal to the probability ratio.

9. The broadband oscillation suppression method based on state space model and routing optimization according to any one of claims 1 to 8, characterized in that: The feature extraction network includes two Mamba blocks and a multilayer perceptron; the two Mamba blocks are arranged in parallel between the input end of the feature extraction network and the multilayer perceptron; the first Mamba block processes the input data, and the second Mamba block processes the input data after dimension flipping. The output of the second Mamba block is dimensionally flipped, superimposed with the output of the first Mamba block and the input data, and then input into the multilayer perceptron to obtain the output of the feature extraction network.

10. A broadband oscillation suppression system, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the broadband oscillation suppression method based on the state space model and route optimization according to any one of claims 1 to 9.

Images