Mobile communication network load prediction and scheduling method and system based on deep learning

Abstract

The invention relates to the technical field of network load prediction and scheduling methods, in particular to a mobile communication network load prediction and scheduling method and system based on deep learning, and the method comprises the steps: obtaining data from a plurality of edge computing nodes, including base station data, user behavior data and environment parameter data; constructing a multi-layer model for load prediction and resource scheduling based on cooperative work of the central system and the edge system; carrying out feature extraction by utilizing a dynamic feature embedding transformation module in the edge computing node; recursive spatial-temporal feature processing is carried out based on the optimized features; generating a network load prediction result through the prediction model; and resource scheduling optimization is performed through a reinforcement learning model based on a prediction result, and multi-dimensional features are nested and mapped to a high-dimensional feature space through dynamic feature embedding of original data. And the nested features are further optimized through a time-varying collaborative tensor optimization algorithm, so that the spatial relevance between different base stations can be enhanced while complex nonlinear features are captured.

Description

Technical Field

[0001] The present invention relates to the technical field of network load prediction and scheduling methods, and more specifically, to a method and system for predicting and scheduling mobile communication network load based on deep learning. Background Art

[0002] Intelligent classification of geological data is a key task in the field of geological exploration and resource assessment. With the rapid development of exploration technology, especially the widespread application of 5G networks, the data traffic in mobile communication networks has grown exponentially. The increase in the number of users and the diversification of application scenarios make network load management a crucial issue in mobile communication networks. Currently, operators usually use traditional load balancing and resource scheduling technologies, such as static threshold control or prediction methods based on historical traffic data, to alleviate network congestion and optimize resource utilization. However, these methods have significant limitations.

[0003] Traditional load forecasting methods mostly rely on time series models, such as the autoregressive integrated moving average model (ARIMA) or simple neural network models. These methods usually assume that network traffic has stationary and linear characteristics, and therefore perform poorly when faced with complex spatiotemporal dynamic changes. At the same time, these methods often ignore the importance of multi-dimensional features such as user behavior data and environmental parameters, resulting in low prediction accuracy. In addition, existing methods have shortcomings in multi-region and multi-base station collaborative optimization, and cannot fully utilize distributed computing resources, which easily leads to resource waste and local network overload.

[0004] In terms of resource scheduling, traditional methods usually adopt rule-driven static policies, such as resource allocation based on predefined thresholds. However, static rules are difficult to adapt to dynamically changing network environments, especially in high-load or burst traffic scenarios, which can easily lead to low resource utilization or reduced user quality of service (QoS). In addition, these methods usually lack effective modeling of the balance between resource allocation and user QoS, resulting in the inability to optimize resource scheduling efficiency and user experience at the same time.

[0005] In response to the above problems, some studies have tried to introduce deep learning and reinforcement learning technologies to improve them. For example, using recurrent neural networks (RNN) for load forecasting or deep Q learning (DQN) to optimize resource allocation. However, these methods still have the following technical problems in practical applications:

[0006] 1. Insufficient modeling of spatiotemporal features: Existing methods usually only perform linear or simple nonlinear modeling on time series data, ignoring the spatial correlation between different base stations and the complex dynamic changes of user behavior, making it difficult to accurately predict the load distribution in complex network environments.

[0007] 2. Low model training efficiency: In the multi-node collaborative optimization process, the traditional centralized training method requires all data to be uploaded to the central system for processing, which not only increases data transmission overhead but also brings the risk of user privacy leakage.

[0008] 3. Inflexible resource scheduling strategies: Although traditional reinforcement learning algorithms can optimize resource allocation to a certain extent, they usually only focus on short-term benefits and lack effective modeling of the long-term balance between user QoS and resource costs. Summary of the invention

[0009] In view of the above problems in the prior art, the present invention proposes a method and system for predicting and scheduling the load of a mobile communication network based on deep learning. Through innovative algorithms and system architecture design, the present invention solves how to accurately model multi-dimensional spatiotemporal features in a complex dynamic network environment to achieve high-precision network load prediction, how to improve model training efficiency in a multi-node distributed environment while protecting user privacy, how to maximize resource utilization while ensuring user QoS through a dynamically optimized resource scheduling strategy, and fundamentally solves the problems of insufficient spatiotemporal feature modeling, low model training efficiency, and inflexible resource scheduling strategies in the prior art, providing a more efficient and accurate solution for load prediction and resource scheduling of mobile communication networks.

[0010] The present invention provides a mobile communication network load prediction and scheduling method based on deep learning, comprising:

[0011] Obtain data from multiple edge computing nodes, including base station data, user behavior data, and environmental parameter data;

[0012] Based on the collaboration between the central system and the edge system, a multi-layer model for load forecasting and resource scheduling is built;

[0013] The dynamic feature embedding transformation module in the edge computing node is used for feature extraction;

[0014] Recursive spatiotemporal feature processing based on optimized features;

[0015] Generate network load prediction results through prediction models;

[0016] Resource scheduling optimization is performed through reinforcement learning model based on prediction results.

[0017] Preferably, the dynamic feature embedding transformation module comprises the following steps:

[0018] The input data is dynamically transformed into high-dimensional features through a multi-layer nested nonlinear kernel mapping algorithm to obtain nested features. The formula is as follows:

[0019]

[0020] in, is the k+1th layer nested feature, λ k is the time decay factor, is the k-th layer nonlinear kernel mapping function, X τ is the input data at time τ, is the k-th layer nested feature.

[0021] As a preference, further, the nested features are optimized by a time-varying collaborative tensor optimization algorithm to obtain a time-varying optimal weight tensor The formula is as follows:

[0022]

[0023] in, is the time-varying optimal weight tensor, is a nested feature tensor, T i is the dynamic mode tensor, is the tensor outer product, and β is the regularization parameter.

[0024] As a preferred method, based on the optimized features, a recursive adaptive graph optimization model is used to process the spatiotemporal correlation of the features to obtain the hidden state vector H t , the formula is as follows:

[0025]

[0026] Among them, H t is the hidden state vector at time t, H t-1 is the hidden state at time t-1, W g Optimize the weight matrix for the graph, G(H t-1 ) is the graph embedding function, W z is the input weight matrix, b g is the bias vector and σ is the nonlinear activation function.

[0027] As a preferred method, the load prediction result is adjusted by using a nonlinear partial differential error feedback optimization algorithm to obtain an adjusted prediction result. The formula is as follows:

[0028]

[0029] Among them, ΔY t is the current error adjustment, α is the second-order partial derivative coefficient, γ is the error amplification coefficient, η is the dynamic feedback coefficient, is the predicted value, is the true value, ΔY t -1 is the error at the previous moment.

[0030] Preferably, the adjusted prediction results are used as input to generate the optimal resource allocation strategy R through the constraint optimization reinforcement learning algorithm. * , the following company

[0031]

[0032] Among them, R * is the optimal resource allocation strategy, U i (R i ) is the utility function of user i, C j (R j ) is the resource cost function, and λ is the penalty coefficient.

[0033] Preferably, the edge computing node further includes a data preprocessing module, which performs standardization processing on the collected base station data and user behavior data to form a feature vector in a unified format, and the feature vector serves as an input of the dynamic feature embedding transformation module.

[0034] Preferably, the central system collaboratively optimizes the training data of multiple edge computing nodes through a federated learning framework, including initialization, distribution and gradient aggregation of model parameters, to build a global model to optimize the prediction performance of edge nodes.

[0035] The deep learning-based mobile communication network load prediction and intelligent scheduling system that executes the method includes a central system and an edge system. The central system is used to achieve global model optimization and cross-regional collaborative modeling. The edge system includes multiple edge computing nodes. The edge computing nodes are used to collect data, execute prediction models and perform resource scheduling.

[0036] Preferably, the central system further integrates the model training data of each edge computing node through a cloud computing resource pool to achieve gradient aggregation across nodes and global model updating.

[0037] The present invention has the following beneficial effects:

[0038] 1. Synergy of dynamic feature embedding and spatiotemporal optimization: This paper proposes a multi-layer nested nonlinear kernel mapping algorithm, which embeds the dynamic features of the original data and maps the multidimensional features into a high-dimensional feature space. The nested features are further optimized by a time-varying collaborative tensor optimization algorithm, which can enhance the spatial correlation between different base stations while capturing complex nonlinear features. This feature modeling method not only solves the problem that traditional methods are difficult to handle dynamic spatiotemporal data, but also significantly improves the accuracy of load prediction.

[0039] 2. Efficient distributed training of the federated learning framework: The present invention adopts a federated learning framework and realizes distributed model training through the coordinated optimization of edge computing nodes and the central system. This framework does not need to upload data centrally, and uses distributed gradient aggregation to optimize the global model, which not only reduces data transmission overhead, but also effectively protects user privacy. At the same time, by dynamically adjusting the learning rate and regularization parameters, the model training efficiency is further improved. Experiments show that the training time of the present invention in a multi-node environment is 38% shorter than that of the traditional centralized method.

[0040] 3. Resource scheduling strategy of constraint optimization reinforcement learning: This paper proposes a constraint optimization reinforcement learning algorithm, which can achieve the optimal decision of resource scheduling in a dynamic network environment. By establishing a balance model of utility function and cost function, the algorithm achieves the optimal balance between user QoS guarantee and resource utilization. At the same time, through the nonlinear partial differential error feedback optimization algorithm, the load prediction results are dynamically adjusted, which effectively improves the adaptability and robustness of resource scheduling.

[0041] 4. Synergy and superposition of algorithm mechanisms: The modules and algorithms of the present invention are closely combined in a complementary and superimposed manner to form a complete prediction and scheduling system. Dynamic feature embedding provides high-quality input for spatiotemporal optimization. The optimized features are further refined through the recursive graph optimization model, and are ultimately used for dynamic decision-making of resource scheduling strategies. Experimental data show that the present invention is significantly superior to the prior art in terms of load prediction accuracy, model training efficiency, and resource scheduling effect.

[0042] The present invention successfully solves many core problems in the prior art through a variety of innovative algorithms and system designs. In particular, it provides a new solution in terms of spatiotemporal feature modeling, multi-node collaborative optimization, and dynamic optimization of resource scheduling. Compared with traditional methods, the present invention not only significantly improves the load prediction and resource scheduling performance of mobile communication networks, but also provides technical support for system optimization in complex dynamic network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a logic block diagram of the overall system architecture of the present invention.

[0044] Figure 2 This is a logic block diagram of the edge computing node of the present invention.

[0045] Figure 3 This is a logical block diagram of the federated learning framework of the present invention.

[0046] Figure 4 This is a logic block diagram of the dynamic feature embedding and load prediction of the present invention.

[0047] Figure 5 This is a logical block diagram of resource scheduling of the present invention. DETAILED DESCRIPTION

[0048] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0049] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0050] Clear reference Figure 1-5 The present invention provides a method for predicting and scheduling mobile communication network load based on deep learning. The overall architecture includes a central system 1 and multiple edge computing nodes 2. The central system 1 works in collaboration with the edge system to undertake global model optimization and local data processing, prediction and resource scheduling tasks respectively. First, the edge computing node 2 obtains multi-dimensional data sources from the vicinity of the operator's base station, including base station data, user behavior data and environmental parameters. The acquired data is processed by the data preprocessing module 3 and standardized into a unified format as the input feature of the model.

[0051] Under the federated learning framework of the central system 1, the model is distributed to each edge computing node 2 for local training, and the local gradient is uploaded to the central system 1 for gradient aggregation, and finally a global optimization model is formed. Through this collaborative approach, a balance between data privacy protection and global prediction accuracy can be achieved. Preferably, in one embodiment, the deployment range of the edge computing node 2 covers a high-density user area, and the data collection frequency of each node is 5 seconds. Experience shows that this collection frequency can reduce data redundancy while ensuring the real-time performance of the model.

[0052] The present invention can effectively improve load prediction accuracy and reduce system delays through multi-node distributed collaborative optimization.

[0053] Preferably, the dynamic feature embedding transformation module of the present invention performs high-dimensional feature transformation on time series data through a multi-layer nested nonlinear kernel mapping algorithm. It is described by the following formula:

[0054]

[0055] in, is the k+1th layer nested feature, λ k K is the time decay factor, and its empirical value is preferably between 0.01 and 0.1 to balance the influence of short-term and long-term characteristics.k It is a kernel mapping function used for nonlinear feature transformation. Common choices include Gaussian kernel or polynomial kernel. The role of this algorithm is to map low-dimensional features to high-dimensional feature space to capture more complex spatiotemporal relationships.

[0056] Through the dynamic nested calculation of the above formula, the model's ability to express nonlinear time series data can be enhanced. According to experimental verification, in an actual network environment, the prediction accuracy of the optimized nested features is improved by about 12% compared to the single-layer feature mapping model.

[0057] Furthermore, the nested feature The optimization is performed through the time-varying cooperative tensor optimization algorithm, and the formula is as follows:

[0058]

[0059] in, is the optimal weight tensor used to optimize feature expression, T i is the dynamic mode tensor, represents the tensor outer product, and β is the regularization parameter. It is preferably between 0.001 and 0.01. This optimization algorithm effectively captures the time-varying dynamic patterns in the features and reduces redundant information through dynamic pattern decomposition.

[0060] In application scenarios, when the network environment fluctuates greatly, the model using this tensor optimization algorithm can adapt to dynamic changes more quickly, and its model convergence time is reduced by 15% compared to traditional linear regression.

[0061]

[0062] Among them, H t is the hidden state vector at time t, H t-1 is the hidden state at time t-1, W g Optimize the weight matrix for the graph, preferably in the range of 0.5 to 1, W z is the input weight matrix, b g is the bias vector, σ is the nonlinear activation function; G(H t-1 ) is a graph embedding function, which is used to map the hidden state of the previous moment to the current graph structure. Through recursive optimization, the model's ability to capture spatiotemporal features is enhanced, further improving the load prediction accuracy.

[0063] Experiments show that in a congested network environment, the prediction error of this model is reduced by about 20% compared with the model without graph optimization. Especially in the case of uneven user distribution, its prediction effect is more significant.

[0064] Finally, the model uses a nonlinear partial differential error feedback optimization algorithm to adjust the load prediction results. It is described by the following formula:

[0065]

[0066] Among them, ΔY t is the current error adjustment value η is the dynamic feedback coefficient, is the predicted value, is the true value, ΔY t -1 is the error at the previous moment; α is the second-order partial derivative coefficient, preferably 0.1 to 0.5, which can balance the short-term and long-term feedback of the error; γ is the error amplification coefficient, which is used to amplify the amplitude of the error adjustment, and the empirical value is 1.5 to 2; η is the dynamic feedback coefficient, preferably 0.3 to 0.7.

[0067] Through this algorithm, accurate feedback can be given to the prediction error. Experimental data show that in multiple actual network tests, this method can effectively reduce the fluctuation of prediction error, with the error reduction rate reaching 25%.

[0068] In a preferred embodiment of the present invention, the adjusted prediction result It is further used for resource scheduling optimization based on reinforcement learning. Specifically, the optimal resource allocation strategy R is determined by the constraint optimization reinforcement learning algorithm. * , the formula is as follows:

[0069]

[0070] Among them, R * represents the optimal resource allocation strategy, U i (R i ) is user i in resource R i The utility function under , the utility function can be dynamically adjusted according to the user's service level requirements and network load. i (R i )Using the logarithmic function form i (R i )=log(1+1)R i ), to reflect the law of diminishing marginal benefits of resources. At the same time, C j (R j ) represents the cost function of resource j, preferably a linear function C j (R j )=c j ·R j , where c j is the unit resource cost. λ is the penalty coefficient, which is used to balance user utility and resource cost. Its empirical value is usually between 0.5 and 1.0.

[0071] Through the above constraint optimization reinforcement learning algorithm, network resources can be dynamically allocated to different users. Experiments show that this method can effectively improve user QoS in high-load scenarios and increase resource utilization by about 18%.

[0072] The edge computing node 2 includes a data preprocessing module 3, which is used to standardize the collected base station data, user behavior data and environmental parameters to generate a feature vector in a unified format. Specifically, the data preprocessing includes the following steps:

[0073] First, the base station data and user behavior data are collected at 5-second intervals. To avoid the influence of data noise, a sliding average filter is preferably used to smooth the data, and the filter window size is 3 to 5 data points. Then, the data is normalized and the feature values ​​are mapped to the interval [0,1] to eliminate the dimensional differences between different features. Preferably, the normalization adopts the following formula:

[0074]

[0075] Where x is the original eigenvalue, x′ is the normalized eigenvalue, min(X) and max(X) represent the minimum and maximum values ​​of feature X, respectively.

[0076] The role of the data preprocessing module is to improve the consistency and accuracy of the model input features, thereby significantly improving the performance of the prediction model. According to experimental data, standardized data can increase the convergence speed of model training by about 15%.

[0077] The central system 1 implements the collaborative optimization of multiple edge computing nodes 2 through the federated learning framework, including the following key processes:

[0078] First, the central system 1 initializes the global model parameter θ0 and distributes it to each edge computing node 2. Each edge computing node is based on the local data set D i Perform local model training and update the local gradient g i The local gradients are then aggregated and the global model parameters are updated by the central system 1 through the following formula:

[0079]

[0080] Where η is the learning rate, preferably in the range of 0.01 to 0.05, i The data weight of each edge node is usually proportional to the amount of node data.

[0081] The advantage of the above-mentioned federated learning framework is that it uses distributed training to improve the accuracy of the global model while protecting user privacy. In practical applications, this framework can improve the prediction accuracy of the global model by more than 10% when data is unevenly distributed.

[0082] The system architecture of the present invention includes a central system 1 and multiple edge computing nodes 2, wherein the central system 1 is mainly responsible for global model optimization and cross-regional collaborative modeling. Preferably, the central system is composed of a cloud computing resource pool, which is used to integrate the data and model gradients uploaded by the edge computing nodes 2 and perform high-performance global optimization.

[0083] The edge computing node 2 is deployed near the operator's base station and is responsible for data collection, preprocessing, prediction and resource scheduling. Each node consists of a data collection module, a dynamic feature embedding transformation module and a resource scheduling module, which respectively undertake functions at different stages.

[0084] The beneficial effect of the overall system architecture is that through the combination of edge computing and cloud computing, efficient collaborative optimization of mobile communication network load prediction and resource scheduling is achieved.

[0085] In a preferred embodiment, the central system 1 integrates the model training data of each edge computing node 2 through the cloud computing resource pool to perform global model optimization. The core of global model optimization is the federated learning algorithm based on gradient aggregation, which specifically includes the following stages:

[0086] 1. Model parameter initialization: distribute the initial model parameters to edge nodes;

[0087] 2. Local model training: Each edge node trains based on local data and calculates gradients;

[0088] 3. Gradient aggregation: The central system receives local gradients uploaded by all nodes and updates the global model parameters through weighted aggregation.

[0089] The above global optimization process can improve the prediction accuracy of the global model without directly accessing local data, ensuring data security and privacy protection.

[0090] In order to verify the superiority of the mobile communication network load prediction and scheduling method based on deep learning of the present invention, this application is tested with the following data, including real-time network traffic data from multiple base stations, user behavior data and environmental parameters. The test indicators include the following four aspects:

[0091] 1. Load forecast accuracy: measures the closeness between the predicted value and the true value, using the mean absolute percentage error (MAPE) as the measurement standard:

[0092]

[0093] in, is the predicted value, is the true value.

[0094] 2. Model training convergence time: measures the time required for the model to reach the target convergence accuracy (error less than 0.01).

[0095] 3. Resource scheduling efficiency: measured by resource utilization rate (RUR), which is defined as follows:

[0096]

[0097] 4. User QoS guarantee: User experience is measured by network latency and throughput. Latency uses the average network latency (Average Latency), and throughput uses the amount of data successfully transmitted per unit time (Throughput)

[0098] In the embodiment, the method of the present invention is adopted, including core algorithms such as dynamic feature embedding transformation, multi-layer spatiotemporal optimization and constraint optimization reinforcement learning. The experimental environment is that multiple edge nodes are deployed in the core area of ​​the city, and the data collection frequency is 5 seconds. The central system performs federated learning model optimization in a cloud environment, using a learning rate η = 0.03 and a regularization parameter β = 0.01.

[0099] The comparative example adopts the existing GRU combined with static feature extraction and centralized training model. No dynamic feature nesting transformation is adopted, and no time-varying cooperative tensor optimization and constraint reinforcement learning resource scheduling strategy is used. The experimental environment is consistent with the embodiment.

[0100] Under the same data and experimental conditions, the test results of the embodiment and the comparative example are shown in Table 1:

[0101] Test indicators Example Comparative Example Load prediction accuracy (MAPE) 5.8％ 13.2％ Model training convergence time (seconds) 420 680 Resource scheduling efficiency (RUR) 92.5％ 78.6％ Average network delay (ms) 23 45 Throughput (Mbps) 152 115

[0102] The above test results are analyzed and explained as follows:

[0103] 1. Load prediction accuracy (MAPE): The dynamic feature embedding transformation and multi-layer spatiotemporal optimization of the present invention significantly improve the model's ability to express complex spatiotemporal features. In the multi-region load prediction task of the CellTrac2024 dataset, the MAPE of the embodiment is only 5.8%, which is 56% lower than the 13.2% of the control example. This shows that the present invention has higher prediction accuracy in complex network traffic prediction scenarios.

[0104] 2. Model training convergence time: Since the present invention adopts federated learning and time-varying collaborative tensor optimization, it can quickly integrate the training results of each edge node to achieve efficient global optimization. Experiments show that the model training convergence time of the embodiment is 420 seconds, which is about 38% higher than the 680 seconds of the comparative example. This not only shortens the model launch time, but also improves the real-time performance of the system.

[0105] 3. Resource scheduling efficiency (RUR): Through the constraint optimization reinforcement learning algorithm, the present invention realizes the dynamic optimal allocation of resources. The test results show that the resource scheduling efficiency of the embodiment is as high as 92.5%, which is significantly better than 78.6% of the comparative example. This means that the present invention can more efficiently utilize limited resources and reduce resource waste under high network load conditions.

[0106] 4. User QoS guarantee: In terms of QoS guarantee, the average network delay of the embodiment is 23ms, which is significantly lower than the 45ms of the comparison, indicating that users experience more smoothly under peak load. At the same time, in terms of throughput, the embodiment reaches 152Mbps, while the comparison is only 115Mbps, reflecting a stronger data transmission capacity. This further proves the significant effect of the present invention in optimizing user experience.

[0107] In summary, the embodiments are significantly better than the control examples in core indicators such as load prediction, model convergence, resource scheduling and QoS guarantee, which fully verifies the superiority of the present invention. The method of the present invention provides an efficient and reliable solution for load prediction and resource scheduling of mobile communication networks through multi-level and multi-module collaborative optimization, and has extremely high promotion value in practical applications.

[0108] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A mobile communication network load prediction and scheduling method based on deep learning, characterized in that: include: Obtain data from multiple edge computing nodes, including base station data, user behavior data, and environmental parameter data; Based on the collaboration between the central system and the edge system, a multi-layer model for load forecasting and resource scheduling is built; The dynamic feature embedding transformation module in the edge computing node is used for feature extraction; Recursive spatiotemporal feature processing based on optimized features; Generate network load prediction results through prediction models; Resource scheduling optimization is performed through reinforcement learning model based on prediction results.

2. The method according to claim 1, characterized in that The dynamic feature embedding transformation module comprises the following steps: The input data is dynamically transformed into high-dimensional features through a multi-layer nested nonlinear kernel mapping algorithm to obtain nested features. The formula is as follows: in, is the k+1th layer nested feature, λ k is the time decay factor, is the k-th layer nonlinear kernel mapping function, X τ is the input data at time τ, is the k-th layer nested feature.

3. The method according to claim 2, characterized in that The nested features are optimized by a time-varying collaborative tensor optimization algorithm to obtain a time-varying optimal weight tensor The formula is as follows: in, is the time-varying optimal weight tensor, is a nested feature tensor, T i is the dynamic mode tensor, is the tensor outer product, and α is the regularization parameter.

4. The method according to claim 3, characterized in that Based on the optimized features, the recursive adaptive graph optimization model is used to process the spatiotemporal correlation of the features to obtain the hidden state vector H t , the formula is as follows: Among them, H t is the hidden state vector at time t, H t-1 is the hidden state at time t-1, W g Optimize the weight matrix for the graph, G(H t-1 ) is the graph embedding function, W z is the input weight matrix, b g is the bias vector, and σ is the nonlinear activation function.

5. The method according to claim 4, characterized in that The load prediction result is adjusted by the nonlinear partial differential error feedback optimization algorithm to obtain the adjusted prediction result. The formula is as follows: Among them, ΔY t is the current error adjustment, α is the second-order partial derivative coefficient, γ is the error amplification coefficient, η is the dynamic feedback coefficient, is the predicted value, is the true value, ΔY t -1 is the error at the previous moment.

6. The method according to claim 5, characterized in that The adjusted prediction results are used as input to generate the optimal resource allocation strategy R through the constraint optimization reinforcement learning algorithm. * , the following company Among them, R * is the optimal resource allocation strategy, U i (R i ) is the utility function of user i, C j (R j ) is the resource cost function, and λ is the penalty coefficient.

7. The method according to claim 1, characterized in that The edge computing node further includes a data preprocessing module, which performs standardization processing on the collected base station data and user behavior data to form a feature vector in a unified format, and the feature vector is used as an input of the dynamic feature embedding transformation module.

8. The method according to claim 1, characterized in that: The central system collaboratively optimizes the training data of multiple edge computing nodes through a federated learning framework, including initialization, distribution and gradient aggregation of model parameters, to build a global model to optimize the prediction performance of edge nodes.

9. A mobile communication network load prediction and intelligent scheduling system based on deep learning that executes the method according to any one of claims 1 to 8, characterized in that: It includes a central system and an edge system. The central system is used to realize global model optimization and cross-regional collaborative modeling. The edge system includes multiple edge computing nodes. The edge computing nodes are used to collect data, execute prediction models and perform resource scheduling.

10. The system according to claim 9, characterized in that The central system further integrates the model training data of each edge computing node through the cloud computing resource pool to achieve cross-node gradient aggregation and global model update.

Images