A smart grid fusion network traffic prediction method based on GNN and LSTM

Abstract

The present invention discloses a method for predicting the flow of a smart grid fusion network based on GNN and LSTM, which relates to the field of smart grid technology and includes the following steps: first, the smart grid processing center collects characteristic data related to terminal and 5G transmission through a power wireless private network, including the current required bandwidth, actual bandwidth, number of connected terminals, quality of service, coordinate position distribution of terminals, coordinate position of base stations, and grouping of base stations; secondly, taking the flow characteristics of base stations as input, a network flow prediction model based on GNN and LSTM is designed, and the model consists of a first part LSTM recursive encoding, a second part GNN spatial context encoding, and a third part LSTM decoding. The method for predicting the flow of a smart grid fusion network based on GNN and LSTM uses GNN and LSTM to complete the flow prediction of the fusion network, which can provide a basis for resource allocation and network optimization of smart grids and improve production efficiency in the power industry.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to a smart grid fusion network traffic prediction method based on GNN and LSTM. Background Art

[0002] Due to its high speed, large capacity, and ultra-low latency, 5G can not only meet the personalized services of individual users, but also meet the differentiated business needs of vertical industries such as smart manufacturing, smart medical care, smart cities, and smart grids. In the 5G vertical industry, the power industry, as an indispensable basic industry for the national economy and people's livelihood, is highly compatible with 5G. For example, 5G's ultra-high reliability and low-latency information transmission characteristics will play an important role in power grid control services such as distribution automation, precise load control, and power differential protection. However, the current 5G network signal coverage is insufficient, the signal penetration is weak, and the traffic charges are too high, which limits the application of 5G networks in power scenarios. The existing power wireless private network has the characteristics of high security and low signal coverage cost. From the perspective of signal coverage, the power wireless private network effectively fills the 5G network signal coverage blind spot. In view of the current situation, a more effective implementation solution is to integrate 5G and power wireless private networks, which can not only save network construction and use costs, but also meet the agile response and precise control needs of massive resources in the integration of "power supply, power grid, load, and energy storage".

[0003] In the converged network, since the 5G public network belongs to the telecommunications operator, the smart grid needs to adopt advanced network management technologies and means to ensure the communication quality of the power business. Among them, the prediction of the base station traffic of the telecommunications 5G public network is of great significance, which can provide a basis for resource allocation and network optimization. For example, when the smart grid processing center predicts that the traffic of the 5G base station will exceed the threshold at the next time point, it can reduce the code flow or enter the silent state of the low-priority business according to the priority of the smart grid terminal business, or even switch to the power private network for business data transmission, so as to ensure the communication quality of the high-priority power business in the 5G bearer network. Predicting the traffic of the 5G public network belongs to the integration of the application layer. The smart grid terminal has dual-mode communication capabilities, that is, it supports the access and communication of wireless private networks and 5G public networks at the same time. The smart grid processing center collects the status information of the terminal through the power private network, including the current required bandwidth of the base station, the actual bandwidth, the number of connected terminals, the service quality, the coordinate location distribution of the terminal, the coordinate location of the base station, and the grouping of the base station. The smart grid processing center statistically analyzes the terminal information collected by the power private network in units of base stations and uses an algorithm model for prediction.

[0004] In order to predict the traffic of base stations, the paper "I. Alawe, A. Ksentini, Y. Hadjadj-Aoul and P. Bertin,"Improving Traffic Forecasting for 5G Core Network Scalability: A Machining Learning Approach," in IEEE Network, vol. 32, no. 6, pp. 42-49, 2018." proposed a method to predict traffic through machine learning technology, and then determine the scale of the 5G core network. It compares the performance of two neural network-based technologies, long short-term memory network (LSTM) and deep learning neural network (DNN), in traffic monitoring. The paper "Y.Fu and X.Wang,"Traffic Prediction-Enabled Energy-EfficientDynamic Computing Resource Allocation in CRAN Based on Deep Learning,"in IEEEOpen Journal of the Communications Society,vol.3,pp.159-175,2022." proposes a method to predict wireless traffic based on a two-dimensional CNN-LSTM model based on time aggregation, which processes traffic data as an image to extract the spatial correlation between wireless network traffic. The paper "G.Jang,N.Kim,T.Ha,C.Lee and S.Cho,"Base Station Switching and Sleep Mode Optimization With LSTM-Based UserPrediction,"in IEEE Access,vol.8,pp.222711-222723,2020." proposes to predict user traffic in base stations in future short time slots based on a long short-term memory (LSTM) model, thereby reducing the overhead generated by information synchronization between users and base stations and minimizing the power consumption of base stations.

[0005] Compared with traditional shallow algorithms, the above-mentioned traffic prediction methods based on deep learning have significantly improved the prediction accuracy. However, they are all aimed at the traffic prediction of a single base station. By using the historical values ​​of the time series of a single base station as training data, the temporal correlation is obtained. However, for 5G, since the coverage range of a base station is usually only about 300 meters, the traffic of adjacent base stations has obvious correlation. Currently, only a few traffic prediction methods consider spatial correlation.

[0006] The paper "C. Zhang, H. Zhang, D. Yuan and M. Zhang," Citywide Cellular Traffic Prediction Based on Densely Connected Convolutional Neural Networks," in IEEE Communications Letters, vol. 22, no. 8, pp. 1656-1659, 2018" divides the area into fixed grids and then uses a two-dimensional convolutional neural network to predict traffic. Compared with traditional time series models, it can better capture the dynamic characteristics of traffic from a spatial perspective. However, in a converged network, 5G base stations only appear in some locations, and it is obviously not suitable to describe them using an image-like data structure. Summary of the invention

[0007] The purpose of the present invention is to provide a smart grid fusion network traffic prediction method based on GNN and LSTM, so as to solve the problem that the existing method for traffic prediction proposed in the above background technology only uses the historical traffic feature information of a single base station, and cannot utilize the traffic correlation between neighboring base stations, resulting in low prediction accuracy.

[0008] To achieve the above object, the present invention provides the following technical solution: a method for predicting network traffic of a smart grid fusion network based on GNN and LSTM, comprising the following steps:

[0009] Step 1: Feature description;

[0010] Step 1.1: Set the length, width and base station coverage of the area covered by the smart grid;

[0011] Step 1.2: Collect the current required bandwidth, actual bandwidth, number of connected terminals, service quality, coordinate position distribution of terminals, coordinate position of base stations, and grouping of base stations to form a feature vector.

[0012] Step 2: Design the first part of the network traffic prediction model based on GNN and LSTM, LSTM recursive encoding. The specific steps are as follows:

[0013] Step 2.1: The input of the entire neural network is The maximum number of base stations in each group is N, and T is the length of the feature sequence. The feature vector of each time step is implied into a high-dimensional embedding space through linear transformation, with an input of 17 dimensions and an output of 512 dimensions. t To express;

[0014] Step 2.2: The current neuron receives the output h of the previous neuront-1 and the high-dimensional embedding x of the base station at the current moment t , calculate the forget gate G forget Select the information to be forgotten, output gate G output Select the information to be remembered and the temporary cell state

[0015] G forget =σ(W f [h t-1 ,x t-1 ]+b f )

[0016] G update =σ(W u [h t-1 ,x t-1 ]+b u )

[0017] G output =σ(W o [h t-1 ,x t-1 ]+b o )

[0018]

[0019] Among them, σ is the Sigmoid activation function, and tanh is also the activation function;

[0020] Step 2.3: Calculate the cell state at the current moment and the output h of the neuron at the current moment t :

[0021]

[0022] h t =G output *tanh(C t ).

[0023] Step 3: Design the second part of the network traffic prediction model based on GNN and LSTM, GNN spatial context encoding. The specific steps are as follows:

[0024] Step 3.1: The graph consists of nodes and edges, symbolically represented by G = {V, E}, where is the set of base stations in the group, E = {e i,j} is the set of edges, when v i and v j When they are directly adjacent, add an edge e between the two base stations. i,j ;

[0025] Step 3.2: Alternately update the features of edges and nodes so that information can be transmitted in the graph. Let the total number of iterations be K. In the kth step, connect v i and v j The edge i,j The feature update is

[0026]

[0027] in f edge is a multilayer perceptron, which updates the edge information according to the node information in turn; when k = 1, the feature vector v of each node (1) is the output h of LSTM t ;

[0028] Step 3.3: In the k+1th step, update the node features based on the original features of the node and the average features of the edges adjacent to the node, as follows:

[0029]

[0030] f node It is also a multi-layer perceptron. After K iterations, it outputs the characteristic information of each base station.

[0031] Step 4: Design the third part of the network traffic prediction model based on GNN and LSTM, LSTM decoding. The specific steps are as follows:

[0032] For each base station of the GNN output, the implicit feature vector It is input into LSTM for decoding, and a double-layer LSTM structure is used to output the predicted traffic of each base station at each time step.

[0033] Step 5: Train and test the model. The specific steps are as follows:

[0034] Step 5.1: Set the learning rate of training to 0.01;

[0035] Step 5.2: Use the mean square error (MSE) between the traffic predicted by the base station and the actual traffic as the learning error function, that is,

[0036]

[0037] where b i is the actual base station traffic, is the predicted base station traffic;

[0038] Step 5.3: Use Adam optimizer to optimize network parameters and gradually reduce the loss value.

[0039] Step 6: System deployment and application, including the following steps:

[0040] Step 6.1: Collect characteristic data of base stations in the area;

[0041] Step 6.2: Input the feature data into the trained network traffic prediction model based on GNN and LSTM;

[0042] Step 6.3: Output the final traffic prediction results for each base station.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. This smart grid fusion network traffic prediction method based on GNN and LSTM uses the power wireless private network to collect terminal 5G transmission-related feature data, including terminal location distribution, carried service type, bandwidth requirements, actual bandwidth, and communication quality. Compared with general methods, the feature information is richer and more complete, thus laying the foundation for obtaining accurate traffic prediction results.

[0045] 2. The smart grid fusion network traffic prediction method based on GNN and LSTM first uses the recurrent neural network LSTM to model the time series characteristics, and then uses the graph neural network to fuse the information of the hidden layer of LSTM encoding, so as to effectively capture the time and spatial correlation of base station traffic. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the process of the present invention;

[0047] Figure 2 This is a schematic diagram of the integration of the power private network and the 5G public network of the present invention;

[0048] Figure 3 This is a schematic diagram of the feature encoding network structure based on GNN and LSTM of the present invention;

[0049] Figure 4 Schematic diagram of the LSTM recursive encoding network structure of the present invention;

[0050] Figure 5 This is a schematic diagram of the LSTM network structure of the base station traffic prediction decoding end of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0052] like Figure 1As shown, the present invention provides a smart grid fusion network traffic prediction method based on GNN and LSTM, the method comprising the following steps:

[0053] Step 1: Feature and problem description;

[0054] Step 1.1: Schematic diagram of the integration of power private network and 5G public network is as follows Figure 2 As shown, the length, width and base station coverage of the area covered by the smart grid are set;

[0055] Step 1.2: Collect the current required bandwidth, actual bandwidth, number of connected terminals, service quality, terminal coordinate location distribution, base station coordinate location, and base station grouping of the base station. For the service quality, it is determined according to the ratio of the base station bandwidth to the actual required bandwidth. For the terminal coordinate location distribution, the number of terminals within 1000 meters is counted at intervals of 100, for a total of 10 dimensions. The collected feature data is merged to obtain a feature vector f with a length of 17, that is, f∈R 17 ;

[0056] Step 1.3: Define the traffic prediction of a single base station. Given a feature sequence F of a single base station with a length of T, predict the required traffic of the base station in the next time step. The problem is defined as:

[0057]

[0058] Among them, the single base station characteristic sequence with a length of T is defined as F = {f 1 ,f 2 ,…,f T};

[0059] Step 1.4: Define group base station traffic prediction. In a converged network, 5G base stations only cover part of the area. Within the coverage area, 5G base stations are usually deployed in groups, and the base stations in the group are adjacent to each other to complete the coverage of the entire local area. Given the characteristics of N base stations sampled simultaneously at the same sampling interval, the goal is to predict the traffic demand for these N base stations in the next S time steps, which is defined as:

[0060]

[0061] Step 2: If Figure 3 As shown in the figure, the first part of the network traffic prediction model based on GNN and LSTM is designed, LSTM recursive encoding, and the specific steps are as follows:

[0062] Step 2.1: The input of the entire neural network is The maximum number of base stations in each group is N. The length of the feature vector corresponding to each base station at a sampling point is 17 dimensions. First, the feature vector of each time step is implicitly projected into a high-dimensional embedding space to describe the relationship between the vectors in the feature vector. This process is implemented through linear transformation. The input is 17 dimensions and the output is 512 dimensions. t To express;

[0063] Step 2.2: If Figure 4 As shown, the high-dimensional feature embedding of each base station at each time step is x t , input to the LSTM module, through the forget gate, update gate and output gate and a memory unit, the fusion information of the sampling point at this moment and the historical sampling point is obtained. The neuron at the current moment receives the output h of the neuron at the previous moment. t-1 and the high-dimensional embedding x of the base station at the current moment t , calculate the forget gate G forget Select the information to be forgotten, output gate G output Select the information to be remembered and the temporary cell state

[0064] G forget =σ(W f [h t-1 ,x t-1 ]+b f )

[0065] G update =σ(W u [h t-1 ,x t-1 ]+b u )

[0066] G output =σ(W o [h t-1 ,x t-1 ]+b o )

[0067]

[0068] Among them, σ is the Sigmoid activation function, and tanh is also the activation function;

[0069] Step 2.3: Calculate the cell state at the current moment and the output h of the neuron at the current moment t :

[0070]

[0071] h t =G output *tanh(C t )

[0072] A 2-layer LSTM structure is used, and the hidden layer of LSTM is 512-dimensional, which integrates the features of the base station time context and the final output h t It is 512 dimensions.

[0073] Step 3: Design the second part of the network traffic prediction model based on GNN and LSTM, GNN spatial context encoding. The specific steps are as follows:

[0074] Step 3.1: After obtaining the temporal context coding h of each base station t After that, since the traffic of adjacent base stations has a strong correlation, each base station is regarded as a node in the graph, and the adjacent base stations are connected through edges. Then, the GNN model is used to complete the message flow between nodes, so that the characteristics of the base station time context can be further integrated with the spatial context information. Assume that the graph consists of nodes and edges, symbolically represented by G = {V, E}, where is the set of base stations in the group, E = {e i,j} is the set of edges, when v i and v j When they are directly adjacent, add an edge e between the two base stations. i,j ;

[0075] Step 3.2: Alternately update the features of edges and nodes so that information can be transmitted in the graph. Let the total number of iterations be K. In the kth step, connect v i and v j The edge i,j The feature update is

[0076]

[0077] in f edge is a multi-layer perceptron, which updates the edge information according to the node information. When k = 1, the feature vector v of each node (1) is the output h of LSTM t ;

[0078] Step 3.3: In the k+1th step, update the node features based on the original features of the node and the average features of the edges adjacent to the node, as follows:

[0079]

[0080] f node It is also a multi-layer perceptron. After K iterations, it outputs the characteristic information of each base station. In the implementation edge and f nodeBoth use a three-layer linear transformation followed by a batch normalization structure, and the empirical value of K is 3.

[0081] Step 4: Design the third part of the network traffic prediction model based on GNN and LSTM, LSTM decoding, such as Figure 5 As shown, the steps are as follows:

[0082] For each base station of the GNN output, the implicit feature vector It is input into LSTM for decoding to output the predicted traffic at each time step. A two-layer LSTM structure is adopted with 512 units.

[0083] Step 5: Train and test the model;

[0084] Step 5.1: Set the learning rate of training to 0.01;

[0085] Step 5.2: Use the mean square error (MSE) between the traffic predicted by the base station and the actual traffic as the learning error function, that is,

[0086]

[0087] where b i is the actual base station traffic, is the predicted base station traffic;

[0088] Step 5.3: Use Adam optimizer to optimize network parameters and gradually reduce the loss value.

[0089] Step 6: System deployment and application, including the following steps:

[0090] Step 6.1: Collect characteristic data of base stations in the area;

[0091] Step 6.2: Input the feature data into the trained network traffic prediction model based on GNN and LSTM;

[0092] Step 6.3: Output the final traffic prediction results for each base station.

[0093] Example

[0094] Assume that the length and width of the area covered by the smart grid are 10,000 meters, the coverage range of the base station is 300 meters, the bandwidth of the base station is 1G, the number of base stations in each group is evenly distributed, and the range is (9,49). These base stations are deployed continuously to completely cover the area. In the entire area, there are 15 groups of base stations. There are three types of terminals in the area. The first type is the mobile terminal type. The mobile terminal conforms to the RWP model. Referring to the flight speed of the drone, the terminal's mobile speed V ranges from 10 to 25 meters per second. The maximum operating time Pmax of the terminal in a place is 1 to 10 minutes. V, Pmax, uniform sampling, the mobile terminal mainly carries inspection services, and the bandwidth required for each terminal is 10M. There are 2,000 mobile terminals in the coverage area. The second type is the video surveillance service of the fixed terminal. The required bandwidth of each terminal is 10M. There are 20 such terminals in the system. The third type is the acquisition terminal, which collects data once every 10 minutes. There are 200 such terminals in the environment, and the bandwidth required for each acquisition is 1M.

[0095] Each terminal sends 5G communication data to the smart grid processing center every 1 minute, including the currently connected base station. The smart grid processing center performs statistics on the base station as a unit to obtain the feature vector of each base station. These data specifically include the current required bandwidth of the base station, the actual bandwidth, the number of connected terminals, the quality of service, the coordinate location distribution of the terminal, the coordinate location of the base station, and the grouping of the base station. For the quality of service, it is determined according to the ratio of the bandwidth of the base station to the actual required bandwidth. For the location distribution of the terminal, the number of terminals within 1000 meters is counted at intervals of 100, and the dimension of the feature vector of each base station is finally obtained to be 17.

[0096] After multiple samplings, the training samples and test samples are divided according to a certain ratio. In order to verify the effectiveness of the method, it is compared with the current mainstream time series learning algorithms, namely Linear RNN, LSTM, and Seq2Seq LSTM. Among them, the linear RNN directly uses linear functions to model the historical state and current input. For Linear RNN and LSTM, the state h corresponding to the last time step t that encapsulates historical information is directly converted to t , decoded into the predicted traffic of the next four consecutive time steps. Seq2Seq LSTM is also based on LSTM. Based on the same training set and test data, the prediction of the traffic of the next four time steps for each method is obtained. Table 1 shows the comparison of the prediction error results of the above three methods and the method proposed in this invention. It can be seen from Table 1 that the prediction result of the method proposed in this invention is significantly improved compared with other methods.

[0097] Table 1 Bandwidth errors of four time steps predicted by different methods

[0098] algorithm Step 1 Step 2 Step 3 Step 4 Linear RNN 19.13 22.53 65.22 80.09 LSTM 13.01 15.41 41.33 48.53 Seq2Seq-LSTM 11.78 13.42 35.77 39.29 GNN-LSTM 8.39 10.02 29.11 31.05

[0099] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A smart grid fusion network traffic prediction method based on GNN and LSTM, characterized in that: Follow these steps: Step 1: The smart grid processing center collects characteristic data related to the terminal and 5G transmission through the power wireless private network. The characteristic data includes the current required bandwidth of the base station, the actual bandwidth, the number of connected terminals, the quality of service, the coordinate position distribution of the terminal, the coordinate position of the base station, and the grouping of the base station; Step 2: Taking the traffic characteristics of the base station as input, a network traffic prediction model based on GNN and LSTM is designed. The model consists of the first part LSTM recursive encoding, the second part GNN spatial context encoding and the third part LSTM decoding. The LSTM recursive encoding includes the following steps: Step 2.11: The input of the entire neural network is The maximum number of base stations in each group is N, T is the length of the feature sequence, and the feature vector of each time step is implicitly mapped to the high-dimensional embedding space x through linear transformation. t , the input is 17 dimensions and the output is 512 dimensions; Step 2.12: The current LSTM neuron receives the output h of the neuron at the previous moment t-1 and the high-dimensional embedding x at the previous moment t-1 , calculate the information G that the forget gate selects to be forgotten forget , the output gate selects the information to be memorized G output With temporary cell state G forget =σ(W f [h t-1 ,x t-1 ]+b f ) G output =σ(W o [h t-1 ,x t-1 ]+b o ) Among them, σ is the Sigmoid activation function, and tanh is also the activation function; Step 2.13: Calculate the cell state at the current moment and the output h of the neuron at the current moment t : G update =σ(W u [h t-1 ,x t-1 ]+b u ) h t =G output *tanh(C t ); GNN spatial context encoding includes the following steps: Step 2.21: The graph is composed of nodes and edges, symbolically represented by G = {V, E}, where is the set of base stations in the group, E={e i，j } is the set of edges, when v i and v j When they are directly adjacent, add an edge e between the two base stations. i，j ; Step 2.22: Alternately update the features of edges and nodes so that information can be transmitted in the graph. Let the total number of iterations be K. In the kth step, connect v i and v j The edge i，j The feature update is in, f edge is a multi-layer perceptron; Step 2.23: In the k+1th step, update the node features based on the original features of the node and the average features of the edges adjacent to the node, as follows: f node It is also a multi-layer perceptron. After K iterations, it outputs the characteristic information of each base station. LSTM decoding is as follows: For the GNN output, the feature information of each base station It is input into LSTM for decoding, and a double-layer LSTM structure is used to output the predicted traffic of each base station at each time step. Step 3: Train the network traffic prediction model based on GNN and LSTM, test it on the test set, and deploy the trained model to the smart grid processing center; Step 4: Finally, the characteristic data of the base stations in the area are collected and input into the trained network traffic prediction model to obtain the final traffic prediction results.

2. According to claim 1, a smart grid fusion network traffic prediction method based on GNN and LSTM is characterized by: The collection of feature data in step 1 includes the following steps: Step 1.1: Set the length and width of the area covered by the smart grid and the base station coverage; Step 1.2: Collect the current required bandwidth, actual bandwidth, number of connected terminals, service quality, coordinate position distribution of terminals, coordinate position of base stations, and grouping of base stations to form a feature vector f.

3. According to the method for predicting network traffic of smart grid fusion based on GNN and LSTM in claim 1, it is characterized by: Training and testing the model in step 3 includes the following steps: Step 3.1: Set the learning rate of training to 0.01; Step 3.2: Use the mean square error (MSE) between the traffic predicted by the base station and the actual traffic as the learning error function, that is, where b i is the actual base station traffic, is the predicted base station traffic; Step 3.3: Use Adam optimizer to optimize network parameters and gradually reduce the loss value.

4. According to the method for predicting network traffic of smart grid fusion based on GNN and LSTM in claim 1, it is characterized in that: The flow prediction in step 4 specifically includes the following steps: Step 4.1: Collect characteristic data of base stations in the area; Step 4.2: Input the feature data into the trained network traffic prediction model based on GNN and LSTM; Step 4.3: Output the final traffic prediction results for each base station.

Images