A wireless edge network-based computing offloading joint optimization method and system

Abstract

This application discloses a joint optimization method and system for compute offloading based on wireless edge networks, belonging to the field of wireless edge network energy consumption optimization. The method includes: determining load state data and load prediction data for local wireless edge nodes based on state sequence data of local wireless edge nodes across multiple state dimensions; determining control decision data for local wireless edge nodes based on the load state data and load prediction data; and adjusting the task processing power and / or task throughput of local wireless edge nodes based on the control decision data. Under the method of this application embodiment, control decisions can be generated based on more comprehensive state information and more reliable prediction results, thereby improving the situation where energy consumption benefits and service quality are difficult to balance due to resource adjustment lag or unreliable predictions, further enhancing the compute offloading and resource management effects of wireless edge nodes.

Description

Technical Field

[0001] This application belongs to the field of energy consumption optimization for wireless edge networks, specifically relating to a joint optimization method, system, device, storage medium, and computer program product for computation offloading based on wireless edge networks. Background Technology

[0002] With the rapid development of mobile internet, the Internet of Things, and smart terminals, massive computing tasks are constantly being generated in wireless access networks. To reduce the computing pressure on terminal devices, reduce energy consumption, and improve service response speed, more and more computing tasks are being offloaded to wireless edge nodes for processing. Wireless edge nodes are typically deployed on the base station side or access point side, possessing certain computing and storage capabilities, and can complete task processing closer to the user side, thereby significantly reducing latency and improving service quality.

[0003] In existing technologies, computational offloading and resource management for wireless edge nodes typically employ scheduling strategies based on static thresholds, short-term load states, or single predictive models. For example, some solutions monitor edge nodes' current CPU utilization, task queue length, or wireless link quality, triggering task offloading or resource adjustments based on preset thresholds; other solutions rely on state inputs at specific moments, attempting to generate control strategies based on reinforcement learning or heuristic algorithms.

[0004] However, scheduling methods based on static thresholds or single-moment states cannot effectively reflect the dynamic changes of wireless edge nodes across multiple state dimensions, leading to delayed or excessive resource adjustments. At the same time, existing algorithm-based control methods often fail to guarantee the reliability of control decisions, making it difficult for the final control decisions to balance both energy consumption benefits and service quality requirements. Summary of the Invention

[0005] This application aims to provide a method, system, device, storage medium, and computer program product for joint optimization of compute offloading based on wireless edge networks, which at least solves the problem of poor energy consumption benefits and service quality effects of compute offloading and resource management at wireless edge nodes.

[0006] In a first aspect, embodiments of this application disclose a joint optimization method for computation offloading based on a wireless edge network, applied to a local wireless edge node, comprising:

[0007] Based on the state sequence data of the local wireless edge node under multiple state dimensions, load state data and load prediction data for the local wireless edge node are determined; the load prediction data is used to characterize the prediction result of the load state change process of the local wireless edge node under a target confidence interval; the target confidence interval matches a preset confidence condition.

[0008] Based on the load status data and the load prediction data, control decision data for the local wireless edge node is determined;

[0009] The task processing power and / or task throughput of the local wireless edge node are adjusted based on the control decision data.

[0010] Secondly, embodiments of this application also disclose a computation offloading joint optimization system based on a wireless edge network, applied to a local wireless edge node, comprising:

[0011] The prediction module is used to determine the load state data and load prediction data of the local wireless edge node based on the state sequence data of the local wireless edge node under multiple state dimensions; the load prediction data is used to characterize the prediction result of the load state change process of the local wireless edge node under a target confidence interval; the target confidence interval matches a preset confidence condition.

[0012] The decision module is used to determine control decision data for the local wireless edge node based on the load status data and the load prediction data.

[0013] An execution module is used to adjust the task processing power and / or task throughput of the local wireless edge node based on the control decision data.

[0014] Thirdly, embodiments of this application also disclose an electronic device, including a processor and a memory, wherein the memory stores a program or instructions that can run on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0015] Fourthly, embodiments of this application also disclose a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0016] Fifthly, embodiments of this application also disclose a computer program product, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps described in the first aspect.

[0017] In summary, in this embodiment, the state sequence data of the local wireless edge node under multiple state dimensions is first processed to determine the load state data and load prediction data. This allows control decisions to no longer rely on the state input at a single moment, but to make judgments based on multi-dimensional and continuous state change trends. The prediction results can reflect the uncertainty of future load to a certain extent, thereby reducing resource adjustment delays or over-adjustments caused by incomplete state information and providing a more reliable reference for subsequent decisions. Furthermore, control decision data can be determined based on the load state data and load prediction data, enabling control decisions to simultaneously consider the current load level. By considering future load trends to enhance decision-making robustness, the control strategy becomes more adaptable to load fluctuations, achieving a more reasonable balance between energy consumption benefits and service quality, and reducing decision imbalances caused by prediction biases or state fluctuations. Ultimately, based on control decision data, the task processing power and throughput of local wireless edge nodes are adjusted, allowing resource scheduling to be adjusted in advance according to predicted trends, reducing the risk of insufficient or wasted resources. This enables wireless edge nodes to maintain relatively stable service quality under dynamic load environments while improving the energy efficiency of the wireless edge network, thus achieving a more optimized trade-off between energy consumption and performance. Therefore, based on the method of this application embodiment, control decisions can be generated on the basis of more comprehensive state information and more reliable prediction results, thereby improving the situation where energy consumption benefits and service quality are difficult to balance due to delayed resource adjustments or unreliable predictions, further enhancing the computational offloading and resource management effects of wireless edge nodes. Attached Figure Description

[0018] In the attached diagram:

[0019] Figure 1 This is a flowchart illustrating the steps of a joint optimization method for computation offloading based on a wireless edge network, as provided in an embodiment of this application.

[0020] Figure 2 This is a flowchart of another computation offloading joint optimization method based on a wireless edge network provided in an embodiment of this application;

[0021] Figure 3 This is a block diagram of a computation offloading joint optimization system based on a wireless edge network, provided in an embodiment of this application.

[0022] Figure 4 This is a block diagram of an electronic device provided in one embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, the "and / or" signifies at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects have an "or" relationship.

[0025] like Figure 1 The image shows a computation offloading joint optimization method based on a wireless edge network provided in this application embodiment, which is applied to a local wireless edge node.

[0026] The method may include the following steps:

[0027] Step 101: Based on the state sequence data of the local wireless edge node under multiple state dimensions, determine the load state data and load prediction data of the local wireless edge node.

[0028] The load prediction data is used to characterize the prediction results of the load state change process of local wireless edge nodes under the target confidence interval; the target confidence interval is matched with the preset confidence conditions.

[0029] In some embodiments of this application, in order to generate data results that reflect the current operating status and future load change trends of the local wireless edge node in subsequent processing, it is necessary to determine load status data and load prediction data based on state sequence data under multiple state dimensions. Specifically, the load status data and load prediction data of the local wireless edge node can be determined based on the state sequence data of the local wireless edge node under multiple state dimensions. The load prediction data is used to characterize the prediction result of the load status change process of the local wireless edge node under a target confidence interval, wherein the target confidence interval matches a preset confidence condition and is used to limit the confidence range of the prediction result in the future change process. The confidence range is usually used to describe the fluctuation range of the prediction result that may occur in the future. Its formation can be based on the analysis of the prediction error distribution, the statistical characteristics of the prediction sequence, or the dispersion of multiple prediction results, so that the prediction result not only includes a single point value, but also includes the upper and lower bounds under the preset confidence condition. This will enable subsequent control decisions to simultaneously refer to the current load level and the future load trend, and to a certain extent reduce the decision bias caused by prediction uncertainty, making the control strategy more robust.

[0030] In a specific example, a local wireless edge node obtains state sequence data for three state dimensions—CPU utilization, task queue length, and wireless link quality—within a target time interval of 10 minutes. Based on this state sequence data, the current load state data can be extracted, and a predictive model can be used to generate load prediction data for several future time points. Ultimately, the system obtains data results that reflect the current load level and future load trends. The predicted future load is given in the form of a target confidence interval that satisfies preset confidence conditions. This ensures that the prediction results not only include the estimated future load value but also the confidence range surrounding that estimated value, providing a more reliable input basis for subsequent control decisions.

[0031] Step 102: Based on load status data and load prediction data, determine the control decision data for the local wireless edge node.

[0032] In some embodiments of this application, to enable local wireless edge nodes to obtain more suitable resource scheduling strategies under dynamic load environments, control decision data for guiding resource adjustments needs to be generated based on load status data and load prediction data. Specifically, control decision data for local wireless edge nodes can be determined based on load status data and load prediction data. Control decision data is typically used to characterize the adjustment methods that local wireless edge nodes should adopt in terms of task processing power or task throughput under the combined effect of current load levels and future load trends, so that the decision-making process can balance energy consumption benefits and service quality. This will provide a clearer basis for subsequent resource adjustment processes, enabling local wireless edge nodes to maintain relatively stable operating performance under load fluctuations.

[0033] In a specific example, the local wireless edge node acquires load status data across three dimensions—CPU utilization, task queue length, and wireless link quality—within a target time interval. Based on this data, it generates load prediction data for several future time points. The current load level and future load trends can then be input into the decision model to determine the control decisions to be made in the next scheduling cycle. Ultimately, the system obtains control decision data that reflects the current load status and future trends, providing a directly usable basis for subsequent adjustments to task processing power or task throughput.

[0034] Step 103: Adjust the task processing power and / or task throughput of the local wireless edge node based on the control decision data.

[0035] In some embodiments of this application, in order for local wireless edge nodes to dynamically adjust resource configuration based on acquired control decision data, the control decision data needs to be transformed into specific adjustment operations on the node's task processing capabilities. Specifically, the task processing power and / or task throughput of the local wireless edge node can be adjusted based on the control decision data. Task processing power is typically used to characterize the scale of computing resources that a node can allocate for task processing per unit time, while task throughput characterizes the number of tasks or data processing volume that a node can complete per unit time. This allows local wireless edge nodes to appropriately increase or decrease resource allocation based on current load status and future load trends, enabling the node to maintain a relatively stable quality of service under load fluctuations and improving energy efficiency to a certain extent.

[0036] In a specific example, the local wireless edge node has already obtained control decision data containing suggestions for adjusting task processing power and task throughput. At this point, the node's processing capacity can be adjusted based on this control decision data. For example, if a future increase in load is predicted, the task processing power can be increased; if a future decrease in load is predicted, the task throughput can be reduced. Ultimately, the system updates the task processing power and task throughput based on the control decision data, enabling the node to respond to load changes with more appropriate resource allocation in subsequent scheduling cycles, providing a more stable operating foundation for the overall computation offloading process.

[0037] In summary, in this embodiment, the state sequence data of the local wireless edge node under multiple state dimensions is first processed to determine the load state data and load prediction data. This allows control decisions to no longer rely on the state input at a single moment, but to make judgments based on multi-dimensional and continuous state change trends. The prediction results can reflect the uncertainty of future load to a certain extent, thereby reducing resource adjustment delays or over-adjustments caused by incomplete state information and providing a more reliable reference for subsequent decisions. Furthermore, control decision data can be determined based on the load state data and load prediction data, enabling control decisions to simultaneously consider the current load level. By considering future load trends to enhance decision-making robustness, the control strategy becomes more adaptable to load fluctuations, achieving a more reasonable balance between energy consumption benefits and service quality, and reducing decision imbalances caused by prediction biases or state fluctuations. Ultimately, based on control decision data, the task processing power and throughput of local wireless edge nodes are adjusted, allowing resource scheduling to be adjusted in advance according to predicted trends, reducing the risk of insufficient or wasted resources. This enables wireless edge nodes to maintain relatively stable service quality under dynamic load environments while improving the energy efficiency of the wireless edge network, thus achieving a more optimized trade-off between energy consumption and performance. Therefore, based on the method of this application embodiment, control decisions can be generated on the basis of more comprehensive state information and more reliable prediction results, thereby improving the situation where energy consumption benefits and service quality are difficult to balance due to delayed resource adjustments or unreliable predictions, further enhancing the computational offloading and resource management effects of wireless edge nodes.

[0038] Figure 2 This is another computation offloading joint optimization method based on wireless edge networks provided in the embodiments of this application, which is applied to local wireless edge nodes.

[0039] The method may include the following steps:

[0040] Step 201: Acquire multiple sets of state data of the local wireless edge node within the target time interval according to multiple state dimensions.

[0041] In some embodiments of this application, in order to construct state sequence data that reflects the changing trends of the local wireless edge node's operating status in subsequent processing, it is necessary to acquire raw state collection data covering multiple state dimensions within the target time interval. Specifically, multiple sets of state collection data of the local wireless edge node within the target time interval can be acquired according to multiple state dimensions. Multiple state dimensions are typically used to characterize different operating characteristics of the wireless edge node in terms of computing resources, communication resources, and task processing, such as CPU utilization, task queue length, wireless link quality, or energy consumption-related indicators. This will provide a more sufficient raw data foundation for subsequent data cleaning and sequence construction, thereby providing more reliable input conditions for determining load status data and load prediction data.

[0042] In a specific example, a local wireless edge node needs to update its computation offload strategy within a target time interval of 10 minutes. At this time, multiple sets of state data can be collected from the wireless edge node at fixed sampling intervals within this time interval, based on three state dimensions: CPU utilization, task queue length, and wireless link quality. Ultimately, the system obtains raw state data covering the three state dimensions with continuous timestamps, providing directly usable data input for subsequent data cleaning and state sequence construction.

[0043] Step 202: Perform data cleaning on each set of state acquisition data to obtain state sequence data of the local wireless edge node under each state dimension.

[0044] In some embodiments of this application, to enable subsequent load status analysis and load prediction based on structured and continuous state sequence data, it is necessary to preprocess the raw state data collected across multiple state dimensions to remove noisy, abnormal, or missing data. Specifically, data cleaning can be performed on each set of state data to obtain state sequence data for the local wireless edge node in each state dimension. Data cleaning is typically used to standardize the format of the raw collected data, detect anomalies, complete missing data, or filter noise, so that the processed data can more accurately reflect the actual operation of the wireless edge node in the corresponding state dimension. This will enable subsequent steps to perform load status extraction and load prediction based on higher quality and more continuous state sequence data, thereby improving the reliability of the prediction results and the stability of the decision input.

[0045] In a specific example, a local wireless edge node collects raw state data for three dimensions—CPU utilization, task queue length, and wireless link quality—within a target time interval of 10 minutes. At this point, each set of collected data can be cleaned, such as removing outliers caused by link jitter during collection, interpolating missing sampling points, and converting inconsistent data into a unified time-series format. Ultimately, the system obtains state-series data with continuous timestamps, low noise, and a uniform format across all three dimensions, providing a directly usable data foundation for subsequent load status extraction and load prediction.

[0046] Step 203: Based on the state sequence data of the local wireless edge node under multiple state dimensions, determine the load state data and load prediction data of the local wireless edge node.

[0047] The load prediction data is used to characterize the prediction results of the load state change process of local wireless edge nodes under the target confidence interval; the target confidence interval is matched with the preset confidence conditions.

[0048] The method shown in this step has been explained in step 101 and will not be repeated here.

[0049] Optionally, step 203 includes the following sub-steps:

[0050] Sub-step 2031 involves fusing the state sequence data of the local wireless edge node across multiple state dimensions according to temporal correlation to obtain the fused sequence data of the local wireless edge node.

[0051] In some embodiments of this application, in order to enable subsequent load status extraction and load prediction to be based on structured and temporally continuous input data, it is necessary to uniformly process the state sequence data under multiple state dimensions according to their temporal correlation. Specifically, the state sequence data of the local wireless edge node under multiple state dimensions can be fused according to temporal correlation to obtain the fused sequence data of the local wireless edge node. Temporal correlation is usually used to describe the synchronous change relationship of different state dimensions on the same time axis, so that state information from different dimensions can be combined under the same time reference to form a joint sequence that can reflect the overall operating status of the node. This will enable the subsequent prediction model to capture cross-dimensional dynamic change features based on more complete input data, providing a more consistent input basis for determining load prediction data and load status data.

[0052] In a specific example, the local wireless edge node has already obtained state sequence data for three state dimensions: CPU utilization, task queue length, and wireless link quality. At this point, based on the correspondence of these sequences at the same timestamp, the data for the three dimensions can be concatenated or combined chronologically to form a fused sequence data containing multiple state dimensions. Ultimately, the system obtains fused sequence data that simultaneously reflects computing resources, task processing status, and communication link conditions, providing a unified and continuous source of input data for subsequent load prediction and load status extraction.

[0053] Optionally, sub-step 2031 includes the following sub-steps:

[0054] Sub-step 20311: Perform time-series data alignment on each state sequence data to obtain aligned sequence data corresponding to each state sequence data.

[0055] In some embodiments of this application, to enable state sequence data from different state dimensions to participate in subsequent fusion processing under a unified time reference, it is necessary to unify the time axis of each state sequence data. Specifically, time-series data alignment can be performed on each state sequence data to obtain aligned sequence data corresponding to each state sequence data. Time-series data alignment is typically used when there are differences in sampling intervals, inconsistent sampling start and end times, or incomplete overlap of timestamps. Through interpolation, resampling, or timestamp matching, different sequences are made to have corresponding data items at the same time point. This allows subsequent splicing operations to be performed on a unified time axis, enabling the fused sequence data to accurately reflect the synchronous changes of local wireless edge nodes across multiple state dimensions.

[0056] In a specific example, the local wireless edge node has already obtained state sequence data across three state dimensions: CPU utilization, task queue length, and wireless link quality. CPU utilization is collected at 1-second intervals, task queue length at 2-second intervals, and wireless link quality using irregular timestamps. The three sequences can then be time-series aligned, for example, by resampling to unify them onto a 1-second time axis and interpolating to fill in missing data points. Ultimately, the system obtains three aligned sequence data points with corresponding data items at the same timestamp, providing directly usable input for subsequent tensor construction.

[0057] Optionally, sub-step 20311 includes the following sub-steps:

[0058] Sub-step 203111: If the sampling density of the state sequence data is greater than or equal to the preset sampling grid width, extract the state sequence sub-data from the state sequence data according to the sampling grid width to use as the resampled data of the state sequence data.

[0059] In some embodiments of this application, to ensure that state sequence data with high sampling density maintains the same temporal resolution as data in other state dimensions during subsequent processing, it is necessary to perform resampling processing on this type of state sequence data based on the sampling grid width. Specifically, when the sampling density of the state sequence data is greater than or equal to a preset sampling grid width, state sequence sub-data can be extracted from the state sequence data according to the sampling grid width to serve as resampled data for the state sequence data. The sampling grid width is typically used to describe the fixed time interval used when performing data extraction or interpolation on the time axis, enabling the resampled data to express state changes at a uniform time step. This ensures that the state sequence data with high sampling density maintains consistency with data in other state dimensions on the time scale, providing a structurally unified input basis for subsequent interpolation, time axis alignment, and fusion processing.

[0060] In a specific example, the local wireless edge node has already obtained state sequence data of CPU utilization, which is collected at 0.2-second intervals, while the preset sampling grid width is 1 second. At this point, sampling points corresponding to each 1-second interval can be extracted from the original CPU utilization sequence according to the 1-second sampling grid width, forming new resampled data. Ultimately, the system obtains a resampled sequence expressing changes in CPU utilization at a 1-second time step, providing directly usable input for subsequent time-scaled processing of data from other state dimensions.

[0061] Sub-step 203112: If the sampling density of the state sequence data is less than the preset sampling grid width, interpolate and supplement the state sequence data according to the sampling grid width to obtain resampled data of the state sequence data.

[0062] In some embodiments of this application, to ensure that state sequence data with low sampling density maintains the same temporal resolution as data from other state dimensions in subsequent processing, it is necessary to perform interpolation supplementation on this type of state sequence data based on the sampling grid width. Specifically, when the sampling density of the state sequence data is less than a preset sampling grid width, interpolation supplementation can be performed on the state sequence data according to the sampling grid width to obtain resampled data of the state sequence data. Interpolation supplementation is typically used to generate new data points between the original sampling points, so that the resampled sequence can express state changes at a uniform time step, thereby compensating for the loss of temporal information caused by excessively large sampling intervals. This will ensure that the state sequence data with low sampling density maintains consistency with data from other state dimensions in terms of time scale, providing a structurally unified input basis for subsequent time axis alignment and fusion processing.

[0063] In a specific example, the local wireless edge node has already obtained state sequence data of the wireless link quality, collected at 3-second intervals, with a preset sampling grid width of 1 second. The original wireless link quality sequence can then be interpolated and supplemented according to the 1-second sampling grid width. For example, two new interpolation points can be generated between adjacent sampling points, ensuring that the sequence has a corresponding value at each 1-second time step. Ultimately, the system obtains a resampled sequence representing the changes in wireless link quality at a 1-second time step, providing directly usable input for subsequent time-scaled processing of data from other state dimensions.

[0064] Sub-step 203113: Based on the temporal correlation analysis of all resampled data, time axis alignment is performed on each resampled data to obtain the aligned sequence data corresponding to each state sequence data.

[0065] In some embodiments of this application, to ensure that all resampled data can participate in subsequent fusion processing under a unified time reference, it is necessary to unify the temporal relationships between the sequences after resampling. Specifically, based on temporal correlation analysis of all resampled data, time-axis alignment can be performed on each resampled data to obtain aligned sequence data corresponding to each state sequence data. Temporal correlation analysis is typically used to identify the correspondence between different resampled sequences on the time axis. The process may include: comparing the timestamps of each resampled sequence to determine their common time point under a unified sampling grid width; and, in the case of slight time offsets or inconsistent sampling start and end times, determining the effective alignment position of each sequence under a unified time reference by analyzing the temporal proximity and change trends between sequences. Time-axis alignment is typically used to synchronize multiple sequences under a unified time reference, ensuring that each time point has corresponding data items from different state dimensions. This will enable the aligned sequence data to accurately represent the synchronization change characteristics of local wireless edge nodes under multiple state dimensions, providing a structurally consistent input basis for subsequent splicing tensor construction and fusion sequence generation.

[0066] In a specific example, the local wireless edge node has already obtained resampled data for three state dimensions: CPU utilization, task queue length, and wireless link quality. At this point, temporal correlation analysis can be performed on the timestamps of the three resampled sequences. For example, this involves identifying their common time points within a uniform sampling grid width and aligning the data from each sequence at these time points. Ultimately, the system obtains three aligned sequence data points with corresponding data items on the same time axis, providing directly usable input for subsequently combining data from multiple state dimensions into a concatenated tensor.

[0067] Sub-step 20312 combines all the aligned sequence data into a splicing tensor to serve as the fused sequence data.

[0068] In some embodiments of this application, to enable aligned sequence data of multiple state dimensions to be input into subsequent prediction models in a unified structural form, these aligned sequence data need to be combined into a tensor structure that can simultaneously express the time and state dimensions according to a preset method. Specifically, all aligned sequence data can be combined into a concatenated tensor as fused sequence data. Concatenated tensors are typically used to represent the time axis and state dimensions in a two-dimensional or multi-dimensional structure, where rows can represent consecutive time steps and columns can represent different state dimensions, so that each row corresponds to the synchronous data of multiple state dimensions at a certain time point. This allows the fused sequence data to present the joint change characteristics of local wireless edge nodes in multiple state dimensions in a structured manner, providing a uniformly formatted and semantically consistent input for subsequent prediction networks.

[0069] In a specific example, the local wireless edge node has already obtained aligned sequence data for three state dimensions: CPU utilization, task queue length, and wireless link quality. These three sequences can then be concatenated chronologically along the column direction, so that each row of the concatenated tensor corresponds to the values ​​of the three state dimensions at the same timestamp, while the columns correspond to CPU utilization, task queue length, and wireless link quality, respectively. The system then obtains a concatenated tensor with time as the rows and state dimensions as the columns. This tensor can be directly input into the subsequent prediction network as fused sequence data to capture cross-dimensional temporal variations.

[0070] Sub-step 2032 involves repeatedly inputting the fused sequence data into the prediction network according to a preset number of predictions to obtain multiple prediction sequence data for the local wireless edge node.

[0071] In each step of inputting the fused sequence data into the prediction network, at least one neuron in the prediction network is randomly masked.

[0072] In some embodiments of this application, to obtain load prediction results that reflect prediction uncertainty, the prediction process needs to be executed multiple times under the same input conditions. This causes the prediction network to exhibit slight structural differences in different prediction rounds, thereby forming multiple discrete prediction sequence data. Specifically, the fused sequence data can be repeatedly input into the prediction network according to a preset number of predictions to obtain multiple prediction sequence data for the local wireless edge node. During each input of the fused sequence data into the prediction network, at least one neuron in the prediction network is randomly masked. Random masking is generally used to describe the processing method in the forward propagation process of the prediction network, where some neurons do not participate in the calculation in a certain prediction, so that the prediction network forms different effective structures in different prediction rounds. The Monte Carlo method is generally used to estimate the statistical characteristics of the target quantity through multiple random samplings. In this step, multiple prediction processes with random masking can be regarded as multiple random samplings of the prediction network, so that multiple prediction sequence data can reflect the output distribution of the prediction model under random perturbations. This will provide the necessary data foundation for subsequent determination of the load prediction data and target confidence interval based on statistical characteristic analysis, so that the prediction results can reflect the uncertainty of future load changes to a certain extent.

[0073] In a specific example, the local wireless edge node has already obtained the fused sequence data. This fused sequence data can then be input into a prediction network containing a random dropout layer, and the forward prediction process is repeated for a preset number of predictions. During each prediction, the dropout layer randomly deactivates some neurons without participating in the computation according to a set deactivation probability. This results in different effective structures in different prediction rounds, leading to variations in the predicted sequence data obtained each time. Ultimately, the system obtains multiple sets of predicted sequence data based on the same input but with random perturbation characteristics. These data can be considered as multiple Monte Carlo sampling results of the prediction network, providing a directly usable data source for subsequent averaging and statistical analysis to determine the load prediction data and target confidence interval.

[0074] In various embodiments of this application, the prediction network model can be implemented using a variety of deep learning architectures that have been widely validated in the prior art and have mature engineering deployment capabilities, including but not limited to fully connected neural networks (FCN), convolutional neural networks (CNN), temporal convolutional networks (TCN), long short-term memory networks (LSTM), gated recurrent unit networks (GRU), and attention-based transformer networks (Transformer). These models provide a stable, interpretable, and scalable structural foundation for processing time-series data, capturing cross-temporal dependencies, extracting multi-dimensional state features, and expressing complex nonlinear mappings, making them suitable for load prediction scenarios in local wireless edge nodes.

[0075] In practical implementation, the prediction network can construct a training dataset based on historical operational data of local wireless edge nodes. This training dataset typically includes continuously collected load state data, task queue length changes, wireless link quality indicators, processor utilization sequences, power state switching records, task arrival rate sequences, and corresponding future load labels. The training dataset can be constructed using a sliding window approach, ensuring that each training sample contains a continuous historical state sequence and its corresponding future prediction target. This allows the prediction network to learn the dynamic changes of nodes under different operating conditions, different service load patterns, and different wireless environments.

[0076] During training, the prediction network can employ a structure with a random deactivation mechanism to achieve uncertainty estimation during the inference phase using Monte Carlo Dropout. The placement of the Dropout layer can be chosen based on the network structure, but in the embodiments of this application, it is preferably placed after the fully connected layer. Fully connected layers typically handle high-level feature combination and decision mapping, and have large parameter sizes and a high risk of overfitting. Therefore, introducing Dropout into this layer can effectively suppress overfitting and enhance the model's generalization ability. Simultaneously, random deactivation in the fully connected layer allows sub-models formed in different prediction rounds to focus on different combinations of input features, enabling the distribution of prediction results to more accurately reflect the model's uncertainty regarding the input data. For convolutional networks, recurrent networks, or attention-based network structures, which may still contain fully connected layers, the aforementioned Dropout strategy is also applicable.

[0077] The training process can employ standard supervised learning methods, such as gradient descent-based optimization algorithms to minimize the loss function. The loss function can include mean squared error loss, weighted time series loss, or other loss forms suitable for time series prediction tasks. After training, the weight parameters of the prediction network can be viewed as a shared set of parameters among multiple potential sub-models, enabling the acquisition of a statistically significant prediction distribution through multiple random deactivation samplings during the inference phase.

[0078] Sub-step 2033: Obtain load prediction data based on the averaging of multiple prediction sequence data, and obtain the target confidence interval based on the statistical characteristic analysis of the load prediction data.

[0079] In some embodiments of this application, in order to obtain comprehensive forecast data reflecting future load change trends from multiple forecast results, and to further quantify the possible fluctuation range of the forecast results in the future, it is necessary to perform statistical processing on multiple forecast sequence data. Specifically, load forecast data can be obtained based on averaging multiple forecast sequence data, and a target confidence interval can be obtained based on statistical feature analysis of the load forecast data. Averaging is typically used to extract the overall trend among multiple forecast results, enabling the load forecast data to reflect, to some extent, the common output characteristics of the forecast model under random disturbances. Statistical feature analysis is typically used to evaluate the dispersion, fluctuation range, or error distribution of the forecast results, so that the target confidence interval can describe the upper and lower bounds of the possible future load under preset confidence conditions. This will enable the load forecast data to contain both a trend estimate of the future load and a quantitative expression of forecast uncertainty, providing a more robust input basis for subsequent control decisions.

[0080] In a specific example, the local wireless edge node has already obtained multiple sets of predicted sequence data generated based on a random deactivation mechanism. These predicted sequence data can then be averaged at each future time point to form load prediction data reflecting the overall trend. Subsequently, statistical characteristic analysis can be performed on this load prediction data, such as calculating the dispersion or fluctuation range of the prediction results at each time point, and determining the target confidence interval based on preset confidence conditions. Ultimately, the system obtains prediction results containing future load estimates and their confidence ranges, providing more reliable reference data for subsequent control decisions.

[0081] Optionally, sub-step 2033 includes the following sub-steps:

[0082] Sub-step 20331: Determine the multi-order estimators of the local wireless edge nodes based on the load prediction data.

[0083] In some embodiments of this application, in order to extract statistics reflecting future load change trends and fluctuation characteristics from load forecast data, it is necessary to calculate multi-order statistical characteristics of the load forecast data. Specifically, multi-order estimators for local wireless edge nodes can be determined based on load forecast data. Multi-order estimators are typically used to describe the characteristics of forecast results under different statistical dimensions. For example, by calculating indicators such as the mean, variance, skewness, or kurtosis of the forecast results, the system can characterize the future load change characteristics from multiple perspectives, such as trend, dispersion, and distribution pattern. This allows the subsequent confidence interval determination process to make inferences based on more comprehensive statistical information, so that the target confidence interval reflects not only the future load trend but also the fluctuation range and uncertainty of the forecast results.

[0084] In a specific example, the local wireless edge node has already obtained multiple sets of predicted sequence data generated based on a random deactivation mechanism, and formed load prediction data through averaging. At this point, multi-order statistical characteristic calculations can be performed on the load prediction data. For example, the mean at each future time point can be calculated to reflect the load trend, the variance can be calculated to reflect the dispersion of the prediction results, or skewness and kurtosis can be further calculated according to application requirements to describe the distribution pattern of the prediction results. Ultimately, the system obtains a multi-order estimator that can characterize the future load change features from multiple statistical dimensions, providing the necessary input foundation for subsequent determination of the target confidence interval based on a statistical distribution model.

[0085] Sub-step 20332: Based on the multi-order estimator and the statistical distribution model corresponding to the state sequence data of the local wireless edge node, determine the target confidence interval.

[0086] In some embodiments of this application, in order to determine the confidence range of future load prediction results under preset confidence conditions, it is necessary to infer the fluctuation characteristics of the load prediction data based on the statistical distribution model corresponding to the multi-order estimator and the state sequence data. Specifically, the target confidence interval can be determined based on the multi-order estimator and the statistical distribution model corresponding to the state sequence data of the local wireless edge node. The statistical distribution model is typically used to describe the probability distribution shape that the prediction result may present at future times, enabling the system to infer the upper and lower bounds of the prediction result based on the distribution characteristics. This will allow the target confidence interval to reflect the reliable range of future load changes under preset confidence conditions, providing a robust reference for subsequent control decisions.

[0087] In a specific example, the local wireless edge node has obtained load forecast data and its multi-order estimates, and determined that the forecast data statistically conforms to a normal distribution model. At this point, the mean of the forecast results can be used as a trend estimate of future load, and the standard deviation of the forecast results can be used as a quantitative indicator of load fluctuation. With a preset confidence level of 95%, the target confidence interval can be determined based on the statistical characteristics of the normal distribution as a range centered on the mean, extended upwards and downwards by two times the standard deviation. Ultimately, the system obtains a target confidence interval that reflects the future load trend and fluctuation range, ensuring that the forecast results include not only point estimates but also a reliable range under the preset confidence level.

[0088] In certain specific cases, the state sequence data of local wireless edge nodes may conform to other statistical distribution models, such as the uniform distribution (UD), exponential distribution (ED), or chi-square distribution (CSD). In such cases, the calculation method of the confidence interval can be adjusted according to the characteristics of the corresponding distribution model based on statistical principles. For example, if the prediction result conforms to the uniform distribution model, the target confidence interval can be determined as the upper and lower bounds of this distribution, so that the calculation method of the confidence interval can be consistent with the characteristics of the distribution model.

[0089] Sub-step 2034: Based on time slicing of the fused sequence data at the target time or time interval, load status data is obtained.

[0090] In some embodiments of this application, in order to extract load status information reflecting the current operating status of local wireless edge nodes from fused sequence data, time slicing processing of the fused sequence data is required at a target time or time interval. Specifically, load status data can be obtained based on time slicing of the fused sequence data at the target time or time interval. Time slicing is typically used to select a data segment corresponding to a certain moment or time interval in a continuous time series, so that the obtained data can reflect the operating status of the node at that time point or time interval. This allows the load status data to present the current load level in a structured manner, providing direct status input for subsequent control decisions.

[0091] In a specific example, the local wireless edge node has already acquired fused sequence data containing multiple state dimensions. At this point, the fused sequence data can be time-sliced ​​according to the target time. For example, if the current load state needs to be obtained, the segment corresponding to the last timestamp of the fused sequence data can be directly extracted. Alternatively, a data segment within a certain time period can be selected as load state data according to application requirements. Ultimately, the system obtains load state data that reflects the node's operating status at the target time or target time interval, providing directly usable state input for subsequent control decisions.

[0092] Step 204: Based on load status data and load prediction data, determine the control decision data for the local wireless edge node.

[0093] The method shown in this step has been explained in step 102 and will not be repeated here.

[0094] Optionally, step 204 includes the following sub-steps:

[0095] Sub-step 2041: Determine the control decision action dataset for the local wireless edge node based on load status data and load prediction data.

[0096] The control decision action dataset contains multiple control decision actions for local wireless edge nodes.

[0097] In some embodiments of this application, to enable local wireless edge nodes to obtain selectable resource adjustment schemes under dynamic load environments, it is necessary to generate multiple executable control decision actions based on the current load status and future load trends. Specifically, a control decision action dataset for the local wireless edge node can be determined based on load status data and load prediction data. This dataset contains multiple control decision actions for the local wireless edge node. Control decision actions typically describe different adjustment methods for the node in terms of task processing power, task throughput, or other adjustable resource parameters, enabling the system to perform a comprehensive benefit evaluation of these actions in subsequent steps. This allows the control decision action dataset to cover a variety of possible strategies, from resource allocation to resource reduction, providing the necessary set of candidate actions for subsequently determining the optimal control decision data.

[0098] In a specific example, the local wireless edge node has already obtained current load status data and load forecast data for several future time points. Based on this data, multiple control decision actions can be generated. For example, if a future load increase is predicted, different actions can be generated, such as increasing task processing power, improving task throughput, or maintaining the current resource configuration; if a future load decrease is predicted, actions can be generated to reduce task processing power or decrease task throughput. Ultimately, the system obtains a dataset of control decision actions containing various resource adjustment methods, providing a directly usable set of candidate actions for subsequent comprehensive benefit assessment based on energy consumption benefit estimation and service quality degradation estimation.

[0099] Optionally, sub-step 2041 includes the following sub-steps:

[0100] Sub-step 20411: Based on the data embedding of load status data and load prediction data under different decision dimensions, determine the fusion input data for the local wireless edge node.

[0101] In some embodiments of this application, in order to enable local wireless edge nodes to accurately represent their current state and future load trends in subsequent decision-making models, it is necessary to perform feature processing on load status data and load prediction data under multiple decision dimensions, and then fuse the processed features according to a preset method. Specifically, the fusion input data for local wireless edge nodes can be determined based on the data embedding of load status data and load prediction data under different decision dimensions. Data embedding is typically used to convert raw numerical, categorical, and time-series data into structured feature vectors, enabling different types of data to express the node's operating state, future load trends, and uncertainties in a unified feature space. This will allow the fusion input data to comprehensively reflect the node's resource pressure at the current moment, future load change patterns, and potential risks, providing a high-quality input foundation for the generation of subsequent control decision actions.

[0102] In a specific example, the local wireless edge node has already obtained load status data and load prediction data. At this point, fused input data can be constructed based on conventional feature engineering methods. For instance, considering that continuous features such as real-time queue length and channel quality indicators are difficult to compare directly under different dimensions, they can be standardized. Considering that discrete features such as node sleep status cannot be directly used as model input, they can be one-hot encoded. Considering that the load prediction curve contains information about future trends and fluctuations, time-series statistical features such as short-term mean, slope of change trend, and fluctuation amplitude can be extracted from it. Furthermore, considering that the subsequent decision model requires a unified input format, the above features can be concatenated and fused in a preset order to form a unified high-dimensional feature vector. Finally, the system obtains fused input data that can simultaneously express the current state of the node and the future load trend, providing a directly usable input basis for subsequent generation of control decision actions.

[0103] Sub-step 20412 involves inputting the fused input data into the decision model to form a control decision action dataset based on the obtained multiple control decision actions.

[0104] In some embodiments of this application, in order for a local wireless edge node to generate multiple selectable control decision actions based on fused input data, the fused input data needs to be input into a preset decision model for inference. Specifically, the fused input data can be input into the decision model to form a control decision action dataset based on the obtained multiple control decision actions. The decision model is typically used to infer the potential behavioral outcomes of the node under different resource configuration methods based on the fused input data, enabling the system to generate executable candidate actions across multiple decision dimensions such as task processing power, task throughput, or sleep depth. This allows the control decision action dataset to cover various strategies from resource allocation to resource deallocation, providing a structured set of candidate actions for subsequent comprehensive benefit evaluation.

[0105] In a specific example, the local wireless edge node has already obtained fused input data containing continuous, discrete, and time-series statistical features. This fused input data can then be input into a pre-defined decision model. Considering that the fused input data already contains information such as the node's current state, future load trends, and uncertainties, the decision model can generate multiple control decision actions based on these features, such as increasing task processing power, decreasing task processing power, adjusting task throughput, entering shallow sleep, or entering deep sleep. Subsequently, the system can organize these actions output by the decision model into a control decision action dataset according to a pre-defined format, making it usable as input for subsequent comprehensive benefit evaluation. Ultimately, the system obtains a control decision action dataset containing various resource adjustment methods, providing a set of directly usable candidate actions for subsequently determining the optimal control decision data.

[0106] Sub-step 20413 updates the control decision action dataset based on the feasibility verification of each control decision action in the control decision action dataset.

[0107] In some embodiments of this application, to ensure that the control decision actions generated by the decision model can be reliably executed in the actual operating environment of the local wireless edge node, it is necessary to perform feasibility verification on each control decision action in the control decision action dataset. Specifically, the control decision action dataset can be updated based on the feasibility verification of the control decision actions. Feasibility verification is typically used to check whether a certain control decision action meets the system's hard rules and operational constraints, such as the node's hardware capabilities, power state switching limits, task processing capacity limits, sleep depth switching conditions, network protocol requirements, or security policy constraints, thereby ensuring that subsequent comprehensive benefit evaluation is based solely on executable candidate actions. Simultaneously, feasibility verification also serves as a pre-assessment of security. Its significance is mainly reflected in the following aspects: On the one hand, feasibility verification can identify and eliminate dangerous actions that may lead to service quality violations before decision generation, ensuring that the system always adheres to the basic service quality requirements when exploring the optimization space, thereby avoiding entering an irrecoverable violation state. On the other hand, by eliminating inactive actions in advance, feasibility verification can significantly reduce the action space required for subsequent computation, improve decision-making efficiency, and focus the model's exploration on action areas with genuine optimization potential. Furthermore, feasibility verification, as a redundant protective layer beyond constrained reinforcement learning, provides additional safety guarantees when the model has not fully converged or predictions are biased, ensuring that the system maintains basic reliability in dynamic environments. At the same time, by clearly defining the boundaries between safe and dangerous actions, feasibility verification can provide clear learning guidance for the model, accelerating its convergence to a safe and efficient policy structure.

[0108] In a specific example, the local wireless edge node has already acquired multiple control decision actions, such as increasing task processing power, decreasing task processing power, entering shallow sleep, or entering deep sleep. At this point, the feasibility of each action can be verified. For example, deep sleep actions can be excluded under high load conditions, further frequency increases can be excluded when processor utilization is near its limit, or actions that reduce throughput can be excluded when network protocols require maintaining a minimum throughput. Ultimately, the system obtains a dataset of control decision actions that can be practically executed under the current operating conditions, providing a filtered set of candidate actions for subsequent comprehensive benefit evaluation.

[0109] Sub-step 2042: Based on the comprehensive benefit assessment of each control decision action under the energy consumption benefit estimate and service quality reduction estimate, determine the control decision data.

[0110] In some embodiments of this application, in order to select a target control strategy that achieves a balance between energy consumption performance and service quality from multiple candidate control decision actions, it is necessary to comprehensively evaluate the effect of each control decision action under different benefit dimensions. Specifically, control decision data can be determined based on the comprehensive benefit evaluation of each control decision action under energy consumption benefit estimation and service quality reduction estimation. Energy consumption benefit estimation is typically used to describe the degree of energy consumption improvement of the local wireless edge node after executing a certain control decision action, while service quality reduction estimation is typically used to describe the potential impact of the action on task processing latency, throughput, or other service indicators. Comprehensive benefit evaluation is typically used to establish a weighted or constrained relationship between energy consumption benefit and service quality change, enabling the system to achieve a suitable balance between energy saving effect and service quality. This will ensure that the final determined control decision data can achieve energy consumption optimization while meeting service quality requirements, providing a reliable decision-making basis for subsequent resource adjustment.

[0111] In a specific example, the local wireless edge node has already obtained a dataset of control decision actions containing various resource adjustment methods. At this point, energy consumption gains and service quality degradation estimates can be performed for each control decision action. For instance, the potential energy increase and service quality improvement resulting from increasing task processing power can be evaluated, while the potential energy decrease and service quality degradation resulting from decreasing task processing power can be evaluated. Subsequently, based on a preset comprehensive benefit evaluation method, a joint analysis of the energy consumption gains and service quality changes for each control decision action can be performed to determine the control decision data with the highest comprehensive benefit. Ultimately, the system obtains the target control decision that achieves a balance between energy consumption performance and service quality, providing directly usable decision input for subsequent adjustments to task processing power or task throughput.

[0112] Optionally, sub-step 2042 includes the following sub-steps:

[0113] Sub-step 20421: Determine the energy consumption benefit estimate and service quality reduction estimate for each control decision action based on each control decision action and the fused input data to the local wireless edge node.

[0114] The fusion input data is obtained by embedding load status data and load prediction data under different decision dimensions.

[0115] In some embodiments of this application, to evaluate the potential effects of different control decision actions on energy consumption and service quality, it is necessary to estimate the benefits and costs of each control decision action based on fused input data. Specifically, energy consumption benefit estimates and service quality reduction estimates for each control decision action can be determined based on each control decision action and the fused input data for the local wireless edge node. The fused input data is obtained by embedding load status data and load prediction data under different decision dimensions. Energy consumption benefit estimates typically describe the degree of improvement in node energy consumption after executing a certain control decision action, while service quality reduction estimates typically describe the potential changes in task processing latency, throughput reduction, or other service metrics caused by the action. This allows the system to perform a comprehensive evaluation based on the joint relationship between benefits and costs in subsequent steps, providing the necessary foundational data for ultimately determining the optimal control decision action.

[0116] In a specific example, the local wireless edge node has already obtained fused input data and multiple candidate control decision actions. In this case, a Double Deep Q network can be used. The Direct-Depth Q-Network (DDQN) estimates the energy consumption benefit and service quality degradation for each control decision action. Considering that the fused input data already contains information such as the current state of nodes, future load trends, and their uncertainties, DDQN can use the fused input data as state input and each control decision action as action input. It estimates the long-term benefit after executing the action through the main Q-network and suppresses overestimation bias through the target Q-network. The system can decompose the Q-value output by DDQN into energy consumption benefit-related and service quality degradation-related components. For example, the positive benefit component of the Q-value can be used as the energy consumption benefit estimate, and the negative cost component as the service quality degradation estimate. Finally, the system obtains energy consumption benefit estimation data and service quality degradation estimation data for each control decision action, providing directly usable input for subsequent comprehensive benefit evaluation based on constraint costs.

[0117] Sub-step 20422: Use each service quality reduction estimate as a constraint cost on the corresponding energy consumption revenue estimate, and determine the decision evaluation value and decision evaluation data for each corresponding control decision action.

[0118] Among them, the decision evaluation value is used to characterize the risk-reward balance evaluation of each corresponding control decision action; the decision evaluation data is used to characterize the confidence level of each corresponding control decision action.

[0119] In some embodiments of this application, to establish a quantifiable risk-reward balance between energy consumption benefits and service quality, each estimated service quality reduction can be used as a constraint cost on the corresponding estimated energy consumption benefits, determining the decision evaluation value and decision evaluation data for each corresponding control decision action. The decision evaluation value characterizes the balance between benefits and risks of the action, while the decision evaluation data characterizes the confidence level of the evaluation result. By uniformly modeling benefits and costs, the system can achieve a dynamic trade-off between energy-saving effects and service quality assurance. Furthermore, the introduction of constrained reinforcement learning can enhance the reliability and optimization capability of the aforementioned modeling process: Unlike traditional reinforcement learning, which relies on the reward function to implicitly handle constraints, constrained reinforcement learning explicitly models service quality degradation as an independent constraint and achieves a rigorous mathematical balance between gains and risks through Lagrange multipliers, thus making it easier to ensure that constraints are satisfied during convergence; In addition, constrained reinforcement learning can maximize energy-saving gains within the boundaries allowed by the service quality threshold, enabling the system to explore deeper optimization potential within the "acceptable risk range"; Its adaptive Lagrange mechanism can also automatically adjust the strategy according to load fluctuations, making the system more conservative in high-risk scenarios and more energy-saving in low-risk scenarios; At the same time, by simultaneously observing the gains estimate, risk costs, and the values ​​of the Lagrange multipliers, the system's decision-making process has higher interpretability, making the control strategy more credible in key application scenarios.

[0120] In a specific example, the local wireless edge node has already obtained energy consumption benefit estimates and service quality degradation estimates for each control decision action. Lagrangian relaxation optimization can then be used to jointly model these two indices. Considering that service quality degradation is a mandatory constraint, a Lagrangian relaxation objective of the following form can be constructed for each control decision action: J = R - λC, where R represents the energy consumption benefit estimate, C represents the service quality degradation estimate, and λ is a dynamically adjusted Lagrange multiplier based on system policy or historical experience, used to control the penalty strength of the service quality constraint. The system can determine the risk-reward balance evaluation based on the J value of different actions and use this value as the decision evaluation value.

[0121] Sub-step 20423: Determine the target control decision action from the control decision action dataset based on the decision evaluation value of each control decision action, and generate control decision data based on the target control decision action and the corresponding decision evaluation data.

[0122] In some embodiments of this application, in order to select the target control strategy that achieves the optimal balance between energy consumption benefits and service quality risks from multiple candidate control decision actions, it is necessary to filter the control decision action dataset based on the obtained decision evaluation values. Specifically, the target control decision action can be determined from the control decision action dataset based on the decision evaluation value of each control decision action, and control decision data is generated based on the target control decision action and the corresponding decision evaluation data. The decision evaluation value is usually used to characterize the comprehensive performance of a control decision action in terms of risk-reward balance, while the decision evaluation data is usually used to characterize the confidence level of the evaluation result. By considering both the evaluation value and the confidence level, the system can achieve robust decision-making between maximizing benefits and controlling risk, so that the final generated control decision data can maintain reliability and interpretability in actual operation.

[0123] In a specific example, the local wireless edge node has already obtained the decision evaluation value and decision evaluation data for each control decision action. At this point, all candidate actions can be sorted based on the decision evaluation values, and the action with the highest evaluation value is selected as the target control decision action. For example, when an action obtains a high risk-reward balance evaluation under Lagrange relaxation optimization, and its corresponding decision evaluation data (such as confidence level) is also at a high level, the system can identify this action as the target control decision action. Subsequently, the final control decision data can be generated based on this target control decision action and its corresponding decision evaluation data, so that the control decision data not only includes the action itself but also a quantitative description of the reliability of the action. Ultimately, the system obtains control decision data that can achieve a robust balance between energy consumption optimization and service quality assurance, providing a directly executable decision basis for the resource adjustment of the local wireless edge node.

[0124] Optionally, in order to generate control decision data based on the target control decision actions and the corresponding decision evaluation data, sub-step 20423 includes the following sub-steps:

[0125] Sub-step 204231: Obtain the target decision evaluation value and target decision evaluation data corresponding to the target control decision action, as well as the fused input data.

[0126] In some embodiments of this application, in order to generate final executable control decision data based on the target control decision action, it is necessary to first obtain key evaluation information related to the target control decision action. Specifically, this can involve obtaining the target decision evaluation value and target decision evaluation data corresponding to the target control decision action, as well as fused input data. The target decision evaluation value is typically used to characterize the comprehensive balance between energy consumption benefits and service quality risks of the target control decision action, while the target decision evaluation data is typically used to characterize the confidence level of the evaluation result. The fused input data is used to provide contextual information about the current state of the node and future load trends, enabling the finally generated control decision data to semantically and completely describe the basis for action selection. This allows the system to record the action itself, the evaluation result, and the contextual information together as control decision data in subsequent steps, providing an interpretable and traceable decision basis for resource adjustment of local wireless edge nodes.

[0127] In a specific example, the local wireless edge node has already obtained decision evaluation values ​​and data for multiple control decision actions based on Lagrange relaxation optimization, and selected a target control decision action based on the evaluation values. At this point, the target decision evaluation value and data corresponding to the target action can be extracted from the obtained evaluation results. Simultaneously, information such as the node's current state, future load trends, and uncertainties can be obtained from the fused input data retained during the feature engineering stage. Ultimately, the system obtains all the necessary information for generating control decision data, providing an input foundation for subsequently recording actions, evaluation values, confidence levels, and contextual features together as control decision data.

[0128] Sub-step 204232: Record at least one of the target decision evaluation value and the fused input data, together with the target decision evaluation data and the target control decision action, as control decision data.

[0129] In some embodiments of this application, to ensure that the final generated control decision data fully expresses the basis for selecting the target control decision action, it is necessary to uniformly record the action itself, its corresponding evaluation information, and contextual features. Specifically, at least one of the target decision evaluation value and the fused input data can be recorded together with the target decision evaluation data and the target control decision action as control decision data. The target decision evaluation value is typically used to characterize the comprehensive balance between energy consumption benefits and service quality risks of the action; the target decision evaluation data is typically used to characterize the confidence level of the evaluation result; and at least one feature in the fused input data is used to provide contextual information when the action is selected, making the control decision data interpretable and traceable. This will enable the final generated control decision data to accurately reflect the basis and background of the action selection in subsequent execution or auditing stages.

[0130] In a specific example, the local wireless edge node has already acquired the target control decision action, the target decision evaluation value, the target decision evaluation data, and the fused input data. At this point, the target decision evaluation value (e.g., a risk-reward balance score obtained based on Lagrange relaxation optimization), at least one feature from the fused input data (e.g., the short-term mean of future load or a predictive uncertainty indicator), the target decision evaluation data (e.g., a confidence level obtained based on DDQN stability), and the target control decision action (e.g., entering shallow sleep or increasing task processing power) can be collectively recorded as control decision data. Ultimately, the system obtains control decision data containing actions, evaluation values, confidence levels, and contextual features, enabling this data to serve as a complete basis for subsequent execution of control strategies or for strategy auditing.

[0131] Step 205: Adjust the task processing power and / or task throughput of the local wireless edge node based on the control decision data.

[0132] The method shown in this step has been explained in step 103 and will not be repeated here.

[0133] Optionally, step 205 includes the following sub-steps:

[0134] Sub-step 2051: Determine the target control decision action for the local wireless edge node from the control decision data, and obtain the real-time load policy of the local wireless edge node.

[0135] In some embodiments of this application, in order for the local wireless edge node to implement the control policy already generated in the actual operating environment, it is necessary to extract the final executable target control decision action from the control decision data and adjust it in conjunction with the node's current operating policy. Specifically, the target control decision action for the local wireless edge node can be determined from the control decision data, and the real-time load policy of the local wireless edge node can be obtained. The target control decision action is typically used to characterize the resource adjustment method that the system should take at the current moment, while the real-time load policy is typically used to describe the node's load management rules, task scheduling policies, or power state constraints in the current operating cycle. By combining the target control decision action with the real-time load policy, the system can ensure that the subsequent resource adjustment process conforms to the prediction-driven optimization direction and meets the node's current operating constraints and service quality requirements.

[0136] In a specific example, the local wireless edge node has already obtained control decision data containing actions, evaluation values, confidence levels, and contextual features. From this data, the system can extract target control decision actions, such as increasing task processing power, decreasing task processing power, adjusting task throughput, or entering shallow sleep. Simultaneously, the system can obtain real-time load policies from the node's operation management module, such as current task scheduling priorities, power state switching limits, minimum quality of service requirements, or network throughput constraints. Ultimately, the system obtains the target control decision actions and real-time load policies, providing the necessary input foundation for subsequent resource adjustments based on their joint execution.

[0137] Sub-step 2052: Adjust the task processing power and / or task throughput based on the target control decision action and real-time load strategy to update the load status data of the local wireless edge node.

[0138] In some embodiments of this application, in order for the local wireless edge node to perform resource adjustment based on target control decision actions, it is necessary to dynamically adjust the task processing power and / or task throughput in conjunction with the node's current real-time load policy. Specifically, the task processing power and / or task throughput can be adjusted based on the target control decision actions and the real-time load policy to update the load status data of the local wireless edge node. Task processing power adjustment is typically used to change the operating frequency of the processor cores or the number of available cores, while task throughput adjustment is typically used to change the task scheduling rate, packet processing rate, or available bandwidth of the wireless link. By following the real-time load policy during resource adjustment, the system can ensure that the adjustment behavior conforms to the prediction-driven optimization direction without violating the node's current quality of service constraints or power state limitations.

[0139] In a specific example, the local wireless edge node has received a target control decision action, such as increasing task processing power or decreasing task throughput. At this point, the system can adjust the node's resource parameters according to the real-time load strategy. For instance, if the target control decision action requires a direct increase in task processing power from a low-power state to a high-power state, but hardware limitations or power state switching rules prevent this from being completed in one go, the system can use a pre-adjustment transition approach to gradually achieve the target adjustment: first, increase the task processing power to an intermediate power level, and then, after meeting stability conditions, further increase it to the target power level. Similarly, if the target action requires a significant reduction in task throughput, but current network protocols or quality of service constraints do not allow for an immediate and substantial decrease, the system can also achieve a smooth transition by reducing throughput in stages. Ultimately, the system can adjust task processing power and / or task throughput while adhering to the real-time load strategy, and record the adjusted operating state as new load state data, providing the latest operational basis for subsequent prediction, evaluation, and decision-making.

[0140] Optionally, the above sub-steps 20421 and 20422 are implemented by a joint decision network. That is, the joint decision network is used to determine the energy consumption benefit estimation data and service quality reduction estimation data for the control decision action based on the control decision action and the fused input data for the local wireless edge node, and to use each service quality reduction estimation data as a constraint cost for the corresponding energy consumption benefit estimation data to determine the decision evaluation value and decision evaluation data for each corresponding control decision action.

[0141] The fusion input data is obtained by embedding load status data and load prediction data under different decision dimensions; the decision evaluation value is used to characterize the risk-reward balance evaluation of each corresponding control decision action; and the decision evaluation data is used to characterize the confidence level of each corresponding control decision action.

[0142] At this point, the computational offloading joint optimization method based on wireless edge networks also includes the following additional steps for adjusting the joint decision network:

[0143] Step 206: Determine the risk identification data corresponding to the target control decision action from the control decision data.

[0144] Among them, the risk identification data includes target decision evaluation data for target control decision actions; the target decision evaluation data is used to characterize the degree of confidence in the target control decision actions.

[0145] In some embodiments of this application, in order to continuously calibrate the inference results of the joint decision-making network and enable it to more accurately reflect the risk characteristics of the actual operating environment during subsequent training or online updates, it is necessary to extract risk information related to the target control decision action from the control decision data. Specifically, risk identification data corresponding to the target control decision action can be determined from the control decision data. This risk identification data includes target decision evaluation data for the target control decision action, which is typically used to characterize the confidence level of the target control decision action in the aforementioned comprehensive benefit assessment. By extracting this confidence level information, the system can use it as a risk feedback signal in subsequent steps to guide the updating of the joint decision-making network, enabling the network to better identify the differences between high-risk and low-risk actions in future inference processes.

[0146] In a specific example, the local wireless edge node has already selected a target control decision action based on a comprehensive benefit assessment and generated control decision data containing the action, evaluation value, confidence level, and contextual features. At this point, the system can extract the target decision evaluation data corresponding to the target action from the control decision data, such as a confidence index formed by Lagrange relaxation optimization and DDQN inference. This confidence index can serve as risk identification data, characterizing the risk level of the target control decision action under current operating conditions. Ultimately, the system obtains risk identification data that can be used for subsequent risk analysis and strategy calibration, enabling subsequent model adjustments to be optimized based on more accurate risk feedback.

[0147] Step 207: Determine the energy consumption assessment data for the target control decision action based on the load status data before the update, the load status data after the update, and the target control decision action.

[0148] Among them, energy consumption assessment data is used to characterize the energy consumption estimate of the target control decision-making action and the compliance level of the action execution.

[0149] In some embodiments of this application, to accurately assess the energy consumption performance of target control decision actions during actual execution, it is necessary to combine the changes in the node's operating state before and after the action execution, and to perform fine-grained measurement of energy consumption during the action execution. Specifically, based on the load state data before the update, the load state data after the update, and the target control decision action, energy consumption assessment data describing the energy consumption performance and execution compliance of the action can be determined. This energy consumption assessment data reflects both the energy consumption changes brought about by the action execution and whether the execution of the action under power state switching rules, task scheduling strategies, and service quality constraints meets system requirements. By jointly analyzing the state changes and energy consumption measurement results, the system can obtain an accurate energy consumption assessment of the target control decision action, providing a reliable basis for subsequent risk identification and strategy calibration.

[0150] In a specific example, the local wireless edge node has already executed a target control decision action, such as switching from a low-power state to a high-power state. The system activates the energy metering unit the instant the action command is issued, performing high-speed, synchronous sampling of the voltage and current of each power rail, ensuring the sampling timestamp is consistent with the system's global clock. Subsequently, the system performs numerical integration on the instantaneous power data collected from the start to the completion of the switch, for example, using the trapezoidal rule to calculate the total energy consumption of the state transition process. After the node enters the target power state, the system continues to collect power data for a period of time, calculates the average power during this phase, and combines this with the node's dwell time in this state to obtain the energy required to maintain that state. The system associates the switching energy consumption and steady-state energy consumption with the corresponding power state and time window, respectively, forming a complete energy consumption record. Based on this, the system also combines load state data before and after the action execution, such as processor core utilization, task queue length, or wireless link throughput, to determine whether the action execution process complies with power state switching rules, task scheduling strategies, or network protocol requirements, thereby determining the compliance level of the action execution. The resulting energy consumption assessment data can simultaneously reflect energy consumption performance and compliance, providing highly reliable input for subsequent strategy optimization.

[0151] Step 208: Generate report data on target control decision actions based on risk identification data and energy consumption assessment data.

[0152] In some embodiments of this application, in order to uniformly express the risk performance and energy consumption performance of target control decision actions during actual execution, it is necessary to generate structured report data based on the aforementioned risk identification data and energy consumption assessment data. Specifically, report data for target control decision actions can be generated based on the risk identification data and energy consumption assessment data. The report data is typically used to comprehensively present the confidence level, risk level, energy consumption performance, and compliance of the action, enabling the system to obtain complete, traceable, and quantifiable feedback information during subsequent strategy calibration or model updates. By uniformly organizing risk information and energy consumption information, the system can form a comprehensive evaluation record of target control decision actions.

[0153] In a specific example, the local wireless edge node has already obtained risk identification data corresponding to the target control decision action, such as confidence indices formed from decision evaluation data, and energy consumption assessment data during the action execution process, such as switching energy consumption, steady-state energy consumption, and execution compliance results. The system can then integrate this information according to a preset format, enabling the report data to simultaneously reflect the risk-reward balance of the action, the energy consumption performance during execution, and whether the action meets the system's operational strategy. The final generated report data can serve as input for subsequent strategy optimization and joint decision-making network updates, allowing the system to continuously improve decision quality based on real execution feedback.

[0154] Step 209: Update the joint decision-making network based on the reported data.

[0155] In some embodiments of this application, to enable the joint decision-making network to continuously improve its ability to predict the benefits and identify the risks of control decisions based on actual execution feedback, it is necessary to update the joint decision-making network using the aforementioned generated report data. Specifically, the joint decision-making network can be updated based on the report data. The report data typically includes risk identification information and energy consumption assessment information for the target control decision action, reflecting the risk level, energy consumption performance, and compliance of the action in the real operating environment. By using this feedback information for network updates, the system can gradually correct the deviations of the joint decision-making network in benefit estimation, risk modeling, and strategy balancing, enabling it to generate more robust and reliable control decisions in subsequent decision-making cycles.

[0156] In a specific example, the local wireless edge node has already generated report data containing risk identification data and energy consumption assessment data. The system can then input this report data into the update module of the joint decision-making network, enabling the network to adjust its internal parameters based on real-world feedback during action execution. For instance, when the report data reflects that the risk level of a certain action is higher than expected or that energy consumption performance deviates from prediction, the system can use the update mechanism to correct the network's benefit estimation or risk modeling method for that action. Ultimately, the joint decision-making network can gradually improve decision quality through continuous feedback loops, making subsequent control decisions more aligned with actual operational needs.

[0157] In one implementation, the joint decision network can be updated using the parameter optimization mechanism of deep Q-networks. The system can update the policy structure of the constrained deep reinforcement learning model based on feature vectors, decision instructions, and performance monitoring reports, through experience replay and online learning algorithms.

[0158] During the policy network update process, the system randomly samples historical transition data from the experience pool, including the state before the action is executed, the action taken, the immediate reward, the state after the action is executed, and constraint violation flags. The target value is calculated using a double-Q learning framework: the main network selects the action with the highest predicted reward in the next state, the target network evaluates the value of that action, and the constraint network provides a risk estimate for that action, ensuring the target value includes both short-term reward and long-term risk-adjusted value. The system optimizes the parameters of the main network through backpropagation and gradient descent, enabling it to more accurately predict future rewards.

[0159] During the update process of the constraint network, the system calibrates the risk model based on the risk events recorded in the report data. If no delay or service quality degradation occurs during task execution, the constraint target value is zero; if a delay or timeout occurs, the constraint target value is set according to the severity of the violation. The system strengthens the learning of real risk events through a weighted loss function, enabling the constraint network to identify potential risk patterns such as load fluctuations, queue backlogs, or link fluctuations, and robustly updates the network parameters through gradient descent.

[0160] Furthermore, the system can adaptively adjust the Lagrange multipliers based on recent constraint violation rates. When the violation rate exceeds a tolerance threshold, the system increases the Lagrange multipliers to strengthen the penalty for risk; when the violation rate is low, the system adjusts the multipliers in a more lenient manner, maintaining a reasonable balance between risk and reward. The new Lagrange multipliers are only used for new metadata generated in subsequent decision-making cycles to ensure the traceability and interpretability of historical records.

[0161] It should also be noted that the aforementioned prediction network for generating load forecasting data, the decision model for performing risk-reward balance modeling, and the various functional modules for data association, action selection, strategy execution, feedback updates, and adaptive adjustment can all be implemented in the form of a computer program product. This computer program product may include complex neural network models, agent-based deep reinforcement learning models, constrained optimization models, feature extraction models, temporal modeling networks, and auxiliary logic modules for performing inference, update, and control instruction generation. The aforementioned computer program product can be stored in a non-transitory computer-readable storage medium and loaded and executed in software on a processor, accelerator, or other programmable hardware.

[0162] During execution, the computer program product can invoke its internal neural network inference engine, policy network, value network, constraint network, Lagrange multiplier update module, data preprocessing module, time-series alignment module, feature fusion module, and action generation module to achieve the functions described in this application, such as load prediction, service quality risk estimation, energy consumption revenue modeling, constraint cost calculation, policy evaluation, action selection, and execution control. Through this programmatic implementation, the system can complete complex inference and decision-making processes in a software-defined manner in dynamic wireless edge environments without relying on specific hardware logic, thus possessing good portability, scalability, and maintainability.

[0163] The specific implementation methods of the aforementioned computer program products, the construction methods of neural networks, the model training process, the parameter update mechanism, the inference execution method, and the details of the program's operation on the processor are all contents that have been fully disclosed and widely used in the prior art, and will not be repeated in this application.

[0164] Furthermore, it should be emphasized that the mechanisms related to data collection, data processing, privacy protection, and legality assurance involved in this application can all be implemented based on existing data governance systems, industry regulatory standards, and common information security technologies. The specific implementation methods of these mechanisms may include, but are not limited to, existing mature technical solutions such as authorization management, access control, data anonymization, encrypted transmission, log auditing, and compliance verification. The above content represents commonly adopted engineering practices in this field, and its implementation paths have been widely disclosed and applied in related technical fields, and do not constitute a technical innovation point of this application. Therefore, this application does not limit its specific implementation methods.

[0165] In summary, in this embodiment, the state sequence data of the local wireless edge node under multiple state dimensions is first processed to determine the load state data and load prediction data. This allows control decisions to no longer rely on the state input at a single moment, but to make judgments based on multi-dimensional and continuous state change trends. The prediction results can reflect the uncertainty of future load to a certain extent, thereby reducing resource adjustment delays or over-adjustments caused by incomplete state information and providing a more reliable reference for subsequent decisions. Furthermore, control decision data can be determined based on the load state data and load prediction data, enabling control decisions to simultaneously consider the current load level. By considering future load trends to enhance decision-making robustness, the control strategy becomes more adaptable to load fluctuations, achieving a more reasonable balance between energy consumption benefits and service quality, and reducing decision imbalances caused by prediction biases or state fluctuations. Ultimately, based on control decision data, the task processing power and throughput of local wireless edge nodes are adjusted, allowing resource scheduling to be adjusted in advance according to predicted trends, reducing the risk of insufficient or wasted resources. This enables wireless edge nodes to maintain relatively stable service quality under dynamic load environments while improving the energy efficiency of the wireless edge network, thus achieving a more optimized trade-off between energy consumption and performance. Therefore, based on the method of this application embodiment, control decisions can be generated on the basis of more comprehensive state information and more reliable prediction results, thereby improving the situation where energy consumption benefits and service quality are difficult to balance due to delayed resource adjustments or unreliable predictions, further enhancing the computational offloading and resource management effects of wireless edge nodes.

[0166] refer to Figure 3 This illustrates a computation offloading joint optimization system 30 based on a wireless edge network, provided in an embodiment of this application, applied to a local wireless edge node, comprising:

[0167] The prediction module 301 is used to determine the load state data and load prediction data of the local wireless edge node based on the state sequence data of the local wireless edge node under multiple state dimensions; the load prediction data is used to characterize the prediction result of the load state change process of the local wireless edge node under the target confidence interval; the target confidence interval matches the preset confidence conditions.

[0168] Decision module 302 is used to determine control decision data for local wireless edge nodes based on load status data and load prediction data;

[0169] The execution module 303 is used to adjust the task processing power and / or task throughput of the local wireless edge node based on control decision data.

[0170] Optionally, the prediction module 301 includes:

[0171] The temporal fusion submodule is used to fuse the state sequence data of the local wireless edge node under multiple state dimensions according to temporal correlation to obtain the fused sequence data of the local wireless edge node.

[0172] The multiple prediction submodule is used to repeatedly input the fused sequence data into the prediction network according to a preset number of predictions to obtain multiple prediction sequence data for the local wireless edge node; during each input of the fused sequence data into the prediction network, at least one neuron in the prediction network is randomly masked.

[0173] The analysis and prediction submodule is used to obtain load prediction data based on the averaging of multiple prediction sequence data, and to obtain the target confidence interval based on the statistical characteristic analysis of the load prediction data.

[0174] The Status Snapshot submodule is used to obtain load status data based on time slice processing of fused sequence data at a target time or time interval.

[0175] Optionally, the timing fusion submodule includes:

[0176] The timing alignment unit is used to align the timing data of each state sequence data to obtain the aligned sequence data corresponding to each state sequence data.

[0177] The temporal fusion unit is used to combine all aligned sequence data into a splicing tensor as fused sequence data.

[0178] Optionally, the timing alignment unit includes:

[0179] The downsampling subunit is used to extract state sequence sub-data from the state sequence data according to the sampling grid width when the sampling density of the state sequence data is greater than or equal to the preset sampling grid width, so as to use the resampled data of the state sequence data;

[0180] The upsampling subunit is used to interpolate and supplement the state sequence data according to the sampling grid width when the sampling density of the state sequence data is less than the preset sampling grid width, so as to obtain resampled data of the state sequence data.

[0181] The alignment subunit is used to align each resampled data point along the time axis based on temporal correlation analysis of all resampled data, so as to obtain aligned sequence data corresponding to each state sequence data.

[0182] Optionally, the analysis and prediction submodule includes:

[0183] A multi-order estimation unit is used to determine multi-order estimates of local wireless edge nodes based on load prediction data.

[0184] The confidence estimation unit is used to determine the target confidence interval based on the multi-order estimators and the statistical distribution model corresponding to the state sequence data of the local wireless edge node.

[0185] Optionally, decision module 302 includes:

[0186] The decision generation submodule is used to determine the control decision action dataset for the local wireless edge node based on load status data and load prediction data; the control decision action dataset contains multiple control decision actions for the local wireless edge node.

[0187] The decision screening submodule is used to determine control decision data based on the comprehensive benefit assessment of each control decision action under the estimation of energy consumption benefits and the estimation of service quality reduction.

[0188] Optionally, the decision generation submodule includes:

[0189] The embedding fusion unit is used to determine the fusion input data for the local wireless edge node based on the data embedding of load status data and load prediction data under different decision dimensions.

[0190] The decision generation unit is used to input the fused input data into the decision model to form a control decision action dataset based on the obtained multiple control decision actions.

[0191] Optionally, the decision generation submodule also includes:

[0192] The feasibility verification unit is used to update the control decision action dataset based on the feasibility verification of each control decision action in the control decision action dataset.

[0193] Optionally, the decision filtering submodule includes:

[0194] The dual decision parameter unit is used to determine the energy consumption benefit estimate and service quality reduction estimate for each control decision action based on each control decision action and the fused input data for the local wireless edge node, respectively; the fused input data is obtained by embedding load status data and load prediction data under different decision dimensions;

[0195] The decision evaluation unit is used to take each service quality reduction estimate as a constraint cost on the corresponding energy consumption revenue estimate, and to determine the decision evaluation value and decision evaluation data for each corresponding control decision action; the decision evaluation value is used to characterize the risk-reward balance evaluation for each corresponding control decision action; the decision evaluation data is used to characterize the confidence level for each corresponding control decision action.

[0196] The decision filtering unit is used to determine the target control decision action from the control decision action dataset based on the decision evaluation value of each control decision action, and to generate control decision data based on the target control decision action and the corresponding decision evaluation data.

[0197] Optionally, the decision filtering unit includes:

[0198] The acquisition subunit is used to acquire the target decision evaluation value and target decision evaluation data corresponding to the target control decision action, as well as the fused input data;

[0199] The packaging subunit is used to record at least one of the target decision evaluation value and the fused input data, together with the target decision evaluation data and the target control decision action, as control decision data.

[0200] Optionally, execution module 303 includes:

[0201] The extraction submodule is used to determine the target control decision action for the local wireless edge node from the control decision data and to obtain the real-time load policy of the local wireless edge node.

[0202] The adjustment submodule is used to adjust the task processing power and / or task throughput based on the target control decision action and real-time load strategy in order to update the load status data of the local wireless edge node.

[0203] Optionally, the compute offloading joint optimization system 30 based on wireless edge networks also includes:

[0204] The first extraction module is used to determine the risk identification data corresponding to the target control decision action from the control decision data; the risk identification data includes target decision evaluation data for the target control decision action; the target decision evaluation data is used to characterize the confidence level of the target control decision action;

[0205] The second extraction module is used to determine the energy consumption assessment data for the target control decision actions based on the load status data before the update, the load status data after the update, and the target control decision actions; the energy consumption assessment data is used to characterize the energy consumption estimate and the compliance level of the action execution for the target control decision actions.

[0206] The data packaging module generates report data for target control decision actions based on risk identification data and energy consumption assessment data. This report data is used to update the joint decision-making network. The joint decision-making network determines energy consumption benefit estimates and service quality reduction estimates for each control decision action based on the control decision action and fused input data from local wireless edge nodes. It also uses each service quality reduction estimate as a constraint cost on the corresponding energy consumption benefit estimate to determine the decision evaluation value and decision evaluation data for each corresponding control decision action. The fused input data is obtained by embedding load status data and load forecast data under different decision dimensions. The decision evaluation value characterizes the risk-benefit balance assessment for each corresponding control decision action. The decision evaluation data characterizes the confidence level for each corresponding control decision action.

[0207] Optionally, the compute offloading joint optimization system 30 based on wireless edge networks also includes:

[0208] The data acquisition module is used to acquire multiple sets of status data of the local wireless edge node within a target time interval according to multiple status dimensions.

[0209] The data cleaning module is used to clean each set of state-collected data to obtain state sequence data of the local wireless edge node in each state dimension.

[0210] In summary, in this embodiment, the state sequence data of the local wireless edge node under multiple state dimensions is first processed to determine the load state data and load prediction data. This allows control decisions to no longer rely on the state input at a single moment, but to make judgments based on multi-dimensional and continuous state change trends. The prediction results can reflect the uncertainty of future load to a certain extent, thereby reducing resource adjustment delays or over-adjustments caused by incomplete state information and providing a more reliable reference for subsequent decisions. Furthermore, control decision data can be determined based on the load state data and load prediction data, enabling control decisions to simultaneously consider the current load level. By considering future load trends to enhance decision-making robustness, the control strategy becomes more adaptable to load fluctuations, achieving a more reasonable balance between energy consumption benefits and service quality, and reducing decision imbalances caused by prediction biases or state fluctuations. Ultimately, based on control decision data, the task processing power and throughput of local wireless edge nodes are adjusted, allowing resource scheduling to be adjusted in advance according to predicted trends, reducing the risk of insufficient or wasted resources. This enables wireless edge nodes to maintain relatively stable service quality under dynamic load environments while improving the energy efficiency of the wireless edge network, thus achieving a more optimized trade-off between energy consumption and performance. Therefore, based on the method of this application embodiment, control decisions can be generated on the basis of more comprehensive state information and more reliable prediction results, thereby improving the situation where energy consumption benefits and service quality are difficult to balance due to delayed resource adjustments or unreliable predictions, further enhancing the computational offloading and resource management effects of wireless edge nodes.

[0211] Reference Figure 4 The electronic device 500 may include one or more of the following components: processing component 502, memory 504, power supply component 506, multimedia component 508, audio component 510, input / output (I / O) interface 512, sensor component 514, and communication component 516.

[0212] Processing component 502 typically controls the overall operation of electronic device 500, such as operations associated with display, telephone calls, data communication, camera operation, and recording. Processing component 502 may include one or more processors 520 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processing component 502 may include one or more modules to facilitate interaction between processing component 502 and other components. For example, processing component 502 may include a multimedia module to facilitate interaction between multimedia component 508 and processing component 502.

[0213] Memory 504 is used to store various types of data to support the operation of electronic device 500. Examples of this data include instructions for any application or method operating on electronic device 500, contact data, phonebook data, messages, pictures, multimedia, etc. Memory 504 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0214] Power supply component 506 provides power to various components of electronic device 500. Power supply component 506 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 500.

[0215] Multimedia component 508 includes an interface that provides an output interface between electronic device 500 and user. In some embodiments, the interface may include a liquid crystal display (LCD) and a touch panel (TP). If the interface includes a touch panel, the interface may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may not only sense the boundaries of touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 508 includes a front-facing camera and / or a rear-facing camera. When electronic device 500 is in an operating mode, such as shooting mode or multimedia mode, the front-facing camera and / or rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0216] Audio component 510 is used to output and / or input audio signals. For example, audio component 510 includes a microphone (MIC) used to receive external audio signals when electronic device 500 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 504 or transmitted via communication component 516. In some embodiments, audio component 510 also includes a speaker for outputting audio signals.

[0217] Input / output (I / O) interface 512 provides an interface between processing component 502 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0218] Sensor assembly 514 includes one or more sensors for providing state assessments of various aspects of electronic device 500. For example, sensor assembly 514 may detect the on / off state of electronic device 500, the relative positioning of components such as the display and keypad of electronic device 500, changes in position of electronic device 500 or a component of electronic device 500, the presence or absence of user contact with electronic device 500, orientation or acceleration / deceleration of electronic device 500, and temperature changes of electronic device 500. Sensor assembly 514 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 514 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 514 may also include an accelerometer, gyroscope, magnetometer, pressure sensor, or temperature sensor.

[0219] Communication component 516 facilitates wired or wireless communication between electronic device 500 and other devices. Electronic device 500 can access wireless networks based on communication standards, such as WiFi, carrier networks (such as 2G, 3G, 4G, or 5G), or combinations thereof. In one exemplary embodiment, communication component 516 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 516 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0220] In an exemplary embodiment, the electronic device 500 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to implement the methods provided in the embodiments of this application.

[0221] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 504 including instructions, which can be executed by a processor 520 of an electronic device 500 to perform the above-described method. For example, the non-transitory storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0222] In an exemplary embodiment, the electronic device 500 may also be provided as a server, including a processing component 502, which further includes one or more processors, and memory resources represented by memory 504 for storing instructions, such as applications, that can be executed by the processing component 502. The applications stored in memory 504 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 502 is configured to execute instructions to perform the methods provided in the embodiments of this application.

[0223] Electronic device 500 may also include a power supply component 506 configured to perform power management of electronic device 500, a wired or wireless communication component 516 configured to connect electronic device 500 to a network, and an input / output (I / O) interface 512. Electronic device 500 may operate on an operating system stored in memory 504, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or similar.

[0224] This application also provides a computer program product, including a computer program, which, when executed by a processor, implements the method of this application embodiment.

[0225] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims below.

[0226] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the claimed technical solutions.

[0227] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0228] It will be readily apparent to those skilled in the art that any combination of the above embodiments is feasible. Therefore, any combination of the above embodiments is an implementation scheme of this application. However, due to space limitations, this specification will not describe them in detail here.

[0229] It should be noted that, unless otherwise expressly stated, the methods provided in the embodiments of this application are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. Based on the above description, the required structure for constructing a system having the solutions of this application is obvious. Furthermore, this application is not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0230] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0231] Similarly, it should be understood that, for the purpose of simplification and aiding understanding of one or more aspects of the application, various features of the application are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the application. However, this disclosure should not be construed as reflecting an intention that the claimed application requires more features than are expressly recited in each of the claimed technical solutions. Rather, as reflected in the claimed technical solutions, the application aspects comprise fewer features than all those in the single embodiment disclosed above. Therefore, the claimed technical solutions following the specific implementation are thus expressly incorporated into that specific implementation, wherein each claimed technical solution is itself a separate embodiment of the application.

[0232] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this application and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this application may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0233] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claimed technical solutions, any one of the claimed embodiments can be used in any combination.

[0234] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components in the methods according to the embodiments of this application. This application can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0235] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the method of the embodiments of the present application.

[0236] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0237] It should be noted that, for the sake of simplicity, the method embodiments of this application are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of this application.

[0238] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims below.

[0239] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the claimed technical solutions.

[0240] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A joint optimization method for computational offloading based on wireless edge networks, characterized in that, Applied to local wireless edge nodes, including: Based on the state sequence data of the local wireless edge node under multiple state dimensions, load state data and load prediction data for the local wireless edge node are determined; the load prediction data is used to characterize the prediction result of the load state change process of the local wireless edge node under a target confidence interval; the target confidence interval matches a preset confidence condition. Based on the load status data and the load prediction data, control decision data for the local wireless edge node is determined; Adjust the task processing power and / or task throughput of the local wireless edge node based on the control decision data; Adjusting the task processing power and / or task throughput of the local wireless edge node based on the control decision data includes: The target control decision action for the local wireless edge node is determined from the control decision data, and the real-time load strategy of the local wireless edge node is obtained. Based on the target control decision action and the real-time load strategy, the task processing power and / or the task throughput are adjusted to update the load status data of the local wireless edge node. The computation offloading joint optimization method based on wireless edge networks also includes: Risk identification data corresponding to the target control decision action is determined from the control decision data; the risk identification data includes target decision evaluation data for the target control decision action. Based on the load status data before the update, the load status data after the update, and the target control decision action, determine the energy consumption assessment data for the target control decision action; Based on the risk identification data and energy consumption assessment data of the target control decision actions, report data for the target control decision actions is generated; the report data is used to update the joint decision network; the joint decision network is used to determine the decision evaluation value and decision evaluation data for each control decision action.

2. The computational offloading joint optimization method based on wireless edge networks as described in claim 1, characterized in that, The process of determining the load status data and load prediction data for the local wireless edge node based on the state sequence data of the local wireless edge node across multiple state dimensions includes: The state sequence data of the local wireless edge node under multiple state dimensions are fused according to temporal correlation to obtain the fused sequence data of the local wireless edge node; The fused sequence data is repeatedly input into the prediction network according to a preset number of predictions to obtain multiple prediction sequence data for the local wireless edge node; during each input of the fused sequence data into the prediction network, at least one neuron in the prediction network is randomly masked. The load prediction data is obtained by averaging multiple predicted sequence data, and the target confidence interval is obtained based on the statistical characteristic analysis of the load prediction data.

3. The computational offloading joint optimization method based on wireless edge networks as described in claim 1, characterized in that, The process of determining the load status data and load prediction data for the local wireless edge node based on the state sequence data of the local wireless edge node across multiple state dimensions includes: The state sequence data of the local wireless edge node under multiple state dimensions are fused according to temporal correlation to obtain the fused sequence data of the local wireless edge node; The load status data is obtained by processing the fused sequence data in time slices at the target time or time interval.

4. The computational offloading joint optimization method based on wireless edge networks as described in claim 1, characterized in that, The step of determining control decision data for the local wireless edge node based on the load status data and the load prediction data includes: The control decision action dataset for the local wireless edge node is determined based on the load status data and the load prediction data; the control decision action dataset contains multiple control decision actions for the local wireless edge node. The control decision data is determined based on a comprehensive benefit assessment of each control decision action under the estimation of energy consumption benefits and service quality reduction.

5. The computational offloading joint optimization method based on wireless edge networks as described in claim 1, characterized in that, The computation offloading joint optimization method based on wireless edge networks also includes: According to the multiple state dimensions, acquire multiple sets of state collection data of the local wireless edge node in the target time interval; Data cleaning is performed on each set of state acquisition data to obtain state sequence data of the local wireless edge node under each state dimension.

6. A computational offloading joint optimization system based on a wireless edge network, characterized in that, Applied to local wireless edge nodes, including: The prediction module is used to determine the load state data and load prediction data of the local wireless edge node based on the state sequence data of the local wireless edge node under multiple state dimensions; the load prediction data is used to characterize the prediction result of the load state change process of the local wireless edge node under a target confidence interval; the target confidence interval matches a preset confidence condition. The decision module is used to determine control decision data for the local wireless edge node based on the load status data and the load prediction data. An execution module is configured to adjust the task processing power and / or task throughput of the local wireless edge node based on the control decision data. The execution module includes: An extraction submodule is used to determine the target control decision action for the local wireless edge node from the control decision data, and to obtain the real-time load policy of the local wireless edge node. The adjustment submodule is used to adjust the task processing power and / or the task throughput based on the target control decision action and the real-time load strategy, so as to update the load status data of the local wireless edge node. The computation offloading joint optimization system based on wireless edge networks also includes: The first extraction module is used to determine risk identification data corresponding to the target control decision action from the control decision data; the risk identification data includes target decision evaluation data for the target control decision action. The second extraction module is used to determine energy consumption assessment data for the target control decision action based on the load status data before the update, the load status data after the update, and the target control decision action. A data packaging module is used to generate report data for the target control decision actions based on risk identification data and energy consumption assessment data; the report data is used to update the joint decision network; the joint decision network is used to determine the decision evaluation value and decision evaluation data for each control decision action.

7. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the computation offloading joint optimization method based on a wireless edge network as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the computation offloading joint optimization method based on a wireless edge network as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product stores a computer program, which, when executed by a processor, implements the steps of the computation offloading joint optimization method based on a wireless edge network as described in any one of claims 1 to 5.

Images