Attention-based model for wireless digital twinning

Abstract

The invention relates to performance prediction in a network. The present disclosure proposes a computer-implemented method for performance prediction in a wireless communication network, the method comprising: obtaining historical performance data of the network, where the historical performance data comprises a plurality of sets of performance data associated with at least one network performance indicator, each group of performance data is collected at a first sampling granularity in a time period, and each piece of performance data included in the group of performance data is collected at a second sampling granularity in the time period; and predicting at least one piece of future performance data based on the plurality of sets of performance data. The invention also provides a wireless digital twinning of a wireless network system.

Description

Technical Field

[0001] This disclosure relates to wireless networks, and more particularly to network optimization and testing in L4-L5 autonomous driving communication networks. To improve the prediction of communication network performance, thereby optimizing network design and configuration, this disclosure proposes a computer-implemented method and a wireless digital twin for a wireless network system. Background Technology

[0002] Wireless network digital twins can be used for network optimization and testing in L4-L5 communication networks for autonomous driving. Figure 1 A schematic diagram of an autonomous driving network hierarchy is shown. Wireless network digital twins simulate different scenarios and configurations and evaluate their impact on network performance and key performance indicators (KPIs) by creating virtual copies of the physical network.

[0003] Using digital twins of wireless networks enables proactive maintenance of communication networks, detecting potential failures and performance degradation before they occur and minimizing downtime and service interruptions. This is particularly important for safety-critical applications such as banking and autonomous vehicles, where even brief communication outages can have serious consequences.

[0004] In real-world scenarios, wireless network digital twins can predict the performance of the communication network of autonomous vehicles through simulation and modeling. To obtain a highly accurate model, there are two challenges: (1) the model needs to have high sensitivity to parameters (e.g., actions), and (2) the data from the real network contains a high level of noise.

[0005] There are two existing approaches to constructing digital twins of wireless network systems. The first approach proposes a method for predicting cellular network throughput. Figure 2 The process of cellular network throughput prediction is illustrated. In this work, multiple machine learning models (e.g., Random Forest (RF), Support Vector Machine (SVM), Multi Layers Perceptron (MLP)) are trained, and then all trained models are evaluated based on several criteria (e.g., Absolute Percentage Error (APE) and Coefficient of Determination (R-squared)). A model is selected based on the evaluation results, and the history length and prediction horizon are determined. However, this scheme only considers short-term predictions (up to 12 seconds) and does not account for parameter sensitivity.

[0006] The second approach proposes a network performance forecasting method. This method assesses how KPIs change with traffic (e.g., users might be interested in the impact of increased traffic on call drop rates), and then constructs a regression model to model the relationship between one KPI and other KPIs. However, the limitation of the second approach is that it relies on assumptions and estimates of traffic growth rates for future forecasts and does not consider parameter sensitivity.

[0007] Therefore, an improved approach is needed to construct digital twins of wireless communication networks. Summary of the Invention

[0008] Given the aforementioned limitations, this disclosure aims to construct an advanced wireless digital twin for network optimization and testing. Specifically, one objective is to provide a highly accurate model for the wireless digital twin. Another objective is to achieve long-term network performance prediction. Yet another objective is to construct a model that is highly sensitive to parameters.

[0009] These and other objectives are achieved by the solutions provided in this disclosure in the independent claims. Advantageous implementations are further defined in the dependent claims.

[0010] A first aspect of this disclosure provides a computer-implemented method for performance prediction in a wireless communication network. The method includes: acquiring historical performance data of the network, wherein the historical performance data includes multiple sets of performance data associated with at least one network performance metric, wherein each set of performance data is collected at a first sampling granularity within a time period, and each piece of performance data included in the set of performance data is collected at a second sampling granularity within the time period; and predicting at least one future performance data based on the multiple sets of performance data.

[0011] This disclosure presents an advanced scheme for constructing wireless digital twins for wireless communication networks. It enables long-term forecasting, particularly the forecasting of long-term KPIs (e.g., traffic and throughput of wireless networks). Specifically, this disclosure proposes capturing network patterns at different time scales, i.e., collecting performance data with different sampling granularities. Forecasting is performed using data captured at different time scales.

[0012] For example, each set of performance data collected at a first sampling granularity can be considered to represent the network's daily pattern. In this example, the time period is considered as one day. The network's daily pattern behavior can be learned and modeled. Additionally, hourly patterns of KPIs are also collected. Each performance data point included in this set of performance data, collected at a second sampling granularity, represents the network's daily pattern.

[0013] In one implementation of the first aspect, the prediction step includes applying at least one machine learning model to the plurality of sets of performance data.

[0014] Optionally, predictions can be performed based on one or more machine learning models.

[0015] In one implementation of the first aspect, the at least one machine learning model includes a first attention-based model and / or a second attention-based model, each model including the following input components: query, key, and value.

[0016] It should be noted that a popular artificial neural network block based on the attention mechanism comprises three types of components: query, key, and value. The attention mechanism calculates the similarity between the query and the key, and uses this similarity as a weight to generate a weighted sum of values. The attention mechanism is a fundamental block of the Transformer model. This disclosure also proposes a novel module for modeling patterns in acquired data, which can be called "AttenInAtten". This AttenInAtten neural network architecture can have two attention layers. Possibly, each attention layer can be based on one of a first attention-based model and a second attention-based model.

[0017] In one implementation of the first aspect, each set of performance data includes state-action pairs indicating each performance data point, each state being associated with at least one network performance metric, and each action being associated with at least one network parameter.

[0018] For example, the collected data can be in a state-action pair format, i.e. ,in, For the i-th state, Let i be the i-th action.

[0019] In one implementation of the first aspect, the at least one future performance data includes a state-action pair, and the method includes: obtaining the action of the at least one future performance data; and using a linear projection of the action of the at least one future performance data as the query to predict the state of the at least one future performance data by applying the first attention-based model to the plurality of sets of performance data.

[0020] To be useful for network optimization, the model designed according to this disclosure further exhibits a high sensitivity to parameters. It is understandable that the state-action pairs from historical data and the network's latest actions (e.g., ...) are highly sensitive to these parameters. It can be used to predict the future state of a network, for example, The value of . Since actions are used as input "queries," the output of the attention mechanism is heavily influenced by the actions (queries). It should be noted that network actions are controlled by, for example, network maintenance personnel. An action that has been performed or will be performed is considered known to the network.

[0021] In one implementation of the first aspect, the step of predicting the state of the at least one future performance data by applying the first attention-based model to the plurality of sets of performance data further includes: using a first linear projection of the state-action pair of each set of performance data as the key, and using a second linear projection of the state-action pair of each set of performance data as the value.

[0022] It should be noted that the linear projection of the state-action pair is used as input to the model to predict the network's next state. It's possible that the linear functions used to generate the linear projections for the input "key" and the input "value" are different.

[0023] In one implementation of the first aspect, the method further includes: applying the second attention-based model to at least a first set of performance data from the plurality of sets of performance data to predict a second set of performance data.

[0024] Optionally, embodiments of this disclosure address long-term forecasting by first solving several short-term forecasting tasks, i.e., the final forecasting task. These short-term forecasting tasks can be referred to as sub-forecasting tasks. For example, using information about the network pattern on the first day, the network pattern for the second day can be forecasted or predicted. Each sub-forecasting task can be represented using an attention mechanism. This can be considered as the first attention layer in the AttenInAtten architecture.

[0025] In one implementation of the first aspect, the method further includes: applying the second attention-based model to at least the first set of performance data and the second set of performance data to predict a third set of performance data.

[0026] Following the previous example, once the network pattern for the second day is known (after the sub-forecasting task), the information from the previous two days can be used to forecast or predict the network pattern for the third day.

[0027] In one implementation of the first aspect, the step of applying the second attention-based model to at least the first set of performance data and the second set of performance data to predict the third set of performance data includes: obtaining the action of the third set of performance data; using a linear projection of the predicted state of the second set of performance data as the query, and predicting the state of the third set of performance data by applying the second attention-based model to at least the first set of performance data and the second set of performance data.

[0028] If it is known, how to... predict Then, according to Further Forecast And so on. Specifically, the query to be input into the second attention-based model to predict the state of the third set of performance data is a linear projection of the state to be predicted in the previous sub-prediction task, i.e., the predicted state of the second set of performance data.

[0029] In one implementation of the first aspect, the step of predicting the state of the third set of performance data by applying the second attention-based model to at least the first set of performance data and the second set of performance data further includes: using a first linear projection of the state-action pair of the first set of performance data and the second set of performance data as the key, and using a second linear projection of the state-action pair of the first set of performance data and the second set of performance data as the value.

[0030] Similar to the first attention-based model, the linear projection of the state-action pair is also used as input to the second attention-based model to predict the state of the sub-prediction task. It is possible that the linear functions used to generate the linear projections for the input "keys" and input "values" are different.

[0031] In one implementation of the first aspect, the method further includes: applying the second attention-based model to the second set of performance data and the third set of performance data to identify the correlation between the second set of performance data and the third set of performance data.

[0032] It is possible that sub-forecasting tasks may be similar, or share some knowledge about how to forecast future states. This similarity can be determined and used to forecast the final forecasting task.

[0033] In one implementation of the first aspect, the method further includes: predicting the at least one future performance data by applying the first attention-based model to the plurality of sets of performance data based on the correlation.

[0034] This disclosure also proposes considering the similarity of sub-forecasting tasks when predicting the final forecasting task. Specifically, the output of the sub-forecasting task (i.e., the first attention layer) is input into a first attention-based model, which can be considered as the second attention layer in the AttenInAtten architecture. Understandably, the second attention layer receives the output of the first attention layer and adds it to the corresponding weights.

[0035] In one implementation of the first aspect, the at least one performance metric includes traffic and / or throughput.

[0036] For example, the KPIs discussed in this disclosure are network traffic or throughput. However, this disclosure can also be applied to other network KPIs.

[0037] A second aspect of this disclosure provides a wireless digital twin of a wireless network system, including a model for implementing the method according to the first aspect or any implementation thereof.

[0038] Therefore, embodiments of this disclosure also propose a wireless digital twin, which is constructed for predicting the performance of wireless communication networks.

[0039] The implementation method of the wireless digital twin in the second aspect can correspond to the implementation method of the computer implementation method in the first aspect. The wireless digital twin and its implementation method in the second aspect achieve the same advantages and effects as the computer implementation method and its implementation method in the first aspect.

[0040] A third aspect of this disclosure provides a computer program product including program code that, when implemented on a processor, performs the method according to the first aspect or any implementation thereof.

[0041] The implementation method of the computer program product in the third aspect can correspond to the implementation method of the computer implementation method in the first aspect. The computer program product and its implementation method in the third aspect achieve the same advantages and effects as the computer implementation method and its implementation method in the first aspect.

[0042] It should be noted that all devices, elements, units, and apparatuses described in this application can be implemented in software or hardware elements or any combination thereof. All steps performed by the various entities described in this application, and functions described as being performed by the various entities, are intended to indicate that the respective entities are adapted or configured to perform the respective steps and functions. Although in the following description of specific embodiments, the specific functions or steps to be performed by external entities are not reflected in the detailed description of the specific elements of the entities performing the specific steps or functions, those skilled in the art will understand that these methods and functions can be implemented in the corresponding software or hardware elements or any combination thereof. Attached Figure Description

[0043] The above aspects and implementations of this disclosure are set forth in the following description of specific embodiments, with reference to the accompanying drawings, in which:

[0044] Figure 1 A schematic diagram of an autonomous driving network hierarchy is shown.

[0045] Figure 2 An exemplary process for predicting throughput in a cellular network is shown;

[0046] Figure 3 A computer implementation method according to an embodiment of the present disclosure is shown;

[0047] Figure 4 The overall model architecture according to embodiments of this disclosure is shown;

[0048] Figure 5 An exemplary Transformer encoder is shown;

[0049] Figure 6 A schematic diagram of the AttenInAtten block according to an embodiment of the present disclosure is shown;

[0050] Figure 7 A schematic diagram of a wireless network and a wireless digital twin according to embodiments of the present disclosure is shown;

[0051] Figure 8 The forecast results according to embodiments of this disclosure are shown;

[0052] Figure 9 The forecast results according to an embodiment of this disclosure are shown. Detailed Implementation

[0053] Illustrative embodiments of a computer implementation method for performance prediction in wireless communication networks, wireless digital twins, and corresponding computer program products are described with reference to the accompanying drawings. While this description provides detailed examples of possible implementations, it should be noted that these details are merely exemplary and in no way limit the scope of this application.

[0054] Furthermore, one embodiment / example may relate to other embodiments / examples. For example, any descriptions mentioned in one embodiment / example, including but not limited to terms, elements, processes, explanations, and / or technical advantages, are applicable to other embodiments / examples.

[0055] In real-world scenarios, wireless network digital twins can predict the performance of communication networks for autonomous vehicles through simulation and modeling. By optimizing network design, configuration, and maintenance, network operators can ensure reliable and secure communication between vehicles and infrastructure.

[0056] The ability to predict KPIs (e.g., traffic and throughput) is crucial for realizing wireless digital twins. A comprehensive and realistic physical network model is needed to accurately predict KPIs. This model should include base stations, antennas, user equipment, and the propagation environment, as well as their interactions and dependencies.

[0057] Typically, such models can be built using a variety of data sources and modeling techniques, including network measurement data, traffic models, and machine learning. Specifically, network operators can collect data from various sources such as network probes, drive tests, and user feedback to characterize network performance and behavior under different conditions. Based on user behavior, application requirements, and network topology, traffic models can predict the size and type of traffic in the network. Machine learning techniques can be used to analyze and model the complex relationships and dependencies between network components and variables, and to make accurate predictions based on historical data.

[0058] Wireless digital twins can use these data sources and modeling techniques to simulate physical networks. This enables accurate predictions of key performance indicators such as traffic and throughput. Therefore, network operators can optimize network design, configuration, and maintenance. Furthermore, it ensures reliable and efficient communication across a variety of applications.

[0059] Figure 3 A computer-implemented method 300 for performance prediction in a wireless communication network according to an embodiment of the present disclosure is illustrated. Method 300 includes step 301: acquiring historical performance data of the network. Specifically, the historical performance data includes multiple sets of performance data associated with at least one network performance metric. Each set of performance data is collected at a first sampling granularity over a time period. Each piece of performance data included in the set is collected at a second sampling granularity over the same time period. Method 300 further includes step 302: predicting at least one future performance data based on the multiple sets of performance data.

[0060] This disclosure presents an advanced scheme for constructing wireless digital twins for wireless communication networks. Embodiments of this disclosure enable long-term forecasting, particularly of long-term KPIs such as traffic and throughput of wireless networks. The main idea of ​​this disclosure is to rely on using data captured at different time scales to predict future data.

[0061] For example, each set of performance data collected at a first sampling granularity can be considered to represent the network's daily pattern. In this example, the time period is considered as one day. The network's daily pattern behavior can be learned and modeled. Additionally, hourly patterns of KPIs are also collected. Each performance data point included in this set of performance data, collected at a second sampling granularity, represents the network's daily pattern.

[0062] Optionally, according to embodiments of this disclosure, each set of performance data includes state-action pairs indicating each performance data point, each state being associated with at least one network performance metric, and each action being associated with at least one network parameter.

[0063] For example, data in a state-action pair format can be represented as ,in, For the i-th state, Let i be the i-th action. For example, multiple sets of performance data can be represented as... If each set of performance data represents the daily pattern of the network, then the state-action pairs... This represents the data for the first day.

[0064] In one example, each performance data point in a set of performance data (i.e., daily data) can represent the hourly pattern of the network. Therefore, each performance data point can be represented as... Where i represents an hour and j represents a day. This example involves 7 days of historical data. It should be understood that this is just an example, and this application is not limited to a specific number of days.

[0065] It's understandable that states can be viewed as vectors. The size of daily states and hourly states differs. For example, for daily states, the state vector shape is 24×1, where 24 represents the 24 hours of a day. For hourly states, if we consider a sampling interval of 5 minutes (12×5 minutes per hour), the state vector shape could be, for example, 12×1.

[0066] Figure 4 The overall model architecture according to an embodiment of this disclosure is shown. To address the long-term forecasting problem, a model with an attention mechanism is designed, and two different branches are used to capture patterns at different time scales (e.g., daily and hourly).

[0067] Optionally, according to embodiments of this disclosure, Figure 3 The prediction step 302 shown may include applying at least one machine learning model to multiple sets of performance data (i.e., daily and hourly patterns of the network).

[0068] In this particular example, a traditional Transformer encoder (branch 1, such as...) is used. Figure 4 The upper part (as shown) represents the hourly pattern of KPIs. A new type of AttenInAtten block (branch 2, as shown) is used. Figure 4 (As shown in the lower half) to learn daily pattern behaviors.

[0069] Figure 4 The Fully Connected Network (FCN) shown represents a fully connected layer that implements an affine transformation from input to output, i.e.:

[0070]

[0071] Figure 5An example of a conventional Transformer encoder is shown. Details are not discussed in this application.

[0072] As previously mentioned, the attention mechanism is a popular block in artificial neural networks, comprising three types of components: query (Q), key (K), and value (V). The attention mechanism calculates the similarity between the query and the key, and uses this similarity as a weight to generate a weighted sum of values. The attention mechanism is a fundamental block of the Transformer model. This disclosure also proposes a novel module, “AttenInAtten,” which can have two attention layers.

[0073] Optionally, according to embodiments of this disclosure, at least one machine learning model includes a first attention-based model and / or a second attention-based model, each model including the following input components: query, key, and value.

[0074] To enhance the impact of actions, this disclosure proposes leveraging an attention mechanism by using the latest action in the state-action pair as an action-sensitive query. Generally, in attention mechanisms, the similarity between the query and the key is calculated, and then a weighted sum of the values ​​is calculated based on the similarity. Since the action is used as the query, the output of the attention mechanism is highly influenced by the action (query). This high sensitivity to parameters is extremely useful for network optimization.

[0075] Figure 6 A detailed schematic diagram of the AttenInAtten block according to an embodiment of this disclosure is shown. It is observed that in long-term forecasting problems, there are many short-term forecasting tasks (also referred to as sub-forecasting tasks). The idea behind this block is that solving the sub-forecasting tasks can contribute to solving the final forecasting task. Sub-forecasting tasks can be represented by functions... Let L be a positive integer. For example, consider how to determine the value of L based on... predict (Right now ), and it is also known how to base it on forecast (Right now And so on, if we need to predict the data for the eighth day (i.e. Then the model can use information from previous sub-prediction tasks to... predict .

[0076] It should be noted that this disclosure proposes using an attention mechanism located in the first attention layer to represent the sub-prediction task. Specifically, the value is the projection of the state-action pair used to predict the next state, the query is the linear projection of the state to be predicted in the sub-prediction, and the key is another linear projection of the state-action pair used to predict the next state. The output of the first attention layer (e.g., Figure 6 In to ) represents the representation of each sub-prediction task. These outputs will become the input to the second attention layer.

[0077] Therefore, according to embodiments of this disclosure, Figure 3 The method 300 shown also includes applying a second attention-based model to at least a first set of performance data from multiple sets of performance data (e.g., ) to predict the second set of performance data (e.g., ).

[0078] Optionally, method 300 may further include applying a second attention-based model to at least the first set of performance data and the second set of performance data to predict a third set of performance data (e.g., ).

[0079] Optionally, according to embodiments of this disclosure, the step of applying a second attention-based model to at least a first set of performance data and a second set of performance data to predict a third set of performance data includes: obtaining the action of the third set of performance data; and using a linear projection of the predicted state of the second set of performance data as a query to predict the state of the third set of performance data by applying the second attention-based model to at least the first set of performance data and the second set of performance data.

[0080] Optionally, according to embodiments of this disclosure, the step of predicting the state of a third set of performance data by applying a second attention-based model to at least a first set of performance data and a second set of performance data further includes: using a first linear projection of the state-action pair of the first set of performance data and the second set of performance data as a key, and using a second linear projection of the state-action pair of the first set of performance data and the second set of performance data as a value.

[0081] In the second attention layer, another attention mechanism is employed. We calculate the last action ( Figure 6 In The similarity between the query (as a query) and all outputs of the first attention layer (as keys and values) is then used as weights to obtain a linear combination of all sub-prediction tasks. The outputs of the second attention layer (...) Figure 6 In The features extracted from all sub-forecasting tasks will be further used to predict the state of the following day. In other words, the final forecasting task can be represented as... It should be noted that the last action ( Figure 6 In ) is used as a query for the predictive model, so action sensitivity will be significant.

[0082] Optionally, according to embodiments of this disclosure, at least one future performance data includes a state-action pair, and method 300 includes: obtaining the action of at least one future performance data; and using a linear projection of the action of at least one future performance data as a query to predict the state of at least one future performance data by applying a first attention-based model to multiple sets of performance data.

[0083] In other words, state-action pairs from historical data and the network's latest actions (e.g., It can be used to predict the future state of a network, for example, The value of . Since actions are used as input "queries," the output of the attention mechanism is greatly influenced by the actions (queries). It should be noted that the network's actions are controlled by, for example, network maintenance personnel. An action that has been performed or will be performed is considered known to the network. The model designed according to this disclosure also exhibits high sensitivity to parameters, which is very useful for network optimization.

[0084] Optionally, according to embodiments of this disclosure, the step of predicting the state of at least one future performance data by applying a first attention-based model to multiple sets of performance data further includes: using a first linear projection of the state-action pair of each set of performance data as a key and using a second linear projection of the state-action pair of each set of performance data as a value.

[0085] Optionally, according to embodiments of this disclosure, Figure 3 The method 300 shown further includes the step of applying a second attention-based model to a second set of performance data and a third set of performance data to identify the correlation (e.g., weights) between the second set of performance data and the third set of performance data.

[0086] Optionally, according to embodiments of this disclosure, Figure 3 The method 300 shown also includes predicting at least one future performance data point by applying a first attention-based model to multiple sets of performance data based on correlation.

[0087] It is worth mentioning further that the KPIs discussed in this disclosure include network traffic and throughput. However, this disclosure can also be applied to other network KPIs.

[0088] Figure 7The following illustrations depict application scenarios of embodiments of this disclosure. The upper portion shows a wireless network system, including a NodeB and a Mobile Network Automatic Engine (MAE) (wireless network management node). The lower portion shows a wireless network digital twin 700 according to embodiments of this disclosure. The wireless network digital twin 700 may include a basic model library, an optimization library, and application features such as traffic prediction models, throughput prediction models, and wireless parameter optimization. The proposed attention-based modeling algorithm is provided in the digital twin as a key capability of the basic model library in the wireless network digital twin 700. The wireless network provides data to the wireless network digital twin 700, while the wireless network digital twin 700 provides modeling and optimization capabilities to the wireless network.

[0089] Attention-based models can be considered part of the Wireless Digital Twin 700 (the product will be a software package). The Wireless Digital Twin essentially comprises a base model / optimization component and an application component, such as... Figure 7 As shown. Attention-based methods will be one of the core capabilities of the basic model module, not only for some applications related to modeling, but also for optimizing libraries / applications.

[0090] The following section discusses a specific implementation of how to use the proposed attention-based model to predict the throughput of wireless communication networks. In this example, the observed data is a time series on an hourly scale, meaning the model obtains throughput data hourly. In addition to throughput, hourly action data can also be acquired. Throughput depends on previous throughput and actions that have been or will be performed. Previous throughput and this action are used when predicting throughput. The prediction model can be expressed as follows: , These are the model parameters that need to be estimated.

[0091] If a week's worth of data is used as input to the model, the model's output is a prediction of the throughput for the following day (i.e., the eighth day). In one example, to train the model, 12 days of data can be used, and the model can be tested using data from the next 3 days. The model we use is similar to... Figure 4 The same applies, including two different branches: (1) the traditional Transformer encoder; and (2) the AttenInAtten model. Here, the throughput / flow dimension is 24, and the action dimension is 38. The input sequence length is 7, the embedding_dim of the attention encoder is 64, and the number of heads in the multi-head attention encoder is 6.

[0092] Table 1 shows the results of our attention-based model, as follows:

[0093] Table 1

[0094]

[0095] It should be noted that APE represents absolute percentage error, and APE@0.25 indicates the proportion of predictions with an APE less than 0.25. To further analyze the proposed method, Figure 8 Some visualizations are shown, illustrating the forecast results for throughput and the autocorrelation function of the forecast error. From Figure 8 As can be seen, most of the predicted curves match the ground truth very well. The prediction residual diagnostics indicate that the residuals are uncorrelated, meaning that there is no information in the residuals applicable to the forecast. This is further proof of the good modeling ability of the proposed model.

[0096] In another specific embodiment, the use of the proposed attention-based model to predict traffic in a wireless communication network is discussed. Similarly, in this example, the observed data is in time-series format on an hourly scale, meaning the model obtains a measured traffic data point every hour. In addition to traffic, hourly action data can also be acquired. Similar to throughput, traffic also depends on previous traffic and actions that have been or will be performed. Previous traffic and this action are used when predicting traffic. The prediction model can be expressed as follows: , These are the model parameters that need to be estimated.

[0097] If you use one week's worth of data as input to the model, the model's output will be the traffic forecast for the following day. In one example, to train the model, you could use 12 days' worth of data and then test the model using data from the next 3 days. The model and... Figure 4 Same as shown.

[0098] Table 2 shows the results of predicting flow using the proposed attention-based model, as follows:

[0099] Table 2

[0100]

[0101] To further analyze the proposed method, Figure 9 Some visualizations are shown, illustrating the autocorrelation function of the flow forecast results and the prediction error. From Figure 9As can be seen, most predicted curves match the ground truth very well. Some flow curves within the prediction time domain exhibit distinctly different behavior compared to historical time periods, and the forecast results closely follow the ground truth, indicating that the proposed attention-based model has good parameter sensitivity. Parameter sensitivity is a key feature when applying the model to downstream optimization tasks. Furthermore, the prediction residual diagnostics show that the residuals are uncorrelated, meaning that there is no information in the residuals that is applied to the forecast. This is further evidence of the good modeling ability of our model.

[0102] In summary, this disclosure proposes using attention mechanisms and multi-timescale modeling methods to construct wireless digital twins. This allows for the simultaneous modeling of both long-term trends and short-term fluctuations. Furthermore, in the wireless digital twin scenario, this disclosure proposes constructing a model that simultaneously considers parameter sensitivity. By considering parameter sensitivity during model construction, the model becomes more suitable for downstream tasks, such as optimizing network parameters.

[0103] This disclosure has been described in conjunction with various embodiments as examples and implementations. However, based on a study of the drawings, this disclosure, and the independent claims, those skilled in the art will be able to understand and implement other variations when practicing the claimed embodiments of this disclosure. In the claims and the description, the word "comprising" does not exclude other elements or steps, and the indefinite articles "a" or "an" do not exclude multiple. A single element or other unit can perform the function of several entities or items recited in the claims. The fact that certain measures are recited in different dependent claims does not in itself mean that a combination of these measures cannot be used in advantageous implementations.

[0104] Furthermore, any method according to embodiments of this disclosure can be implemented in a computer program having code means, which, when run by a processing means, causes the processing means to perform the steps of the method. The computer program is included in a computer-readable medium of a computer program product. A computer-readable medium can substantially include any memory, such as read-only memory (ROM), programmable read-only memory (PROM), erasable PROM (EPROM), flash memory, electrically erasable PROM (EEPROM), or a hard disk drive.

[0105] Furthermore, those skilled in the art will recognize that embodiments of the proposed wireless digital twin 700 and corresponding computer program products include the necessary communication capabilities in the form of functions, devices, units, elements, etc., for executing the scheme. Other examples of such devices, units, elements, and functions include: processors, memories, buffers, control logic, encoders, decoders, rate matchers, de-rate matchers, mapping units, multipliers, decision units, selection units, switches, interleavers, deinterleavers, modulators, demodulators, inputs, outputs, antennas, amplifiers, receiving units, transmitting units, DSPs, trellis-coded modulation (TCM) encoders, TCM decoders, power supply units, power feeders, communication interfaces, communication protocols, etc., these components being appropriately arranged together to execute the scheme.

[0106] Specifically, the processors of the Wireless Digital Twin 700 and its corresponding computer program product may include, for example, a Central Processing Unit (CPU), a processing unit, processing circuitry, a processor, an Application Specific Integrated Circuit (ASIC), a microprocessor, or one or more instances of other processing logic capable of interpreting and executing instructions. Therefore, the term "processor" can refer to a processing circuitry system that includes multiple processing circuits (e.g., any, some, or all of the aforementioned processing circuits). The processing circuitry system can also perform data processing functions for inputting, outputting, and processing data, including data buffering and device control functions such as call processing control or user interface control.

Claims

1. A computer-implemented method (300) for performance prediction in a wireless communication network, the method (300) comprising: (301) Obtain historical performance data of the network, wherein the historical performance data includes multiple sets of performance data associated with at least one network performance indicator, wherein each set of performance data is collected at a first sampling granularity within a time period, and each piece of performance data included in the set of performance data is collected at a second sampling granularity within the time period; and Based on the multiple sets of performance data, predict (302) at least one future performance data.

2. The method (300) according to claim 1, wherein, The prediction (302) includes applying at least one machine learning model to the multiple sets of performance data.

3. The method (300) according to claim 2, wherein, The at least one machine learning model includes a first attention-based model and / or a second attention-based model, each model including the following input components: query, key, and value.

4. The method (300) according to any one of claims 1 to 3, wherein, Each set of performance data includes state-action pairs indicating each performance data point, with each state associated with at least one network performance metric and each action associated with at least one network parameter.

5. The method (300) according to claims 3 and 4, wherein, The at least one piece of future performance data includes a state-action pair, and the method (300) includes: The action of obtaining at least one piece of future performance data; and Using the linear projection of the action of the at least one future performance data point as the query, the state of the at least one future performance data point is predicted by applying the first attention-based model to the multiple sets of performance data.

6. The method (300) according to claim 5, wherein, Predicting the state of the at least one future performance data point by applying the first attention-based model to the multiple sets of performance data further includes: The first linear projection of the state-action pair for each set of performance data is used as the key, and the second linear projection of the state-action pair for each set of performance data is used as the value.

7. The method (300) according to claim 3 or any one of claims 4 to 6 dependent on claim 3, wherein, The method (300) includes: The second attention-based model is applied to at least the first set of performance data from the plurality of sets of performance data to predict the second set of performance data.

8. The method (300) according to claim 7, wherein, The method (300) includes: The second attention-based model is applied to at least the first set of performance data and the second set of performance data to predict the third set of performance data.

9. The method (300) according to claim 8, wherein, Applying the second attention-based model to at least the first set of performance data and the second set of performance data to predict the third set of performance data includes: The action of acquiring the third set of performance data; and The state of the third set of performance data is predicted by using a linear projection of the predicted state of the second set of performance data as the query, and by applying the second attention-based model to at least the first set of performance data and the second set of performance data.

10. The method (300) according to claim 9, wherein, Predicting the state of the third set of performance data by applying the second attention-based model to at least the first set of performance data and the second set of performance data further includes: The first linear projection of the state-action pair of the first set of performance data and the second set of performance data is used as the key, and the second linear projection of the state-action pair of the first set of performance data and the second set of performance data is used as the value.

11. The method (300) according to any one of claims 8 to 10, wherein, The method (300) includes: The second attention-based model is applied to the second set of performance data and the third set of performance data to identify the correlation between the second set of performance data and the third set of performance data.

12. The method (300) according to claim 11, wherein, The method (300) includes: The first attention-based model is applied to the multiple sets of performance data based on the correlation to predict the at least one future performance data point.

13. The method (300) according to any one of claims 1 to 12, wherein, The at least one performance metric includes flow rate and / or throughput.

14. A wireless digital twin (700) of a wireless network system, comprising a model for implementing the method (300) according to any one of claims 1 to 13.

15. A computer program product comprising program code for performing the method (300) according to any one of claims 1 to 13 when implemented on a processor.

Images