Power optical communication link traffic prediction method based on hybrid spatiotemporal deep learning model
By adopting a hybrid space-time deep learning model in the power communication network, combining DCNN, GRU and LSTM, and integrating time and space characteristics, the problem of insufficient traffic prediction accuracy in the power communication network in the prior art is solved, and higher prediction accuracy is achieved and network resource optimization.
Patent Information
- Application Number
- CN202410431436.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-11
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-04-11
AI Technical Summary
The existing flow prediction methods of power communication networks are difficult to adapt to the dynamic changes in the network structure, resulting in insufficient prediction accuracy and reliability.
The power optical communication link traffic prediction method based on a hybrid space-time deep learning model is adopted, combining deep convolutional neural network (DCNN), gated cycle unit (GRU) and long and short-term memory model (LSTM), and constructing spatio-time graphs and integration time With spatial features, improve prediction accuracy and response speed.
It significantly improves the accuracy and stability of power communication network traffic prediction, optimizes network resource configuration, and enhances the ability and robustness of dynamic changes in network traffic.
Smart Images

Figure CN118200169B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power optical communication networks, and in particular to a method for predicting power optical communication link traffic based on a hybrid spatiotemporal deep learning model. Background Art
[0002] With the rapid development and expansion of power communication networks, the power industry is facing increasingly complex data processing challenges. The power communication network is the nerve center of the power system. It not only supports the daily operation of the power system, but is also crucial to ensuring the safe, stable and efficient operation of the power system. With the transformation of the power system towards intelligence and digitalization, the power communication network is gradually developing in the direction of bandwidth, intelligence and business IP (Internet Protocol). This change is accompanied by the rapid growth of power business types and business volumes, as well as the uneven distribution of network traffic caused by sudden and uncertain business, which puts higher requirements on ensuring the high-quality transmission services of power business.
[0003] In this context, accurate and timely link traffic prediction becomes particularly important. Link traffic prediction is crucial for optimizing network resource allocation, ensuring efficient network operation, and preventing potential network congestion. Traditional traffic prediction methods mainly rely on time series analysis, such as ARIMA (Auto-Regressive Integrated Moving Average Model) and seasonal decomposition. However, these methods often ignore the complex spatial relationship between nodes in the power communication network. When predicting the business traffic of the power network, they only analyze and predict through time series, while ignoring the dynamic changes of the network topology. In the actual power network, the topology of the network is time-varying and complex, and its changes will directly affect the routing and transmission time of the data packet. Therefore, ignoring these changes will significantly reduce the accuracy and reliability of the prediction. Summary of the invention
[0004] In order to solve the problem that the existing power communication network prediction method is difficult to adapt to the dynamic topological structure relationship of the network structure, the present invention provides a power optical communication link flow prediction method based on a hybrid spatiotemporal deep learning model (Deep Convolution Neural Network-Gated Recurrent Unit, DCNN-GRU). The purpose of the present invention is to improve the prediction accuracy and optimize the response speed of the model to rapidly changing data to ensure the security and reliability of the network.
[0005] A method for predicting the flow rate of power optical communication links based on a hybrid spatiotemporal deep learning model is implemented by the following steps:
[0006] Step 1: Obtain the input traffic data set of the prediction model;
[0007] Step 2: Use the traffic data set obtained in step 1 to construct an adjacency matrix and a time feature matrix, and construct a spatiotemporal graph based on the adjacency matrix and the time feature matrix;
[0008] Step 3: construct a DCNN-GRU prediction model, use the spatiotemporal graph, i.e., the spatiotemporal feature matrix, as the input of the DCNN-GRU prediction model, and obtain spatial features and temporal features through DCNN-GRU prediction, i.e., predict the output traffic data matrix;
[0009] Step 4: Divide the traffic data set to perform training optimization of the DCNN-GRU prediction model and LSTM model integration;
[0010] Step 5: Use the weighted average method to integrate the DCNN-GRU prediction model and the LSTM model to output the traffic data matrix for the next time period.
[0011] Beneficial effects of the present invention:
[0012] 1. Based on the existing spatiotemporal graph convolutional neural network prediction model, the present invention proposes a DCNN-GRU (Deep Convolution Neural Network-Gated Recurrent Unit) prediction model. This model takes into account the changing network topology structure on the basis of prediction based on time series. By integrating the data of time dimension and space dimension, the model will more comprehensively capture the spatiotemporal characteristics in the power communication network, effectively predict the flow distribution and dynamic changes, thereby improving the accuracy of the prediction. In the actual power communication network, the present invention improves the network efficiency and stability and optimizes the network resource configuration by improving the accuracy of the prediction of network link flow.
[0013] 2. The method of the present invention combines the Deep Convolution Neural Network (DCNN), the Gated Recurrent Unit (GRU) and the Long Short-Term Memory (LSTM) model. Through the effective integration of temporal dynamics and spatial structure information by the model, the spatiotemporal characteristics of the network are fully captured, so that in a complex power communication network environment, the distribution and dynamic changes of traffic can be effectively predicted, and the prediction accuracy can be improved. It provides more comprehensive spatiotemporal information required for traffic prediction, which not only improves the ability to capture the spatial characteristics of network traffic data, but also enhances the processing efficiency of time series data.
[0014] 3. The present invention adopts a model integration strategy to combine the DCNN-GRU model and the independent long short-term memory (LSTM) model. This integrated model can outperform any single model. After combining the dynamic capture ability of the DCNN-GRU model and the long-term dependency processing ability of the LSTM model, the integrated model achieves higher prediction accuracy and reliability through weighted averaging. This unique fusion strategy provides more comprehensive information, enabling the integrated model to respond quickly to dynamic changes in network traffic, thereby enhancing the robustness of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 It is a flow chart of the method for predicting the flow rate of a power optical communication link based on a hybrid spatiotemporal deep learning model according to the present invention;
[0016] Figure 2 It is a schematic diagram of the DCNN-GRU prediction model in the present invention;
[0017] Figure 3 A comparison chart of the performance of the three models DCNN-GRU+LSTM, LSTM and GRU in data throughput prediction and the actual value;
[0018] Figure 4 This is a comparison chart of the root mean square error of the three models DCNN-GRU+LSTM, LSTM and GRU on the test set;
[0019] Figure 5 This is a comparison chart of the mean absolute errors of the three models DCNN-GRU+LSTM, LSTM and GRU on the test set. DETAILED DESCRIPTION
[0020] Combination Figures 1 to 5 This embodiment describes a method for predicting the flow rate of a power optical communication link based on a hybrid spatiotemporal deep learning model, which is implemented by the following steps:
[0021] Step 1: Data preparation stage, including data collection and data processing;
[0022] Step 1.1: Data collection;
[0023] Various data required for subsequent predictions are obtained in the State Grid's network management system NMS (Network Management System). Link flow data is used to record flow information on each link, such as the number of packets, the number of bytes, and the data rate; network topology data is used to describe the physical structure or logical connection of the network; historical flow data is used to capture time dependencies; device status data and external factor data mainly include CPU usage, device temperature, power demand data, weather data, etc.
[0024] Step 1.2: Data processing;
[0025] The collected data is cleaned to remove missing and duplicate values, and all missing values of the feature are filled with the average value of the feature; duplicate values in the data set are deleted. This allows cleaner and more useful data to be obtained, so that the time and space feature matrices can be better constructed later, and the processed data can be used as traffic data.
[0026] Step 2: Obtain the input of the DCNN-GRU prediction model based on the traffic data in step 1;
[0027] Step 2.1: Construct the adjacency matrix;
[0028] Assuming there are n nodes in total, create an n×n matrix A. Construct a weighted network topology graph (Network Topology) based on the relationship between network nodes, G Topo =(V Topo ,E Topo ,A); where V Topo is the set of all network nodes, E Topo is the set of edges, i.e., the connection relationship between adjacent network nodes, and A is the adjacency matrix of the network topology graph, with all elements initially set to 0. Traverse the collected topological data, and for each link, determine the two nodes it connects. If there is a fiber link between node i and node j, then A ij and A ji All are set to 1, that is, A ij =A ji =1, otherwise A ij =A ji =0, i,j∈[1,n], n is the number of nodes.
[0029] The degree of the topological graph is expressed as follows. The degree of a node i is denoted as D ii is the number of edges connected to it;
[0030] The link information in the preprocessed data set is used as the attribute feature of the node, and X is used to represent the network topology graph G Topo The feature matrix of the midpoint.
[0031] Step 2.2: Construct the time feature matrix;
[0032] Create an n×t matrix based on the spatial range and time window, and fill in the data at the corresponding position of each node and time step to build a time feature matrix.
[0033] For example, when the number of nodes is 10, the time window is 5, and random data flow is used for filling, the n×t matrix is as follows:
[0034]
[0035] Each row in the matrix represents a node, each column represents a time point, and each element in the matrix represents the traffic of a node at a time point. time It can be expressed as:
[0036] F time =[f1,f2,...,f t ]
[0037] Among them, f t Represents the node flow characteristic value at time step t.
[0038] f t =[ξ1,ξ2,…,ξ n ]
[0039] Among them, ξ n Denotes the eigenvector f t The eigenvalue of node n.
[0040] The time feature matrix F is normalized using the minimum-maximum normalization method. time Normalize and scale all eigenvalues in the matrix to the range of 0-1. In this way, all eigenvalues will be converted to a uniform scale, which is convenient for subsequent data processing and analysis. The normalization formula is as follows:
[0041]
[0042] Among them, ξ i is the original eigenvalue, ξ norm is the normalized eigenvalue, ξ max and min are the maximum and minimum values of the feature in the entire dataset, respectively.
[0043] Step 3: Construct a space-time graph; construct a space-time graph based on the adjacency matrix and time feature matrix obtained in step 2;
[0044] The electric optical communication network is regarded as a spatial-temporal graph, and the connection relationship between nodes constitutes the network edge. There is a series of time series data {R1, R2, …, R t}, then we can construct a series of graphs {G1,G2,…,G t}, each graph G τDescribe the network topology and node status at the corresponding time point, τ∈[1,t].
[0045] The time step is set according to the frequency of traffic data collection, and a separate layer is created for each time step. In each time layer, the corresponding nodes are connected according to the actual network topology, and the weight of the edge is used to reflect the change of traffic in the time dimension. In addition, it is necessary to connect the corresponding nodes in adjacent time layers to show the continuity of the nodes over time, and assign key network indicators such as packet loss rate as attributes to each node and edge, so as to comprehensively analyze the dynamic performance of the network. Use the spatiotemporal feature matrix, that is, the time feature matrix and the spatial feature matrix, that is, the time feature matrix obtained in step 2 and the adjacency matrix representing the spatial features, to represent the spatiotemporal graph input into the prediction model, where the elements of the matrix contain feature information in spatial position and time.
[0046] Step 4: Construct a DCNN-GRU prediction model, which includes a DCNN module and a GRU module; the schematic diagram of the DCNN module and the GRU module is as follows: Figure 2 As shown,
[0047] In this embodiment, the DCNN module mainly includes the following network structure layers:
[0048] Input layer: responsible for determining the dimension of the input features and receiving the spatiotemporal graph, i.e. the spatiotemporal feature matrix, as input;
[0049] DCNN graph convolution layer: Use the topological information of the power optical communication network to learn features and capture the dependencies between nodes. By applying Laplace transform to the input data, it simulates the diffusion process of information on the graph structure;
[0050] Activation function layer: In order to improve the expressiveness of the model, the graph convolution layer is followed by a ReLU (Rectified Linear Unit) activation function to introduce nonlinear elements;
[0051] GRU layer: Gated recurrent unit layer, mainly used to process time series data, can effectively capture the short-term and long-term trends of traffic data and the dependencies between them;
[0052] Fully connected layer: used for dimensionality reduction and integration of features, providing necessary features for prediction tasks and used for final prediction. Activation functions are added between each layer to enhance the nonlinear expression ability of the model;
[0053] Dropout layer: used to regularize the neural network. In order to improve the generalization ability of the model, a dropout layer is added between the fully connected layers. By randomly discarding some nodes during the training process, overfitting can be prevented.
[0054] Output layer: determines the dimension of the model output. In addition, activation functions such as sigmoid can be used to limit the predicted values to a specific range.
[0055] Step 5: Integrate the adjacency matrix into a weighted adjacency matrix and perform feature diffusion, input the DCNN-GRU prediction model to obtain the output temporal and spatial features;
[0056] In this step, the weighted adjacency matrix, i.e., the weighted adjacency matrix, is input into the DCNN module as a key, and diffused through multiple layers of graph convolutional layers to obtain the output spatial features; the temporal feature matrix is input into the GRU module to obtain the output temporal features.
[0057] Step 5.1: Learn the weights of the adjacency matrix;
[0058] Before the features are input into the model, a weight matrix needs to be learned. This implementation uses the Adam (Adaptive Moment Estimation) optimization algorithm to perform adaptive weight adjustment:
[0059] (1) Initial setting: Set the initial value of the weight matrix W and initialize the two cumulative variables m and v to 0, which are used to store the first-order momentum of the gradient (i.e., the exponential moving average of the historical gradient) and the second-order momentum (i.e., the exponential moving average of the square of the gradient), respectively;
[0060] (2) Define the loss function: Use mean squared error (MSE) as the loss function to measure the difference between the prediction and the true value of the current adjacency matrix weight:
[0061]
[0062] Where q is the total number of data points, It is The true value of the data point, It is The predicted value of a data point. is the difference between the true value and the predicted value.
[0063] (3) Parameter setting: Set an appropriate learning rate α and determine the two hyperparameters β1 and β2 of the Adam algorithm. β1 and β2 are used to estimate the exponential decay rates of m and v, respectively, and are usually set to 0.9 and 0.999.
[0064] (4) Iterative update:
[0065] In each iteration, the gradient g of the weight matrix W with respect to the loss function is calculated through forward propagation and back propagation;
[0066] Update the momentum estimate:
[0067] m N =β1×m N-1 +(1-β1)×g N
[0068] v N =β2×v N-1 +(1-β1)×g N 2
[0069] Among them, m N and v N are the variables m and v, g when the number of iterations is N N It is the gradient of the matrix W with respect to the loss function when the number of iterations is N;
[0070] Because m N and v N are initialized to 0, so they will bias the initial estimates, and they are corrected for the bias:
[0071]
[0072]
[0073] in, and is the deviation-corrected m N and v N , and are β1 and β2 for N iterations respectively; apply the modified momentum value to update the weights:
[0074]
[0075] Among them, W N is the weight matrix for N iterations, W N+1 is the weight matrix for the next iteration, α represents the learning rate, and ε is a very small constant added for numerical stability, usually 1×10 -8 .
[0076] (5) Stop condition: When the change in the loss function in several consecutive iterations is less than a preset threshold η, the iteration is stopped.
[0077] Step 5.2: Input the weighted adjacency matrix into the DCNN-GRU prediction model and perform graph convolution operation;
[0078] In this embodiment, a dynamic deep convolutional neural network DCNN is used to extract spatial features, which is different from the traditional spatial temporal graph convolutional network (STGCN). In this embodiment, the original spatial convolution operation is optimized to DCNN operation. With the help of DCNN's unique diffusion mechanism, this embodiment can extract spatial features in the graph G Topo The information can be effectively transmitted in the network, and the information between nodes can be widely diffused, so as to deeply capture the complex dependencies between nodes. Compared with the prior art, the present invention not only makes full use of the structural topological information of the graph, but also combines the characteristic information of the nodes or edges, thereby significantly improving the accuracy of the prediction and solving the problem of insufficient prediction accuracy in the prior art.
[0079] In this implementation, DCNN relies on the topological structure of the power optical communication network to construct the propagation matrix. Through a series of diffusion convolutional layers, each node can obtain updated feature representations from its neighboring nodes, achieving powerful feature extraction and information integration capabilities, laying the foundation for accurate prediction.
[0080] In the DCNN module, the graph structure can be represented by the Laplacian Matrix and further optimized by the Regularized Laplacian Matrix. The Laplacian matrix L is expressed as follows:
[0081] L=DA
[0082] Where D is the degree matrix of the graph and A is the adjacency matrix;
[0083] Regularized Laplace matrix L reg :
[0084]
[0085] Where I is the identity matrix, is the inverse square root of the degree matrix. The elements of this matrix are given by the following formula:
[0086]
[0087] in, Represents the elements of the regularized adjacency matrix, corresponding to node i and node j in the graph, deg(i) and deg(j) are the degrees of nodes i and j, that is, the number of edges directly connecting nodes i and j.
[0088] The Laplace diffusion process is initiated by giving an initial signal x, which can be any feature representation on the graph. The diffusion method can be simulated by the following formula:
[0089]
[0090] Among them, θ is the time parameter that controls the diffusion speed and range, and x′(θ) is the result of the signal diffusion or evolution on the graph after time θ. Represented by the regularized Laplace matrix L reg and diffusion time θ to simulate the diffusion process;
[0091] In this embodiment, the specific process of using the DCNN module to extract spatial features is as follows:
[0092] Known network topology G Topo , whose adjacency matrix is A. The existing weight matrix W is used to weight the adjacency matrix to obtain a weighted adjacency matrix A'. The weighted weighted adjacency matrix A' can better reflect the connection strength and importance between nodes. The convolution operation can be expressed as:
[0093]
[0094] Among them, H (l+1) is the node representation of the l+1th layer, σ is the sigmoid activation function, is the degree matrix after adding the self-loop The inverse square root of represents the adjacency matrix plus self-loops, is the normalized adjacency matrix, which is the key part of graph convolution. (l) is the node representation of the lth layer, W (l) is the weight of the lth layer. In this way, each node can obtain a new feature representation based on its neighboring nodes.
[0095] After being processed by successive graph convolutional layers, the feature representation of nodes becomes richer and more complex, capable of capturing and representing deeper dependencies.
[0096] The mapping after two graph convolutional layers is as follows:
[0097]
[0098] Among them, X represents the feature matrix of the node, A′ is the weighted adjacency matrix, f(X,A′) is the output function of the two graph convolutional networks, σ here is the external activation function sigmoid, which acts on the output of the last layer of the graph convolutional network, and ReLU is the internal activation function used to increase the nonlinear ability of the model. It is a preprocessed weighted adjacency matrix, W0 is the weight matrix learned by the first layer of graph convolution, and W1 is the weight matrix learned by the second layer of graph convolution.
[0099] In this process, the graph convolution operation is closely linked to the Laplace diffusion process, which works together to transform the graph signal, ensuring the effective diffusion and nonlinear transformation of the signal on the graph, thereby extracting features that are useful for prediction tasks.
[0100] Step 5.3: Fusion GRU module processes the time feature matrix to obtain the time features for network traffic prediction;
[0101] Let F time is the time feature matrix, n is the number of nodes, and t is the time step. The features of each node not only contain the information of its spatially adjacent nodes, but also its own historical information at past time points.
[0102] In order to process time series, we first set an initialized hidden state h0 for each node of the sequence, usually set to a zero vector, and define the relevant parameters of GRU, including the reset gate, update gate, and weights and biases of the candidate hidden state that control information retention.
[0103] like Figure 2 As shown, in this implementation, in each time step t of the time series, the following operations are performed:
[0104] (1) Calculate the reset gate r t : Use the hidden state h of the previous time step t-1 and the input u at the current time step t To calculate the reset gate r t , which determines how much past information needs to be ignored;
[0105] r t =σ(W r ·[h t-1 ,u t ]+b r )
[0106] Where σ is the sigmoid activation function, W r represents the weight matrix associated with the reset gate, b r is the bias vector that resets the gate.
[0107] (2) Calculate the update gate z t : Similarly, using h t-1 and u t To calculate the update gate z t , which controls how much previous information will be retained in the current state;
[0108] z t =σ(W z·[h t-1 ,u t ]+b z )
[0109] Where σ is the sigmoid activation function, W z represents the weight matrix associated with the update gate, b z is the bias vector for the update gate.
[0110] (3) Calculate candidate hidden states Combined with the output r of the reset gate t and h t-1 ,u t To calculate the candidate hidden state It contains new information that may be updated at the current time step;
[0111]
[0112] Among them, tanh (hyperbolic tangent function) is the hyperbolic tangent activation function, W is the weight matrix, and b is the bias vector.
[0113] (4) Update the hidden state h t : Use update gate z t To combine the hidden state h of the previous time step t-1 and candidate hidden states Generate a new hidden state h for the current time step t :
[0114]
[0115] Finally, the hidden state h of the last time window observed is used T To predict network traffic, we can obtain the time characteristics of network traffic prediction. This method of processing spatiotemporal features combined with GRU can effectively predict future changes in network traffic, thus playing an important role in traffic management and network planning.
[0116] Step 6: Fuse the temporal features and spatial features obtained in step 5 and output the predicted traffic matrix;
[0117] After completing the spatiotemporal feature extraction, the temporal and spatial features are fused and reduced in dimension through a fully connected layer. This process not only integrates the various features, but also reduces the dimensions, preparing to capture the key patterns in the traffic data. Subsequently, in order to enhance the model's ability to adapt to new data and prevent overfitting, a dropout layer is introduced. This layer randomly discards some neurons during the training process, thereby enhancing the generalization of the model.
[0118] Finally, after processing at the output layer, the predicted traffic data matrix is obtained. This matrix reflects in detail the traffic prediction values of each spatial node in the network over time, where each row of the matrix corresponds to a node, each column corresponds to a time point, and each element represents the traffic prediction value of the corresponding node at the corresponding time point. This output format intuitively shows the distribution of predicted traffic in the network, which is convenient for subsequent traffic analysis and network optimization decisions.
[0119] In this embodiment, step 7 is also included: training and verifying the prediction model; and step 8 is model evaluation and optimization;
[0120] According to the results of step 6, in order to avoid problems such as insufficient flexibility of a single model, an integrated learning strategy is used in training to further improve model performance.
[0121] Step 7.1: Dataset division;
[0122] First, the dataset is divided into training set, validation set and test set with a ratio of 8:1:1 to ensure the independence of the three sets. The training set is used for model training, the validation set is used to adjust the parameters and hyperparameters of the model, and the test set is used to evaluate the final prediction performance of the model.
[0123] Step 7.2: Training optimization;
[0124] In model training, an integrated learning strategy is adopted, combining the advantages of the hybrid spatiotemporal deep learning DCNN-GRU model and the Long Short-Term Memory model. The LSTM long short-term memory network prediction model has flexible memory capabilities and can learn complex patterns in time series data. The combination of the two optimizes the model's response speed to rapid changes in time series data, enabling the model to quickly adjust predictions when traffic changes suddenly.
[0125] During the training process, the DCNN-GRU model is first randomly initialized and the loss function is defined. Here, the mean squared error (MSE) is used as the loss function to quantify the difference between the predicted value and the actual value and guide the optimization direction of the model parameters.
[0126] In this embodiment, the MSE is the average of the sum of squares of the prediction errors at all data points. The smaller the MSE value, the better the prediction effect of the model. When the MSE is 0, it means that the predicted value is completely consistent with the actual value. To ensure consistency in training, the Adam optimization algorithm used in learning the weight matrix is still used in the training of the DCNN-GRU prediction model;
[0127] The Adam optimization algorithm is used to train the model, and hyperparameters such as the network width are set. The loss on the validation set is monitored, and when the loss no longer decreases or decreases slightly, the learning rate is decayed or an early stopping strategy is implemented.
[0128] In this embodiment, the Adam optimization algorithm uses the moving average gradient to calculate the momentum, which helps to accelerate the convergence of the gradient descent in the relevant direction, thereby offsetting high-frequency oscillations.
[0129] When the LSTM model is trained independently, the same data normalization, Adam optimization algorithm, and time step and window size matching the DCNN-GRU prediction model are used.
[0130] Use the training data set for training and monitor the performance on the validation data set. Use the weighted average method to obtain a more robust model performance evaluation. The average performance index of each model on the validation set will be used to calculate the weight of each model in the final ensemble.
[0131] The process of assigning weights using the weighted average method is as follows:
[0132] (1) Evaluate the performance of the prediction models: Evaluate the performance of the two models independently on the same validation set and choose an appropriate metric, such as the root mean squared error (RMSE), to measure the prediction accuracy of each model.
[0133] (2) Calculate weights based on performance indicators: Use RMSE. A lower RMSE indicates better performance. The weight of each model can be set to the inverse of its RMSE. For example, if the RMSE of DCNN-GRU is 0.06 and the RMSE of LSTM is 0.07, the weights can be initially set as: LSTM is
[0134] (3) Normalize the calculated weights: For example, the initial weight of DCNN-GRU is 17 and that of LSTM is 14. The normalized weights are and
[0135] The output results of the two models are weighted averaged according to the assigned weights. After the above steps, the prediction results of the two models and their respective performance weights on the validation set have been obtained. This implementation proposes the following method to calculate the weighted average prediction results:
[0136]
[0137] Among them, the prediction Ais the prediction result of the DCNN-GRU model, predicting B is the prediction result of the LSTM model, ω A is the weight of the DCNN-GRU model, ω B is the weight of the LSTM model, ω A ′ is the standardized weight of the DCNN-GRU model, ω B ′ is the standardized weight of the LSTM prediction model, and the final prediction result is obtained. This weighted average method can balance the advantages of the two models and reduce the prediction error that may exist in a single model. The final prediction result can capture the rapid changes in the time dimension and the complex dependencies in the spatial dimension, thereby achieving higher accuracy in network traffic prediction.
[0138] Step 8: Model evaluation and optimization;
[0139] The root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the prediction performance of the model. The prediction results of the model on the test set are compared with the true values to analyze the causes of the errors, such as data anomalies, model structure, feature selection, etc.
[0140] The model is deployed to the network management system NMS to obtain network link traffic data in real time, perform real-time traffic prediction, and provide a reference for network traffic management and scheduling.
[0141] Based on the feedback from actual applications, the model is further optimized and adjusted, such as adding more features, adjusting the network structure, updating training data, etc.
[0142] like Figures 3 to 5 As shown, compared with the LSTM model and the GRU model, the performance of the method of the present invention in data throughput prediction is closer to the true value, and the root mean square error and mean absolute error on the test set are also smaller.
[0143] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0144] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
Claims
1. A method for predicting the flow rate of power optical communication links based on a hybrid spatiotemporal deep learning model, characterized by: The method is implemented by the following steps: Step 1: Obtain the input traffic data set of the prediction model; Step 2: Use the traffic data set obtained in step 1 to construct an adjacency matrix and a time feature matrix, and construct a spatiotemporal graph based on the adjacency matrix and the time feature matrix; the specific process is: Process the traffic data set described in step 1 to construct an adjacency matrix and a time feature matrix; Construct a weighted network topology graph G based on the relationship between network nodes Topo =(V Topo ,E Topo ,A), where V Topo is the set of all network nodes, E Topo is the set of edges, A is the adjacency matrix of the network topology, and all elements are initially set to 0; traverse the collected topological data, for each link, determine the two nodes it connects, if there is a fiber link between node i and node j, then A ij and A ji All are set to 1, that is, A ij =A ji =1, otherwise A ij =A ji =0, i,j∈[1,n], n is the number of nodes; The degree of node i is denoted as D ii is the number of edges connected to it; The link information in the processed traffic data set is used as the attribute feature of the node, and X is used to represent the network topology graph G Topo The feature matrix of the midpoint; Construct a time feature matrix; the time feature matrix F time It is expressed as: F time =[f1,f2,...,f t ] Among them, f t represents the node traffic feature vector at time step t; f t =[ξ1,ξ2,…,ξ n ] Among them, ξ n is the feature vector f t The eigenvalue of the middle node n; The minimum and maximum normalization method is used to normalize the time feature matrix F time Normalize and scale all eigenvalues in the matrix to the range of 0-1; The specific process of constructing the space-time graph is as follows: The power optical communication network is regarded as a space-time network graph. Each space-time network graph G τ Describe the network topology and node status at the corresponding time point; The time step is set according to the collection frequency of the traffic data set, and an independent time layer is created for each time step. In each time layer, the corresponding nodes are connected according to the actual network topology structure, and the changes in traffic in the time dimension are reflected through the weight of the edge; the corresponding same nodes in adjacent time layers are connected, and the key network indicator of packet loss rate is assigned as an attribute to each node and edge; The time feature matrix and the adjacency matrix used to express the spatial features are used as a spatiotemporal feature matrix, that is, the spatiotemporal graph is represented by the spatiotemporal feature matrix and inputted into the prediction model; Step 3: construct a DCNN-GRU prediction model, use the spatiotemporal graph, i.e., the spatiotemporal feature matrix, as the input of the DCNN-GRU prediction model, obtain spatial features and temporal features through DCNN-GRU prediction, and predict the output traffic data matrix; Step 4: Divide the traffic data set to perform training optimization of the DCNN-GRU prediction model and LSTM model integration; Step 5: Use the weighted average method to integrate the DCNN-GRU prediction model and the LSTM model, output the traffic data matrix in the next time period, and complete the traffic prediction of the power optical communication link.
2. According to claim 1, the method for predicting the flow rate of a power optical communication link based on a hybrid spatiotemporal deep learning model is characterized in that: In step one, the traffic data set includes link traffic data, network topology data, historical traffic data, device status data, and external factor data.
3. The method for predicting power optical communication link flow based on a hybrid spatiotemporal deep learning model according to claim 1 is characterized in that: In step 3, the adjacency matrix is weighted and used as the input of the DCNN module in the DCNN-GRU prediction model, and the spatial features are extracted through diffusion of the multi-layer graph convolution layer in the DCNN module; The temporal feature matrix is used as the input of the GRU module in the DCNN-GRU prediction model, and the temporal features are extracted by the GRU module; The spatial features and temporal features are passed through the fully connected layer and output layer of the DCNN-GRU prediction model to obtain a predicted traffic data matrix.
4. The method for predicting power optical communication link flow based on a hybrid spatiotemporal deep learning model according to claim 3 is characterized in that: The specific process of extracting spatial features is as follows: The adjacency matrix A is weighted by using the weight matrix W to obtain the weighted adjacency matrix A′, and the weighted adjacency matrix A′ is input into the DCNN module to perform a graph convolution operation and output spatial features; The mapping after two graph convolutional layers is as follows: In the formula, X is the feature matrix of the node, A′ is the weighted adjacency matrix, f(X,A′) is the output function of the graph convolutional network, σ is the external activation function sigmoid, which acts on the output of the last layer of the graph convolutional network, and ReLU is the internal activation function. is the preprocessed weighted adjacency matrix, W0 is the weight matrix learned by the first layer of graph convolution, and W1 is the weight matrix learned by the second layer of graph convolution.
5. The method for predicting power optical communication link flow based on a hybrid spatiotemporal deep learning model according to claim 3 is characterized in that: The specific process of obtaining time features is as follows: First, an initialized hidden state h0 is set for each node of the time series, and the relevant parameters of GRU are defined, including the reset gate, update gate, and weights and biases of the candidate hidden states that control information retention; At each time step t of the time series, the following operations are performed by the GRU module: Calculate the reset gate r t : Use the hidden state h of the previous time step t-1 and the input u of the current time step t Calculate the reset gate r t ; r t =σ(W r ·[h t-1 ,u t ]+b r ) Among them, σ is the sigmoid activation function, W r represents the weight matrix associated with the reset gate, b r is the bias vector for resetting the gate; Calculate the update gate z t : Use h t-1 and u t Calculate the update gate z t ; z t =σ(W z ·[h t-1 ,u t ]+b z ) Where σ is the sigmoid activation function, W z represents the weight matrix associated with the update gate, b z is the bias vector of the update gate; Compute candidate hidden states Combined with the output r of the reset gate t and h t-1 ,u t Compute candidate hidden states Among them, tanh is the hyperbolic tangent activation function, W is the weight matrix, and b is the bias vector; Update hidden state h t : Use update gate z t Combined with the hidden state h of the previous time step t-1 and candidate hidden states Generate a new hidden state h for the current time step t : Finally, the hidden state h of the last time window observed is used T Make predictions on network traffic and obtain the time characteristics of network traffic prediction.
6. The method for predicting power optical communication link flow based on a hybrid spatiotemporal deep learning model according to claim 1 is characterized in that: In step 4, the DCNN-GRU prediction model is randomly initialized and the mean square error (MSE) is used as the loss function to quantify the difference between the predicted value and the actual value; the Adam optimization algorithm is used for training; The LSTM model was trained using data normalization, the Adam optimization algorithm, and a time step and window size that matched the DCNN-GRU prediction model.
7. The method for predicting power optical communication link flow based on a hybrid spatiotemporal deep learning model according to claim 1 is characterized in that: In step 5, the weighted average method is used to calculate the prediction results of the DCNN-GRU prediction model and the LSTM model, which can be expressed as follows: In the formula, prediction A is the prediction result of the DCNN-GRU prediction model, prediction B is the prediction result of the LSTM model, ω A is the weight of the DCNN-GRU prediction model, ω B is the weight of the LSTM model, ω A ′ is the standardized weight of the DCNN-GRU prediction model, ω B ′ is the standardized weight of the LSTM model.
8. The method for predicting power optical communication link flow based on a hybrid spatiotemporal deep learning model according to claim 1 is characterized in that: In step five, the root mean square error and mean absolute error are used to evaluate the prediction performance of the DCNN-GRU prediction model and the LSTM model.
