Wireless channel prediction method based on machine learning and geographical environment information

Abstract

The invention discloses a wireless channel prediction method based on machine learning and geographical environment information, and belongs to the field of intelligent channel prediction. According to the method, topographic feature information and channel state information values in a topographic image are extracted from low level to high level through a multilayer structure of a convolutional neural network, and the topographic feature information and the channel state information values are coded into high-dimensional state vectors to serve as input of a full-connection neural network; and the full-connection neural network analyzes and learns a complex relationship between the topographic feature high-dimensional state vector and a channel state information CSI value of a corresponding region, so that outdoor wireless channel prediction is realized. According to the method, the accuracy of the deep neural channel prediction network is improved, and the unmanned aerial vehicle base stations can be deployed more efficiently in remote areas or disaster sites lacking infrastructures.

Description

Technical Field

[0001] The invention relates to a wireless channel prediction method based on machine learning and geographical environment information, belonging to the technical field of intelligent channel prediction. Background Art

[0002] In modern wireless communication systems, achieving widespread and efficient network coverage has always been a challenge, especially in remote areas or disaster sites where infrastructure is lacking. Traditional base station layouts rely on existing infrastructure, such as communication towers, which have high deployment costs and poor flexibility, making it difficult to quickly adapt to communication needs in emergencies or special environments. With the rapid development of Unmanned Aerial Vehicle (UAV) technology, drone-assisted communication has become an emerging solution that can be quickly deployed in environments without infrastructure support to provide temporary wireless communication services. However, the effective deployment of drone base stations requires accurate wireless channel information to optimize coverage and communication quality, especially in outdoor scenarios with complex and unknown geographical environments.

[0003] Existing channel prediction methods are mainly divided into channel prediction based on deterministic models and intelligent channel prediction based on machine learning. Although the prediction method based on deterministic modeling accurately describes the channel propagation characteristics in a specific environment, it is helpful for accurate communication system design and performance evaluation in a known environment; however, its computational complexity is high, it requires a large amount of environmental information, and it is difficult to apply to situations where environmental information is incomplete or changing. In unknown complex environments, in order to solve the limitations of the deterministic modeling channel prediction method, channel prediction research has gradually developed from deterministic channel modeling to predictive channel modeling driven by large amounts of data. In this regard, machine learning enables computers to discover patterns from data and handle nonlinear complex relationships. The relationship between terrain environmental factors and channel state information is discovered from a large amount of data, and a more accurate description of them provides more reliable and efficient channel information for UAV-assisted communication systems.

[0004] However, when deploying drone base stations in remote areas or disaster sites lacking infrastructure, the complexity of the unknown terrain (slope, slope direction, terrain height, etc.) and the height of the drone base station have a decisive effect on the performance of the communication system. Therefore, how to achieve more accurate channel prediction in remote areas or disaster sites lacking infrastructure in unknown terrain has become a key issue for the rapid deployment of drone base stations. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a wireless channel prediction method based on machine learning and geographic environment information, aiming to solve the technical problem that in an unknown complex environment, deterministic modeling channel prediction has limitations, which makes it difficult to quickly deploy drone base stations.

[0006] In order to solve the above technical problems, the technical solution of the present invention is to provide a wireless channel prediction method based on machine learning and geographical environment information, comprising the following steps:

[0007] Step S1: Acquire a scene image of the drone base station to be deployed;

[0008] Step S2: inputting the scene image into a trained convolutional neural network module, and outputting a high-dimensional state vector of terrain features;

[0009] Step S3: inputting the high-dimensional state vector into a trained fully connected neural network channel prediction model to obtain a channel prediction result of the scene image;

[0010] Step S4: Based on the channel prediction results, the evaluation of the above model is completed to achieve rapid deployment of drone base stations in the scene.

[0011] The step S1 is specifically as follows:

[0012] The digital terrain images of different areas are obtained through the digital elevation model DEM, and the contour density IL, slope S and slope aspect A information of each terrain image are extracted from the DEM using the geographic information system GIS;

[0013] Select multiple points in different terrain images to deploy ground receiving points;

[0014] The terrain feature information of each ground point, including the contour line density IL, slope S and slope aspect A, and the CSI information at the corresponding UAV altitude, including the channel capacity C, ground node receiving power P and delay D, are combined with the UAV altitude H to form a seven-dimensional data set (IL, S, A, C, P, D, H);

[0015] By repeating the above operations at multiple different points in different scenes, n sample data are obtained to construct a data set covering different drone base station heights and terrain features.

[0016] The step S2 is specifically as follows:

[0017] The input terrain image and its corresponding initial channel state information C, P, D and the UAV deployment height H are input into the convolutional neural network as joint features;

[0018] In the second convolutional layer, the convolution kernel performs intermediate abstraction on the low-level features extracted in the first layer, gradually learning the medium-scale relationship between terrain features and channel delay D, as well as the distribution trend of channel capacity C at different drone heights H. After the second convolutional layer, the maximum pooling layer is used to spatially compress the intermediate features and extract key channel-terrain correlation characteristics.

[0019] The pooled feature map is input into the third convolutional layer to perform high-level abstraction and fusion on the global relationship between terrain features IL, S, A and channel features C, P, D and drone height H. The spatial size of the feature map is further compressed through the maximum pooling layer, and the feature representation is concentrated in a smaller spatial range to reduce data redundancy, retain the global information of high-level features, and represent it with a high-dimensional state vector.

[0020] The step S3 is specifically as follows:

[0021] The high-dimensional state vector output by the convolutional neural network is input into the input layer of the fully connected neural network. The fully connected network gradually learns the potential characteristic patterns of the high-dimensional state vector through layer-by-layer processing, and predicts the channel states C, P, D and the optimal UAV deployment height H;

[0022] Perform preliminary feature combination on the input high-dimensional state vector, capture the complex nonlinear relationship between terrain features IL, S, A and channel states C, P, D and drone height H through nonlinear activation function, and generate preliminary comprehensive high-dimensional representation;

[0023] The initially extracted comprehensive features are passed to the fully connected layer to further abstract high-dimensional features and extract the deep correlation between terrain features and channel performance indicators;

[0024] The features are further compressed and optimized, the terrain features and the global pattern of channel performance are integrated, and the influence of the UAV height H on the channel state is combined to generate a feature representation with predictive capabilities for the channel state and the optimal UAV deployment height;

[0025] The comprehensive features extracted from the hidden layer are mapped to specific prediction values ​​of the channel states C, P, D and the UAV deployment height H, providing channel performance prediction for the deployment of UAV base stations and the optimal deployment height of UAVs.

[0026] The step S4 is specifically as follows:

[0027] During the model training phase, the back propagation algorithm is used to dynamically adjust the weights and biases of the convolutional network and the fully connected network to calculate the channel state values ​​predicted by the model. and drone altitude The error with its true value (C, P, D, H), the loss function Loss uses the mean square error, the formula is:

[0028]

[0029] Where n is the total number of samples, (C i ,P i ,D i ,H i) represents the actual channel state and drone height of the ith sample, The error signal is propagated back to the hidden layer and convolutional layer through the chain rule to dynamically adjust the weights and biases of each layer.

[0030] In the model optimization stage, the gradient descent method is used to iteratively update the network parameters. The update formula is:

[0031]

[0032] Among them, w t is the current weight, η is the learning rate, is the gradient of the loss function with respect to the weight;

[0033] The absolute error AE and mean absolute error MAE are used for evaluation:

[0034] The absolute error AE is calculated as follows:

[0035]

[0036] Among them, y is the actual value, is the predicted value;

[0037] The mean absolute error MAE is calculated as follows:

[0038]

[0039] Among them, n is the total number of samples, y i and are the actual value and predicted value of the i-th sample respectively;

[0040] Based on the channel prediction results, the model generates the optimal deployment strategies for different areas.

[0041] The beneficial effects of the present invention are: compared with the prior art, the present invention extracts terrain environment information from terrain images through a convolutional neural network, thereby improving the accuracy of the deep neural channel prediction network, which helps to more efficiently deploy drone base stations in remote areas or disaster sites that lack infrastructure. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] To make the implementation process of the embodiment of the present invention clearer, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a partial embodiment of the present invention, rather than all the embodiments.

[0043] Figure 1 It is a schematic diagram of the implementation process of the present invention;

[0044] Figure 2It is a scene image where a drone base station needs to be deployed in an embodiment of the present invention;

[0045] Figure 3 is the terrain feature information of the scene image in the embodiment of the present invention, Figure 3 (a) is a DEM digitized terrain image of the scene image, Figure 3 (b) Figure 3 (a) The scene image containing terrain height information is produced. Figure 3 (c) Figure 3 (a) The image containing slope information is produced. Figure 3 (d) Figure 3 (a) The resulting image contains terrain trend information;

[0046] Figure 4 It is a channel state information diagram of the communication users and drone base stations at different heights in a certain ground area in the scene image;

[0047] Figure 5 This is a structural diagram of the neural network model of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0049] Example 1: Figure 1 As shown, the wireless channel prediction method based on machine learning and geographical environment information has the following specific steps:

[0050] Step S1: Obtain a scene image of the drone base station to be deployed.

[0051] Obtain different scene images and their corresponding terrain feature information and channel state information values, and perform labeling operations on the scene images. Specifically, the data style in the constructed dataset is (IL, S, A, C, P, D, H). The first three columns represent the terrain feature information of the contour line density IL, slope S and slope direction A of the ground point; the seventh column of data H represents the deployment height of the drone. In order to reduce the amount of data, this patent considers three different heights of 800 meters, 1000 meters and 2000 meters; the 4th, 5th and 6th columns respectively represent the three CSI information of the channel capacity C, the ground node receiving power P and the delay D between the drone and the ground node at a certain height.

[0052] Acquire terrain images: obtain digital terrain images of different areas through the digital elevation model DEM, and use the geographic information system GIS to extract the contour density IL, slope S and slope aspect A information of each terrain image from the DEM.

[0053] Channel state information acquisition: Use Wireless Insite software to select multiple points in different terrain images to deploy ground receiving points. At each ground point, deploy drone base stations at altitudes of 800 meters, 1000 meters, and 2000 meters. For each combination of ground point and drone altitude, use the Wireless Insite simulation tool to measure and record the three CSI values ​​of channel capacity C, received power P, and delay D.

[0054] Dataset construction: Combine the terrain feature information (IL, S, A) of each ground point with the CSI information (C, P, D) at the corresponding drone altitude and the drone altitude (H) to form a seven-dimensional data point (IL, S, A, C, P, D, H). By repeating the above operation at multiple different points in different scenes, n sample data are obtained to construct a dataset covering different drone base station altitudes and terrain features.

[0055] like Figure 2 Shown is an image of a scene where a drone base station needs to be deployed. The present invention aims to deploy drones in remote areas or disaster sites that lack infrastructure to achieve extensive and efficient network coverage.

[0056] like Figure 3 Shown is the terrain feature information of the scene image;

[0057] in Figure 3 (a) is a DEM digitized terrain image of the scene image, Figure 3 (b) Figure 3 (a) The scene image containing terrain height information is produced. Figure 3 (c) Figure 3 (a) The image containing slope information is produced. Figure 3 (d) Figure 3 (a) The resulting image contains terrain trend information;

[0058] Figure 4 This is a channel state information diagram of the different height distributions of communication users and drone base stations in a certain ground area in the scene image. It can be seen that when the height distribution of drones is different, there will be line-of-sight transmission channels and non-line-of-sight transmission channels between mobile users and drone base stations;

[0059] in accordance with Figure 2 What you get Figure 3 and Figure 4The relationship between the different distribution heights of mobile users and drone base stations in different areas of the same scene is different, so as to construct data sets (training data sets, evaluation data sets and test data sets) of different channel state information and different scene terrain features at different drone base station heights; according to the regional channel state information corresponding to different terrain features and drone base station heights, these scene images are labeled to obtain the data sets. The constructed convolutional neural network is trained using the data sets to obtain a convolutional neural network model that can extract features from terrain feature images.

[0060] In an embodiment of the present invention, there are multiple methods for obtaining the scene image of the drone base station to be deployed, the terrain feature information of the scene image, and the channel state information of the communication users and the drone base station distributed at different heights in a certain ground area in the scene image, without limiting the specific acquisition method.

[0061] Step S2: Input the scene image into a trained convolutional neural network module, and output a high-dimensional state vector of terrain features.

[0062] Specifically, the convolutional neural network model extracts the multi-level relationship between terrain features and CSI layer by layer by inputting the constructed seven-dimensional data set, and generates a high-dimensional state vector;

[0063] First, the input terrain image and its corresponding initial channel state information C, P, D and the drone deployment height H are fed into the convolutional neural network as joint features. The first convolutional layer (Conv2D-32x3x3, ReLU) extracts low-level visual features of the terrain (such as contour boundaries, local changes in slope, and the directionality of slope) through 32 3×3 filters, and preliminarily learns the local relationship between these terrain features and the channel state information. For example, the network can capture the distribution trend of weak signals between the contour-dense area and the channel capacity C, as well as the impact of the drone height H on the channel received power P. The convolution operation of this layer enables the local features of the terrain image to be initially integrated with the channel state information. After being processed by the ReLU activation function, the network's ability to model local nonlinear relationships is enhanced. Subsequently, the extracted features are reduced in dimension through the maximum pooling layer (MaxPool2D-2x2, stride 2), highlighting the key feature areas and reducing the computational complexity.

[0064] Next, in the second convolutional layer (Conv2D-64x3x3, ReLU), the convolution kernel performs intermediate-level abstraction on the low-level features extracted in the first layer. At this point, the network can gradually learn the mesoscale relationship between terrain features and channel delay D, as well as the distribution trend of channel capacity C at different drone heights H. For example, in areas with steeper slopes A, the network can capture the obvious fluctuations in channel received power P that may occur, and further encode these intermediate features into feature maps of the convolution output. After the second convolutional layer, the intermediate features are spatially compressed again through the maximum pooling layer to extract more critical channel-terrain correlation characteristics.

[0065] Subsequently, the pooled feature map is input into the third convolutional layer (Conv2D-128x3x3, ReLU). In this layer, the network performs high-level abstraction and fusion of the global relationship between terrain features IL, S, A and channel features C, P, D and drone height H. For example, the network can integrate the complex shape of the dense contour area, the spatial fluctuation of the slope, and the overall impact of the drone height H on the channel capacity C and delay D to form a high-level representation of the coupling relationship between terrain and channel characteristics. The feature extraction results of this layer can more comprehensively reflect the impact of terrain information on the channel state distribution, and further compress the spatial size of the feature map through the maximum pooling layer (MaxPool2D-2x2, stride 2), concentrating the feature representation into a smaller spatial range, thereby reducing data redundancy and retaining the global information of high-level features, represented by a high-dimensional state vector.

[0066] Step S3: inputting the high-dimensional state vector into a trained fully connected neural network channel prediction model to obtain a channel prediction result of the scene image.

[0067] The fully connected neural network channel prediction model completes the prediction of channel state information and optimal UAV deployment height by inputting the high-dimensional state vector generated in step S2. The specific implementation is as follows:

[0068] First, the high-dimensional state vector output by the convolutional neural network is input into the input layer of the fully connected neural network. This high-dimensional state vector is a comprehensive representation of terrain feature information (contour line density IL, slope S, slope direction A), channel state information (channel capacity C, received power P, delay D) and drone height H, encapsulating the multi-level correlation between the features. The fully connected network gradually learns the potential feature patterns of the high-dimensional state vector through layer-by-layer processing, and predicts the channel states C, P, D and the optimal drone deployment height H.

[0069] In the fully connected layer (Dense-1024, ReLU), the network performs preliminary feature combination on the input high-dimensional state vector, and captures the complex nonlinear relationship between the terrain features IL, S, A and the channel states C, P, D and the drone height H through the nonlinear activation function (ReLU). The function of this layer is to integrate the original terrain features and channel features in the input vector to generate a preliminary comprehensive high-dimensional representation.

[0070] Next, the initially extracted comprehensive features are passed to the fully connected layer (Dense-512, ReLU). In this layer, the network further abstracts high-dimensional features and extracts the deep correlation between terrain features and channel performance indicators. For example, this layer can learn the local impact of high-density contour areas on channel capacity, as well as the nonlinear relationship between slope variation and received power and delay.

[0071] In the fully connected layer (Dense-256, ReLU), the network further compresses and optimizes the features to generate highly abstract low-dimensional feature representations that integrate the global patterns of terrain features and channel performance, combined with the impact of the UAV altitude H on the channel state. For example, this layer can learn the global impact of terrain structure on channel states (such as delay D) from high-dimensional features, thereby generating feature representations that have predictive capabilities for channel states and optimal UAV deployment altitudes.

[0072] Finally, in the output layer (Dense-4, Linear), the network maps the comprehensive features extracted by the hidden layer into specific prediction values ​​of the channel state (C, P, D) and the drone deployment height (H). The linear activation function ensures that the range of the predicted values ​​conforms to the actual physical meaning, such as the channel capacity and received power are positive values, and the drone height is within a reasonable range (such as 800 meters to 2000 meters). Through the output of this layer, the model can provide channel performance predictions and the optimal deployment height of drones for the subsequent deployment of drone base stations.

[0073] Step S4: Based on the channel prediction results, the evaluation of the above model is completed to achieve rapid deployment of drone base stations in the scene.

[0074] First, during the model training phase, the backpropagation algorithm (BP) is used to dynamically adjust the weights and biases of the convolutional network and the fully connected network to minimize the error between the predicted value and the actual channel state value. Specifically, the channel state value predicted by the calculation model and drone altitude The error with its true value (C, P, D, H), the loss function uses the mean square error (MSE), the formula is:

[0075]

[0076] Where n is the total number of samples, (C i ,P i ,D i ,H i ) represents the actual channel state and drone height of the ith sample, is the corresponding model prediction value. The error signal is back-propagated to the hidden layer and convolution layer through the chain rule, dynamically adjusting the weights and biases of each layer.

[0077] Secondly, in the model optimization stage, the gradient descent method (such as Adam optimizer) is used to iteratively update the network parameters, gradually reduce the error function value, and optimize the model performance. The parameter update formula is:

[0078]

[0079] Among them, w t is the current weight, η is the learning rate, is the gradient of the loss function with respect to the weight. Through multiple iterative updates, the model can effectively learn the complex relationship between terrain features, channel status, and drone altitude, improving channel prediction accuracy and deployment efficiency.

[0080] In order to verify the accuracy and reliability of the model, the absolute error AE and mean absolute error MAE are used for evaluation; the absolute error AE is calculated as follows:

[0081]

[0082] Among them, y is the actual value, is the predicted value;

[0083] The mean absolute error MAE is calculated as follows:

[0084]

[0085] Among them, n is the total number of samples, y i and are the actual value and predicted value of the i-th sample, respectively. Through the calculation of AE and MAE, the prediction accuracy of the model for channel status and drone altitude is comprehensively evaluated.

[0086] Finally, based on the channel prediction results and the recommendations for drone deployment height, the model can generate the optimal deployment strategy for different areas. For example, in areas with low channel capacity C and complex terrain, it is recommended to deploy low-altitude drone base stations to increase the received power P; in areas with large slopes and significant channel delay D, it is recommended to deploy higher drone base stations to improve delay performance; and in areas with flat terrain and large coverage requirements, medium-altitude drone base stations are recommended to balance coverage and performance. By comprehensively analyzing the distribution relationship between terrain characteristics, channel status, and drone height, the model achieves precise deployment, which is particularly suitable for the needs of improving communication quality in remote areas or disaster relief scenarios.

[0087] The above are the implementation steps of the present invention, which include the preparation of the data set used for training and evaluation, the model structure and the functions of each module, and the implementation of model training and channel result prediction;

[0088] A complete channel prediction process is as follows: Figure 1 As shown in the figure, the scene image of the drone base station to be deployed is obtained, and the image containing the terrain feature information and channel state information of the image is input into the pre-trained model. The model structure is as follows Figure 5 As shown, the channel state prediction results in different areas in the scene are obtained.

[0089] The wireless channel prediction method based on machine learning and geographic environment information in the embodiment of the present invention can complete real-time prediction of the channel by acquiring the terrain feature image in the unknown scene, which is used for the deployment of UAV base stations; the above is a detailed description of the specific implementation methods of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation methods. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A wireless channel prediction method based on machine learning and geographical environment information, characterized in that: The following steps are involved: Step S1: Acquire a scene image of the drone base station to be deployed; Step S2: inputting the scene image into a trained convolutional neural network module, and outputting a high-dimensional state vector of terrain features; Step S3: inputting the high-dimensional state vector into a trained fully connected neural network channel prediction model to obtain a channel prediction result of the scene image; Step S4: Based on the channel prediction results, the evaluation of the above model is completed to achieve rapid deployment of drone base stations in the scene.

2. The wireless channel prediction method based on machine learning and geographic environment information according to claim 1, characterized in that: The step S1 is specifically as follows: The digital terrain images of different areas are obtained through the digital elevation model DEM, and the contour density IL, slope S and slope aspect A information of each terrain image are extracted from the DEM using the geographic information system GIS; Select multiple points in different terrain images to deploy ground receiving points; The terrain feature information of each ground point, including the contour line density IL, slope S and slope aspect A, and the CSI information at the corresponding UAV altitude, including the channel capacity C, ground node receiving power P and delay D, are combined with the UAV altitude H to form a seven-dimensional data set (IL, S, A, C, P, D, H); By repeating the above operations at multiple different points in different scenes, n sample data are obtained to construct a data set covering different drone base station heights and terrain features.

3. The wireless channel prediction method based on machine learning and geographic environment information according to claim 1, characterized in that: The step S2 is specifically as follows: The input terrain image and its corresponding initial channel state information C, P, D and the UAV deployment height H are input into the convolutional neural network as joint features; In the second convolutional layer, the convolution kernel performs intermediate abstraction on the low-level features extracted in the first layer, gradually learning the medium-scale relationship between terrain features and channel delay D, as well as the distribution trend of channel capacity C at different drone heights H. After the second convolutional layer, the maximum pooling layer is used to spatially compress the intermediate features and extract key channel-terrain correlation characteristics. The pooled feature map is input into the third convolutional layer to perform high-level abstraction and fusion on the global relationship between terrain features IL, S, A and channel features C, P, D and drone height H. The spatial size of the feature map is further compressed through the maximum pooling layer, and the feature representation is concentrated in a smaller spatial range to reduce data redundancy, retain the global information of high-level features, and represent it with a high-dimensional state vector.

4. The wireless channel prediction method based on machine learning and geographic environment information according to claim 1, characterized in that: The step S3 is specifically as follows: The high-dimensional state vector output by the convolutional neural network is input into the input layer of the fully connected neural network. The fully connected network gradually learns the potential characteristic patterns of the high-dimensional state vector through layer-by-layer processing, and predicts the channel states C, P, D and the optimal UAV deployment height H; Perform preliminary feature combination on the input high-dimensional state vector, capture the complex nonlinear relationship between terrain features IL, S, A and channel states C, P, D and drone height H through nonlinear activation function, and generate preliminary comprehensive high-dimensional representation; The initially extracted comprehensive features are passed to the fully connected layer to further abstract high-dimensional features and extract the deep correlation between terrain features and channel performance indicators; The features are further compressed and optimized, the terrain features and the global pattern of channel performance are integrated, and the influence of the UAV height H on the channel state is combined to generate a feature representation with predictive capabilities for the channel state and the optimal UAV deployment height; The comprehensive features extracted from the hidden layer are mapped to specific prediction values ​​of the channel states C, P, D and the UAV deployment height H, providing channel performance prediction for the deployment of UAV base stations and the optimal deployment height of UAVs.

5. The wireless channel prediction method based on machine learning and geographic environment information according to claim 1, characterized in that: The step S4 is specifically as follows: During the model training phase, the back propagation algorithm is used to dynamically adjust the weights and biases of the convolutional network and the fully connected network to calculate the channel state values ​​predicted by the model. and drone altitude The error with its true value (C, P, D, H), the loss function Loss uses the mean square error, the formula is: Where n is the total number of samples, (C i ,P i ,D i ,H i ) represents the actual channel state and drone height of the ith sample, The error signal is propagated back to the hidden layer and convolutional layer through the chain rule to dynamically adjust the weights and biases of each layer. In the model optimization stage, the gradient descent method is used to iteratively update the network parameters. The update formula is: Among them, w t is the current weight, η is the learning rate, is the gradient of the loss function with respect to the weight; the absolute error AE and the mean absolute error MAE are used for evaluation: The absolute error AE is calculated as follows: Among them, y is the actual value, is the predicted value; The mean absolute error MAE is calculated as follows: Among them, n is the total number of samples, y i and are the actual value and predicted value of the i-th sample respectively; Based on the channel prediction results, the model generates the optimal deployment strategies for different areas.

Images