Model training method, apparatus and device

Abstract

This invention provides a model training method, apparatus, and device. The method includes: acquiring multiple sets of sample communication features corresponding to multiple sample base stations; combining the multiple sets of sample communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor; determining multiple training samples based on the three-dimensional communication feature tensor and each acquisition time; determining a base station spatial topology map in the communication feature space of each sample based on the three-dimensional communication feature tensor and the multiple sample communication features; and training an initial graph-spatiotemporal model based on the multiple training samples and the base station spatial topology map to obtain a communication traffic prediction model. The model training method, apparatus, and device provided by this invention are used to improve the accuracy of communication traffic prediction models.

Description

Technical Field

[0001] This invention relates to the fields of communication technology and communication traffic prediction technology, and in particular to a model training method, apparatus and equipment. Background Technology

[0002] Traffic forecasting is a key component of communication network management. Accurate traffic forecasting can help operators improve service quality, alleviate network congestion, and reduce operating expenses.

[0003] In related technologies, a recurrent neural network (RNN) can be trained based on a sequence of communication traffic characteristics at different acquisition times to obtain a communication traffic prediction model; and communication traffic can be predicted based on the communication traffic prediction model.

[0004] In the aforementioned related technologies, training an RNN with a sequence of communication traffic features collected at different times results in a poor accuracy of the obtained communication traffic prediction model. Summary of the Invention

[0005] This invention provides a model training method, apparatus, and device to address the shortcomings of poor accuracy in existing communication traffic prediction models and to improve the accuracy of communication traffic prediction models.

[0006] In a first aspect, the present invention provides a model training method, comprising:

[0007] Obtain multiple sets of sample communication features corresponding to each of the multiple sample base stations; wherein each set of sample communication features includes multiple sample communication features;

[0008] Based on the acquisition time of each of the multiple sets of sample communication features, the multiple sets of sample communication features are combined to obtain a three-dimensional communication feature tensor.

[0009] Based on the three-dimensional communication feature tensor and each acquisition time, multiple training samples are determined;

[0010] Based on the three-dimensional communication feature tensor and the multiple sample communication features, a base station spatial topology map is determined in the communication feature space of each sample.

[0011] Based on the multiple training samples and the base station spatial topology map, the initial graph spatiotemporal model is trained to obtain the communication traffic prediction model.

[0012] According to a model training method provided by the present invention, determining multiple training samples based on the three-dimensional communication feature tensor and each acquisition time includes:

[0013] For each acquisition time, the time slice corresponding to the acquisition time in the three-dimensional communication feature tensor is determined as the training sample corresponding to the acquisition time; wherein, the time slice includes a set of sample communication features of each sample base station at the acquisition time.

[0014] According to a model training method provided by the present invention, determining the base station spatial topology map in the communication feature space of each sample based on the three-dimensional communication feature tensor and the multiple sets of sample communication features includes:

[0015] From the three-dimensional communication feature tensor, feature slices corresponding to the communication features of each sample are obtained; wherein, the feature slices include the communication features of each sample base station at each acquisition time;

[0016] Based on the sample communication characteristics of any two sample base stations among the plurality of sample base stations at each collection time, determine the edge weight between the two sample base stations;

[0017] Based on the edge weights between any two sample base stations, a base station spatial topology graph is constructed under the communication feature space of each sample.

[0018] According to a model training method provided by the present invention, the two sample base stations include a first sample base station and a second sample base station;

[0019] Based on the sample communication characteristics of the two sample base stations at each collection time, the edge weight between the two sample base stations is determined, including:

[0020] pass The sample communication features of the first sample base station and the sample communication features of the second sample base station at each collection time are processed to obtain the edge weights between the first sample base station and the second sample base station.

[0021] in, The edge weight between the first sample base station and the second sample base station is represented by k, which represents the sample communication feature, and X is the edge weight between the first sample base station and the second sample base station. k Y represents the vector composed of the sample communication features of the first sample base station at each acquisition time. k This represents a vector composed of the sample communication features of the second sample base station at each acquisition time. X represents k standard deviation Y represents k Standard deviation, COV(X) k ,Y k ) represents X k and Y k The covariance between them.

[0022] According to a model training method provided by the present invention, the two sample base stations include a first sample base station and a second sample base station;

[0023] Based on the sample communication characteristics of the two sample base stations at each collection time, the edge weight between the two sample base stations is determined, including:

[0024] By using D(X, Y) = min P ∑ (i，j)∈P d(x i y j The sample communication features of the first sample base station and the sample communication features of the second sample base station at each collection time are processed to obtain the edge weights between the first sample base station and the second sample base station.

[0025] Where D(X, Y) represents the edge weight between the first sample base station and the second sample base station, X represents the vector composed of the sample communication features of the first sample base station at each collection time, Y represents the vector composed of the sample communication features of the second sample base station at each collection time, P represents the set of index pairs composed of the indices of elements in X and the indices of elements in Y, (i, j) represents an index pair in P, x i Let y represent the i-th element in X. j Min represents the j-th element in Y. P · indicates the minimum value operation, d(x) i y j ) represents x i With y j The degree of difference between them, ∑· represents the summation operation.

[0026] According to a model training method provided by the present invention, the initial graph spatiotemporal model includes an initial graph attention network and an initial gated recurrent unit;

[0027] The step of training an initial graph spatiotemporal model based on the multiple training samples and the base station spatial topology map to obtain a communication traffic prediction model includes:

[0028] The multiple training samples and the base station spatial topology map under each communication feature space are input into the initial graph attention network to obtain the predicted features of each sample base station;

[0029] The predicted features of each sample base station are input into the initial gated loop unit to obtain the communication traffic prediction results corresponding to each of the multiple sample base stations;

[0030] Based on the communication traffic prediction results and tag communication traffic corresponding to each of the multiple sample base stations, the model parameters in the initial graph attention network and the initial gated recurrent unit are updated to obtain the communication traffic prediction model.

[0031] Secondly, the present invention provides a communication traffic prediction method, comprising:

[0032] Obtain multiple sets of communication features corresponding to each of the multiple base stations; where each set of communication features includes multiple communication features.

[0033] Based on the acquisition time of each of the multiple sets of communication features, the multiple sets of communication features are combined to obtain a three-dimensional communication feature tensor;

[0034] Based on the three-dimensional communication feature tensor and the multiple communication features, a base station spatial topology map is determined in each communication feature space;

[0035] The multiple communication features and the base station spatial topology map are input into the communication traffic prediction model to obtain the communication traffic prediction results of each of the multiple base stations; wherein, the communication traffic prediction model is the communication traffic prediction model described in the first aspect above.

[0036] Thirdly, the present invention provides a model training apparatus, the apparatus comprising:

[0037] The acquisition module is used to acquire multiple sets of sample communication features corresponding to each of the multiple sample base stations; wherein each set of sample communication features includes multiple sample communication features;

[0038] The combination module is used to combine the multiple sets of sample communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor.

[0039] The determination module is used to determine multiple training samples based on the three-dimensional communication feature tensor and each acquisition time.

[0040] The graph construction module is used to determine the base station spatial topology map in the communication feature space of each sample based on the three-dimensional communication feature tensor and the multiple sample communication features;

[0041] The training module is used to train the initial spatiotemporal model based on the multiple training samples and the base station spatial topology map to obtain a communication traffic prediction model.

[0042] According to a model training apparatus provided by the present invention, the determining module is specifically used for:

[0043] For each acquisition time, the time slice corresponding to the acquisition time in the three-dimensional communication feature tensor is determined as the training sample corresponding to the acquisition time; wherein, the time slice includes a set of sample communication features of each sample base station at the acquisition time.

[0044] According to a model training apparatus provided by the present invention, the graph construction module is specifically used for:

[0045] From the three-dimensional communication feature tensor, feature slices corresponding to the communication features of each sample are obtained; wherein, the feature slices include the communication features of each sample base station at each acquisition time;

[0046] Based on the sample communication characteristics of any two sample base stations among the plurality of sample base stations at each collection time, determine the edge weight between the two sample base stations;

[0047] Based on the edge weights between any two sample base stations, a base station spatial topology graph is constructed under the communication feature space of each sample.

[0048] According to a model training apparatus provided by the present invention, the two sample base stations include a first sample base station and a second sample base station; the graph construction module is specifically used for: through The sample communication features of the first sample base station and the sample communication features of the second sample base station at each collection time are processed to obtain the edge weights between the first sample base station and the second sample base station.

[0049] in, The edge weight between the first sample base station and the second sample base station is represented by k, which represents the sample communication feature, and X is the edge weight between the first sample base station and the second sample base station. k Y represents the vector composed of the sample communication features of the first sample base station at each acquisition time. k σ represents the vector composed of the sample communication features of the second sample base station at each acquisition time. X k represents X k Standard deviation, σ Y k represents Y k Standard deviation, COV(X) k Y k ) represents X k and Y k The covariance between them.

[0050] According to a model training apparatus provided by the present invention, the two base stations include a first sample base station and a second sample base station; the graph construction module is specifically used to: [implement a method] using D(X, Y) = min[…]. P ∑ (i，j)∈P d(xi y j The sample communication features of the first sample base station and the sample communication features of the second sample base station at each collection time are processed to obtain the edge weights between the first sample base station and the second sample base station.

[0051] Where D(X, Y) represents the edge weight between the first sample base station and the second sample base station, X represents the vector composed of the sample communication features of the first sample base station at each collection time, Y represents the vector composed of the sample communication features of the second sample base station at each collection time, P represents the set of index pairs composed of the indices of elements in X and the indices of elements in Y, (i, j) represents an index pair in P, x i Let y represent the i-th element in X. j Min represents the j-th element in Y. P · indicates the minimum value operation, d(x) i y j ) represents x i With y j The degree of difference between them, ∑· represents the summation operation.

[0052] According to a model training apparatus provided by the present invention, the initial graph spatiotemporal model includes an initial graph attention network and an initial gated recurrent unit; the training module is specifically used for:

[0053] The multiple training samples and the base station spatial topology map under each communication feature space are input into the initial graph attention network to obtain the predicted features of each sample base station;

[0054] The predicted features of each sample base station are input into the initial gated loop unit to obtain the communication traffic prediction results corresponding to each of the multiple sample base stations;

[0055] Based on the communication traffic prediction results and tag communication traffic corresponding to each of the multiple sample base stations, the model parameters in the initial graph attention network and the initial gated recurrent unit are updated to obtain the communication traffic prediction model.

[0056] Fourthly, the present invention provides a communication traffic prediction device, the device comprising:

[0057] The acquisition module is used to acquire multiple sets of communication features corresponding to each of the multiple base stations; wherein each set of communication features includes multiple communication features.

[0058] The combination module is used to combine the multiple sets of communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor.

[0059] The graph construction module is used to determine the base station spatial topology map under each communication feature space based on the three-dimensional communication feature tensor and the multiple communication features;

[0060] The prediction module is used to input the multiple communication features and the base station spatial topology map into the communication traffic prediction model to obtain the communication traffic prediction results of each of the multiple base stations; wherein, the communication traffic prediction model is the communication traffic prediction model described in the first aspect above.

[0061] Fifthly, the present invention also provides a training device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the model training method as described in the first aspect above.

[0062] In a sixth aspect, the present invention also provides a central server, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the communication traffic prediction method as described in the second aspect above.

[0063] In a seventh aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the model training method as described in the first aspect and the communication traffic prediction method as described in the second aspect.

[0064] Eighthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the model training method as described in the first aspect and the communication traffic prediction method as described in the second aspect.

[0065] The model training method, apparatus, and equipment provided by this invention acquire multiple sets of sample communication features corresponding to multiple sample base stations, combine the multiple sets of sample communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor, determine multiple training samples based on the three-dimensional communication feature tensor and each acquisition time, determine the base station spatial topology map under the communication feature space of each sample based on the three-dimensional communication feature tensor and the multiple sample communication features, and train the initial graph spatiotemporal model based on the multiple training samples and the base station spatial topology map to obtain a communication traffic prediction model, thereby improving the accuracy of the communication traffic prediction model. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0067] Figure 1 This is a flowchart illustrating the model training method provided in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the three-dimensional communication feature tensor provided in an embodiment of the present invention;

[0069] Figure 3 This is a schematic diagram of a time slice provided in an embodiment of the present invention;

[0070] Figure 4 This is a schematic flowchart of the method for determining the spatial topology map of a base station provided in an embodiment of the present invention;

[0071] Figure 5 This is a schematic diagram of a feature slice provided in an embodiment of the present invention;

[0072] Figure 6 This is a schematic diagram of the method for obtaining a communication traffic prediction model provided in an embodiment of the present invention;

[0073] Figure 7 This is a schematic diagram of the structure of the initial graph spatiotemporal model provided in an embodiment of the present invention;

[0074] Figure 8 This is a flowchart illustrating the communication traffic prediction method provided in an embodiment of the present invention;

[0075] Figure 9 This is a schematic diagram of the structure of the model training device provided in an embodiment of the present invention;

[0076] Figure 10 This is a schematic diagram of the communication traffic prediction device provided in an embodiment of the present invention;

[0077] Figure 11 This is a schematic diagram of the physical structure of the training device provided in an embodiment of the present invention;

[0078] Figure 12 This is a schematic diagram of the physical structure of the central server provided in an embodiment of the present invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0080] In this invention, the term "comprising" and its variations can refer to a non-limiting inclusion; the term "or" and its variations can refer to "and / or". The terms "first", "second", etc., in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. In this invention, "at least one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0081] In related technologies, training an RNN with a sequence of communication traffic features collected at different times results in a poor accuracy of the obtained communication traffic prediction model.

[0082] To improve the accuracy of communication traffic prediction models, this invention provides a model training method. Below, in conjunction with... Figures 1 to 7 The embodiments describe the model training method provided by the present invention.

[0083] Figure 1 This is a flowchart illustrating the model training method provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0084] Step 101: Obtain multiple sets of sample communication features corresponding to each of the multiple sample base stations; wherein each set of sample communication features includes multiple sample communication features.

[0085] Optionally, the model training method provided by this invention can be executed by a training device or a model training apparatus installed on a training device, or it can be a central server or a model training apparatus installed on a central server. The model training apparatus can be implemented through a combination of software and / or hardware. For example, the training device can be a desktop computer, a laptop computer, or a tablet computer.

[0086] The central server and the base station cluster can communicate via wired or wireless networks. Wired networks can include, for example, coaxial cable, twisted pair, and fiber optic cables. Wireless networks can include, for example, Wi-Fi and Bluetooth.

[0087] The base station cluster includes multiple sample base stations. Each sample base station serves its corresponding serving cell. The sample base stations can collect the communication characteristics of the serving cell in real time, obtain multiple sets of sample communication characteristics, and send these sets of characteristics to the central server.

[0088] Optionally, when the execution entity is a training device, the training device can obtain multiple sets of sample communication features corresponding to each of the multiple sample base stations from the central server.

[0089] One sample communication feature corresponds to one category; multiple sample communication features correspond to multiple categories; the multiple categories include at least two of the following:

[0090] The number of times a Radio Resource Control (RRC) connection has been successfully established;

[0091] The number of times an Evolved Radio Access Bearer (E-RAB) connection is successfully established;

[0092] Uplink physical resource block (PRB) utilization;

[0093] Downlink PRB utilization rate;

[0094] Physical Uplink Shared Channel (PUSCH) utilization;

[0095] Average number of users;

[0096] Uplink traffic;

[0097] Downlink traffic.

[0098] Step 102: Based on the acquisition time of each of the multiple sets of sample communication features, combine the multiple sets of sample communication features to obtain the three-dimensional communication feature tensor.

[0099] For each sample base station, the collection times for each set of sample communication features are different, and the time interval between two adjacent collection times is equal to a preset duration. The preset duration can be any duration, such as 15 minutes, 1 hour, 1 day, 1 week, or 1 month. For example, if the number of sample communication features for this sample base station is 3, the collection time for the first set of sample communication features is T1, the collection time for the second set of sample communication features is T2, and the collection time for the third set of sample communication features is T3. The time interval between T1 and T2 is the same as the time interval between T2 and T3.

[0100] Step 103: Based on the three-dimensional communication feature tensor and each acquisition time, determine multiple training samples.

[0101] Step 104: Based on the three-dimensional communication feature tensor and multiple sample communication features, determine the base station spatial topology map under the communication feature space of each sample.

[0102] The spatial topology of the base station is an undirected graph.

[0103] Step 105: Based on multiple training samples and the spatial topology map of the base station, train the initial spatiotemporal model to obtain the communication traffic prediction model.

[0104] exist Figure 1 In this embodiment, based on a three-dimensional communication feature tensor obtained by combining multiple sets of sample communication features and the acquisition times of each set of sample communication features, multiple training samples are determined. These training samples can reflect the dependence of communication traffic on temporal and feature information. Based on the three-dimensional communication feature tensor and the multiple sample communication features, a base station spatial topology map is determined. The base station spatial topology map can reflect the dependence of communication traffic on spatial information. Therefore, based on the base station spatial topology map and the multiple training samples, the initial spatiotemporal model is trained, enabling the obtained communication traffic prediction model to take into account the dependence of spatial, temporal, and feature information, thereby improving the accuracy of the communication traffic prediction model.

[0105] Below, in conjunction with Figure 2 This paper describes the three-dimensional communication feature tensor with dimensions of 3×3×3.

[0106] Figure 2 This is a schematic diagram of the three-dimensional communication feature tensor provided in an embodiment of the present invention. For example... Figure 2 As shown, exemplarily, multiple sample base stations include S1, S2, and S3, multiple collection times include T1, T2, and T3, and each group of sample communication features includes sample communication features corresponding to category A, category B, and category C. Category A can be, for example, uplink traffic, category B can be, for example, downlink traffic, and category C can be, for example, the average number of users.

[0107] For example, the set of sample communication features corresponding to sample base station S1 at time T1 is [A11, B11, C11], the set of sample communication features corresponding to sample base station S2 at time T1 is [A21, B21, C21], and the set of sample communication features corresponding to sample base station S3 at time T1 is [A23, B23, C23].

[0108] In some embodiments, multiple training samples are determined based on the three-dimensional communication feature tensor and each acquisition time, including:

[0109] For each acquisition time, the time slice corresponding to the acquisition time in the three-dimensional communication feature tensor is determined as the training sample corresponding to the acquisition time; wherein, the time slice includes a set of sample communication features of each sample base station at the acquisition time.

[0110] Figure 3 This is a schematic diagram of a time slice provided in an embodiment of the present invention. Figure 2 On the basis of, such as Figure 3 As shown, the time slice corresponding to acquisition time T1 includes:

[0111] A set of sample communication features [A11, B11, C11] of sample base station S1 at acquisition time T1;

[0112] A set of sample communication features of sample base station S2 at acquisition time T1 [A21, B21, C21];

[0113] A set of sample communication features [A31, B31, C31] of sample base station S3 at acquisition time T1.

[0114] The time slice corresponding to acquisition time T2 includes:

[0115] A set of sample communication features [A12, B12, C12] of sample base station S1 at acquisition time T2;

[0116] A set of sample communication features of sample base station S2 at acquisition time T2 [A22, B22, C22];

[0117] A set of sample communication features [A32, B32, C32] of sample base station S3 at acquisition time T2.

[0118] The time slice corresponding to acquisition time T3 includes:

[0119] A set of sample communication features of sample base station S1 at acquisition time T3 [A13, B13, C13];

[0120] A set of sample communication characteristics of sample base station S2 at acquisition time T3 [A23, B23, C23];

[0121] A set of sample communication features [A33, B33, C33] of sample base station S3 at acquisition time T3.

[0122] In this embodiment of the invention, for each acquisition time, the time slice corresponding to the acquisition time in the three-dimensional communication feature tensor is determined as the training sample corresponding to the acquisition time. This allows the communication traffic prediction model obtained by training the initial graph spatiotemporal model with the training sample to capture the dependence of communication traffic on time information and feature information, thereby improving the accuracy of the communication traffic prediction model.

[0123] Figure 4 This is a schematic flowchart of a method for determining a base station spatial topology map provided in an embodiment of the present invention. Figure 4 As shown, the method includes:

[0124] Step 401: Obtain the feature slices corresponding to the communication features of each sample from the three-dimensional communication feature tensor; wherein, the feature slices include the sample communication features of each sample base station at each acquisition time.

[0125] Figure 5 This is a schematic diagram of a feature slice provided in an embodiment of the present invention. Figure 2 On the basis of, such as Figure 5 As shown, the three-dimensional communication feature tensor includes three feature slices: the feature slice corresponding to the communication feature of sample A, the feature slice corresponding to the communication feature of sample B, and the feature slice corresponding to the communication feature of sample C.

[0126] For example, for the communication features of a sample of category A, the corresponding feature slices include:

[0127] Sample communication characteristics of sample base station S1 from collection time T1 to collection time T3 [A11, A12, A13];

[0128] Sample communication characteristics of sample base station S2 from collection time T1 to collection time T3 [A21, A22, A23];

[0129] Sample communication characteristics of base station S3 from acquisition time T1 to acquisition time T3 [A31, A32, A33].

[0130] Step 402: Based on the sample communication characteristics of any two sample base stations at each collection time, determine the edge weight between the two sample base stations.

[0131] In some embodiments, the two sample base stations include a first sample base station and a second sample base station; the edge weight between the first sample base station and the second sample base station can be determined by the following methods 1 and 2.

[0132] Method 1, through The sample communication characteristics of the first sample base station at each collection time and the sample communication characteristics of the second sample base station at each collection time are processed to obtain the edge weight between the first sample base station and the second sample base station.

[0133] in, X represents the edge weight between the first sample base station and the second sample base station, k represents the sample communication feature, and X represents the edge weight between the first sample base station and the second sample base station. k Y represents a vector composed of the sample communication features of the first sample base station at each acquisition time.k This represents a vector composed of the sample communication features of the second sample base station at each acquisition time. X represents k standard deviation Y represents k Standard deviation, COV(X) k ,Y k ) represents X k and Y k The covariance between them.

[0134] In Method 1, the edge weights between two sample base stations are determined by the Pearson correlation coefficient. The Pearson correlation coefficient is a statistic that measures the strength of the linear relationship between the vectors composed of sample communication features of any two sample base stations at each acquisition time. The Pearson correlation coefficient ranges from [-1, 1].

[0135] When there is a perfect positive correlation between the vector composed of the sample communication features of the first sample base station at each acquisition time and the vector composed of the sample communication features of the second sample base station at each acquisition time, the Pearson correlation coefficient is 1; when there is a perfect positive correlation between the vector composed of the sample communication features of the first sample base station at each acquisition time and the vector composed of the sample communication features of the second sample base station at each acquisition time, the Pearson correlation coefficient is -1; when there is a perfect positive correlation between the vector composed of the sample communication features of the first sample base station at each acquisition time and the vector composed of the sample communication features of the second sample base station at each acquisition time, the Pearson correlation coefficient is 0.

[0136] In this embodiment of the invention, based on the Pearson correlation coefficient, the sample communication features of two sample base stations at each collection time are processed to obtain the edge weights between the two sample base stations. Based on the edge weights between any two sample base stations, a base station spatial topology graph under the sample communication feature space is constructed. This allows the communication traffic prediction model obtained by training the initial graph spatiotemporal model based on the base station spatial topology graph to take into account the dependence of communication traffic on spatial information and feature information, thereby improving the accuracy of the communication traffic prediction model.

[0137] Method 2, using D(X,Y) = min P ∑ (i,j)∈P d(x i ,y j The sample communication characteristics of the first sample base station at each collection time and the sample communication characteristics of the second sample base station at each collection time are processed to obtain the edge weight between the first sample base station and the second sample base station.

[0138] Where D(X,Y) represents the edge weight between the first sample base station and the second sample base station, X represents the vector composed of the sample communication features of the first sample base station at each acquisition time, Y represents the vector composed of the sample communication features of the second sample base station at each acquisition time, P represents the set of index pairs composed of the indices of elements in X and the indices of elements in Y, and (i,j) represents an index pair in P. i Let y represent the i-th element in X. j Min represents the j-th element in Y. P · indicates the minimum value operation, d(x) i ,y j ) represents x i With y j The degree of difference between them, ∑· represents the summation operation.

[0139] Optionally, x i With y j The degree of difference between them can be Euclidean distance, Manhattan distance, or other similarity measures.

[0140] For example, the indices of elements in X include 0, 1, and 2; the indices of elements in Y include 0, 1, and 2; and the indices of elements in P include (0,0), (0,1), (0,2), (1,0), (1,1), (1,2), (2,0), (2,1), and (2,2).

[0141] It should be noted that P defines the mapping of elements of X to elements of Y.

[0142] In Method 2, the edge weights between two sample base stations are determined by Dynamic Time Warping (DTW). DTW can calculate the distance between the vectors composed of sample communication features of any two sample base stations at each acquisition time, measure the similarity, time shift, and shape change between the vectors composed of sample communication features of any two sample base stations at each acquisition time, and thus analyze the similarity between the vectors composed of sample communication features of any two sample base stations at each acquisition time and find similar patterns.

[0143] In this embodiment of the invention, based on DTW, the sample communication features of two sample base stations at each collection time are processed to obtain the edge weights between the two sample base stations. Based on the edge weights between any two sample base stations, a base station spatial topology graph under the sample communication feature space is constructed. This allows the communication traffic prediction model obtained by training the initial graph spatiotemporal model based on the base station spatial topology graph to take into account the dependence of communication traffic on spatial information and feature information, thereby improving the accuracy of the communication traffic prediction model.

[0144] Step 403: Based on the edge weights between any two sample base stations, construct the base station spatial topology map under the communication feature space of each sample.

[0145] The nodes in the base station spatial topology map are each sample base station.

[0146] Figure 6 This is a schematic diagram of the method for obtaining a communication traffic prediction model provided in an embodiment of the present invention. Figure 6 As shown, the initial graph spatiotemporal model includes an initial graph attention network (GAT) and an initial gated recurrent unit (GRU). The method includes:

[0147] Step 601: Input multiple training samples and base station spatial topology maps under each communication feature space into the initial graph attention network to obtain the predicted features of each sample base station.

[0148] Below, in conjunction with Figure 7 The initial graph spatiotemporal model provided in the embodiments of the present invention will be described by way of example.

[0149] Figure 7 This is a schematic diagram of the initial graph spatiotemporal model provided in an embodiment of the present invention. For example... Figure 7 As shown, the initial graph spatiotemporal model includes the initial GAT and the initial GRU.

[0150] The initial GAT can calculate the relationship weights between each node in the base station spatial topology graph and use a self-attention mechanism to determine the importance of each node during aggregation.

[0151] The self-attention mechanism is implemented through a feedforward neural network. First, the vector corresponding to each node is mapped to a hidden layer. Then, the softmax() activation function is used to calculate the weights between nodes. Based on the weights, the vectors corresponding to two adjacent nodes are weighted and averaged to obtain a new feature representation for each node.

[0152] In step 601, the graph attention network can handle the relationship between any two nodes in the base station spatial topology graph, and can perform finer-grained control when learning node representations.

[0153] Optionally, in the initial GAT, a multi-head attention mechanism can also be used to process multiple training samples and base station spatial topology maps under each communication feature space.

[0154] The initial GRU includes variables such as hidden states, input gates, and reset gates, which can capture the temporal dependencies of communication features and generate long-term and short-term memories of communication feature sequences. Input gates and reset gates are used to regulate the data flow, controlling how much new input should be added to the previous hidden state and how much of the previous hidden state should be forgotten.

[0155] Step 602: Input the prediction features of each sample base station into the initial gated loop unit to obtain the communication traffic prediction results corresponding to each of the multiple sample base stations.

[0156] Step 603: Based on the communication traffic prediction results and tag communication traffic corresponding to each of the multiple sample base stations, update the model parameters in the initial graph attention network and the initial gated recurrent unit to obtain the communication traffic prediction model.

[0157] For example, in Figure 2 Based on this, with T1 being May 15, 2023, T2 being May 16, 2023, and T3 being May 17, 2023, for each sample base station, the tag communication traffic of the sample base station is a set of actual communication characteristics corresponding to the sample base station on May 18, 2023, and the communication traffic prediction result of the sample base station is a set of predicted communication characteristics obtained by predicting the communication characteristics of the sample base station on May 18, 2023.

[0158] Optionally, the model parameters in the initial graph attention network and the initial gated recurrent unit can be updated based on the error between the communication traffic prediction results corresponding to each of the multiple sample base stations and the tag communication traffic.

[0159] Optionally, the error can be the mean-square error (MSE).

[0160] exist Figure 6 In this embodiment, if a gated recurrent unit is used alone to process multiple training samples and base station spatial topology maps under each communication feature space, matrix multiplication may confuse the independent importance of the input communication features when the input passes through the linear layer of the gated recurrent unit, resulting in the inability to capture spatial dependencies and feature dependencies. Therefore, training a model that includes a graph attention network and a gated recurrent unit can avoid the problem of failing to capture spatial dependencies and feature dependencies, thereby improving the accuracy of communication traffic prediction.

[0161] Figure 8 This is a flowchart illustrating the communication traffic prediction method provided in an embodiment of the present invention. Figure 8 As shown, the method includes:

[0162] Step 801: Obtain multiple sets of communication features corresponding to each of the multiple base stations; wherein each set of communication features includes multiple communication features.

[0163] Optionally, the model training method provided by this invention can be executed by a central server or a communication traffic prediction device located within the central server. The communication traffic prediction device can be implemented through a combination of software and / or hardware.

[0164] Step 802: Based on the acquisition time of each of the multiple sets of communication features, combine the multiple sets of communication features to obtain the three-dimensional communication feature tensor.

[0165] Step 803: Based on the three-dimensional communication feature tensor and multiple communication features, determine the base station spatial topology map under each communication feature space.

[0166] Step 804: Input multiple communication features and base station spatial topology maps into the communication traffic prediction model to obtain the communication traffic prediction results for each of the multiple base stations.

[0167] The communication traffic prediction model is as described above. Figures 1 to 7 The communication traffic prediction model described in any of the embodiments.

[0168] Other related technologies, such as traffic prediction models based on convolutional neural networks (CNNs), fail to capture the dependence of feature information, resulting in poor accuracy in predicting traffic. Similarly, traffic prediction models based on causal structure learning, such as graph attention networks and graph convolutional neural networks (GCNs), have limitations in capturing the dependence of spatial information, leading to poor accuracy in predicting traffic.

[0169] exist Figure 8 In this embodiment, multiple sets of communication features corresponding to each of the multiple base stations are acquired. Based on the acquisition time of each set of communication features, the multiple sets of communication features are combined to obtain a three-dimensional communication feature tensor. Based on the three-dimensional communication feature tensor and multiple communication features, the spatial topology map of the base stations under each communication feature space is determined. The multiple communication features and the spatial topology map of the base stations are input into the communication traffic prediction model to obtain the communication traffic prediction results of each of the multiple base stations. This approach can take into account the dependence of communication traffic on spatial information, temporal information and feature information, thereby improving the accuracy of communication traffic prediction.

[0170] Figure 9 This is a schematic diagram of the structure of the model training device provided in an embodiment of the present invention. Figure 9As shown, the model training device includes:

[0171] The acquisition module 910 is used to acquire multiple sets of sample communication features corresponding to each of the multiple sample base stations; wherein each set of sample communication features includes multiple sample communication features.

[0172] The combination module 920 is used to combine the communication features of multiple sets of samples based on their respective acquisition times to obtain a three-dimensional communication feature tensor.

[0173] The determination module 930 is used to determine multiple training samples based on the three-dimensional communication feature tensor and each acquisition time.

[0174] Graph construction module 940 is used to determine the base station spatial topology map in the communication feature space of each sample based on the three-dimensional communication feature tensor and multiple sample communication features.

[0175] Training module 950 is used to train the initial spatiotemporal model based on multiple training samples and base station spatial topology map to obtain a communication traffic prediction model.

[0176] It should be noted that the model training device can implement the above-mentioned model training method and achieve the same technical effect. Therefore, the beneficial effects of this embodiment will not be described in detail here.

[0177] In some embodiments, the determining module 930 is specifically used for:

[0178] For each acquisition time, the time slice corresponding to the acquisition time in the three-dimensional communication feature tensor is determined as the training sample corresponding to the acquisition time; wherein, the time slice includes a set of sample communication features of each sample base station at the acquisition time.

[0179] In some embodiments, the graph construction module 940 is specifically used for:

[0180] The feature slices corresponding to the communication features of each sample are obtained from the three-dimensional communication feature tensor; wherein, the feature slices include the sample communication features of each sample base station at each acquisition time;

[0181] Based on the sample communication characteristics of any two sample base stations at each collection time, the edge weight between the two sample base stations is determined.

[0182] Based on the edge weights between any two sample base stations, a base station spatial topology graph is constructed under the communication feature space of each sample.

[0183] In some embodiments, the two sample base stations include a first sample base station and a second sample base station; the graph construction module 940 is specifically used for:

[0184] pass The sample communication characteristics of the first sample base station at each collection time and the sample communication characteristics of the second sample base station at each collection time are processed to obtain the edge weight between the first sample base station and the second sample base station.

[0185] in, X represents the edge weight between the first sample base station and the second sample base station, k represents the sample communication feature, and X represents the edge weight between the first sample base station and the second sample base station. k Y represents a vector composed of the sample communication features of the first sample base station at each acquisition time. k This represents a vector composed of the sample communication features of the second sample base station at each acquisition time. X represents k standard deviation Y represents k Standard deviation, COV(X) k Y k ) represents the covariance between Xk and Yk.

[0186] In some embodiments, the two sample base stations include a first sample base station and a second sample base station; the graph construction module 940 is specifically used for:

[0187] By using D(X, Y) = min P ∑ (i，j)∈P d(x i y j The sample communication characteristics of the first sample base station at each collection time and the sample communication characteristics of the second sample base station at each collection time are processed to obtain the edge weight between the first sample base station and the second sample base station.

[0188] Where D(X, Y) represents the edge weight between the first sample base station and the second sample base station, X represents the vector composed of the sample communication features of the first sample base station at each collection time, Y represents the vector composed of the sample communication features of the second sample base station at each collection time, P represents the set of index pairs composed of the indices of elements in X and the indices of elements in Y, (i, j) represents an index pair in P, x i Let y represent the i-th element in X. j Min represents the j-th element in Y. P · indicates the minimum value operation, d(x) i y j ) represents x i With y j The degree of difference between them, ∑· represents the summation operation.

[0189] In some embodiments, the initial graph spatiotemporal model includes an initial graph attention network and an initial gated recurrent unit; the training module 950 is specifically used for:

[0190] Multiple training samples and base station spatial topology maps under each communication feature space are input into the initial graph attention network to obtain the predicted features of each sample base station.

[0191] The predicted features of each sample base station are input into the initial gated loop unit to obtain the communication traffic prediction results corresponding to each of the multiple sample base stations.

[0192] Figure 10 This is a schematic diagram of the communication traffic prediction device provided in an embodiment of the present invention. Figure 10 As shown, the communication traffic prediction device includes:

[0193] The acquisition module 1010 is used to acquire multiple sets of communication features corresponding to multiple base stations; wherein each set of communication features includes multiple communication features.

[0194] The combination module 1020 is used to combine multiple sets of communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor.

[0195] Graph construction module 1030 is used to determine the base station spatial topology map under each communication feature space based on the three-dimensional communication feature tensor and multiple communication features;

[0196] Prediction module 1040 is used to input multiple communication features and base station spatial topology maps into a communication traffic prediction model to obtain communication traffic prediction results for each of the multiple base stations; wherein, the communication traffic prediction model is as described above. Figures 1 to 7 The communication traffic prediction model described in any of the embodiments.

[0197] It should be noted that the communication traffic prediction device can implement the above-mentioned communication traffic prediction method and achieve the same technical effect. Therefore, the beneficial effects of this embodiment will not be described in detail here.

[0198] Figure 11 This is a schematic diagram of the physical structure of the training device provided in an embodiment of the present invention. Figure 11As shown, the training device may include: a processor 1110, a communications interface 1120, a memory 1130, and a communication bus 1140. The processor 1110, communications interface 1120, and memory 1130 communicate with each other via the communication bus 1140. The processor 1110 can call logical instructions in the memory 1130 to execute a model training method. This method includes: acquiring multiple sets of sample communication features corresponding to multiple sample base stations; wherein each set of sample communication features includes multiple sample communication features; combining the multiple sets of sample communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor; determining multiple training samples based on the three-dimensional communication feature tensor and each acquisition time; determining a base station spatial topology map under the communication feature space of each sample based on the three-dimensional communication feature tensor and the multiple sample communication features; and training an initial spatiotemporal model based on the multiple training samples and the base station spatial topology map to obtain a communication traffic prediction model.

[0199] Figure 12 This is a schematic diagram of the physical structure of the central server provided in an embodiment of the present invention. Figure 12 As shown, the central server may include: a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240. The processor 1210, communications interface 1220, and memory 1230 communicate with each other via the communication bus 1240. The processor 1210 can call logical instructions in the memory 1230 to execute a communication traffic prediction method. This method includes: acquiring multiple sets of communication features corresponding to multiple base stations; wherein each set of communication features includes multiple communication features; combining the multiple sets of communication features based on their respective acquisition times to obtain a three-dimensional communication feature tensor; determining a base station spatial topology map under each communication feature space based on the three-dimensional communication feature tensor and the multiple communication features; and inputting the multiple communication features and the base station spatial topology map into a communication traffic prediction model to obtain the communication traffic prediction results for each of the multiple base stations; wherein the communication traffic prediction model is as described above. Figures 1 to 7 The communication traffic prediction model described in any of the embodiments.

[0200] Furthermore, the logical instructions in the aforementioned memories 1130 and 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0201] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the above-mentioned model training method or communication traffic prediction method.

[0202] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the above-described model training method or communication traffic prediction method.

[0203] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0204] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A model training method, characterized in that, include: Obtain multiple sets of sample communication features corresponding to each of the multiple sample base stations; wherein each set of sample communication features includes multiple sample communication features; Based on the acquisition time of each of the multiple sets of sample communication features, the multiple sets of sample communication features are combined to obtain a three-dimensional communication feature tensor. The three dimensions of the three-dimensional communication feature tensor correspond to the sample base station, the acquisition time, and the category of the sample communication feature, respectively. Based on the three-dimensional communication feature tensor and each acquisition time, multiple training samples are determined; Based on the three-dimensional communication feature tensor and the multiple sample communication features, a base station spatial topology graph is determined in the communication feature space of each sample. The base station spatial topology graph is an undirected graph with each sample base station as a node, and the edge weight between any two nodes in the base station spatial topology graph is determined based on the sample communication features of the corresponding two sample base stations at each acquisition time. Based on the multiple training samples and the base station spatial topology map, the initial graph spatiotemporal model is trained to obtain a communication traffic prediction model. The initial graph spatiotemporal model includes an initial graph attention network and an initial gated recurrent unit.

2. The model training method according to claim 1, characterized in that, Based on the three-dimensional communication feature tensor and each acquisition time, multiple training samples are determined, including: For each acquisition time, the time slice corresponding to the acquisition time in the three-dimensional communication feature tensor is determined as the training sample corresponding to the acquisition time; wherein, the time slice includes a set of sample communication features of each sample base station at the acquisition time.

3. The model training method according to claim 2, characterized in that, The step of determining the base station spatial topology map in the communication feature space of each sample based on the three-dimensional communication feature tensor and the multiple sets of sample communication features includes: From the three-dimensional communication feature tensor, feature slices corresponding to the communication features of each sample are obtained; wherein, the feature slices include the communication features of each sample base station at each acquisition time; Based on the sample communication characteristics of any two sample base stations among the plurality of sample base stations at each collection time, determine the edge weight between the two sample base stations; Based on the edge weights between any two sample base stations, a base station spatial topology graph is constructed under the communication feature space of each sample.

4. The model training method according to claim 3, characterized in that, The two sample base stations include a first sample base station and a second sample base station; Based on the sample communication characteristics of the two sample base stations at each collection time, the edge weight between the two sample base stations is determined, including: pass The sample communication features of the first sample base station and the sample communication features of the second sample base station at each collection time are processed to obtain the edge weights between the first sample base station and the second sample base station. in, This represents the edge weight between the first sample base station and the second sample base station. This indicates the communication characteristics of the sample. This represents a vector composed of the sample communication features of the first sample base station at each acquisition time. This represents a vector composed of the sample communication features of the second sample base station at each acquisition time. express standard deviation express standard deviation express and The covariance between them.

5. The model training method according to claim 3, characterized in that, The two sample base stations include a first sample base station and a second sample base station; Based on the sample communication characteristics of the two sample base stations at each collection time, the edge weight between the two sample base stations is determined, including: pass The sample communication features of the first sample base station and the sample communication features of the second sample base station at each collection time are processed to obtain the edge weights between the first sample base station and the second sample base station. in, This represents the edge weight between the first sample base station and the second sample base station. This represents a vector composed of the sample communication features of the first sample base station at each acquisition time. This represents a vector composed of the sample communication features of the second sample base station at each acquisition time. express index of element in A set of index pairs consisting of the indices of the elements in the set. express One of the index pairs, express The first in One element, express The first in One element, This indicates the minimum value operation. express and The degree of difference between them This indicates a summation operation.

6. The model training method according to any one of claims 2 to 5, characterized in that, The step of training an initial graph spatiotemporal model based on the multiple training samples and the base station spatial topology map to obtain a communication traffic prediction model includes: The multiple training samples and the base station spatial topology map under each communication feature space are input into the initial graph attention network to obtain the predicted features of each sample base station; The predicted features of each sample base station are input into the initial gated loop unit to obtain the communication traffic prediction results corresponding to each of the multiple sample base stations; Based on the communication traffic prediction results and tag communication traffic corresponding to each of the multiple sample base stations, the model parameters in the initial graph attention network and the initial gated recurrent unit are updated to obtain the communication traffic prediction model.

7. A method for predicting communication traffic, characterized in that, include: Obtain multiple sets of communication features corresponding to each of the multiple base stations; where each set of communication features includes multiple communication features. Based on the acquisition time of each of the multiple sets of communication features, the multiple sets of communication features are combined to obtain a three-dimensional communication feature tensor. The three dimensions of the three-dimensional communication feature tensor correspond to the base station, the acquisition time, and the category of the communication feature, respectively. Based on the three-dimensional communication feature tensor and the multiple communication features, a base station spatial topology graph is determined under each communication feature space. The base station spatial topology graph is an undirected graph with each base station as a node, and the edge weight between any two nodes in the base station spatial topology graph is determined based on the communication features of the corresponding two base stations at each acquisition time. The plurality of communication features and the spatial topology map of the base stations are input into the communication traffic prediction model to obtain the communication traffic prediction results of each of the plurality of base stations; wherein, the communication traffic prediction model is the communication traffic prediction model according to any one of claims 1 to 6 above.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the model training method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the model training method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the model training method as described in any one of claims 1 to 6.

Images