A Multi-Band Spectrum Prediction Method Based on Dynamic Graphs

Abstract

The present application provides a multi-band spectrum prediction method based on a dynamic graph. The method includes: Step 1, constructing a multi-band dynamic graph structure of a sliding time window; Step 2, constructing a multi-band spectrum tensor based on historical data; Step 3, a time-frequency dual-channel spectrum prediction model based on the dynamic graph; Step 4, implementing the multi-band spectrum prediction method based on the dynamic graph. The present application constructs a dynamic graph for the multi-band spectrum correlation of different time windows, and uses a method of dual-channel extraction of time features and frequency band features for multi-band prediction, effectively avoiding the problem of poor model generalization ability caused by traditional prediction methods ignoring the dynamic characteristics of the spectrum environment.

Description

Technical Field

[0001] This application relates to the field of wireless communication technologies, and particularly to a multi-band spectrum prediction method based on a dynamic graph. Background Art

[0002] With the progress of wireless communication technologies, the demand for spectrum has also been growing rapidly. In recent decades, communication technologies have experienced rapid development from 3G to 4G and then to 5G, and more and more advanced communication devices have been designed and widely used. This has led to the scarcity of spectrum resources. On the other hand, the low-frequency spectrum utilization rate of existing static allocation strategies has further exacerbated the shortage of spectrum resources. How to dynamically utilize scarce spectrum resources has become the focus of attention of researchers. In order to dynamically and effectively utilize idle spectrum resources, researchers have proposed the concept of dynamic spectrum access, that is, without introducing interference, underutilized licensed frequency bands are dynamically shared by a large number of unlicensed users to achieve the reuse of spectrum resources. Accurately predicting spectrum usage is the key prerequisite for realizing spectrum sharing and dynamic spectrum access. Spectrum prediction is a method of inferring future channel occupancy status, duty cycle, power spectrum, and other information based on historical spectrum detection data. By predicting spectrum holes and temporarily allocating them to unlicensed users, the overall utilization rate of the spectrum can be improved, which is a widely accepted method. Existing research results show that the use of spectrum has a high degree of temporal correlation and frequency-domain correlation. Making full use of the temporal and frequency correlations obtained from historical observation data is the key to achieving high-precision spectrum prediction.

[0003] In spectrum analysis, a single-band spectrum can be regarded as a data sequence with time characteristics, while a multi-band spectrum is a collection of multiple such time data sequences. Initially, researchers predicted the spectrum trend by combining time-frequency domain methods with a convolutional neural network (CNN) model. Later, researchers found that this method does not produce satisfactory results when dealing with multi-band data. Therefore, researchers proposed using a graph structure to associate frequency bands that are far apart.

[0004] Although this method has shown some improvements, due to the openness of the spectrum environment and the inherent time-dynamic characteristics of spectrum data, such as Figure 1As shown. Three frequency bands are randomly selected from the Electrosense data in the range of 790 - 820 MHz for analysis. In time window 1, there is a positive correlation among three frequency bands: Band 1, Band 2, and Band 3. In time window 2, the trends between Band 1 and Band 3 are relatively similar, but they show uncorrelated or even negatively correlated behavior with Band 2. It can be seen from the figure that the relationships among multiple frequency bands exhibit time-varying characteristics. However, the simple static graphs in previous studies ignored this feature and were not sufficient to fully capture the time-varying global characteristics in multi-band spectral data. Therefore, the current research focus has shifted to effectively extracting the global characteristics with time-varying properties among multi-band frequencies, which is an important direction for improving the accuracy of spectral prediction. Summary of the Invention

[0005] This application provides a multi-band spectral prediction method based on a dynamic graph, which can be used to solve the problem of ignoring the dynamic information of multi-band correlations in existing research.

[0006] This application provides a multi-band spectral prediction method based on a dynamic graph, and the method includes:

[0007] Step 1, constructing a multi-band dynamic graph structure with a sliding time window;

[0008] Step 2, constructing a multi-band spectral tensor based on historical data;

[0009] Step 3, a time-frequency dual-channel spectral prediction model based on the dynamic graph;

[0010] Step 4, implementing the multi-band spectral prediction method based on the dynamic graph.

[0011] Further, constructing a multi-band dynamic graph structure with a sliding time window includes:

[0012] The purpose of this step is to capture the time-varying characteristics on multiple frequency bands, and a dynamic graph is constructed using the sliding window method. The sliding window is applied to the frequency bands, and by calculating the similarity relationships between frequency bands within each window, a dynamic graph that changes over time is constructed. This dynamic graph can not only intuitively display the changing trend of the similarity between frequency bands but also provide in-depth insights into the time-varying characteristics of multi-bands.

[0013] Since the correlation relationships among multi-bands in different time windows are dynamically changing, a dynamic graph of multi-bands is constructed through a sliding event window to capture the multi-band dynamic relationships;

[0014] Constructing a dynamic graph through a time sliding window where represents a set of nodes, ε represents a set of edges, and each frequency band f i is considered as a node v in the dynamic graph i, the similarity between frequency bands forms the edges of the graph; calculate the correlation between a frequency band and all other frequency bands except this frequency band, and then select the k frequency bands with the highest correlation with this frequency band to establish connection edges; the similarity between frequency bands f1 and f n in the time window w i is transformed into:

[0015]

[0016] w i =[t p , t p + 1,..., t q

[0017] where and represent the spectral states of frequency bands f1 and f n in the time window w i respectively; use the Euclidean distance as a metric to measure the similarity between two frequency bands, where α is defaulted to 2, and eps represents a small number to avoid division by zero, with the default setting of 1e-6; t p to t q represent the size of the time window.

[0018] Furthermore, step 2, construct a multi-band spectral tensor based on historical data, including:

[0019] Since the spectral usage has a certain temporal regularity in adjacent time periods, arrange the power spectral values of each frequency band in time sequence as a time tensor for processing;

[0020] Analyze the frequency band with a frequency range of A - B MHz, with a frequency band width of C MHz, to form a total of (B - A) / C frequency band data; then, analyze the power spectral values of each frequency band in chronological order, with T minutes as a time unit, and form a vector of size X i ∈R n×c containing all frequency bands within each time unit, where c = 1, n = (B - A) / C, and then form a tensor arranged in time sequence:

[0021]

[0022] where m represents time.

[0023] Furthermore, step 3, establish a time-frequency dual-channel spectral prediction model based on a dynamic graph, including:

[0024] ​The time-frequency dual-channel spectrum prediction model based on a dynamic graph first uses the k-nearest neighbor algorithm to model the similarity relationship between multiple frequency bands, and constructs a dynamic graph using the sliding window method to capture the time-varying features on multiple frequency bands. Subsequently, a dual-channel feature extraction module is used to extract global and local features respectively. Finally, an adaptive fusion mechanism is used to weight and fuse these two types of features. The model separates the two modules to ensure that they do not interfere with each other, avoids feature redundancy, and effectively solves the problem of making full use of global and local features between frequency bands.

[0025] Building a time-frequency dual-channel spectrum prediction model based on a dynamic graph includes three modules, namely a dynamic graph construction module, a global feature extraction module, and a local feature extraction module;

[0026] Among them, the dynamic graph construction model generates a graph structure for each multi-frequency band within each time window, representing the similarity between multi-frequency bands. The multi-band spectrum is divided into several frequency bands according to the frequency band interval of CMHz, and the power spectral values of each frequency band are arranged in chronological order, and then the sliding time window t p to t q is performed. Correlation analysis is carried out on the multi-band power spectra within each time window, and the k most relevant nodes are selected using the k-nearest neighbor algorithm to establish connection relationships. In this way, a graph structure G = <V, E> is dynamically formed for each time window;

[0027] Then the constructed graph structure G = <V, E> and node features are input into two parallel modules, namely a global feature extraction module and a local feature extraction module;

[0028] And the extracted global features and local features are fused, and finally the time-frequency dual-channel spectrum prediction model is output;

[0029] In the global feature extraction module, a dynamic graph convolutional module is used to adaptively learn and update the graph structure, effectively process and analyze the dynamic changes between frequency bands, and a long short-term memory (LSTM) network is used to extract time features; in the local feature extraction module, local features of adjacent regions between multiple frequency bands are extracted; a gate fusion mechanism is used to weight and fuse these two types of features; finally, the power spectral density values for the next d time units are output.

[0030] Furthermore, the global feature extraction module includes three dynamic graph convolutional layers and two LSTM layers;

[0031] Extract the global-level features of multiple frequency bands within the graph structure using the global feature extraction module; global correlation refers to the extensive and non-local interdependencies existing among all time-frequency objects in the frequency domain and time dimension. The interdependencies not only consider the interactions between adjacent frequency bands in the frequency domain but also the mutual influences of all frequency bands on a long time scale. When modeling the global features of multiple frequency bands, a dynamic graph structure is adopted, which varies dynamically according to multiple frequency bands within different time windows. Given the powerful capabilities demonstrated by graph convolutional networks in the field of graph learning, a dynamic graph convolutional network (DGCN) is used to capture the information in the graph structure, thereby extracting the graph structure features of multiple frequency bands.

[0032] The dynamic graph convolutional layer aggregates the features of adjacent nodes and updates the features of nodes in different time windows as follows:

[0033]

[0034] where is the graph structure updated and aggregated through one layer of dynamic convolution operation; ψ agg and ψ update are the learnable weights of the aggregation and update operations respectively;

[0035] For each node, the aggregation and update operations are as follows:

[0036]

[0037] is the set of neighbor nodes of v i , and v j is a neighbor node of v i ; To adaptively update the node features, a fine-grained edge attention module is used to weight and aggregate the node edges, and the operation is as follows:

[0038]

[0039] where represents the node feature of v i through the attention module, and Att() represents the function of node weighted aggregation; among them, Att(v i ) represents the weighted aggregation of its own features as follows:

[0040]

[0041] where, represents the original feature of the node, μ i,i represents the attention weight; ζ() represents the convolutional layer with a 1×1 convolutional kernel parameterized by ψ agg ; The attention-weighted aggregation of the adjacent nodes of v i is expressed as

[0042]

[0043] Among them represents connecting two features; the attention weights are as follows:

[0044]

[0045] where u() represents the convolutional layer of a 1×1 convolutional kernel parameterized by ψ update ;

[0046] In addition, considering that the features carried by each node are the power spectral values of the corresponding frequency bands arranged in time series, time information is extracted from the corresponding node features through two LSTM layers to obtain the feature vector χ g :

[0047]

[0048] where ⊙ represents the Hadamard product operator, c i-1 represents the memory information of the previous time step, c i represents the memory information of the current time step, h i is the output of the LSTM cell, h i-1 is the output of the previous time step; δ represents the weight matrix, and b represents the bias vector;

[0049] Through the global feature extraction module, the dynamic graph convolution is responsible for processing the structural information of the graph, while the LSTM layer is used for the temporal evolution of the node features, so as to comprehensively capture the global features of the multi-band data with time variations.

[0050] Furthermore, the local feature extraction module is used to extract the local feature information of the multi-band spectral data. The local feature extraction module includes two convolutional layers and two LSTM layers;

[0051] Input the multi-band feature vector into the convolutional layer, and use the convolutional kernel to traverse the feature vector to generate a new feature vector η, effectively extracting the information of the current frequency band and its surrounding frequency bands. The generation of the feature vector is as follows:

[0052]

[0053] where δ represents the convolutional kernel weight vector, b represents the offset vector, and λ represents the activation function, represents the convolution operation;

[0054] Then, input the feature vector η into the LSTM to extract time information and obtain the feature X c ;

[0055] Finally, an adaptive gating mechanism is used to fuse the multi-band global features extracted by the global feature extraction module and the local features extracted by the local feature extraction module to obtain feature X f :

[0056]

[0057] where γ represents the learned allocation weight, represents the Sigmoid function, χ g and χ c represent the weights of the multi-band global features and the multi-band local features respectively, and b represents the bias.

[0058] Furthermore, in step 4, a multi-band spectrum prediction method based on a dynamic graph is implemented, including:

[0059] The task of this application is to predict the power spectral density values of n frequency bands in the next d time units The power spectral density values are determined by training the prediction model F:

[0060]

[0061] First, select the frequency band bandwidth as A - B MHz, and then use the power spectral values of the past m time units to predict the power spectral values of the next d time units; in the process of constructing the graph structure, the k-nearest neighbor algorithm is used to select k nodes with the strongest correlation to construct the connection relationship; through continuous iterative optimization, the algorithm converges, and then the prediction result is output; compare the power spectral values obtained by the multi-band spectrum prediction module based on the dynamic graph in the next d time units with the true power spectral values of the current time unit, and evaluate the prediction accuracy by calculating the error between the true value and the predicted value. The prediction performance evaluation indicators include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE);

[0062] The root mean square error is a commonly used indicator to measure the difference between the model prediction value and the actual observed value. It is obtained by calculating the mean of the squares of the differences between the prediction value and the actual observed value and taking the square root:

[0063]

[0064] The mean absolute error is used to evaluate the fitting degree of the model on the given data; it is obtained by calculating the mean of the absolute values of the differences between the prediction value and the actual observed value:

[0065]

[0066] The mean absolute error is sensitive to the relative error, which refers to the mean absolute error between the predicted value and the true value. The determination method is as follows:

[0067]

[0068] X m+i represents the true value of the spectrum at the (m + i)-th time unit, represents the predicted value of the spectrum at the (m + i)-th time unit, and s represents the predicted time step; the smaller the values of the three evaluation metrics, the smaller the error between the predicted result obtained by the model and the true value, and thus the better the prediction effect of the proposed model; the prediction effect of the model is obtained by calculating the three error metrics.

[0069] Compared with the prior art, the remarkable advantages of the present invention are:

[0070] (1) By constructing a dynamic graph to capture the temporal dynamics of global features in multiple frequency bands and introducing a dynamic graph convolutional module that can adaptively learn and update the graph, for each time window, a dedicated graph is dynamically learned, thereby improving the generalization ability of the model.

[0071] (2) The present invention adopts a dual-channel network, which can simultaneously extract global and local features of multiple frequency bands, thereby overcoming the limitation of relying only on a single type of feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a dynamic correlation graph between multiple frequency bands provided by an embodiment of the present application.

[0073] Figure 2 is a multi-band prediction model graph based on a dynamic graph provided by an embodiment of the present application.

[0074] Figure 3 is a long short-term memory network graph provided by an embodiment of the present application.

[0075] Figure 4 is a node selection analysis graph for constructing a dynamic graph based on k-nearest neighbors provided by an embodiment of the present application.

[0076] Figure 5 is a result graph of the multi-band spectrum prediction method proposed by the present invention in different prediction steps provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0077] To make the objectives, technical solutions, and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the accompanying drawings.

[0078] Next, the embodiments of the present application will be introduced with reference to the accompanying drawings first.

[0079] Next, the embodiments of the present application will be introduced with reference to the accompanying drawings first.

[0080] The present invention uses a directed graph model to describe the spatial correlation between user service behaviors and considers the importance differences in different regions. In addition, a spatio-temporal fusion structure is used to extract spatial features and temporal features simultaneously, so as to more comprehensively understand the characteristics of user business behaviors and make more accurate predictions. The present invention uses a large amount of historical user business behavior data of multiple days and multiple regions to predict the user business behavior state in the future for multiple days.

[0081] The present application constructs a dynamic graph for the multi-band correlation of different time windows. In the graph, each frequency band is regarded as a node, and the correlation between each frequency band and other frequency bands constructs an edge. According to the k-nearest neighbor algorithm, the k frequency bands with the highest correlation are selected as having connectivity, and vice versa, there is no connection relationship. And the historical spectrum data of each node is arranged in chronological order to form a time correlation tensor of multiple frequency bands. Then, the graph structure and the time correlation tensor of the nodes are input into a two-channel prediction model, and finally the prediction result is output.

[0082] The steps for multi-band spectrum prediction using the method of the present invention are as follows: Step 1, construct a multi-band dynamic graph structure of a sliding time window; Step 2, construct a multi-band spectrum tensor based on historical data; Step 3, a time-frequency two-channel spectrum prediction model based on the dynamic graph; Step 4, evaluate the performance of the multi-band spectrum prediction method based on the dynamic graph. Each step is specifically elaborated below:

[0083] Step 1, construct a multi-band dynamic graph structure of a sliding time window.

[0084] The purpose of this step is to capture the time-varying features on multiple frequency bands and construct a dynamic graph using the sliding window method. The sliding window is applied to the frequency bands, and by calculating the similarity relationship between frequency bands within each window, a dynamic graph that changes over time is constructed. This dynamic graph can not only intuitively display the changing trend of the similarity between frequency bands but also provide in-depth insights into the time-varying features of multiple frequency bands.

[0085] Since the correlation relationship between multiple frequency bands within different time windows is dynamically changing, a dynamic graph is constructed for multiple frequency bands by sliding the event window to capture the dynamic relationship of multiple frequency bands; a dynamic graph is constructed through the time sliding window where represents a set of nodes, and ε represents a set of edges. Specifically, each frequency band f i is considered as a node v in the dynamic graph i, the similarity between frequency bands forms the edges of the graph. Specifically, the correlation is calculated between a frequency band and all other frequency bands except this one, and then the k frequency bands with the highest correlation with this frequency band are selected to establish connection edges. Frequency band f1 and f n In the time window w i The similarity is formulated as:

[0086]

[0087] w i = [t p , t p +1,..., t q

[0088] where and represent the spectral states of frequency bands f1 and f n in the time window w i respectively. The Euclidean distance is used as a metric to measure the similarity between two frequency bands, where α is defaulted to 2, and eps represents a small number to avoid division by zero, with the default setting of 1e-6. t p to t q represent the size of the time window. Then, for each node v i , find its k most similar nodes to add edges.

[0089] Step 2, construct a multi-band spectral tensor based on historical data.

[0090] Since the spectral usage has a certain temporal regularity in adjacent time periods, the power spectral values of each frequency band are arranged in time sequence as a time tensor for processing;

[0091] Analyze the frequency band with a frequency range of A - B MHz, with a frequency band width of C MHz, to form a total of (B - A) / C frequency band data. Then, analyze the power spectral values of each frequency band in chronological order, with T minutes as a time unit, and form a vector of size X i ∈R n×c containing all frequency bands within each time unit, where c = 1, n = 59, and then form a tensor arranged in time sequence:

[0092]

[0093] where m represents time.

[0094] Step 3, establish a time-frequency dual-channel spectral prediction model based on a dynamic graph.

[0095] ​The time-frequency dual-channel spectrum prediction model based on a dynamic graph first uses the k-nearest neighbor algorithm to model the similarity relationships between multiple frequency bands and constructs a dynamic graph using a sliding window method to capture the time-varying features on multiple frequency bands. Subsequently, a dual-channel feature extraction module is used to extract global and local features respectively. In the global feature extraction module, a dynamic graph convolutional module is used to adaptively learn and update the graph structure, effectively processing and analyzing the dynamic changes between frequency bands, while LSTM is used to extract temporal features. The local feature extraction module is used to extract the features of adjacent regions between multiple frequency bands. Finally, an adaptive fusion mechanism is used to weight and fuse these two types of features. The model separates the two modules to ensure that they do not interfere with each other, avoiding feature redundancy, and effectively solving the problem of making full use of global and local features between frequency bands.

[0096] As shown in the time-frequency dual-channel spectrum prediction model based on a dynamic graph Figure 2 below, the DGCN-LSTM part represents the global feature extraction module, which is used to extract the global features of multiple frequency bands; the CNN-LSTM part represents the local feature extraction module, which is used to extract the local features of multiple frequency bands; the feature fusion part represents the fusion module, which is used to adaptively fuse the multi-band global information and local information features extracted by the global feature extraction module and the local feature extraction module.

[0097] The following specifically elaborates on each part of the time-frequency dual-channel spectrum prediction model based on a dynamic graph:

[0098] Global feature extraction module:

[0099] The global feature extraction module is used to extract the global-level features of multiple frequency bands within the graph structure; global correlation refers to the extensive and non-local interdependencies that exist between all time-frequency objects in the frequency domain and time dimensions. This global correlation not only considers the interactions between adjacent frequency bands in the frequency domain but also the mutual influences of all frequency bands on a long time scale. When modeling the global features of multiple frequency bands, a dynamic graph structure is adopted to change dynamically according to multiple frequency bands within different time windows. Given the powerful capabilities demonstrated by graph convolutional networks in the field of graph learning, a dynamic graph convolutional network (DGCN) is used to capture the information in the graph structure, thereby extracting the graph structure features of multiple frequency bands.

[0100] Specifically, the graph structure and the feature matrix are input into the dynamic graph convolution to adaptively update and aggregate the node features in different time windows. The specific process is described as follows: First, the features of adjacent nodes are aggregated through the dynamic graph convolutional layer:

[0101]

[0102] where is the graph structure updated and aggregated through one layer of dynamic convolution operation. ψagg and ψ update are learnable weights for aggregation and update operations, respectively.

[0103] Specifically, for each node, the aggregation and update operations are as follows:

[0104]

[0105] is the set of neighbor nodes of v i , and v j is a neighbor node of v i . To adaptively update the node features, a fine-grained edge attention module is used to weight and aggregate the node edges, and the operation is as follows:

[0106]

[0107] where represents the node feature of v i passing through the attention module, and Att() represents the function of node weighted aggregation. Among them, Att(v i ) represents the weighted aggregation of its own features, which is specifically represented as follows:

[0108]

[0109] where represents the original feature of the node, and μ i,i represents the attention weight. ζ() represents the convolutional layer with a 1×1 convolutional kernel parameterized by ψ agg . The attention weighted aggregation of the adjacent nodes of v i is represented as

[0110]

[0111] where represents connecting two features. The attention weight is specifically represented as follows:

[0112]

[0113] where u() represents the convolutional layer with a 1×1 convolutional kernel parameterized by ψ update . In addition, considering that the features carried by each node are the power spectral values of frequency bands arranged in time series, time information is extracted from these node features through two LSTM layers to obtain the feature vector χ g .

[0114]

[0115] where ⊙ represents the Hadamard product operator, and c i-1 represents the memory information of the previous time step, and c i represents the memory information of the current time step, h i is the output of the LSTM cell, and h i-1 is the output of the previous time step. δ represents the weight matrix, and b represents the bias vector. Through the global feature extraction module, the dynamic graph convolution is responsible for processing the structural information of the graph, while the LSTM layer focuses on the temporal evolution of the node features, thus comprehensively capturing the global features of the multi-band data with time variations.

[0116] Local feature extraction module:

[0117] In the local feature extraction module, convolutional layers and long short-term memory units are used to extract features in the multi-band local range. It includes two convolutional layers and two long short-term memory units.

[0118] Input the multi-band feature vector into the convolutional layer, and use the convolutional kernel to traverse the feature vector to generate a new feature vector η, effectively extracting the information of the current frequency band and its surrounding frequency bands. The specific process of feature vector generation is as follows:

[0119]

[0120] where δ represents the convolutional kernel weight vector, b represents the offset vector, λ represents the activation function, represents the convolutional operation. Then, input the feature vector η into the LSTM to extract the temporal information and obtain the feature χ c .

[0121] Feature fusion module:

[0122] Finally, use the adaptive gating mechanism to fuse the multi-band global features extracted by the global feature extraction module and the local features extracted by the local feature extraction module to obtain the feature χ f :

[0123]

[0124] where γ represents the learned allocation weight, represents the Sigmoid function, χ g and χ c represent the weights of the multi-band global feature and the multi-band local feature respectively, and b represents the bias.

[0125] Step 4, implement the multi-band spectrum prediction method based on the dynamic graph.

[0126] The task of this application is to predict the power spectral density values of these n frequency bands in the next d time units This problem is solved by training a prediction model F, and the specific operation process is as follows:

[0127]

[0128] Compare the power spectral values obtained by the multi-band spectrum prediction module based on the dynamic graph in the next d time units with the true power spectral values of this time unit, and evaluate the prediction accuracy by calculating the error between the true value and the predicted value. The prediction performance evaluation indicators include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE);

[0129] Root Mean Squared Error is a commonly used indicator to measure the difference between the predicted values of the model and the actual observed values. It is obtained by calculating the mean of the squares of the differences between the predicted values and the actual observed values and taking the square root:

[0130]

[0131] Mean Absolute Error is used to evaluate the fitting degree of the model on the given data; it is obtained by calculating the average of the absolute values of the differences between the predicted values and the actual observed values:

[0132]

[0133] Mean Absolute Percentage Error is sensitive to relative errors and refers to the average absolute error between the predicted value and the true value. The specific calculation is as follows:

[0134]

[0135] Among these three indicators, X m+i represents the true value of the spectrum at the m + i-th time unit, represents the predicted value of the spectrum at the m + i-th time unit, and s represents the predicted time step; the smaller the values of the three evaluation indicators, the smaller the error between the predicted result obtained by the model and the true value, and thus the better the prediction effect of the proposed model.

[0136] The following further elaborates on this application in combination with specific embodiments.

[0137] In this embodiment, the data we used comes from the Electrosense open-source data platform. The selected frequency band bandwidth is 790 - 820 MHz, the data acquisition time resolution is 15 minutes, and the frequency resolution is 0.5 MHz. Therefore, there are a total of 59 frequency bands in this dataset. In this experiment, we use the power spectral values of the past 12 hours to predict the power spectral values for the next 3 hours. We divide the dataset into a training set, a validation set, and a test set in a ratio of 7:1:2. In the experiment, the hyperparameters of the model are set as follows: MiniBatchSize is set to 4, MaxEpochs is set to 200, InitialLearnRate is set to 0.0005, and Optimizers is set to Adam.

[0138] In the process of constructing the graph structure, the k-nearest neighbor algorithm is used to select k nodes with the strongest correlation to construct the connection relationship. Therefore, the selection of k is very important. According to the calculation principle of graph convolution, the value of k should not be too large. So we conducted experiments on k from 3 to 18, and the experimental results are as Figure 3 shown. It can be seen from the figure that when k is 12, the prediction effect is the best. Therefore, for each node, the top 12 nodes with the strongest correlation to it are selected to set the connectivity. After continuous iterative optimization, the algorithm converges, and then the prediction result is output.

[0139] Comparison scheme: The method proposed in the present invention is compared with the models often used in prediction tasks. The comparison models are: GCN-LSTM, CNN-LSTM, LSTM, and CONVLSTM.

[0140] The GCN-LSTM model mainly includes two layers of graph convolution and two layers of long short-term memory units. The difference from the global feature extraction module of the present invention is that the GCN-LSTM model has no dynamic characteristics and can only extract features from static graphs.

[0141] The CNN-LSTM model includes two layers of convolutional layers and two layers of long short-term memory units.

[0142] The LSTM model only contains two layers of long short-term memory units.

[0143] The CONVLSTM model is a model used to extract the features of two-dimensional matrices and is often used in prediction tasks.

[0144] The prediction effect achieved by the model proposed in this application was compared with other models, and the comparison results were reflected by prediction evaluation indicators. The prediction effects achieved by the five models and the comparison indicators are shown in Table 1. It can be seen from the table that the model proposed in this application achieved the best prediction effect. And in order to more intuitively show the advantages of the proposed model over other models, the comparison under different prediction steps was shown, such as Figure 4 as shown

[0145] Table 1: Comparison results of the dynamic multi-band prediction model and the comparison model in prediction indicators

[0146]

[0147]

[0148] To verify the effectiveness of the dynamic graph, ablation analysis was carried out on the spectrum prediction method based on the dynamic graph proposed in this paper. The spectrum prediction based on the dynamic graph and the static graph was compared, and the comparison results are shown in Table 2. The results show that the dynamic graph can better capture the characteristics of multi-band spectra.

[0149] Table 2: Comparison results of the dynamic multi-band prediction model and the static graph model in prediction indicators

[0150] MAE RMSE MAPE Static graph 0.7848 1.4393 6.6491％ Dynamic graph 0.7477 1.4022 6.4177％

[0151] The significant advantages of the present invention are as follows:

[0152] (1) By constructing a dynamic graph to capture the temporal dynamics of global features in multiple frequency bands, and introducing a dynamic graph convolutional module that can adaptively learn and update the graph. For each time window, a dedicated graph is dynamically learned, thereby improving the generalization ability of the model.

[0153] (2) The present invention adopts a dual-channel network, which can simultaneously extract global and local features of multiple frequency bands, thereby overcoming the limitation of relying only on a single type of feature.

[0154] The embodiments of the present application described above do not constitute a limitation on the protection scope of the present application.

Claims

1. A multi-band spectrum prediction method based on a dynamic graph, characterized in that, The method includes: Step 1, constructing a multi-band dynamic graph structure of a sliding time window; Step 2, constructing a multi-band spectrum tensor based on historical data; Step 3, a time-frequency dual-channel spectrum prediction model based on a dynamic graph; Step 4, implementing a multi-band spectrum prediction method based on a dynamic graph; Constructing a multi-band dynamic graph structure of a sliding time window includes: Since the correlation relationship between multi-bands within different time windows is dynamically changing, a dynamic graph is constructed for multi-bands through a sliding event window to capture the multi-band dynamic relationship; Construct a dynamic graph through a time-sliding window where represents a set of nodes, ε represents a set of edges, and each frequency band f i is considered as a node v in the dynamic graph i , the similarity between frequency bands forms the edges of the graph; calculate the correlation between a frequency band and all other frequency bands except this frequency band, and then select the k frequency bands with the highest correlation with this frequency band to establish connection edges; the similarity between frequency bands f1 and f n in the time window w i is transformed into: w i = [t p , t p + 1,..., t q ​ Among them and respectively represent the spectral states of frequency bands f1 and f n in time window w i ; The Euclidean distance is used as a metric to measure the similarity between two frequency bands, where α defaults to 2, eps represents a small number to avoid division by zero, and the default setting is 1e-6; t p to t q represents the size of the time window.

2. The method according to claim 1, wherein Step 2, constructing a multi-band spectrum tensor based on historical data, includes: Analyze the frequency band with a frequency range of A - B MHz, with a frequency band width of C MHz, to form data for a total of (B - A) / C frequency bands; then, analyze the power spectral values of each frequency band in chronological order, with T minutes as a time unit, and form a vector of size X containing all frequency bands within each time unit i ∈R n×c where c = 1, n = (B - A) / C, and then form a tensor arranged in chronological order: X = {X1, X2,..., X m} ∈ R n×m×c where m represents time.

3. The method according to claim 1, wherein Step 3, establishing a time-frequency dual-channel spectrum prediction model based on a dynamic graph, includes: Establishing a time-frequency dual-channel spectrum prediction model based on a dynamic graph includes three modules, namely a dynamic graph construction module, a global feature extraction module, and a local feature extraction module; Among them, the dynamic graph construction model generates a graph structure for each multi-band within a time window, representing the similarity between multi-bands. The multi-band spectrum is divided into several frequency bands at intervals of CMHz, and the power spectrum values of each frequency band are arranged in chronological order, and then a sliding time window t p to t q is set. Correlation analysis is performed on the multi-band power spectra within each time window, and the k most relevant nodes are selected using the k-nearest neighbor algorithm to establish connection relationships. In this way, a graph structure G = <V, E> is dynamically formed for each time window; Then, the constructed graph structure G = <V, E> and node features X = {X1, X2,..., X m} ∈ R n×m×c are input into two parallel modules, namely the global feature extraction module and the local feature extraction module; And the extracted global features and local features are fused, and finally a time-frequency dual-channel spectrum prediction model is output; In the global feature extraction module, a dynamic graph convolution module is used to adaptively learn and update the graph structure, and a long short-term memory (LSTM) network is used to extract time features; in the local feature extraction module, local features of adjacent regions between multiple frequency bands are extracted; a gate fusion mechanism is used to weight and fuse these two types of features; finally, the power spectral density values for the next d time units are output.

4. The method according to claim 3, characterized in that, The global feature extraction module includes three dynamic graph convolutional layers and two LSTM layers; The dynamic graph convolutional layer aggregates the features of adjacent nodes and updates the features of nodes in different time windows as follows: Among them is a graph structure updated and aggregated through a layer of dynamic convolution operations; ψ agg and ψ update are learnable weights for the aggregation and update operations respectively; For each node, the aggregation and update operations are as follows: is v i is the set of neighbor nodes of v, and v j is v i is a neighbor node of v; to adaptively update the node features, a fine-grained edge attention module is used to weight and aggregate the node edges, and the operations are as follows: Among them represents the node feature of v passing through the attention module i , Att() represents the function of node weighted aggregation; among them, Att(v i ) represents the weighted aggregation of its own features, as follows: Among them, represents the original features of the node, μ i,i represents the attention weight; δ() represents a convolutional layer with a 1×1 convolutional kernel parameterized by ψ agg ; v i The attention-weighted aggregation of adjacent nodes of is expressed as Among them indicates connecting two features; the attention weights are as follows: where u() represents the convolutional layer of a 1×1 convolutional kernel parameterized by ψ update ; Extract temporal information from the corresponding node features through two LSTM layers to obtain the feature vector X g : where ⊙ represents the Hadamard product operator, c i-1 represents the memory information of the previous time step, c i represents the memory information of the current time step, h i is the output of the LSTM cell, h i-1 is the output of the previous time step; δ represents the weight matrix, and b represents the bias vector; Through the global feature extraction module, the dynamic graph convolution is responsible for processing the structural information of the graph, while the LSTM layer is used for the temporal evolution of node features, so as to comprehensively capture the global features of multi-band data with time variations.

5. The method according to claim 4, wherein The local feature extraction module is used to extract local feature information of multi-band spectrum data. The local feature extraction module includes two convolutional layers and two LSTM layers; Input the multi-band feature vector X = X1, X2,..., X m into the convolutional layer, and use the convolutional kernel to traverse the feature vector to generate a new feature vector η, and the feature vector is generated as follows: where δ represents the convolutional kernel weight vector, b represents the bias vector, and λ represents the activation function, represents the convolution operation; Then, the feature vector η is input into the LSTM to extract temporal information and obtain the feature X c ; Finally, use the adaptive gating mechanism to fuse the multi-band global features extracted by the global feature extraction module with the local features extracted by the local feature extraction module to obtain feature X f : Among them, γ represents the learned assignment weight, θ represents the Sigmoid function, X g and X c represent the weights of the multi-band global feature and the multi-band local feature respectively, and b represents the bias.

6. The method according to claim 1, wherein Step 4, implementing a multi-band spectrum prediction method based on a dynamic graph, includes: Predict the power spectral density values of n frequency bands in the next d time units Determine the power spectral density values by training a prediction model F: First, the frequency band bandwidth is selected as A - B MHz, and then the power spectrum values of the past m time units are used to predict the power spectrum values of the next d time units; in the process of constructing the graph structure, the k-nearest neighbor algorithm is used to select k nodes with the strongest correlation to construct the connection relationship; through continuous iterative optimization, the algorithm converges, and then the prediction result is output; finally, the power spectrum values obtained by the multi-band spectrum prediction module based on the dynamic graph for the next d time units are compared with the true power spectrum values at the current time, and the prediction accuracy is evaluated by calculating the error between the true value and the predicted value. The prediction performance evaluation indicators include the mean absolute error, the root mean square error, and the mean absolute percentage error; The root mean square error is obtained by calculating the mean of the squares of the differences between the predicted values and the actual observed values and taking the square root: The mean absolute error is obtained by calculating the average of the absolute values of the differences between the predicted values and the actual observed values: The mean absolute error is sensitive to the relative error, which refers to the mean absolute error between the predicted value and the true value. The determination method is as follows: X m+i represents the true value of the spectrum at the (m + i)-th time unit, represents the predicted value of the spectrum at the (m + i)-th time unit, and s represents the predicted time step; the prediction effect of the model is obtained by calculating three error metrics.

Images