A method for suppressing broadband oscillations based on a state space model and route optimization

Abstract

The present application relates to the technical field of power transmission, and in particular to a wide frequency oscillation suppression method based on a state space model and routing optimization. First, the collected multivariate time series metadata is preprocessed, and the processed data and the corresponding suppressor cooperation scheme are taken as input and the corresponding label is sent into a deep learning model, and the output data corresponds to the suppression scheme of the wide frequency oscillation phenomenon. In the present application, based on the selective state space model combined with the dynamic routing mechanism, the oscillation data with different characteristics is modeled differently to improve the fitting ability of the model to the data, which can preserve the data sequence dependence ability of the RNN model while realizing the linear complexity sequence modeling ability of the time series data, so as to provide efficient and accurate modeling ability for long time series power data. The present application solves the problem that the prior art cannot accurately, effectively and timely suppress various complex wide frequency oscillations in the power grid.

Description

Technical Field

[0001] This invention relates to the field of power transmission technology, and in particular to a broadband oscillation suppression method based on state-space model and routing optimization. Background Technology

[0002] In high-voltage and high-efficiency power systems, the interaction between power generation equipment, transmission networks, and power loads can cause unstable oscillations in the frequency range of several Hz to several kiloHz. Due to the influence of internal or external factors, the electrical quantities (such as voltage, current, and power) generated by the interaction between power electronic devices and the power grid fluctuate periodically over time, and the oscillation frequency varies over a wide range. This dynamic process is called broadband oscillation in power systems. Broadband oscillations have new characteristics such as strong time-varying nature, strong nonlinearity, frequency range drifting over a large time interval, and oscillation energy propagating across a wide area. If divergent oscillations occur, they may even lead to major safety accidents such as power system disconnection or grid collapse, and in recent years have become a bottleneck problem restricting the safety of high-voltage and high-efficiency power systems.

[0003] To address broadband oscillation problems, academia and industry have achieved remarkable research results in oscillation suppression. However, existing broadband oscillation suppression techniques based on control theory, intelligent optimization, and intelligent control methods mainly focus on control optimization and parameter tuning within the suppressor. In reality, broadband oscillation suppression requires the cooperation of oscillation suppressors across various frequency bands, locations, and types. Existing broadband oscillation suppression techniques struggle to make accurate and timely judgments when faced with the various complex broadband oscillation problems occurring in power grids. Therefore, it is necessary to provide a method that can automatically provide suppression schemes for various broadband oscillation problems occurring 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 in the power grid, this invention proposes a broadband oscillation suppression method based on state-space model and routing optimization. This method is rationally designed, highly accurate, and further improves computational efficiency, thereby ensuring timely processing of broadband oscillations.

[0005] The present invention proposes a broadband oscillation suppression method based on state-space model and route optimization. First, a broadband oscillation suppression model is constructed, which generates suppressor parameters based on the normalization result of multivariate time-series data collected when the system experiences broadband oscillation.

[0006] The broadband oscillation suppression model consists of 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 features from the input data.

[0007] The feature analysis network consists of multiple layers, each with multiple Mamba blocks. From the first to the penultimate layer, each Mamba block is backed by a corresponding routing network. The Mamba blocks process the input data. The routing network calculates the probability of each Mamba block in the next layer, filters the transmission objects, and calculates the weight of each transmission object. The output of the Mamba blocks is distributed to each transmission object according to the weight.

[0008] 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 transmission objects to calculate weights; the output of the feature extraction network is distributed to the transmission objects in the first layer of the feature analysis network according to the weights for processing; in the feature analysis network, each transmission object distributes its output to the transmission objects in the next layer according to the weights calculated by the corresponding routing network; when a Mamba block is selected multiple times, it is calculated independently for each input, and the sum of all calculated tensors is used as the output;

[0009] The output of the last layer's passed object is weighted and then converted into suppressor parameters by the output module.

[0010] Preferably, the broadband oscillation suppression model is trained on the dataset {multivariate time series data samples, suppressor parameters}, where the multivariate time series data samples are collected from the time period in which the broadband oscillation occurs, and the suppressor parameters are at the time after the multivariate time series data and the broadband oscillation is suppressed.

[0011] During the training of the broadband oscillation suppression model, the loss function L is used as a reference. g Reverse update the gated network based on the total loss function L total The feature extraction network, feature analysis network, and output module are updated in reverse order.

[0012] L g Includes the loss L, representing the coupling value between the prediction error and the gating network-directed load. gt And the load balancing loss L of the load of each Mamba block in the first layer of the balanced gating network-oriented characteristic analysis network. gb ;

[0013] L total The task loss L includes evaluating the prediction error. t Load balancing loss L b And routing network fluctuation loss L c .

[0014] Preferred:

[0015] ;

[0016] Among them, y j The input training sample x jThe 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 the training sample x j The first-level e-th Mamba block M allocated 1e Weights; Ey e (x j (x) is the input sample for the model. j And with Mamba block M 1e B represents the predicted label output by the model when passing the object; B represents the training batch size.

[0017] Preferred:

[0018] ;

[0019] Among them, U e Indicates the Mamba block M in the current round 1e Utilization rate.

[0020] Preferred:

[0021] ;

[0022] Among them, K i This represents the number of Mamba blocks contained in the i-th layer of the feature analysis network; M i,h This represents the h-th transmitted object in the i-th layer of the feature analysis network. This indicates the processing of training sample x. j When passing object M i,h _B_ is the utilization rate; B is the training batch size; I represents the number of layers in the feature analysis network; G_ i This represents the set of Mamba blocks at level i that are passed as objects.

[0023] Preferred:

[0024] ;

[0025] Where, x r For training samples; P jih The input sample for the wideband oscillation suppression model is x. j At that time, the object M transmitted in the i-th layer of the feature analysis network i,h The probability vector output by the routing network; P rih The input sample for the wideband oscillation suppression model is x. r At that time, the object M transmitted in the i-th layer of the feature analysis network i,h The probability vector output by the routing network; This is a hyperparameter.

[0026] Preferably, the routing network includes a first linear layer, a ReLU activation function, a second linear layer, and a Softmax function connected in sequence. The output of a Mamba block is first reduced in dimensionality by the first linear layer, then activated by the ReLU function, and then passed through the second linear layer to map the dimension to the number of Mamba blocks in the next layer. It is then normalized by the Softmax function to obtain the probability distribution P(x) of the next layer Mamba blocks. The routing network selects the transmission object based on probability and distributes the weight of the corresponding Mamba block to the transmission object according to the probability ratio of the transmission object.

[0027] Preferably, the gated network includes sequentially connected convolutional layers, ReLU activation functions, average pooling layers, linear layers, and softmax layers. The output of the feature extraction network is first reduced in dimensionality by convolutional layers, then flattened by global average pooling after passing through the ReLU activation function, followed by passing through linear layers to obtain the routing weights of each Mamba block in the first layer of the feature analysis network, and then normalized by the softmax network to obtain the probability of each Mamba block in the first layer of the feature analysis network. The sum of the weights of each transmitted object in the first layer of the feature analysis network is 1, and the weight ratio is equal to the probability ratio.

[0028] Preferably, the feature extraction network includes two Mamba blocks and a multilayer perceptron; the two Mamba blocks are arranged side by side between the input of the feature extraction network and the multilayer perceptron; the first Mamba block processes the input data, the second Mamba block processes the input data after dimension flipping, and the output of the second Mamba block, after dimension flipping, is 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.

[0029] The present invention proposes a broadband oscillation suppression system, comprising a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor executes the computer program to implement the broadband oscillation suppression method based on state-space model and routing optimization.

[0030] The advantages of this invention are:

[0031] (1) This invention provides a broadband oscillation suppression method based on state-space model and route optimization. First, the collected multivariate time-series data is preprocessed. The processed data and the corresponding suppressor cooperation scheme are used as inputs and corresponding labels and fed into a deep learning model. The output is the suppression scheme for the broadband oscillation phenomenon corresponding to the data. In this invention, the multivariate time-series information in the transmission line is modeled based on the selective state-space model, namely the Mamba block. It can achieve the ability to model the linear complexity of time-series data while preserving the data order dependency capability of the RNN-like model. Therefore, it can provide efficient and accurate modeling capability for long-term power data.

[0032] (2) The present invention further introduces a dynamic routing mechanism similar to an expert network, which can be used to model various types of broadband oscillation problems in transmission lines. By using a gated network and a dynamic routing mechanism to process the captured time-series features, different features can be modeled separately to improve accuracy, and sparse network activation can be used to reduce the required computing resources.

[0033] (3) The loss function of the present invention is composed 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 enable similar samples to choose the same routing route, so as to maximize data utilization and improve the prediction accuracy of the model.

[0034] (4) In the feature extraction, the present invention adopts a multi-channel dimension transformation feature fusion method for broadband oscillation time series data, and captures the correlation between multi-channel data and the correlation between single-channel data context. It can utilize the combined features of multi-channel data and capture the time series transformation features of single-channel data. Therefore, it has a more comprehensive and accurate feature extraction capability for multi-dimensional time series data. Attached Figure Description

[0035] Figure 1 This is a structural diagram of the broadband oscillation suppression model proposed in this invention;

[0036] Figure 2 This is a flowchart of the training method for the broadband oscillation suppression model proposed in this invention;

[0037] Figure 3 This is a flowchart of the broadband oscillation model method proposed in this invention;

[0038] Figure 4 This refers to the Mamba block structure in the embodiment;

[0039] Figure 5 This is a training convergence graph of the broadband oscillation model in the embodiment. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] This embodiment proposes a broadband oscillation data preprocessing method for normalizing multivariate time series data.

[0042] Let the multivariate time series data when the system exhibits wideband oscillation be denoted as X={X1,X2,…,Xn,…,XN}; 1≤n≤N, where Xn represents the time series data of the nth data source; N is the number of data sources, which include voltage data and current data, etc.

[0043] Xn={x n,1 ,x n,2 ,……, };

[0044] in, Represents the nth data source t n The value at time t n ≥1; x n,1 Let x represent the value of the nth data source at the first moment. n,2 This represents the value of the nth data source at the second time step;

[0045] The Z-score normalization method is used to normalize the data Xn from each data source. Let the normalized data of the multivariate time series data X be denoted as X'={X1',X2',…,Xn',…,XN'}.

[0046] Xn'={x' n,1 ,x' n,2 ,……, };

[0047] =( -µ n ) / σ n；

[0048] Where, µ n Let σ be the mean of the data source Xn. n Let Xn be the variance of the data source.

[0049] After normalizing the data from each data source, the data Xn' is time-series aligned and then divided into blocks according to length L and spliced ​​together to form a tensor x. tens ∈R N×L .

[0050] In this embodiment, a training dataset {multivariate time series data, suppressor parameters} for a broadband oscillation suppression model is constructed. The construction method is as follows:

[0051] First, multivariate time-series data containing broadband oscillations are extracted from the system's historical data. Normalized data X' is obtained using the aforementioned broadband oscillation data preprocessing method and time-aligned. Then, the broadband oscillation suppression moment and corresponding suppressor parameters are located from the time length corresponding to X'. These suppressor parameters include the target parameters of suppressors at different locations in the system, forming the system's suppressor collaborative control scheme. In specific implementation, data from each data source at the broadband oscillation suppression moment are extracted as multi-source data labels. Data from each data source with a time-series length L is extracted backward from this moment to form a multivariate time-series data sample, i.e., tensor x. tens ∈R N×L This allows us to construct a dataset {multivariate time series data samples, suppressor parameters}; the suppressor parameters are the true labels of the multivariate time series data samples.

[0052] That is, multivariate time series data refers to power system data when broadband oscillations occur, while the true label refers to the suppressor parameters adopted to solve broadband oscillations, which are located after the multivariate time series data has ended and broadband oscillations have been suppressed.

[0053] In practice, software simulation can also be used to simulate the multivariate time series data corresponding to the broadband oscillation suppression time, thereby obtaining the corresponding suppressor parameters and enriching the number of dataset samples.

[0054] 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.

[0055] The broadband oscillation suppression model consists of a sequentially connected feature extraction network, a gating network, a feature analysis network, and an output module.

[0056] The feature extraction network consists of two Mamba blocks and a multilayer perceptron (MLP). The two Mamba blocks are placed side-by-side between the input of the feature extraction network and the MLP, and the output of the MLP is the output of the feature extraction network. In the feature extraction network, the first Mamba block processes the input sample X... in ∈R N×L The second Mamba block processes the sample X after its dimensions are flipped. inThe output of the second Mamba block is processed, and after dimension flipping, it is combined with the output of the first Mamba block and the sample X. in After dimension stacking, the input is fed into a multilayer perceptron to obtain the output of the feature extraction network. ∈R N×L .

[0057] The feature analysis network consists of multiple layers, each with multiple Mamba blocks. Each Mamba block in the first to penultimate layers has a corresponding routing network at its backend. The Mamba blocks process the input data, and the routing network calculates the probability of each Mamba block in the next layer. The routing network selects the transmission objects from the next layer's Mamba blocks based on the probability and calculates the weight of each transmission object. The output of the Mamba blocks is then distributed to each transmission object according to the weight.

[0058] Obviously, in a feature analysis network, if there are too few Mamba blocks, it cannot fit the data; if there are too many, the data utilization is low and the computational cost is high. The size of the feature analysis network can be designed according to the amount of data and the requirements. In this embodiment, a better design is to take the number of layers of the feature analysis network in the range (4, 16) and the number of Mamba blocks in each layer in the range (6, 16).

[0059] 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 to balance the load and avoid the dynamic routing mechanism from being overly concentrated on a few Mamba blocks.

[0060] The routing network consists of a first linear layer, a ReLU activation function, a second linear layer, and a Softmax function, all connected sequentially. The output of a Mamba block is first reduced in dimensionality by the first linear layer, then activated by the ReLU function, and finally passed through the second linear layer to map the dimension to the number of Mamba blocks in the next layer. The output is then normalized using the Softmax function to obtain the probability distribution P(x) of the next Mamba block. The formula is as follows:

[0061] P(x) = Softmax{W p2 [ReLu(W p1 y M +ε)]}

[0062] Among them, W p1 W represents the parameters of the first linear layer. p2 y represents the parameters of the second linear layer. M ε represents the output of the corresponding Mamba block; ε represents the introduced noise, which is the product of a randomized normalized tensor and a preset noise ratio. The noise ratio is less than 1, and can be 0.1.

[0063] 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, it selects the transmission objects in the first layer based on the probabilities and assigns weights according to the probability ratio of each transmission object in the first layer. The formula is as follows:

[0064] Let the probability of the gating network being assigned to each Mamba block in layer 1 be denoted as { Let g∈G1}, where 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 transit objects, and the set of transit objects in the first layer is denoted as G1, G1∈G1; then, the weight of the h-th transit object in G1 is denoted as C(M 1,h M 1,h ∈G1;

[0065] (1);

[0066] in, This indicates that the gating network passes the h-th object M to the first layer. 1,h The probability of allocation.

[0067] Passing object M 1,h The output is denoted as y. 1,h =C(M 1,h ) , This is the output of the feature extraction network.

[0068] In the feature analysis network, let the set of Mamba blocks in layer i be denoted as Gi; the set of Mamba blocks in layer i+1 be denoted as G(i+1); and the set of Mamba blocks in layer i that are passed on be denoted as G... i G i ∈Gi; the set of Mamba blocks at level i+1 that are passed as objects 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.

[0069] In the feature analysis network, the h-th transit object M in the i-th layer i,h The routing network is assigned to the g-th Mamba block at layer (i+1). (i+1)g The probability is denoted as M i,h ∈G i ; =1;

[0070] Passing object M i,h The weight is denoted as C(M). i,h ), passing object Mi,h The selected collection of 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 This refers to the h-th object in the (i+1)th layer.

[0071] M i+1,h The weights are:

[0072] C(M i+1,h )= (2);

[0073] M i,h The output is denoted as y. i,h M i+1,h The output is y i+1,h :

[0074] y i+1,h = (3);

[0075] in, Indicates that M is selected. i+1,h The set of objects to be passed in the i-th layer.

[0076] The gated network consists of sequentially connected convolutional layers, ReLU activation functions, average pooling layers, linear layers, and softmax layers. The output of the feature extraction network is first reduced in dimensionality by convolutional layers, then flattened by global average pooling after passing through the ReLU activation function. Next, it passes through linear layers to obtain the routing weights of each Mamba block in the first layer of the corresponding feature analysis network. Finally, it passes through a softmax network to normalize the weights, thus obtaining the probabilities of each Mamba block in the first layer of the feature analysis network.

[0077] Specifically, when selecting transmission objects, gating networks and routing networks can choose based on probability ranking, i.e., selecting a set number of transmission objects with the highest probability, i.e., the TOP-K mechanism; or they can choose based on a probability threshold, i.e., selecting all Mamba blocks with a probability greater than a set threshold as transmission objects.

[0078] In the Mamba block of the last layer of the feature analysis network, the passing object performs a weighted calculation of its own output as the output of the feature analysis network, and then inputs it into the output module. The output module generates the predicted data y∈R. N .

[0079] 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 adjust its dimensions, then passes through the SoftMax activation function to obtain an activation value vector. This activation value vector is then passed through the second linear layer to be transformed into suppressor parameters, i.e., the predicted data y.

[0080] Reference Figure 2 The training method for the above broadband oscillation suppression model is as follows:

[0081] S1. Initialize model parameters;

[0082] 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 The loss function L of the gated network g ;

[0083] (4);

[0084] Among them, L t The task loss represents the magnitude of the difference between the predicted label and the true label; L b The load balancing penalty is an additional cost incurred to prevent routing networks from concentrating on the same Mamba block; L c The routing network fluctuation loss is to avoid the initial instability of the routing network, which can easily lead to frequent jumps when selecting multiple lower-level networks; λ1 and λ2 are both hyperparameters, which can be taken in the range (0, 0.5], and are set to 0.1 in subsequent embodiments.

[0085] (5);

[0086] Where B is the training batch size, i.e., the number of training samples in each round; y j The input training sample x j The true label, Indicates training sample x j The predicted label, i.e., x j Predicted values ​​generated when inputting a wideband oscillation suppression model; ||.|| 2 This represents the square of the L2 norm.

[0087] (6);

[0088] Among them, K i This represents the number of Mamba blocks contained in the i-th layer of the feature analysis network; This indicates the processing of training sample x. j When, as the passing object M i,h The utilization rate of Mamba blocks; that is, M i,hThe total number of times the object being passed through the (i-1)th layer of the feature analysis network is selected; in this embodiment, only one gating network is set, and when i=1, =1, because the first-layer transmission object is selected by the gating network and can only be selected once. In practice, multiple gating networks can be set up in parallel, so the first-layer Mamba block may be selected as the transmission object by multiple gating networks. Greater than 1.

[0089] (7);

[0090] Where, x r For training samples; P jih The input sample for the wideband oscillation suppression model is x. j At that time, the object M transmitted in the i-th layer of the feature analysis network i,h The probability vector output by the routing network; P rih The input sample for the wideband oscillation suppression model is x. r At that time, the object M transmitted in the i-th layer of the feature analysis network i,h The probability vector output by the routing network; The hyperparameter is set to represent the similarity between two samples; It can take values ​​in the interval (0,1), and in the following embodiment, it takes the value 0.5.

[0091] (8);

[0092] Among them, L gt L represents the coupling value between the error of the predicted label (i.e., the final output of the model) and the true label, and is the main loss determined by the downstream task; gb This represents the load balancing loss, used to balance the load of each Mamba block in the first layer of a gated network-guided feature analysis network; The hyperparameter is set and can take values ​​in the range (0, 0.5). In the subsequent embodiment, the value is 0.14.

[0093] (9);

[0094] Among them, y j 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 the training sample x j The first-level e-th Mamba block M allocated 1e Weights; Ey e (x j (x) is the input sample for the model. jAnd with Mamba block M 1e The predicted label output by the model when passing an object.

[0095] When training sample x j After inputting the model, M 1e If selected as the transmission object, the weight w is calculated according to formula (1). j ,e =C(M 1e If M 1e If w is not selected as the transmission object, then j ,e =0.

[0096] (10);

[0097] Among them, U e Indicates the Mamba block M in the current round 1e The utilization rate, i.e., when processing the current batch of B training samples, M 1e The number of times it is selected as the transmission object.

[0098] S3, according to L g Reverse update the gating network, based on L total The feature extraction network, feature analysis network, and output module are updated in reverse.

[0099] S4. Repeat steps S2-S3 until the convergence condition is met.

[0100] 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 satisfied after N consecutive iterations. total The range is less than the set difference.

[0101] Reference Figure 3 In this embodiment, both the input and output of the broadband oscillation suppression model are normalized data. Therefore, in practical applications, it is first necessary to normalize the multivariate time-series data collected when the broadband oscillation occurs, and then extract the input data X. in ∈R N×L Input the broadband oscillation suppression model, obtain the output of the broadband oscillation suppression model, and after inverse normalization, obtain the execution parameters of each data element.

[0102] The above-mentioned broadband oscillation suppression model and method are verified in conjunction with specific embodiments below.

[0103] This embodiment takes broadband oscillation data from the power grid as an example. Voltage, current, active power, current converter parameters, and damping ratios of each oscillation mode collected from the line are used as data sources. Through time-series alignment, raw multivariate time-series data is constructed. Then, the aforementioned broadband oscillation data preprocessing method is used to construct a dataset {multivariate time-series data, suppressor parameters}. The multivariate time-series data in the dataset are data with the same dimensions obtained after normalization of the multivariate data. The suppressor parameters are the set of target parameters for multiple suppressors in the system, representing the coordinated adjustment strategy of the suppressors in the system.

[0104] In this embodiment, the wideband oscillation suppression model provided by this invention is used to extract and fuse different features between multi-channel and single-channel data from multivariate time-series data. The obtained features are input into a gating network to obtain the initial routing direction, and then the obtained features are input into a deep neural network composed of the main Mamba block and the routing network to obtain the prediction result. All Mamba blocks employ... Figure 4 The structure is shown. In this embodiment, the feature analysis network has 8 layers, with 8 Mamba blocks in each layer.

[0105] In this embodiment, the dataset is divided into a training set and a test set. The training set is fed into a 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 in this invention is as follows: Figure 5 As shown.

[0106] To verify the effectiveness of the broadband oscillation suppression model proposed in this invention, the broadband oscillation suppression model proposed in this invention (here named DRmamba model) is compared with the Long Short-Term Memory Network (LSTM), the Multi-Head Attention Mechanism Model (Transformer), and the Transvariable Information Fusion Linear Model (TSMixer) through simulation.

[0107] The learning datasets for the three comparison models are {original multivariate time series data, suppressor parameters}. To ensure the uniformity of the verification conditions, the data samples in the learning dataset {original multivariate time series data, suppressor parameters} in this embodiment correspond one-to-one with the data samples in the dataset {multivariate time series data, suppressor parameters}, and the collection 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).

[0108] The training dataset {original multivariate time series data, suppressor parameters} and the dataset {multivariate time series data, suppressor parameters} are both split into training set and test set;

[0109] The three comparative models were trained on the training set {original multivariate time series data, suppressor parameters}, with mean squared error loss as the loss function; after training, they were tested on the test set {original multivariate time series data, suppressor parameters}.

[0110] The test accuracy of the model of this invention and three comparative 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:

[0111] Table 1

[0112] ;

[0113] Combination Figure 2 As shown in Table 1, although this algorithm converges more slowly, its prediction accuracy after convergence is significantly superior.

[0114] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes 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 illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0115] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0116] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A broadband oscillation suppression method based on state-space model and route optimization, characterized in that, A broadband oscillation suppression model is constructed based on the normalization results of multivariate time-series data collected when the system experiences broadband oscillations, generating suppressor parameters. The broadband oscillation suppression model consists of a sequentially connected feature extraction network, a gating network, a feature analysis network, and an output module; the feature extraction module extracts features from the input data. The feature analysis network has multiple layers, each with multiple Mamba blocks. Each Mamba block from the first to the penultimate layer has a corresponding routing network at its backend. The Mamba blocks process the input data. The routing network calculates the probability of each Mamba block in the next layer and filters the transmission objects, calculating the weight of each transmission object. The output of the Mamba blocks is distributed to each transmission object according to 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 selects the transmission objects to calculate weights; the output of the feature extraction network is assigned to the transmission objects in the first layer of the feature analysis network according to the weights for processing. In the feature analysis network, each transmission object distributes its output to the next layer transmission object according to the weights calculated by the corresponding routing network; when a Mamba block is selected multiple times, it is calculated independently for each input, and the sum of all calculated tensors is used as the output. The output of the last layer's passed object is weighted and then converted into suppressor parameters by the output module.

2. The broadband oscillation suppression method based on state-space model and route optimization as described in 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 oscillations occur, and the suppressor parameters are at the time after the multivariate time series data and when broadband oscillations are suppressed. During the training of the broadband oscillation suppression model, the loss function L is used as a reference. g The gated network is updated in reverse, based on the total loss function L. total The feature extraction network, feature analysis network, and output module are updated in reverse order. L g Includes the loss L, representing the coupling value between the prediction error and the gating network-directed load. gt And the load balancing loss L of the load of each Mamba block in the first layer of the balanced gating network-oriented characteristic analysis network. gb ; L total The task loss L includes evaluating the 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 route optimization as described in claim 2, characterized in that: ； Among them, y j 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 the training sample x j The first-level e-th Mamba block M is allocated 1e Weights; Ey e (x j (x) is the input sample for the model. j And with Mamba block M 1e B represents the predicted label output by the model when passing the object; B represents the training batch size.

4. The broadband oscillation suppression method based on state-space model and route optimization as described in claim 3, characterized in that: ； 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 as described in claim 2, characterized in that: ； Among them, K i This represents the number of Mamba blocks contained in the i-th layer of the feature analysis network; M i,h This represents the h-th transmitted object in the i-th layer of the feature analysis network. This indicates the processing of training sample x. j When passing object M i,h _B_ is the utilization rate; B is the training batch size; I represents the number of layers in the feature analysis network; G_ i This represents the set of Mamba blocks at level i that are passed as objects.

6. The broadband oscillation suppression method based on state-space model and route optimization as described in claim 5, characterized in that: ； Where, x r For training samples; P jih The input sample for the wideband oscillation suppression model is x. j At that time, the object M transmitted in the i-th layer of the feature analysis network i,h The probability vector output by the routing network; P rih The input sample for the wideband oscillation suppression model is x. r At that time, the object M transmitted in the i-th layer of the feature analysis network i,h The probability vector output by the routing network; This is a hyperparameter.

7. The broadband oscillation suppression method based on state-space model and route optimization as described in 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, all connected sequentially. The output of a Mamba block is first reduced in dimensionality by the first linear layer, then activated by the ReLU function, and then passed through the second linear layer to map the dimension to the number of Mamba blocks in the next layer. Finally, it is normalized by the Softmax function to obtain the probability distribution P(x) of the next layer Mamba blocks. The routing network selects the transmission object based on probability and distributes the weight of the corresponding Mamba block to the transmission object according to the probability ratio of the transmission object.

8. The broadband oscillation suppression method based on state-space model and route optimization as described in claim 7, characterized in that, The gated network consists of sequentially connected convolutional layers, ReLU activation functions, average pooling layers, linear layers, and softmax layers. The output of the feature extraction network is first reduced in dimensionality by convolutional layers, then flattened by ReLU activation functions and global average pooling, and then passed through linear layers to obtain the routing weights of each Mamba block in the first layer of the feature analysis network. Finally, it is normalized by the softmax network to obtain the probability of each Mamba block in the first layer of the feature analysis network. The sum of the weights of each transmitted 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 route optimization as described in any one of claims 1-8, characterized in that, The feature extraction network consists of two Mamba blocks and a multilayer perceptron. The two Mamba blocks are arranged side by side between the input 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, after dimension flipping, is 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, It includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to implement the broadband oscillation suppression method based on state-space model and routing optimization as described in any one of claims 1-9.

Images